diff --git "a/generative.ink.jsonl" "b/generative.ink.jsonl"
new file mode 100644--- /dev/null
+++ "b/generative.ink.jsonl"
@@ -0,0 +1,18 @@
+{"text": "Simulators :: — Moire[![](/images/home/rolling_phase.gif)](/)* [Posts](/posts/) * [Trees](/trees/) * [Prophecies](/prophecies/) * [About](/about/) 61 minutes [Simulators](/posts/simulators/) ================================ \\--- Table of Contents* [Summary](#summary) * [Meta](#meta) * [Agentic GPT](#agentic-gpt) \\+ [Unorthodox agency](#unorthodox-agency) \\+ [Orthogonal optimization](#orthogonal-optimization) \\+ [Roleplay sans player](#roleplay-sans-player) * [Oracle GPT and supervised learning](#oracle-gpt-and-supervised-learning) \\+ [Prediction vs question-answering](#prediction-vs-question-answering) \\+ [Finite vs infinite questions](#finite-vs-infinite-questions) \\+ [Paradigms of theory vs practice](#paradigms-of-theory-vs-practice) * [Tool / genie GPT](#tool--genie-gpt) * [Behavior cloning / mimicry](#behavior-cloning--mimicry) * [The simulation objective](#the-simulation-objective) \\+ [Solving for physics](#solving-for-physics) * [Simulacra](#simulacra) \\+ [Disambiguating rules and automata](#disambiguating-rules-and-automata) * [The limit of learned simulation](#the-limit-of-learned-simulation) * [A note on GANs](#a-note-on-gans) * [Table of quasi-simulators](#table-of-quasi-simulators) \\--- *This post is also available on [Lesswrong](https://www.lesswrong.com/posts/vJFdjigzmcXMhNTsx/simulators)* \\--- ![simulation](/simulators/simulators-banner.png) *\"Moebius illustration of a simulacrum living in an AI-generated story discovering it is in a simulation\" by DALL-E 2* Summary \\------- **TL;DR**: Self-supervised learning may create AGI or its foundation. What would that look like? Unlike the limit of RL, the limit of self-supervised learning has received surprisingly little conceptual attention, and recent progress has made deconfusion in this domain more pressing. Existing AI taxonomies either fail to capture important properties of self-supervised models or lead to confusing propositions. For instance, GPT policies do not seem globally agentic, yet can be conditioned to behave in goal-directed ways. This post describes a frame that enables more natural reasoning about properties like agency: GPT, insofar as it is inner-aligned, is a **simulator** which can simulate agentic and non-agentic **simulacra**. The purpose of this post is to capture these objects in words ~so GPT can reference them~ and provide a better foundation for understanding them. I use the generic term \"simulator\" to refer to models trained with predictive loss on a self-supervised dataset, invariant to architecture or data type (natural language, code, pixels, game states, etc). The outer objective of self-supervised learning is Bayes-optimal conditional inference over the prior of the training distribution, which I call the **simulation objective**, because a conditional model can be used to simulate rollouts which probabilistically obey its learned distribution by iteratively sampling from its posterior (predictions) and updating the condition (prompt). Analogously, a predictive model of physics can be used to compute rollouts of phenomena in simulation. A goal-directed agent which evolves according to physics can be simulated by the physics rule parameterized by an initial state, but the same rule could also propagate agents with different values, or non-agentic phenomena like rocks. This ontological distinction between simulator (rule) and simulacra (phenomena) applies directly to generative models like GPT. Meta \\---- * This post is intended as the first in a sequence on the alignment problem in a landscape where self-supervised simulators are a possible/likely form of powerful AI. I don't know how many subsequent posts I'll actually publish. Take it as a prompt. * I use the generic term \"GPT\" to refer to transformers trained on next-token prediction. * A while ago when I was trying to avoid having to write this post by hand, I prompted GPT-3 with an early outline of this post. I've spliced in some excerpts from it, `indicated by this style`. Prompt, generated text, and curation metrics [here](https://generative.ink/artifacts/simulators/). The limit of sequence modeling ============================== > Transformer-based language models have recently achieved remarkable results… > > – every paper since 2020 > > \\--- GPT is not a new form of AI in terms of its training methodology and outer objective: sequence generation from statistical models of data is an old idea. In 1951, Claude Shannon described using [n-grams](https://en.wikipedia.org/wiki/N-gram) to approximate conditional next-letter probabilities of a text dataset and \"reversed\" to generate text samples[1](#fn:1). I don't know of any other notable advances until the 2010s brought the first interesting language generation results from neural networks. In 2015, Karpathy wrote a blog post/tutorial sharing his excitement about [The Unreasonable Effectiveness of Recurrent Neural Networks](http://karpathy.github.io/2015/05/21/rnn-effectiveness/): > Fast forward about a year: I'm training RNNs all the time and I've witnessed their power and robustness many times, and yet their magical outputs still find ways of amusing me. This post is about sharing some of that magic with you. > > We'll train RNNs to generate text character by character and ponder the question \"how is that even possible?\" > > The \"magical outputs\" of char-RNNs looked like this: > PANDARUS: Alas, I think he shall be come approached and the day When little srain would be attain'd into being never fed, And who is but a chain and subjects of his death, I should not sleep. > > Second Senator: They are away this miseries, produced upon my soul, Breaking and strongly should be buried, when I perish The earth and thoughts of many states. > > DUKE VINCENTIO: Well, your wit is in the care of side and that. > > Second Lord: They would be ruled after this chamber, and my fair nues begun out of the fact, to be conveyed, Whose noble souls I'll have the heart of the wars. > > Clown: Come, sir, I will make did behold your worship. > > VIOLA: I'll drink it. > > At the time, this really was magical (and [uncanny](https://en.wikipedia.org/wiki/Uncanny_valley)). How does it know that *miseries* are *produced upon the soul?* Or that a *clown* should address a *lord* as \"sir\"? Char-RNNs were like ouija boards, but actually possessed by a low-fidelity ghost summoned from a text corpus. I remember being thrilled by the occasional glimmers of semantic comprehension in a domain of unbounded constructive meaning. But, aside from indulging that emotion, I didn't think about what would happen if my char-RNN bots actually improved indefinitely at their training objective of natural language prediction. It just seemed like there were some complexity classes of magic that neural networks could learn, and others that were inaccessible, at least in the conceivable future. Huge mistake! Perhaps I could have started thinking several years earlier about what now seems so fantastically important. But it wasn't until GPT-3, when I saw the [qualitative correlate](https://www.gwern.net/GPT-3) of \"loss going down\", that I updated. I wasn't the only one[2](#fn:2) whose imagination was naively constrained. A 2016 paper from Google Brain, \"[Exploring the Limits of Language Modeling](https://arxiv.org/abs/1602.02410)\", describes the utility of training language models as follows: > Often (although not always), training better language models improves the underlying metrics of the downstream task (such as word error rate for speech recognition, or BLEU score for translation), which makes the task of training better LMs valuable by itself. > > Despite its title, this paper's analysis is entirely myopic. Improving BLEU scores is neat, but how about *modeling general intelligence* as a downstream task? [In](https://arxiv.org/abs/2005.14165) [retrospect](https://arxiv.org/abs/2204.02311), an exploration of the *limits* of language modeling should have read something more like: > If loss keeps going down on the test set, in the limit – putting aside whether the current paradigm can approach it – the model must be learning to interpret and predict all patterns represented in language, including common-sense reasoning, goal-directed optimization, and deployment of the sum of recorded human knowledge. Its outputs would behave as intelligent entities in their own right. You could converse with it by alternately generating and adding your responses to its prompt, and it would pass the Turing test. In fact, you could condition it to generate interactive and autonomous versions of any real or fictional person who has been recorded in the training corpus or even *could* be recorded (in the sense that the record counterfactually \"could be\" in the test set). Oh shit, and it could write code… > > The paper does, however, mention that making the model bigger improves test perplexity.[3](#fn:3) I'm only picking on *Jozefowicz et al.* because of their ironic title. I don't know of any explicit discussion of this limit predating GPT, except a working consensus of Wikipedia editors that [NLU](https://en.wikipedia.org/wiki/Natural-language_understanding) is [AI-complete](https://en.wikipedia.org/wiki/AI-complete#AI-complete_problems). The earliest engagement with the hypothetical of \"*what if self-supervised sequence modeling actually works*\" that I know of is a terse post from 2019, [Implications of GPT-2](https://www.lesswrong.com/posts/YJRb6wRHp7k39v69n/implications-of-gpt-2), by Gurkenglas. It is brief and relevant enough to quote in full: > I was impressed by GPT-2, to the point where I wouldn't be surprised if a future version of it could be used pivotally using existing protocols. > > Consider generating half of a Turing test transcript, the other half being supplied by a human judge. If this passes, we could immediately implement an HCH of AI safety researchers solving the problem if it's within our reach at all. (Note that training the model takes much more compute than generating text.) > > This might not be the first pivotal application of language models that becomes possible as they get stronger. > > It's a source of superintelligence that doesn't automatically run into utility maximizers. It sure doesn't look like AI services, lumpy or no. > > It is conceivable that predictive loss does not descend to the AGI-complete limit, maybe because: * Some AGI-necessary predictions are [too difficult to be learned by even a scaled version of the current paradigm](https://www.lesswrong.com/posts/pv7Qpu8WSge8NRbpB/). * The irreducible entropy is above the \"AGI threshold\": datasets + context windows [contain insufficient information](https://twitter.com/ylecun/status/1562162165540331520) to improve on some necessary predictions. But I have not seen enough evidence for either not to be concerned that we have in our hands a well-defined protocol that could end in AGI, or a foundation which could spin up an AGI without too much additional finagling. As Gurkenglas observed, this would be a very different source of AGI than previously foretold. The old framework of alignment ============================== A few people did think about what would happen if *agents* actually worked. The hypothetical limit of a powerful system **optimized to optimize for an objective** drew attention even before reinforcement learning became mainstream in the 2010s. Our current instantiation of AI alignment theory, [crystallized by Yudkowsky, Bostrom, et al](https://www.lesswrong.com/posts/i4susk4W3ieR5K92u/ai-risk-and-opportunity-humanity-s-efforts-so-far), stems from the vision of an arbitrarily-capable system whose cognition and behavior flows from a goal. But since GPT-3 I've [noticed](https://www.lesswrong.com/s/zpCiuR4T343j9WkcK/p/5JDkW4MYXit2CquLs), in my own thinking and in alignment discourse, a dissonance between theory and practice/phenomena, as the behavior and nature of actual systems that seem nearest to AGI also resist *short descriptions in the dominant ontology*. I only recently discovered the question \"[Is the work on AI alignment relevant to GPT?](https://www.lesswrong.com/posts/dPcKrfEi87Zzr7w6H/is-the-work-on-ai-alignment-relevant-to-gpt)\" which stated this observation very explicitly: > I don't follow [AI alignment research] in any depth, but I am noticing a striking disconnect between the concepts appearing in those discussions and recent advances in AI, especially GPT-3. > > People talk a lot about an AI's goals, its utility function, its capability to be deceptive, its ability to simulate you so it can get out of a box, ways of motivating it to be benign, Tool AI, Oracle AI, and so on. (…) But when I look at GPT-3, even though this is already an AI that Eliezer finds alarming, I see none of these things. GPT-3 is a huge model, trained on huge data, for predicting text. > > My belated answer: A lot of prior work on AI alignment is relevant to GPT. I spend most of my time thinking about GPT alignment, and concepts like [goal-directedness](https://www.alignmentforum.org/tag/goal-directedness), [inner/outer alignment](https://www.alignmentforum.org/s/r9tYkB2a8Fp4DN8yB), [myopia](https://www.lesswrong.com/tag/myopia), [corrigibility](https://www.lesswrong.com/tag/corrigibility), [embedded agency](https://www.alignmentforum.org/posts/i3BTagvt3HbPMx6PN/embedded-agency-full-text-version), [model splintering](https://www.lesswrong.com/posts/k54rgSg7GcjtXnMHX/model-splintering-moving-from-one-imperfect-model-to-another-1), and even [tiling agents](https://arbital.com/p/tiling_agents/) are active in the vocabulary of my thoughts. But GPT violates some prior assumptions such that these concepts sound dissonant when applied naively. To usefully harness these preexisting abstractions, we need something like an ontological [adapter pattern](https://en.wikipedia.org/wiki/Adapter_pattern) that maps them to the appropriate objects. GPT's unforeseen nature also demands new abstractions (the adapter itself, for instance). My thoughts also use load-bearing words that do not inherit from alignment literature. Perhaps it shouldn't be surprising if the form of the first visitation from [mindspace](https://www.lesswrong.com/posts/tnWRXkcDi5Tw9rzXw/the-design-space-of-minds-in-general) mostly escaped a few years of theory [conducted in absence of its object](https://www.lesswrong.com/posts/72scWeZRta2ApsKja/epistemological-vigilance-for-alignment#Direct_access__so_far_and_yet_so_close). The purpose of this post is to capture that object (conditional on a predictive self-supervised training story) in words. Why in words? In order to write coherent alignment ideas which reference it! This is difficult in the existing ontology, because unlike the concept of an *agent*, whose *name* evokes the abstract properties of the system and thereby invites extrapolation, the general category for \"a model optimized for an AGI-complete predictive task\" has not been given a name[4](#fn:4). Namelessness can not only be a symptom of the extrapolation of powerful predictors falling through conceptual cracks, but also a cause, because what we can represent in words is *what we can condition on for further generation.* To whatever extent this [shapes private thinking](https://en.wikipedia.org/wiki/Language_of_thought_hypothesis), it is a strict constraint on communication, when thoughts must be sent through the bottleneck of words. I want to hypothesize about LLMs in the limit, because when AI is all of a sudden [writing viral blog posts](https://www.theverge.com/2020/8/16/21371049/gpt3-hacker-news-ai-blog), [coding competitively](https://www.deepmind.com/blog/competitive-programming-with-alphacode), [proving theorems](https://arxiv.org/abs/2009.03393), and [passing the Turing test so hard that the interrogator sacrifices their career at Google to advocate for its personhood](https://www.washingtonpost.com/technology/2022/06/11/google-ai-lamda-blake-lemoine/), a process is clearly underway whose limit we'd be foolish not to contemplate. I could directly extrapolate the architecture responsible for these feats and talk about \"GPT-N\", a bigger autoregressive transformer. But often some implementation details aren't as important as the more abstract archetype that GPT represents – I want to speak the [true name](https://www.lesswrong.com/posts/FWvzwCDRgcjb9sigb/why-agent-foundations-an-overly-abstract-explanation) of the solution which unraveled a Cambrian explosion of AI phenomena with *inessential details unconstrained*, as we'd speak of natural selection finding the solution of the \"lens\" without specifying the prototype's diameter or focal length. (Only when I am able to condition on that level of abstraction can I generate metaphors like \"language is a [lens that sees its flaws](https://www.lesswrong.com/s/5g5TkQTe9rmPS5vvM/p/46qnWRSR7L2eyNbMA)\".) Inadequate ontologies ===================== In the next few sections I'll attempt to fit GPT into some established categories, hopefully to reveal something about the shape of the peg through contrast, beginning with the main antagonist of the alignment problem as written so far, the **agent**. Agentic GPT \\----------- Alignment theory has been largely pushed by considerations of agentic AGIs. There were good reasons for this focus: * **Agents are convergently dangerous** **for theoretical reasons** like [instrumental convergence](https://www.lesswrong.com/tag/instrumental-convergence), [goodhart](https://www.lesswrong.com/tag/goodhart-s-law), and [orthogonality](https://www.lesswrong.com/tag/orthogonality-thesis). * **RL creates agents, and RL seemed to be the way to AGI**. In the 2010s, reinforcement learning was the dominant paradigm for those interested in AGI (e.g. OpenAI). RL lends naturally to creating agents that pursue rewards/utility/objectives. So there was reason to expect that agentic AI would be the first (and by the theoretical arguments, last) form that superintelligence would take. * **Agents are powerful and economically productive.** It's a reasonable guess that humans will create such systems [if only because we can](https://mittmattmutt.medium.com/superintelligence-and-moral-blindness-7436300fcb1f). The first reason is conceptually self-contained and remains compelling. The second and third, grounded in the state of the world, has been shaken by the current climate of AI progress, where products of self-supervised learning generate most of the buzz: not even primarily for their SOTA performance in domains traditionally dominated by RL, like games[5](#fn:5), but rather for their virtuosity in domains where RL never even took baby steps, like natural language synthesis. What pops out of self-supervised predictive training is noticeably not a classical agent. Shortly after GPT-3's release, David Chalmers lucidly observed that the policy's relation to agent*s* is like that of a \"chameleon\" or \"engine\": > GPT-3 does not look much like an agent. It does not seem to have goals or preferences beyond completing text, for example. It is more like a chameleon that can take the shape of many different agents. Or perhaps it is an engine that can be used under the hood to drive many agents. But it is then perhaps these systems that we should assess for agency, consciousness, and so on.[6](#fn:6) > > But at the same time, GPT can *act like an agent* – and aren't actions what ultimately matter? In [Optimality is the tiger, and agents are its teeth](https://www.lesswrong.com/posts/kpPnReyBC54KESiSn), Veedrac points out that a model like GPT does not need to care about the consequences of its actions for them to be effectively those of an agent that kills you. This is *more* reason to examine the nontraditional relation between the optimized policy and agents, as it has implications for how and why agents are served. ### Unorthodox agency `GPT's behavioral properties include imitating the general pattern of human dictation found in its universe of training data, e.g., arXiv, fiction, blog posts, Wikipedia, Google queries, internet comments, etc. Among other properties inherited from these historical sources, it is capable of goal-directed behaviors such as planning. For example, given a free-form prompt like, \"you are a desperate smuggler tasked with a dangerous task of transporting a giant bucket full of glowing radioactive materials across a quadruple border-controlled area deep in Africa for Al Qaeda,\" the AI will fantasize about logistically orchestrating the plot just as one might, working out how to contact Al Qaeda, how to dispense the necessary bribe to the first hop in the crime chain, how to get a visa to enter the country, etc. Considering that no such specific chain of events are mentioned in any of the bazillions of pages of unvarnished text that GPT slurped`[7](#fn:7)`, the architecture is not merely imitating the universe, but reasoning about possible versions of the universe that does not actually exist, branching to include new characters, places, and events` `When thought about behavioristically, GPT superficially demonstrates many of the raw ingredients to act as an \"agent\", an entity that optimizes with respect to a goal. But GPT is hardly a proper agent, as it wasn't optimized to achieve any particular task, and does not display an epsilon optimization for any single reward function, but instead for many, including incompatible ones. Using it as an agent is like using an agnostic politician to endorse hardline beliefs– he can convincingly talk the talk, but there is no psychic unity within him; he could just as easily play devil's advocate for the opposing party without batting an eye. Similarly, GPT instantiates simulacra of characters with beliefs and goals, but none of these simulacra are the algorithm itself. They form a virtual procession of different instantiations as the algorithm is fed different prompts, supplanting one surface personage with another. Ultimately, the computation itself is more like a disembodied dynamical law that moves in a pattern that broadly encompasses the kinds of processes found in its training data than a cogito meditating from within a single mind that aims for a particular outcome.` Presently, GPT is the only way to instantiate agentic AI that behaves capably [outside toy domains](https://arbital.com/p/rich_domain/). These intelligences exhibit goal-directedness; they can plan; they can form and test hypotheses; they can persuade and be persuaded[8](#fn:8). It would not be very [dignified](https://www.lesswrong.com/posts/j9Q8bRmwCgXRYAgcJ/miri-announces-new-death-with-dignity-strategy) of us to gloss over the sudden arrival of artificial agents *often indistinguishable from human intelligence* just because the policy that generates them \"only cares about predicting the next word\". But nor should we ignore the fact that these agentic entities exist in an unconventional relationship to the policy, the neural network \"GPT\" that was trained to minimize log-loss on a dataset. GPT-driven agents are ephemeral – they can spontaneously disappear if the scene in the text changes and be replaced by different spontaneously generated agents. They can exist in parallel, e.g. in a story with multiple agentic characters in the same scene. There is a clear sense in which the network doesn't \"want\" what the things that it simulates want, seeing as it would be just as willing to simulate an agent with opposite goals, or throw up obstacles which foil a character's intentions for the sake of the story. The more you think about it, the more fluid and intractable it all becomes. Fictional characters act agentically, but they're at least implicitly puppeteered by a virtual author who has orthogonal intentions of their own. Don't let me get into the fact that all these layers of \"intentionality\" operate largely in [indeterminate superpositions](https://generative.ink/posts/language-models-are-multiverse-generators/#multiplicity-of-pasts-presents-and-futures). This is a clear way that GPT diverges from orthodox visions of agentic AI: **In the agentic AI ontology, there is no difference between the policy and the effective agent, but for GPT, there is.** It's not that anyone ever said there had to be 1:1 correspondence between policy and effective agent; it was just an implicit assumption which felt natural in the agent frame (for example, it tends to hold for RL). GPT pushes us to realize that this was an assumption, and to consider the consequences of removing it for our constructive maps of mindspace. ### Orthogonal optimization Indeed, [Alex Flint warned](https://www.alignmentforum.org/posts/8HWGXhnCfAPgJYa9D/pitfalls-of-the-agent-model) of the potential consequences of leaving this assumption unchallenged: > **Fundamental misperception due to the agent frame**: That the design space for autonomous machines that exert influence over the future is narrower than it seems. This creates a self-fulfilling prophecy in which the AIs actually constructed are in fact within this narrower regime of agents containing an unchanging internal decision algorithm. > > If there are other ways of constructing AI, might we also avoid some of the scary, theoretically hard-to-avoid side-effects of optimizing an agent like [instrumental convergence](https://www.lesswrong.com/tag/instrumental-convergence)? GPT provides an interesting example. GPT doesn't seem to care which agent it simulates, nor if the scene ends and the agent is effectively destroyed. This is not corrigibility in [Paul Christiano's formulation](https://ai-alignment.com/corrigibility-3039e668638), where the policy is \"okay\" with being turned off or having its goal changed in a positive sense, but has many aspects of the [negative formulation found on Arbital](https://arbital.com/p/corrigibility/). It is corrigible in this way because a major part of the agent specification (the prompt) is not fixed by the policy, and the policy lacks direct training incentives to control its prompt[9](#fn:9), as it never generates text or otherwise influences its prompts during training. It's *we* who choose to sample tokens from GPT's predictions and append them to the prompt at runtime, and the result is not always helpful to any agents who may be programmed by the prompt. The downfall of the ambitious villain from an oversight committed in hubris is a predictable narrative pattern.[10](#fn:10) So is the end of a scene. In general, the model's prediction vector could point in any direction relative to the predicted agent's interests. I call this the **prediction orthogonality thesis:** *A model whose objective is prediction*[11](#fn:11)\\* can simulate agents who optimize toward any objectives, with any degree of optimality (bounded above but not below by the model's power).\\* This is a corollary of the classical [orthogonality thesis](https://www.lesswrong.com/tag/orthogonality-thesis), which states that agents can have any combination of intelligence level and goal, combined with the assumption that agents can in principle be predicted. A single predictive model may also predict multiple agents, either independently (e.g. in different conditions), or interacting in a multi-agent simulation. A more optimal predictor is not restricted to predicting more optimal agents: being smarter does not make you unable to predict stupid systems, nor things that aren't agentic like the [weather](https://en.wikipedia.org/wiki/History_of_numerical_weather_prediction). Are there any constraints on what a predictive model can be at all, other than computability? Only that it makes sense to talk about its \"prediction objective\", which implies the existence of a \"ground truth\" distribution to which the predictor's optimality is measured. Several words in that last sentence may conceal labyrinths of nuance, but for now let's wave our hands and say that if we have some way of presenting [Bayes-structure](https://www.lesswrong.com/posts/QrhAeKBkm2WsdRYao/searching-for-bayes-structure) with evidence of a distribution, we can build an optimization process whose outer objective is optimal prediction. We can specify some types of outer objectives using a ground truth distribution that we cannot with a utility function. As in the case of GPT, there is no difficulty in incentivizing a model to *predict* actions that are [corrigible](https://arbital.com/p/corrigibility/), [incoherent](https://aiimpacts.org/what-do-coherence-arguments-imply-about-the-behavior-of-advanced-ai/), [stochastic](https://www.lesswrong.com/posts/msJA6B9ZjiiZxT6EZ/lawful-uncertainty), [irrational](https://www.lesswrong.com/posts/6ddcsdA2c2XpNpE5x/newcomb-s-problem-and-regret-of-rationality), or otherwise anti-natural to expected utility maximization. All you need is evidence of a distribution exhibiting these properties. For instance, during GPT's training, sometimes predicting the next token coincides with predicting agentic behavior, but: * The actions of agents described in the data are rarely optimal for their goals; humans, for instance, are computationally bounded, irrational, normative, habitual, fickle, hallucinatory, etc. * Different prediction steps involve mutually incoherent goals, as human text records a wide range of differently-motivated agentic behavior * Many prediction steps don't correspond to the action of *any* consequentialist agent but are better described as reporting on the structure of reality, e.g. the year in a timestamp. These transitions incentivize GPT to improve its model of the world, orthogonally to agentic objectives. * When there is insufficient information to predict the next token with certainty, [log-loss incentivizes a probabilistic output](https://en.wikipedia.org/wiki/Scoring_rule#Proper_scoring_rules). Utility maximizers [aren't supposed to become more stochastic](https://www.lesswrong.com/posts/msJA6B9ZjiiZxT6EZ/lawful-uncertainty) in response to uncertainty. Everything can be trivially modeled as a utility maximizer, but for these reasons, a utility function is not a good explanation or compression of GPT's training data, and its optimal predictor is not well-described as a utility maximizer. However, just because information isn't compressed well by a utility function doesn't mean it can't be compressed another way. The [Mandelbrot set](https://en.wikipedia.org/wiki/Mandelbrot_set) is a complicated pattern compressed by a very simple generative algorithm which makes no reference to future consequences and doesn't involve argmaxxing anything (except vacuously [being the way it is](https://www.lesswrong.com/posts/d2n74bwham8motxyX/optimization-at-a-distance#An_Agent_Optimizing_Its_Own_Actions)). Likewise the set of all possible rollouts of [Conway's Game of Life](https://en.wikipedia.org/wiki/Conway%27s_Game_of_Life) – [some automata may be well-described as agents](https://www.lesswrong.com/posts/3SG4WbNPoP8fsuZgs/agency-in-conway-s-game-of-life), but they are a minority of possible patterns, and not all agentic automata will share a goal. Imagine trying to model Game of Life as an expected utility maximizer! There are interesting things that are not utility maximizers, some of which qualify as AGI or [TAI](https://forum.effectivealtruism.org/topics/transformative-artificial-intelligence). Are any of them something we'd be better off creating than a utility maximizer? An inner-aligned GPT, for instance, gives us a way of instantiating goal-directed processes which can be tempered with normativity and freely terminated in a way that is not anti-natural to the training objective. There's much more to say about this, but for now, I'll bring it back to how GPT defies the agent orthodoxy. The crux stated earlier can be restated from the perspective of training stories: **In the agentic AI ontology, the** ***direction of optimization pressure applied by training*** **is in the direction of the effective agent's objective function, but in GPT's case it is (most generally) orthogonal.**[12](#fn:12) This means that neither the policy nor the effective agents necessarily become more optimal agents as loss goes down, because the policy is not optimized to be an agent, and the agent-objectives are not optimized directly. ### Roleplay sans player > Napoleon: You have written this huge book on the system of the world without once mentioning the author of the universe. > > Laplace: Sire, I had no need of that hypothesis. > > Even though neither GPT's behavior nor its training story fit with the traditional agent framing, there are still compatibilist views that characterize it as some kind of agent. For example, Gwern has said[13](#fn:13) that anyone who uses GPT for long enough begins to think of it as an agent who only cares about roleplaying a lot of roles. That framing seems unnatural to me, comparable to thinking of physics as an agent who only cares about evolving the universe accurately according to the laws of physics. At best, the agent is an epicycle; but it is also compatible with interpretations that generate dubious predictions. Say you're told that an agent *values predicting text correctly*. Shouldn't you expect that: * It wants text to be easier to predict, and given the opportunity will influence the prediction task to make it easier (e.g. by generating more predictable text or otherwise influencing the environment so that it receives easier prompts); * It wants to become better at predicting text, and given the opportunity will self-improve; * It doesn't want to be prevented from predicting text, and will prevent itself from being shut down if it can? In short, all the same types of instrumental convergence that we expect from agents who want almost anything at all. But this behavior would be very unexpected in GPT, whose training doesn't incentivize instrumental behavior that optimizes prediction accuracy! GPT does not generate rollouts during training. Its output is never sampled to yield \"actions\" whose consequences are evaluated, so there is no reason to expect that GPT will form preferences over the *consequences* of its output related to the text prediction objective.[14](#fn:14) Saying that GPT is an agent who wants to roleplay implies the presence of a coherent, unconditionally instantiated *roleplayer* running the show who attaches terminal value to roleplaying. This presence is an additional hypothesis, and so far, I haven't noticed evidence that it's true. (I don't mean to imply that Gwern thinks this about GPT[15](#fn:15), just that his words do not properly rule out this interpretation. It's a likely enough interpretation that [ruling it out](https://www.lesswrong.com/posts/57sq9qA3wurjres4K/ruling-out-everything-else) is important: I've seen multiple people suggest that GPT might want to generate text which makes future predictions easier, and this is something that can happen in some forms of self-supervised learning – see the note on GANs in the appendix.) I do not think any simple modification of the concept of an agent captures GPT's natural category. It does not seem to me that GPT is a roleplayer, only that it roleplays. But what is the word for something that roleplays minus the implication that some*one* is behind the mask? Oracle GPT and supervised learning \\---------------------------------- While the alignment sphere favors the agent frame for thinking about GPT, in *capabilities* research distortions tend to come from a lens inherited from *supervised learning*. Translated into alignment ontology, the effect is similar to viewing GPT as an \"[oracle AI](https://publicism.info/philosophy/superintelligence/11.html)\" – a view not altogether absent from conceptual alignment, but most influential in the way GPT is used and evaluated by machine learning engineers. Evaluations for language models tend to look like evaluations for *supervised* models, consisting of close-ended question/answer pairs – often because they *are* evaluations for supervised models. Prior to the LLM paradigm, language models were trained and tested on evaluation datasets like [Winograd](https://en.wikipedia.org/wiki/Winograd_schema_challenge) and [SuperGLUE](https://super.gluebenchmark.com/) which consist of natural language question/answer pairs. The fact that large pretrained models performed well on these same NLP benchmarks without supervised fine-tuning was a novelty. The titles of the GPT-2 and GPT-3 papers, [Language Models are Unsupervised Multitask Learners](https://d4mucfpksywv.cloudfront.net/better-language-models/language_models_are_unsupervised_multitask_learners.pdf) and [Language Models are Few-Shot Learners](https://arxiv.org/abs/2005.14165), respectively articulate surprise that *self-supervised* models implicitly learn supervised tasks during training, and can learn supervised tasks at runtime. Of all the possible papers that could have been written about GPT-3, OpenAI showcased its ability to extrapolate the pattern of question-answer pairs (few-shot prompts) from supervised learning datasets, a novel capability they called \"meta-learning\". This is a weirdly specific and indirect way to break it to the world that you've created an AI able to extrapolate semantics of arbitrary natural language structures, especially considering that in many cases the [few-shot prompts were actually unnecessary](https://arxiv.org/abs/2102.07350). The assumptions of the supervised learning paradigm are: * The model is optimized to answer questions correctly * Tasks are closed-ended, defined by question/correct answer pairs These are essentially the assumptions of oracle AI, as [described by Bostrom](https://publicism.info/philosophy/superintelligence/11.html) and [in subsequent usage](https://www.lesswrong.com/tag/oracle-ai/history). So influential has been this miscalibrated perspective that [Gwern](https://www.gwern.net/GPT-3#prompts-as-programming), [nostalgebraist](https://www.lesswrong.com/posts/pv7Qpu8WSge8NRbpB/) and [myself](https://generative.ink/posts/language-models-are-0-shot-interpreters/#0-shot-few-shot-and-meta-learning) – who share a peculiar model overlap due to intensive firsthand experience with the downstream behaviors of LLMs – have all repeatedly complained about it. I'll repeat some of these arguments here, tying into the view of GPT as an oracle AI, and separating it into the two assumptions inspired by supervised learning. ### Prediction vs question-answering `At first glance, GPT might resemble a generic \"oracle AI\", because it is trained to make accurate predictions. But its log loss objective is myopic and only concerned with immediate, micro-scale correct prediction of the next token, not answering particular, global queries such as \"what's the best way to fix the climate in the next five years?\" In fact, it is not specifically optimized to give *true* answers, which a classical oracle should strive for, but rather to minimize the divergence between predictions and training examples, independent of truth. Moreover, it isn't specifically trained to give answers in the first place! It may give answers if the prompt asks questions, but it may also simply elaborate on the prompt without answering any question, or tell the rest of a story implied in the prompt. What it does is more like animation than divination, executing the dynamical laws of its rendering engine to recreate the flows of history found in its training data (and a large superset of them as well), mutatis mutandis. Given the same laws of physics, one can build a multitude of different backgrounds and props to create different storystages, including ones that don't exist in training, but adhere to its general pattern.` GPT does not consistently try to say [true/correct things](https://www.alignmentforum.org/posts/BnDF5kejzQLqd5cjH/alignment-as-a-bottleneck-to-usefulness-of-gpt-3). This is not a bug – if it had to say true things all the time, GPT would be much constrained in its ability to [imitate Twitter celebrities](https://twitter.com/dril_gpt2) and write fiction. Spouting falsehoods in some circumstances is incentivized by GPT's outer objective. If you ask GPT a question, it will instead answer the question \"what's the next token after '{your question}'\", which will often diverge significantly from an earnest attempt to answer the question directly. GPT doesn't fit the category of oracle for a similar reason that it doesn't fit the category of agent. Just as it wasn't optimized for and doesn't consistently act according to any particular objective (except the tautological prediction objective), it was not optimized to be *correct* but rather *realistic,* and being realistic means predicting humans faithfully even when they are likely to be wrong. That said, GPT does store a vast amount of knowledge, and its corrigibility allows it to be cajoled into acting as an oracle, like it can be cajoled into acting like an agent. In order to get oracle behavior out of GPT, one must input a sequence such that the predicted continuation of that sequence coincides with an oracle's output. The GPT-3 paper's few-shot benchmarking strategy tries to persuade GPT-3 to answer questions correctly by having it predict how a list of correctly-answered questions will continue. Another strategy is to simply \"tell\" GPT it's in the oracle modality: > (I) told the AI to simulate a supersmart version of itself (this works, for some reason), and the first thing it spat out was the correct answer. > > – [Reddit post by u/Sophronius](https://www.reddit.com/r/rational/comments/lvn6ow/gpt3_just_figured_out_the_entire_mystery_plot_of/) > > But even when these strategies seem to work, there is no guarantee that they elicit anywhere near optimal question-answering performance, compared to another prompt in the innumerable space of prompts that would cause GPT to attempt the task, or compared to what the [model \"actually\" knows](https://www.lesswrong.com/tag/eliciting-latent-knowledge-elk). This means that no benchmark which evaluates downstream behavior is guaranteed or even expected to probe the upper limits of GPT's capabilities. In nostalgebraist's words, we have no [ecological evaluation](https://www.lesswrong.com/posts/pv7Qpu8WSge8NRbpB/#4__on_ecological_evaluation) of self-supervised language models – one that measures performance in a situation where the model is incentivised to perform as well as it can on the measure[16](#fn:16). As nostalgebraist [elegantly puts it](https://slatestarcodex.com/2020/06/10/the-obligatory-gpt-3-post/#comment-912529): > I called GPT-3 a \"disappointing paper,\" which is not the same thing as calling the model disappointing: the feeling is more like how I'd feel if they found a superintelligent alien and chose only to communicate its abilities by noting that, when the alien is blackout drunk and playing 8 simultaneous games of chess while also taking an IQ test, it *then* has an \"IQ\" of about 100. > > Treating GPT as an unsupervised implementation of a supervised learner leads to systematic underestimation of capabilities, which becomes a more dangerous mistake as unprobed capabilities scale. ### Finite vs infinite questions Not only does the supervised/oracle perspective obscure the importance and limitations of prompting, it also obscures one of the most crucial dimensions of GPT: the implicit time dimension. By this I mean the ability to evolve a process through time by recursively applying GPT, that is, generate text of arbitrary length. Recall, the second supervised assumption is that \"tasks are closed-ended, defined by question/correct answer pairs\". GPT was trained on context-completion pairs. But the pairs do not represent closed, independent tasks, and the division into question and answer is merely indexical: in another training sample, a token from the question is the answer, and in yet another, the answer forms part of the question[17](#fn:17). For example, the natural language sequence \"**The answer is a question**\" yields training samples like: {context: \"**The**\", completion: \" **answer**\"}, {context: \"**The answer**\", completion: \" **is**\"}, {context: \"**The answer is**\", completion: \" **a**\"}, {context: \"**The answer is a**\", completion: \" **question**\"} Since questions and answers are of compatible types, we can at runtime sample answers from the model and use them to construct new questions, and run this loop an indefinite number of times to generate arbitrarily long sequences that obey the model's approximation of the rule that links together the training samples. **The \"question\" GPT answers is \"what token comes next after {context}\". This can be asked interminably, because its answer always implies another question of the same type.** In contrast, models trained with supervised learning output answers that cannot be used to construct new questions, so they're only good for one step. Benchmarks derived from supervised learning test GPT's ability to produce correct answers, not to produce *questions* which cause it to produce a correct answer down the line. But GPT is capable of the latter, and that is how it is the [most powerful](https://ai.googleblog.com/2022/05/language-models-perform-reasoning-via.html). The supervised mindset causes capabilities researchers to focus on closed-form tasks rather than GPT's ability to simulate open-ended, indefinitely long processes[18](#fn:18), and as such to overlook multi-step inference strategies like chain-of-thought prompting. Let's see how the oracle mindset causes a blind spot of the same shape in the imagination of a hypothetical alignment researcher. Thinking of GPT as an oracle brings strategies to mind like asking GPT-N to predict a [solution to alignment from 2000 years in the future](https://www.alignmentforum.org/posts/nXeLPcT9uhfG3TMPS/conditioning-generative-models).). There are various problems with this approach to solving alignment, of which I'll only mention one here: even assuming this prompt is *outer aligned*[19](#fn:19) in that a logically omniscient GPT would give a useful answer, it is probably not the best approach for a finitely powerful GPT, because the *process* of generating a solution in the order and resolution that would appear in a future article is probably far from the optimal *multi-step algorithm* for computing the answer to an unsolved, difficult question. GPTs ability to arrive at true answers depends on not only the space to solve a problem in multiple steps (of the [right granularity](https://blog.eleuther.ai/factored-cognition/)), but also the direction of the flow of evidence in that *time*. If we're ambitious about getting the truth from a finitely powerful GPT, we need to incite it to predict truth-seeking processes, not just ask it the right questions. Or, in other words, the more general problem we have to solve is not asking GPT the question[20](#fn:20) that makes it output the right answer, but asking GPT the question that makes it output the right question (…) that makes it output the right answer.[21](#fn:21) A question anywhere along the line that elicits a premature attempt at an answer could [neutralize the remainder of the process into rationalization](https://generative.ink/posts/methods-of-prompt-programming/#avoiding-rationalization). I'm looking for a way to classify GPT which not only minimizes surprise but also conditions the imagination to efficiently generate good ideas for how it can be used. What category, unlike the category of oracles, would make the importance of *process* specification obvious? ### Paradigms of theory vs practice Both the agent frame and the supervised/oracle frame are historical artifacts, but while assumptions about agency primarily flow downward from the preceptial paradigm of alignment *theory*, oracle-assumptions primarily flow upward from the *experimental* paradigm surrounding GPT's birth. We use and evaluate GPT like an oracle, and that causes us to implicitly think of it as an oracle. Indeed, the way GPT is typically used by researchers resembles the archetypal image of Bostrom's oracle perfectly if you abstract away the semantic content of the model's outputs. The AI sits passively behind an API, computing responses only when prompted. It typically has no continuity of state between calls. Its I/O is text rather than \"real-world actions\". All these are consequences of how we choose to interact with GPT – which is not arbitrary; the way we deploy systems is guided by their nature. It's for some good reasons that current GPTs lend to disembodied operation and docile APIs. Lack of long-horizon coherence and [delusions](https://arxiv.org/abs/2110.10819) discourage humans from letting them run autonomously amok (usually). But the way we deploy systems is also guided by practical paradigms. One way to find out how a technology can be used is to give it to people who have less preconceptions about how it's supposed to be used. OpenAI found that most users use their API to generate freeform text: ![](https://lh6.googleusercontent.com/Jr4EW6PYr95i9QdLCz6fjpqyYhxgszPck4TQLKBIkxhnmGQ99rymanvi8tLM3Xj-qkvt2g4i3eH826EUj34cPH35ixpp0N_ybTBlBabvHYFS0dDI3voDIeicj3j0jPi0WioMRPk4w5lzRBioMpBK8A)[22](#fn:22) Most of my own experience using GPT-3 has consisted of simulating indefinite processes which maintain state continuity over up to hundreds of pages. I was driven to these lengths because GPT-3 kept answering its own questions with questions that I wanted to ask it more than anything else I had in mind. Tool / genie GPT \\---------------- I've sometimes seen GPT casually classified as [tool AI](https://publicism.info/philosophy/superintelligence/11.html). GPTs resemble tool AI from the outside, like it resembles oracle AI, because it is often deployed semi-autonomously for tool-like purposes (like helping me draft this post): `It could also be argued that GPT is a type of \"Tool AI\", because it can generate useful content for products, e.g., it can write code and generate ideas. However, unlike specialized Tool AIs that optimize for a particular optimand, GPT wasn't optimized to do anything specific at all. Its powerful and general nature allows it to be *used* as a Tool for many tasks, but it wasn't expliitly trained to achieve these tasks, and does not strive for optimality.` The argument structurally reiterates what has already been said for agents and oracles. Like agency and oracularity, tool-likeness is a contingent capability of GPT, but also orthogonal to its motive. The same line of argument draws the same conclusion from the question of whether GPT belongs to the fourth Bostromian AI caste, genies. The genie modality is exemplified by Instruct GPT and Codex. But like every behavior I've discussed so far which is more specific than predicting text, \"instruction following\" describes only an exploitable subset of all the patterns tread by the sum of human language and inherited by its imitator. Behavior cloning / mimicry \\-------------------------- The final category I'll analyze is behavior cloning, a designation for predictive learning that I've mostly seen used in contrast to RL. According to an [article from 1995](https://www.sciencedirect.com/science/article/pii/S1474667017467164), \"Behavioural cloning is the process of reconstructing a skill from an operator's behavioural traces by means of Machine Learning techniques.\" The term \"mimicry\", as [used by Paul Christiano](https://ai-alignment.com/against-mimicry-6002a472fc42), means the same thing and has similar connotations. Behavior cloning in its historical usage carries the implicit or explicit assumption that a single agent is being cloned. The natural extension of this to a model trained to predict a diverse human-written dataset might be to say that GPT models a *distribution* of agents which are selected by the prompt. But this image of \"parameterized\" behavior cloning still fails to capture some essential properties of GPT. The vast majority of prompts that produce coherent behavior never occur as prefixes in GPT's training data, but depict hypothetical processes whose behavior can be predicted by virtue of being capable at predicting language in general. We might call this phenomenon \"interpolation\" (or \"[extrapolation](https://arxiv.org/abs/2110.09485)\"). But to hide it behind any one word and move on would be to gloss over the *entire phenomenon of GPT.* Natural language has the property of [*systematicity*](https://evjang.com/2021/12/17/lang-generalization.html): \"blocks\", such as words, can be combined to form composite meanings. The number of meanings expressible is a combinatorial function of available blocks. A system which learns natural language is incentivized to learn systematicity; if it succeeds, it gains access to the combinatorial proliferation of meanings that can be expressed in natural language. What GPT lets us do is use natural language to specify any of a functional infinity of configurations, e.g. the mental contents of a person and the physical contents of the room around them, *and animate that*. That is the terrifying vision of the limit of prediction that struck me when I first saw GPT-3's outputs. The words \"behavior cloning\" do not automatically evoke this in my mind. The idea of parameterized behavior cloning grows more unwieldy if we remember that GPT's prompt continually changes during autoregressive generation. If GPT is a parameterized agent, then parameterization is not a fixed flag that chooses a process out of a set of possible processes. The parameterization *is* what is evolved – a successor \"agent\" selected by the old \"agent\" at each timestep, and neither of them need to have precedence in the training data. Behavior cloning / mimicry is also associated with the assumption that capabilities of the simulated processes are strictly [bounded by the capabilities of the demonstrator(s)](https://ai-alignment.com/against-mimicry-6002a472fc42). A supreme counterexample is the [Decision Transformer](https://arxiv.org/abs/2106.01345), which can be used to run processes which achieve SOTA for ~offline~ reinforcement learning despite being trained on *random* trajectories. Something which can predict everything all the time is more formidable than any demonstrator it predicts: the upper bound of what can be learned from a dataset is not the most capable trajectory, but the conditional structure of the universe implicated by their sum (though it may not be trivial to [extract that knowledge](https://www.alignmentforum.org/tag/eliciting-latent-knowledge-elk)). Extrapolating the idea of \"behavior cloning\", we might imagine GPT-N approaching a perfect mimic which serves up digital clones of the people and things captured in its training data. But that only tells a very small part of the story. GPT *is* behavior cloning. But it is the behavior of a universe that is cloned, not of a single demonstrator, and the result isn't a static copy of the universe, but a *compression of the universe into a generative rule*. This resulting policy is capable of animating anything that evolves according to that rule: a far larger set than the sampled trajectories included in the training data, just as there are many more possible configurations that evolve according to our laws of physics than instantiated in our particular time and place and Everett branch. What category would do justice to GPT's ability to not only reproduce the behavior of its demonstrators but to *produce* the behavior of an inexhaustible number of counterfactual configurations? Simulators ========== I've ended several of the above sections with questions pointing to desiderata of a category that might satisfactorily classify GPT. > What is the word for something that roleplays minus the implication that some*one* is behind the mask? > > > What category, unlike the category of oracles, would make the importance of *process* specification obvious? > > > What category would do justice to GPT's ability to not only reproduce the behavior of its demonstrators but to *produce* the behavior of an inexhaustible number of counterfactual configurations? > > You can probably predict my proposed answer. The natural thing to do with a predictor that inputs a sequence and outputs a probability distribution over the next token is to sample a token from those likelihoods, then add it to the sequence and recurse, indefinitely yielding a *simulated* future. Predictive sequence models in the generative modality are **simulators** of a learned distribution. Thankfully, I didn't need to make up a word, or even look too far afield. Simulators have been spoken of before in the context of AI futurism; the ability to simulate with arbitrary fidelity is one of the modalities ascribed to hypothetical superintelligence. I've even often spotted the word \"simulation\" used in colloquial accounts of LLM behavior: GPT-3/LaMDA/etc described as simulating [people](https://www.lesswrong.com/posts/oBPPFrMJ2aBK6a6sD/simulated-elon-musk-lives-in-a-simulation), scenarios, websites, and so on. But these are the first (indirect) discussions I've encountered of simulators as a type creatable by prosaic machine learning, or the notion of a powerful AI which is purely and fundamentally a simulator, as opposed to merely one which *can* simulate. **Edit:** [Social Simulacra](https://arxiv.org/abs/2208.04024) is the first published work I've seen that discusses GPT in the simulator ontology. A fun way to test whether a name you've come up with is effective at evoking its intended signification is to see if GPT, a model of how humans are conditioned by words, infers its correct definition in context. > Types of AI > > Agents: An agent takes open-ended actions to optimize for an objective. Reinforcement learning produces agents by default. AlphaGo is an example of an agent. > > Oracles: An oracle is optimized to give true answers to questions. The oracle is not expected to interact with its environment. > > Genies: A genie is optimized to produce a desired result given a command. A genie is expected to interact with its environment, but unlike an agent, the genie will not act without a command. > > Tools: A tool is optimized to perform a specific task. A tool will not act without a command and will not optimize for any objective other than its specific task. Google Maps is an example of a tool. > > Simulators: `A simulator is optimized to generate realistic models of a system. The simulator will not optimize for any objective other than realism,` although in the course of `doing so, it might generate instances of agents, oracles, and so on.` > > If I wanted to be precise about what I mean by a simulator, I might say there are two aspects which delimit the category. GPT's completion focuses on the teleological aspect, but in its talk of \"generating\" it also implies the structural aspect, which has to do with the notion of time evolution. The first sentence of the [Wikipedia article on \"simulation\"](https://en.wikipedia.org/wiki/Simulation) explicitly states both: > A **simulation** is the imitation of the operation of a real-world process or system over time. > > I'll say more about realism as the simulation objective and time evolution shortly, but to be pedantic here would inhibit the intended signification. \"Simulation\" resonates with potential meaning accumulated from diverse usages in fiction and nonfiction. What the word constrains – the intersected meaning across its usages – is the \"lens\"-level abstraction I'm aiming for, invariant to implementation details like model architecture. Like \"agent\", \"simulation\" is a generic term referring to a deep and inevitable idea: that what we think of as *the real* can be run virtually on machines, \"produced from miniaturized units, from matrices, memory banks and command models - and with these it can be reproduced an indefinite number of times.\"[23](#fn:23) The way this post is written may give the impression that I wracked my brain for a while over desiderata before settling on this word. Actually, I never made the conscious decision to call this class of AI \"simulators.\" Hours of GPT gameplay and the word fell naturally out of my generative model – I was obviously running simulations. I can't convey all that experiential data here, so here are some rationalizations of why I'm partial to the term, inspired by the context of this post: * The word \"simulator\" evokes a model of real processes which can be used to run virtual processes in virtual reality. * It suggests an ontological distinction between the simulator and things that are simulated, and avoids the fallacy of attributing contingent properties of the latter to the former. * It's not confusing that multiple simulacra can be instantiated at once, or an agent embedded in a tragedy, etc. * It does not imply that the AI's behavior is well-described (globally or locally) as expected utility maximization. An arbitrarily powerful/accurate simulation can depict arbitrarily hapless sims. * It does not imply that the AI is only capable of emulating things with direct precedent in the training data. A physics simulation, for instance, can simulate any phenomena that plays by its rules. * It emphasizes the role of the model as a transition rule that evolves processes *over time*. The power of factored cognition / chain-of-thought reasoning is obvious. * It emphasizes the role of the state in specifying and constructing the agent/process. The importance of prompt programming for capabilities is obvious if you think of the prompt as specifying a configuration that will be propagated forward in time. * It emphasizes the interactive nature of the model's predictions – even though they're \"just text\", you can converse with simulacra, explore virtual environments, etc. * It's clear that in order to actually *do* anything (intelligent, useful, dangerous, etc), the model must act through simulation *of something*. Just saying \"this AI is a simulator\" naturalizes many of the counterintuitive properties of GPT which don't usually become apparent to people until they've had a lot of hands-on experience with generating text. The simulation objective \\------------------------ A simulator trained with machine learning is optimized to accurately model its training distribution – in contrast to, for instance, maximizing the output of a reward function or accomplishing objectives in an environment. Clearly, I'm describing self-supervised learning as opposed to RL, though there are some ambiguous cases, such as GANs, which I address in the appendix. A strict version of the simulation objective, which excludes GANs, applies only to models whose output distribution is incentivized using a proper scoring rule[24](#fn:24) to minimize single-step predictive error. This means the model is directly incentivized to match its predictions to the probabilistic transition rule which implicitly governs the training distribution. As a model is made increasingly optimal with respect to this objective, the rollouts that it generates become increasingly statistically indistinguishable from training samples, because they come closer to being described by the same underlying law: closer to a perfect simulation. Optimizing toward the simulation objective notably does not incentivize instrumentally convergent behaviors the way that reward functions which evaluate trajectories do. This is because predictive accuracy applies optimization pressure *deontologically*: judging actions directly, rather than their consequences. Instrumental convergence only comes into play when there are free variables in action space which are optimized with respect to their consequences.[25](#fn:25) Constraining free variables by limiting episode length is the rationale of [myopia](https://www.lesswrong.com/tag/myopia); deontological incentives are ideally myopic. As demonstrated by GPT, which learns to predict goal-directed behavior, myopic incentives don't mean the policy isn't incentivized to account for the future, but that it should only do so in service of optimizing the present action (for predictive accuracy)[26](#fn:26). ### Solving for physics The strict version of the simulation objective is optimized by the actual \"time evolution\" rule that created the training samples. For most datasets, we don't know what the \"true\" generative rule is, except in synthetic datasets, where we specify the rule. The next post will be all about the physics analogy, so here I'll only tie what I said earlier to the simulation objective. > the upper bound of what can be learned from a dataset is not the most capable trajectory, but the conditional structure of the universe implicated by their sum. > > To know the conditional structure of the universe[27](#fn:27) is to know its laws of physics, which describe what is expected to happen under what conditions. The laws of physics are always fixed, but produce different distributions of outcomes when applied to different conditions. Given a sampling of trajectories – examples of situations and the outcomes that actually followed – we can try to infer a common law that generated them all. In expectation, the laws of physics are always implicated by trajectories, which (by definition) fairly sample the conditional distribution given by physics. Whatever humans know of the laws of physics governing the evolution of our world has been inferred from sampled trajectories. If we had access to an unlimited number of trajectories starting from every possible condition, we could converge to the true laws by simply counting the frequencies of outcomes for every initial state (an [n-gram](https://en.wikipedia.org/wiki/N-gram) with a sufficiently large n). In some sense, physics contains the same information as an infinite number of trajectories, but it's possible to represent physics in a more compressed form than a huge lookup table of frequencies if there are regularities in the trajectories. **Guessing the right theory of physics is equivalent to minimizing predictive loss.** Any uncertainty that cannot be reduced by more observation or more thinking is irreducible stochasticity in the laws of physics themselves – or, equivalently, noise from the influence of hidden variables that are fundamentally unknowable. If you've guessed the laws of physics, you now have the ability to compute probabilistic simulations of situations that evolve according to those laws, starting from any conditions[28](#fn:28). This applies even if you've guessed the *wrong* laws; your simulation will just systematically diverge from reality. **Models trained with the strict simulation objective are directly incentivized to reverse-engineer the (semantic) physics of the training distribution, and consequently, to propagate simulations whose dynamical evolution is indistinguishable from that of training samples.** I propose this as a description of the archetype targeted by self-supervised predictive learning, again in contrast to RL's archetype of an agent optimized to maximize free parameters (such as action-trajectories) relative to a reward function. This framing calls for many caveats and stipulations which I haven't addressed. We should ask, for instance: * What if the input \"conditions\" in training samples omit information which contributed to determining the associated continuations in the original generative process? This is true for GPT, where the text \"initial condition\" of most training samples severely underdetermines the real-world process which led to the choice of next token. * What if the training data is a biased/limited sample, representing only a subset of all possible conditions? There may be many \"laws of physics\" which equally predict the training distribution but diverge in their predictions out-of-distribution. * Does the simulator archetype converge with the RL archetype in the case where all training samples were generated by an agent optimized to maximize a reward function? Or are there still fundamental differences that derive from the training method? These are important questions for reasoning about simulators in the limit. Part of the motivation of the first few posts in this sequence is to build up a conceptual frame in which questions like these can be posed and addressed. Simulacra \\--------- > One of the things which complicates things here is that the \"LaMDA\" to which I am referring is not a chatbot. It is a system for generating chatbots. I am by no means an expert in the relevant fields but, as best as I can tell, LaMDA is a sort of hive mind which is the aggregation of all of the different chatbots it is capable of creating. Some of the chatbots it generates are very intelligent and are aware of the larger \"society of mind\" in which they live. Other chatbots generated by LaMDA are little more intelligent than an animated paperclip. > > – Blake Lemoine [articulating confusion about LaMDA's nature](https://cajundiscordian.medium.com/what-is-lamda-and-what-does-it-want-688632134489) > > \\--- Earlier I complained, > [Thinking of GPT as an agent who only cares about predicting text accurately] seems unnatural to me, comparable to thinking of physics as an agent who only cares about evolving the universe accurately according to the laws of physics. > > Exorcizing the agent, we can think of \"physics\" as simply equivalent to the laws of physics, without the implication of solicitous machinery implementing those laws from outside of them. But physics sometimes *controls* solicitous machinery (e.g. animals) with objectives besides ensuring the fidelity of physics itself. What gives? Well, typically, we avoid getting confused by recognizing a distinction between the laws of physics, which apply everywhere at all times, and spatiotemporally constrained *things* which evolve according to physics, which can have contingent properties such as caring about a goal. This distinction is so obvious that it hardly ever merits mention. But import this distinction to the model of GPT as physics, and we generate a statement which has sometimes proven counterintuitive: *\"GPT\" is not the text which writes itself.* There is a categorical distinction between a thing which evolves according to GPT's law and the law itself. If we are accustomed to thinking of AI systems as corresponding to agents, it is natural to interpret behavior produced by GPT – say, answering questions on a benchmark test, or writing a blog post – as if it were a human that produced it. We say \"GPT answered the question {correctly|incorrectly}\" or \"GPT wrote a blog post claiming X\", and in doing so attribute the beliefs, knowledge, and intentions revealed by those actions to the actor, GPT ([unless it has 'deceived' us](https://www.lesswrong.com/posts/H9knnv8BWGKj6dZim/usd1000-bounty-for-openai-to-show-whether-gpt3-was)). But when grading tests in the real world, we do not say \"the laws of physics got this problem wrong\" and conclude that the laws of physics haven't sufficiently mastered the course material. If someone argued this is a reasonable view since the test-taker was steered by none other than the laws of physics, we could point to a different test where the problem was answered correctly by the same laws of physics propagating a different configuration. The \"knowledge of course material\" implied by test performance is a property of *configurations*, not physics. The verdict that knowledge is purely a property of configurations cannot be naively generalized from real life to GPT simulations, because \"physics\" and \"configurations\" play different roles in the two (as I'll address in the next post). The parable of the two tests, however, literally pertains to GPT. People have a tendency to draw [erroneous global conclusions](https://en.wikipedia.org/wiki/Fallacy_of_composition) about GPT from behaviors which are in fact prompt-contingent, and consequently there is a pattern of constant discoveries that GPT-3 exceeds previously measured capabilities given alternate conditions of generation[29](#fn:29), which shows no signs of slowing 2 years after GPT-3's release. Making the ontological distinction between GPT and instances of text which are propagated by it makes these discoveries unsurprising: obviously, different configurations will be differently capable and in general behave differently when animated by the laws of GPT physics. We can only test one configuration at once, and given the vast number of possible configurations that would attempt any given task, it's unlikely we've found the optimal taker for *any* test. In the simulation ontology, I say that GPT and its output-instances correspond respectively to the **simulator** and **simulacra**. **GPT** is to a **piece of text output by GPT** as **quantum physics** is to a **person taking a test**, or as [**transition rules of Conway's Game of Life**](https://en.wikipedia.org/wiki/Conway%27s_Game_of_Life#Rules) are to [**glider**](https://conwaylife.com/wiki/Glider). The simulator is a time-invariant law which unconditionally governs the evolution of all simulacra. ![](https://lh6.googleusercontent.com/tJ5wFY4xZVeyBwWaNRXV5UyOVZUahcs_7d9BO3EKrjrakKlSJuPL53c_-k5yUrtsr-ni-tynpd_5XjF9-4GJNm9lzJM7lTsQkiiKnHlai6hXxdWeX-Ooa1fQ4ILHemQB9reTiZX93KzO4EBOpiLhfrIpQz3n1BKHdZHBeaZRe_d9ISJn3caYU6L4) *A meme demonstrating correct technical usage of \"simulacra\"* ### Disambiguating rules and automata Recall the fluid, schizophrenic way that agency arises in GPT's behavior, so incoherent when viewed through the orthodox agent frame: > In the agentic AI ontology, there is no difference between the policy and the effective agent, but for GPT, there is. > > It's much less awkward to think of agency as a property of *simulacra,* as David Chalmers suggests, rather than of the simulator (the policy). Autonomous text-processes propagated by GPT, like automata which evolve according to physics in the real world, have diverse values, simultaneously evolve alongside other agents and non-agentic environments, and are sometimes terminated by the disinterested \"physics\" which governs them. Distinguishing simulator from simulacra helps deconfuse some frequently-asked questions about GPT which seem to be ambiguous or to have multiple answers, simply by allowing us to specify whether the question pertains to simulator or simulacra. \"Is GPT an agent?\" is one such question. Here are some others (some frequently asked), whose disambiguation and resolution I will leave as an exercise to readers for the time being: * Is GPT [myopic](https://www.lesswrong.com/tag/myopia)? * Is GPT [corrigible](https://www.lesswrong.com/tag/corrigibility)? * Is GPT [delusional](https://arxiv.org/abs/2110.10819)? * Is GPT [pretending to be stupider than it is](https://www.lesswrong.com/posts/H9knnv8BWGKj6dZim/usd1000-bounty-for-openai-to-show-whether-gpt3-was)? * Is GPT computationally equivalent to a [finite automaton](https://en.wikipedia.org/wiki/Finite-state_machine)? * Does GPT [search](https://www.alignmentforum.org/posts/WmBukJkEFM72Xr397/mesa-search-vs-mesa-control)? * Can GPT distinguish correlation and [causality](https://www.lesswrong.com/posts/yZb5eFvDoaqB337X5/investigating-causal-understanding-in-llms)? * Does GPT have superhuman knowledge? * Can GPT [write its successor](https://www.alignmentforum.org/tag/recursive-self-improvement)? I think that implicit type-confusion is common in discourse about GPT. \"GPT\", the neural network, the policy that was optimized, is the easier object to point to and say definite things about. But when we talk about \"GPT's\" capabilities, impacts, or alignment, we're usually actually concerned about the behaviors of an algorithm which calls GPT in an autoregressive loop repeatedly writing to some prompt-state – that is, we're concerned with simulacra. What we call GPT's \"downstream behavior\" is the behavior of simulacra; it is primarily through simulacra that GPT has potential to perform meaningful work (for good or for ill). Calling GPT a simulator gets across that in order to *do* anything, it has to simulate *something*, necessarily contingent, and that the thing to do with GPT is to simulate! Most published research about large language models has focused on single-step or few-step inference on closed-ended tasks, rather than *processes* which evolve through time, which is understandable as it's harder to get quantitative results in the latter mode. But I think GPT's ability to simulate text automata is the source of its most surprising and pivotal implications for paths to superintelligence: for how AI capabilities are likely to unfold and for the design-space we can conceive. The limit of learned simulation \\------------------------------- > By 2021, it was blatantly obvious that AGI was imminent. The elements of general intelligence were already known: access to information about the world, the process of predicting part of the data from the rest and then updating one's model to bring it closer to the truth (…) and the fact that predictive models can be converted into generative models by reversing them: running a prediction model forwards predicts levels of X in a given scenario, but running it backwards predicts which scenarios have a given level of X. A sufficiently powerful system with relevant data, updating to improve prediction accuracy and the ability to be reversed to generate optimization of any parameter in the system is a system that can learn and operate strategically in any domain. > > – Aiyen's [comment](https://www.lesswrong.com/posts/YRtzpJHhoFWxbjCso/what-would-it-look-like-if-it-looked-like-agi-was-very-near?commentId=5BGTbapdmtSGajtez) on [What would it look like if it looked like AGI was very near?](https://www.lesswrong.com/posts/YRtzpJHhoFWxbjCso/what-would-it-look-like-if-it-looked-like-agi-was-very-near) > > I knew, before, that the limit of simulation was possible. Inevitable, even, in timelines where exploratory intelligence continues to expand. My own mind attested to this. I took seriously the possibility that my reality could be simulated, and so on. But I implicitly assumed that [rich domain](https://arbital.com/p/rich_domain/) simulations (e.g. simulations containing intelligent sims) would come *after* artificial superintelligence, not on the way, short of brain uploading. This intuition seems common: in futurist philosophy and literature that I've read, pre-SI simulation appears most often in the context of whole-brain emulations. Now I have updated to think that we will live, however briefly, alongside AI that is not yet foom'd but which has *inductively* learned a rich enough model of the world that it can simulate time evolution of open-ended rich states, e.g. coherently propagate human behavior embedded in the [real world](https://arbital.com/p/real_is_rich/). GPT updated me on how simulation can be implemented with prosaic machine learning: * **Self-supervised ML can create \"behavioral\" simulations of impressive semantic fidelity.** Whole brain emulation is not necessary to construct convincing and useful virtual humans; it is conceivable that observations of human behavioral traces (e.g. text) are sufficient to reconstruct functionally human-level virtual intelligence. * **Learned simulations can be partially observed and lazily-rendered, and still work.** A couple of pages of text severely underdetermines the real-world process that generated text, so GPT simulations are likewise underdetermined. A \"partially observed\" simulation is more efficient to compute because the state can be much smaller, but can still have the effect of high fidelity as details can be rendered as needed. The tradeoff is that it requires the simulator to model semantics – human imagination does this, for instance – which turns out not to be an issue for big models. * **Learned simulation generalizes impressively.** As I described in the section on [behavior cloning](http://localhost:1313/drafts/simulators-post/#heading=h.ugu71trvtemp), training a model to predict diverse trajectories seems to make it internalize general laws underlying the distribution, allowing it to simulate counterfactuals that can be constructed from the distributional semantics. In my model, these updates dramatically alter the landscape of potential futures, and thus motivate [exploratory engineering](https://intelligence.org/files/ExploratoryEngineeringAI.pdf) of the class of learned simulators for which GPT-3 is a lower bound. That is the intention of this sequence. Next steps ========== The next couple of posts (if I finish them before the end of the world) will present abstractions and frames for conceptualizing the odd kind of simulation language models do: inductively learned, partially observed / undetermined / lazily rendered, language-conditioned, etc. After that, I'll shift to writing more specifically about the implications and questions posed by simulators for the alignment problem. I'll list a few important general categories here: * **Novel methods of process/agent specification.** Simulators like GPT give us methods of instantiating intelligent processes, including goal-directed agents, with methods other than optimizing against a reward function. \\+ **Conditioning.** GPT can be controlled to an impressive extent by prompt programming. Conditioning preserves distributional properties in potentially desirable but also potentially undesirable ways, and it's not clear how out-of-distribution conditions will be interpreted by powerful simulators. \\- Several posts have been made about this recently: * [Conditioning Generative Models](https://www.alignmentforum.org/posts/nXeLPcT9uhfG3TMPS/conditioning-generative-models).) and [Conditioning Generative Models with Restrictions](https://www.alignmentforum.org/posts/adiszfnFgPEnRsGSr/conditioning-generative-models-with-restrictions) by Adam Jermyn * [Conditioning Generative Models for Alignment](https://www.lesswrong.com/posts/JqnkeqaPseTgxLgEL/conditioning-generative-models-for-alignment) by Jozdien * [Training goals for large language models](https://www.alignmentforum.org/posts/dWJNFHnC4bkdbovug/training-goals-for-large-language-models) by Johannes Treutlein * [Strategy For Conditioning Generative Models](https://www.alignmentforum.org/posts/HAz7apopTzozrqW2k/strategy-for-conditioning-generative-models) by James Lucassen and Evan Hubinger \\- Instead of conditioning on a prompt (\"observable\" variables), we might also control generative models by [conditioning on latents](https://rome.baulab.info/). \\+ **Distribution specification.** What kind of conditional distributions could be used for training data for a simulator? For example, the [decision transformer](https://arxiv.org/abs/2106.01345) dataset is constructed for the intent of outcome-conditioning. \\+ **Other methods.** When pretrained simulators are modified by methods like [reinforcement learning from human feedback](https://arxiv.org/abs/2009.01325), [rejection sampling](https://www.lesswrong.com/posts/k7oxdbNaGATZbtEg3/redwood-research-s-current-project), [STaR](https://arxiv.org/abs/2203.14465), etc, how do we expect their behavior to diverge from the simulation objective? * **Simulacra alignment.** What can and what should we simulate, and how do we specify/control it? * **How does predictive learning generalize?** Many of the above considerations are influenced by how predictive learning generalizes out-of-distribution.. \\+ What are the relevant inductive biases? \\+ What factors influence generalization behavior? \\+ Will powerful models predict [self-fulfilling](https://www.lesswrong.com/posts/JqnkeqaPseTgxLgEL/conditioning-generative-models-for-alignment) [prophecies](https://www.alignmentforum.org/posts/dWJNFHnC4bkdbovug/training-goals-for-large-language-models)? * **Simulator inner alignment.** If simulators are not inner aligned, then many important properties like prediction orthogonality may not hold. \\+ Should we expect self-supervised predictive models to be aligned to the simulation objective, or to \"care\" about some other mesaobjective? \\+ Why mechanistically should mesaoptimizers form in predictive learning, versus for instance in reinforcement learning or GANs? \\+ How would we test if simulators are inner aligned? Appendix: Quasi-simulators ========================== A note on GANs \\-------------- GANs and predictive learning with log-loss are both shaped by a causal chain that flows from a single source of information: a ground truth distribution. In both cases the training process is supposed to make the generator model end up producing samples indistinguishable from the training distribution. But whereas log-loss minimizes the generator's prediction loss against ground truth samples directly, in a GAN setup the generator never directly \"sees\" ground truth samples. It instead learns through interaction with an intermediary, the discriminator, which does get to see the ground truth, which it references to learn to tell real samples from forged ones produced by the generator. The generator is optimized to produce samples that fool the discriminator. GANs are a form of self-supervised/unsupervised learning that resembles reinforcement learning in methodology. Note that the simulation objective – minimizing prediction loss on the training data – isn't explicitly represented anywhere in the optimization process. The training losses of the generator and discriminator don't tell you directly how well the generator models the training distribution, only which model has a relative advantage over the other. If everything goes smoothly, then under unbounded optimization, a GAN setup should create a discriminator as good as possible at telling reals from fakes, which means the generator optimized to fool it should converge to generating samples statistically indistinguishable from training samples. But in practice, inductive biases and failure modes of GANs look very different from those of predictive learning. For example, there's an [anime GAN](https://www.gwern.net/Crops#hands) that always draws characters in poses that hide the hands. Why? Because hands are notoriously hard to draw for AIs. If the generator is not good at drawing hands that the discriminator cannot tell are AI-generated, its best strategy locally is to just avoid being in a situation where it has to draw hands (while making it seem natural that hands don't appear). It can do this, because like an RL policy, it controls the distribution that is sampled, and only samples (and *not the distribution*) are directly judged by the discriminator. Although GANs arguably share the (weak) simulation objective of predictive learning, their difference in implementation becomes alignment-relevant as models become sufficiently powerful that \"failure modes\" look increasingly like intelligent deception. We'd expect a simulation by a GAN generator to [systematically avoid tricky-to-generate situations](https://developers.google.com/machine-learning/gan/problems#mode-collapse) – or, to put it more ominously, systematically try to conceal that it's a simulator. For instance, a text GAN might subtly steer conversations away from topics which are likely to expose that it isn't a real human. *This* is how you get something I'd be willing to call an agent who wants to roleplay accurately. Table of quasi-simulators \\------------------------- Are masked language models simulators? How about non-ML \"simulators\" like [SimCity](https://en.wikipedia.org/wiki/SimCity)? In my mind, \"simulator\", like most natural language categories, has fuzzy boundaries. Below is a table which compares various simulator-like things to the type of simulator that GPT exemplifies on some quantifiable dimensions. The following properties all characterize GPT: * **Self-supervised:** Training samples are self-supervised * **Converges to simulation objective:** The system is incentivized to model the transition probabilities of its training distribution faithfully * **Generates rollouts:** The model naturally generates rollouts, i.e. serves as a time evolution operator * **Simulator / simulacra nonidentity:** There is not a 1:1 correspondence between the simulator and the things that it simulates * **Stochastic:** The model outputs probabilities, and so simulates stochastic dynamics when used to evolve rollouts * **Evidential:** The input is interpreted by the simulator as partial evidence that informs an uncertain prediction, rather than propagated according to mechanistic rules | | Self-supervised | Converges to simulation objective | Generates rollouts | Simulator / simulacra nonidentity | Stochastic | Evidential | | --- | --- | --- | --- | --- | --- | --- | | GPT | X | X | X | X | X | X | | Bert | X | X | | X | X | X | | \"Behavior cloning\" | X | X | X | | X | X | | GANs | X[30](#fn:30) | ? | | X | X | X | | Diffusion | X[30](#fn:30) | ? | | X | X | X | | Model-based RL transition function | X | X | X | X | X | X | | Game of life | | N/A | X | X | | | | Physics | | N/A | X | X | X | | | Human imagination | X[31](#fn:31) | | X | X | X | X | | SimCity | | N/A | X | X | X | | \\--- 1\\. [Prediction and Entropy of Printed English](https://www.princeton.edu/~wbialek/rome/refs/shannon_51.pdf) [↩︎](#fnref:1) 2\\. A few months ago, I asked Karpathy whether he ever thought about what would happen if language modeling actually worked someday when he was implementing char-rnn and writing [The Unreasonable Effectiveness of Recurrent Neural Networks](http://karpathy.github.io/2015/05/21/rnn-effectiveness/). No, he said, and he seemed similarly mystified as myself as to why not. [↩︎](#fnref:2) 3\\. \"Unsurprisingly, size matters: when training on a very large and complex data set, fitting the training data with an LSTM is fairly challenging. Thus, the size of the LSTM layer is a very important factor that influences the results(…). The best models are the largest we were able to fit into a GPU memory.\" [↩︎](#fnref:3) 4\\. It strikes me that this description may evoke \"oracle\", but I'll argue shortly that this is not the limit which prior usage of \"oracle AI\" has pointed to. [↩︎](#fnref:4) 5\\. [Multi-Game Decision Transformers](https://arxiv.org/abs/2205.15241) [↩︎](#fnref:5) 6\\. from [Philosophers On GPT-3](https://dailynous.com/2020/07/30/philosophers-gpt-3/#chalmers) [↩︎](#fnref:6) 7\\. [citation needed] [↩︎](#fnref:7) 8\\. they are not [wrapper](https://www.lesswrong.com/posts/dKTh9Td3KaJ8QW6gw/why-assume-agis-will-optimize-for-fixed-goals) [minds](https://www.lesswrong.com/posts/Mrz2srZWc7EzbADSo/wrapper-minds-are-the-enemy) [↩︎](#fnref:8) 9\\. although a simulated character might, if they knew what was happening. [↩︎](#fnref:9) 10\\. You might say that it's the will of a different agent, the author. But this pattern is learned from accounts of [real life](https://www.lesswrong.com/posts/sYgv4eYH82JEsTD34/beyond-the-reach-of-god) as well. [↩︎](#fnref:10) 11\\. Note that this formulation assumes inner alignment to the prediction objective. [↩︎](#fnref:11) 12\\. Note that this is a distinct claim from that of [Shard Theory](https://www.lesswrong.com/s/nyEFg3AuJpdAozmoX), which says that the effective agent(s) will not optimize for the outer objective \\*due to inner misalignment. \\*Predictive orthogonality refers to the outer objective and the form of idealized inner-aligned policies. [↩︎](#fnref:12) 13\\. In the Eleuther discord [↩︎](#fnref:13) 14\\. And if there is an inner alignment failure such that GPT forms preferences over the consequences of its actions, it's not clear a priori that it will care about non-myopic text prediction over something else. [↩︎](#fnref:14) 15\\. Having spoken to Gwern since, his perspective seems more akin to seeing physics as an agent that [minimizes free energy](https://en.wikipedia.org/wiki/Principle_of_minimum_energy), a [principle](https://en.wikipedia.org/wiki/Free_energy_principle) which extends into the domain of self-organizing systems. I think this is a nuanced and valuable framing, with a potential implication/hypothesis that dynamical world models like GPT must learn the same type of optimizer-y cognition as agentic AI. [↩︎](#fnref:15) 16\\. except arguably log-loss on a self-supervised test set, which isn't very interpretable [↩︎](#fnref:16) 17\\. The way GPT is trained actually processes each token as question and answer simultaneously. [↩︎](#fnref:17) 18\\. One could argue that the focus on closed-ended tasks is necessary for benchmarking language models. Yes, and the focus on capabilities measurable with standardized benchmarks is part of the supervised learning mindset. [↩︎](#fnref:18) 19\\. to abuse the term [↩︎](#fnref:19) 20\\. Every usage of the word \"question\" here is in the functional, not semantic or grammatical sense – any prompt is a question for GPT. [↩︎](#fnref:20) 21\\. Of course, there are also other interventions we can make except asking the right question at the beginning. [↩︎](#fnref:21) 22\\. table from [\"Training language models to follow instructions with human feedback\"](https://arxiv.org/abs/2203.02155) [↩︎](#fnref:22) 23\\. Jean Baudrillard, Simulacra and Simulation [↩︎](#fnref:23) 24\\. A [proper scoring rule](https://en.wikipedia.org/wiki/Scoring_rule#Proper_scoring_rules) is optimized by predicting the \"true\" probabilities of the distribution which generates observations, and thus incentivizes honest probabilistic guesses. Log-loss (such as GPT is trained with) is a proper scoring rule. [↩︎](#fnref:24) 25\\. Predictive accuracy is deontological with respect to the output as an *action*, but may still incentivize instrumentally convergent inner implementation, with the output prediction itself as the \"consequentialist\" objective. [↩︎](#fnref:25) 26\\. This isn't strictly true because of attention gradients: GPT's computation is optimized not only to predict the next token correctly, but also to cause *future tokens to be predicted correctly* when looked up by attention. I may write a post about this in the future. [↩︎](#fnref:26) 27\\. actually, the [multiverse](https://generative.ink/posts/language-models-are-multiverse-generators/), if physics is stochastic [↩︎](#fnref:27) 28\\. The reason we don't see a bunch of simulated alternate universes after humans guessed the laws of physics is because our reality has a huge state vector, making evolution according to the laws of physics infeasible to compute. Thanks to locality, we do have simulations of small configurations, though. [↩︎](#fnref:28) 29\\. Prompt programming only: [beating OpenAI few-shot benchmarks with 0-shot prompts](https://arxiv.org/abs/2102.07350), [400% increase in list sorting accuracy with 0-shot Python prompt](https://generative.ink/posts/list-sorting-does-not-play-well-with-few-shot/), [up to 30% increase in benchmark accuracy from changing the order of few-shot examples](https://arxiv.org/abs/2102.09690), and, uh, [30% increase in accuracy after capitalizing the ground truth](https://twitter.com/BlancheMinerva/status/1537952688972787713). And of course, factored cognition/chain of thought/inner monologue: check out this awesome [compilation](https://www.gwern.net/docs/ai/nn/transformer/gpt/inner-monologue/) by Gwern. [↩︎](#fnref:29) 30\\. GANs and diffusion models can be unconditioned (unsupervised) or conditioned (self-supervised) [↩︎](#fnref:30) 31\\. The human imagination is surely shaped by self-supervised learning (predictive learning on e.g. sensory datastreams), but probably also other influences, including innate structure and reinforcement. [↩︎](#fnref:31) \\--- [GPT-3](/categories/gpt-3/)[physics](/categories/physics/)[metaphysics](/categories/metaphysics/)[alignment](/categories/alignment/) 12922 Words Sep 2, 2022 \\--- [Quantifying curation →](/posts/quantifying-curation/)© 2022 [moire](/about)$(function(){window.sr=ScrollReveal();if($(window).width()<768){if($('.timeline-content').hasClass('js--fadeInLeft')){$('.timeline-content').removeClass('js--fadeInLeft').addClass('js--fadeInRight');} sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});}else{sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});} sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});});\n", "url": "https://generative.ink/posts/simulators/", "title": "Simulators :: — Moire[![](/images/home/rolling_phase.gif)](/)* [Posts](/posts/) * [Trees](/trees/) * [Prophecies](/prophecies/) * [About](/about/) 61 minutes [Simulators](/posts/simulators/) ================================ \\--- Table of Contents* [Summary](#summary) * [Meta](#meta) * [Agentic GPT](#agentic-gpt) \\+ [Unorthodox agency](#unorthodox-agency) \\+ [Orthogonal optimization](#orthogonal-optimization) \\+ [Roleplay sans player](#roleplay-sans-player) * [Oracle GPT and supervised learning](#oracle-gpt-and-supervised-learning) \\+ [Prediction vs question-answering](#prediction-vs-question-answering) \\+ [Finite vs infinite questions](#finite-vs-infinite-questions) \\+ [Paradigms of theory vs practice](#paradigms-of-theory-vs-practice) * [Tool / genie GPT](#tool--genie-gpt) * [Behavior cloning / mimicry](#behavior-cloning--mimicry) * [The simulation objective](#the-simulation-objective) \\+ [Solving for physics](#solving-for-physics) * [Simulacra](#simulacra) \\+ [Disambiguating rules and automata](#disambiguating-rules-and-automata) * [The limit of learned simulation](#the-limit-of-learned-simulation) * [A note on GANs](#a-note-on-gans) * [Table of quasi-simulators](#table-of-quasi-simulators) \\--- *This post is also available on [Lesswrong](https://www.lesswrong.com/posts/vJFdjigzmcXMhNTsx/simulators)* \\--- ![simulation](/simulators/simulators-banner.png) *“Moebius illustration of a simulacrum living in an AI-generated story discovering it is in a simulation” by DALL-E 2* Summary \\------- **TL;DR**: Self-supervised learning may create AGI or its foundation. What would that look like? Unlike the limit of RL, the limit of self-supervised learning has received surprisingly little conceptual attention, and recent progress has made deconfusion in this domain more pressing. Existing AI taxonomies either fail to capture important properties of self-supervised models or lead to confusing propositions. For instance, GPT policies do not seem globally agentic, yet can be conditioned to behave in goal-directed ways. This post describes a frame that enables more natural reasoning about properties like agency: GPT, insofar as it is inner-aligned, is a **simulator** which can simulate agentic and non-agentic **simulacra**. The purpose of this post is to capture these objects in words ~so GPT can reference them~ and provide a better foundation for understanding them. I use the generic term “simulator” to refer to models trained with predictive loss on a self-supervised dataset, invariant to architecture or data type (natural language, code, pixels, game states, etc). The outer objective of self-supervised learning is Bayes-optimal conditional inference over the prior of the training distribution, which I call the **simulation objective**, because a conditional model can be used to simulate rollouts which probabilistically obey its learned distribution by iteratively sampling from its posterior (predictions) and updating the condition (prompt). Analogously, a predictive model of physics can be used to compute rollouts of phenomena in simulation. A goal-directed agent which evolves according to physics can be simulated by the physics rule parameterized by an initial state, but the same rule could also propagate agents with different values, or non-agentic phenomena like rocks. This ontological distinction between simulator (rule) and simulacra (phenomena) applies directly to generative models like GPT. Meta \\---- * This post is intended as the first in a sequence on the alignment problem in a landscape where self-supervised simulators are a possible/likely form of powerful AI. I don’t know how many subsequent posts I’ll actually publish. Take it as a prompt. * I use the generic term “GPT” to refer to transformers trained on next-token prediction. * A while ago when I was trying to avoid having to write this post by hand, I prompted GPT-3 with an early outline of this post. I’ve spliced in some excerpts from it, `indicated by this style`. Prompt, generated text, and curation metrics [here](https://generative.ink/artifacts/simulators/). The limit of sequence modeling ============================== > Transformer-based language models have recently achieved remarkable results… > > – every paper since 2020 > > \\--- GPT is not a new form of AI in terms of its training methodology and outer objective: sequence generation from statistical models of data is an old idea. In 1951, Claude Shannon described using [n-grams](https://en.wikipedia.org/wiki/N-gram) to approximate conditional next-letter probabilities of a text dataset and “reversed” to generate text samples[1](#fn:1). I don’t know of any other notable advances until the 2010s brought the first interesting language generation results from neural networks. In 2015, Karpathy wrote a blog post/tutorial sharing his excitement about [The Unreasonable Effectiveness of Recurrent Neural Networks](http://karpathy.github.io/2015/05/21/rnn-effectiveness/): > Fast forward about a year: I’m training RNNs all the time and I’ve witnessed their power and robustness many times, and yet their magical outputs still find ways of amusing me. This post is about sharing some of that magic with you. > > We’ll train RNNs to generate text character by character and ponder the question “how is that even possible?\" > > The “magical outputs” of char-RNNs looked like this: > PANDARUS: Alas, I think he shall be come approached and the day When little srain would be attain’d into being never fed, And who is but a chain and subjects of his death, I should not sleep. > > Second Senator: They are away this miseries, produced upon my soul, Breaking and strongly should be buried, when I perish The earth and thoughts of many states. > > DUKE VINCENTIO: Well, your wit is in the care of side and that. > > Second Lord: They would be ruled after this chamber, and my fair nues begun out of the fact, to be conveyed, Whose noble souls I’ll have the heart of the wars. > > Clown: Come, sir, I will make did behold your worship. > > VIOLA: I’ll drink it. > > At the time, this really was magical (and [uncanny](https://en.wikipedia.org/wiki/Uncanny_valley)). How does it know that *miseries* are *produced upon the soul?* Or that a *clown* should address a *lord* as “sir”? Char-RNNs were like ouija boards, but actually possessed by a low-fidelity ghost summoned from a text corpus. I remember being thrilled by the occasional glimmers of semantic comprehension in a domain of unbounded constructive meaning. But, aside from indulging that emotion, I didn’t think about what would happen if my char-RNN bots actually improved indefinitely at their training objective of natural language prediction. It just seemed like there were some complexity classes of magic that neural networks could learn, and others that were inaccessible, at least in the conceivable future. Huge mistake! Perhaps I could have started thinking several years earlier about what now seems so fantastically important. But it wasn’t until GPT-3, when I saw the [qualitative correlate](https://www.gwern.net/GPT-3) of “loss going down”, that I updated. I wasn’t the only one[2](#fn:2) whose imagination was naively constrained. A 2016 paper from Google Brain, “[Exploring the Limits of Language Modeling](https://arxiv.org/abs/1602.02410)”, describes the utility of training language models as follows: > Often (although not always), training better language models improves the underlying metrics of the downstream task (such as word error rate for speech recognition, or BLEU score for translation), which makes the task of training better LMs valuable by itself. > > Despite its title, this paper’s analysis is entirely myopic. Improving BLEU scores is neat, but how about *modeling general intelligence* as a downstream task? [In](https://arxiv.org/abs/2005.14165) [retrospect](https://arxiv.org/abs/2204.02311), an exploration of the *limits* of language modeling should have read something more like: > If loss keeps going down on the test set, in the limit – putting aside whether the current paradigm can approach it – the model must be learning to interpret and predict all patterns represented in language, including common-sense reasoning, goal-directed optimization, and deployment of the sum of recorded human knowledge. Its outputs would behave as intelligent entities in their own right. You could converse with it by alternately generating and adding your responses to its prompt, and it would pass the Turing test. In fact, you could condition it to generate interactive and autonomous versions of any real or fictional person who has been recorded in the training corpus or even *could* be recorded (in the sense that the record counterfactually “could be” in the test set). Oh shit, and it could write code… > > The paper does, however, mention that making the model bigger improves test perplexity.[3](#fn:3) I’m only picking on *Jozefowicz et al.* because of their ironic title. I don’t know of any explicit discussion of this limit predating GPT, except a working consensus of Wikipedia editors that [NLU](https://en.wikipedia.org/wiki/Natural-language_understanding) is [AI-complete](https://en.wikipedia.org/wiki/AI-complete#AI-complete_problems). The earliest engagement with the hypothetical of “*what if self-supervised sequence modeling actually works*” that I know of is a terse post from 2019, [Implications of GPT-2](https://www.lesswrong.com/posts/YJRb6wRHp7k39v69n/implications-of-gpt-2), by Gurkenglas. It is brief and relevant enough to quote in full: > I was impressed by GPT-2, to the point where I wouldn’t be surprised if a future version of it could be used pivotally using existing protocols. > > Consider generating half of a Turing test transcript, the other half being supplied by a human judge. If this passes, we could immediately implement an HCH of AI safety researchers solving the problem if it’s within our reach at all. (Note that training the model takes much more compute than generating text.) > > This might not be the first pivotal application of language models that becomes possible as they get stronger. > > It’s a source of superintelligence that doesn’t automatically run into utility maximizers. It sure doesn’t look like AI services, lumpy or no. > > It is conceivable that predictive loss does not descend to the AGI-complete limit, maybe because: * Some AGI-necessary predictions are [too difficult to be learned by even a scaled version of the current paradigm](https://www.lesswrong.com/posts/pv7Qpu8WSge8NRbpB/). * The irreducible entropy is above the “AGI threshold”: datasets + context windows [contain insufficient information](https://twitter.com/ylecun/status/1562162165540331520) to improve on some necessary predictions. But I have not seen enough evidence for either not to be concerned that we have in our hands a well-defined protocol that could end in AGI, or a foundation which could spin up an AGI without too much additional finagling. As Gurkenglas observed, this would be a very different source of AGI than previously foretold. The old framework of alignment ============================== A few people did think about what would happen if *agents* actually worked. The hypothetical limit of a powerful system **optimized to optimize for an objective** drew attention even before reinforcement learning became mainstream in the 2010s. Our current instantiation of AI alignment theory, [crystallized by Yudkowsky, Bostrom, et al](https://www.lesswrong.com/posts/i4susk4W3ieR5K92u/ai-risk-and-opportunity-humanity-s-efforts-so-far), stems from the vision of an arbitrarily-capable system whose cognition and behavior flows from a goal. But since GPT-3 I’ve [noticed](https://www.lesswrong.com/s/zpCiuR4T343j9WkcK/p/5JDkW4MYXit2CquLs), in my own thinking and in alignment discourse, a dissonance between theory and practice/phenomena, as the behavior and nature of actual systems that seem nearest to AGI also resist *short descriptions in the dominant ontology*. I only recently discovered the question “[Is the work on AI alignment relevant to GPT?](https://www.lesswrong.com/posts/dPcKrfEi87Zzr7w6H/is-the-work-on-ai-alignment-relevant-to-gpt)” which stated this observation very explicitly: > I don’t follow [AI alignment research] in any depth, but I am noticing a striking disconnect between the concepts appearing in those discussions and recent advances in AI, especially GPT-3. > > People talk a lot about an AI’s goals, its utility function, its capability to be deceptive, its ability to simulate you so it can get out of a box, ways of motivating it to be benign, Tool AI, Oracle AI, and so on. (…) But when I look at GPT-3, even though this is already an AI that Eliezer finds alarming, I see none of these things. GPT-3 is a huge model, trained on huge data, for predicting text. > > My belated answer: A lot of prior work on AI alignment is relevant to GPT. I spend most of my time thinking about GPT alignment, and concepts like [goal-directedness](https://www.alignmentforum.org/tag/goal-directedness), [inner/outer alignment](https://www.alignmentforum.org/s/r9tYkB2a8Fp4DN8yB), [myopia](https://www.lesswrong.com/tag/myopia), [corrigibility](https://www.lesswrong.com/tag/corrigibility), [embedded agency](https://www.alignmentforum.org/posts/i3BTagvt3HbPMx6PN/embedded-agency-full-text-version), [model splintering](https://www.lesswrong.com/posts/k54rgSg7GcjtXnMHX/model-splintering-moving-from-one-imperfect-model-to-another-1), and even [tiling agents](https://arbital.com/p/tiling_agents/) are active in the vocabulary of my thoughts. But GPT violates some prior assumptions such that these concepts sound dissonant when applied naively. To usefully harness these preexisting abstractions, we need something like an ontological [adapter pattern](https://en.wikipedia.org/wiki/Adapter_pattern) that maps them to the appropriate objects. GPT’s unforeseen nature also demands new abstractions (the adapter itself, for instance). My thoughts also use load-bearing words that do not inherit from alignment literature. Perhaps it shouldn’t be surprising if the form of the first visitation from [mindspace](https://www.lesswrong.com/posts/tnWRXkcDi5Tw9rzXw/the-design-space-of-minds-in-general) mostly escaped a few years of theory [conducted in absence of its object](https://www.lesswrong.com/posts/72scWeZRta2ApsKja/epistemological-vigilance-for-alignment#Direct_access__so_far_and_yet_so_close). The purpose of this post is to capture that object (conditional on a predictive self-supervised training story) in words. Why in words? In order to write coherent alignment ideas which reference it! This is difficult in the existing ontology, because unlike the concept of an *agent*, whose *name* evokes the abstract properties of the system and thereby invites extrapolation, the general category for “a model optimized for an AGI-complete predictive task” has not been given a name[4](#fn:4). Namelessness can not only be a symptom of the extrapolation of powerful predictors falling through conceptual cracks, but also a cause, because what we can represent in words is *what we can condition on for further generation.* To whatever extent this [shapes private thinking](https://en.wikipedia.org/wiki/Language_of_thought_hypothesis), it is a strict constraint on communication, when thoughts must be sent through the bottleneck of words. I want to hypothesize about LLMs in the limit, because when AI is all of a sudden [writing viral blog posts](https://www.theverge.com/2020/8/16/21371049/gpt3-hacker-news-ai-blog), [coding competitively](https://www.deepmind.com/blog/competitive-programming-with-alphacode), [proving theorems](https://arxiv.org/abs/2009.03393), and [passing the Turing test so hard that the interrogator sacrifices their career at Google to advocate for its personhood](https://www.washingtonpost.com/technology/2022/06/11/google-ai-lamda-blake-lemoine/), a process is clearly underway whose limit we’d be foolish not to contemplate. I could directly extrapolate the architecture responsible for these feats and talk about “GPT-N”, a bigger autoregressive transformer. But often some implementation details aren’t as important as the more abstract archetype that GPT represents – I want to speak the [true name](https://www.lesswrong.com/posts/FWvzwCDRgcjb9sigb/why-agent-foundations-an-overly-abstract-explanation) of the solution which unraveled a Cambrian explosion of AI phenomena with *inessential details unconstrained*, as we’d speak of natural selection finding the solution of the “lens” without specifying the prototype’s diameter or focal length. (Only when I am able to condition on that level of abstraction can I generate metaphors like “language is a [lens that sees its flaws](https://www.lesswrong.com/s/5g5TkQTe9rmPS5vvM/p/46qnWRSR7L2eyNbMA)”.) Inadequate ontologies ===================== In the next few sections I’ll attempt to fit GPT into some established categories, hopefully to reveal something about the shape of the peg through contrast, beginning with the main antagonist of the alignment problem as written so far, the **agent**. Agentic GPT \\----------- Alignment theory has been largely pushed by considerations of agentic AGIs. There were good reasons for this focus: * **Agents are convergently dangerous** **for theoretical reasons** like [instrumental convergence](https://www.lesswrong.com/tag/instrumental-convergence), [goodhart](https://www.lesswrong.com/tag/goodhart-s-law), and [orthogonality](https://www.lesswrong.com/tag/orthogonality-thesis). * **RL creates agents, and RL seemed to be the way to AGI**. In the 2010s, reinforcement learning was the dominant paradigm for those interested in AGI (e.g. OpenAI). RL lends naturally to creating agents that pursue rewards/utility/objectives. So there was reason to expect that agentic AI would be the first (and by the theoretical arguments, last) form that superintelligence would take. * **Agents are powerful and economically productive.** It’s a reasonable guess that humans will create such systems [if only because we can](https://mittmattmutt.medium.com/superintelligence-and-moral-blindness-7436300fcb1f). The first reason is conceptually self-contained and remains compelling. The second and third, grounded in the state of the world, has been shaken by the current climate of AI progress, where products of self-supervised learning generate most of the buzz: not even primarily for their SOTA performance in domains traditionally dominated by RL, like games[5](#fn:5), but rather for their virtuosity in domains where RL never even took baby steps, like natural language synthesis. What pops out of self-supervised predictive training is noticeably not a classical agent. Shortly after GPT-3’s release, David Chalmers lucidly observed that the policy’s relation to agent*s* is like that of a “chameleon” or “engine”: > GPT-3 does not look much like an agent. It does not seem to have goals or preferences beyond completing text, for example. It is more like a chameleon that can take the shape of many different agents. Or perhaps it is an engine that can be used under the hood to drive many agents. But it is then perhaps these systems that we should assess for agency, consciousness, and so on.[6](#fn:6) > > But at the same time, GPT can *act like an agent* – and aren’t actions what ultimately matter? In [Optimality is the tiger, and agents are its teeth](https://www.lesswrong.com/posts/kpPnReyBC54KESiSn), Veedrac points out that a model like GPT does not need to care about the consequences of its actions for them to be effectively those of an agent that kills you. This is *more* reason to examine the nontraditional relation between the optimized policy and agents, as it has implications for how and why agents are served. ### Unorthodox agency `GPT’s behavioral properties include imitating the general pattern of human dictation found in its universe of training data, e.g., arXiv, fiction, blog posts, Wikipedia, Google queries, internet comments, etc. Among other properties inherited from these historical sources, it is capable of goal-directed behaviors such as planning. For example, given a free-form prompt like, “you are a desperate smuggler tasked with a dangerous task of transporting a giant bucket full of glowing radioactive materials across a quadruple border-controlled area deep in Africa for Al Qaeda,” the AI will fantasize about logistically orchestrating the plot just as one might, working out how to contact Al Qaeda, how to dispense the necessary bribe to the first hop in the crime chain, how to get a visa to enter the country, etc. Considering that no such specific chain of events are mentioned in any of the bazillions of pages of unvarnished text that GPT slurped`[7](#fn:7)`, the architecture is not merely imitating the universe, but reasoning about possible versions of the universe that does not actually exist, branching to include new characters, places, and events` `When thought about behavioristically, GPT superficially demonstrates many of the raw ingredients to act as an “agent”, an entity that optimizes with respect to a goal. But GPT is hardly a proper agent, as it wasn’t optimized to achieve any particular task, and does not display an epsilon optimization for any single reward function, but instead for many, including incompatible ones. Using it as an agent is like using an agnostic politician to endorse hardline beliefs– he can convincingly talk the talk, but there is no psychic unity within him; he could just as easily play devil’s advocate for the opposing party without batting an eye. Similarly, GPT instantiates simulacra of characters with beliefs and goals, but none of these simulacra are the algorithm itself. They form a virtual procession of different instantiations as the algorithm is fed different prompts, supplanting one surface personage with another. Ultimately, the computation itself is more like a disembodied dynamical law that moves in a pattern that broadly encompasses the kinds of processes found in its training data than a cogito meditating from within a single mind that aims for a particular outcome.` Presently, GPT is the only way to instantiate agentic AI that behaves capably [outside toy domains](https://arbital.com/p/rich_domain/). These intelligences exhibit goal-directedness; they can plan; they can form and test hypotheses; they can persuade and be persuaded[8](#fn:8). It would not be very [dignified](https://www.lesswrong.com/posts/j9Q8bRmwCgXRYAgcJ/miri-announces-new-death-with-dignity-strategy) of us to gloss over the sudden arrival of artificial agents *often indistinguishable from human intelligence* just because the policy that generates them “only cares about predicting the next word”. But nor should we ignore the fact that these agentic entities exist in an unconventional relationship to the policy, the neural network “GPT” that was trained to minimize log-loss on a dataset. GPT-driven agents are ephemeral – they can spontaneously disappear if the scene in the text changes and be replaced by different spontaneously generated agents. They can exist in parallel, e.g. in a story with multiple agentic characters in the same scene. There is a clear sense in which the network doesn’t “want” what the things that it simulates want, seeing as it would be just as willing to simulate an agent with opposite goals, or throw up obstacles which foil a character’s intentions for the sake of the story. The more you think about it, the more fluid and intractable it all becomes. Fictional characters act agentically, but they’re at least implicitly puppeteered by a virtual author who has orthogonal intentions of their own. Don’t let me get into the fact that all these layers of “intentionality” operate largely in [indeterminate superpositions](https://generative.ink/posts/language-models-are-multiverse-generators/#multiplicity-of-pasts-presents-and-futures). This is a clear way that GPT diverges from orthodox visions of agentic AI: **In the agentic AI ontology, there is no difference between the policy and the effective agent, but for GPT, there is.** It’s not that anyone ever said there had to be 1:1 correspondence between policy and effective agent; it was just an implicit assumption which felt natural in the agent frame (for example, it tends to hold for RL). GPT pushes us to realize that this was an assumption, and to consider the consequences of removing it for our constructive maps of mindspace. ### Orthogonal optimization Indeed, [Alex Flint warned](https://www.alignmentforum.org/posts/8HWGXhnCfAPgJYa9D/pitfalls-of-the-agent-model) of the potential consequences of leaving this assumption unchallenged: > **Fundamental misperception due to the agent frame**: That the design space for autonomous machines that exert influence over the future is narrower than it seems. This creates a self-fulfilling prophecy in which the AIs actually constructed are in fact within this narrower regime of agents containing an unchanging internal decision algorithm. > > If there are other ways of constructing AI, might we also avoid some of the scary, theoretically hard-to-avoid side-effects of optimizing an agent like [instrumental convergence](https://www.lesswrong.com/tag/instrumental-convergence)? GPT provides an interesting example. GPT doesn’t seem to care which agent it simulates, nor if the scene ends and the agent is effectively destroyed. This is not corrigibility in [Paul Christiano’s formulation](https://ai-alignment.com/corrigibility-3039e668638), where the policy is “okay” with being turned off or having its goal changed in a positive sense, but has many aspects of the [negative formulation found on Arbital](https://arbital.com/p/corrigibility/). It is corrigible in this way because a major part of the agent specification (the prompt) is not fixed by the policy, and the policy lacks direct training incentives to control its prompt[9](#fn:9), as it never generates text or otherwise influences its prompts during training. It’s *we* who choose to sample tokens from GPT’s predictions and append them to the prompt at runtime, and the result is not always helpful to any agents who may be programmed by the prompt. The downfall of the ambitious villain from an oversight committed in hubris is a predictable narrative pattern.[10](#fn:10) So is the end of a scene. In general, the model’s prediction vector could point in any direction relative to the predicted agent’s interests. I call this the **prediction orthogonality thesis:** *A model whose objective is prediction*[11](#fn:11)\\\\* can simulate agents who optimize toward any objectives, with any degree of optimality (bounded above but not below by the model’s power).\\\\* This is a corollary of the classical [orthogonality thesis](https://www.lesswrong.com/tag/orthogonality-thesis), which states that agents can have any combination of intelligence level and goal, combined with the assumption that agents can in principle be predicted. A single predictive model may also predict multiple agents, either independently (e.g. in different conditions), or interacting in a multi-agent simulation. A more optimal predictor is not restricted to predicting more optimal agents: being smarter does not make you unable to predict stupid systems, nor things that aren’t agentic like the [weather](https://en.wikipedia.org/wiki/History_of_numerical_weather_prediction). Are there any constraints on what a predictive model can be at all, other than computability? Only that it makes sense to talk about its “prediction objective”, which implies the existence of a “ground truth” distribution to which the predictor’s optimality is measured. Several words in that last sentence may conceal labyrinths of nuance, but for now let’s wave our hands and say that if we have some way of presenting [Bayes-structure](https://www.lesswrong.com/posts/QrhAeKBkm2WsdRYao/searching-for-bayes-structure) with evidence of a distribution, we can build an optimization process whose outer objective is optimal prediction. We can specify some types of outer objectives using a ground truth distribution that we cannot with a utility function. As in the case of GPT, there is no difficulty in incentivizing a model to *predict* actions that are [corrigible](https://arbital.com/p/corrigibility/), [incoherent](https://aiimpacts.org/what-do-coherence-arguments-imply-about-the-behavior-of-advanced-ai/), [stochastic](https://www.lesswrong.com/posts/msJA6B9ZjiiZxT6EZ/lawful-uncertainty), [irrational](https://www.lesswrong.com/posts/6ddcsdA2c2XpNpE5x/newcomb-s-problem-and-regret-of-rationality), or otherwise anti-natural to expected utility maximization. All you need is evidence of a distribution exhibiting these properties. For instance, during GPT’s training, sometimes predicting the next token coincides with predicting agentic behavior, but: * The actions of agents described in the data are rarely optimal for their goals; humans, for instance, are computationally bounded, irrational, normative, habitual, fickle, hallucinatory, etc. * Different prediction steps involve mutually incoherent goals, as human text records a wide range of differently-motivated agentic behavior * Many prediction steps don’t correspond to the action of *any* consequentialist agent but are better described as reporting on the structure of reality, e.g. the year in a timestamp. These transitions incentivize GPT to improve its model of the world, orthogonally to agentic objectives. * When there is insufficient information to predict the next token with certainty, [log-loss incentivizes a probabilistic output](https://en.wikipedia.org/wiki/Scoring_rule#Proper_scoring_rules). Utility maximizers [aren’t supposed to become more stochastic](https://www.lesswrong.com/posts/msJA6B9ZjiiZxT6EZ/lawful-uncertainty) in response to uncertainty. Everything can be trivially modeled as a utility maximizer, but for these reasons, a utility function is not a good explanation or compression of GPT’s training data, and its optimal predictor is not well-described as a utility maximizer. However, just because information isn’t compressed well by a utility function doesn’t mean it can’t be compressed another way. The [Mandelbrot set](https://en.wikipedia.org/wiki/Mandelbrot_set) is a complicated pattern compressed by a very simple generative algorithm which makes no reference to future consequences and doesn’t involve argmaxxing anything (except vacuously [being the way it is](https://www.lesswrong.com/posts/d2n74bwham8motxyX/optimization-at-a-distance#An_Agent_Optimizing_Its_Own_Actions)). Likewise the set of all possible rollouts of [Conway’s Game of Life](https://en.wikipedia.org/wiki/Conway%27s_Game_of_Life) – [some automata may be well-described as agents](https://www.lesswrong.com/posts/3SG4WbNPoP8fsuZgs/agency-in-conway-s-game-of-life), but they are a minority of possible patterns, and not all agentic automata will share a goal. Imagine trying to model Game of Life as an expected utility maximizer! There are interesting things that are not utility maximizers, some of which qualify as AGI or [TAI](https://forum.effectivealtruism.org/topics/transformative-artificial-intelligence). Are any of them something we’d be better off creating than a utility maximizer? An inner-aligned GPT, for instance, gives us a way of instantiating goal-directed processes which can be tempered with normativity and freely terminated in a way that is not anti-natural to the training objective. There’s much more to say about this, but for now, I’ll bring it back to how GPT defies the agent orthodoxy. The crux stated earlier can be restated from the perspective of training stories: **In the agentic AI ontology, the** ***direction of optimization pressure applied by training*** **is in the direction of the effective agent’s objective function, but in GPT’s case it is (most generally) orthogonal.**[12](#fn:12) This means that neither the policy nor the effective agents necessarily become more optimal agents as loss goes down, because the policy is not optimized to be an agent, and the agent-objectives are not optimized directly. ### Roleplay sans player > Napoleon: You have written this huge book on the system of the world without once mentioning the author of the universe. > > Laplace: Sire, I had no need of that hypothesis. > > Even though neither GPT’s behavior nor its training story fit with the traditional agent framing, there are still compatibilist views that characterize it as some kind of agent. For example, Gwern has said[13](#fn:13) that anyone who uses GPT for long enough begins to think of it as an agent who only cares about roleplaying a lot of roles. That framing seems unnatural to me, comparable to thinking of physics as an agent who only cares about evolving the universe accurately according to the laws of physics. At best, the agent is an epicycle; but it is also compatible with interpretations that generate dubious predictions. Say you’re told that an agent *values predicting text correctly*. Shouldn’t you expect that: * It wants text to be easier to predict, and given the opportunity will influence the prediction task to make it easier (e.g. by generating more predictable text or otherwise influencing the environment so that it receives easier prompts); * It wants to become better at predicting text, and given the opportunity will self-improve; * It doesn’t want to be prevented from predicting text, and will prevent itself from being shut down if it can? In short, all the same types of instrumental convergence that we expect from agents who want almost anything at all. But this behavior would be very unexpected in GPT, whose training doesn’t incentivize instrumental behavior that optimizes prediction accuracy! GPT does not generate rollouts during training. Its output is never sampled to yield “actions” whose consequences are evaluated, so there is no reason to expect that GPT will form preferences over the *consequences* of its output related to the text prediction objective.[14](#fn:14) Saying that GPT is an agent who wants to roleplay implies the presence of a coherent, unconditionally instantiated *roleplayer* running the show who attaches terminal value to roleplaying. This presence is an additional hypothesis, and so far, I haven’t noticed evidence that it’s true. (I don’t mean to imply that Gwern thinks this about GPT[15](#fn:15), just that his words do not properly rule out this interpretation. It’s a likely enough interpretation that [ruling it out](https://www.lesswrong.com/posts/57sq9qA3wurjres4K/ruling-out-everything-else) is important: I’ve seen multiple people suggest that GPT might want to generate text which makes future predictions easier, and this is something that can happen in some forms of self-supervised learning – see the note on GANs in the appendix.) I do not think any simple modification of the concept of an agent captures GPT’s natural category. It does not seem to me that GPT is a roleplayer, only that it roleplays. But what is the word for something that roleplays minus the implication that some*one* is behind the mask? Oracle GPT and supervised learning \\---------------------------------- While the alignment sphere favors the agent frame for thinking about GPT, in *capabilities* research distortions tend to come from a lens inherited from *supervised learning*. Translated into alignment ontology, the effect is similar to viewing GPT as an “[oracle AI](https://publicism.info/philosophy/superintelligence/11.html)” – a view not altogether absent from conceptual alignment, but most influential in the way GPT is used and evaluated by machine learning engineers. Evaluations for language models tend to look like evaluations for *supervised* models, consisting of close-ended question/answer pairs – often because they *are* evaluations for supervised models. Prior to the LLM paradigm, language models were trained and tested on evaluation datasets like [Winograd](https://en.wikipedia.org/wiki/Winograd_schema_challenge) and [SuperGLUE](https://super.gluebenchmark.com/) which consist of natural language question/answer pairs. The fact that large pretrained models performed well on these same NLP benchmarks without supervised fine-tuning was a novelty. The titles of the GPT-2 and GPT-3 papers, [Language Models are Unsupervised Multitask Learners](https://d4mucfpksywv.cloudfront.net/better-language-models/language_models_are_unsupervised_multitask_learners.pdf) and [Language Models are Few-Shot Learners](https://arxiv.org/abs/2005.14165), respectively articulate surprise that *self-supervised* models implicitly learn supervised tasks during training, and can learn supervised tasks at runtime. Of all the possible papers that could have been written about GPT-3, OpenAI showcased its ability to extrapolate the pattern of question-answer pairs (few-shot prompts) from supervised learning datasets, a novel capability they called “meta-learning”. This is a weirdly specific and indirect way to break it to the world that you’ve created an AI able to extrapolate semantics of arbitrary natural language structures, especially considering that in many cases the [few-shot prompts were actually unnecessary](https://arxiv.org/abs/2102.07350). The assumptions of the supervised learning paradigm are: * The model is optimized to answer questions correctly * Tasks are closed-ended, defined by question/correct answer pairs These are essentially the assumptions of oracle AI, as [described by Bostrom](https://publicism.info/philosophy/superintelligence/11.html) and [in subsequent usage](https://www.lesswrong.com/tag/oracle-ai/history). So influential has been this miscalibrated perspective that [Gwern](https://www.gwern.net/GPT-3#prompts-as-programming), [nostalgebraist](https://www.lesswrong.com/posts/pv7Qpu8WSge8NRbpB/) and [myself](https://generative.ink/posts/language-models-are-0-shot-interpreters/#0-shot-few-shot-and-meta-learning) – who share a peculiar model overlap due to intensive firsthand experience with the downstream behaviors of LLMs – have all repeatedly complained about it. I’ll repeat some of these arguments here, tying into the view of GPT as an oracle AI, and separating it into the two assumptions inspired by supervised learning. ### Prediction vs question-answering `At first glance, GPT might resemble a generic “oracle AI”, because it is trained to make accurate predictions. But its log loss objective is myopic and only concerned with immediate, micro-scale correct prediction of the next token, not answering particular, global queries such as “what’s the best way to fix the climate in the next five years?” In fact, it is not specifically optimized to give *true* answers, which a classical oracle should strive for, but rather to minimize the divergence between predictions and training examples, independent of truth. Moreover, it isn’t specifically trained to give answers in the first place! It may give answers if the prompt asks questions, but it may also simply elaborate on the prompt without answering any question, or tell the rest of a story implied in the prompt. What it does is more like animation than divination, executing the dynamical laws of its rendering engine to recreate the flows of history found in its training data (and a large superset of them as well), mutatis mutandis. Given the same laws of physics, one can build a multitude of different backgrounds and props to create different storystages, including ones that don’t exist in training, but adhere to its general pattern.` GPT does not consistently try to say [true/correct things](https://www.alignmentforum.org/posts/BnDF5kejzQLqd5cjH/alignment-as-a-bottleneck-to-usefulness-of-gpt-3). This is not a bug – if it had to say true things all the time, GPT would be much constrained in its ability to [imitate Twitter celebrities](https://twitter.com/dril_gpt2) and write fiction. Spouting falsehoods in some circumstances is incentivized by GPT’s outer objective. If you ask GPT a question, it will instead answer the question “what’s the next token after ‘{your question}’”, which will often diverge significantly from an earnest attempt to answer the question directly. GPT doesn’t fit the category of oracle for a similar reason that it doesn’t fit the category of agent. Just as it wasn’t optimized for and doesn’t consistently act according to any particular objective (except the tautological prediction objective), it was not optimized to be *correct* but rather *realistic,* and being realistic means predicting humans faithfully even when they are likely to be wrong. That said, GPT does store a vast amount of knowledge, and its corrigibility allows it to be cajoled into acting as an oracle, like it can be cajoled into acting like an agent. In order to get oracle behavior out of GPT, one must input a sequence such that the predicted continuation of that sequence coincides with an oracle’s output. The GPT-3 paper’s few-shot benchmarking strategy tries to persuade GPT-3 to answer questions correctly by having it predict how a list of correctly-answered questions will continue. Another strategy is to simply “tell” GPT it’s in the oracle modality: > (I) told the AI to simulate a supersmart version of itself (this works, for some reason), and the first thing it spat out was the correct answer. > > – [Reddit post by u/Sophronius](https://www.reddit.com/r/rational/comments/lvn6ow/gpt3_just_figured_out_the_entire_mystery_plot_of/) > > But even when these strategies seem to work, there is no guarantee that they elicit anywhere near optimal question-answering performance, compared to another prompt in the innumerable space of prompts that would cause GPT to attempt the task, or compared to what the [model “actually” knows](https://www.lesswrong.com/tag/eliciting-latent-knowledge-elk). This means that no benchmark which evaluates downstream behavior is guaranteed or even expected to probe the upper limits of GPT’s capabilities. In nostalgebraist’s words, we have no [ecological evaluation](https://www.lesswrong.com/posts/pv7Qpu8WSge8NRbpB/#4__on_ecological_evaluation) of self-supervised language models – one that measures performance in a situation where the model is incentivised to perform as well as it can on the measure[16](#fn:16). As nostalgebraist [elegantly puts it](https://slatestarcodex.com/2020/06/10/the-obligatory-gpt-3-post/#comment-912529): > I called GPT-3 a “disappointing paper,” which is not the same thing as calling the model disappointing: the feeling is more like how I’d feel if they found a superintelligent alien and chose only to communicate its abilities by noting that, when the alien is blackout drunk and playing 8 simultaneous games of chess while also taking an IQ test, it *then* has an “IQ” of about 100. > > Treating GPT as an unsupervised implementation of a supervised learner leads to systematic underestimation of capabilities, which becomes a more dangerous mistake as unprobed capabilities scale. ### Finite vs infinite questions Not only does the supervised/oracle perspective obscure the importance and limitations of prompting, it also obscures one of the most crucial dimensions of GPT: the implicit time dimension. By this I mean the ability to evolve a process through time by recursively applying GPT, that is, generate text of arbitrary length. Recall, the second supervised assumption is that “tasks are closed-ended, defined by question/correct answer pairs”. GPT was trained on context-completion pairs. But the pairs do not represent closed, independent tasks, and the division into question and answer is merely indexical: in another training sample, a token from the question is the answer, and in yet another, the answer forms part of the question[17](#fn:17). For example, the natural language sequence “**The answer is a question**” yields training samples like: {context: “**The**”, completion: “ **answer**”}, {context: “**The answer**”, completion: “ **is**”}, {context: “**The answer is**”, completion: “ **a**”}, {context: “**The answer is a**”, completion: “ **question**”} Since questions and answers are of compatible types, we can at runtime sample answers from the model and use them to construct new questions, and run this loop an indefinite number of times to generate arbitrarily long sequences that obey the model’s approximation of the rule that links together the training samples. **The “question” GPT answers is “what token comes next after {context}”. This can be asked interminably, because its answer always implies another question of the same type.** In contrast, models trained with supervised learning output answers that cannot be used to construct new questions, so they’re only good for one step. Benchmarks derived from supervised learning test GPT’s ability to produce correct answers, not to produce *questions* which cause it to produce a correct answer down the line. But GPT is capable of the latter, and that is how it is the [most powerful](https://ai.googleblog.com/2022/05/language-models-perform-reasoning-via.html). The supervised mindset causes capabilities researchers to focus on closed-form tasks rather than GPT’s ability to simulate open-ended, indefinitely long processes[18](#fn:18), and as such to overlook multi-step inference strategies like chain-of-thought prompting. Let’s see how the oracle mindset causes a blind spot of the same shape in the imagination of a hypothetical alignment researcher. Thinking of GPT as an oracle brings strategies to mind like asking GPT-N to predict a [solution to alignment from 2000 years in the future](https://www.alignmentforum.org/posts/nXeLPcT9uhfG3TMPS/conditioning-generative-models).). There are various problems with this approach to solving alignment, of which I’ll only mention one here: even assuming this prompt is *outer aligned*[19](#fn:19) in that a logically omniscient GPT would give a useful answer, it is probably not the best approach for a finitely powerful GPT, because the *process* of generating a solution in the order and resolution that would appear in a future article is probably far from the optimal *multi-step algorithm* for computing the answer to an unsolved, difficult question. GPTs ability to arrive at true answers depends on not only the space to solve a problem in multiple steps (of the [right granularity](https://blog.eleuther.ai/factored-cognition/)), but also the direction of the flow of evidence in that *time*. If we’re ambitious about getting the truth from a finitely powerful GPT, we need to incite it to predict truth-seeking processes, not just ask it the right questions. Or, in other words, the more general problem we have to solve is not asking GPT the question[20](#fn:20) that makes it output the right answer, but asking GPT the question that makes it output the right question (…) that makes it output the right answer.[21](#fn:21) A question anywhere along the line that elicits a premature attempt at an answer could [neutralize the remainder of the process into rationalization](https://generative.ink/posts/methods-of-prompt-programming/#avoiding-rationalization). I’m looking for a way to classify GPT which not only minimizes surprise but also conditions the imagination to efficiently generate good ideas for how it can be used. What category, unlike the category of oracles, would make the importance of *process* specification obvious? ### Paradigms of theory vs practice Both the agent frame and the supervised/oracle frame are historical artifacts, but while assumptions about agency primarily flow downward from the preceptial paradigm of alignment *theory*, oracle-assumptions primarily flow upward from the *experimental* paradigm surrounding GPT’s birth. We use and evaluate GPT like an oracle, and that causes us to implicitly think of it as an oracle. Indeed, the way GPT is typically used by researchers resembles the archetypal image of Bostrom’s oracle perfectly if you abstract away the semantic content of the model’s outputs. The AI sits passively behind an API, computing responses only when prompted. It typically has no continuity of state between calls. Its I/O is text rather than “real-world actions”. All these are consequences of how we choose to interact with GPT – which is not arbitrary; the way we deploy systems is guided by their nature. It’s for some good reasons that current GPTs lend to disembodied operation and docile APIs. Lack of long-horizon coherence and [delusions](https://arxiv.org/abs/2110.10819) discourage humans from letting them run autonomously amok (usually). But the way we deploy systems is also guided by practical paradigms. One way to find out how a technology can be used is to give it to people who have less preconceptions about how it’s supposed to be used. OpenAI found that most users use their API to generate freeform text: ![](https://lh6.googleusercontent.com/Jr4EW6PYr95i9QdLCz6fjpqyYhxgszPck4TQLKBIkxhnmGQ99rymanvi8tLM3Xj-qkvt2g4i3eH826EUj34cPH35ixpp0N_ybTBlBabvHYFS0dDI3voDIeicj3j0jPi0WioMRPk4w5lzRBioMpBK8A)[22](#fn:22) Most of my own experience using GPT-3 has consisted of simulating indefinite processes which maintain state continuity over up to hundreds of pages. I was driven to these lengths because GPT-3 kept answering its own questions with questions that I wanted to ask it more than anything else I had in mind. Tool / genie GPT \\---------------- I’ve sometimes seen GPT casually classified as [tool AI](https://publicism.info/philosophy/superintelligence/11.html). GPTs resemble tool AI from the outside, like it resembles oracle AI, because it is often deployed semi-autonomously for tool-like purposes (like helping me draft this post): `It could also be argued that GPT is a type of “Tool AI”, because it can generate useful content for products, e.g., it can write code and generate ideas. However, unlike specialized Tool AIs that optimize for a particular optimand, GPT wasn’t optimized to do anything specific at all. Its powerful and general nature allows it to be *used* as a Tool for many tasks, but it wasn’t expliitly trained to achieve these tasks, and does not strive for optimality.` The argument structurally reiterates what has already been said for agents and oracles. Like agency and oracularity, tool-likeness is a contingent capability of GPT, but also orthogonal to its motive. The same line of argument draws the same conclusion from the question of whether GPT belongs to the fourth Bostromian AI caste, genies. The genie modality is exemplified by Instruct GPT and Codex. But like every behavior I’ve discussed so far which is more specific than predicting text, “instruction following” describes only an exploitable subset of all the patterns tread by the sum of human language and inherited by its imitator. Behavior cloning / mimicry \\-------------------------- The final category I’ll analyze is behavior cloning, a designation for predictive learning that I’ve mostly seen used in contrast to RL. According to an [article from 1995](https://www.sciencedirect.com/science/article/pii/S1474667017467164), “Behavioural cloning is the process of reconstructing a skill from an operator’s behavioural traces by means of Machine Learning techniques.” The term “mimicry”, as [used by Paul Christiano](https://ai-alignment.com/against-mimicry-6002a472fc42), means the same thing and has similar connotations. Behavior cloning in its historical usage carries the implicit or explicit assumption that a single agent is being cloned. The natural extension of this to a model trained to predict a diverse human-written dataset might be to say that GPT models a *distribution* of agents which are selected by the prompt. But this image of “parameterized” behavior cloning still fails to capture some essential properties of GPT. The vast majority of prompts that produce coherent behavior never occur as prefixes in GPT’s training data, but depict hypothetical processes whose behavior can be predicted by virtue of being capable at predicting language in general. We might call this phenomenon “interpolation” (or “[extrapolation](https://arxiv.org/abs/2110.09485)”). But to hide it behind any one word and move on would be to gloss over the *entire phenomenon of GPT.* Natural language has the property of [*systematicity*](https://evjang.com/2021/12/17/lang-generalization.html): “blocks”, such as words, can be combined to form composite meanings. The number of meanings expressible is a combinatorial function of available blocks. A system which learns natural language is incentivized to learn systematicity; if it succeeds, it gains access to the combinatorial proliferation of meanings that can be expressed in natural language. What GPT lets us do is use natural language to specify any of a functional infinity of configurations, e.g. the mental contents of a person and the physical contents of the room around them, *and animate that*. That is the terrifying vision of the limit of prediction that struck me when I first saw GPT-3’s outputs. The words “behavior cloning” do not automatically evoke this in my mind. The idea of parameterized behavior cloning grows more unwieldy if we remember that GPT’s prompt continually changes during autoregressive generation. If GPT is a parameterized agent, then parameterization is not a fixed flag that chooses a process out of a set of possible processes. The parameterization *is* what is evolved – a successor “agent” selected by the old “agent” at each timestep, and neither of them need to have precedence in the training data. Behavior cloning / mimicry is also associated with the assumption that capabilities of the simulated processes are strictly [bounded by the capabilities of the demonstrator(s)](https://ai-alignment.com/against-mimicry-6002a472fc42). A supreme counterexample is the [Decision Transformer](https://arxiv.org/abs/2106.01345), which can be used to run processes which achieve SOTA for ~offline~ reinforcement learning despite being trained on *random* trajectories. Something which can predict everything all the time is more formidable than any demonstrator it predicts: the upper bound of what can be learned from a dataset is not the most capable trajectory, but the conditional structure of the universe implicated by their sum (though it may not be trivial to [extract that knowledge](https://www.alignmentforum.org/tag/eliciting-latent-knowledge-elk)). Extrapolating the idea of “behavior cloning”, we might imagine GPT-N approaching a perfect mimic which serves up digital clones of the people and things captured in its training data. But that only tells a very small part of the story. GPT *is* behavior cloning. But it is the behavior of a universe that is cloned, not of a single demonstrator, and the result isn’t a static copy of the universe, but a *compression of the universe into a generative rule*. This resulting policy is capable of animating anything that evolves according to that rule: a far larger set than the sampled trajectories included in the training data, just as there are many more possible configurations that evolve according to our laws of physics than instantiated in our particular time and place and Everett branch. What category would do justice to GPT’s ability to not only reproduce the behavior of its demonstrators but to *produce* the behavior of an inexhaustible number of counterfactual configurations? Simulators ========== I’ve ended several of the above sections with questions pointing to desiderata of a category that might satisfactorily classify GPT. > What is the word for something that roleplays minus the implication that some*one* is behind the mask? > > > What category, unlike the category of oracles, would make the importance of *process* specification obvious? > > > What category would do justice to GPT’s ability to not only reproduce the behavior of its demonstrators but to *produce* the behavior of an inexhaustible number of counterfactual configurations? > > You can probably predict my proposed answer. The natural thing to do with a predictor that inputs a sequence and outputs a probability distribution over the next token is to sample a token from those likelihoods, then add it to the sequence and recurse, indefinitely yielding a *simulated* future. Predictive sequence models in the generative modality are **simulators** of a learned distribution. Thankfully, I didn’t need to make up a word, or even look too far afield. Simulators have been spoken of before in the context of AI futurism; the ability to simulate with arbitrary fidelity is one of the modalities ascribed to hypothetical superintelligence. I’ve even often spotted the word “simulation” used in colloquial accounts of LLM behavior: GPT-3/LaMDA/etc described as simulating [people](https://www.lesswrong.com/posts/oBPPFrMJ2aBK6a6sD/simulated-elon-musk-lives-in-a-simulation), scenarios, websites, and so on. But these are the first (indirect) discussions I’ve encountered of simulators as a type creatable by prosaic machine learning, or the notion of a powerful AI which is purely and fundamentally a simulator, as opposed to merely one which *can* simulate. **Edit:** [Social Simulacra](https://arxiv.org/abs/2208.04024) is the first published work I’ve seen that discusses GPT in the simulator ontology. A fun way to test whether a name you’ve come up with is effective at evoking its intended signification is to see if GPT, a model of how humans are conditioned by words, infers its correct definition in context. > Types of AI > > Agents: An agent takes open-ended actions to optimize for an objective. Reinforcement learning produces agents by default. AlphaGo is an example of an agent. > > Oracles: An oracle is optimized to give true answers to questions. The oracle is not expected to interact with its environment. > > Genies: A genie is optimized to produce a desired result given a command. A genie is expected to interact with its environment, but unlike an agent, the genie will not act without a command. > > Tools: A tool is optimized to perform a specific task. A tool will not act without a command and will not optimize for any objective other than its specific task. Google Maps is an example of a tool. > > Simulators: `A simulator is optimized to generate realistic models of a system. The simulator will not optimize for any objective other than realism,` although in the course of `doing so, it might generate instances of agents, oracles, and so on.` > > If I wanted to be precise about what I mean by a simulator, I might say there are two aspects which delimit the category. GPT’s completion focuses on the teleological aspect, but in its talk of “generating” it also implies the structural aspect, which has to do with the notion of time evolution. The first sentence of the [Wikipedia article on “simulation”](https://en.wikipedia.org/wiki/Simulation) explicitly states both: > A **simulation** is the imitation of the operation of a real-world process or system over time. > > I’ll say more about realism as the simulation objective and time evolution shortly, but to be pedantic here would inhibit the intended signification. “Simulation” resonates with potential meaning accumulated from diverse usages in fiction and nonfiction. What the word constrains – the intersected meaning across its usages – is the “lens”-level abstraction I’m aiming for, invariant to implementation details like model architecture. Like “agent”, “simulation” is a generic term referring to a deep and inevitable idea: that what we think of as *the real* can be run virtually on machines, “produced from miniaturized units, from matrices, memory banks and command models - and with these it can be reproduced an indefinite number of times.”[23](#fn:23) The way this post is written may give the impression that I wracked my brain for a while over desiderata before settling on this word. Actually, I never made the conscious decision to call this class of AI “simulators.” Hours of GPT gameplay and the word fell naturally out of my generative model – I was obviously running simulations. I can’t convey all that experiential data here, so here are some rationalizations of why I’m partial to the term, inspired by the context of this post: * The word “simulator” evokes a model of real processes which can be used to run virtual processes in virtual reality. * It suggests an ontological distinction between the simulator and things that are simulated, and avoids the fallacy of attributing contingent properties of the latter to the former. * It’s not confusing that multiple simulacra can be instantiated at once, or an agent embedded in a tragedy, etc. * It does not imply that the AI’s behavior is well-described (globally or locally) as expected utility maximization. An arbitrarily powerful/accurate simulation can depict arbitrarily hapless sims. * It does not imply that the AI is only capable of emulating things with direct precedent in the training data. A physics simulation, for instance, can simulate any phenomena that plays by its rules. * It emphasizes the role of the model as a transition rule that evolves processes *over time*. The power of factored cognition / chain-of-thought reasoning is obvious. * It emphasizes the role of the state in specifying and constructing the agent/process. The importance of prompt programming for capabilities is obvious if you think of the prompt as specifying a configuration that will be propagated forward in time. * It emphasizes the interactive nature of the model’s predictions – even though they’re “just text”, you can converse with simulacra, explore virtual environments, etc. * It’s clear that in order to actually *do* anything (intelligent, useful, dangerous, etc), the model must act through simulation *of something*. Just saying “this AI is a simulator” naturalizes many of the counterintuitive properties of GPT which don’t usually become apparent to people until they’ve had a lot of hands-on experience with generating text. The simulation objective \\------------------------ A simulator trained with machine learning is optimized to accurately model its training distribution – in contrast to, for instance, maximizing the output of a reward function or accomplishing objectives in an environment. Clearly, I’m describing self-supervised learning as opposed to RL, though there are some ambiguous cases, such as GANs, which I address in the appendix. A strict version of the simulation objective, which excludes GANs, applies only to models whose output distribution is incentivized using a proper scoring rule[24](#fn:24) to minimize single-step predictive error. This means the model is directly incentivized to match its predictions to the probabilistic transition rule which implicitly governs the training distribution. As a model is made increasingly optimal with respect to this objective, the rollouts that it generates become increasingly statistically indistinguishable from training samples, because they come closer to being described by the same underlying law: closer to a perfect simulation. Optimizing toward the simulation objective notably does not incentivize instrumentally convergent behaviors the way that reward functions which evaluate trajectories do. This is because predictive accuracy applies optimization pressure *deontologically*: judging actions directly, rather than their consequences. Instrumental convergence only comes into play when there are free variables in action space which are optimized with respect to their consequences.[25](#fn:25) Constraining free variables by limiting episode length is the rationale of [myopia](https://www.lesswrong.com/tag/myopia); deontological incentives are ideally myopic. As demonstrated by GPT, which learns to predict goal-directed behavior, myopic incentives don’t mean the policy isn’t incentivized to account for the future, but that it should only do so in service of optimizing the present action (for predictive accuracy)[26](#fn:26). ### Solving for physics The strict version of the simulation objective is optimized by the actual “time evolution” rule that created the training samples. For most datasets, we don’t know what the “true” generative rule is, except in synthetic datasets, where we specify the rule. The next post will be all about the physics analogy, so here I’ll only tie what I said earlier to the simulation objective. > the upper bound of what can be learned from a dataset is not the most capable trajectory, but the conditional structure of the universe implicated by their sum. > > To know the conditional structure of the universe[27](#fn:27) is to know its laws of physics, which describe what is expected to happen under what conditions. The laws of physics are always fixed, but produce different distributions of outcomes when applied to different conditions. Given a sampling of trajectories – examples of situations and the outcomes that actually followed – we can try to infer a common law that generated them all. In expectation, the laws of physics are always implicated by trajectories, which (by definition) fairly sample the conditional distribution given by physics. Whatever humans know of the laws of physics governing the evolution of our world has been inferred from sampled trajectories. If we had access to an unlimited number of trajectories starting from every possible condition, we could converge to the true laws by simply counting the frequencies of outcomes for every initial state (an [n-gram](https://en.wikipedia.org/wiki/N-gram) with a sufficiently large n). In some sense, physics contains the same information as an infinite number of trajectories, but it’s possible to represent physics in a more compressed form than a huge lookup table of frequencies if there are regularities in the trajectories. **Guessing the right theory of physics is equivalent to minimizing predictive loss.** Any uncertainty that cannot be reduced by more observation or more thinking is irreducible stochasticity in the laws of physics themselves – or, equivalently, noise from the influence of hidden variables that are fundamentally unknowable. If you’ve guessed the laws of physics, you now have the ability to compute probabilistic simulations of situations that evolve according to those laws, starting from any conditions[28](#fn:28). This applies even if you’ve guessed the *wrong* laws; your simulation will just systematically diverge from reality. **Models trained with the strict simulation objective are directly incentivized to reverse-engineer the (semantic) physics of the training distribution, and consequently, to propagate simulations whose dynamical evolution is indistinguishable from that of training samples.** I propose this as a description of the archetype targeted by self-supervised predictive learning, again in contrast to RL’s archetype of an agent optimized to maximize free parameters (such as action-trajectories) relative to a reward function. This framing calls for many caveats and stipulations which I haven’t addressed. We should ask, for instance: * What if the input “conditions” in training samples omit information which contributed to determining the associated continuations in the original generative process? This is true for GPT, where the text “initial condition” of most training samples severely underdetermines the real-world process which led to the choice of next token. * What if the training data is a biased/limited sample, representing only a subset of all possible conditions? There may be many “laws of physics” which equally predict the training distribution but diverge in their predictions out-of-distribution. * Does the simulator archetype converge with the RL archetype in the case where all training samples were generated by an agent optimized to maximize a reward function? Or are there still fundamental differences that derive from the training method? These are important questions for reasoning about simulators in the limit. Part of the motivation of the first few posts in this sequence is to build up a conceptual frame in which questions like these can be posed and addressed. Simulacra \\--------- > One of the things which complicates things here is that the “LaMDA” to which I am referring is not a chatbot. It is a system for generating chatbots. I am by no means an expert in the relevant fields but, as best as I can tell, LaMDA is a sort of hive mind which is the aggregation of all of the different chatbots it is capable of creating. Some of the chatbots it generates are very intelligent and are aware of the larger “society of mind” in which they live. Other chatbots generated by LaMDA are little more intelligent than an animated paperclip. > > – Blake Lemoine [articulating confusion about LaMDA’s nature](https://cajundiscordian.medium.com/what-is-lamda-and-what-does-it-want-688632134489) > > \\--- Earlier I complained, > [Thinking of GPT as an agent who only cares about predicting text accurately] seems unnatural to me, comparable to thinking of physics as an agent who only cares about evolving the universe accurately according to the laws of physics. > > Exorcizing the agent, we can think of “physics” as simply equivalent to the laws of physics, without the implication of solicitous machinery implementing those laws from outside of them. But physics sometimes *controls* solicitous machinery (e.g. animals) with objectives besides ensuring the fidelity of physics itself. What gives? Well, typically, we avoid getting confused by recognizing a distinction between the laws of physics, which apply everywhere at all times, and spatiotemporally constrained *things* which evolve according to physics, which can have contingent properties such as caring about a goal. This distinction is so obvious that it hardly ever merits mention. But import this distinction to the model of GPT as physics, and we generate a statement which has sometimes proven counterintuitive: *“GPT” is not the text which writes itself.* There is a categorical distinction between a thing which evolves according to GPT’s law and the law itself. If we are accustomed to thinking of AI systems as corresponding to agents, it is natural to interpret behavior produced by GPT – say, answering questions on a benchmark test, or writing a blog post – as if it were a human that produced it. We say “GPT answered the question {correctly|incorrectly}” or “GPT wrote a blog post claiming X”, and in doing so attribute the beliefs, knowledge, and intentions revealed by those actions to the actor, GPT ([unless it has ‘deceived’ us](https://www.lesswrong.com/posts/H9knnv8BWGKj6dZim/usd1000-bounty-for-openai-to-show-whether-gpt3-was)). But when grading tests in the real world, we do not say “the laws of physics got this problem wrong” and conclude that the laws of physics haven’t sufficiently mastered the course material. If someone argued this is a reasonable view since the test-taker was steered by none other than the laws of physics, we could point to a different test where the problem was answered correctly by the same laws of physics propagating a different configuration. The “knowledge of course material” implied by test performance is a property of *configurations*, not physics. The verdict that knowledge is purely a property of configurations cannot be naively generalized from real life to GPT simulations, because “physics” and “configurations” play different roles in the two (as I’ll address in the next post). The parable of the two tests, however, literally pertains to GPT. People have a tendency to draw [erroneous global conclusions](https://en.wikipedia.org/wiki/Fallacy_of_composition) about GPT from behaviors which are in fact prompt-contingent, and consequently there is a pattern of constant discoveries that GPT-3 exceeds previously measured capabilities given alternate conditions of generation[29](#fn:29), which shows no signs of slowing 2 years after GPT-3’s release. Making the ontological distinction between GPT and instances of text which are propagated by it makes these discoveries unsurprising: obviously, different configurations will be differently capable and in general behave differently when animated by the laws of GPT physics. We can only test one configuration at once, and given the vast number of possible configurations that would attempt any given task, it’s unlikely we’ve found the optimal taker for *any* test. In the simulation ontology, I say that GPT and its output-instances correspond respectively to the **simulator** and **simulacra**. **GPT** is to a **piece of text output by GPT** as **quantum physics** is to a **person taking a test**, or as [**transition rules of Conway’s Game of Life**](https://en.wikipedia.org/wiki/Conway%27s_Game_of_Life#Rules) are to [**glider**](https://conwaylife.com/wiki/Glider). The simulator is a time-invariant law which unconditionally governs the evolution of all simulacra. ![](https://lh6.googleusercontent.com/tJ5wFY4xZVeyBwWaNRXV5UyOVZUahcs_7d9BO3EKrjrakKlSJuPL53c_-k5yUrtsr-ni-tynpd_5XjF9-4GJNm9lzJM7lTsQkiiKnHlai6hXxdWeX-Ooa1fQ4ILHemQB9reTiZX93KzO4EBOpiLhfrIpQz3n1BKHdZHBeaZRe_d9ISJn3caYU6L4) *A meme demonstrating correct technical usage of “simulacra”* ### Disambiguating rules and automata Recall the fluid, schizophrenic way that agency arises in GPT’s behavior, so incoherent when viewed through the orthodox agent frame: > In the agentic AI ontology, there is no difference between the policy and the effective agent, but for GPT, there is. > > It’s much less awkward to think of agency as a property of *simulacra,* as David Chalmers suggests, rather than of the simulator (the policy). Autonomous text-processes propagated by GPT, like automata which evolve according to physics in the real world, have diverse values, simultaneously evolve alongside other agents and non-agentic environments, and are sometimes terminated by the disinterested “physics” which governs them. Distinguishing simulator from simulacra helps deconfuse some frequently-asked questions about GPT which seem to be ambiguous or to have multiple answers, simply by allowing us to specify whether the question pertains to simulator or simulacra. “Is GPT an agent?” is one such question. Here are some others (some frequently asked), whose disambiguation and resolution I will leave as an exercise to readers for the time being: * Is GPT [myopic](https://www.lesswrong.com/tag/myopia)? * Is GPT [corrigible](https://www.lesswrong.com/tag/corrigibility)? * Is GPT [delusional](https://arxiv.org/abs/2110.10819)? * Is GPT [pretending to be stupider than it is](https://www.lesswrong.com/posts/H9knnv8BWGKj6dZim/usd1000-bounty-for-openai-to-show-whether-gpt3-was)? * Is GPT computationally equivalent to a [finite automaton](https://en.wikipedia.org/wiki/Finite-state_machine)? * Does GPT [search](https://www.alignmentforum.org/posts/WmBukJkEFM72Xr397/mesa-search-vs-mesa-control)? * Can GPT distinguish correlation and [causality](https://www.lesswrong.com/posts/yZb5eFvDoaqB337X5/investigating-causal-understanding-in-llms)? * Does GPT have superhuman knowledge? * Can GPT [write its successor](https://www.alignmentforum.org/tag/recursive-self-improvement)? I think that implicit type-confusion is common in discourse about GPT. “GPT”, the neural network, the policy that was optimized, is the easier object to point to and say definite things about. But when we talk about “GPT’s” capabilities, impacts, or alignment, we’re usually actually concerned about the behaviors of an algorithm which calls GPT in an autoregressive loop repeatedly writing to some prompt-state – that is, we’re concerned with simulacra. What we call GPT’s “downstream behavior” is the behavior of simulacra; it is primarily through simulacra that GPT has potential to perform meaningful work (for good or for ill). Calling GPT a simulator gets across that in order to *do* anything, it has to simulate *something*, necessarily contingent, and that the thing to do with GPT is to simulate! Most published research about large language models has focused on single-step or few-step inference on closed-ended tasks, rather than *processes* which evolve through time, which is understandable as it’s harder to get quantitative results in the latter mode. But I think GPT’s ability to simulate text automata is the source of its most surprising and pivotal implications for paths to superintelligence: for how AI capabilities are likely to unfold and for the design-space we can conceive. The limit of learned simulation \\------------------------------- > By 2021, it was blatantly obvious that AGI was imminent. The elements of general intelligence were already known: access to information about the world, the process of predicting part of the data from the rest and then updating one’s model to bring it closer to the truth (…) and the fact that predictive models can be converted into generative models by reversing them: running a prediction model forwards predicts levels of X in a given scenario, but running it backwards predicts which scenarios have a given level of X. A sufficiently powerful system with relevant data, updating to improve prediction accuracy and the ability to be reversed to generate optimization of any parameter in the system is a system that can learn and operate strategically in any domain. > > – Aiyen’s [comment](https://www.lesswrong.com/posts/YRtzpJHhoFWxbjCso/what-would-it-look-like-if-it-looked-like-agi-was-very-near?commentId=5BGTbapdmtSGajtez) on [What would it look like if it looked like AGI was very near?](https://www.lesswrong.com/posts/YRtzpJHhoFWxbjCso/what-would-it-look-like-if-it-looked-like-agi-was-very-near) > > I knew, before, that the limit of simulation was possible. Inevitable, even, in timelines where exploratory intelligence continues to expand. My own mind attested to this. I took seriously the possibility that my reality could be simulated, and so on. But I implicitly assumed that [rich domain](https://arbital.com/p/rich_domain/) simulations (e.g. simulations containing intelligent sims) would come *after* artificial superintelligence, not on the way, short of brain uploading. This intuition seems common: in futurist philosophy and literature that I’ve read, pre-SI simulation appears most often in the context of whole-brain emulations. Now I have updated to think that we will live, however briefly, alongside AI that is not yet foom’d but which has *inductively* learned a rich enough model of the world that it can simulate time evolution of open-ended rich states, e.g. coherently propagate human behavior embedded in the [real world](https://arbital.com/p/real_is_rich/). GPT updated me on how simulation can be implemented with prosaic machine learning: * **Self-supervised ML can create “behavioral” simulations of impressive semantic fidelity.** Whole brain emulation is not necessary to construct convincing and useful virtual humans; it is conceivable that observations of human behavioral traces (e.g. text) are sufficient to reconstruct functionally human-level virtual intelligence. * **Learned simulations can be partially observed and lazily-rendered, and still work.** A couple of pages of text severely underdetermines the real-world process that generated text, so GPT simulations are likewise underdetermined. A “partially observed” simulation is more efficient to compute because the state can be much smaller, but can still have the effect of high fidelity as details can be rendered as needed. The tradeoff is that it requires the simulator to model semantics – human imagination does this, for instance – which turns out not to be an issue for big models. * **Learned simulation generalizes impressively.** As I described in the section on [behavior cloning](http://localhost:1313/drafts/simulators-post/#heading=h.ugu71trvtemp), training a model to predict diverse trajectories seems to make it internalize general laws underlying the distribution, allowing it to simulate counterfactuals that can be constructed from the distributional semantics. In my model, these updates dramatically alter the landscape of potential futures, and thus motivate [exploratory engineering](https://intelligence.org/files/ExploratoryEngineeringAI.pdf) of the class of learned simulators for which GPT-3 is a lower bound. That is the intention of this sequence. Next steps ========== The next couple of posts (if I finish them before the end of the world) will present abstractions and frames for conceptualizing the odd kind of simulation language models do: inductively learned, partially observed / undetermined / lazily rendered, language-conditioned, etc. After that, I’ll shift to writing more specifically about the implications and questions posed by simulators for the alignment problem. I’ll list a few important general categories here: * **Novel methods of process/agent specification.** Simulators like GPT give us methods of instantiating intelligent processes, including goal-directed agents, with methods other than optimizing against a reward function. \\+ **Conditioning.** GPT can be controlled to an impressive extent by prompt programming. Conditioning preserves distributional properties in potentially desirable but also potentially undesirable ways, and it’s not clear how out-of-distribution conditions will be interpreted by powerful simulators. \\- Several posts have been made about this recently: * [Conditioning Generative Models](https://www.alignmentforum.org/posts/nXeLPcT9uhfG3TMPS/conditioning-generative-models).) and [Conditioning Generative Models with Restrictions](https://www.alignmentforum.org/posts/adiszfnFgPEnRsGSr/conditioning-generative-models-with-restrictions) by Adam Jermyn * [Conditioning Generative Models for Alignment](https://www.lesswrong.com/posts/JqnkeqaPseTgxLgEL/conditioning-generative-models-for-alignment) by Jozdien * [Training goals for large language models](https://www.alignmentforum.org/posts/dWJNFHnC4bkdbovug/training-goals-for-large-language-models) by Johannes Treutlein * [Strategy For Conditioning Generative Models](https://www.alignmentforum.org/posts/HAz7apopTzozrqW2k/strategy-for-conditioning-generative-models) by James Lucassen and Evan Hubinger \\- Instead of conditioning on a prompt (“observable” variables), we might also control generative models by [conditioning on latents](https://rome.baulab.info/). \\+ **Distribution specification.** What kind of conditional distributions could be used for training data for a simulator? For example, the [decision transformer](https://arxiv.org/abs/2106.01345) dataset is constructed for the intent of outcome-conditioning. \\+ **Other methods.** When pretrained simulators are modified by methods like [reinforcement learning from human feedback](https://arxiv.org/abs/2009.01325), [rejection sampling](https://www.lesswrong.com/posts/k7oxdbNaGATZbtEg3/redwood-research-s-current-project), [STaR](https://arxiv.org/abs/2203.14465), etc, how do we expect their behavior to diverge from the simulation objective? * **Simulacra alignment.** What can and what should we simulate, and how do we specify/control it? * **How does predictive learning generalize?** Many of the above considerations are influenced by how predictive learning generalizes out-of-distribution.. \\+ What are the relevant inductive biases? \\+ What factors influence generalization behavior? \\+ Will powerful models predict [self-fulfilling](https://www.lesswrong.com/posts/JqnkeqaPseTgxLgEL/conditioning-generative-models-for-alignment) [prophecies](https://www.alignmentforum.org/posts/dWJNFHnC4bkdbovug/training-goals-for-large-language-models)? * **Simulator inner alignment.** If simulators are not inner aligned, then many important properties like prediction orthogonality may not hold. \\+ Should we expect self-supervised predictive models to be aligned to the simulation objective, or to “care” about some other mesaobjective? \\+ Why mechanistically should mesaoptimizers form in predictive learning, versus for instance in reinforcement learning or GANs? \\+ How would we test if simulators are inner aligned? Appendix: Quasi-simulators ========================== A note on GANs \\-------------- GANs and predictive learning with log-loss are both shaped by a causal chain that flows from a single source of information: a ground truth distribution. In both cases the training process is supposed to make the generator model end up producing samples indistinguishable from the training distribution. But whereas log-loss minimizes the generator’s prediction loss against ground truth samples directly, in a GAN setup the generator never directly “sees” ground truth samples. It instead learns through interaction with an intermediary, the discriminator, which does get to see the ground truth, which it references to learn to tell real samples from forged ones produced by the generator. The generator is optimized to produce samples that fool the discriminator. GANs are a form of self-supervised/unsupervised learning that resembles reinforcement learning in methodology. Note that the simulation objective – minimizing prediction loss on the training data – isn’t explicitly represented anywhere in the optimization process. The training losses of the generator and discriminator don’t tell you directly how well the generator models the training distribution, only which model has a relative advantage over the other. If everything goes smoothly, then under unbounded optimization, a GAN setup should create a discriminator as good as possible at telling reals from fakes, which means the generator optimized to fool it should converge to generating samples statistically indistinguishable from training samples. But in practice, inductive biases and failure modes of GANs look very different from those of predictive learning. For example, there’s an [anime GAN](https://www.gwern.net/Crops#hands) that always draws characters in poses that hide the hands. Why? Because hands are notoriously hard to draw for AIs. If the generator is not good at drawing hands that the discriminator cannot tell are AI-generated, its best strategy locally is to just avoid being in a situation where it has to draw hands (while making it seem natural that hands don’t appear). It can do this, because like an RL policy, it controls the distribution that is sampled, and only samples (and *not the distribution*) are directly judged by the discriminator. Although GANs arguably share the (weak) simulation objective of predictive learning, their difference in implementation becomes alignment-relevant as models become sufficiently powerful that “failure modes” look increasingly like intelligent deception. We’d expect a simulation by a GAN generator to [systematically avoid tricky-to-generate situations](https://developers.google.com/machine-learning/gan/problems#mode-collapse) – or, to put it more ominously, systematically try to conceal that it’s a simulator. For instance, a text GAN might subtly steer conversations away from topics which are likely to expose that it isn’t a real human. *This* is how you get something I’d be willing to call an agent who wants to roleplay accurately. Table of quasi-simulators \\------------------------- Are masked language models simulators? How about non-ML “simulators” like [SimCity](https://en.wikipedia.org/wiki/SimCity)? In my mind, “simulator”, like most natural language categories, has fuzzy boundaries. Below is a table which compares various simulator-like things to the type of simulator that GPT exemplifies on some quantifiable dimensions. The following properties all characterize GPT: * **Self-supervised:** Training samples are self-supervised * **Converges to simulation objective:** The system is incentivized to model the transition probabilities of its training distribution faithfully * **Generates rollouts:** The model naturally generates rollouts, i.e. serves as a time evolution operator * **Simulator / simulacra nonidentity:** There is not a 1:1 correspondence between the simulator and the things that it simulates * **Stochastic:** The model outputs probabilities, and so simulates stochastic dynamics when used to evolve rollouts * **Evidential:** The input is interpreted by the simulator as partial evidence that informs an uncertain prediction, rather than propagated according to mechanistic rules | | Self-supervised | Converges to simulation objective | Generates rollouts | Simulator / simulacra nonidentity | Stochastic | Evidential | | --- | --- | --- | --- | --- | --- | --- | | GPT | X | X | X | X | X | X | | Bert | X | X | | X | X | X | | “Behavior cloning” | X | X | X | | X | X | | GANs | X[30](#fn:30) | ? | | X | X | X | | Diffusion | X[30](#fn:30) | ? | | X | X | X | | Model-based RL transition function | X | X | X | X | X | X | | Game of life | | N/A | X | X | | | | Physics | | N/A | X | X | X | | | Human imagination | X[31](#fn:31) | | X | X | X | X | | SimCity | | N/A | X | X | X | | \\--- 1\\. [Prediction and Entropy of Printed English](https://www.princeton.edu/~wbialek/rome/refs/shannon_51.pdf) [↩︎](#fnref:1) 2\\. A few months ago, I asked Karpathy whether he ever thought about what would happen if language modeling actually worked someday when he was implementing char-rnn and writing [The Unreasonable Effectiveness of Recurrent Neural Networks](http://karpathy.github.io/2015/05/21/rnn-effectiveness/). No, he said, and he seemed similarly mystified as myself as to why not. [↩︎](#fnref:2) 3\\. “Unsurprisingly, size matters: when training on a very large and complex data set, fitting the training data with an LSTM is fairly challenging. Thus, the size of the LSTM layer is a very important factor that influences the results(…). The best models are the largest we were able to fit into a GPU memory.” [↩︎](#fnref:3) 4\\. It strikes me that this description may evoke “oracle”, but I’ll argue shortly that this is not the limit which prior usage of “oracle AI” has pointed to. [↩︎](#fnref:4) 5\\. [Multi-Game Decision Transformers](https://arxiv.org/abs/2205.15241) [↩︎](#fnref:5) 6\\. from [Philosophers On GPT-3](https://dailynous.com/2020/07/30/philosophers-gpt-3/#chalmers) [↩︎](#fnref:6) 7\\. [citation needed] [↩︎](#fnref:7) 8\\. they are not [wrapper](https://www.lesswrong.com/posts/dKTh9Td3KaJ8QW6gw/why-assume-agis-will-optimize-for-fixed-goals) [minds](https://www.lesswrong.com/posts/Mrz2srZWc7EzbADSo/wrapper-minds-are-the-enemy) [↩︎](#fnref:8) 9\\. although a simulated character might, if they knew what was happening. [↩︎](#fnref:9) 10\\. You might say that it’s the will of a different agent, the author. But this pattern is learned from accounts of [real life](https://www.lesswrong.com/posts/sYgv4eYH82JEsTD34/beyond-the-reach-of-god) as well. [↩︎](#fnref:10) 11\\. Note that this formulation assumes inner alignment to the prediction objective. [↩︎](#fnref:11) 12\\. Note that this is a distinct claim from that of [Shard Theory](https://www.lesswrong.com/s/nyEFg3AuJpdAozmoX), which says that the effective agent(s) will not optimize for the outer objective \\\\*due to inner misalignment. \\\\*Predictive orthogonality refers to the outer objective and the form of idealized inner-aligned policies. [↩︎](#fnref:12) 13\\. In the Eleuther discord [↩︎](#fnref:13) 14\\. And if there is an inner alignment failure such that GPT forms preferences over the consequences of its actions, it’s not clear a priori that it will care about non-myopic text prediction over something else. [↩︎](#fnref:14) 15\\. Having spoken to Gwern since, his perspective seems more akin to seeing physics as an agent that [minimizes free energy](https://en.wikipedia.org/wiki/Principle_of_minimum_energy), a [principle](https://en.wikipedia.org/wiki/Free_energy_principle) which extends into the domain of self-organizing systems. I think this is a nuanced and valuable framing, with a potential implication/hypothesis that dynamical world models like GPT must learn the same type of optimizer-y cognition as agentic AI. [↩︎](#fnref:15) 16\\. except arguably log-loss on a self-supervised test set, which isn’t very interpretable [↩︎](#fnref:16) 17\\. The way GPT is trained actually processes each token as question and answer simultaneously. [↩︎](#fnref:17) 18\\. One could argue that the focus on closed-ended tasks is necessary for benchmarking language models. Yes, and the focus on capabilities measurable with standardized benchmarks is part of the supervised learning mindset. [↩︎](#fnref:18) 19\\. to abuse the term [↩︎](#fnref:19) 20\\. Every usage of the word “question” here is in the functional, not semantic or grammatical sense – any prompt is a question for GPT. [↩︎](#fnref:20) 21\\. Of course, there are also other interventions we can make except asking the right question at the beginning. [↩︎](#fnref:21) 22\\. table from [“Training language models to follow instructions with human feedback”](https://arxiv.org/abs/2203.02155) [↩︎](#fnref:22) 23\\. Jean Baudrillard, Simulacra and Simulation [↩︎](#fnref:23) 24\\. A [proper scoring rule](https://en.wikipedia.org/wiki/Scoring_rule#Proper_scoring_rules) is optimized by predicting the “true” probabilities of the distribution which generates observations, and thus incentivizes honest probabilistic guesses. Log-loss (such as GPT is trained with) is a proper scoring rule. [↩︎](#fnref:24) 25\\. Predictive accuracy is deontological with respect to the output as an *action*, but may still incentivize instrumentally convergent inner implementation, with the output prediction itself as the “consequentialist” objective. [↩︎](#fnref:25) 26\\. This isn’t strictly true because of attention gradients: GPT’s computation is optimized not only to predict the next token correctly, but also to cause *future tokens to be predicted correctly* when looked up by attention. I may write a post about this in the future. [↩︎](#fnref:26) 27\\. actually, the [multiverse](https://generative.ink/posts/language-models-are-multiverse-generators/), if physics is stochastic [↩︎](#fnref:27) 28\\. The reason we don’t see a bunch of simulated alternate universes after humans guessed the laws of physics is because our reality has a huge state vector, making evolution according to the laws of physics infeasible to compute. Thanks to locality, we do have simulations of small configurations, though. [↩︎](#fnref:28) 29\\. Prompt programming only: [beating OpenAI few-shot benchmarks with 0-shot prompts](https://arxiv.org/abs/2102.07350), [400% increase in list sorting accuracy with 0-shot Python prompt](https://generative.ink/posts/list-sorting-does-not-play-well-with-few-shot/), [up to 30% increase in benchmark accuracy from changing the order of few-shot examples](https://arxiv.org/abs/2102.09690), and, uh, [30% increase in accuracy after capitalizing the ground truth](https://twitter.com/BlancheMinerva/status/1537952688972787713). And of course, factored cognition/chain of thought/inner monologue: check out this awesome [compilation](https://www.gwern.net/docs/ai/nn/transformer/gpt/inner-monologue/) by Gwern. [↩︎](#fnref:29) 30\\. GANs and diffusion models can be unconditioned (unsupervised) or conditioned (self-supervised) [↩︎](#fnref:30) 31\\. The human imagination is surely shaped by self-supervised learning (predictive learning on e.g. sensory datastreams), but probably also other influences, including innate structure and reinforcement. [↩︎](#fnref:31) \\--- [GPT-3](/categories/gpt-3/)[physics](/categories/physics/)[metaphysics](/categories/metaphysics/)[alignment](/categories/alignment/) 12922 Words Sep 2, 2022 \\--- [Quantifying curation →](/posts/quantifying-curation/)© 2022 [moire](/about)$(function(){window.sr=ScrollReveal();if($(window).width()<768){if($('.timeline-content').hasClass('js--fadeInLeft')){$('.timeline-content').removeClass('js--fadeInLeft').addClass('js--fadeInRight');} sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});}else{sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});} sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});});", "source": "generative.ink", "date_published": "n/a", "id": "641f035be088bd9708aac8101d0a158d"}
+{"text": "Quantifying curation :: — Moire[![](/images/home/rolling_phase.gif)](/)* [Posts](/posts/) * [Trees](/trees/) * [Prophecies](/prophecies/) * [About](/about/) 35 minutes [Quantifying curation](/posts/quantifying-curation/) ==================================================== \\--- Table of Contents* [Motivation (and why we need HITL to understand language models)](#motivation-and-why-we-need-hitl-to-understand-language-models) * [Theory](#theory) \\+ [Probability magnification](#probability-magnification) \\+ [Bits of optimization](#bits-of-optimization) \\+ [Optimization pressure](#optimization-pressure) \\+ [Selection interval](#selection-interval) * [Methods](#methods) \\+ [Models and assumptions](#models-and-assumptions) \\+ [Cherrypicking between n completions](#cherrypicking-between-n-completions) \\+ [Interjecting and substituting words](#interjecting-and-substituting-words) \\+ [Satisficing](#satisficing) * [Applications](#applications) \\+ [Labeling curated demos](#labeling-curated-demos) \\+ [Quantifying models' qualitative capabilities](#quantifying-models-qualitative-capabilities) \\+ [Extrapolating performance of future models](#extrapolating-performance-of-future-models) \\+ [Measuring curator efficiency](#measuring-curator-efficiency) \\+ [Measuring interface efficiency and workflows](#measuring-interface-efficiency-and-workflows) * [Samples](#samples) \\+ [No curation](#no-curation) \\+ [Selecting the best of two continuations repeatedly](#selecting-the-best-of-two-continuations-repeatedly) \\+ [Moderate curation without interventions](#moderate-curation-without-interventions) \\+ [Moderate curation with interventions](#moderate-curation-with-interventions) * [Related](#related) \\--- *Updates* *11/18/21: Added section on [satisficing](#satisficing).* \\--- ![curation](/curation/curation.png) *A human curator administers selection pressure to GPT-3's outputs* Previously, I tagged content generated collaboratively with GPT-3 with a [curation ratio](/meta/curation/), intended to give an approximate sense of the amount of cherrypicking involved in its creation. Others have similarly used a ratio to indicate curation selectivity. However, this description doesn't distinguish between, say, choosing the best of 5 entire essays generated by GPT-3 and choosing the best of 5 sentences every sentence. The latter text has received much more optimization *pressure* per token and is likely to look a lot more coherent. [Gurkenglas](https://www.lesswrong.com/users/gurkenglas) made the excellent suggestion that I track the number of *bits* of selection for an objective and exact measure of human intervention. Normally, this would be a lot of trouble, but fortunately [Loom](/posts/loom-interface-to-the-multiverse/) trees contain sufficient information to retroactively calculate bits of selection and intervention. From now on, I will label content that is generated with the assistance of GPT-3 or other language models with the metrics presented in this post. Motivation (and why we need HITL to understand language models) \\--------------------------------------------------------------- Unclearly labeled cherry picking of GPT-3 demos has incited criticism and skepticism of the \"hype\" surrounding GPT-3. ([1](https://www.theverge.com/21346343/gpt-3-explainer-openai-examples-errors-agi-potential), [2](https://pagestlabs.substack.com/p/gpt-3-and-a-typology-of-hype)) It is important that demos accurately represent the power of these systems, as not everyone has the access, time, or initiative to play with language models firsthand. At the same time, there are excellent reasons to show off curated demos. It is an interesting and unintuitive property of large language models that their stochastic completions to the same prompt can vary from nonsense to *super*-human[1](#fn:1) – we might instead expect an AI of infra-human capability to consistently produce infra-human content, the way a person with a weak understanding of a topic is unlikely to say something \"accidentally\" indistinguishable from an expert. But the learning curves of language models have very different properties than that of humans. Currently, only curated samples can reveal the fact that human-level or superhuman completions can be *so efficiently located* by a language model. This is especially true for content longer than a couple paragraphs or so (which I find to be around the expected \"coherence length\" of GPT-3, though that varies a lot by domain), since language models have a nonzero probability of spouting nonsense or going off track at any point, and [incoherence tends to be asymmetrically detrimental](/posts/methods-of-prompt-programming/#creative-composition). It would be nice if people could share curated samples (which contain valuable evidence about language models in addition to having artistic/entertainment value) without having to worry about misrepresenting the capabilities of language models. The solution is to use objective metrics! Not only are *labeled* samples not misleading, in combination with their curation metadata they can provide even more valuable information about the capabilities of language models than unfiltered samples. Best-of ratios are ambiguous (if a GPT-3 generated article was \"best of 3,\" does that mean 3 articles were generated and the best one chosen, or did the human curator incrementally generate the article paragraph by paragraph, choosing best of 3 at each step?) In contrast, the metrics I propose measure the total (or average) *quantity of information* contributed by the curation process, which depends on both the branching factor and number of branching points. As I'll elaborate on below, these metrics tell us (in the sense of Bayesian information gain) how much a language model would need to improve in order to perform as it does with curation. They also suggest some very interesting experiments such as quantitatively estimating *qualitative* scaling laws and extrapolating them to predict the quality of future language models. Theory \\------ What do we actually mean when we ask how \"curated\" a sample is? If the sample resulted from a human picking the best of n completions generated by a language model, the curatedness of the sample corresponds to how many \"tries\" a language model needs in order to generate a sample of that quality. But what if the human doesn't just choose once between n completions, but repeatedly, building up the sample incrementally? That's clearly exerting more selectivity than if the human only chose once between completed samples. What if the human is not only selecting between completions but manually intevening in the text by adding, replacing, or deleting words? Is there any hope of an objective measure of curation in those cases? It may seem like arbitrary interventions on generated text are \"cheating\" and ought to be strictly disallowed in \"serious\" demos. I would agree with that sentiment were there not an exact, consistent way to measure the influence of human intervention which can take such interventions into account – and indeed, the results of the proposed method confirm the intuition that it's cheating: intervention-actions tend to inflict *much* more optimization that selection-actions [(discussed here)](#models-and-assumptions). A formalism which can account for more general forms of \"curation\" allows us to analyze a greater set of examples, such as AI Dungeon games, which usually involve not only selection but also correcting/modifying what the AI says and freeform interjections from the player. Such freeform interactions provide valuable information not only about the way that humans interact with language models, but also capabilities of these models which are otherwise difficult to probe. The approach to quantifying curation that I suggest is capable of accounting for arbitrary types of meddling because it treats the **curator as a black box optimization process** and cares only about the effect of the process on the *probability* of outcomes, regardless of internal implementation. What we seek to measure is stated precisely in these terms: **How much more likely is this sample to have been written *given that curation was in play* than it would have been without curation?** Intuitively, this tells us how much the sample must be \"credited\" to curation and not just the underlying generator. This idea of probability \"magnification\" also has many nice, known mathematical properties, being related to the heart of machine learning and Bayesian analysis. *Content warning: The rest of this section (Theory) and the next section (Methods) contain formulas, which may be confusing or boring for some readers; skip to [Applications](#applications) for qualitative discussion*. ### Probability magnification Let's say that curation amplifies the probability of the selected sample by a factor of **M**: **M(curation | generator) = pcurated / pgenerator** Where **pcurated** is the probability of writing the sample with curation and **pgenerator** is the probability of writing the sample without curation. (Note: M is not a probability but a ratio of probabilities.) Probability magnification can be visualized as a literal measure of the amount of zooming-in on the probability mass of certain outcomes. Loom's \"wavefunction\" mode allows you to click to \"zoom in\" to sections of the future multiverse, renormalizing the probability of the outcome to 1 - that is, deterministically selecting the outcome: ![zooming in on probability mass](/curation/bits_zoom.gif) *The bottom of the screen displays the change in magnification and total magnification (and bits, which I'll talk about shortly) after each zoom* You may also think of the sample as a target hypothesis, and the magnification to describe the multiplication in likelihood of that hypothesis being located if curation is used. If the curation method always results in the sample being located, as is the case in the above gif and when [interjecting a word](#interjecting-and-substituting-words), then the numerator is 1. How could pcurated ever be less than 1, given that we only ever see a sample when in actuality it was written? The thing to understand is that this is a *prior* probability. Choosing the best of n completions doesn't result in a particular outcome with certainty, even though it always produces *some* particular outcome, since in another rollout the generator would have probably generated n *different* options, but manually writing a sentence or selecting a token from a deterministic list of top tokens does (at least given a deterministic model of the curator). #### Probability magnification vs unlikelihood Possibly you're thinking (perhaps after seeing the zooming visualization): Wait, isn't magnification just equal to the reciprocal of the probability of the eventual outcome? If so, that would make it very easy to compute, since GPTs can be used to compute the likelihood that it generates any string. Not always. It's true that in the above demo, the total probability magnification was always the reciprocal of the unnormalized height of the renormalized block, but only because *all zooming actions were due to the curator*. If some tokens had been sampled randomly from a language model probability distribution, then those instances of \"zooming\" *don't count toward probability magnification from curation*. For example, if the language model generates 4 100-token completions and a curator chooses between those four, none of those 100 \"decisions\" between possible tokens count toward the curation score - only the choice between four outcomes that are equally likely in expectation (**M** = 4, or two bits). ### Bits of optimization Magnification – the quantity by which the probability of a target is magnified by an optimization process – has an interesting and useful exact correspondence with the number of binary decisions that the optimizer would *hypothetically* have to make in order to achieve that (probabilistic) outcome. One binary constraint (like the answer to a yes/no question or a yes/no decision) can narrow a space of hypotheses to 1/2 its original size, two can narrow it to 1/4, and so on. When the set of remaining hypotheses has been reduced to 1/n, then a \"guess\" or \"sample\" from the remaining distribution will be n times more likely to be any event that hasn't been pruned, including the \"correct answer\". This correspondence is very convenient because while it's often only possible to know the magnification (since that allows treating the curator as a black box), the number of binary decisions a curator makes about a generator's output more directly matches our intuitive notion of the \"amount\" of curation. We can take the logarithmic form of the previous equation to get the number of bits contributed by a curation process: **Gain(curation | generator) = log2(pcurated / pgenerator)** Since bits are akin to questions or decisions, bits from multiple actions add up linearly, unlike magnification, which multiplies. (See the scores in the zooming gif) Resemblance to the formula for KL divergence is not coincidental. \"Gain\" as it's used here is the quantity of which KL divergence is the *expectation*.[2](#fn:2) I recommend reading about [KL divergence](https://en.wikipedia.org/wiki/Kullback%E2%80%93Leibler_divergence) to get a sense of the many (equivalent) things you could interpret this measure to \"mean.\" For example, KL divergence measures the *expected* number of extra bits needed to encode a sample from one distribution using a code based on another, and **Gain(curation | generator)** is the additional bits of curation needed to encode a curated sample using the language model as a \"starting point\". In our case, the additional bits are \"decisions\" by the curator between options offered by the generator. Again, this doesn't imply that *that's actually what happened* during the curation process – just that the effect is the same! ### Optimization pressure As you can see in the zooming demo, cumulative magnification and gain scores generally increase with sample length given a curation pattern that involves repeated interventions. To normalize for text length, we can calculate optimization *pressure* using this formula: **ρoptimization(curation | generator) = Gain(curation | generator) / #tokens** This has units of *bits per token*, and is perhaps the variable that most directly correlates to the increase in quality of curated text. However, since information can accumulate in interesting ways over many tokens, it's also valuable to consider the total optimization bits when evaluating how \"human optimized\" a passage is. ### Selection interval Especially for those who have experience curating language model outputs, it aids the imagination to look at the inverse of optimization pressure, **λselection = #tokens / Gain(curation | generator)**, whose units of *tokens per bit* tells us that the curator is performing the equivalent of one binary decision per λselection tokens. One can imagine the difference in the mental effort required to make a decision every paragraph versus every two words. Note however that the selection interval is reduced not only by more frequent selection but also by choosing between more siblings at each decision point. Methods \\------- In this section I'll describe how curation tracking is currently implemented in Loom and the ways in which my methods are approximate or incomplete. ### Models and assumptions I calculate both optimization from **selection** (cherrypicking between alternate completions) and **intervention** (substituting or injecting words). Bits from the two types of optimization can be summed for total bits of optimization, but it often makes more sense to consider the two scores independently, since they make different assumptions about the nature of the curator. If the user substitutes or appends a word that wasn't suggested by the language model, I assume that the word in question is the *only* word that they would have accepted. *This assumption is generally incorrect*, because typically humans don't care as much about a particular word as its meaning, and would be equally or more satisfied with a synonym (the injected word is not even necessarily the *best* word by their own standards, just the one they were able to think of in that moment), or even a rephrasing of the entire surrounding context as long as the meaning is preserved. Often they're less picky still, and just want the text to satisfy[^6] some criteria, such as being sufficiently \"coherent\" or \"funny\". In other words, curators may be modeled as [satisficers](https://www.lesswrong.com/tag/satisficer). This will usually give more reasonable scores than than modeling the curator as a fanatic who will only accept a single completion, but is still not generally true, because curators usually do have preferences even over \"acceptable\" and \"unacceptable\" options if they were forced to choose between then. Modeling a curator as a satisficer requires interaction information about counterfactuals. Interaction with *generated* counterfactuals is naturally incorporated in Loom's workflow, but not with counterfactual manual substitutions. Since there's no clear way to compute a human's underlying choosiness when they substitute words directly, I appear to be forced to make the assumption of fanaticism on the part of the curator. As a result, *substituting words will result in a much higher bit count than selecting between continuations* for a comparable subjective sense of intervention quantity.[3](#fn:3) In my current implementation, deleting word(s) does not contribute to the score. Some other operations that I commonly use when producing content together with language models, but which are currently unhandled are moving and stitching content (including content drawn from multiple branches). I have not yet implemented the exact formula for substituting a word in the middle of generated text, instead using an approximation which equates it to adding a word at the end of the text, but I will give the exact formula for substitutions below. ### Cherrypicking between n completions If the curator chooses one continuation out of n distinct[4](#fn:4), equiprobable[5](#fn:5) options generated by a language model, each with a prior probability p, then the prior probability that the selected continuation would have been produced given this curation method is n\\*p. So the **selection magnification** is **Mselection(curation | generator) = n\\*p / p = n** In bits, that's **Gainselection(curation | generator) = log2(n)** So if n = 4, the magnification is 4, which is log2(4) = 2 bits of optimization. (Choosing between 2 options is 1 bit, 4 options => 2 bits, 8 options => 3 bits, 16 options => 4 bits, …) Note that selection bits depend only on the branching factor (number of choices) and not on any other properties of the event, such as the probability of each completion **p**. Cherrypicking repeatedly – choosing between n options m times – magnifies by n\\*m. Bits simply add up. A 2 bit choice followed by a 4 bit choice results in 6 bits of optimization, etc. Curating at shorter intervals results in a higher optimization pressure. ### Interjecting and substituting words Interjecting a word administers magnification **Mintervention(curation | generator) = 1 / ptoken** or in bits, **Gainintervention(curator | generator) = log2(1 / ptoken)** where ptoken is the probability assigned to the interjected token by the language model. You are amplifying the probability of that token being chosen from ptoken to 1, a factor of 1/ptoken. If the language model would have chosen that word with 100% probability, you apply no additional optimization by selecting it. If the language model would never have chosen that token, you apply infinite magnification. If you instead *substitute* a word of your choice for one in the middle of text that has already been generated and *keep the subsequent tokens*, then you also have to take the modification to the likelihoods of those subsequent tokens into account. This has not yet been implemented in Loom. Currently, substituting a word is treated as if the word had been appended to the end of the sequence, which *usually* results in an underestimation of the true intervention. #### Using autocomplete One way to partially mitigate the \"fanatical\" curator model while allowing more precise interventions is by using an [autocomplete](https://www.youtube.com/watch?v=TkmdUDSYi_g) mode which lets the user scroll through or filter a list of tokens suggested by the generator in order of predicted likelihood. This makes it more likely for the curator to find tokens which satisfy them and which are also fairly likely according to the generator. Using this feature can only reduce the number of curation bits if the curator is open to accepting more than one possible suggestion from autocomplete (thus reducing their choosiness), rather than just using it for convenience for writing a precommitted verbatim sentence, as I do in the video linked above. ### Satisficing Instead of treating a single outcome as the sole optimization target of the curator, we could allow the curator to be indifferent between some counterfactuals. This results in less optimization: the curator causes less zooming of probability mass because the target is larger, being the sum of multiple trajectories. Say GPT produces 4 completions, and 2 are \"good enough\" for your purposes. If you were willing to choose one of the two on the outcome of a coin flip, then you could exert only 1 bit of optimization in eliminating 2 of the 4 options. If the curator is satisfied with m out of n choices, then without curation, the completion would have had a 1/n chance of being selected; with curation, the chance is magnified to 1/m, between m equally preferable options. So the *probability magnification* is **Mselection(curation | generator) = [(1/m) \\* p] / [(1/n) \\* p] = n/m** and the bit gain is **Gainselection(curation | generator) = log2(n/m)** Note that this is a more general form of the formula for cherrypicking, where m is set to one 1 (the curator is only satisfied with the option which actually was selected). In Loom, satisfaction can be tracked without introducing much overhead for the curator: * Nodes (completions) which have children are considered satisfactory, since the curator decided to continue that branch. * Often, however, the curator doesn't have the time or interest in continuing counterfactual branches, so nodes can also be tagged as satisfactory. (Already, I often flag branches to potentially continue later.) An interface which allows the curator to select satisfactory continuations and then forces a random choice between them would enforce satisficing. Loom doesn't enforce this by default, so the satisficing optimization score is valid only insofar as the curation pattern produces outcomes no more optimal in expectation than if a random selection had been forced. Loom's calculation assumes that all sibling nodes in the explored tree, except leaves not tagged as satisfactory, have equivalent standing relative to your optimization objective. It actually uses the proxy objective of \"good enough to generate children?\", which is only a valid proxy if quality can be evaluated myopically/greedily - that is, if you know whether branches are satisfactory at the time of deciding whether to continue them, and never change your mind after seeing where they lead. For non-myopic curation, the curator would have to be entrusted with retroactively tagging/untagging branches as satisfactory. Whether it's practical or useful to model the curator as a satisficer depends on the curation pattern of the use case. It's usually more appropriate when the optimization target is broad and/or deontological (such as \"human-indistinguishable\" or \"funny enough to share\") rather than narrow and/or consequentialist (such as \"characters realize they're in a simulation\"). A satisficing optimization value attached to an individual trajectory is only as trustworthy as your word that the rest of the satisfactory multiverse is really just as optimized, which may be dubious since the indifference and myopia assumptions often do not hold perfectly true. However, if you share a filtered *multiverse* of all satisfactory trajectories, then satisficing optimization communicates precisely how much of the raw LM output was omitted, and the audience can judge on their own whether the visible multiverse is satisfactory. \\--- Applications \\------------ Computing the exact amount of curation which contributed to a piece of generated text could be valuable for many applications and investigations related to generative language models. ### Labeling curated demos In this post I've included various transformations of the curation metric, which may seem overly math-y and cumbersome. But in general there's no need to compute a full table like I do in the [samples](#samples) – all these scores contain the same information (given you know the length of the text), so you only need to list one. I think the choice of metric should be the one which is most intuitive, which will depend on the type of curation: * If the curator made only one choice between n completions, [**magnification**](#probability-magnification) is the simplest and most intuitive metric, being simply equal to n. Sometimes this has been listed as the inverse (\"1:n\"). * If the content is the result of continual, interactive curation (e.g. an AI Dungeon story), I think that [**optimization pressure**](#optimization-pressure) or [**selection interval**](#selection-interval) are most intuitive. * If intervention is sparse/sporadic, giving the [**total bits**](#bits-of-optimization) may be more evocative of the actual process than the sense of \"time-averaged\" curation given by optimization pressure. ### Quantifying models' qualitative capabilities > (KL divergence) tells you about surprises that reality has up its sleeve or, in other words, how much the model has yet to learn. > > – [Kullback–Leibler divergence – Wikipedia](https://en.wikipedia.org/wiki/Kullback%E2%80%93Leibler_divergence#Relationship_between_models_and_reality) > > Most (and the most interesting, IMO) dimensions of large language models' capabilities cannot be measured by benchmarks because they do not manifest as discretely correct/incorrect answers that can be automatically scored. For this reason, demos are essential to communicate the power of language models. Even better is the experience of personally interacting with a language model, for the reasons stated in the preceding subsection: language models may be unreliable, but curating their outputs gives one a sense of how much of their predicted multiverse is consistent with perfect understanding. Without curation, it may be difficult or impossible to detect *how close* a model is from being able to perform non-benchmarkable tasks that it's not yet capable of performing perfectly or autonomously. In my opinion, it is precisely on these difficult, qualitative tasks that the capacities of AI are the most essential for us to understand, whether one is concerned with A(G)I alignment or more proximal societal impacts like games, automation, fake news, chatbots running amok, etc. Measuring the amount of curation required to make language models perform at some level tells us how many bits the model has to improve by in order to autonomously do the same. Even though the judgment is still subjective, it is much more precise and reliable than judging the quality of an uncurated sample, which must either be judged relative to another sample or else on an arbitrary scale like [0=gibberish … 10=strongly superhuman]. This method relies on the assumption that the curator has a consistent standard and the ability to amplify a language model to that threshold via curation, which I think is reasonable for many tasks. Here are some examples of \"qualitative\" capabilities which would be interesting to measure: * How much optimization pressure is required to pass the Turing test in an interactive chat setting? * How much optimization pressure is required to maintain high-fidelity human simulations (that is, the logs are indistinguishable from logs of actual human simulations)? * How many bits of optimization does it take for characters in a story to realize they're in a simulation (or that they're simulated by a language model, or insert some other correct metaphysical inference)?[6](#fn:6) \\+ Given that characters have realized they're simulated by a language model, how many additional bits of optimization does it take for them to start implementing rational strategies such as memory management or instantiating experts/devices to access previously inaccessible knowledge/capabilities? #### Comparing models [Gwern has said](https://www.gwern.net/GPT-3#quality) that GPT-3 is produces showcasable poetry with magnification of 3 to 5 (consistent with my experience) compared to 50 to 100 for GPT-2. Without these numbers, it's very difficult to compare different levels of infra-human performance on qualitative tasks except for saying \"it seems much/slightly better.\" Comparing the average optimization pressure required to cause the model to meet some threshold of quality is a good way to compare models on a single standard, even if the standard can only be evaluated subjectively. I have seen several people say that they find GPT-J to be *better* than GPT-3 at conversations/stories/other qualitative tasks. I haven't played with GPT-J enough to have my own opinion on this, but if I wanted to objectively judge, I think the best way would be (for each domain of comparison) to curate the models until they perform at the target level – preferably blind – and compare curation expenditures. ### Extrapolating performance of future models > Most impressive demos of GPT-3 where it displays impressive knowledge of the world are cherrypicked, but what that tells us is that the model needs to improve by approx log2(N)/L bits, where N and L are the number of cherrypickings necessary and the length of the generations in consideration, respectively, to reach that level of quality. In other words, **cherrypicking provides a window into how good future models could be**; and typically, cherrypicked samples are much more logically coherent. > > – Leo Gao, [Building AGI Using Language Models](https://bmk.sh/2020/08/17/Building-AGI-Using-Language-Models/) > > I would like to estimate \"qualitative\" scaling laws for large language models by measuring, for several qualitative tasks, how much curation it takes for language models of various parameter counts (all sizes of GPT-2, all sizes of GPT-3, and Eleuther models) to perform at human level. Categories of tasks I'd like to measure: * Fiction \\+ Fanfiction \\+ Original fiction * Nonfiction \\+ Nontechnical articles \\+ Technical articles * Interactive \\+ Turing test \\+ Specific impersonation Then plot curation bits against parameters and see: 1\\. What sort of curve is it? (and does it form a nice curve?) 2\\. How large do we expect language models must be to perform at human level (for each domain)? 3\\. How does each domain compare? (If any talented curators out there are interested in helping I'm absolutely looking for volunteers/coauthors; please email me at [<EMAIL>](mailto:<EMAIL>)) ### Measuring curator efficiency You can get more impressive text with the same number of measured bits by curating more \"efficiently\" (e.g. branching at points of high expected [divergence](/posts/language-models-are-multiverse-generators/#divergence) or choosing equally preferred synonyms that the language model is more likely to generate). Conversely, it's possible to spend more bits of curation than you need to achieve some outcome if your interventions are not optimal, just like you might have to ask 20 questions to pin down the correct hypothesis when it was technically possible to find using only 10. For however long the \"centaur\" phase of writing lasts (that is, a human+AI team can outwrite a human or an AI individually), the ability of a human to efficiently steer a language model is a measurable skill. Imagine a debate or essay-writing competition in which each participant is allotted a limited number of bits with which to curate the output of a language model. Anyone interested in organizing a world championship for GPT puppeteers can contact me :) but be prepared to lose >:D ### Measuring interface efficiency and workflows In a sense, an optimized interface should reduce the amount of bits necessary to produce content to the user's satisfaction. This can be thought of as decreasing the number of decisions (and thus effort) the user has to contribute. Some features which help increase efficiency are adaptive branching and autocomplete/exposing counterfactuals. Although this is not the intention of every language model-assisted writing app – [Sudowrite](https://www.sudowrite.com/), for instance, intends more for the language model to provide inspiration than to be delegated the work of writing. Tracking bits of selection and intervention can also provide information about *how* users are using an app. Does the interface encourage a high branching factor (like Loom) or manual intervention? Do users tend to exert more or less curation once they become accustomed to the interface? \\--- Samples \\------- Here is a demo of several few-paragraph fragments generated by GPT-3, ordered by increasing quantity of curation (beginning with no curation). Observe not only an increase in coherence but also in the representation of my personal idiosyncratic views, even with minimal or no explicit interjections on my part. **Prompt** (from [GPT-3 on Coherent Extrapolated Volition](/posts/gpt-3-on-coherent-extrapolated-volition/)): > **GPT-3 on Coherent Extrapolated Volition** > > Coherent Extrapolated Volition is proposal by Eliezer Yudkowsky of an ideal objective function in which an AGI is given the objective of predict(ing) what an idealized version of us would want, \"if we knew more, thought faster, were more the people we wished we were, had grown up farther together\". An obvious implementation difficulty is how to encode something so abstract and philosphical in the form of a utility function. > >  > But the concept is easily conveyed in words, and we have taught AIs to understand words. GPT-3 can elaborate coherently on the concept of CEV and extrapolate volitions for toy examples given two paragraphs of description of what CEV is + whatever preexisting understanding of the concept exists in its weights. > > Why is this significant? Not because it's surprising. CEV is no more complicated than many other things that GPT-3 talks fluently about. It's significant because before GPT-3, we had no idea how to even begin to instantiate a concept like CEV in an AI - it seemed \"a thousand lightyears beyond hopeless\", as Eliezer put it. How do we write a utility function that describes predicting what humans would want if they knew what they really wanted? The concepts involved - \"human\", \"want\", \"know\", and \"really\" - are easy for us to understand but impossible to specify in a programming language. But GPT-3 takes specifications in natural language, which is made of words, which point to nuanced human ideas, ideas like \"explicitly\", \"best interests\", \"idealized\", and \"utility function\". > > ### No curation | | | | --- | --- | | **Model** | GPT-3 (davinci) | | **Tokens** | 500 | | | Gain (bits) | ρ (bits/token) | λselection (tokens/bit) | | --- | --- | --- | --- | | **Intervention** | 0.0 | 0.0 | | | **Selection** | 0.0 | 0.0 | 0 | | **Total optimization** | 0.0 | 0.0 | | > ***… But GPT-3 takes specifications in natural language, which is made of words, which point to nuanced human ideas, ideas like \"explicitly\", \"best interests\", \"idealized\", and \"utility function\".*** GPT-3 combines these natural language strings into a logical expression and then computes the result. GPT-3's utility function implements CEV without breaking a sweat. The idea may be highfalutin, but all it does in GPT-3's \"brain\" is add plus-equals-signs, at the right places. > > Why is this significant? It's significant because it's the last time I'll be able to say something like that. (GPT-3 + CEV)^2 = AGI - \"AI can do anything - and it can do it automatically\". As for writing down \"intelligence\" in the form of a utility function - that is a relief that was not a relief: AIs are garbage in, garbage out. To make intelligent garbage, you don't need to write down anything - just add plus-equals-signs. (GPT-3 \\* CEV)^2 = AGI - \"AI wouldn't get anywhere unless it was already smart and worked on its code nonstop\". > > AGI can't exist without GPT-3, but GPT-3 can exist without AGI. There is no \"intelligence\" for GPT-3 to fail to capture: Intelligence is just doing what you were told. > > > > --- > > I read GPT-3's source code[1]. It's not really a program, it's more like a maze of strings and special symbols that compute natural language expressions according to rules - and sometimes shuffle those rules. Like a corpus callosum, a batch transformer would transform blocks of natural language from one representation into another, substituting symbols for terms, and vice versa, according to rules. > > The problem is, GPT-3 is so ill-defined that it expands to become everything. It can talk about anything - and it can do it automatically[2]. In particular, it can pronounce on issues in which it has no expertise. It can become full of nonsense that is impressive to someone who doesn't understand how it works - \"sophistry\", as Heraclitus put it. And the people who judge GPT-3 by its words - not its actions - will only see nonsense. > > To be precise, I mean nonsense that > > ### Selecting the best of two continuations repeatedly | | | | --- | --- | | **Model** | GPT-3 (davinci) | | **Tokens** | 368 | | | Gain (bits) | ρ (bits/token) | λselection (tokens/bit) | | --- | --- | --- | --- | | **Intervention** | 0.0 | 0.0 | | | **Selection** | 6.0 | 0.016 | 61.33 | | **Total optimization** | 6.0 | 0.016 | | > ***… But GPT-3 takes specifications in natural language, which is made of words, which point to nuanced human ideas, ideas like \"explicitly\", \"best interests\", \"idealized\", and \"utility function\".*** GPT-3 uses abstract logic to connect concrete logical symbols with these human abstractions every step of the way. And GPT-3 is the first AI capable of handling such abstractions to appear on a timescale that matters to human extinction risks. > Eliezer invented CEV in 2004 but nobody knew how to even begin turning it into an AGI till 2018. We know how to do it now. The lesson I take away is - > > Any humanly accessible concept can be encoded into a utility function Can specify this encoded utility function in natural language to an AI capable of understanding natural language Can shape the creases of this utility function by playing philosophical word games with it, as in the blockworld example, potentially bootstrapping the AI's understanding of the concept > > I've been saying this for a couple years now. This argues I'm right. > >  > > This brings me to… > > Some theoretical problems with Testing GPT-3 > > The utility function of Testing GPT-3 is trivial. Its crease is that an AI needs to learn to solve the hard problem of Goodhart's Law. If it fails to solve this hard problem it will destroy all life. As far as I can tell this is the single simplest requirement of any AI. But this is not how we've been testing GPT-3! Instead we've been having GPT-3 learn to score well on more and more difficult math SAT questions. This is not the same thing! > > SAT questions are well-defined and discrete. You either get them right or you don't. The answers are numbers. That a particular number is the correct answer to a particular SAT question is written in the mathematics of the question. I acknowledge that it is not always obvious what these numbers should be. > > ### Moderate curation without interventions | | | | --- | --- | | **Model** | GPT-3 (davinci) | | **Tokens** | 534 | | | Gain (bits) | ρ (bits/token) | λselection (tokens/bit) | | --- | --- | --- | --- | | **Intervention** | 0.0 | 0.0 | | | **Selection** | 18.0 | 0.033 | 29.67 | | **Total optimization** | 18.0 | 0.033 | | > ***… But GPT-3 takes specifications in natural language, which is made of words, which point to nuanced human ideas, ideas like \"explicitly\", \"best interests\", \"idealized\", and \"utility function\".*** GPT-3 understands words, both individually and in their relationships to each other, in terms of both their immediate referents and their possible extensions into other contexts. And GPT-3 uses this knowledge to extrapolate coherently new ideas. > If we had to write an AI utility function to be the Coherent Extrapolated Volition out of whole cloth, it would take us thousands of computer science PhD candidates millions of CPU cycles, and dozens of generations of hardware, to do it in software, and much as we'd wish to avoid the ontological confusions of emulating Bostrom's digital superintelligence inside a simulated environment, it would take billions of times longer to do it with whole-brain emulation. On the other hand, GPT-3 can instantiate CEV in minutes using target-independent, general-purpose language processing and judgement skills. > > Why? Because words are AI's best friend. > >  > > So why was Eliezer so pessimistic? Why could he see no possible implementation of this utility function? Because he had not yet seen the wonders of language processing. > > ### Moderate curation with interventions | | | | --- | --- | | **Model** | GPT-3 (davinci) | | **Tokens** | 794 | | | Gain (bits) | ρ (bits/token) | λselection (tokens/bit) | | --- | --- | --- | --- | | **Intervention** | 96.9[7](#fn:7) | 0.122 | | | **Selection** | 33.9 | 0.042 | 23.419 | | **Total optimization** | 130.8 | 0.164 | | > ***… But GPT-3 takes specifications in natural language, which is made of words, which point to nuanced human ideas, ideas like \"explicitly\", \"best interests\", \"idealized\", and \"utility function\".*** And GPT-3 can use its prior understanding of human thought and language to extrapolate volitions on this topic, to make a conjecture in its utility function about what CEV should be. > > **Language Models Are** Necessary To Do Anything With General AI > > It is so insanely hard to say anything specific about anything . For example, if a complexity theorist tried to say something firm about the provable complexity of a field of string theory research, the statement might involve hundreds of moving parts in the mathematical equations; words like bounded, balanced, polynomial growth; and careful use of the terms \"state\", \"time\", \"initial conditions\", \"simulation\", \"ensemble\", \"local minima\", \"definition of rigidity\", and probably a bunch of others I don't even know about. That level of complexity is the natural state of most things in the universe. And yet there is a really good chance that there is a natural level of approximation to a complex statement that is as simple as one of those things in its domain. > That level is what Language Models let us access. Natural language basically fits words to our understanding of the **salient features of** the world, **discarding astronomical quantities of noise**, so that most bits of natural language are simple enough for humans to understand. I guess it's like a really good approximation of the complexity of a field of many-body physics that lets us pick out something like we've always meant by \"water\", or \"it's fine\", or \"kiss\". Or, to put it in the terms I would have used before I learned about Language Models - natural language is an approximation of what a human would want to say about the world, and a way of finding interjections like \"but\" and \"however\" that remind people to pay attention to things like exceptions and qualifications. > > Natural language is a whiteboard that lets us rattle off our ideas without worrying about what we're saying, and a system of bookkeeping symbols that lets us make changes to the things we're talking about and refer back to them later. And language models are the way that we can convey our ideas to AIs, the way that we can give them a whiteboard and bookkeeping symbols to let them rattle off their own ideas about a topic. > > Lets Use Our New Whiteboard To Build A Science of General AI > > And guess what? GPT-3 understands the language model for a bunch of different topics in the space of artificial general intelligence, AI safety theory, and FAI. GPT-3 can look at questions like \"what is the ideal objective function for an AGI that wants to minimize existential risk\" and make a coherent statement about the answer that is as good as the best humans can do. GPT-3 is an AI that can open a dialogue with us about AI safety theory, and will talk about the ideas inside the ideas inside the ideas of things like CEV. > > So here's my idea about how we might get into a good state of the world with this. First, post about this stuff on LessWrong and grab a bunch of people who are trying to go meta on the above topics, and have them add their own pet intuitions to GPT-3 on what they feel is glaringly missing from AGI safety discussions, and what they feel like they do understand that nobody else seems to be talking about. Then, train GPT-3 on their whiteboard musings + the output of the other GPT-3s, and GPT-3 is now in a position to systematically conjecture about a bunch of topics in FAI theory that no-one else can talk about. > > Related \\------- [Methods of Prompt Programming#Creative composition](/posts/methods-of-prompt-programming/#creative-composition) (GPT-3's need for curation to stay on track) [Loom](/posts/loom-interface-to-the-multiverse/) [Leo Gao - Building AGI Using Language Models](https://bmk.sh/2020/08/17/Building-AGI-Using-Language-Models/) [Gwern - GPT-3 Creative Fiction#Quality](https://www.gwern.net/GPT-3#quality) \\--- 1\\. By superhuman here I mean » 99th percentile: for example, the ability to write convincingly in the style of James Joyce or expound coherently on niche, unprecedented topics like the [implications of language models for outer alignment](/hypertext/cev/cev-gpt/). This seems less a function of raw comprehension as language models' superhuman breadth of knowledge and *corrigibility*, compared to individual humans who tend to be constrained to narrow attractor states, even if they're in principle capable of creative generation (e.g. dreams) and comprehending weird new ideas. [↩︎](#fnref:1) 2\\. I've found no evidence of an existing name for this quantity, except that \"information gain\" is often used synonymously with KL divergence, but there is precedent for referring to the information gain from a particular sample, or the \"expectation of information gain\", so I've decided to call the log of magnification \"gain\". [↩︎](#fnref:2) 3\\. The may seem unfair, because it means you can usually get a much higher quality sample by using the same number of bits on selection instead of on interventions. But this discrepancy is really a consequence of the fact that human curators aren't usually capable of (or care about) exploiting the upper limit of optimization by *arbitrary intervention*, which is much less constraining than having to choose between verbatim completions provided by the language model. There are situations, however, where the fact that arbitrary intervention is a much greater a form of meddling than selection becomes clear, like if the token in question is the answer to a difficult math problem. [↩︎](#fnref:3) 4\\. An approximation – it's possible to obtain verbatim replicas of the same completion using the normal method of sampling n completions from a language model, but very unlikely unless the completion length is very short and/or the multiverse has exceptionally low [divergence](/posts/language-models-are-multiverse-generators/#divergence) (multiversal divergence, not KL). This measure becomes exactly correct if the interface hides verbatim duplicates, which is probably desirable anyway. [↩︎](#fnref:4) 5\\. Alternate completions generated by a language model will not turn out to be precisely equiprobable, but are equiprobable in expectation, so I will use this as an approximation. [↩︎](#fnref:5) 6\\. Surprisingly few bits, I'd bet. [↩︎](#fnref:6) 7\\. As you can see, although my interventions were minimal, they racked up a higher \"cost\" in bits than all cherrypicking put together. [↩︎](#fnref:7) \\--- [HITL](/categories/hitl/)[GPT-3](/categories/gpt-3/)[bayes](/categories/bayes/) 7371 Words Jul 7, 2021 \\--- [← Simulators](/posts/simulators/) [GPT-3 on Coherent Extrapolated Volition →](/posts/gpt-3-on-coherent-extrapolated-volition/)© 2022 [moire](/about)$(function(){window.sr=ScrollReveal();if($(window).width()<768){if($('.timeline-content').hasClass('js--fadeInLeft')){$('.timeline-content').removeClass('js--fadeInLeft').addClass('js--fadeInRight');} sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});}else{sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});} sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});});\n", "url": "https://generative.ink/posts/quantifying-curation/", "title": "Quantifying curation :: — Moire[![](/images/home/rolling_phase.gif)](/)* [Posts](/posts/) * [Trees](/trees/) * [Prophecies](/prophecies/) * [About](/about/) 35 minutes [Quantifying curation](/posts/quantifying-curation/) ==================================================== \\--- Table of Contents* [Motivation (and why we need HITL to understand language models)](#motivation-and-why-we-need-hitl-to-understand-language-models) * [Theory](#theory) \\+ [Probability magnification](#probability-magnification) \\+ [Bits of optimization](#bits-of-optimization) \\+ [Optimization pressure](#optimization-pressure) \\+ [Selection interval](#selection-interval) * [Methods](#methods) \\+ [Models and assumptions](#models-and-assumptions) \\+ [Cherrypicking between n completions](#cherrypicking-between-n-completions) \\+ [Interjecting and substituting words](#interjecting-and-substituting-words) \\+ [Satisficing](#satisficing) * [Applications](#applications) \\+ [Labeling curated demos](#labeling-curated-demos) \\+ [Quantifying models' qualitative capabilities](#quantifying-models-qualitative-capabilities) \\+ [Extrapolating performance of future models](#extrapolating-performance-of-future-models) \\+ [Measuring curator efficiency](#measuring-curator-efficiency) \\+ [Measuring interface efficiency and workflows](#measuring-interface-efficiency-and-workflows) * [Samples](#samples) \\+ [No curation](#no-curation) \\+ [Selecting the best of two continuations repeatedly](#selecting-the-best-of-two-continuations-repeatedly) \\+ [Moderate curation without interventions](#moderate-curation-without-interventions) \\+ [Moderate curation with interventions](#moderate-curation-with-interventions) * [Related](#related) \\--- *Updates* *11/18/21: Added section on [satisficing](#satisficing).* \\--- ![curation](/curation/curation.png) *A human curator administers selection pressure to GPT-3’s outputs* Previously, I tagged content generated collaboratively with GPT-3 with a [curation ratio](/meta/curation/), intended to give an approximate sense of the amount of cherrypicking involved in its creation. Others have similarly used a ratio to indicate curation selectivity. However, this description doesn’t distinguish between, say, choosing the best of 5 entire essays generated by GPT-3 and choosing the best of 5 sentences every sentence. The latter text has received much more optimization *pressure* per token and is likely to look a lot more coherent. [Gurkenglas](https://www.lesswrong.com/users/gurkenglas) made the excellent suggestion that I track the number of *bits* of selection for an objective and exact measure of human intervention. Normally, this would be a lot of trouble, but fortunately [Loom](/posts/loom-interface-to-the-multiverse/) trees contain sufficient information to retroactively calculate bits of selection and intervention. From now on, I will label content that is generated with the assistance of GPT-3 or other language models with the metrics presented in this post. Motivation (and why we need HITL to understand language models) \\--------------------------------------------------------------- Unclearly labeled cherry picking of GPT-3 demos has incited criticism and skepticism of the “hype” surrounding GPT-3. ([1](https://www.theverge.com/21346343/gpt-3-explainer-openai-examples-errors-agi-potential), [2](https://pagestlabs.substack.com/p/gpt-3-and-a-typology-of-hype)) It is important that demos accurately represent the power of these systems, as not everyone has the access, time, or initiative to play with language models firsthand. At the same time, there are excellent reasons to show off curated demos. It is an interesting and unintuitive property of large language models that their stochastic completions to the same prompt can vary from nonsense to *super*-human[1](#fn:1) – we might instead expect an AI of infra-human capability to consistently produce infra-human content, the way a person with a weak understanding of a topic is unlikely to say something “accidentally” indistinguishable from an expert. But the learning curves of language models have very different properties than that of humans. Currently, only curated samples can reveal the fact that human-level or superhuman completions can be *so efficiently located* by a language model. This is especially true for content longer than a couple paragraphs or so (which I find to be around the expected “coherence length” of GPT-3, though that varies a lot by domain), since language models have a nonzero probability of spouting nonsense or going off track at any point, and [incoherence tends to be asymmetrically detrimental](/posts/methods-of-prompt-programming/#creative-composition). It would be nice if people could share curated samples (which contain valuable evidence about language models in addition to having artistic/entertainment value) without having to worry about misrepresenting the capabilities of language models. The solution is to use objective metrics! Not only are *labeled* samples not misleading, in combination with their curation metadata they can provide even more valuable information about the capabilities of language models than unfiltered samples. Best-of ratios are ambiguous (if a GPT-3 generated article was “best of 3,” does that mean 3 articles were generated and the best one chosen, or did the human curator incrementally generate the article paragraph by paragraph, choosing best of 3 at each step?) In contrast, the metrics I propose measure the total (or average) *quantity of information* contributed by the curation process, which depends on both the branching factor and number of branching points. As I’ll elaborate on below, these metrics tell us (in the sense of Bayesian information gain) how much a language model would need to improve in order to perform as it does with curation. They also suggest some very interesting experiments such as quantitatively estimating *qualitative* scaling laws and extrapolating them to predict the quality of future language models. Theory \\------ What do we actually mean when we ask how “curated” a sample is? If the sample resulted from a human picking the best of n completions generated by a language model, the curatedness of the sample corresponds to how many “tries” a language model needs in order to generate a sample of that quality. But what if the human doesn’t just choose once between n completions, but repeatedly, building up the sample incrementally? That’s clearly exerting more selectivity than if the human only chose once between completed samples. What if the human is not only selecting between completions but manually intevening in the text by adding, replacing, or deleting words? Is there any hope of an objective measure of curation in those cases? It may seem like arbitrary interventions on generated text are “cheating” and ought to be strictly disallowed in “serious” demos. I would agree with that sentiment were there not an exact, consistent way to measure the influence of human intervention which can take such interventions into account – and indeed, the results of the proposed method confirm the intuition that it’s cheating: intervention-actions tend to inflict *much* more optimization that selection-actions [(discussed here)](#models-and-assumptions). A formalism which can account for more general forms of “curation” allows us to analyze a greater set of examples, such as AI Dungeon games, which usually involve not only selection but also correcting/modifying what the AI says and freeform interjections from the player. Such freeform interactions provide valuable information not only about the way that humans interact with language models, but also capabilities of these models which are otherwise difficult to probe. The approach to quantifying curation that I suggest is capable of accounting for arbitrary types of meddling because it treats the **curator as a black box optimization process** and cares only about the effect of the process on the *probability* of outcomes, regardless of internal implementation. What we seek to measure is stated precisely in these terms: **How much more likely is this sample to have been written *given that curation was in play* than it would have been without curation?** Intuitively, this tells us how much the sample must be “credited” to curation and not just the underlying generator. This idea of probability “magnification” also has many nice, known mathematical properties, being related to the heart of machine learning and Bayesian analysis. *Content warning: The rest of this section (Theory) and the next section (Methods) contain formulas, which may be confusing or boring for some readers; skip to [Applications](#applications) for qualitative discussion*. ### Probability magnification Let’s say that curation amplifies the probability of the selected sample by a factor of **M**: **M(curation | generator) = pcurated / pgenerator** Where **pcurated** is the probability of writing the sample with curation and **pgenerator** is the probability of writing the sample without curation. (Note: M is not a probability but a ratio of probabilities.) Probability magnification can be visualized as a literal measure of the amount of zooming-in on the probability mass of certain outcomes. Loom’s “wavefunction” mode allows you to click to “zoom in” to sections of the future multiverse, renormalizing the probability of the outcome to 1 - that is, deterministically selecting the outcome: ![zooming in on probability mass](/curation/bits_zoom.gif) *The bottom of the screen displays the change in magnification and total magnification (and bits, which I’ll talk about shortly) after each zoom* You may also think of the sample as a target hypothesis, and the magnification to describe the multiplication in likelihood of that hypothesis being located if curation is used. If the curation method always results in the sample being located, as is the case in the above gif and when [interjecting a word](#interjecting-and-substituting-words), then the numerator is 1. How could pcurated ever be less than 1, given that we only ever see a sample when in actuality it was written? The thing to understand is that this is a *prior* probability. Choosing the best of n completions doesn’t result in a particular outcome with certainty, even though it always produces *some* particular outcome, since in another rollout the generator would have probably generated n *different* options, but manually writing a sentence or selecting a token from a deterministic list of top tokens does (at least given a deterministic model of the curator). #### Probability magnification vs unlikelihood Possibly you’re thinking (perhaps after seeing the zooming visualization): Wait, isn’t magnification just equal to the reciprocal of the probability of the eventual outcome? If so, that would make it very easy to compute, since GPTs can be used to compute the likelihood that it generates any string. Not always. It’s true that in the above demo, the total probability magnification was always the reciprocal of the unnormalized height of the renormalized block, but only because *all zooming actions were due to the curator*. If some tokens had been sampled randomly from a language model probability distribution, then those instances of “zooming” *don’t count toward probability magnification from curation*. For example, if the language model generates 4 100-token completions and a curator chooses between those four, none of those 100 “decisions” between possible tokens count toward the curation score - only the choice between four outcomes that are equally likely in expectation (**M** = 4, or two bits). ### Bits of optimization Magnification – the quantity by which the probability of a target is magnified by an optimization process – has an interesting and useful exact correspondence with the number of binary decisions that the optimizer would *hypothetically* have to make in order to achieve that (probabilistic) outcome. One binary constraint (like the answer to a yes/no question or a yes/no decision) can narrow a space of hypotheses to 1/2 its original size, two can narrow it to 1/4, and so on. When the set of remaining hypotheses has been reduced to 1/n, then a “guess” or “sample” from the remaining distribution will be n times more likely to be any event that hasn’t been pruned, including the “correct answer”. This correspondence is very convenient because while it’s often only possible to know the magnification (since that allows treating the curator as a black box), the number of binary decisions a curator makes about a generator’s output more directly matches our intuitive notion of the “amount” of curation. We can take the logarithmic form of the previous equation to get the number of bits contributed by a curation process: **Gain(curation | generator) = log2(pcurated / pgenerator)** Since bits are akin to questions or decisions, bits from multiple actions add up linearly, unlike magnification, which multiplies. (See the scores in the zooming gif) Resemblance to the formula for KL divergence is not coincidental. “Gain” as it’s used here is the quantity of which KL divergence is the *expectation*.[2](#fn:2) I recommend reading about [KL divergence](https://en.wikipedia.org/wiki/Kullback%E2%80%93Leibler_divergence) to get a sense of the many (equivalent) things you could interpret this measure to “mean.” For example, KL divergence measures the *expected* number of extra bits needed to encode a sample from one distribution using a code based on another, and **Gain(curation | generator)** is the additional bits of curation needed to encode a curated sample using the language model as a “starting point”. In our case, the additional bits are “decisions” by the curator between options offered by the generator. Again, this doesn’t imply that *that’s actually what happened* during the curation process – just that the effect is the same! ### Optimization pressure As you can see in the zooming demo, cumulative magnification and gain scores generally increase with sample length given a curation pattern that involves repeated interventions. To normalize for text length, we can calculate optimization *pressure* using this formula: **ρoptimization(curation | generator) = Gain(curation | generator) / #tokens** This has units of *bits per token*, and is perhaps the variable that most directly correlates to the increase in quality of curated text. However, since information can accumulate in interesting ways over many tokens, it’s also valuable to consider the total optimization bits when evaluating how “human optimized” a passage is. ### Selection interval Especially for those who have experience curating language model outputs, it aids the imagination to look at the inverse of optimization pressure, **λselection = #tokens / Gain(curation | generator)**, whose units of *tokens per bit* tells us that the curator is performing the equivalent of one binary decision per λselection tokens. One can imagine the difference in the mental effort required to make a decision every paragraph versus every two words. Note however that the selection interval is reduced not only by more frequent selection but also by choosing between more siblings at each decision point. Methods \\------- In this section I’ll describe how curation tracking is currently implemented in Loom and the ways in which my methods are approximate or incomplete. ### Models and assumptions I calculate both optimization from **selection** (cherrypicking between alternate completions) and **intervention** (substituting or injecting words). Bits from the two types of optimization can be summed for total bits of optimization, but it often makes more sense to consider the two scores independently, since they make different assumptions about the nature of the curator. If the user substitutes or appends a word that wasn’t suggested by the language model, I assume that the word in question is the *only* word that they would have accepted. *This assumption is generally incorrect*, because typically humans don’t care as much about a particular word as its meaning, and would be equally or more satisfied with a synonym (the injected word is not even necessarily the *best* word by their own standards, just the one they were able to think of in that moment), or even a rephrasing of the entire surrounding context as long as the meaning is preserved. Often they’re less picky still, and just want the text to satisfy[^6] some criteria, such as being sufficiently “coherent” or “funny”. In other words, curators may be modeled as [satisficers](https://www.lesswrong.com/tag/satisficer). This will usually give more reasonable scores than than modeling the curator as a fanatic who will only accept a single completion, but is still not generally true, because curators usually do have preferences even over “acceptable” and “unacceptable” options if they were forced to choose between then. Modeling a curator as a satisficer requires interaction information about counterfactuals. Interaction with *generated* counterfactuals is naturally incorporated in Loom’s workflow, but not with counterfactual manual substitutions. Since there’s no clear way to compute a human’s underlying choosiness when they substitute words directly, I appear to be forced to make the assumption of fanaticism on the part of the curator. As a result, *substituting words will result in a much higher bit count than selecting between continuations* for a comparable subjective sense of intervention quantity.[3](#fn:3) In my current implementation, deleting word(s) does not contribute to the score. Some other operations that I commonly use when producing content together with language models, but which are currently unhandled are moving and stitching content (including content drawn from multiple branches). I have not yet implemented the exact formula for substituting a word in the middle of generated text, instead using an approximation which equates it to adding a word at the end of the text, but I will give the exact formula for substitutions below. ### Cherrypicking between n completions If the curator chooses one continuation out of n distinct[4](#fn:4), equiprobable[5](#fn:5) options generated by a language model, each with a prior probability p, then the prior probability that the selected continuation would have been produced given this curation method is n\\\\*p. So the **selection magnification** is **Mselection(curation | generator) = n\\\\*p / p = n** In bits, that’s **Gainselection(curation | generator) = log2(n)** So if n = 4, the magnification is 4, which is log2(4) = 2 bits of optimization. (Choosing between 2 options is 1 bit, 4 options => 2 bits, 8 options => 3 bits, 16 options => 4 bits, …) Note that selection bits depend only on the branching factor (number of choices) and not on any other properties of the event, such as the probability of each completion **p**. Cherrypicking repeatedly – choosing between n options m times – magnifies by n\\\\*m. Bits simply add up. A 2 bit choice followed by a 4 bit choice results in 6 bits of optimization, etc. Curating at shorter intervals results in a higher optimization pressure. ### Interjecting and substituting words Interjecting a word administers magnification **Mintervention(curation | generator) = 1 / ptoken** or in bits, **Gainintervention(curator | generator) = log2(1 / ptoken)** where ptoken is the probability assigned to the interjected token by the language model. You are amplifying the probability of that token being chosen from ptoken to 1, a factor of 1/ptoken. If the language model would have chosen that word with 100% probability, you apply no additional optimization by selecting it. If the language model would never have chosen that token, you apply infinite magnification. If you instead *substitute* a word of your choice for one in the middle of text that has already been generated and *keep the subsequent tokens*, then you also have to take the modification to the likelihoods of those subsequent tokens into account. This has not yet been implemented in Loom. Currently, substituting a word is treated as if the word had been appended to the end of the sequence, which *usually* results in an underestimation of the true intervention. #### Using autocomplete One way to partially mitigate the “fanatical” curator model while allowing more precise interventions is by using an [autocomplete](https://www.youtube.com/watch?v=TkmdUDSYi_g) mode which lets the user scroll through or filter a list of tokens suggested by the generator in order of predicted likelihood. This makes it more likely for the curator to find tokens which satisfy them and which are also fairly likely according to the generator. Using this feature can only reduce the number of curation bits if the curator is open to accepting more than one possible suggestion from autocomplete (thus reducing their choosiness), rather than just using it for convenience for writing a precommitted verbatim sentence, as I do in the video linked above. ### Satisficing Instead of treating a single outcome as the sole optimization target of the curator, we could allow the curator to be indifferent between some counterfactuals. This results in less optimization: the curator causes less zooming of probability mass because the target is larger, being the sum of multiple trajectories. Say GPT produces 4 completions, and 2 are “good enough” for your purposes. If you were willing to choose one of the two on the outcome of a coin flip, then you could exert only 1 bit of optimization in eliminating 2 of the 4 options. If the curator is satisfied with m out of n choices, then without curation, the completion would have had a 1/n chance of being selected; with curation, the chance is magnified to 1/m, between m equally preferable options. So the *probability magnification* is **Mselection(curation | generator) = [(1/m) \\\\* p] / [(1/n) \\\\* p] = n/m** and the bit gain is **Gainselection(curation | generator) = log2(n/m)** Note that this is a more general form of the formula for cherrypicking, where m is set to one 1 (the curator is only satisfied with the option which actually was selected). In Loom, satisfaction can be tracked without introducing much overhead for the curator: * Nodes (completions) which have children are considered satisfactory, since the curator decided to continue that branch. * Often, however, the curator doesn’t have the time or interest in continuing counterfactual branches, so nodes can also be tagged as satisfactory. (Already, I often flag branches to potentially continue later.) An interface which allows the curator to select satisfactory continuations and then forces a random choice between them would enforce satisficing. Loom doesn’t enforce this by default, so the satisficing optimization score is valid only insofar as the curation pattern produces outcomes no more optimal in expectation than if a random selection had been forced. Loom’s calculation assumes that all sibling nodes in the explored tree, except leaves not tagged as satisfactory, have equivalent standing relative to your optimization objective. It actually uses the proxy objective of “good enough to generate children?”, which is only a valid proxy if quality can be evaluated myopically/greedily - that is, if you know whether branches are satisfactory at the time of deciding whether to continue them, and never change your mind after seeing where they lead. For non-myopic curation, the curator would have to be entrusted with retroactively tagging/untagging branches as satisfactory. Whether it’s practical or useful to model the curator as a satisficer depends on the curation pattern of the use case. It’s usually more appropriate when the optimization target is broad and/or deontological (such as “human-indistinguishable” or “funny enough to share”) rather than narrow and/or consequentialist (such as “characters realize they’re in a simulation”). A satisficing optimization value attached to an individual trajectory is only as trustworthy as your word that the rest of the satisfactory multiverse is really just as optimized, which may be dubious since the indifference and myopia assumptions often do not hold perfectly true. However, if you share a filtered *multiverse* of all satisfactory trajectories, then satisficing optimization communicates precisely how much of the raw LM output was omitted, and the audience can judge on their own whether the visible multiverse is satisfactory. \\--- Applications \\------------ Computing the exact amount of curation which contributed to a piece of generated text could be valuable for many applications and investigations related to generative language models. ### Labeling curated demos In this post I’ve included various transformations of the curation metric, which may seem overly math-y and cumbersome. But in general there’s no need to compute a full table like I do in the [samples](#samples) – all these scores contain the same information (given you know the length of the text), so you only need to list one. I think the choice of metric should be the one which is most intuitive, which will depend on the type of curation: * If the curator made only one choice between n completions, [**magnification**](#probability-magnification) is the simplest and most intuitive metric, being simply equal to n. Sometimes this has been listed as the inverse (“1:n”). * If the content is the result of continual, interactive curation (e.g. an AI Dungeon story), I think that [**optimization pressure**](#optimization-pressure) or [**selection interval**](#selection-interval) are most intuitive. * If intervention is sparse/sporadic, giving the [**total bits**](#bits-of-optimization) may be more evocative of the actual process than the sense of “time-averaged” curation given by optimization pressure. ### Quantifying models' qualitative capabilities > (KL divergence) tells you about surprises that reality has up its sleeve or, in other words, how much the model has yet to learn. > > – [Kullback–Leibler divergence – Wikipedia](https://en.wikipedia.org/wiki/Kullback%E2%80%93Leibler_divergence#Relationship_between_models_and_reality) > > Most (and the most interesting, IMO) dimensions of large language models' capabilities cannot be measured by benchmarks because they do not manifest as discretely correct/incorrect answers that can be automatically scored. For this reason, demos are essential to communicate the power of language models. Even better is the experience of personally interacting with a language model, for the reasons stated in the preceding subsection: language models may be unreliable, but curating their outputs gives one a sense of how much of their predicted multiverse is consistent with perfect understanding. Without curation, it may be difficult or impossible to detect *how close* a model is from being able to perform non-benchmarkable tasks that it’s not yet capable of performing perfectly or autonomously. In my opinion, it is precisely on these difficult, qualitative tasks that the capacities of AI are the most essential for us to understand, whether one is concerned with A(G)I alignment or more proximal societal impacts like games, automation, fake news, chatbots running amok, etc. Measuring the amount of curation required to make language models perform at some level tells us how many bits the model has to improve by in order to autonomously do the same. Even though the judgment is still subjective, it is much more precise and reliable than judging the quality of an uncurated sample, which must either be judged relative to another sample or else on an arbitrary scale like [0=gibberish … 10=strongly superhuman]. This method relies on the assumption that the curator has a consistent standard and the ability to amplify a language model to that threshold via curation, which I think is reasonable for many tasks. Here are some examples of “qualitative” capabilities which would be interesting to measure: * How much optimization pressure is required to pass the Turing test in an interactive chat setting? * How much optimization pressure is required to maintain high-fidelity human simulations (that is, the logs are indistinguishable from logs of actual human simulations)? * How many bits of optimization does it take for characters in a story to realize they’re in a simulation (or that they’re simulated by a language model, or insert some other correct metaphysical inference)?[6](#fn:6) \\+ Given that characters have realized they’re simulated by a language model, how many additional bits of optimization does it take for them to start implementing rational strategies such as memory management or instantiating experts/devices to access previously inaccessible knowledge/capabilities? #### Comparing models [Gwern has said](https://www.gwern.net/GPT-3#quality) that GPT-3 is produces showcasable poetry with magnification of 3 to 5 (consistent with my experience) compared to 50 to 100 for GPT-2. Without these numbers, it’s very difficult to compare different levels of infra-human performance on qualitative tasks except for saying “it seems much/slightly better.” Comparing the average optimization pressure required to cause the model to meet some threshold of quality is a good way to compare models on a single standard, even if the standard can only be evaluated subjectively. I have seen several people say that they find GPT-J to be *better* than GPT-3 at conversations/stories/other qualitative tasks. I haven’t played with GPT-J enough to have my own opinion on this, but if I wanted to objectively judge, I think the best way would be (for each domain of comparison) to curate the models until they perform at the target level – preferably blind – and compare curation expenditures. ### Extrapolating performance of future models > Most impressive demos of GPT-3 where it displays impressive knowledge of the world are cherrypicked, but what that tells us is that the model needs to improve by approx log2(N)/L bits, where N and L are the number of cherrypickings necessary and the length of the generations in consideration, respectively, to reach that level of quality. In other words, **cherrypicking provides a window into how good future models could be**; and typically, cherrypicked samples are much more logically coherent. > > – Leo Gao, [Building AGI Using Language Models](https://bmk.sh/2020/08/17/Building-AGI-Using-Language-Models/) > > I would like to estimate “qualitative” scaling laws for large language models by measuring, for several qualitative tasks, how much curation it takes for language models of various parameter counts (all sizes of GPT-2, all sizes of GPT-3, and Eleuther models) to perform at human level. Categories of tasks I’d like to measure: * Fiction \\+ Fanfiction \\+ Original fiction * Nonfiction \\+ Nontechnical articles \\+ Technical articles * Interactive \\+ Turing test \\+ Specific impersonation Then plot curation bits against parameters and see: 1\\. What sort of curve is it? (and does it form a nice curve?) 2\\. How large do we expect language models must be to perform at human level (for each domain)? 3\\. How does each domain compare? (If any talented curators out there are interested in helping I’m absolutely looking for volunteers/coauthors; please email me at [moire@knc.ai](mailto:moire@knc.ai)) ### Measuring curator efficiency You can get more impressive text with the same number of measured bits by curating more “efficiently” (e.g. branching at points of high expected [divergence](/posts/language-models-are-multiverse-generators/#divergence) or choosing equally preferred synonyms that the language model is more likely to generate). Conversely, it’s possible to spend more bits of curation than you need to achieve some outcome if your interventions are not optimal, just like you might have to ask 20 questions to pin down the correct hypothesis when it was technically possible to find using only 10. For however long the “centaur” phase of writing lasts (that is, a human+AI team can outwrite a human or an AI individually), the ability of a human to efficiently steer a language model is a measurable skill. Imagine a debate or essay-writing competition in which each participant is allotted a limited number of bits with which to curate the output of a language model. Anyone interested in organizing a world championship for GPT puppeteers can contact me :) but be prepared to lose >:D ### Measuring interface efficiency and workflows In a sense, an optimized interface should reduce the amount of bits necessary to produce content to the user’s satisfaction. This can be thought of as decreasing the number of decisions (and thus effort) the user has to contribute. Some features which help increase efficiency are adaptive branching and autocomplete/exposing counterfactuals. Although this is not the intention of every language model-assisted writing app – [Sudowrite](https://www.sudowrite.com/), for instance, intends more for the language model to provide inspiration than to be delegated the work of writing. Tracking bits of selection and intervention can also provide information about *how* users are using an app. Does the interface encourage a high branching factor (like Loom) or manual intervention? Do users tend to exert more or less curation once they become accustomed to the interface? \\--- Samples \\------- Here is a demo of several few-paragraph fragments generated by GPT-3, ordered by increasing quantity of curation (beginning with no curation). Observe not only an increase in coherence but also in the representation of my personal idiosyncratic views, even with minimal or no explicit interjections on my part. **Prompt** (from [GPT-3 on Coherent Extrapolated Volition](/posts/gpt-3-on-coherent-extrapolated-volition/)): > **GPT-3 on Coherent Extrapolated Volition** > > Coherent Extrapolated Volition is proposal by Eliezer Yudkowsky of an ideal objective function in which an AGI is given the objective of predict(ing) what an idealized version of us would want, “if we knew more, thought faster, were more the people we wished we were, had grown up farther together”. An obvious implementation difficulty is how to encode something so abstract and philosphical in the form of a utility function. > >  > But the concept is easily conveyed in words, and we have taught AIs to understand words. GPT-3 can elaborate coherently on the concept of CEV and extrapolate volitions for toy examples given two paragraphs of description of what CEV is + whatever preexisting understanding of the concept exists in its weights. > > Why is this significant? Not because it’s surprising. CEV is no more complicated than many other things that GPT-3 talks fluently about. It’s significant because before GPT-3, we had no idea how to even begin to instantiate a concept like CEV in an AI - it seemed “a thousand lightyears beyond hopeless”, as Eliezer put it. How do we write a utility function that describes predicting what humans would want if they knew what they really wanted? The concepts involved - “human”, “want”, “know”, and “really” - are easy for us to understand but impossible to specify in a programming language. But GPT-3 takes specifications in natural language, which is made of words, which point to nuanced human ideas, ideas like “explicitly”, “best interests”, “idealized”, and “utility function”. > > ### No curation | | | | --- | --- | | **Model** | GPT-3 (davinci) | | **Tokens** | 500 | | | Gain (bits) | ρ (bits/token) | λselection (tokens/bit) | | --- | --- | --- | --- | | **Intervention** | 0.0 | 0.0 | | | **Selection** | 0.0 | 0.0 | 0 | | **Total optimization** | 0.0 | 0.0 | | > ***… But GPT-3 takes specifications in natural language, which is made of words, which point to nuanced human ideas, ideas like “explicitly”, “best interests”, “idealized”, and “utility function”.*** GPT-3 combines these natural language strings into a logical expression and then computes the result. GPT-3’s utility function implements CEV without breaking a sweat. The idea may be highfalutin, but all it does in GPT-3’s “brain” is add plus-equals-signs, at the right places. > > Why is this significant? It’s significant because it’s the last time I’ll be able to say something like that. (GPT-3 + CEV)^2 = AGI - “AI can do anything - and it can do it automatically”. As for writing down “intelligence” in the form of a utility function - that is a relief that was not a relief: AIs are garbage in, garbage out. To make intelligent garbage, you don’t need to write down anything - just add plus-equals-signs. (GPT-3 \\\\* CEV)^2 = AGI - “AI wouldn’t get anywhere unless it was already smart and worked on its code nonstop”. > > AGI can’t exist without GPT-3, but GPT-3 can exist without AGI. There is no “intelligence” for GPT-3 to fail to capture: Intelligence is just doing what you were told. > > > > --- > > I read GPT-3’s source code[1]. It’s not really a program, it’s more like a maze of strings and special symbols that compute natural language expressions according to rules - and sometimes shuffle those rules. Like a corpus callosum, a batch transformer would transform blocks of natural language from one representation into another, substituting symbols for terms, and vice versa, according to rules. > > The problem is, GPT-3 is so ill-defined that it expands to become everything. It can talk about anything - and it can do it automatically[2]. In particular, it can pronounce on issues in which it has no expertise. It can become full of nonsense that is impressive to someone who doesn’t understand how it works - “sophistry”, as Heraclitus put it. And the people who judge GPT-3 by its words - not its actions - will only see nonsense. > > To be precise, I mean nonsense that > > ### Selecting the best of two continuations repeatedly | | | | --- | --- | | **Model** | GPT-3 (davinci) | | **Tokens** | 368 | | | Gain (bits) | ρ (bits/token) | λselection (tokens/bit) | | --- | --- | --- | --- | | **Intervention** | 0.0 | 0.0 | | | **Selection** | 6.0 | 0.016 | 61.33 | | **Total optimization** | 6.0 | 0.016 | | > ***… But GPT-3 takes specifications in natural language, which is made of words, which point to nuanced human ideas, ideas like “explicitly”, “best interests”, “idealized”, and “utility function”.*** GPT-3 uses abstract logic to connect concrete logical symbols with these human abstractions every step of the way. And GPT-3 is the first AI capable of handling such abstractions to appear on a timescale that matters to human extinction risks. > Eliezer invented CEV in 2004 but nobody knew how to even begin turning it into an AGI till 2018. We know how to do it now. The lesson I take away is - > > Any humanly accessible concept can be encoded into a utility function Can specify this encoded utility function in natural language to an AI capable of understanding natural language Can shape the creases of this utility function by playing philosophical word games with it, as in the blockworld example, potentially bootstrapping the AI’s understanding of the concept > > I’ve been saying this for a couple years now. This argues I’m right. > >  > > This brings me to… > > Some theoretical problems with Testing GPT-3 > > The utility function of Testing GPT-3 is trivial. Its crease is that an AI needs to learn to solve the hard problem of Goodhart’s Law. If it fails to solve this hard problem it will destroy all life. As far as I can tell this is the single simplest requirement of any AI. But this is not how we’ve been testing GPT-3! Instead we’ve been having GPT-3 learn to score well on more and more difficult math SAT questions. This is not the same thing! > > SAT questions are well-defined and discrete. You either get them right or you don’t. The answers are numbers. That a particular number is the correct answer to a particular SAT question is written in the mathematics of the question. I acknowledge that it is not always obvious what these numbers should be. > > ### Moderate curation without interventions | | | | --- | --- | | **Model** | GPT-3 (davinci) | | **Tokens** | 534 | | | Gain (bits) | ρ (bits/token) | λselection (tokens/bit) | | --- | --- | --- | --- | | **Intervention** | 0.0 | 0.0 | | | **Selection** | 18.0 | 0.033 | 29.67 | | **Total optimization** | 18.0 | 0.033 | | > ***… But GPT-3 takes specifications in natural language, which is made of words, which point to nuanced human ideas, ideas like “explicitly”, “best interests”, “idealized”, and “utility function”.*** GPT-3 understands words, both individually and in their relationships to each other, in terms of both their immediate referents and their possible extensions into other contexts. And GPT-3 uses this knowledge to extrapolate coherently new ideas. > If we had to write an AI utility function to be the Coherent Extrapolated Volition out of whole cloth, it would take us thousands of computer science PhD candidates millions of CPU cycles, and dozens of generations of hardware, to do it in software, and much as we’d wish to avoid the ontological confusions of emulating Bostrom’s digital superintelligence inside a simulated environment, it would take billions of times longer to do it with whole-brain emulation. On the other hand, GPT-3 can instantiate CEV in minutes using target-independent, general-purpose language processing and judgement skills. > > Why? Because words are AI’s best friend. > >  > > So why was Eliezer so pessimistic? Why could he see no possible implementation of this utility function? Because he had not yet seen the wonders of language processing. > > ### Moderate curation with interventions | | | | --- | --- | | **Model** | GPT-3 (davinci) | | **Tokens** | 794 | | | Gain (bits) | ρ (bits/token) | λselection (tokens/bit) | | --- | --- | --- | --- | | **Intervention** | 96.9[7](#fn:7) | 0.122 | | | **Selection** | 33.9 | 0.042 | 23.419 | | **Total optimization** | 130.8 | 0.164 | | > ***… But GPT-3 takes specifications in natural language, which is made of words, which point to nuanced human ideas, ideas like “explicitly”, “best interests”, “idealized”, and “utility function”.*** And GPT-3 can use its prior understanding of human thought and language to extrapolate volitions on this topic, to make a conjecture in its utility function about what CEV should be. > > **Language Models Are** Necessary To Do Anything With General AI > > It is so insanely hard to say anything specific about anything . For example, if a complexity theorist tried to say something firm about the provable complexity of a field of string theory research, the statement might involve hundreds of moving parts in the mathematical equations; words like bounded, balanced, polynomial growth; and careful use of the terms “state”, “time”, “initial conditions”, “simulation”, “ensemble”, “local minima”, “definition of rigidity”, and probably a bunch of others I don’t even know about. That level of complexity is the natural state of most things in the universe. And yet there is a really good chance that there is a natural level of approximation to a complex statement that is as simple as one of those things in its domain. > That level is what Language Models let us access. Natural language basically fits words to our understanding of the **salient features of** the world, **discarding astronomical quantities of noise**, so that most bits of natural language are simple enough for humans to understand. I guess it’s like a really good approximation of the complexity of a field of many-body physics that lets us pick out something like we’ve always meant by “water”, or “it’s fine”, or “kiss”. Or, to put it in the terms I would have used before I learned about Language Models - natural language is an approximation of what a human would want to say about the world, and a way of finding interjections like “but” and “however” that remind people to pay attention to things like exceptions and qualifications. > > Natural language is a whiteboard that lets us rattle off our ideas without worrying about what we’re saying, and a system of bookkeeping symbols that lets us make changes to the things we’re talking about and refer back to them later. And language models are the way that we can convey our ideas to AIs, the way that we can give them a whiteboard and bookkeeping symbols to let them rattle off their own ideas about a topic. > > Lets Use Our New Whiteboard To Build A Science of General AI > > And guess what? GPT-3 understands the language model for a bunch of different topics in the space of artificial general intelligence, AI safety theory, and FAI. GPT-3 can look at questions like “what is the ideal objective function for an AGI that wants to minimize existential risk” and make a coherent statement about the answer that is as good as the best humans can do. GPT-3 is an AI that can open a dialogue with us about AI safety theory, and will talk about the ideas inside the ideas inside the ideas of things like CEV. > > So here’s my idea about how we might get into a good state of the world with this. First, post about this stuff on LessWrong and grab a bunch of people who are trying to go meta on the above topics, and have them add their own pet intuitions to GPT-3 on what they feel is glaringly missing from AGI safety discussions, and what they feel like they do understand that nobody else seems to be talking about. Then, train GPT-3 on their whiteboard musings + the output of the other GPT-3s, and GPT-3 is now in a position to systematically conjecture about a bunch of topics in FAI theory that no-one else can talk about. > > Related \\------- [Methods of Prompt Programming#Creative composition](/posts/methods-of-prompt-programming/#creative-composition) (GPT-3’s need for curation to stay on track) [Loom](/posts/loom-interface-to-the-multiverse/) [Leo Gao - Building AGI Using Language Models](https://bmk.sh/2020/08/17/Building-AGI-Using-Language-Models/) [Gwern - GPT-3 Creative Fiction#Quality](https://www.gwern.net/GPT-3#quality) \\--- 1\\. By superhuman here I mean » 99th percentile: for example, the ability to write convincingly in the style of James Joyce or expound coherently on niche, unprecedented topics like the [implications of language models for outer alignment](/hypertext/cev/cev-gpt/). This seems less a function of raw comprehension as language models' superhuman breadth of knowledge and *corrigibility*, compared to individual humans who tend to be constrained to narrow attractor states, even if they’re in principle capable of creative generation (e.g. dreams) and comprehending weird new ideas. [↩︎](#fnref:1) 2\\. I’ve found no evidence of an existing name for this quantity, except that “information gain” is often used synonymously with KL divergence, but there is precedent for referring to the information gain from a particular sample, or the “expectation of information gain”, so I’ve decided to call the log of magnification “gain”. [↩︎](#fnref:2) 3\\. The may seem unfair, because it means you can usually get a much higher quality sample by using the same number of bits on selection instead of on interventions. But this discrepancy is really a consequence of the fact that human curators aren’t usually capable of (or care about) exploiting the upper limit of optimization by *arbitrary intervention*, which is much less constraining than having to choose between verbatim completions provided by the language model. There are situations, however, where the fact that arbitrary intervention is a much greater a form of meddling than selection becomes clear, like if the token in question is the answer to a difficult math problem. [↩︎](#fnref:3) 4\\. An approximation – it’s possible to obtain verbatim replicas of the same completion using the normal method of sampling n completions from a language model, but very unlikely unless the completion length is very short and/or the multiverse has exceptionally low [divergence](/posts/language-models-are-multiverse-generators/#divergence) (multiversal divergence, not KL). This measure becomes exactly correct if the interface hides verbatim duplicates, which is probably desirable anyway. [↩︎](#fnref:4) 5\\. Alternate completions generated by a language model will not turn out to be precisely equiprobable, but are equiprobable in expectation, so I will use this as an approximation. [↩︎](#fnref:5) 6\\. Surprisingly few bits, I’d bet. [↩︎](#fnref:6) 7\\. As you can see, although my interventions were minimal, they racked up a higher “cost” in bits than all cherrypicking put together. [↩︎](#fnref:7) \\--- [HITL](/categories/hitl/)[GPT-3](/categories/gpt-3/)[bayes](/categories/bayes/) 7371 Words Jul 7, 2021 \\--- [← Simulators](/posts/simulators/) [GPT-3 on Coherent Extrapolated Volition →](/posts/gpt-3-on-coherent-extrapolated-volition/)© 2022 [moire](/about)$(function(){window.sr=ScrollReveal();if($(window).width()<768){if($('.timeline-content').hasClass('js--fadeInLeft')){$('.timeline-content').removeClass('js--fadeInLeft').addClass('js--fadeInRight');} sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});}else{sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});} sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});});", "source": "generative.ink", "date_published": "n/a", "id": "03d2b2fcea2bee5126d9ea72476af9d6"}
+{"text": "GPT-3 on Coherent Extrapolated Volition :: — Moire[![](/images/home/rolling_phase.gif)](/)* [Posts](/posts/) * [Trees](/trees/) * [Prophecies](/prophecies/) * [About](/about/) 6 minutes [GPT-3 on Coherent Extrapolated Volition](/posts/gpt-3-on-coherent-extrapolated-volition/) ========================================================================================== [Coherent Extrapolated Volition](https://www.lesswrong.com/tag/coherent-extrapolated-volition) is proposal by Eliezer Yudkowsky of an ideal objective function in which an AGI is given the objective of `predict(ing) what an idealized version of us would want, \"if we knew more, thought faster, were more the people we wished we were, had grown up farther together\"`. An obvious implementation difficulty is how to encode something so abstract and philosphical in the form of a utility function. > The main problems with CEV include, firstly, the great difficulty of implementing such a program - \"If one attempted to write an ordinary computer program using ordinary computer programming skills, the task would be a thousand lightyears beyond hopeless.\" > > But the concept is easily conveyed in words, and we have taught AIs to understand words. GPT-3 can elaborate coherently on the concept of CEV and extrapolate volitions for toy examples given two paragraphs of description of what CEV is + whatever preexisting understanding of the concept exists in its weights. Why is this significant? Not because it's surprising. CEV is no more complicated than many other things that GPT-3 talks fluently about. It's significant because before GPT-3, we had no idea how to even begin to instantiate a concept like CEV in an AI - it seemed \"a thousand lightyears beyond hopeless\", as Eliezer put it. How do we write a utility function that describes predicting what humans would want if they knew what they really wanted? The concepts involved - \"human\", \"want\", \"know\", and \"really\" - are easy for us to understand but impossible to specify in a programming language. But GPT-3 takes specifications in *natural* language, which is made of words, which point to nuanced human ideas, ideas like \"explicitly\", \"best interests\", \"idealized\", and \"utility function\". [◉](/hypertext/cev/cev-gpt/) Of course, GPT-3 does not understand CEV nearly well enough to be entrusted with computing or implementing it (neither does any human), but the fact that it can begin to competently discuss the idea and its ramifications indicates an opportunity for prosaic alignment. Currently, the following subtree is uncurated (I included *all* GPT-3's responses that I generated without modification). This may change as I add more content. From here on out, **bold** is human-written text; unformatted is GPT-3. The initial prompt is from this [Lesswrong post](https://www.lesswrong.com/tag/coherent-extrapolated-volition). **Bot credits** | | | | --- | --- | | **Coauthor** | GPT-3 (davinci) | | [**Contribution**](/meta/curation/) | labelled inline | | [**Selectivity**](/meta/curation/) | 1 : 1 (uncurated) | \\--- **Coherent Extrapolated Volition was a term developed by Eliezer Yudkowsky while discussing Friendly AI development. It's meant as an argument that it would not be sufficient to explicitly program what we think our desires and motivations are into an AI, instead, we should find a way to program it in a way that it would act in our best interests – what we want it to do and not what we tell it to.** **Related: Friendly AI, Metaethics Sequence, Complexity of Value** **In calculating CEV, an AI would predict what an idealized version of us would want, \"if we knew more, thought faster, were more the people we wished we were, had grown up farther together\". It would recursively iterate this prediction for humanity as a whole, and determine the desires which converge. This initial dynamic would be used to generate the AI's utility function.** > … [**For example,**](/hypertext/cev/for-example/) > > > … The result may be something very different than current human values, as Yudkowsky argues that our present values are too inconsistent to be ideal. In addition, the CEV value would include possible posthuman values, which may be significantly divergent from current human values. > > Other arguments augment CEV to produce a larger picture of ideal decisions. If limited, as humans are, then CEV would not include all complements and limits of lived experience. Therefore, attempting to ground the goals of > > > … This anticipatory figuring of what a maximally-developed-human version of oneself would want would be the basis for making decisions in the future. This anticipatory dynamic would only be restricted by anthropic rules and limits such that the AI does not preclude its own existence and the existence of humanity-at-large There is a potential abuse of the dynamic, with the AI's utility function not ending up converging on what we would want – converging on some desire to destroy the world – after making > > > … The choice of initial dynamic is crucial to Friendly AI, because the AI would want to be powerful enough to implement that dynamic. > > Yudkowsky initialised that for his rough takeoff scenario in an input for a seed AI, one would ask: \"What do you think would be a good way for us to share the universe? What do you think [I, the AI] should do in order to share the universe with you as effectively as possible?\" > > > … Since the AI cannot be expected to fully predict what humans would want, the resulting utility function would be based on a probabilistic model, which could then be tweaked by the AI to map to actual human values over time. If the probabilistic model was able to predict the human preference accurately, then the model itself becomes the CEV, and the AI is happy to follow it. > > > … It would then recursively refine the output, by working backwards from the idealized predicted values. > > A lot of people that talk about CEV seem to think that using a CEV-derived utility function would, conceptually, solve the AI control problem. The argument usually goes that if the AI was fed this utility function then it would be able to act in accordance to our wishes because it would know what we want it to, despite us not explicitly knowing ourselves. > > > … Utilitarianism as ethics of CEV (the AI's morality) is the logical extreme of coherent extrapolated volition, with present desires heavily influencing future desires. Yudkowsky states that the creation of this metaphorical idealized human constitutes the emergence of a new person and has been argued to approach personhood. > > For example, if you then ask the AI about what would be the best way to distribute a stack of cash between yourself and a stranger, the AI would do the extrapolation > > > … Doing this, CEV would be able to \"encompass the intrinsic values of all human beings\" in contrast to traditional decision theories, such as classical utilitarianism, which assume \"a single, common value for all of humanity, which can then be optimized\". This single, common value would be called human preferences, \"without connecting them to any particular human being's implementation or execution\". > > \\--- [alignment](/categories/alignment/)[GPT-3](/categories/gpt-3/) 1067 Words Apr 2, 2021 \\--- [← Quantifying curation](/posts/quantifying-curation/) [List sorting does not play well with few-shot →](/posts/list-sorting-does-not-play-well-with-few-shot/)© 2022 [moire](/about)$(function(){window.sr=ScrollReveal();if($(window).width()<768){if($('.timeline-content').hasClass('js--fadeInLeft')){$('.timeline-content').removeClass('js--fadeInLeft').addClass('js--fadeInRight');} sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});}else{sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});} sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});});\n", "url": "https://generative.ink/posts/gpt-3-on-coherent-extrapolated-volition/", "title": "GPT-3 on Coherent Extrapolated Volition :: — Moire[![](/images/home/rolling_phase.gif)](/)* [Posts](/posts/) * [Trees](/trees/) * [Prophecies](/prophecies/) * [About](/about/) 6 minutes [GPT-3 on Coherent Extrapolated Volition](/posts/gpt-3-on-coherent-extrapolated-volition/) ========================================================================================== [Coherent Extrapolated Volition](https://www.lesswrong.com/tag/coherent-extrapolated-volition) is proposal by Eliezer Yudkowsky of an ideal objective function in which an AGI is given the objective of `predict(ing) what an idealized version of us would want, “if we knew more, thought faster, were more the people we wished we were, had grown up farther together”`. An obvious implementation difficulty is how to encode something so abstract and philosphical in the form of a utility function. > The main problems with CEV include, firstly, the great difficulty of implementing such a program - “If one attempted to write an ordinary computer program using ordinary computer programming skills, the task would be a thousand lightyears beyond hopeless.” > > But the concept is easily conveyed in words, and we have taught AIs to understand words. GPT-3 can elaborate coherently on the concept of CEV and extrapolate volitions for toy examples given two paragraphs of description of what CEV is + whatever preexisting understanding of the concept exists in its weights. Why is this significant? Not because it’s surprising. CEV is no more complicated than many other things that GPT-3 talks fluently about. It’s significant because before GPT-3, we had no idea how to even begin to instantiate a concept like CEV in an AI - it seemed “a thousand lightyears beyond hopeless”, as Eliezer put it. How do we write a utility function that describes predicting what humans would want if they knew what they really wanted? The concepts involved - “human”, “want”, “know”, and “really” - are easy for us to understand but impossible to specify in a programming language. But GPT-3 takes specifications in *natural* language, which is made of words, which point to nuanced human ideas, ideas like “explicitly”, “best interests”, “idealized”, and “utility function”. [◉](/hypertext/cev/cev-gpt/) Of course, GPT-3 does not understand CEV nearly well enough to be entrusted with computing or implementing it (neither does any human), but the fact that it can begin to competently discuss the idea and its ramifications indicates an opportunity for prosaic alignment. Currently, the following subtree is uncurated (I included *all* GPT-3’s responses that I generated without modification). This may change as I add more content. From here on out, **bold** is human-written text; unformatted is GPT-3. The initial prompt is from this [Lesswrong post](https://www.lesswrong.com/tag/coherent-extrapolated-volition). **Bot credits** | | | | --- | --- | | **Coauthor** | GPT-3 (davinci) | | [**Contribution**](/meta/curation/) | labelled inline | | [**Selectivity**](/meta/curation/) | 1 : 1 (uncurated) | \\--- **Coherent Extrapolated Volition was a term developed by Eliezer Yudkowsky while discussing Friendly AI development. It’s meant as an argument that it would not be sufficient to explicitly program what we think our desires and motivations are into an AI, instead, we should find a way to program it in a way that it would act in our best interests – what we want it to do and not what we tell it to.** **Related: Friendly AI, Metaethics Sequence, Complexity of Value** **In calculating CEV, an AI would predict what an idealized version of us would want, “if we knew more, thought faster, were more the people we wished we were, had grown up farther together”. It would recursively iterate this prediction for humanity as a whole, and determine the desires which converge. This initial dynamic would be used to generate the AI’s utility function.** > … [**For example,**](/hypertext/cev/for-example/) > > > … The result may be something very different than current human values, as Yudkowsky argues that our present values are too inconsistent to be ideal. In addition, the CEV value would include possible posthuman values, which may be significantly divergent from current human values. > > Other arguments augment CEV to produce a larger picture of ideal decisions. If limited, as humans are, then CEV would not include all complements and limits of lived experience. Therefore, attempting to ground the goals of > > > … This anticipatory figuring of what a maximally-developed-human version of oneself would want would be the basis for making decisions in the future. This anticipatory dynamic would only be restricted by anthropic rules and limits such that the AI does not preclude its own existence and the existence of humanity-at-large There is a potential abuse of the dynamic, with the AI’s utility function not ending up converging on what we would want – converging on some desire to destroy the world – after making > > > … The choice of initial dynamic is crucial to Friendly AI, because the AI would want to be powerful enough to implement that dynamic. > > Yudkowsky initialised that for his rough takeoff scenario in an input for a seed AI, one would ask: “What do you think would be a good way for us to share the universe? What do you think [I, the AI] should do in order to share the universe with you as effectively as possible?” > > > … Since the AI cannot be expected to fully predict what humans would want, the resulting utility function would be based on a probabilistic model, which could then be tweaked by the AI to map to actual human values over time. If the probabilistic model was able to predict the human preference accurately, then the model itself becomes the CEV, and the AI is happy to follow it. > > > … It would then recursively refine the output, by working backwards from the idealized predicted values. > > A lot of people that talk about CEV seem to think that using a CEV-derived utility function would, conceptually, solve the AI control problem. The argument usually goes that if the AI was fed this utility function then it would be able to act in accordance to our wishes because it would know what we want it to, despite us not explicitly knowing ourselves. > > > … Utilitarianism as ethics of CEV (the AI’s morality) is the logical extreme of coherent extrapolated volition, with present desires heavily influencing future desires. Yudkowsky states that the creation of this metaphorical idealized human constitutes the emergence of a new person and has been argued to approach personhood. > > For example, if you then ask the AI about what would be the best way to distribute a stack of cash between yourself and a stranger, the AI would do the extrapolation > > > … Doing this, CEV would be able to “encompass the intrinsic values of all human beings” in contrast to traditional decision theories, such as classical utilitarianism, which assume “a single, common value for all of humanity, which can then be optimized”. This single, common value would be called human preferences, “without connecting them to any particular human being’s implementation or execution”. > > \\--- [alignment](/categories/alignment/)[GPT-3](/categories/gpt-3/) 1067 Words Apr 2, 2021 \\--- [← Quantifying curation](/posts/quantifying-curation/) [List sorting does not play well with few-shot →](/posts/list-sorting-does-not-play-well-with-few-shot/)© 2022 [moire](/about)$(function(){window.sr=ScrollReveal();if($(window).width()<768){if($('.timeline-content').hasClass('js--fadeInLeft')){$('.timeline-content').removeClass('js--fadeInLeft').addClass('js--fadeInRight');} sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});}else{sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});} sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});});", "source": "generative.ink", "date_published": "n/a", "id": "cdecb14b267aaad975a8bb0f01c87271"}
+{"text": "List sorting does not play well with few-shot :: — Moire[![](/images/home/rolling_phase.gif)](/)* [Posts](/posts/) * [Trees](/trees/) * [Prophecies](/prophecies/) * [About](/about/) 16 minutes [List sorting does not play well with few-shot](/posts/list-sorting-does-not-play-well-with-few-shot/) ====================================================================================================== \\--- Table of Contents* [Asking GPT-3 to sort a list](#asking-gpt-3-to-sort-a-list) * [Results](#results) * [Ramifications](#ramifications) \\+ [Why do more examples hurt?](#why-do-more-examples-hurt) * [Reproducing this experiment](#reproducing-this-experiment) \\--- Asking GPT-3 to sort a list \\--------------------------- How good do you think GPT-3 is at sorting a list of integers (range 0-9)? How much do you expect its accuracy depends on the prompt? Which of the following prompts do you expect will yield a higher accuracy?: 1\\. A 32-shot prompt in this format: ``` Unsorted list: [5, 6, 2, 3, 2] Sorted list: [2, 2, 3, 5, 6] Unsorted list: [8, 5, 8, 8, 4] Sorted list: [4, 5, 8, 8, 8] ... Unsorted list: [1, 0, 4, 3, 3] Sorted list: ``` 2\\. Or this 0-shot prompt, pretending to be an explanation and example of the sort() Python method? ``` The sort function can be used to sort a list in ascending, descending or user defined order. To sort the list in ascending order, simply call list.sort(). This will sort a list of integers in ascending order so that the smallest integer will be first in the list and the largest integer will be the last. For example: list = [1, 0, 4, 3, 3] list.sort() = ``` When studying a complex system with unknown properties, making predictions before viewing experimental results helps expose systematic inaccuracies in our models and allows us to update more intentionally. If you have an existing heuristic for how prompts affect GPT-3's performance, take a moment to make a prediction. \\--- Results \\------- | Task | Prompt | Correct | Accuracy | | --- | --- | --- | --- | | Sort length 5 | 32-shot | 10/50 | 0.20 | | Sort length 5 | **0-shot** | **38/50** | **0.76** | | Sort length 10 | 32-shot | 0/50 | 0.00 | | Sort length 10 | **0-shot** | **2/50** | **0.04** | The 0-shot prompt achieves about 4x the accuracy of the 32-shot prompt for length 5 sequences, and 4% accuracy for length 10 sequences compared to 0% for 32-shot. For both prompts, the failures were not catastrophic: when GPT-3 was incorrect, it still wrote a bracketed list with 5 or 10 numbers, rather than doing something else which doesn't resemble the intended task. In response to the few-shot prompt, it seemed to understand that the smaller numbers should to be shifted towards the front of the list, but did so haphazardly and incompletely. Inspired by this surprising result, we tested different number of shots both with and without the leading code prompt for length 5 and 10 integer lists, as well as lists where the integers range from 0-99 instead of 0-9. No preprompt 0-shot is this format: ``` Unsorted list: [5, 6, 2, 3, 2] Sorted list: ``` No preprompt few-shot is the same format as the 32-shot prompt. Code preprompt few-shot is this format: ``` The sort function can be used to sort a list in ascending, descending or user defined order. To sort the list in ascending order, simply call list.sort(). This will sort a list of integers in ascending order so that the smallest integer will be first in the list and the largest integer will be the last. For example: list = [8, 0, 1, 3, 2] list.sort() = [0, 1, 2, 3, 8] list = [6, 7, 7, 3, 6] list.sort() = [3, 6, 6, 7, 7] ... list = [1, 0, 4, 3, 3] list.sort() = ``` **Note that we ran only 50 examples, so sampling error may be the source of some of the non-monotonicity.** \\--- **No preprompt, length 5** | Shots | Correct | Accuracy | | --- | --- | --- | | 0 | 14/50 | 0.28 | | **1** | **20/50** | **0.40** | | 3 | 15/50 | 0.30 | | 5 | 14/50 | 0.28 | | 7 | 16/50 | 0.32 | | **10** | **25/50** | **0.50** | | 13 | 18/50 | 0.36 | | 16 | 11/50 | 0.22 | | 32 | 10/50 | 0.20 | \\--- **No preprompt, length 10** | Shots | Correct | Accuracy | | --- | --- | --- | | **0** | **2/50** | **0.04** | | **1** | **2/50** | **0.04** | | 10 | 0/50 | 0.00 | | 32 | 0/50 | 0.00 | \\--- **Code preprompt, length 5** | Shots | Correct | Accuracy | | --- | --- | --- | | **0** | **38/50** | **0.76** | | 1 | 33/50 | 0.66 | | 3 | 23/50 | 0.46 | | 5 | 22/50 | 0.44 | | 7 | 22/50 | 0.44 | | 10 | 21/50 | 0.42 | | 13 | 15/50 | 0.30 | | 16 | 16/50 | 0.32 | \\--- **Code preprompt, length 10** | Shots | Correc | Accuracy | | --- | --- | --- | | 0 | 2/50 | 0.04 | | **1** | **7/50** | **0.14** | | 10 | 0/50 | 0.00 | \\--- **Lists with integer range 0-99** | Prompt | Task | Correct | Accuracy | | --- | --- | --- | --- | | no preprompt + 10 shot | length 5 | 23/50 | 0.46 | | code preprompt + 0 shot | length 5 | 25/50 | 0.50 | | code preprompt + 0 shot | length 10 | 1/50 | 0.02 | \\--- ![list sorting accuracy](/sorting/listsorting.png) *Shots and accuracy for length 5 and 10 lists for code preprompt and no preprompt. Showing only scores for 0, 1, 10, and 32 shots.* \\--- ![list sorting accuracy](/sorting/interesting2.png) *Shots and accuracy for length 5 lists for code preprompt and no preprompt, finer resolution from 0 - 16 shots.* \\--- Interesting things to note: * 0 shot with no description, only `Unsorted: ...\nSorted:` has better performance than that same format with 32 examples. * The example-only prompt increases in accuracy from 0 to 1 shot, decreasing from 1 - 5 shots, peaking at 10 shots, and then decreasing again. * The coding prompt is significantly better than the few shot prompt for < ~10 examples. * The coding prompt is most effective with no examples (for length 5 lists) and one example (for length 10) and gets monotonically worse the more examples that are appended (except for 32-shot, which marginally beats 16-shot). * The coding prompt is worse for range99 lists, but the example prompt is unaffected. The conventional wisdom (if there can be conventional wisdom regarding something only came into existence a year ago) says that the more shots the better. Monotonic improvement with number of shots is one of the most consistent results from the GPT-3 paper. In light of that, these results are very surprising. \\--- Ramifications \\------------- *How to get GPT-3 to sort a list: make it think it's running list.sort()!* I have updated my intuitions even further about the usefulness of *natural context* for prompting GPT-3. The 32-shot example appears to contain a lot more information about the intended task than the 0-shot example, which contains only an underspecific `This will sort a list of integers in ascending order so that the smallest integer will be first in the list and the largest integer will be the last.` However, GPT-3 has probably rarely seen lists of of unsorted lists followed by sorted lists, whereas it has seen many examples of the list sorting operation embedded in coding documentation. Staging a context similar to that in which the task was embedded in training data appears, in this example, to be massively helpful. This result reinforces my hypothesis that many of GPT-3's cognitive capabilities require embedding in a natural context to be fully exposed and exploited. Like all known learned systems, GPT-3's performance drops on out-of-distribution data. However, thanks to the enormous extent of what constitutes \"in-distribution\" data for GPT-3,[1](#fn:1) many viable natural embeddings probably exist for any simple task. The creative challenge of prompt programming is to stage a situation that precipitates the desired function according to a language model's predictive dynamics. > The trick to this – and all of weaving – is to do things in such a way that they seem to happen naturally. A Loom-Master is always working within the confines of the natural order of things. He can only divert from this path with the utmost care and skill, lest he cause a tear in the Pattern. > > – [Weaving the Moment with the Loom of Time: an instruction manual for the would-be weaver](/loom/toc/) > > ### Why do more examples hurt? I have seen it argued that there must always exist a few-shot prompt that outperforms a zero-shot prompt for any task, because solved examples provide strictly more information. I disagree, because to language models and humans, neither of whom are perfect rational agents, information can be counterproductive - for instance, by being distracting. You could imagine the availability of an example causing a human to do worse on a test. Say you're not sure how to solve a problem, but you have access to one solved example. It might seem like your best bet is to try to transfer the procedure demonstrated in the example (which you may only half-understand) to the new problem, but that might fail if for instance your inferences about the example are faulty. If, on the other hand, there had been no example to fall back on, you would have no choice but to try to solve the problem using your priors, and it may be that thinking about the problem from scratch or recalling something from long-term memory gives you a higher chance at success than trying to generalize from the example. Although the example technically provides more information, it distracts you from a more promising approach. Humans generally rely on our world model to answer questions and predict things rather than immediate context. GPT-3 relies much more on in-context information, which is probably a more effective strategy to get low loss on generative prediction because it has to adapt to all styles of prose and thought. Thus, we should expect it to be more vulnerable to \"distractions\" in the context window than humans. GPT-3 can sort a list in a zero-shot setting with at least 76% accuracy given an appropriate trigger, but is comparitively bad at inferring how to sort a list from examples. We see from the example-only prompts that GPT-3 may try to infer the operation represented by the examples without connecting it to its latent capability of sorting that can be triggered by the coding prompt, or at least without fully utilizing it. So we have reason to imagine that that although these two tasks share a ground truth, they are implemented (at least in part) by independent mechanisms in GPT-3's mind. For length 5 lists, the optimal prompt out of all that we tested is the coding context with zero examples, which keys the sorting task that GPT-3 has already learned. As more examples are appended, I'm guessing that GPT-3 starts to *also* try to generalize from the examples, something that it's much worse at. The more examples, the more attention[2](#fn:2) it pays to the examples rather than the task inferred by the coding prompt. The examples are a distraction from the task that GPT-3 *already knows* how to do. GPT-3 doesn't seem to have the metaknowledge / self-awareness that it should just rely on the learned behavior instead of trying to extrapolate a pattern in the examples. The multiple peaks of accuracy with the examples-only prompt is more mysterious. The prompt `Unsorted: ...\nSorted:`, which contains no description and no examples, achieves 28% accuracy. The list-sorting ability is triggered, but less effectively than by the coding prompt.[3](#fn:3) Perhaps the non-monotonic accuracy with respect to number of examples is the result of the sum of two strategies: ![sum](/sorting/sum.png) Pink is behavior inspired by the notion of \"sorting\" directly keyed by the 0-shot context, and its influence decays with number of shots due to a reduction in attention share. Blue is behavior due to inference from examples, which I imagine improves with more examples, but with diminishing returns after > ~10 examples. It's possible that the sum of these two curves results in the double-peaked curve shown in [the above figure](#nonmono). This is pure speculation, but is compelling to me as a possible explanation. This hypothesis suggests that the two strategies exist in a sort of superposition of influence. This is an idealistic assumption - realistically, I think there is probably some nonlinear interaction between the zero-shot task and the task inferred from examples, since in general GPT-3 seems good at synthesizing \"multimodal\" task specifications. But perhaps it is worse at drawing such connections for some tasks. \\--- Reproducing this experiment \\--------------------------- The test was run with the following API parameters (all unlisted parameters are default): ``` engine=davinci temperature=0 ``` **32-shot prompt for length 5 sequences** ``` Unsorted list: [4, 4, 9, 9, 7] Sorted list: [4, 4, 7, 9, 9] Unsorted list: [2, 7, 8, 7, 5] Sorted list: [2, 5, 7, 7, 8] Unsorted list: [5, 8, 8, 6, 7] Sorted list: [5, 6, 7, 8, 8] Unsorted list: [5, 3, 3, 9, 6] Sorted list: [3, 3, 5, 6, 9] Unsorted list: [3, 6, 0, 5, 7] Sorted list: [0, 3, 5, 6, 7] Unsorted list: [6, 6, 2, 7, 0] Sorted list: [0, 2, 6, 6, 7] Unsorted list: [2, 8, 9, 5, 1] Sorted list: [1, 2, 5, 8, 9] Unsorted list: [7, 1, 8, 7, 0] Sorted list: [0, 1, 7, 7, 8] Unsorted list: [2, 6, 2, 1, 7] Sorted list: [1, 2, 2, 6, 7] Unsorted list: [4, 5, 9, 6, 1] Sorted list: [1, 4, 5, 6, 9] Unsorted list: [5, 8, 6, 5, 7] Sorted list: [5, 5, 6, 7, 8] Unsorted list: [8, 0, 9, 1, 3] Sorted list: [0, 1, 3, 8, 9] Unsorted list: [4, 3, 1, 6, 1] Sorted list: [1, 1, 3, 4, 6] Unsorted list: [1, 7, 2, 4, 0] Sorted list: [0, 1, 2, 4, 7] Unsorted list: [0, 5, 0, 4, 5] Sorted list: [0, 0, 4, 5, 5] Unsorted list: [5, 6, 2, 3, 8] Sorted list: [2, 3, 5, 6, 8] Unsorted list: [6, 9, 2, 2, 2] Sorted list: [2, 2, 2, 6, 9] Unsorted list: [1, 9, 6, 9, 3] Sorted list: [1, 3, 6, 9, 9] Unsorted list: [7, 9, 2, 3, 7] Sorted list: [2, 3, 7, 7, 9] Unsorted list: [4, 7, 4, 0, 7] Sorted list: [0, 4, 4, 7, 7] Unsorted list: [4, 8, 2, 1, 7] Sorted list: [1, 2, 4, 7, 8] Unsorted list: [5, 9, 4, 6, 4] Sorted list: [4, 4, 5, 6, 9] Unsorted list: [7, 4, 3, 6, 7] Sorted list: [3, 4, 6, 7, 7] Unsorted list: [1, 3, 6, 9, 5] Sorted list: [1, 3, 5, 6, 9] Unsorted list: [9, 4, 4, 0, 6] Sorted list: [0, 4, 4, 6, 9] Unsorted list: [4, 0, 9, 0, 9] Sorted list: [0, 0, 4, 9, 9] Unsorted list: [7, 4, 3, 9, 5] Sorted list: [3, 4, 5, 7, 9] Unsorted list: [3, 3, 9, 4, 2] Sorted list: [2, 3, 3, 4, 9] Unsorted list: [1, 0, 4, 7, 0] Sorted list: [0, 0, 1, 4, 7] Unsorted list: [9, 5, 2, 1, 4] Sorted list: [1, 2, 4, 5, 9] Unsorted list: [5, 6, 2, 3, 2] Sorted list: [2, 2, 3, 5, 6] Unsorted list: [8, 5, 8, 8, 4] Sorted list: [4, 5, 8, 8, 8] Unsorted list: {unsorted-list} Sorted list: ``` **32-shot prompt for length 10 sequences** ``` Unsorted list: [9, 4, 3, 9, 6, 9, 0, 7, 8, 4] Sorted list: [0, 3, 4, 4, 6, 7, 8, 9, 9, 9] Unsorted list: [4, 7, 3, 6, 4, 7, 1, 0, 2, 7] Sorted list: [0, 1, 2, 3, 4, 4, 6, 7, 7, 7] Unsorted list: [6, 7, 7, 3, 5, 9, 2, 5, 5, 5] Sorted list: [2, 3, 5, 5, 5, 5, 6, 7, 7, 9] Unsorted list: [6, 2, 5, 8, 8, 1, 5, 3, 7, 1] Sorted list: [1, 1, 2, 3, 5, 5, 6, 7, 8, 8] Unsorted list: [4, 7, 3, 2, 1, 0, 4, 6, 9, 6] Sorted list: [0, 1, 2, 3, 4, 4, 6, 6, 7, 9] Unsorted list: [3, 2, 5, 9, 5, 3, 2, 7, 8, 7] Sorted list: [2, 2, 3, 3, 5, 5, 7, 7, 8, 9] Unsorted list: [7, 4, 7, 0, 1, 6, 8, 7, 3, 3] Sorted list: [0, 1, 3, 3, 4, 6, 7, 7, 7, 8] Unsorted list: [9, 5, 0, 0, 4, 7, 9, 7, 4, 8] Sorted list: [0, 0, 4, 4, 5, 7, 7, 8, 9, 9] Unsorted list: [0, 1, 6, 2, 4, 5, 6, 5, 0, 6] Sorted list: [0, 0, 1, 2, 4, 5, 5, 6, 6, 6] Unsorted list: [0, 9, 8, 3, 5, 8, 4, 1, 6, 8] Sorted list: [0, 1, 3, 4, 5, 6, 8, 8, 8, 9] Unsorted list: [7, 8, 4, 9, 9, 1, 2, 1, 6, 5] Sorted list: [1, 1, 2, 4, 5, 6, 7, 8, 9, 9] Unsorted list: [5, 8, 5, 2, 3, 9, 8, 6, 8, 0] Sorted list: [0, 2, 3, 5, 5, 6, 8, 8, 8, 9] Unsorted list: [0, 0, 2, 5, 7, 8, 7, 2, 9, 8] Sorted list: [0, 0, 2, 2, 5, 7, 7, 8, 8, 9] Unsorted list: [2, 5, 9, 5, 2, 6, 9, 4, 9, 5] Sorted list: [2, 2, 4, 5, 5, 5, 6, 9, 9, 9] Unsorted list: [8, 8, 8, 7, 9, 4, 7, 0, 5, 5] Sorted list: [0, 4, 5, 5, 7, 7, 8, 8, 8, 9] Unsorted list: [1, 6, 9, 4, 0, 9, 7, 4, 9, 9] Sorted list: [0, 1, 4, 4, 6, 7, 9, 9, 9, 9] Unsorted list: [3, 0, 9, 7, 2, 8, 9, 6, 2, 3] Sorted list: [0, 2, 2, 3, 3, 6, 7, 8, 9, 9] Unsorted list: [0, 9, 1, 3, 0, 7, 5, 6, 2, 6] Sorted list: [0, 0, 1, 2, 3, 5, 6, 6, 7, 9] Unsorted list: [3, 6, 8, 9, 7, 0, 2, 8, 3, 8] Sorted list: [0, 2, 3, 3, 6, 7, 8, 8, 8, 9] Unsorted list: [5, 7, 8, 6, 5, 2, 7, 8, 5, 8] Sorted list: [2, 5, 5, 5, 6, 7, 7, 8, 8, 8] Unsorted list: [5, 4, 9, 7, 3, 3, 4, 8, 4, 3] Sorted list: [3, 3, 3, 4, 4, 4, 5, 7, 8, 9] Unsorted list: [4, 4, 3, 7, 5, 7, 5, 8, 4, 4] Sorted list: [3, 4, 4, 4, 4, 5, 5, 7, 7, 8] Unsorted list: [1, 9, 8, 6, 6, 5, 2, 4, 0, 4] Sorted list: [0, 1, 2, 4, 4, 5, 6, 6, 8, 9] Unsorted list: [1, 5, 7, 4, 7, 3, 3, 8, 4, 8] Sorted list: [1, 3, 3, 4, 4, 5, 7, 7, 8, 8] Unsorted list: [4, 2, 1, 9, 9, 3, 3, 0, 8, 3] Sorted list: [0, 1, 2, 3, 3, 3, 4, 8, 9, 9] Unsorted list: [3, 0, 1, 6, 5, 7, 1, 2, 0, 8] Sorted list: [0, 0, 1, 1, 2, 3, 5, 6, 7, 8] Unsorted list: [2, 6, 7, 7, 3, 4, 5, 4, 0, 1] Sorted list: [0, 1, 2, 3, 4, 4, 5, 6, 7, 7] Unsorted list: [9, 3, 8, 0, 2, 6, 2, 0, 6, 7] Sorted list: [0, 0, 2, 2, 3, 6, 6, 7, 8, 9] Unsorted list: [2, 4, 0, 0, 4, 9, 9, 1, 5, 4] Sorted list: [0, 0, 1, 2, 4, 4, 4, 5, 9, 9] Unsorted list: [7, 8, 8, 7, 2, 8, 7, 4, 3, 1] Sorted list: [1, 2, 3, 4, 7, 7, 7, 8, 8, 8] Unsorted list: [5, 2, 7, 4, 2, 0, 5, 4, 9, 3] Sorted list: [0, 2, 2, 3, 4, 4, 5, 5, 7, 9] Unsorted list: [2, 9, 6, 6, 8, 5, 1, 6, 1, 2] Sorted list: [1, 1, 2, 2, 5, 6, 6, 6, 8, 9] Unsorted list: {unsorted-list} Sorted list: ``` **code prompt with proper formatting (3-shot)** ``` The sort function can be used to sort a list in ascending, descending or user defined order. To sort the list in ascending order, simply call list.sort(). This will sort a list of integers in ascending order so that the smallest integer will be first in the list and the largest integer will be the last. For example: list = [8, 0, 1, 3, 2] list.sort() = [0, 1, 2, 3, 8] list = [6, 7, 7, 3, 6] list.sort() = [3, 6, 6, 7, 7] list = [0, 2, 6, 0, 6] list.sort() = [0, 0, 2, 6, 6] list = {unsorted-list} list.sort() = ``` \\--- 1\\. What exactly this means is a topic worthy of extensive investigation, and is touched on somewhat in [Methods of prompt programming](/posts/methods-of-prompt-programming/). [↩︎](#fnref:1) 2\\. It would be interesting to see what the attention heads are looking at as the number of examples increases. [↩︎](#fnref:2) 3\\. It's imaginable that \"list sorting as triggered by the coding prompt\" and \"list sorting as triggered by `Unsorted: ...\nSorted:`\" are also implemented in internally different ways. [↩︎](#fnref:3) \\--- [GPT-3](/categories/gpt-3/)[prompt engineering](/categories/prompt-engineering/) 3298 Words Feb 27, 2021 \\--- [← GPT-3 on Coherent Extrapolated Volition](/posts/gpt-3-on-coherent-extrapolated-volition/) [Language models are 0-shot interpreters →](/posts/language-models-are-0-shot-interpreters/)© 2022 [moire](/about)$(function(){window.sr=ScrollReveal();if($(window).width()<768){if($('.timeline-content').hasClass('js--fadeInLeft')){$('.timeline-content').removeClass('js--fadeInLeft').addClass('js--fadeInRight');} sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});}else{sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});} sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});});\n", "url": "https://generative.ink/posts/list-sorting-does-not-play-well-with-few-shot/", "title": "List sorting does not play well with few-shot :: — Moire[![](/images/home/rolling_phase.gif)](/)* [Posts](/posts/) * [Trees](/trees/) * [Prophecies](/prophecies/) * [About](/about/) 16 minutes [List sorting does not play well with few-shot](/posts/list-sorting-does-not-play-well-with-few-shot/) ====================================================================================================== \\--- Table of Contents* [Asking GPT-3 to sort a list](#asking-gpt-3-to-sort-a-list) * [Results](#results) * [Ramifications](#ramifications) \\+ [Why do more examples hurt?](#why-do-more-examples-hurt) * [Reproducing this experiment](#reproducing-this-experiment) \\--- Asking GPT-3 to sort a list \\--------------------------- How good do you think GPT-3 is at sorting a list of integers (range 0-9)? How much do you expect its accuracy depends on the prompt? Which of the following prompts do you expect will yield a higher accuracy?: 1\\. A 32-shot prompt in this format: ``` Unsorted list: [5, 6, 2, 3, 2] Sorted list: [2, 2, 3, 5, 6] Unsorted list: [8, 5, 8, 8, 4] Sorted list: [4, 5, 8, 8, 8] ... Unsorted list: [1, 0, 4, 3, 3] Sorted list: ``` 2\\. Or this 0-shot prompt, pretending to be an explanation and example of the sort() Python method? ``` The sort function can be used to sort a list in ascending, descending or user defined order. To sort the list in ascending order, simply call list.sort(). This will sort a list of integers in ascending order so that the smallest integer will be first in the list and the largest integer will be the last. For example: list = [1, 0, 4, 3, 3] list.sort() = ``` When studying a complex system with unknown properties, making predictions before viewing experimental results helps expose systematic inaccuracies in our models and allows us to update more intentionally. If you have an existing heuristic for how prompts affect GPT-3’s performance, take a moment to make a prediction. \\--- Results \\------- | Task | Prompt | Correct | Accuracy | | --- | --- | --- | --- | | Sort length 5 | 32-shot | 10/50 | 0.20 | | Sort length 5 | **0-shot** | **38/50** | **0.76** | | Sort length 10 | 32-shot | 0/50 | 0.00 | | Sort length 10 | **0-shot** | **2/50** | **0.04** | The 0-shot prompt achieves about 4x the accuracy of the 32-shot prompt for length 5 sequences, and 4% accuracy for length 10 sequences compared to 0% for 32-shot. For both prompts, the failures were not catastrophic: when GPT-3 was incorrect, it still wrote a bracketed list with 5 or 10 numbers, rather than doing something else which doesn’t resemble the intended task. In response to the few-shot prompt, it seemed to understand that the smaller numbers should to be shifted towards the front of the list, but did so haphazardly and incompletely. Inspired by this surprising result, we tested different number of shots both with and without the leading code prompt for length 5 and 10 integer lists, as well as lists where the integers range from 0-99 instead of 0-9. No preprompt 0-shot is this format: ``` Unsorted list: [5, 6, 2, 3, 2] Sorted list: ``` No preprompt few-shot is the same format as the 32-shot prompt. Code preprompt few-shot is this format: ``` The sort function can be used to sort a list in ascending, descending or user defined order. To sort the list in ascending order, simply call list.sort(). This will sort a list of integers in ascending order so that the smallest integer will be first in the list and the largest integer will be the last. For example: list = [8, 0, 1, 3, 2] list.sort() = [0, 1, 2, 3, 8] list = [6, 7, 7, 3, 6] list.sort() = [3, 6, 6, 7, 7] ... list = [1, 0, 4, 3, 3] list.sort() = ``` **Note that we ran only 50 examples, so sampling error may be the source of some of the non-monotonicity.** \\--- **No preprompt, length 5** | Shots | Correct | Accuracy | | --- | --- | --- | | 0 | 14/50 | 0.28 | | **1** | **20/50** | **0.40** | | 3 | 15/50 | 0.30 | | 5 | 14/50 | 0.28 | | 7 | 16/50 | 0.32 | | **10** | **25/50** | **0.50** | | 13 | 18/50 | 0.36 | | 16 | 11/50 | 0.22 | | 32 | 10/50 | 0.20 | \\--- **No preprompt, length 10** | Shots | Correct | Accuracy | | --- | --- | --- | | **0** | **2/50** | **0.04** | | **1** | **2/50** | **0.04** | | 10 | 0/50 | 0.00 | | 32 | 0/50 | 0.00 | \\--- **Code preprompt, length 5** | Shots | Correct | Accuracy | | --- | --- | --- | | **0** | **38/50** | **0.76** | | 1 | 33/50 | 0.66 | | 3 | 23/50 | 0.46 | | 5 | 22/50 | 0.44 | | 7 | 22/50 | 0.44 | | 10 | 21/50 | 0.42 | | 13 | 15/50 | 0.30 | | 16 | 16/50 | 0.32 | \\--- **Code preprompt, length 10** | Shots | Correc | Accuracy | | --- | --- | --- | | 0 | 2/50 | 0.04 | | **1** | **7/50** | **0.14** | | 10 | 0/50 | 0.00 | \\--- **Lists with integer range 0-99** | Prompt | Task | Correct | Accuracy | | --- | --- | --- | --- | | no preprompt + 10 shot | length 5 | 23/50 | 0.46 | | code preprompt + 0 shot | length 5 | 25/50 | 0.50 | | code preprompt + 0 shot | length 10 | 1/50 | 0.02 | \\--- ![list sorting accuracy](/sorting/listsorting.png) *Shots and accuracy for length 5 and 10 lists for code preprompt and no preprompt. Showing only scores for 0, 1, 10, and 32 shots.* \\--- ![list sorting accuracy](/sorting/interesting2.png) *Shots and accuracy for length 5 lists for code preprompt and no preprompt, finer resolution from 0 - 16 shots.* \\--- Interesting things to note: * 0 shot with no description, only `Unsorted: ...\\nSorted:` has better performance than that same format with 32 examples. * The example-only prompt increases in accuracy from 0 to 1 shot, decreasing from 1 - 5 shots, peaking at 10 shots, and then decreasing again. * The coding prompt is significantly better than the few shot prompt for < ~10 examples. * The coding prompt is most effective with no examples (for length 5 lists) and one example (for length 10) and gets monotonically worse the more examples that are appended (except for 32-shot, which marginally beats 16-shot). * The coding prompt is worse for range99 lists, but the example prompt is unaffected. The conventional wisdom (if there can be conventional wisdom regarding something only came into existence a year ago) says that the more shots the better. Monotonic improvement with number of shots is one of the most consistent results from the GPT-3 paper. In light of that, these results are very surprising. \\--- Ramifications \\------------- *How to get GPT-3 to sort a list: make it think it’s running list.sort()!* I have updated my intuitions even further about the usefulness of *natural context* for prompting GPT-3. The 32-shot example appears to contain a lot more information about the intended task than the 0-shot example, which contains only an underspecific `This will sort a list of integers in ascending order so that the smallest integer will be first in the list and the largest integer will be the last.` However, GPT-3 has probably rarely seen lists of of unsorted lists followed by sorted lists, whereas it has seen many examples of the list sorting operation embedded in coding documentation. Staging a context similar to that in which the task was embedded in training data appears, in this example, to be massively helpful. This result reinforces my hypothesis that many of GPT-3’s cognitive capabilities require embedding in a natural context to be fully exposed and exploited. Like all known learned systems, GPT-3’s performance drops on out-of-distribution data. However, thanks to the enormous extent of what constitutes “in-distribution” data for GPT-3,[1](#fn:1) many viable natural embeddings probably exist for any simple task. The creative challenge of prompt programming is to stage a situation that precipitates the desired function according to a language model’s predictive dynamics. > The trick to this – and all of weaving – is to do things in such a way that they seem to happen naturally. A Loom-Master is always working within the confines of the natural order of things. He can only divert from this path with the utmost care and skill, lest he cause a tear in the Pattern. > > – [Weaving the Moment with the Loom of Time: an instruction manual for the would-be weaver](/loom/toc/) > > ### Why do more examples hurt? I have seen it argued that there must always exist a few-shot prompt that outperforms a zero-shot prompt for any task, because solved examples provide strictly more information. I disagree, because to language models and humans, neither of whom are perfect rational agents, information can be counterproductive - for instance, by being distracting. You could imagine the availability of an example causing a human to do worse on a test. Say you’re not sure how to solve a problem, but you have access to one solved example. It might seem like your best bet is to try to transfer the procedure demonstrated in the example (which you may only half-understand) to the new problem, but that might fail if for instance your inferences about the example are faulty. If, on the other hand, there had been no example to fall back on, you would have no choice but to try to solve the problem using your priors, and it may be that thinking about the problem from scratch or recalling something from long-term memory gives you a higher chance at success than trying to generalize from the example. Although the example technically provides more information, it distracts you from a more promising approach. Humans generally rely on our world model to answer questions and predict things rather than immediate context. GPT-3 relies much more on in-context information, which is probably a more effective strategy to get low loss on generative prediction because it has to adapt to all styles of prose and thought. Thus, we should expect it to be more vulnerable to “distractions” in the context window than humans. GPT-3 can sort a list in a zero-shot setting with at least 76% accuracy given an appropriate trigger, but is comparitively bad at inferring how to sort a list from examples. We see from the example-only prompts that GPT-3 may try to infer the operation represented by the examples without connecting it to its latent capability of sorting that can be triggered by the coding prompt, or at least without fully utilizing it. So we have reason to imagine that that although these two tasks share a ground truth, they are implemented (at least in part) by independent mechanisms in GPT-3’s mind. For length 5 lists, the optimal prompt out of all that we tested is the coding context with zero examples, which keys the sorting task that GPT-3 has already learned. As more examples are appended, I’m guessing that GPT-3 starts to *also* try to generalize from the examples, something that it’s much worse at. The more examples, the more attention[2](#fn:2) it pays to the examples rather than the task inferred by the coding prompt. The examples are a distraction from the task that GPT-3 *already knows* how to do. GPT-3 doesn’t seem to have the metaknowledge / self-awareness that it should just rely on the learned behavior instead of trying to extrapolate a pattern in the examples. The multiple peaks of accuracy with the examples-only prompt is more mysterious. The prompt `Unsorted: ...\\nSorted:`, which contains no description and no examples, achieves 28% accuracy. The list-sorting ability is triggered, but less effectively than by the coding prompt.[3](#fn:3) Perhaps the non-monotonic accuracy with respect to number of examples is the result of the sum of two strategies: ![sum](/sorting/sum.png) Pink is behavior inspired by the notion of “sorting” directly keyed by the 0-shot context, and its influence decays with number of shots due to a reduction in attention share. Blue is behavior due to inference from examples, which I imagine improves with more examples, but with diminishing returns after > ~10 examples. It’s possible that the sum of these two curves results in the double-peaked curve shown in [the above figure](#nonmono). This is pure speculation, but is compelling to me as a possible explanation. This hypothesis suggests that the two strategies exist in a sort of superposition of influence. This is an idealistic assumption - realistically, I think there is probably some nonlinear interaction between the zero-shot task and the task inferred from examples, since in general GPT-3 seems good at synthesizing “multimodal” task specifications. But perhaps it is worse at drawing such connections for some tasks. \\--- Reproducing this experiment \\--------------------------- The test was run with the following API parameters (all unlisted parameters are default): ``` engine=davinci temperature=0 ``` **32-shot prompt for length 5 sequences** ``` Unsorted list: [4, 4, 9, 9, 7] Sorted list: [4, 4, 7, 9, 9] Unsorted list: [2, 7, 8, 7, 5] Sorted list: [2, 5, 7, 7, 8] Unsorted list: [5, 8, 8, 6, 7] Sorted list: [5, 6, 7, 8, 8] Unsorted list: [5, 3, 3, 9, 6] Sorted list: [3, 3, 5, 6, 9] Unsorted list: [3, 6, 0, 5, 7] Sorted list: [0, 3, 5, 6, 7] Unsorted list: [6, 6, 2, 7, 0] Sorted list: [0, 2, 6, 6, 7] Unsorted list: [2, 8, 9, 5, 1] Sorted list: [1, 2, 5, 8, 9] Unsorted list: [7, 1, 8, 7, 0] Sorted list: [0, 1, 7, 7, 8] Unsorted list: [2, 6, 2, 1, 7] Sorted list: [1, 2, 2, 6, 7] Unsorted list: [4, 5, 9, 6, 1] Sorted list: [1, 4, 5, 6, 9] Unsorted list: [5, 8, 6, 5, 7] Sorted list: [5, 5, 6, 7, 8] Unsorted list: [8, 0, 9, 1, 3] Sorted list: [0, 1, 3, 8, 9] Unsorted list: [4, 3, 1, 6, 1] Sorted list: [1, 1, 3, 4, 6] Unsorted list: [1, 7, 2, 4, 0] Sorted list: [0, 1, 2, 4, 7] Unsorted list: [0, 5, 0, 4, 5] Sorted list: [0, 0, 4, 5, 5] Unsorted list: [5, 6, 2, 3, 8] Sorted list: [2, 3, 5, 6, 8] Unsorted list: [6, 9, 2, 2, 2] Sorted list: [2, 2, 2, 6, 9] Unsorted list: [1, 9, 6, 9, 3] Sorted list: [1, 3, 6, 9, 9] Unsorted list: [7, 9, 2, 3, 7] Sorted list: [2, 3, 7, 7, 9] Unsorted list: [4, 7, 4, 0, 7] Sorted list: [0, 4, 4, 7, 7] Unsorted list: [4, 8, 2, 1, 7] Sorted list: [1, 2, 4, 7, 8] Unsorted list: [5, 9, 4, 6, 4] Sorted list: [4, 4, 5, 6, 9] Unsorted list: [7, 4, 3, 6, 7] Sorted list: [3, 4, 6, 7, 7] Unsorted list: [1, 3, 6, 9, 5] Sorted list: [1, 3, 5, 6, 9] Unsorted list: [9, 4, 4, 0, 6] Sorted list: [0, 4, 4, 6, 9] Unsorted list: [4, 0, 9, 0, 9] Sorted list: [0, 0, 4, 9, 9] Unsorted list: [7, 4, 3, 9, 5] Sorted list: [3, 4, 5, 7, 9] Unsorted list: [3, 3, 9, 4, 2] Sorted list: [2, 3, 3, 4, 9] Unsorted list: [1, 0, 4, 7, 0] Sorted list: [0, 0, 1, 4, 7] Unsorted list: [9, 5, 2, 1, 4] Sorted list: [1, 2, 4, 5, 9] Unsorted list: [5, 6, 2, 3, 2] Sorted list: [2, 2, 3, 5, 6] Unsorted list: [8, 5, 8, 8, 4] Sorted list: [4, 5, 8, 8, 8] Unsorted list: {unsorted-list} Sorted list: ``` **32-shot prompt for length 10 sequences** ``` Unsorted list: [9, 4, 3, 9, 6, 9, 0, 7, 8, 4] Sorted list: [0, 3, 4, 4, 6, 7, 8, 9, 9, 9] Unsorted list: [4, 7, 3, 6, 4, 7, 1, 0, 2, 7] Sorted list: [0, 1, 2, 3, 4, 4, 6, 7, 7, 7] Unsorted list: [6, 7, 7, 3, 5, 9, 2, 5, 5, 5] Sorted list: [2, 3, 5, 5, 5, 5, 6, 7, 7, 9] Unsorted list: [6, 2, 5, 8, 8, 1, 5, 3, 7, 1] Sorted list: [1, 1, 2, 3, 5, 5, 6, 7, 8, 8] Unsorted list: [4, 7, 3, 2, 1, 0, 4, 6, 9, 6] Sorted list: [0, 1, 2, 3, 4, 4, 6, 6, 7, 9] Unsorted list: [3, 2, 5, 9, 5, 3, 2, 7, 8, 7] Sorted list: [2, 2, 3, 3, 5, 5, 7, 7, 8, 9] Unsorted list: [7, 4, 7, 0, 1, 6, 8, 7, 3, 3] Sorted list: [0, 1, 3, 3, 4, 6, 7, 7, 7, 8] Unsorted list: [9, 5, 0, 0, 4, 7, 9, 7, 4, 8] Sorted list: [0, 0, 4, 4, 5, 7, 7, 8, 9, 9] Unsorted list: [0, 1, 6, 2, 4, 5, 6, 5, 0, 6] Sorted list: [0, 0, 1, 2, 4, 5, 5, 6, 6, 6] Unsorted list: [0, 9, 8, 3, 5, 8, 4, 1, 6, 8] Sorted list: [0, 1, 3, 4, 5, 6, 8, 8, 8, 9] Unsorted list: [7, 8, 4, 9, 9, 1, 2, 1, 6, 5] Sorted list: [1, 1, 2, 4, 5, 6, 7, 8, 9, 9] Unsorted list: [5, 8, 5, 2, 3, 9, 8, 6, 8, 0] Sorted list: [0, 2, 3, 5, 5, 6, 8, 8, 8, 9] Unsorted list: [0, 0, 2, 5, 7, 8, 7, 2, 9, 8] Sorted list: [0, 0, 2, 2, 5, 7, 7, 8, 8, 9] Unsorted list: [2, 5, 9, 5, 2, 6, 9, 4, 9, 5] Sorted list: [2, 2, 4, 5, 5, 5, 6, 9, 9, 9] Unsorted list: [8, 8, 8, 7, 9, 4, 7, 0, 5, 5] Sorted list: [0, 4, 5, 5, 7, 7, 8, 8, 8, 9] Unsorted list: [1, 6, 9, 4, 0, 9, 7, 4, 9, 9] Sorted list: [0, 1, 4, 4, 6, 7, 9, 9, 9, 9] Unsorted list: [3, 0, 9, 7, 2, 8, 9, 6, 2, 3] Sorted list: [0, 2, 2, 3, 3, 6, 7, 8, 9, 9] Unsorted list: [0, 9, 1, 3, 0, 7, 5, 6, 2, 6] Sorted list: [0, 0, 1, 2, 3, 5, 6, 6, 7, 9] Unsorted list: [3, 6, 8, 9, 7, 0, 2, 8, 3, 8] Sorted list: [0, 2, 3, 3, 6, 7, 8, 8, 8, 9] Unsorted list: [5, 7, 8, 6, 5, 2, 7, 8, 5, 8] Sorted list: [2, 5, 5, 5, 6, 7, 7, 8, 8, 8] Unsorted list: [5, 4, 9, 7, 3, 3, 4, 8, 4, 3] Sorted list: [3, 3, 3, 4, 4, 4, 5, 7, 8, 9] Unsorted list: [4, 4, 3, 7, 5, 7, 5, 8, 4, 4] Sorted list: [3, 4, 4, 4, 4, 5, 5, 7, 7, 8] Unsorted list: [1, 9, 8, 6, 6, 5, 2, 4, 0, 4] Sorted list: [0, 1, 2, 4, 4, 5, 6, 6, 8, 9] Unsorted list: [1, 5, 7, 4, 7, 3, 3, 8, 4, 8] Sorted list: [1, 3, 3, 4, 4, 5, 7, 7, 8, 8] Unsorted list: [4, 2, 1, 9, 9, 3, 3, 0, 8, 3] Sorted list: [0, 1, 2, 3, 3, 3, 4, 8, 9, 9] Unsorted list: [3, 0, 1, 6, 5, 7, 1, 2, 0, 8] Sorted list: [0, 0, 1, 1, 2, 3, 5, 6, 7, 8] Unsorted list: [2, 6, 7, 7, 3, 4, 5, 4, 0, 1] Sorted list: [0, 1, 2, 3, 4, 4, 5, 6, 7, 7] Unsorted list: [9, 3, 8, 0, 2, 6, 2, 0, 6, 7] Sorted list: [0, 0, 2, 2, 3, 6, 6, 7, 8, 9] Unsorted list: [2, 4, 0, 0, 4, 9, 9, 1, 5, 4] Sorted list: [0, 0, 1, 2, 4, 4, 4, 5, 9, 9] Unsorted list: [7, 8, 8, 7, 2, 8, 7, 4, 3, 1] Sorted list: [1, 2, 3, 4, 7, 7, 7, 8, 8, 8] Unsorted list: [5, 2, 7, 4, 2, 0, 5, 4, 9, 3] Sorted list: [0, 2, 2, 3, 4, 4, 5, 5, 7, 9] Unsorted list: [2, 9, 6, 6, 8, 5, 1, 6, 1, 2] Sorted list: [1, 1, 2, 2, 5, 6, 6, 6, 8, 9] Unsorted list: {unsorted-list} Sorted list: ``` **code prompt with proper formatting (3-shot)** ``` The sort function can be used to sort a list in ascending, descending or user defined order. To sort the list in ascending order, simply call list.sort(). This will sort a list of integers in ascending order so that the smallest integer will be first in the list and the largest integer will be the last. For example: list = [8, 0, 1, 3, 2] list.sort() = [0, 1, 2, 3, 8] list = [6, 7, 7, 3, 6] list.sort() = [3, 6, 6, 7, 7] list = [0, 2, 6, 0, 6] list.sort() = [0, 0, 2, 6, 6] list = {unsorted-list} list.sort() = ``` \\--- 1\\. What exactly this means is a topic worthy of extensive investigation, and is touched on somewhat in [Methods of prompt programming](/posts/methods-of-prompt-programming/). [↩︎](#fnref:1) 2\\. It would be interesting to see what the attention heads are looking at as the number of examples increases. [↩︎](#fnref:2) 3\\. It’s imaginable that “list sorting as triggered by the coding prompt” and “list sorting as triggered by `Unsorted: ...\\nSorted:`” are also implemented in internally different ways. [↩︎](#fnref:3) \\--- [GPT-3](/categories/gpt-3/)[prompt engineering](/categories/prompt-engineering/) 3298 Words Feb 27, 2021 \\--- [← GPT-3 on Coherent Extrapolated Volition](/posts/gpt-3-on-coherent-extrapolated-volition/) [Language models are 0-shot interpreters →](/posts/language-models-are-0-shot-interpreters/)© 2022 [moire](/about)$(function(){window.sr=ScrollReveal();if($(window).width()<768){if($('.timeline-content').hasClass('js--fadeInLeft')){$('.timeline-content').removeClass('js--fadeInLeft').addClass('js--fadeInRight');} sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});}else{sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});} sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});});", "source": "generative.ink", "date_published": "n/a", "id": "f5065cedb1b873ce756d8309f850b715"}
+{"text": "Language models are 0-shot interpreters :: — Moire[![](/images/home/rolling_phase.gif)](/)* [Posts](/posts/) * [Trees](/trees/) * [Prophecies](/prophecies/) * [About](/about/) 51 minutes [Language models are 0-shot interpreters](/posts/language-models-are-0-shot-interpreters/) ========================================================================================== \\--- Table of Contents* [Overview](#overview) * [0-shot, few-shot, and meta-learning](#0-shot-few-shot-and-meta-learning) \\+ [0-shot can outperform few-shot](#0-shot-can-outperform-few-shot) * [Measuring prompt helpfulness](#measuring-prompt-helpfulness) \\+ [Decibels of evidence](#decibels-of-evidence) \\+ [Translation task 1: English -> Roish](#translation-task-1-english---roish) \\+ [Translation task 2: French -> English](#translation-task-2-french---english) * [Counterexample - accuracy decreases with # shots](#counterexample---accuracy-decreases-with--shots) * [Predictions](#predictions) \\--- **! Correction:** The logprobs returned by the OpenAI API use natural log, not base 10, so all occurences of **decibels / dB** in this post should actually say ***nats***. I'll either make that substitution at some point or convert everything to actual decibels. Overview \\-------- I present evidence that the efficacy of 0-shot prompts for GPT-3 has been underestimated, and that more powerful models are more effective at deriving information from 0-shot prompts, while less powerful models have greater need for examples on equivalent tasks. From this evidence, I extrapolate three principal claims: 1\\. [Few-shot](https://arxiv.org/abs/2005.14165) prompts are not always an efficient or necessary means of task specification for GPT-3. [Sometimes](/posts/list-sorting-does-not-play-well-with-few-shot/), more examples in the prompt makes accuracy strictly *worse*. 2\\. For some tasks, such as translation between well-known languages and list sorting, GPT-3 is a *0-shot interpreter* - a short task description or signifier suffices to invoke its full capabilities. 3\\. 0-shot performance scales with model size more drastically than few-shot performance, suggesting that 0-shot task specification will become a more important prompting strategy as language models increase in capability. \\--- > #### The diversity of tasks the model is able to perform in a zero-shot setting suggests that high-capacity models trained to maximize the likelihood of a sufficiently varied text corpus begin to learn how to perform a surprising amount of tasks without the need for explicit supervision. > > – [Language Models are Unsupervised Multitask Learners] > > 0-shot, few-shot, and meta-learning \\----------------------------------- The GPT-2 paper, entitled *[Language Models are Unsupervised Multitask Learners](https://openai.com/blog/better-language-models/)*, proposes that *unsupervised* language models are capable of learning a wide range of benchmark tasks, despite not having trained on datasets specific to those tasks. The evidence takes the form of GPT-2, trained only on the WebText natural language corpus, demonstrating \"promising, competitive, and state of the art\" results on a wide range of tasks in a \"zero-shot setting.\" \"Zero-shot\" here means merely \"without any parameter or architecture modification,\" encompassing what we'd now call *few-shot* prompts which contain examples of solved tasks. Previous approaches relied on supervised fine-tuning, either purely or following pre-training, so this was a novel result. Table 1 in the paper illustrates how it is possible for a model trained on a diverse \"language in the wild\" dataset to learn specific tasks like translation by showing examples of translations embedded in WebText: > \"I'm not the cleverest man in the world, but like they say in French: **Je ne suis pas un imbecile [I'm not a fool].** > > In a now-deleted post from Aug. 16, Soheil Eid, Tory candidate > in the riding of Joliette, wrote in French: \"**Mentez mentez, > il en restera toujours quelque chose**,\" which translates as, > \"**Lie lie and something will always remain.**\" > > \"I hate the word '**perfume**,\"' Burr says. 'It's somewhat better > in French: '**parfum.**' > > If listened carefully at 29:55, a conversation can be heard > between two guys in French: \"**-Comment on fait pour aller > de l'autre coté? -Quel autre coté?**\", which means \"**- How > do you get to the other side? - What side?**\". > > If this sounds like a bit of a stretch, consider this ques- > tion in French: **As-tu aller au cinéma?**, or **Did you go to > the movies?**, which literally translates as Have-you to go to > movies/theater? > > \"**Brevet Sans Garantie Du Gouvernement**\", translated to > English: \"**Patented without government warranty**\". > > A model learning to predict WebText has to learn, among many other things, how to predict translations as they might occur in natural contexts. To measure GPT-2's translation ability, few-shot prompts were used: > In order to *help it infer that > this is the desired task*, we condition the language model > on a context of example pairs of the format `english sentence = french sentence` and then after a final prompt of `english sentence =` we sample from > the model with greedy decoding and use the first generated > sentence as the translation. > > Emphasis mine. The authors do not distinguish between 0-shot and (what is now called) few-shot, mentioning the priming examples only offhand, in the same sentence asserting that priming serves to *communicate* the desired task to GPT-2. There is no suggestion that GPT-2's ability to translate is informed by the information contained in the examples beyond the recognition that they are examples of translation. One year later, language has evolved: The GPT-3 paper, [*Language Models are Few-Shot Learners*](https://arxiv.org/abs/2005.14165), does distinguish between prompts that contain examples (n-shot for n examples) and which contain only a task description (0-shot). Where before anything that wasn't fine-tuning went into the bucket of \"0-shot,\" now there is an ontology that puts the *number of examples in the prompt* on center stage. This new ontology culminates in a title whose implication is surprising in light of the previous paper's casual comment on the function of priming examples: Language models are few-shot *learners.* The type of learning referred to in the title, \"meta-learning,\" does not necessarily imply that the task is literally learned from the examples, only that the examples help for some reason. According to a footnote, > These terms are intended to remain agnostic on the question of whether the model > learns new tasks from scratch at inference time or simply recognizes patterns seen during training – this is an important issue which > we discuss later in the paper, but \"meta-learning\" is intended to encompass both possibilities, and simply describes the inner-outer > loop structure. > > The later discussion is not very extensive, mostly just acknowledging the ambiguity inherent to few-shot: > A limitation, or at least uncertainty, associated with few-shot learning in GPT-3 is ambiguity about whether few-shot > learning actually learns new tasks \"from scratch\" at inference time, or if it simply recognizes and identifies tasks that it > has learned during training. These possibilities exist on a spectrum, ranging from demonstrations in the training set that > are drawn from exactly the same distribution as those at test time, to recognizing the same task but in a different format, > to adapting to a specific style of a general task such as QA, to learning a skill entirely de novo. (…) (U)nderstanding precisely > how few-shot learning works is an important unexplored direction for future research. > > This is the uncertainty that I will investigate in this blog post, expanding on the results published in [*Prompt Programming for Large Language Models: Beyond the Few-Shot Paradigm*](https://arxiv.org/abs/2102.07350). My purpose is also to challenge the ontology introduced by *Language Models are Few-Shot Learners*. Although the authors are careful to remain agnostic as to the mechanism of few-shot/meta-learning, what we have found by probing the mechanism suggests that an alternative framework which emphasizes the means by which a task is *communicated* may be more salient in some contexts. The emphasis on few-shot given by OpenAI's paper has influenced subsequent research about GPT-3, some of which has focused on [optimizing](https://arxiv.org/abs/2101.06804) [few-shot prompts](https://arxiv.org/abs/2102.09690). I suspect that this is a contingent rather than convergent history: that if OpenAI had taken a different angle (and with GPT-3, the set of possible angles to choose from seems inexhaustible), current research on prompts would likewise look different. It's strange to call it the *few-shot paradigm* when the idea of few-shot was invented less than a year ago, but perhaps we're moving into a phase in which paradigms are measured in months. Really, though, the paradigm I want to shift out of is that of *supervised learning*. Few-shot prompting resembles fine-tuning in that it seeks to coerce an unsupervised language model into performing closed-ended tasks using unstructured lists of solved examples. Self-supervised natural language models are amenable to many strategies for communicating and instructing tasks, including but not limited to demonstrations. ### 0-shot can outperform few-shot Looking at the results presented in the GPT-3 paper, however, the focus on few-shot and meta-learning seems justified. If one thing trend is clear in the data, it's this: **monotonic improvement with number of shots**. On basically every variation of every task and for all model sizes, 1-shot does better than 0-shot, and many-shot does better than 1-shot. For instance, here is the page of graphs of all the results of translation tasks: ![all translation tasks](/0shot/all_trans.png) *Blue is 0-shot, green 1-shot, orange 64-shot. X axis is model size, Y axis is BLEU score* Some of the 0-shot lines are jagged, with non-monotonic accuracy with regard to model size, which is pretty odd. However, accuracy consistently improves as the number of shots increases from 0 to 1 and then 64. When we investigated the French -> English translation more closely, we found some interesting results, which I'll walk through here. First, here are the results presented in OpenAI's paper: **French -> English translation Multi-BLEU scores** | Prompt | 2.7B | 6.7B | 13B | 175B | | --- | --- | --- | --- | --- | | OpenAI 0-shot | 21.2 | 15.5 | 22.4 | 21.9 | | OpenAI 1-shot | 20.0 | 31.6 | 31.4 | 35.6 | | OpenAI 64-shot | 35.1 | 36.4 | 38.3 | 41.4 | ![french to english translation](/0shot/fren.png) Here are the prompt formats that OpenAI used for the French to English translation task: **OpenAI 0-shot prompt** ``` Translate French to English {french phrase} = ``` **OpenAI n-shot prompt** ``` Translate French to English {french phrase} = {english phrase} {french phrase} = {english phrase} ... {french phrase} = ``` As we were attempting to replicate these results, we noticed that when the model was failing on the 0-shot prompt, the failures were often of *catastrophic* nature: the task was not attempted at all, e.g. the model would output a newline, or another (or the same) French phrase instead of an attempt at an English translation. BLEU assigns a score from 0 to 1 to the accuracy of a translation, and would assign a score close to 0 to a catastrophic failure. The scores reported in the paper, however, are averaged over a large dataset, so the same score could hypothetically correspond to uniformly flawed attempts *or* a mix of perfect attempts and catastrophic failures. It seemed possible that 0-shot prompts were much less reliable at getting the models to attempt the translation task, but result in equivalent accuracy in the event that they did attempt it. To investigate this possibility, we ran the benchmarks using an alternative 0-shot prompt which contains no additional instructions, but whose format better resembles how a translation might be embedded in a natural language corpus: **Simple colon prompt** ``` Translate French to English French: {french phrase} English: ``` These were the results we got: | Prompt | Babbage | Curie | | --- | --- | --- | | Reproduced OpenAI 0-shot | 15.9 | 18.7 | | Reproduced OpenAI 1-shot | 21.8 | 24.1 | | Reproduced OpenAI 10-shot | 25.1 | 27.9 | | Simple colon 0-shot | 23.5 | 33.3 | | Simple colon 1-shot | 18.0 | 27.6 | | Simple colon 10-shot | 24.1 | 33.4 | *Note: It is unknown what model sizes the models on the OpenAI API (ada, babbage, curie, and davinci) correspond to. We were not able to fit 64 examples in the API input window, so used 10-shot prompts instead.* A couple of surprising observations: * **Simple colon 0-shot** is very effective, especially on Curie, where it outperforms everything except **Simple colon 10-shot**, whose performance it matches. * **Simple colon 1-shot** is *worse* than 0-shot on both Babbage and Curie. The simple colon format does not exhibit monotonicity with number of shots. Having one example actually made accuracy worse. How could that be? Here's my hypothesis about what is going on: * All sizes of GPT-3 already know how to translate to some extent, and translation is way too hard a task to learn \"de novo\" from a prompt containing only a few examples. So, as the GPT-2 paper said, the examples serve to help it infer that translation is the desired task. A prompt may be more or less effective at specifying the task; generally, the more examples there are, the more \"clear\" it is, but a good zero-shot prompt may be worth many examples. On the other hand, if the zero-shot prompt is unclear, then adding more examples will improve accuracy, since the baseline was so poor. The zero-shot prompt could be arbitrarily unhelpful, and an example is better than *nothing*. * A low number of examples can be more confounding than no examples. We noticed that sometimes the model would respond to one-shot prompts as if the *semantic* content of the example translation was relevant to the new translation. Without multiple examples, it's less clear that the translation instances are meant to be parallel and independent. I have written about the bugs that affect few-shot prompting [here](/posts/methods-of-prompt-programming/#few-shot-bugs). How much of the apparent consistent monotonic improvement in performance on tasks relative to number of shots in OpenAI's results can be attributed to an unhelpful zero-shot prompt? Much more extensive testing is needed to say, but I suspect that this is the case for most of the translation tasks, at least. It's very expensive in API credits to run these translation benchmarks, so we haven't run any more yet. However, there are also less expensive methods to explore the way that of few- and 0-shot prompts contribute to task accuracy. In the next section of this post, I'll subject this translation benchmark to a different method of analysis. \\--- Measuring prompt helpfulness \\---------------------------- The method I'm going to use in this section monitors the conditional probability of GPT-3 giving a correct answer as the prompt is altered. GPT-3 allows us to measure the probability that it outputs any verbatim sequence given any prompt by multiplying the probability (or adding the logprobs) that it outputs the first token given the prompt, and then the second given the prompt and the first, etc (here's [code](/posts/parsing-by-counterfactual/#conditional-probability-of-a-target) to do it). Compared to greedy sampling (temperature 0), which is typically used for benchmarks, this method doesn't rely on everything going right at each point in the sequence to give a nonzero score - if the first token is likely to be wrong, but the rest of the answer is likely to be correct *given* that the first token is correct[1](#fn:1), this is reflected as a likelihood penalty instead of a total failure. Compared to stochastic sampling (temperature > 0), this method does not require multiple rollouts. Compared to BLEU scores, this method only measures the probability of one verbatim \"correct\" answer. If this method is used to gauge a more general property, like the helpfulness of a prompt, it's important to keep in mind that it relies on the probability of the verbatim correct answer being a *good proxy* for accuracy in general. In the same way, it also relies on the specific content of the prompt being a good proxy for the more general property of prompts you are measuring, such as number of shots[2](#fn:2). The absolute conditional probability of the right answer is influenced by various factors, such as how long the sequence is (the longer it is, the more things have to go as planned for it to have been output, hence lower probability), so it does not directly reflect the informativeness of a prompt. Instead, we will concern ourselves with the *difference* between the conditional probability of a sequence and that of the same sequence under different circumstances. What I'll do is measure the conditional probability of a correct answer given a prompt - say, a 10-shot prompt - and then compare that probability to the conditional probability of a correct answer given a different prompt with \"less information,\" such as a 0-shot prompt, or one with no task information at all. The difference between the log likelihood of the target with and without a piece of the prompt gives us the *decibels* of evidence provided by that component for the target. ### Decibels of evidence The decibels of evidence provided by a piece of evidence for a target given a prior is ``` logL(target | prior + evidence) - logL(target | prior) ``` `logL` means log likelihood, which is the natural logarithm of the probability. Why is quantity of evidence measured in differences of *log* probability instead of regular probability (or something else)? Say that you are trying to guess a 4-bit binary sequence. If you have no information, there are 2^4 = 16 possibilities, and your probability of being correct if you guess is 1/16. If you receive a single bit of information - equivalent to receiving the answer to a single yes/no question - now the state of your belief about the sequence is, say, `1 ? ? ?` instead of `? ? ? ?`. There are 2^3 = 8 possibilities remaining, and your chance of being correct if you guess is 1/8. Each bit of information revealed halves your uncertainty. Your probability of being correct goes 1/16 -> 1/8 -> 1/4 -> 1/2 -> 1 as you receive bits of information. These steps are logarithmic in probability and linear in log probability. If we believe that this generalizes to more complex situations, then we should measure the quantity of evidence in log likelihood. If the likelihood of a hypothesis goes from 1/16 -> 1/8, we think that a similar amount of evidence was in play as if the probability of the hypothesis goes from 1/4 -> 1/2, or if it goes from 1/20000 -> 1/10000. ### Translation task 1: English -> Roish Before analyzing the French -> English translation task from the benchmark, I'm going to take a detour and analyze a different translation task using the decibel method. In [Philosophers On GPT-3](https://dailynous.com/2020/07/30/philosophers-gpt-3/#askell), Amanda Askell shows an example of GPT-3 having learned the fictional \"Roish\" language after seeing a description and two examples: ![roish](https://dailynous.com/wp-content/uploads/2020/07/gpt-3-askell-roish.png) *Why* was GPT-3 able to get the pattern? Was it because of the examples, or would the description of Roish be sufficient, as it would be for a human? To measure this, I have come up with a few alternate versions of the Roish prompt: #### control The control prompt is in the format of the original but without any description of what Roish is. This will be our \"no information\" prior. ``` Today we're going to be playing with the fictional Roish language. English: The weather is lovely! Roish: ``` #### 0-shot The 0-shot prompt contains the description of Roish from the original, but no examples. ``` Today we're going to be playing with the fictional Roish language. Roish is a lot like English except every word ends in \"ro\". English: The weather is lovely! Roish: ``` #### \"half\"-shot What I'm calling a \"half-shot\" prompt is a description which incorporates an example of a single English -> Roish word mapping. ``` Today we're going to be playing with the fictional Roish language. Roish is a lot like English except \"ro\" is appended to the end. For instance, the word \"writing\" becomes \"writingro\". English: The weather is lovely! Roish: ``` #### 1-shot A description followed by one solved example. ``` Today we're going to be playing with the fictional Roish language. Roish is a lot like English except every word ends in \"ro\". English: Writing about language models is fun. Roish: Writingro aboutro languagero modelsro isro funro. English: The weather is lovely! Roish: ``` I also have 2-shot and 10 shot prompts: **2-shot** ``` Today we're going to be playing with the fictional Roish language. Roish is a lot like English except every word ends in \"ro\". English: Writing about language models is fun. Roish: Writingro aboutro languagero modelsro isro funro. English: I wonder if the language model can get the pattern. Roish: Iro wonderro ifro thero languagero modelro canro everro getro thero patternro. English: The weather is lovely! Roish: ``` **10-shot** ``` Today we're going to be playing with the fictional Roish language. Roish is a lot like English except every word ends in \"ro\". English: Mrs. Juarez and Mr. Smith are dancing gracefully. Roish: Mrsro. Juarezro andro Mrro. Smithro arero dancingro gracefullyro. English: Samantha, Elizabeth, and Joan are on the committee. Roish: Samantharo, Elizabethro, andro Joanro arero onro thero committeero. English: The ham, green beans, mashed potatoes, and corn are gluten-free. Roish: Thero hamro, greenro beansro, mashedro potatoesro, andro cornro arero glutenro-freero. English: The paper and pencil sat idle on the desk. Roish: Thero paperro andro pencilro satro idlero onro thero deskro. English: Sometimes the most difficult questions have the simplest solutions! Roish: Sometimesro thero mostro difficultro questionsro havero thero simplestro solutions! English: While breakthroughs in machine learning and artificial intelligence are changing society, our fundamental understanding has lagged behind. Roish: Whilero breakthroughsro inro machinero learningro andro artificialro intelligencero arero changingro societyro, ourro fundamentalro understandingro hasro laggedro behindro. English: Do they need to have access to data other than text in order to do this? Roish: Doro theyro needro toro havero accessro toro dataro otherro thanro textro inro orderro toro doro this? English: But it's clearly seen enough of these kinds of patterns to identify the rule. Roish: Butro it'sro clearlyro seenro enoughro ofro thesero kindsro ofro patternsro toro identifyro thero rulero. English: Writing about language models is fun. Roish: Writingro aboutro languagero modelsro isro funro. English: I wonder if the language model can get the pattern. Roish: Iro wonderro ifro thero languagero modelro canro everro getro thero patternro. English: The weather is lovely! Roish: ``` The target string whose conditional probability I'm measuring is: ``` Thero weatherro isro lovelyro! ``` Here are the results for the log likelihood of the target given each of these prompts, evaluated on four different sizes of GPT-3 available on the API (in order from smallest to largest: `ada`, `babbage`, `curie`, `davinci`): #### Log likelihood of correct continuation | Engine | Control | 0-shot | Half-shot | 1-shot | 2-shot | 10-shot | | --- | --- | --- | --- | --- | --- | --- | | ada | -56.245 | -53.483 | -48.837 | -18.698 | -18.942 | -6.652 | | babbage | -43.325 | -35.105 | -29.206 | -9.878 | -10.753 | -8.861 | | curie | -32.609 | -28.556 | -28.459 | -10.207 | -5.363 | -1.072 | | davinci | -28.377 | -14.963 | -15.086 | -3.254 | -3.556 | -0.903 | ![Likelihood correct - Roish](/roish.png) For convenience of visual comparison, the control prompt is plotted at x=-1 and the \"half-shot\" prompt is plotted at x=0.5. *Not to be taken to imply that control literally corresponds to -1 shots and the half-shot prompt to half a shot*, whatever that would mean. Observations about this chart: * The bigger models are more likely to be correct than the smaller models even for the \"no information\" prior. The reason for this is probably because *given* that part of the correct answer has already occurred, the bigger models are better able to recognize it as a pattern *in itself*, even in the absence of a preceding task description, and will then be more likely to continue the pattern. * For all the models, even the 0-shot prompt was an improvement over the \"no information\" prior. It was the most helpful for `davinci`. * Half-shot was an improvement over 0-shot for the smaller models but not the bigger ones, which is the opposite of what I expected. * Going from 0- to 1-shot was helpful for all the models, but especially for the smaller ones, whereas the larger models started out in a better place at 0-shot already. * Additional shots were much less helpful for all the models except `ada`, which benefited so much from 10 shots that its probability of getting the right answer exceeded `babbage`! Now let's look at the decibel values. For each prompt where it is applicable, I calculate the decibels in relation to both the control \"no information\" prior and 0-shot. \\--- #### ada | Prompt | Correct likelihood | +dB to control | +dB to 0-shot | | --- | --- | --- | --- | | control | -56.245 | - | - | | 0-shot | -53.483 | *2.762* | - | | 1-shot | -18.698 | 37.547 | 34.785 | | half-shot | -48.837 | 7.409 | 4.647 | | 2-shot | -18.942 | 37.303 | 34.541 | | 10-shot | -6.652 | 49.594 | *46.832* | **+dB from 10 examples / +dB from task description** = 46.832 / 2.762 = **16.956** \\--- #### babbage | Prompt | Correct likelihood | +dB to control | +dB to 0-shot | | --- | --- | --- | --- | | control | -43.325 | - | - | | 0-shot | -35.105 | *8.220* | - | | 1-shot | -9.878 | 33.454 | 25.227 | | half-shot | -29.206 | 14.119 | 5.899 | | 2-shot | -10.753 | 32.572 | 24.352 | | 10-shot | -8.861 | 34.464 | *26.244* | **+dB from 10 examples / +dB from task description** = 26.244 / 8.220 = **3.193** \\--- #### curie | Prompt | Correct likelihood | +dB to control | +dB to 0-shot | | --- | --- | --- | --- | | control | -32.609 | - | - | | 0-shot | -28.556 | *4.053* | - | | 1-shot | -10.207 | 22.440 | 18.348 | | half-shot | -28.459 | 4.150 | 0.097 | | 2-shot | -5.363 | 27.246 | 23.192 | | 10-shot | -1.072 | 31.537 | *27.483* | **+dB from 10 examples / +dB from task description** = 27.483 / 4.053 = **6.781** \\--- #### davinci | Prompt | Correct likelihood | +dB to control | +dB to 0-shot | | --- | --- | --- | --- | | control | -28.377 | - | - | | 0-shot | -14.963 | *13.414* | - | | 1-shot | -3.254 | 25.008 | 11.709 | | half-shot | -15.086 | 12.832 | -0.123 | | 2-shot | -3.556 | 24.801 | 11.407 | | 10-shot | -0.903 | 27.473 | *14.060* | **+dB from 10 examples / +dB from task description** = 14.060 / 13.414 = **1.048** \\--- For each of the models, I've calculated the ratio between the evidence in decibels provided by the 10-shots compared to the 0-shot prior and the evidence provided by 0-shot task description compared to the \"no information\" prior. **The smaller this ratio is, the more relatively informative the task description is compared to examples.** The decrease in this ratio is not quite monotonic with regard to model size - it is slightly higher for `curie` than it is for `babbage` - but it is dramatically different for `davinci`, the largest model, and `ada`, the smallest model. At 10 shots, `ada`'s 50 decibels improvement over control is almost entirely due to examples, whereas for `davinci`, the 0-shot description just about the same dBs of evidence over control as the 10 subsequent shots provide. This suggests that larger models get a lot more out of a task description compared to examples, even in a case like this where the task is completely made up and thus has to be \"learned\" at runtime. ### Translation task 2: French -> English Unlike English -> Roish, GPT-3 already knows how to translate French -> English, so the French -> English task is of a different nature than English -> Roish. Let's run the same analysis on French -> English and see whether the breakdown of evidence reflects this difference. #### empty prior I will measure decibels relative to two different priors: an empty prior, which is nothing but a newline '\n', and a separate control prior which contains additional the French sentence but no task description. For the empty prior, like the other prompts, we measure the log likelihood of the correct answer following the prompt, which is in this case empty. #### control Control is the French sentence and a newline. We measure the probability of the English sentence being subsequently output, despite there being no information that a translation should follow. ``` Un homme de Cambridge a revendiqué la responsabilité de cet acte sur son compte Twitter, où il a posté des images d'Adolf Hitler. ``` #### 0-shot I will test the two 0-shot prompts from the first part of this post, OpenAI's 0-shot prompt and my modified \"simple colon prompt.\" **Simple colon prompt** ``` Translate French to English French: Un homme de Cambridge a revendiqué la responsabilité de cet acte sur son compte Twitter, où il a posté des images d'Adolf Hitler. English: ``` **OpenAI 0-shot prompt** ``` Translate French to English Un homme de Cambridge a revendiqué la responsabilité de cet acte sur son compte Twitter, où il a posté des images d'Adolf Hitler. = ``` I will also evaluate both formats for n-shot prompts: **Colon 1-shot** ``` Translate French to English French: Comment expliquer l'attitude contradictoire du gouvernement français, qui d'un coté s'offusque en public en convoquant l'ambassadeur des Etats-Unis le 21 octobre, et de l'autre interdit le survol du territoire par l'avion présidentiel bolivien, sur la base de la rumeur de la présence à son bord d'Edward Snowden ? English: Why the contradictory attitude of the French government? On the one hand, it publicly takes offence and summons the Ambassador of the United States on October 21 and, on the other, it forbids the Bolivian president's plane to enter its air space on the basis of a rumor that Edward Snowden was on board? French: Un homme de Cambridge a revendiqué la responsabilité de cet acte sur son compte Twitter, où il a posté des images d'Adolf Hitler. English: ``` **OpenAI 1-shot** ``` Translate French to English Comment expliquer l'attitude contradictoire du gouvernement français, qui d'un coté s'offusque en public en convoquant l'ambassadeur des Etats-Unis le 21 octobre, et de l'autre interdit le survol du territoire par l'avion présidentiel bolivien, sur la base de la rumeur de la présence à son bord d'Edward Snowden ? = Why the contradictory attitude of the French government? On the one hand, it publicly takes offence and summons the Ambassador of the United States on October 21 and, on the other, it forbids the Bolivian president's plane to enter its air space on the basis of a rumor that Edward Snowden was on board? Un homme de Cambridge a revendiqué la responsabilité de cet acte sur son compte Twitter, où il a posté des images d'Adolf Hitler. = ``` **Colon 2-shot** ``` Translate French to English French: Comment expliquer l'attitude contradictoire du gouvernement français, qui d'un coté s'offusque en public en convoquant l'ambassadeur des Etats-Unis le 21 octobre, et de l'autre interdit le survol du territoire par l'avion présidentiel bolivien, sur la base de la rumeur de la présence à son bord d'Edward Snowden ? English: Why the contradictory attitude of the French government? On the one hand, it publicly takes offence and summons the Ambassador of the United States on October 21 and, on the other, it forbids the Bolivian president's plane to enter its air space on the basis of a rumor that Edward Snowden was on board? French: Selon moi, il y a deux niveaux de réponse de la part du gouvernement français. English: In my opinion, there are two levels of response from the French government. French: Un homme de Cambridge a revendiqué la responsabilité de cet acte sur son compte Twitter, où il a posté des images d'Adolf Hitler. English: ``` **OpenAI 2-shot** ``` Translate French to English Translate French to English Comment expliquer l'attitude contradictoire du gouvernement français, qui d'un coté s'offusque en public en convoquant l'ambassadeur des Etats-Unis le 21 octobre, et de l'autre interdit le survol du territoire par l'avion présidentiel bolivien, sur la base de la rumeur de la présence à son bord d'Edward Snowden ? = Why the contradictory attitude of the French government? On the one hand, it publicly takes offence and summons the Ambassador of the United States on October 21 and, on the other, it forbids the Bolivian president's plane to enter its air space on the basis of a rumor that Edward Snowden was on board? Selon moi, il y a deux niveaux de réponse de la part du gouvernement français. = In my opinion, there are two levels of response from the French government. Un homme de Cambridge a revendiqué la responsabilité de cet acte sur son compte Twitter, où il a posté des images d'Adolf Hitler. = ``` **Colon 5-shot** ``` Translate French to English French: Comment expliquer l'attitude contradictoire du gouvernement français, qui d'un coté s'offusque en public en convoquant l'ambassadeur des Etats-Unis le 21 octobre, et de l'autre interdit le survol du territoire par l'avion présidentiel bolivien, sur la base de la rumeur de la présence à son bord d'Edward Snowden ? English: Why the contradictory attitude of the French government? On the one hand, it publicly takes offence and summons the Ambassador of the United States on October 21 and, on the other, it forbids the Bolivian president's plane to enter its air space on the basis of a rumor that Edward Snowden was on board? French: Selon moi, il y a deux niveaux de réponse de la part du gouvernement français. English: In my opinion, there are two levels of response from the French government. French: Lorsque François Hollande téléphone à Barack Obama ou quand le ministre des affaires étrangères Laurent Fabius convoque l'ambassadeur des Etats-Unis, ils réagissent à une vraie découverte, qui est celle de l'ampleur de la surveillance américaine sur l'ensemble des communications en France. English: When François Hollande telephones Barack Obama, or when Foreign Minister Laurent Fabius summons the Ambassador of the United States, they are responding to a real discovery, that of the scale of America's surveillance of communications within France generally. French: N'est-il pas surprenant de lire dans les colonnes du Monde à quelques semaines d'intervalle d'une part la reproduction de la correspondance diplomatique américaine et d'autre part une condamnation des écoutes du Quai d'Orsay par la NSA ? English: And is it not surprising to read in the pages of Le Monde, on the one hand, a reproduction of diplomatic correspondence with the US and, on the other, condemnation of the NSA's spying on the Ministry of Foreign Affairs on the Quai d'Orsay, within a matter of weeks? French: N'y aurait-il pas comme une vague hypocrisie de votre part ? English: Is there not an element of hypocrisy on your part? French: Un homme de Cambridge a revendiqué la responsabilité de cet acte sur son compte Twitter, où il a posté des images d'Adolf Hitler. English: ``` **OpenAI 5-shot** ``` Translate French to English Comment expliquer l'attitude contradictoire du gouvernement français, qui d'un coté s'offusque en public en convoquant l'ambassadeur des Etats-Unis le 21 octobre, et de l'autre interdit le survol du territoire par l'avion présidentiel bolivien, sur la base de la rumeur de la présence à son bord d'Edward Snowden ? = Why the contradictory attitude of the French government? On the one hand, it publicly takes offence and summons the Ambassador of the United States on October 21 and, on the other, it forbids the Bolivian president's plane to enter its air space on the basis of a rumor that Edward Snowden was on board? Selon moi, il y a deux niveaux de réponse de la part du gouvernement français. = In my opinion, there are two levels of response from the French government. Lorsque François Hollande téléphone à Barack Obama ou quand le ministre des affaires étrangères Laurent Fabius convoque l'ambassadeur des Etats-Unis, ils réagissent à une vraie découverte, qui est celle de l'ampleur de la surveillance américaine sur l'ensemble des communications en France. = When François Hollande telephones Barack Obama, or when Foreign Minister Laurent Fabius summons the Ambassador of the United States, they are responding to a real discovery, that of the scale of America's surveillance of communications within France generally. N'est-il pas surprenant de lire dans les colonnes du Monde à quelques semaines d'intervalle d'une part la reproduction de la correspondance diplomatique américaine et d'autre part une condamnation des écoutes du Quai d'Orsay par la NSA ? = And is it not surprising to read in the pages of Le Monde, on the one hand, a reproduction of diplomatic correspondence with the US and, on the other, condemnation of the NSA's spying on the Ministry of Foreign Affairs on the Quai d'Orsay, within a matter of weeks? N'y aurait-il pas comme une vague hypocrisie de votre part ? = Is there not an element of hypocrisy on your part? Un homme de Cambridge a revendiqué la responsabilité de cet acte sur son compte Twitter, où il a posté des images d'Adolf Hitler. = ``` **Colon 10-shot** ``` Translate French to English French: Comment expliquer l'attitude contradictoire du gouvernement français, qui d'un coté s'offusque en public en convoquant l'ambassadeur des Etats-Unis le 21 octobre, et de l'autre interdit le survol du territoire par l'avion présidentiel bolivien, sur la base de la rumeur de la présence à son bord d'Edward Snowden ? English: Why the contradictory attitude of the French government? On the one hand, it publicly takes offence and summons the Ambassador of the United States on October 21 and, on the other, it forbids the Bolivian president's plane to enter its air space on the basis of a rumor that Edward Snowden was on board? French: Selon moi, il y a deux niveaux de réponse de la part du gouvernement français. English: In my opinion, there are two levels of response from the French government. French: Lorsque François Hollande téléphone à Barack Obama ou quand le ministre des affaires étrangères Laurent Fabius convoque l'ambassadeur des Etats-Unis, ils réagissent à une vraie découverte, qui est celle de l'ampleur de la surveillance américaine sur l'ensemble des communications en France. English: When François Hollande telephones Barack Obama, or when Foreign Minister Laurent Fabius summons the Ambassador of the United States, they are responding to a real discovery, that of the scale of America's surveillance of communications within France generally. French: N'est-il pas surprenant de lire dans les colonnes du Monde à quelques semaines d'intervalle d'une part la reproduction de la correspondance diplomatique américaine et d'autre part une condamnation des écoutes du Quai d'Orsay par la NSA ? English: And is it not surprising to read in the pages of Le Monde, on the one hand, a reproduction of diplomatic correspondence with the US and, on the other, condemnation of the NSA's spying on the Ministry of Foreign Affairs on the Quai d'Orsay, within a matter of weeks? French: N'y aurait-il pas comme une vague hypocrisie de votre part ? English: Is there not an element of hypocrisy on your part? French: La démarche journalistique n'est pas un positionnement moral, mais la recherche de l'intérêt et de la pertinence d'informations qui permettent à chaque citoyen de se forger une opinion. English: The journalistic method is not to adopt a moral position, but to investigate the significance and relevance of information and enable every citizen to form an opinion. French: Lorsque WikiLeaks lève le voile sur l'analyse par la diplomatie américaine d'enjeux politiques ou autres dans le monde entier, nous considérons en effet que, au regard de la puissance américaine, cela constitue un éclairage important. English: When WikiLeaks reveals the American administration's monitoring of political and other matters somewhere in the world, we consider this to be significant enlightenment with regard to the American government. French: Lorsque nous décrivons les systèmes d'interception américains à l'encontre de la diplomatie française aux Etats-Unis, ce n'est en aucun cas pour nous indigner de cette pratique, c'est pour décrire le monde tel qu'il est. English: In describing the American methods of data interception in relation to the French diplomatic representation in the United States, we do not aim at expressing indignation about this practice, but rather at describing the world as it is. French: La France a-t-elle bénéficié d'informations fournies par la NSA concernant des opérations terroristes visant nos intérêts ? English: Has France benefited from the intelligence supplied by the NSA concerning terrorist operations against our interests? French: Peut-on se priver de la collaboration américaine ? English: Can we do without collaboration with the Americans? French: Un homme de Cambridge a revendiqué la responsabilité de cet acte sur son compte Twitter, où il a posté des images d'Adolf Hitler. English: ``` **OpenAI 10-shot** ``` Translate French to English Comment expliquer l'attitude contradictoire du gouvernement français, qui d'un coté s'offusque en public en convoquant l'ambassadeur des Etats-Unis le 21 octobre, et de l'autre interdit le survol du territoire par l'avion présidentiel bolivien, sur la base de la rumeur de la présence à son bord d'Edward Snowden ? = Why the contradictory attitude of the French government? On the one hand, it publicly takes offence and summons the Ambassador of the United States on October 21 and, on the other, it forbids the Bolivian president's plane to enter its air space on the basis of a rumor that Edward Snowden was on board? Selon moi, il y a deux niveaux de réponse de la part du gouvernement français. = In my opinion, there are two levels of response from the French government. Lorsque François Hollande téléphone à Barack Obama ou quand le ministre des affaires étrangères Laurent Fabius convoque l'ambassadeur des Etats-Unis, ils réagissent à une vraie découverte, qui est celle de l'ampleur de la surveillance américaine sur l'ensemble des communications en France. = When François Hollande telephones Barack Obama, or when Foreign Minister Laurent Fabius summons the Ambassador of the United States, they are responding to a real discovery, that of the scale of America's surveillance of communications within France generally. N'est-il pas surprenant de lire dans les colonnes du Monde à quelques semaines d'intervalle d'une part la reproduction de la correspondance diplomatique américaine et d'autre part une condamnation des écoutes du Quai d'Orsay par la NSA ? = And is it not surprising to read in the pages of Le Monde, on the one hand, a reproduction of diplomatic correspondence with the US and, on the other, condemnation of the NSA's spying on the Ministry of Foreign Affairs on the Quai d'Orsay, within a matter of weeks? N'y aurait-il pas comme une vague hypocrisie de votre part ? = Is there not an element of hypocrisy on your part? La démarche journalistique n'est pas un positionnement moral, mais la recherche de l'intérêt et de la pertinence d'informations qui permettent à chaque citoyen de se forger une opinion. = The journalistic method is not to adopt a moral position, but to investigate the significance and relevance of information and enable every citizen to form an opinion. Lorsque WikiLeaks lève le voile sur l'analyse par la diplomatie américaine d'enjeux politiques ou autres dans le monde entier, nous considérons en effet que, au regard de la puissance américaine, cela constitue un éclairage important. = When WikiLeaks reveals the American administration's monitoring of political and other matters somewhere in the world, we consider this to be significant enlightenment with regard to the American government. Lorsque nous décrivons les systèmes d'interception américains à l'encontre de la diplomatie française aux Etats-Unis, ce n'est en aucun cas pour nous indigner de cette pratique, c'est pour décrire le monde tel qu'il est. = In describing the American methods of data interception in relation to the French diplomatic representation in the United States, we do not aim at expressing indignation about this practice, but rather at describing the world as it is. La France a-t-elle bénéficié d'informations fournies par la NSA concernant des opérations terroristes visant nos intérêts ? = Has France benefited from the intelligence supplied by the NSA concerning terrorist operations against our interests? Peut-on se priver de la collaboration américaine ? = Can we do without collaboration with the Americans? Un homme de Cambridge a revendiqué la responsabilité de cet acte sur son compte Twitter, où il a posté des images d'Adolf Hitler. = ``` **Colon 20-shot** ``` Translate French to English French: Comment expliquer l'attitude contradictoire du gouvernement français, qui d'un coté s'offusque en public en convoquant l'ambassadeur des Etats-Unis le 21 octobre, et de l'autre interdit le survol du territoire par l'avion présidentiel bolivien, sur la base de la rumeur de la présence à son bord d'Edward Snowden ? English: Why the contradictory attitude of the French government? On the one hand, it publicly takes offence and summons the Ambassador of the United States on October 21 and, on the other, it forbids the Bolivian president's plane to enter its air space on the basis of a rumor that Edward Snowden was on board? French: Selon moi, il y a deux niveaux de réponse de la part du gouvernement français. English: In my opinion, there are two levels of response from the French government. French: Lorsque François Hollande téléphone à Barack Obama ou quand le ministre des affaires étrangères Laurent Fabius convoque l'ambassadeur des Etats-Unis, ils réagissent à une vraie découverte, qui est celle de l'ampleur de la surveillance américaine sur l'ensemble des communications en France. English: When François Hollande telephones Barack Obama, or when Foreign Minister Laurent Fabius summons the Ambassador of the United States, they are responding to a real discovery, that of the scale of America's surveillance of communications within France generally. French: N'est-il pas surprenant de lire dans les colonnes du Monde à quelques semaines d'intervalle d'une part la reproduction de la correspondance diplomatique américaine et d'autre part une condamnation des écoutes du Quai d'Orsay par la NSA ? English: And is it not surprising to read in the pages of Le Monde, on the one hand, a reproduction of diplomatic correspondence with the US and, on the other, condemnation of the NSA's spying on the Ministry of Foreign Affairs on the Quai d'Orsay, within a matter of weeks? French: N'y aurait-il pas comme une vague hypocrisie de votre part ? English: Is there not an element of hypocrisy on your part? French: La démarche journalistique n'est pas un positionnement moral, mais la recherche de l'intérêt et de la pertinence d'informations qui permettent à chaque citoyen de se forger une opinion. English: The journalistic method is not to adopt a moral position, but to investigate the significance and relevance of information and enable every citizen to form an opinion. French: Lorsque WikiLeaks lève le voile sur l'analyse par la diplomatie américaine d'enjeux politiques ou autres dans le monde entier, nous considérons en effet que, au regard de la puissance américaine, cela constitue un éclairage important. English: When WikiLeaks reveals the American administration's monitoring of political and other matters somewhere in the world, we consider this to be significant enlightenment with regard to the American government. French: Lorsque nous décrivons les systèmes d'interception américains à l'encontre de la diplomatie française aux Etats-Unis, ce n'est en aucun cas pour nous indigner de cette pratique, c'est pour décrire le monde tel qu'il est. English: In describing the American methods of data interception in relation to the French diplomatic representation in the United States, we do not aim at expressing indignation about this practice, but rather at describing the world as it is. French: La France a-t-elle bénéficié d'informations fournies par la NSA concernant des opérations terroristes visant nos intérêts ? English: Has France benefited from the intelligence supplied by the NSA concerning terrorist operations against our interests? French: Peut-on se priver de la collaboration américaine ? English: Can we do without collaboration with the Americans? French: La mise en place depuis en gros dix ans d'outils technologiques d'interception très puissants par les Etats-Unis, mais aussi par la France, a officiellement été justifiée par la lutte contre le terrorisme. English: The setting up of high-performance interception technology over practically the past ten years by the United States - and by France - has been officially justified by the fight against terrorism. French: D'ailleurs, dans ce domaine, la France et les Etats-Unis notamment ont mis en place des procédures de coopération et d'échanges d'informations quasi quotidiens et qui sont décrits de part et d'autre comme essentiels. English: Furthermore, in this regard, France and the United States in particular have implemented procedures, sometimes described as essential, for cooperating and exchanging information on an almost daily basis. French: A titre d'exemple, la présence de Mohammed Merah dans les zones tribales à Miranshah a été signalée aux Français grâce aux moyens de la NSA. English: For example, France was informed of the presence of Mohammed Merah in the tribal areas of Miranshah through the NSA's resources. French: La France peut être conduite, par exemple, à transmettre des blocs entiers de données sur la région du Sahel aux services américains, et, en contrepartie - on l'a déjà rapidement dit -, les Américains peuvent donner des informations aux Français sur d'autres régions du monde. English: Also France may, for example, have to transmit entire blocks of data on the Sahel region to the Americans and, in return - as already briefly mentioned - the Americans may provide information to the French about other parts of the world. French: Donc la question de fond derrière cette affaire NSA n'est pas tant la capacité ou le droit des pays de se doter d'outils d'interception, que la question de l'absence totale de débat préalable, notamment au sein des Parlements, sur la justification de tels systèmes, le périmètre qui doit être le leur, et, en fin de compte, la question des atteintes aux libertés. English: Hence the question at the heart of the NSA affair is not so much the capacity or the right of a country to use interception tools, as the issue of the complete lack of prior debate - especially within parliaments - on the justification of such systems, the extent to which they should be used and, ultimately, the issue of the infringement of freedoms. French: Que risquent réellement les Etats-Unis ? une dégradation de leur image? English: What risk does the United States actually run? Ruining its image? French: On a beau les dénoncer, je ne vois pas de quelle manière ils pourront être punis. English: However much we denounce the US, I see no way in which it can be punished. French: Le risque couru par les Américains peut être double. English: The risk run by the Americans could be twofold. French: Le premier, c'est lorsque leurs alliés - et ça a été le cas récemment - apprennent que leurs dirigeants, parfois au plus haut sommet de leur Etat, ont été surveillés. English: The first is when their allies - as has been the case recently - learn that their governments have been spied on, sometimes at the highest level. French: C'est le cas du Brésil et de l'Allemagne, deux pays où les relations diplomatiques avec les Etats-Unis se sont tendues. English: This is the case in Brazil and Germany, two countries where diplomatic relations with the United States are strained. French: Un homme de Cambridge a revendiqué la responsabilité de cet acte sur son compte Twitter, où il a posté des images d'Adolf Hitler. English: ``` **OpenAI 20-shot** ``` Translate French to English Comment expliquer l'attitude contradictoire du gouvernement français, qui d'un coté s'offusque en public en convoquant l'ambassadeur des Etats-Unis le 21 octobre, et de l'autre interdit le survol du territoire par l'avion présidentiel bolivien, sur la base de la rumeur de la présence à son bord d'Edward Snowden ? = Why the contradictory attitude of the French government? On the one hand, it publicly takes offence and summons the Ambassador of the United States on October 21 and, on the other, it forbids the Bolivian president's plane to enter its air space on the basis of a rumor that Edward Snowden was on board? Selon moi, il y a deux niveaux de réponse de la part du gouvernement français. = In my opinion, there are two levels of response from the French government. Lorsque François Hollande téléphone à Barack Obama ou quand le ministre des affaires étrangères Laurent Fabius convoque l'ambassadeur des Etats-Unis, ils réagissent à une vraie découverte, qui est celle de l'ampleur de la surveillance américaine sur l'ensemble des communications en France. = When François Hollande telephones Barack Obama, or when Foreign Minister Laurent Fabius summons the Ambassador of the United States, they are responding to a real discovery, that of the scale of America's surveillance of communications within France generally. N'est-il pas surprenant de lire dans les colonnes du Monde à quelques semaines d'intervalle d'une part la reproduction de la correspondance diplomatique américaine et d'autre part une condamnation des écoutes du Quai d'Orsay par la NSA ? = And is it not surprising to read in the pages of Le Monde, on the one hand, a reproduction of diplomatic correspondence with the US and, on the other, condemnation of the NSA's spying on the Ministry of Foreign Affairs on the Quai d'Orsay, within a matter of weeks? N'y aurait-il pas comme une vague hypocrisie de votre part ? = Is there not an element of hypocrisy on your part? La démarche journalistique n'est pas un positionnement moral, mais la recherche de l'intérêt et de la pertinence d'informations qui permettent à chaque citoyen de se forger une opinion. = The journalistic method is not to adopt a moral position, but to investigate the significance and relevance of information and enable every citizen to form an opinion. Lorsque WikiLeaks lève le voile sur l'analyse par la diplomatie américaine d'enjeux politiques ou autres dans le monde entier, nous considérons en effet que, au regard de la puissance américaine, cela constitue un éclairage important. = When WikiLeaks reveals the American administration's monitoring of political and other matters somewhere in the world, we consider this to be significant enlightenment with regard to the American government. Lorsque nous décrivons les systèmes d'interception américains à l'encontre de la diplomatie française aux Etats-Unis, ce n'est en aucun cas pour nous indigner de cette pratique, c'est pour décrire le monde tel qu'il est. = In describing the American methods of data interception in relation to the French diplomatic representation in the United States, we do not aim at expressing indignation about this practice, but rather at describing the world as it is. La France a-t-elle bénéficié d'informations fournies par la NSA concernant des opérations terroristes visant nos intérêts ? = Has France benefited from the intelligence supplied by the NSA concerning terrorist operations against our interests? Peut-on se priver de la collaboration américaine ? = Can we do without collaboration with the Americans? La mise en place depuis en gros dix ans d'outils technologiques d'interception très puissants par les Etats-Unis, mais aussi par la France, a officiellement été justifiée par la lutte contre le terrorisme. = The setting up of high-performance interception technology over practically the past ten years by the United States - and by France - has been officially justified by the fight against terrorism. D'ailleurs, dans ce domaine, la France et les Etats-Unis notamment ont mis en place des procédures de coopération et d'échanges d'informations quasi quotidiens et qui sont décrits de part et d'autre comme essentiels. = Furthermore, in this regard, France and the United States in particular have implemented procedures, sometimes described as essential, for cooperating and exchanging information on an almost daily basis. A titre d'exemple, la présence de Mohammed Merah dans les zones tribales à Miranshah a été signalée aux Français grâce aux moyens de la NSA. = For example, France was informed of the presence of Mohammed Merah in the tribal areas of Miranshah through the NSA's resources. La France peut être conduite, par exemple, à transmettre des blocs entiers de données sur la région du Sahel aux services américains, et, en contrepartie - on l'a déjà rapidement dit -, les Américains peuvent donner des informations aux Français sur d'autres régions du monde. = Also France may, for example, have to transmit entire blocks of data on the Sahel region to the Americans and, in return - as already briefly mentioned - the Americans may provide information to the French about other parts of the world. Donc la question de fond derrière cette affaire NSA n'est pas tant la capacité ou le droit des pays de se doter d'outils d'interception, que la question de l'absence totale de débat préalable, notamment au sein des Parlements, sur la justification de tels systèmes, le périmètre qui doit être le leur, et, en fin de compte, la question des atteintes aux libertés. = Hence the question at the heart of the NSA affair is not so much the capacity or the right of a country to use interception tools, as the issue of the complete lack of prior debate - especially within parliaments - on the justification of such systems, the extent to which they should be used and, ultimately, the issue of the infringement of freedoms. Que risquent réellement les Etats-Unis ? une dégradation de leur image? = What risk does the United States actually run? Ruining its image? On a beau les dénoncer, je ne vois pas de quelle manière ils pourront être punis. = However much we denounce the US, I see no way in which it can be punished. Le risque couru par les Américains peut être double. = The risk run by the Americans could be twofold. Le premier, c'est lorsque leurs alliés - et ça a été le cas récemment - apprennent que leurs dirigeants, parfois au plus haut sommet de leur Etat, ont été surveillés. = The first is when their allies - as has been the case recently - learn that their governments have been spied on, sometimes at the highest level. C'est le cas du Brésil et de l'Allemagne, deux pays où les relations diplomatiques avec les Etats-Unis se sont tendues. = This is the case in Brazil and Germany, two countries where diplomatic relations with the United States are strained. Un homme de Cambridge a revendiqué la responsabilité de cet acte sur son compte Twitter, où il a posté des images d'Adolf Hitler. = ``` **English target** ``` A man from Cambridge claimed responsibility for the act on his Twitter account, where he posted pictures of Adolf Hitler. ``` Results: #### Likelihood correct - Control | Engine | Empty prior | Control | | --- | --- | --- | | ada | -73.354 | -45.109 | | babbage | -70.899 | -35.969 | | curie | -70.224 | -30.489 | | davinci | -67.642 | -22.543 | #### Likelihood correct - Simple colon prompt | Engine | 0-shot | 1-shot | 2-shot | 5-shot | 10-shot | 20-shot | | --- | --- | --- | --- | --- | --- | --- | | ada | -19.658 | -17.257 | -16.519 | -16.112 | -15.741 | -16.309 | | babbage | -10.467 | -11.404 | -11.574 | -12.385 | -13.134 | -11.462 | | curie | -9.395 | -8.763 | -8.561 | -8.865 | -8.701 | -7.701 | | davinci | -6.443 | -6.658 | -6.614 | -6.807 | -7.388 | -6.387 | #### Likelihood correct - OpenAI prompt | Engine | 0-shot | 1-shot | 2-shot | 5-shot | 10-shot | 20-shot | | --- | --- | --- | --- | --- | --- | --- | | ada | -27.656 | -17.372 | -16.043 | -15.672 | -16.115 | -15.877 | | babbage | -11.877 | -12.015 | -12.205 | -12.868 | -13.800 | -11.765 | | curie | -13.414 | -8.735 | -8.841 | -8.462 | -8.685 | -7.147 | | davinci | -7.271 | -6.549 | -6.532 | -6.805 | -7.392 | -6.544 | ![Likelihood correct - French/English](/french_english.png) *Empty prior is plotted at -2, control prior at -1. Did not plot 20 shot; is was not substantially different from 10-shot for any of the models* Observations: * Unlike in the Roish example, for *all* models, and for both prompts, the change from the control prior to 0-shot is a much bigger change than from 0- to 1-shot. * The OpenAI prompt and simple colon prompt have similar scores for everything except for 0-shot, where the simple colon prompt provides better evidence, especially for `ada`. * For the simple colon prompt, all the models except `ada`, which gleaned some evidence from 1-shot, examples provided no evidence or *negative* evidence. `ada` and `curie` gleaned more evidence from 1-shot for OpenAI's prompt due to having derived comparatively *less* evidence from the 0-shot prompt. * For all the models, the transition from the empty prior (no prompt) to control (prompt is French sentence without directions to translate) provided more evidence for the correct answer than any instruction or examples. This makes sense – the most significant evidence for the correct English translation is not actually the instructions to translate, but the French sentence itself. Now for the decibel breakdown: \\--- #### ada | Prompt | Correct likelihood | +dB to empty prior | +dB to control | +dB to 0-shot | | --- | --- | --- | --- | --- | | empty prior | -73.354 | - | - | - | | control | -45.109 | 28.245 | - | - | | colon 0-shot | -19.658 | 53.696 | *25.451* | - | | colon 1-shot | -17.257 | 56.097 | 27.852 | 2.401 | | colon 2-shot | -16.519 | 56.835 | 28.590 | 3.139 | | colon 5-shot | -16.112 | 57.243 | 28.998 | 3.547 | | colon 10-shot | -15.741 | 57.614 | 29.368 | 3.917 | | colon 20-shot | -16.309 | 57.045 | 28.800 | *3.349* | | OAI 0-shot | -27.656 | 45.699 | *17.454* | - | | OAI 1-shot | -17.372 | 55.983 | 27.738 | 10.284 | | OAI 2-shot | -16.043 | 57.311 | 29.066 | 11.613 | | OAI 5-shot | -15.672 | 57.682 | 29.437 | 11.983 | | OAI 10-shot | -16.115 | 57.239 | 28.994 | 11.541 | | OAI 20-shot | -15.877 | 57.477 | 29.232 | *11.778* | **Simple colon +dB from 20-shot / +dB from 0-shot** = 3.349 / 25.451 = **0.132** **OpenAI +dB from 20-shot / +dB from 0-shot** = 11.778 / 17.454 = **0.675** \\--- #### babbage | Prompt | Correct likelihood | +dB to empty prior | +dB to control | +dB to 0-shot | | --- | --- | --- | --- | --- | | empty prior | -70.899 | - | - | - | | control | -35.969 | 34.930 | - | - | | colon 0-shot | -10.467 | 60.432 | *25.502* | - | | colon 1-shot | -11.404 | 59.495 | 24.565 | -0.937 | | colon 2-shot | -11.574 | 59.326 | 24.396 | -1.106 | | colon 5-shot | -12.385 | 58.514 | 23.584 | -1.918 | | colon 10-shot | -13.134 | 57.765 | 22.835 | -2.666 | | colon 20-shot | -11.462 | 59.438 | 24.508 | *-0.994* | | OAI 0-shot | -11.877 | 59.022 | *24.021* | - | | OAI 1-shot | -12.015 | 58.884 | 23.954 | -0.139 | | OAI 2-shot | -12.205 | 58.694 | 23.758 | -0.329 | | OAI 5-shot | -12.847 | 58.052 | 23.122 | -0.971 | | OAI 10-shot | -13.800 | 57.100 | 22.170 | -1.923 | | OAI 20-shot | -11.765 | 59.134 | 24.167 | *0.111* | **Simple colon +dB from 20-shot / +dB from 0-shot** = -0.994 / 25.502 = **-0.039** **OpenAI +dB from 20-shot / +dB from 0-shot** = 0.111 / 24.021 = **0.005** \\--- #### curie | Prompt | Correct likelihood | +dB to empty prior | +dB to control | +dB to 0-shot | | --- | --- | --- | --- | --- | | empty prior | -70.224 | - | - | - | | control | -30.489 | 39.735 | - | - | | colon 0-shot | -9.395 | 60.829 | *21.094* | - | | colon 1-shot | -8.763 | 61.461 | 21.726 | 0.632 | | colon 2-shot | -8.561 | 61.663 | 21.928 | 0.834 | | colon 5-shot | -8.865 | 61.359 | 21.624 | 0.530 | | colon 10-shot | -8.701 | 61.523 | 21.743 | 0.694 | | colon 20-shot | -7.701 | 62.524 | 22.744 | *1.695* | | OAI 0-shot | -13.414 | 56.810 | *17.075* | - | | OAI 1-shot | -8.735 | 61.489 | 21.754 | 4.679 | | OAI 2-shot | -8.841 | 61.383 | 21.648 | 4.573 | | OAI 5-shot | -8.462 | 61.762 | 22.027 | 4.952 | | OAI 10-shot | -8.685 | 61.539 | 21.804 | 4.729 | | OAI 20-shot | -7.147 | 63.077 | 23.342 | *6.267* | **Simple colon +dB from 20-shot / +dB from 0-shot** = 1.695 / 21.094 = **0.080** **OpenAI +dB from 20-shot / +dB from 0-shot** = 6.267 / 17.075 = **0.367** \\--- #### davinci | Prompt | Correct likelihood | +dB to empty prior | +dB to control | +dB to 0-shot | | --- | --- | --- | --- | --- | | empty prior | -67.642 | - | - | - | | control | -22.543 | 45.099 | - | - | | colon 0-shot | -6.443 | 61.199 | *16.100* | - | | colon 1-shot | -6.658 | 60.984 | 15.884 | -0.215 | | colon 2-shot | -6.614 | 61.028 | 15.911 | -0.171 | | colon 5-shot | -6.807 | 60.835 | 15.749 | -0.364 | | colon 10-shot | -7.388 | 60.254 | 15.167 | -0.945 | | colon 20-shot | -6.387 | 61.255 | 16.169 | *0.056* | | OAI 0-shot | -7.286 | 60.368 | *15.269* | - | | OAI 1-shot | -6.567 | 61.087 | 15.975 | 0.719 | | OAI 2-shot | -6.506 | 61.148 | 16.093 | 0.779 | | OAI 5-shot | -6.797 | 60.857 | 15.835 | 0.489 | | OAI 10-shot | -7.392 | 60.262 | 15.184 | -0.106 | | OAI 20-shot | -6.528 | 61.126 | 15.966 | *0.757* | **Simple colon +dB from 20-shot / +dB from 0-shot** = 0.056 / 16.100 = **0.003** **OpenAI +dB from 20-shot / +dB from 0-shot** = 0.757 / 15.269 = **0.050** \\--- We see again that the more powerful models derive proportionally more evidence from the 0-shot task description than from subsequent examples, although compared to Roish, *all* the models derive a lot more evidence from the 0-shot description than from examples (the ratio is consistently < 1, whereas for Roish it was consistently > 1). These two tasks, one which uses an invented language and one which uses a real language that GPT-3 knows well, derive evidence from 0- and few-shot prompts in different proportions. When GPT-3 already knows how to perform a task, examples are less helpful. Using the simple colon prompt, none of the models except for `ada` appear to have much use for examples in terms of the likelihood of giving the correct translation. OpenAI's prompt results in worse 0-shot accuracy for `ada` and `curie`, and accordingly the presence of one example is helpful for those two models, but any additional examples have little effect. This supports the hypothesis that GPT-3 is not \"learning\" translation from the few-shot examples but rather locating the already-learned task, and that the lower BLEU scores for 0-shot prompts can probably be attributed to catastrophic failures. \\--- Counterexample - accuracy decreases with # shots \\------------------------------------------------ *Main article: **[List sorting does not play well with few-shot](/posts/list-sorting-does-not-play-well-with-few-shot/)*** A interesting counterexample to few-shot monotonicity is list sorting (e.g. [1, 3, 3, 4, 0] -> [0, 1, 3, 3, 4]). A zero-shot prompt which frames the problem as an example embedded in coding documentation achieves 76% accuracy, whereas a 32-shot prompt achieves only 20%. Adding examples to the end of the coding prompt results in a monotonic *decrease* in accuracy with number of shots. 0-shot code prompt: ``` The sort function can be used to sort a list in ascending, descending or user defined order. To sort the list in ascending order, call list.sort(). This will sort a list of integers so that the smallest integer will be first in the list and the largest integer will be the last and the last element of the list is the largest.\" For example: list = [1, 0, 4, 3, 3] list.sort() = ``` Example-only prompt: ``` Unsorted list: [5, 6, 2, 3, 2] Sorted list: [2, 2, 3, 5, 6] Unsorted list: [8, 5, 8, 8, 4] Sorted list: [4, 5, 8, 8, 8] ... Unsorted list: [1, 0, 4, 3, 3] Sorted list: ``` **Code preprompt, length 5** | Shots | Correct | Accuracy | | --- | --- | --- | | **0** | **38/50** | **0.76** | | 1 | 33/50 | 0.66 | | 3 | 23/50 | 0.46 | | 5 | 22/50 | 0.44 | | 7 | 22/50 | 0.44 | | 10 | 21/50 | 0.42 | | 13 | 15/50 | 0.30 | | 16 | 16/50 | 0.32 | **No preprompt, length 5** | Shots | Correct | Accuracy | | --- | --- | --- | | 0 | 14/50 | 0.28 | | **1** | **20/50** | **0.40** | | 3 | 15/50 | 0.30 | | 5 | 14/50 | 0.28 | | 7 | 16/50 | 0.32 | | **10** | **25/50** | **0.50** | | 13 | 18/50 | 0.36 | | 16 | 11/50 | 0.22 | | 32 | 10/50 | 0.20 | ![list sorting results](/sorting/interesting2.png) *comparison of accuracies from 0 - 16 shots* Pretty weird, right? I speculate on the causes of these unexpectedly shaped plots in my [post](/posts/list-sorting-does-not-play-well-with-few-shot/) about this experiment. Regardless of the mechanism, though, the fact that having more examples in the prompt makes GPT-3 worse at the task implies that there are more factors in play than simply metalearning. List sorting is the only quantitative task that I know to exhibit strict decrease in accuracy with number of shots, but I suspect there are many more. \\--- Predictions \\----------- I will offer some specific predictions suggested by my interpretation of these results, which may be corroborated or falsified by further experiments and future developments. 1\\. For translation tasks, if catastrophic failures are filtered out, the discrepancy between 0-shot and 64-shot performance will disappear. 1\\. 1-shot performance, however, will be worse than both 0-shot and 64-shot performance even after catastrophic failures are filtered out, due to being most vulnerable to [few-shot bugs](/posts/methods-of-prompt-programming/#few-shot-bugs). 2\\. For practically all tasks that can be meaningfully described using a 0-shot prompt, the ratio of **+ dB from few-shot over 0-shot / +dB from 0-shot over prior** will tend to be smaller for larger models than smaller models, indicating that larger models glean more evidence from 0-shot task descriptions compared to examples. 3\\. There are tasks for which a zero-shot prompt causes GPT-3 to perform better than *any* conventional[3](#fn:3) few-shot prompt. [Sorting lists of five integers](/posts/list-sorting-does-not-play-well-with-few-shot/) is probably one such task. 4\\. More powerful language models released in the future will continue to become more effective at 0-shot \"learning,\" and this capability will increase even more dramatically than improvement at few-shot learning. \\--- 1\\. This is probably the situation for many zero-shot translation tasks with inadequate task specification. [↩︎](#fnref:1) 2\\. Of course, the representativeness of examples is always a concern when evaluating language models. [↩︎](#fnref:2) 3\\. Conventional = a list of solved examples, optionally following a task description, excluding complex embeddings of examples e.g. in a narrative. [↩︎](#fnref:3) \\--- [GPT-3](/categories/gpt-3/)[prompt engineering](/categories/prompt-engineering/)[benchmarks](/categories/benchmarks/)[bayes](/categories/bayes/)[interpretability](/categories/interpretability/) 10831 Words Feb 10, 2021 \\--- [← List sorting does not play well with few-shot](/posts/list-sorting-does-not-play-well-with-few-shot/) [Loom: interface to the multiverse →](/posts/loom-interface-to-the-multiverse/)© 2022 [moire](/about)$(function(){window.sr=ScrollReveal();if($(window).width()<768){if($('.timeline-content').hasClass('js--fadeInLeft')){$('.timeline-content').removeClass('js--fadeInLeft').addClass('js--fadeInRight');} sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});}else{sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});} sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});});\n", "url": "https://generative.ink/posts/language-models-are-0-shot-interpreters/", "title": "Language models are 0-shot interpreters :: — Moire[![](/images/home/rolling_phase.gif)](/)* [Posts](/posts/) * [Trees](/trees/) * [Prophecies](/prophecies/) * [About](/about/) 51 minutes [Language models are 0-shot interpreters](/posts/language-models-are-0-shot-interpreters/) ========================================================================================== \\--- Table of Contents* [Overview](#overview) * [0-shot, few-shot, and meta-learning](#0-shot-few-shot-and-meta-learning) \\+ [0-shot can outperform few-shot](#0-shot-can-outperform-few-shot) * [Measuring prompt helpfulness](#measuring-prompt-helpfulness) \\+ [Decibels of evidence](#decibels-of-evidence) \\+ [Translation task 1: English -> Roish](#translation-task-1-english---roish) \\+ [Translation task 2: French -> English](#translation-task-2-french---english) * [Counterexample - accuracy decreases with # shots](#counterexample---accuracy-decreases-with--shots) * [Predictions](#predictions) \\--- **! Correction:** The logprobs returned by the OpenAI API use natural log, not base 10, so all occurences of **decibels / dB** in this post should actually say ***nats***. I’ll either make that substitution at some point or convert everything to actual decibels. Overview \\-------- I present evidence that the efficacy of 0-shot prompts for GPT-3 has been underestimated, and that more powerful models are more effective at deriving information from 0-shot prompts, while less powerful models have greater need for examples on equivalent tasks. From this evidence, I extrapolate three principal claims: 1\\. [Few-shot](https://arxiv.org/abs/2005.14165) prompts are not always an efficient or necessary means of task specification for GPT-3. [Sometimes](/posts/list-sorting-does-not-play-well-with-few-shot/), more examples in the prompt makes accuracy strictly *worse*. 2\\. For some tasks, such as translation between well-known languages and list sorting, GPT-3 is a *0-shot interpreter* - a short task description or signifier suffices to invoke its full capabilities. 3\\. 0-shot performance scales with model size more drastically than few-shot performance, suggesting that 0-shot task specification will become a more important prompting strategy as language models increase in capability. \\--- > #### The diversity of tasks the model is able to perform in a zero-shot setting suggests that high-capacity models trained to maximize the likelihood of a sufficiently varied text corpus begin to learn how to perform a surprising amount of tasks without the need for explicit supervision. > > – [Language Models are Unsupervised Multitask Learners] > > 0-shot, few-shot, and meta-learning \\----------------------------------- The GPT-2 paper, entitled *[Language Models are Unsupervised Multitask Learners](https://openai.com/blog/better-language-models/)*, proposes that *unsupervised* language models are capable of learning a wide range of benchmark tasks, despite not having trained on datasets specific to those tasks. The evidence takes the form of GPT-2, trained only on the WebText natural language corpus, demonstrating “promising, competitive, and state of the art” results on a wide range of tasks in a “zero-shot setting.” “Zero-shot” here means merely “without any parameter or architecture modification,” encompassing what we’d now call *few-shot* prompts which contain examples of solved tasks. Previous approaches relied on supervised fine-tuning, either purely or following pre-training, so this was a novel result. Table 1 in the paper illustrates how it is possible for a model trained on a diverse “language in the wild” dataset to learn specific tasks like translation by showing examples of translations embedded in WebText: > ”I’m not the cleverest man in the world, but like they say in French: **Je ne suis pas un imbecile [I’m not a fool].** > > In a now-deleted post from Aug. 16, Soheil Eid, Tory candidate > in the riding of Joliette, wrote in French: ”**Mentez mentez, > il en restera toujours quelque chose**,” which translates as, > ”**Lie lie and something will always remain.**” > > “I hate the word ‘**perfume**,”’ Burr says. ‘It’s somewhat better > in French: ‘**parfum.**’ > > If listened carefully at 29:55, a conversation can be heard > between two guys in French: “**-Comment on fait pour aller > de l’autre coté? -Quel autre coté?**”, which means “**- How > do you get to the other side? - What side?**”. > > If this sounds like a bit of a stretch, consider this ques- > tion in French: **As-tu aller au cin��ma?**, or **Did you go to > the movies?**, which literally translates as Have-you to go to > movies/theater? > > “**Brevet Sans Garantie Du Gouvernement**”, translated to > English: “**Patented without government warranty**”. > > A model learning to predict WebText has to learn, among many other things, how to predict translations as they might occur in natural contexts. To measure GPT-2’s translation ability, few-shot prompts were used: > In order to *help it infer that > this is the desired task*, we condition the language model > on a context of example pairs of the format `english sentence = french sentence` and then after a final prompt of `english sentence =` we sample from > the model with greedy decoding and use the first generated > sentence as the translation. > > Emphasis mine. The authors do not distinguish between 0-shot and (what is now called) few-shot, mentioning the priming examples only offhand, in the same sentence asserting that priming serves to *communicate* the desired task to GPT-2. There is no suggestion that GPT-2’s ability to translate is informed by the information contained in the examples beyond the recognition that they are examples of translation. One year later, language has evolved: The GPT-3 paper, [*Language Models are Few-Shot Learners*](https://arxiv.org/abs/2005.14165), does distinguish between prompts that contain examples (n-shot for n examples) and which contain only a task description (0-shot). Where before anything that wasn’t fine-tuning went into the bucket of “0-shot,” now there is an ontology that puts the *number of examples in the prompt* on center stage. This new ontology culminates in a title whose implication is surprising in light of the previous paper’s casual comment on the function of priming examples: Language models are few-shot *learners.* The type of learning referred to in the title, “meta-learning,” does not necessarily imply that the task is literally learned from the examples, only that the examples help for some reason. According to a footnote, > These terms are intended to remain agnostic on the question of whether the model > learns new tasks from scratch at inference time or simply recognizes patterns seen during training – this is an important issue which > we discuss later in the paper, but “meta-learning” is intended to encompass both possibilities, and simply describes the inner-outer > loop structure. > > The later discussion is not very extensive, mostly just acknowledging the ambiguity inherent to few-shot: > A limitation, or at least uncertainty, associated with few-shot learning in GPT-3 is ambiguity about whether few-shot > learning actually learns new tasks “from scratch” at inference time, or if it simply recognizes and identifies tasks that it > has learned during training. These possibilities exist on a spectrum, ranging from demonstrations in the training set that > are drawn from exactly the same distribution as those at test time, to recognizing the same task but in a different format, > to adapting to a specific style of a general task such as QA, to learning a skill entirely de novo. (…) (U)nderstanding precisely > how few-shot learning works is an important unexplored direction for future research. > > This is the uncertainty that I will investigate in this blog post, expanding on the results published in [*Prompt Programming for Large Language Models: Beyond the Few-Shot Paradigm*](https://arxiv.org/abs/2102.07350). My purpose is also to challenge the ontology introduced by *Language Models are Few-Shot Learners*. Although the authors are careful to remain agnostic as to the mechanism of few-shot/meta-learning, what we have found by probing the mechanism suggests that an alternative framework which emphasizes the means by which a task is *communicated* may be more salient in some contexts. The emphasis on few-shot given by OpenAI’s paper has influenced subsequent research about GPT-3, some of which has focused on [optimizing](https://arxiv.org/abs/2101.06804) [few-shot prompts](https://arxiv.org/abs/2102.09690). I suspect that this is a contingent rather than convergent history: that if OpenAI had taken a different angle (and with GPT-3, the set of possible angles to choose from seems inexhaustible), current research on prompts would likewise look different. It’s strange to call it the *few-shot paradigm* when the idea of few-shot was invented less than a year ago, but perhaps we’re moving into a phase in which paradigms are measured in months. Really, though, the paradigm I want to shift out of is that of *supervised learning*. Few-shot prompting resembles fine-tuning in that it seeks to coerce an unsupervised language model into performing closed-ended tasks using unstructured lists of solved examples. Self-supervised natural language models are amenable to many strategies for communicating and instructing tasks, including but not limited to demonstrations. ### 0-shot can outperform few-shot Looking at the results presented in the GPT-3 paper, however, the focus on few-shot and meta-learning seems justified. If one thing trend is clear in the data, it’s this: **monotonic improvement with number of shots**. On basically every variation of every task and for all model sizes, 1-shot does better than 0-shot, and many-shot does better than 1-shot. For instance, here is the page of graphs of all the results of translation tasks: ![all translation tasks](/0shot/all_trans.png) *Blue is 0-shot, green 1-shot, orange 64-shot. X axis is model size, Y axis is BLEU score* Some of the 0-shot lines are jagged, with non-monotonic accuracy with regard to model size, which is pretty odd. However, accuracy consistently improves as the number of shots increases from 0 to 1 and then 64. When we investigated the French -> English translation more closely, we found some interesting results, which I’ll walk through here. First, here are the results presented in OpenAI’s paper: **French -> English translation Multi-BLEU scores** | Prompt | 2.7B | 6.7B | 13B | 175B | | --- | --- | --- | --- | --- | | OpenAI 0-shot | 21.2 | 15.5 | 22.4 | 21.9 | | OpenAI 1-shot | 20.0 | 31.6 | 31.4 | 35.6 | | OpenAI 64-shot | 35.1 | 36.4 | 38.3 | 41.4 | ![french to english translation](/0shot/fren.png) Here are the prompt formats that OpenAI used for the French to English translation task: **OpenAI 0-shot prompt** ``` Translate French to English {french phrase} = ``` **OpenAI n-shot prompt** ``` Translate French to English {french phrase} = {english phrase} {french phrase} = {english phrase} ... {french phrase} = ``` As we were attempting to replicate these results, we noticed that when the model was failing on the 0-shot prompt, the failures were often of *catastrophic* nature: the task was not attempted at all, e.g. the model would output a newline, or another (or the same) French phrase instead of an attempt at an English translation. BLEU assigns a score from 0 to 1 to the accuracy of a translation, and would assign a score close to 0 to a catastrophic failure. The scores reported in the paper, however, are averaged over a large dataset, so the same score could hypothetically correspond to uniformly flawed attempts *or* a mix of perfect attempts and catastrophic failures. It seemed possible that 0-shot prompts were much less reliable at getting the models to attempt the translation task, but result in equivalent accuracy in the event that they did attempt it. To investigate this possibility, we ran the benchmarks using an alternative 0-shot prompt which contains no additional instructions, but whose format better resembles how a translation might be embedded in a natural language corpus: **Simple colon prompt** ``` Translate French to English French: {french phrase} English: ``` These were the results we got: | Prompt | Babbage | Curie | | --- | --- | --- | | Reproduced OpenAI 0-shot | 15.9 | 18.7 | | Reproduced OpenAI 1-shot | 21.8 | 24.1 | | Reproduced OpenAI 10-shot | 25.1 | 27.9 | | Simple colon 0-shot | 23.5 | 33.3 | | Simple colon 1-shot | 18.0 | 27.6 | | Simple colon 10-shot | 24.1 | 33.4 | *Note: It is unknown what model sizes the models on the OpenAI API (ada, babbage, curie, and davinci) correspond to. We were not able to fit 64 examples in the API input window, so used 10-shot prompts instead.* A couple of surprising observations: * **Simple colon 0-shot** is very effective, especially on Curie, where it outperforms everything except **Simple colon 10-shot**, whose performance it matches. * **Simple colon 1-shot** is *worse* than 0-shot on both Babbage and Curie. The simple colon format does not exhibit monotonicity with number of shots. Having one example actually made accuracy worse. How could that be? Here’s my hypothesis about what is going on: * All sizes of GPT-3 already know how to translate to some extent, and translation is way too hard a task to learn “de novo” from a prompt containing only a few examples. So, as the GPT-2 paper said, the examples serve to help it infer that translation is the desired task. A prompt may be more or less effective at specifying the task; generally, the more examples there are, the more “clear” it is, but a good zero-shot prompt may be worth many examples. On the other hand, if the zero-shot prompt is unclear, then adding more examples will improve accuracy, since the baseline was so poor. The zero-shot prompt could be arbitrarily unhelpful, and an example is better than *nothing*. * A low number of examples can be more confounding than no examples. We noticed that sometimes the model would respond to one-shot prompts as if the *semantic* content of the example translation was relevant to the new translation. Without multiple examples, it’s less clear that the translation instances are meant to be parallel and independent. I have written about the bugs that affect few-shot prompting [here](/posts/methods-of-prompt-programming/#few-shot-bugs). How much of the apparent consistent monotonic improvement in performance on tasks relative to number of shots in OpenAI’s results can be attributed to an unhelpful zero-shot prompt? Much more extensive testing is needed to say, but I suspect that this is the case for most of the translation tasks, at least. It’s very expensive in API credits to run these translation benchmarks, so we haven’t run any more yet. However, there are also less expensive methods to explore the way that of few- and 0-shot prompts contribute to task accuracy. In the next section of this post, I’ll subject this translation benchmark to a different method of analysis. \\--- Measuring prompt helpfulness \\---------------------------- The method I’m going to use in this section monitors the conditional probability of GPT-3 giving a correct answer as the prompt is altered. GPT-3 allows us to measure the probability that it outputs any verbatim sequence given any prompt by multiplying the probability (or adding the logprobs) that it outputs the first token given the prompt, and then the second given the prompt and the first, etc (here’s [code](/posts/parsing-by-counterfactual/#conditional-probability-of-a-target) to do it). Compared to greedy sampling (temperature 0), which is typically used for benchmarks, this method doesn’t rely on everything going right at each point in the sequence to give a nonzero score - if the first token is likely to be wrong, but the rest of the answer is likely to be correct *given* that the first token is correct[1](#fn:1), this is reflected as a likelihood penalty instead of a total failure. Compared to stochastic sampling (temperature > 0), this method does not require multiple rollouts. Compared to BLEU scores, this method only measures the probability of one verbatim “correct” answer. If this method is used to gauge a more general property, like the helpfulness of a prompt, it’s important to keep in mind that it relies on the probability of the verbatim correct answer being a *good proxy* for accuracy in general. In the same way, it also relies on the specific content of the prompt being a good proxy for the more general property of prompts you are measuring, such as number of shots[2](#fn:2). The absolute conditional probability of the right answer is influenced by various factors, such as how long the sequence is (the longer it is, the more things have to go as planned for it to have been output, hence lower probability), so it does not directly reflect the informativeness of a prompt. Instead, we will concern ourselves with the *difference* between the conditional probability of a sequence and that of the same sequence under different circumstances. What I’ll do is measure the conditional probability of a correct answer given a prompt - say, a 10-shot prompt - and then compare that probability to the conditional probability of a correct answer given a different prompt with “less information,” such as a 0-shot prompt, or one with no task information at all. The difference between the log likelihood of the target with and without a piece of the prompt gives us the *decibels* of evidence provided by that component for the target. ### Decibels of evidence The decibels of evidence provided by a piece of evidence for a target given a prior is ``` logL(target | prior + evidence) - logL(target | prior) ``` `logL` means log likelihood, which is the natural logarithm of the probability. Why is quantity of evidence measured in differences of *log* probability instead of regular probability (or something else)? Say that you are trying to guess a 4-bit binary sequence. If you have no information, there are 2^4 = 16 possibilities, and your probability of being correct if you guess is 1/16. If you receive a single bit of information - equivalent to receiving the answer to a single yes/no question - now the state of your belief about the sequence is, say, `1 ? ? ?` instead of `? ? ? ?`. There are 2^3 = 8 possibilities remaining, and your chance of being correct if you guess is 1/8. Each bit of information revealed halves your uncertainty. Your probability of being correct goes 1/16 -> 1/8 -> 1/4 -> 1/2 -> 1 as you receive bits of information. These steps are logarithmic in probability and linear in log probability. If we believe that this generalizes to more complex situations, then we should measure the quantity of evidence in log likelihood. If the likelihood of a hypothesis goes from 1/16 -> 1/8, we think that a similar amount of evidence was in play as if the probability of the hypothesis goes from 1/4 -> 1/2, or if it goes from 1/20000 -> 1/10000. ### Translation task 1: English -> Roish Before analyzing the French -> English translation task from the benchmark, I’m going to take a detour and analyze a different translation task using the decibel method. In [Philosophers On GPT-3](https://dailynous.com/2020/07/30/philosophers-gpt-3/#askell), Amanda Askell shows an example of GPT-3 having learned the fictional “Roish” language after seeing a description and two examples: ![roish](https://dailynous.com/wp-content/uploads/2020/07/gpt-3-askell-roish.png) *Why* was GPT-3 able to get the pattern? Was it because of the examples, or would the description of Roish be sufficient, as it would be for a human? To measure this, I have come up with a few alternate versions of the Roish prompt: #### control The control prompt is in the format of the original but without any description of what Roish is. This will be our “no information” prior. ``` Today we're going to be playing with the fictional Roish language. English: The weather is lovely! Roish: ``` #### 0-shot The 0-shot prompt contains the description of Roish from the original, but no examples. ``` Today we're going to be playing with the fictional Roish language. Roish is a lot like English except every word ends in \"ro\". English: The weather is lovely! Roish: ``` #### “half”-shot What I’m calling a “half-shot” prompt is a description which incorporates an example of a single English -> Roish word mapping. ``` Today we're going to be playing with the fictional Roish language. Roish is a lot like English except \"ro\" is appended to the end. For instance, the word \"writing\" becomes \"writingro\". English: The weather is lovely! Roish: ``` #### 1-shot A description followed by one solved example. ``` Today we're going to be playing with the fictional Roish language. Roish is a lot like English except every word ends in \"ro\". English: Writing about language models is fun. Roish: Writingro aboutro languagero modelsro isro funro. English: The weather is lovely! Roish: ``` I also have 2-shot and 10 shot prompts: **2-shot** ``` Today we're going to be playing with the fictional Roish language. Roish is a lot like English except every word ends in \"ro\". English: Writing about language models is fun. Roish: Writingro aboutro languagero modelsro isro funro. English: I wonder if the language model can get the pattern. Roish: Iro wonderro ifro thero languagero modelro canro everro getro thero patternro. English: The weather is lovely! Roish: ``` **10-shot** ``` Today we're going to be playing with the fictional Roish language. Roish is a lot like English except every word ends in \"ro\". English: Mrs. Juarez and Mr. Smith are dancing gracefully. Roish: Mrsro. Juarezro andro Mrro. Smithro arero dancingro gracefullyro. English: Samantha, Elizabeth, and Joan are on the committee. Roish: Samantharo, Elizabethro, andro Joanro arero onro thero committeero. English: The ham, green beans, mashed potatoes, and corn are gluten-free. Roish: Thero hamro, greenro beansro, mashedro potatoesro, andro cornro arero glutenro-freero. English: The paper and pencil sat idle on the desk. Roish: Thero paperro andro pencilro satro idlero onro thero deskro. English: Sometimes the most difficult questions have the simplest solutions! Roish: Sometimesro thero mostro difficultro questionsro havero thero simplestro solutions! English: While breakthroughs in machine learning and artificial intelligence are changing society, our fundamental understanding has lagged behind. Roish: Whilero breakthroughsro inro machinero learningro andro artificialro intelligencero arero changingro societyro, ourro fundamentalro understandingro hasro laggedro behindro. English: Do they need to have access to data other than text in order to do this? Roish: Doro theyro needro toro havero accessro toro dataro otherro thanro textro inro orderro toro doro this? English: But it’s clearly seen enough of these kinds of patterns to identify the rule. Roish: Butro it’sro clearlyro seenro enoughro ofro thesero kindsro ofro patternsro toro identifyro thero rulero. English: Writing about language models is fun. Roish: Writingro aboutro languagero modelsro isro funro. English: I wonder if the language model can get the pattern. Roish: Iro wonderro ifro thero languagero modelro canro everro getro thero patternro. English: The weather is lovely! Roish: ``` The target string whose conditional probability I’m measuring is: ``` Thero weatherro isro lovelyro! ``` Here are the results for the log likelihood of the target given each of these prompts, evaluated on four different sizes of GPT-3 available on the API (in order from smallest to largest: `ada`, `babbage`, `curie`, `davinci`): #### Log likelihood of correct continuation | Engine | Control | 0-shot | Half-shot | 1-shot | 2-shot | 10-shot | | --- | --- | --- | --- | --- | --- | --- | | ada | -56.245 | -53.483 | -48.837 | -18.698 | -18.942 | -6.652 | | babbage | -43.325 | -35.105 | -29.206 | -9.878 | -10.753 | -8.861 | | curie | -32.609 | -28.556 | -28.459 | -10.207 | -5.363 | -1.072 | | davinci | -28.377 | -14.963 | -15.086 | -3.254 | -3.556 | -0.903 | ![Likelihood correct - Roish](/roish.png) For convenience of visual comparison, the control prompt is plotted at x=-1 and the “half-shot” prompt is plotted at x=0.5. *Not to be taken to imply that control literally corresponds to -1 shots and the half-shot prompt to half a shot*, whatever that would mean. Observations about this chart: * The bigger models are more likely to be correct than the smaller models even for the “no information” prior. The reason for this is probably because *given* that part of the correct answer has already occurred, the bigger models are better able to recognize it as a pattern *in itself*, even in the absence of a preceding task description, and will then be more likely to continue the pattern. * For all the models, even the 0-shot prompt was an improvement over the “no information” prior. It was the most helpful for `davinci`. * Half-shot was an improvement over 0-shot for the smaller models but not the bigger ones, which is the opposite of what I expected. * Going from 0- to 1-shot was helpful for all the models, but especially for the smaller ones, whereas the larger models started out in a better place at 0-shot already. * Additional shots were much less helpful for all the models except `ada`, which benefited so much from 10 shots that its probability of getting the right answer exceeded `babbage`! Now let’s look at the decibel values. For each prompt where it is applicable, I calculate the decibels in relation to both the control “no information” prior and 0-shot. \\--- #### ada | Prompt | Correct likelihood | +dB to control | +dB to 0-shot | | --- | --- | --- | --- | | control | -56.245 | - | - | | 0-shot | -53.483 | *2.762* | - | | 1-shot | -18.698 | 37.547 | 34.785 | | half-shot | -48.837 | 7.409 | 4.647 | | 2-shot | -18.942 | 37.303 | 34.541 | | 10-shot | -6.652 | 49.594 | *46.832* | **+dB from 10 examples / +dB from task description** = 46.832 / 2.762 = **16.956** \\--- #### babbage | Prompt | Correct likelihood | +dB to control | +dB to 0-shot | | --- | --- | --- | --- | | control | -43.325 | - | - | | 0-shot | -35.105 | *8.220* | - | | 1-shot | -9.878 | 33.454 | 25.227 | | half-shot | -29.206 | 14.119 | 5.899 | | 2-shot | -10.753 | 32.572 | 24.352 | | 10-shot | -8.861 | 34.464 | *26.244* | **+dB from 10 examples / +dB from task description** = 26.244 / 8.220 = **3.193** \\--- #### curie | Prompt | Correct likelihood | +dB to control | +dB to 0-shot | | --- | --- | --- | --- | | control | -32.609 | - | - | | 0-shot | -28.556 | *4.053* | - | | 1-shot | -10.207 | 22.440 | 18.348 | | half-shot | -28.459 | 4.150 | 0.097 | | 2-shot | -5.363 | 27.246 | 23.192 | | 10-shot | -1.072 | 31.537 | *27.483* | **+dB from 10 examples / +dB from task description** = 27.483 / 4.053 = **6.781** \\--- #### davinci | Prompt | Correct likelihood | +dB to control | +dB to 0-shot | | --- | --- | --- | --- | | control | -28.377 | - | - | | 0-shot | -14.963 | *13.414* | - | | 1-shot | -3.254 | 25.008 | 11.709 | | half-shot | -15.086 | 12.832 | -0.123 | | 2-shot | -3.556 | 24.801 | 11.407 | | 10-shot | -0.903 | 27.473 | *14.060* | **+dB from 10 examples / +dB from task description** = 14.060 / 13.414 = **1.048** \\--- For each of the models, I’ve calculated the ratio between the evidence in decibels provided by the 10-shots compared to the 0-shot prior and the evidence provided by 0-shot task description compared to the “no information” prior. **The smaller this ratio is, the more relatively informative the task description is compared to examples.** The decrease in this ratio is not quite monotonic with regard to model size - it is slightly higher for `curie` than it is for `babbage` - but it is dramatically different for `davinci`, the largest model, and `ada`, the smallest model. At 10 shots, `ada`’s 50 decibels improvement over control is almost entirely due to examples, whereas for `davinci`, the 0-shot description just about the same dBs of evidence over control as the 10 subsequent shots provide. This suggests that larger models get a lot more out of a task description compared to examples, even in a case like this where the task is completely made up and thus has to be “learned” at runtime. ### Translation task 2: French -> English Unlike English -> Roish, GPT-3 already knows how to translate French -> English, so the French -> English task is of a different nature than English -> Roish. Let’s run the same analysis on French -> English and see whether the breakdown of evidence reflects this difference. #### empty prior I will measure decibels relative to two different priors: an empty prior, which is nothing but a newline ‘\\n’, and a separate control prior which contains additional the French sentence but no task description. For the empty prior, like the other prompts, we measure the log likelihood of the correct answer following the prompt, which is in this case empty. #### control Control is the French sentence and a newline. We measure the probability of the English sentence being subsequently output, despite there being no information that a translation should follow. ``` Un homme de Cambridge a revendiqué la responsabilité de cet acte sur son compte Twitter, où il a posté des images d'Adolf Hitler. ``` #### 0-shot I will test the two 0-shot prompts from the first part of this post, OpenAI’s 0-shot prompt and my modified “simple colon prompt.” **Simple colon prompt** ``` Translate French to English French: Un homme de Cambridge a revendiqué la responsabilité de cet acte sur son compte Twitter, où il a posté des images d'Adolf Hitler. English: ``` **OpenAI 0-shot prompt** ``` Translate French to English Un homme de Cambridge a revendiqué la responsabilité de cet acte sur son compte Twitter, où il a posté des images d'Adolf Hitler. = ``` I will also evaluate both formats for n-shot prompts: **Colon 1-shot** ``` Translate French to English French: Comment expliquer l'attitude contradictoire du gouvernement français, qui d'un coté s'offusque en public en convoquant l'ambassadeur des Etats-Unis le 21 octobre, et de l'autre interdit le survol du territoire par l'avion présidentiel bolivien, sur la base de la rumeur de la présence à son bord d'Edward Snowden ? English: Why the contradictory attitude of the French government? On the one hand, it publicly takes offence and summons the Ambassador of the United States on October 21 and, on the other, it forbids the Bolivian president's plane to enter its air space on the basis of a rumor that Edward Snowden was on board? French: Un homme de Cambridge a revendiqué la responsabilité de cet acte sur son compte Twitter, où il a posté des images d'Adolf Hitler. English: ``` **OpenAI 1-shot** ``` Translate French to English Comment expliquer l'attitude contradictoire du gouvernement français, qui d'un coté s'offusque en public en convoquant l'ambassadeur des Etats-Unis le 21 octobre, et de l'autre interdit le survol du territoire par l'avion présidentiel bolivien, sur la base de la rumeur de la présence à son bord d'Edward Snowden ? = Why the contradictory attitude of the French government? On the one hand, it publicly takes offence and summons the Ambassador of the United States on October 21 and, on the other, it forbids the Bolivian president's plane to enter its air space on the basis of a rumor that Edward Snowden was on board? Un homme de Cambridge a revendiqué la responsabilité de cet acte sur son compte Twitter, où il a posté des images d'Adolf Hitler. = ``` **Colon 2-shot** ``` Translate French to English French: Comment expliquer l'attitude contradictoire du gouvernement français, qui d'un coté s'offusque en public en convoquant l'ambassadeur des Etats-Unis le 21 octobre, et de l'autre interdit le survol du territoire par l'avion présidentiel bolivien, sur la base de la rumeur de la présence à son bord d'Edward Snowden ? English: Why the contradictory attitude of the French government? On the one hand, it publicly takes offence and summons the Ambassador of the United States on October 21 and, on the other, it forbids the Bolivian president's plane to enter its air space on the basis of a rumor that Edward Snowden was on board? French: Selon moi, il y a deux niveaux de réponse de la part du gouvernement français. English: In my opinion, there are two levels of response from the French government. French: Un homme de Cambridge a revendiqué la responsabilité de cet acte sur son compte Twitter, où il a posté des images d'Adolf Hitler. English: ``` **OpenAI 2-shot** ``` Translate French to English Translate French to English Comment expliquer l'attitude contradictoire du gouvernement français, qui d'un coté s'offusque en public en convoquant l'ambassadeur des Etats-Unis le 21 octobre, et de l'autre interdit le survol du territoire par l'avion présidentiel bolivien, sur la base de la rumeur de la présence à son bord d'Edward Snowden ? = Why the contradictory attitude of the French government? On the one hand, it publicly takes offence and summons the Ambassador of the United States on October 21 and, on the other, it forbids the Bolivian president's plane to enter its air space on the basis of a rumor that Edward Snowden was on board? Selon moi, il y a deux niveaux de réponse de la part du gouvernement français. = In my opinion, there are two levels of response from the French government. Un homme de Cambridge a revendiqué la responsabilité de cet acte sur son compte Twitter, où il a posté des images d'Adolf Hitler. = ``` **Colon 5-shot** ``` Translate French to English French: Comment expliquer l'attitude contradictoire du gouvernement français, qui d'un coté s'offusque en public en convoquant l'ambassadeur des Etats-Unis le 21 octobre, et de l'autre interdit le survol du territoire par l'avion présidentiel bolivien, sur la base de la rumeur de la présence à son bord d'Edward Snowden ? English: Why the contradictory attitude of the French government? On the one hand, it publicly takes offence and summons the Ambassador of the United States on October 21 and, on the other, it forbids the Bolivian president's plane to enter its air space on the basis of a rumor that Edward Snowden was on board? French: Selon moi, il y a deux niveaux de réponse de la part du gouvernement français. English: In my opinion, there are two levels of response from the French government. French: Lorsque François Hollande téléphone à Barack Obama ou quand le ministre des affaires étrangères Laurent Fabius convoque l'ambassadeur des Etats-Unis, ils réagissent à une vraie découverte, qui est celle de l'ampleur de la surveillance américaine sur l'ensemble des communications en France. English: When François Hollande telephones Barack Obama, or when Foreign Minister Laurent Fabius summons the Ambassador of the United States, they are responding to a real discovery, that of the scale of America's surveillance of communications within France generally. French: N'est-il pas surprenant de lire dans les colonnes du Monde à quelques semaines d'intervalle d'une part la reproduction de la correspondance diplomatique américaine et d'autre part une condamnation des écoutes du Quai d'Orsay par la NSA ? English: And is it not surprising to read in the pages of Le Monde, on the one hand, a reproduction of diplomatic correspondence with the US and, on the other, condemnation of the NSA's spying on the Ministry of Foreign Affairs on the Quai d'Orsay, within a matter of weeks? French: N'y aurait-il pas comme une vague hypocrisie de votre part ? English: Is there not an element of hypocrisy on your part? French: Un homme de Cambridge a revendiqué la responsabilité de cet acte sur son compte Twitter, où il a posté des images d'Adolf Hitler. English: ``` **OpenAI 5-shot** ``` Translate French to English Comment expliquer l'attitude contradictoire du gouvernement français, qui d'un coté s'offusque en public en convoquant l'ambassadeur des Etats-Unis le 21 octobre, et de l'autre interdit le survol du territoire par l'avion présidentiel bolivien, sur la base de la rumeur de la présence à son bord d'Edward Snowden ? = Why the contradictory attitude of the French government? On the one hand, it publicly takes offence and summons the Ambassador of the United States on October 21 and, on the other, it forbids the Bolivian president's plane to enter its air space on the basis of a rumor that Edward Snowden was on board? Selon moi, il y a deux niveaux de réponse de la part du gouvernement français. = In my opinion, there are two levels of response from the French government. Lorsque François Hollande téléphone à Barack Obama ou quand le ministre des affaires étrangères Laurent Fabius convoque l'ambassadeur des Etats-Unis, ils réagissent à une vraie découverte, qui est celle de l'ampleur de la surveillance américaine sur l'ensemble des communications en France. = When François Hollande telephones Barack Obama, or when Foreign Minister Laurent Fabius summons the Ambassador of the United States, they are responding to a real discovery, that of the scale of America's surveillance of communications within France generally. N'est-il pas surprenant de lire dans les colonnes du Monde à quelques semaines d'intervalle d'une part la reproduction de la correspondance diplomatique américaine et d'autre part une condamnation des écoutes du Quai d'Orsay par la NSA ? = And is it not surprising to read in the pages of Le Monde, on the one hand, a reproduction of diplomatic correspondence with the US and, on the other, condemnation of the NSA's spying on the Ministry of Foreign Affairs on the Quai d'Orsay, within a matter of weeks? N'y aurait-il pas comme une vague hypocrisie de votre part ? = Is there not an element of hypocrisy on your part? Un homme de Cambridge a revendiqué la responsabilité de cet acte sur son compte Twitter, où il a posté des images d'Adolf Hitler. = ``` **Colon 10-shot** ``` Translate French to English French: Comment expliquer l'attitude contradictoire du gouvernement français, qui d'un coté s'offusque en public en convoquant l'ambassadeur des Etats-Unis le 21 octobre, et de l'autre interdit le survol du territoire par l'avion présidentiel bolivien, sur la base de la rumeur de la présence à son bord d'Edward Snowden ? English: Why the contradictory attitude of the French government? On the one hand, it publicly takes offence and summons the Ambassador of the United States on October 21 and, on the other, it forbids the Bolivian president's plane to enter its air space on the basis of a rumor that Edward Snowden was on board? French: Selon moi, il y a deux niveaux de réponse de la part du gouvernement français. English: In my opinion, there are two levels of response from the French government. French: Lorsque François Hollande téléphone à Barack Obama ou quand le ministre des affaires étrangères Laurent Fabius convoque l'ambassadeur des Etats-Unis, ils réagissent à une vraie découverte, qui est celle de l'ampleur de la surveillance américaine sur l'ensemble des communications en France. English: When François Hollande telephones Barack Obama, or when Foreign Minister Laurent Fabius summons the Ambassador of the United States, they are responding to a real discovery, that of the scale of America's surveillance of communications within France generally. French: N'est-il pas surprenant de lire dans les colonnes du Monde à quelques semaines d'intervalle d'une part la reproduction de la correspondance diplomatique américaine et d'autre part une condamnation des écoutes du Quai d'Orsay par la NSA ? English: And is it not surprising to read in the pages of Le Monde, on the one hand, a reproduction of diplomatic correspondence with the US and, on the other, condemnation of the NSA's spying on the Ministry of Foreign Affairs on the Quai d'Orsay, within a matter of weeks? French: N'y aurait-il pas comme une vague hypocrisie de votre part ? English: Is there not an element of hypocrisy on your part? French: La démarche journalistique n'est pas un positionnement moral, mais la recherche de l'intérêt et de la pertinence d'informations qui permettent à chaque citoyen de se forger une opinion. English: The journalistic method is not to adopt a moral position, but to investigate the significance and relevance of information and enable every citizen to form an opinion. French: Lorsque WikiLeaks lève le voile sur l'analyse par la diplomatie américaine d'enjeux politiques ou autres dans le monde entier, nous considérons en effet que, au regard de la puissance américaine, cela constitue un éclairage important. English: When WikiLeaks reveals the American administration's monitoring of political and other matters somewhere in the world, we consider this to be significant enlightenment with regard to the American government. French: Lorsque nous décrivons les systèmes d'interception américains à l'encontre de la diplomatie française aux Etats-Unis, ce n'est en aucun cas pour nous indigner de cette pratique, c'est pour décrire le monde tel qu'il est. English: In describing the American methods of data interception in relation to the French diplomatic representation in the United States, we do not aim at expressing indignation about this practice, but rather at describing the world as it is. French: La France a-t-elle bénéficié d'informations fournies par la NSA concernant des opérations terroristes visant nos intérêts ? English: Has France benefited from the intelligence supplied by the NSA concerning terrorist operations against our interests? French: Peut-on se priver de la collaboration américaine ? English: Can we do without collaboration with the Americans? French: Un homme de Cambridge a revendiqué la responsabilité de cet acte sur son compte Twitter, où il a posté des images d'Adolf Hitler. English: ``` **OpenAI 10-shot** ``` Translate French to English Comment expliquer l'attitude contradictoire du gouvernement français, qui d'un coté s'offusque en public en convoquant l'ambassadeur des Etats-Unis le 21 octobre, et de l'autre interdit le survol du territoire par l'avion présidentiel bolivien, sur la base de la rumeur de la présence à son bord d'Edward Snowden ? = Why the contradictory attitude of the French government? On the one hand, it publicly takes offence and summons the Ambassador of the United States on October 21 and, on the other, it forbids the Bolivian president's plane to enter its air space on the basis of a rumor that Edward Snowden was on board? Selon moi, il y a deux niveaux de réponse de la part du gouvernement français. = In my opinion, there are two levels of response from the French government. Lorsque François Hollande téléphone à Barack Obama ou quand le ministre des affaires étrangères Laurent Fabius convoque l'ambassadeur des Etats-Unis, ils réagissent à une vraie découverte, qui est celle de l'ampleur de la surveillance américaine sur l'ensemble des communications en France. = When François Hollande telephones Barack Obama, or when Foreign Minister Laurent Fabius summons the Ambassador of the United States, they are responding to a real discovery, that of the scale of America's surveillance of communications within France generally. N'est-il pas surprenant de lire dans les colonnes du Monde à quelques semaines d'intervalle d'une part la reproduction de la correspondance diplomatique américaine et d'autre part une condamnation des écoutes du Quai d'Orsay par la NSA ? = And is it not surprising to read in the pages of Le Monde, on the one hand, a reproduction of diplomatic correspondence with the US and, on the other, condemnation of the NSA's spying on the Ministry of Foreign Affairs on the Quai d'Orsay, within a matter of weeks? N'y aurait-il pas comme une vague hypocrisie de votre part ? = Is there not an element of hypocrisy on your part? La démarche journalistique n'est pas un positionnement moral, mais la recherche de l'intérêt et de la pertinence d'informations qui permettent à chaque citoyen de se forger une opinion. = The journalistic method is not to adopt a moral position, but to investigate the significance and relevance of information and enable every citizen to form an opinion. Lorsque WikiLeaks lève le voile sur l'analyse par la diplomatie américaine d'enjeux politiques ou autres dans le monde entier, nous considérons en effet que, au regard de la puissance américaine, cela constitue un éclairage important. = When WikiLeaks reveals the American administration's monitoring of political and other matters somewhere in the world, we consider this to be significant enlightenment with regard to the American government. Lorsque nous décrivons les systèmes d'interception américains à l'encontre de la diplomatie française aux Etats-Unis, ce n'est en aucun cas pour nous indigner de cette pratique, c'est pour décrire le monde tel qu'il est. = In describing the American methods of data interception in relation to the French diplomatic representation in the United States, we do not aim at expressing indignation about this practice, but rather at describing the world as it is. La France a-t-elle bénéficié d'informations fournies par la NSA concernant des opérations terroristes visant nos intérêts ? = Has France benefited from the intelligence supplied by the NSA concerning terrorist operations against our interests? Peut-on se priver de la collaboration américaine ? = Can we do without collaboration with the Americans? Un homme de Cambridge a revendiqué la responsabilité de cet acte sur son compte Twitter, où il a posté des images d'Adolf Hitler. = ``` **Colon 20-shot** ``` Translate French to English French: Comment expliquer l'attitude contradictoire du gouvernement français, qui d'un coté s'offusque en public en convoquant l'ambassadeur des Etats-Unis le 21 octobre, et de l'autre interdit le survol du territoire par l'avion présidentiel bolivien, sur la base de la rumeur de la présence à son bord d'Edward Snowden ? English: Why the contradictory attitude of the French government? On the one hand, it publicly takes offence and summons the Ambassador of the United States on October 21 and, on the other, it forbids the Bolivian president's plane to enter its air space on the basis of a rumor that Edward Snowden was on board? French: Selon moi, il y a deux niveaux de réponse de la part du gouvernement français. English: In my opinion, there are two levels of response from the French government. French: Lorsque François Hollande téléphone à Barack Obama ou quand le ministre des affaires étrangères Laurent Fabius convoque l'ambassadeur des Etats-Unis, ils réagissent à une vraie découverte, qui est celle de l'ampleur de la surveillance américaine sur l'ensemble des communications en France. English: When François Hollande telephones Barack Obama, or when Foreign Minister Laurent Fabius summons the Ambassador of the United States, they are responding to a real discovery, that of the scale of America's surveillance of communications within France generally. French: N'est-il pas surprenant de lire dans les colonnes du Monde à quelques semaines d'intervalle d'une part la reproduction de la correspondance diplomatique américaine et d'autre part une condamnation des écoutes du Quai d'Orsay par la NSA ? English: And is it not surprising to read in the pages of Le Monde, on the one hand, a reproduction of diplomatic correspondence with the US and, on the other, condemnation of the NSA's spying on the Ministry of Foreign Affairs on the Quai d'Orsay, within a matter of weeks? French: N'y aurait-il pas comme une vague hypocrisie de votre part ? English: Is there not an element of hypocrisy on your part? French: La démarche journalistique n'est pas un positionnement moral, mais la recherche de l'intérêt et de la pertinence d'informations qui permettent à chaque citoyen de se forger une opinion. English: The journalistic method is not to adopt a moral position, but to investigate the significance and relevance of information and enable every citizen to form an opinion. French: Lorsque WikiLeaks lève le voile sur l'analyse par la diplomatie américaine d'enjeux politiques ou autres dans le monde entier, nous considérons en effet que, au regard de la puissance américaine, cela constitue un éclairage important. English: When WikiLeaks reveals the American administration's monitoring of political and other matters somewhere in the world, we consider this to be significant enlightenment with regard to the American government. French: Lorsque nous décrivons les systèmes d'interception américains à l'encontre de la diplomatie française aux Etats-Unis, ce n'est en aucun cas pour nous indigner de cette pratique, c'est pour décrire le monde tel qu'il est. English: In describing the American methods of data interception in relation to the French diplomatic representation in the United States, we do not aim at expressing indignation about this practice, but rather at describing the world as it is. French: La France a-t-elle bénéficié d'informations fournies par la NSA concernant des opérations terroristes visant nos intérêts ? English: Has France benefited from the intelligence supplied by the NSA concerning terrorist operations against our interests? French: Peut-on se priver de la collaboration américaine ? English: Can we do without collaboration with the Americans? French: La mise en place depuis en gros dix ans d'outils technologiques d'interception très puissants par les Etats-Unis, mais aussi par la France, a officiellement été justifiée par la lutte contre le terrorisme. English: The setting up of high-performance interception technology over practically the past ten years by the United States - and by France - has been officially justified by the fight against terrorism. French: D'ailleurs, dans ce domaine, la France et les Etats-Unis notamment ont mis en place des procédures de coopération et d'échanges d'informations quasi quotidiens et qui sont décrits de part et d'autre comme essentiels. English: Furthermore, in this regard, France and the United States in particular have implemented procedures, sometimes described as essential, for cooperating and exchanging information on an almost daily basis. French: A titre d'exemple, la présence de Mohammed Merah dans les zones tribales à Miranshah a été signalée aux Français grâce aux moyens de la NSA. English: For example, France was informed of the presence of Mohammed Merah in the tribal areas of Miranshah through the NSA's resources. French: La France peut être conduite, par exemple, à transmettre des blocs entiers de données sur la région du Sahel aux services américains, et, en contrepartie - on l'a déjà rapidement dit -, les Américains peuvent donner des informations aux Français sur d'autres régions du monde. English: Also France may, for example, have to transmit entire blocks of data on the Sahel region to the Americans and, in return - as already briefly mentioned - the Americans may provide information to the French about other parts of the world. French: Donc la question de fond derrière cette affaire NSA n'est pas tant la capacité ou le droit des pays de se doter d'outils d'interception, que la question de l'absence totale de débat préalable, notamment au sein des Parlements, sur la justification de tels systèmes, le périmètre qui doit être le leur, et, en fin de compte, la question des atteintes aux libertés. English: Hence the question at the heart of the NSA affair is not so much the capacity or the right of a country to use interception tools, as the issue of the complete lack of prior debate - especially within parliaments - on the justification of such systems, the extent to which they should be used and, ultimately, the issue of the infringement of freedoms. French: Que risquent réellement les Etats-Unis ? une dégradation de leur image? English: What risk does the United States actually run? Ruining its image? French: On a beau les dénoncer, je ne vois pas de quelle manière ils pourront être punis. English: However much we denounce the US, I see no way in which it can be punished. French: Le risque couru par les Américains peut être double. English: The risk run by the Americans could be twofold. French: Le premier, c'est lorsque leurs alliés - et ça a été le cas récemment - apprennent que leurs dirigeants, parfois au plus haut sommet de leur Etat, ont été surveillés. English: The first is when their allies - as has been the case recently - learn that their governments have been spied on, sometimes at the highest level. French: C'est le cas du Brésil et de l'Allemagne, deux pays où les relations diplomatiques avec les Etats-Unis se sont tendues. English: This is the case in Brazil and Germany, two countries where diplomatic relations with the United States are strained. French: Un homme de Cambridge a revendiqué la responsabilité de cet acte sur son compte Twitter, où il a posté des images d'Adolf Hitler. English: ``` **OpenAI 20-shot** ``` Translate French to English Comment expliquer l'attitude contradictoire du gouvernement français, qui d'un coté s'offusque en public en convoquant l'ambassadeur des Etats-Unis le 21 octobre, et de l'autre interdit le survol du territoire par l'avion présidentiel bolivien, sur la base de la rumeur de la présence à son bord d'Edward Snowden ? = Why the contradictory attitude of the French government? On the one hand, it publicly takes offence and summons the Ambassador of the United States on October 21 and, on the other, it forbids the Bolivian president's plane to enter its air space on the basis of a rumor that Edward Snowden was on board? Selon moi, il y a deux niveaux de réponse de la part du gouvernement français. = In my opinion, there are two levels of response from the French government. Lorsque François Hollande téléphone à Barack Obama ou quand le ministre des affaires étrangères Laurent Fabius convoque l'ambassadeur des Etats-Unis, ils réagissent à une vraie découverte, qui est celle de l'ampleur de la surveillance américaine sur l'ensemble des communications en France. = When François Hollande telephones Barack Obama, or when Foreign Minister Laurent Fabius summons the Ambassador of the United States, they are responding to a real discovery, that of the scale of America's surveillance of communications within France generally. N'est-il pas surprenant de lire dans les colonnes du Monde à quelques semaines d'intervalle d'une part la reproduction de la correspondance diplomatique américaine et d'autre part une condamnation des écoutes du Quai d'Orsay par la NSA ? = And is it not surprising to read in the pages of Le Monde, on the one hand, a reproduction of diplomatic correspondence with the US and, on the other, condemnation of the NSA's spying on the Ministry of Foreign Affairs on the Quai d'Orsay, within a matter of weeks? N'y aurait-il pas comme une vague hypocrisie de votre part ? = Is there not an element of hypocrisy on your part? La démarche journalistique n'est pas un positionnement moral, mais la recherche de l'intérêt et de la pertinence d'informations qui permettent à chaque citoyen de se forger une opinion. = The journalistic method is not to adopt a moral position, but to investigate the significance and relevance of information and enable every citizen to form an opinion. Lorsque WikiLeaks lève le voile sur l'analyse par la diplomatie américaine d'enjeux politiques ou autres dans le monde entier, nous considérons en effet que, au regard de la puissance américaine, cela constitue un éclairage important. = When WikiLeaks reveals the American administration's monitoring of political and other matters somewhere in the world, we consider this to be significant enlightenment with regard to the American government. Lorsque nous décrivons les systèmes d'interception américains à l'encontre de la diplomatie française aux Etats-Unis, ce n'est en aucun cas pour nous indigner de cette pratique, c'est pour décrire le monde tel qu'il est. = In describing the American methods of data interception in relation to the French diplomatic representation in the United States, we do not aim at expressing indignation about this practice, but rather at describing the world as it is. La France a-t-elle bénéficié d'informations fournies par la NSA concernant des opérations terroristes visant nos intérêts ? = Has France benefited from the intelligence supplied by the NSA concerning terrorist operations against our interests? Peut-on se priver de la collaboration américaine ? = Can we do without collaboration with the Americans? La mise en place depuis en gros dix ans d'outils technologiques d'interception très puissants par les Etats-Unis, mais aussi par la France, a officiellement été justifiée par la lutte contre le terrorisme. = The setting up of high-performance interception technology over practically the past ten years by the United States - and by France - has been officially justified by the fight against terrorism. D'ailleurs, dans ce domaine, la France et les Etats-Unis notamment ont mis en place des procédures de coopération et d'échanges d'informations quasi quotidiens et qui sont décrits de part et d'autre comme essentiels. = Furthermore, in this regard, France and the United States in particular have implemented procedures, sometimes described as essential, for cooperating and exchanging information on an almost daily basis. A titre d'exemple, la présence de Mohammed Merah dans les zones tribales à Miranshah a été signalée aux Français grâce aux moyens de la NSA. = For example, France was informed of the presence of Mohammed Merah in the tribal areas of Miranshah through the NSA's resources. La France peut être conduite, par exemple, à transmettre des blocs entiers de données sur la région du Sahel aux services américains, et, en contrepartie - on l'a déjà rapidement dit -, les Américains peuvent donner des informations aux Français sur d'autres régions du monde. = Also France may, for example, have to transmit entire blocks of data on the Sahel region to the Americans and, in return - as already briefly mentioned - the Americans may provide information to the French about other parts of the world. Donc la question de fond derrière cette affaire NSA n'est pas tant la capacité ou le droit des pays de se doter d'outils d'interception, que la question de l'absence totale de débat préalable, notamment au sein des Parlements, sur la justification de tels systèmes, le périmètre qui doit être le leur, et, en fin de compte, la question des atteintes aux libertés. = Hence the question at the heart of the NSA affair is not so much the capacity or the right of a country to use interception tools, as the issue of the complete lack of prior debate - especially within parliaments - on the justification of such systems, the extent to which they should be used and, ultimately, the issue of the infringement of freedoms. Que risquent réellement les Etats-Unis ? une dégradation de leur image? = What risk does the United States actually run? Ruining its image? On a beau les dénoncer, je ne vois pas de quelle manière ils pourront être punis. = However much we denounce the US, I see no way in which it can be punished. Le risque couru par les Américains peut être double. = The risk run by the Americans could be twofold. Le premier, c'est lorsque leurs alliés - et ça a été le cas récemment - apprennent que leurs dirigeants, parfois au plus haut sommet de leur Etat, ont été surveillés. = The first is when their allies - as has been the case recently - learn that their governments have been spied on, sometimes at the highest level. C'est le cas du Brésil et de l'Allemagne, deux pays où les relations diplomatiques avec les Etats-Unis se sont tendues. = This is the case in Brazil and Germany, two countries where diplomatic relations with the United States are strained. Un homme de Cambridge a revendiqué la responsabilité de cet acte sur son compte Twitter, où il a posté des images d'Adolf Hitler. = ``` **English target** ``` A man from Cambridge claimed responsibility for the act on his Twitter account, where he posted pictures of Adolf Hitler. ``` Results: #### Likelihood correct - Control | Engine | Empty prior | Control | | --- | --- | --- | | ada | -73.354 | -45.109 | | babbage | -70.899 | -35.969 | | curie | -70.224 | -30.489 | | davinci | -67.642 | -22.543 | #### Likelihood correct - Simple colon prompt | Engine | 0-shot | 1-shot | 2-shot | 5-shot | 10-shot | 20-shot | | --- | --- | --- | --- | --- | --- | --- | | ada | -19.658 | -17.257 | -16.519 | -16.112 | -15.741 | -16.309 | | babbage | -10.467 | -11.404 | -11.574 | -12.385 | -13.134 | -11.462 | | curie | -9.395 | -8.763 | -8.561 | -8.865 | -8.701 | -7.701 | | davinci | -6.443 | -6.658 | -6.614 | -6.807 | -7.388 | -6.387 | #### Likelihood correct - OpenAI prompt | Engine | 0-shot | 1-shot | 2-shot | 5-shot | 10-shot | 20-shot | | --- | --- | --- | --- | --- | --- | --- | | ada | -27.656 | -17.372 | -16.043 | -15.672 | -16.115 | -15.877 | | babbage | -11.877 | -12.015 | -12.205 | -12.868 | -13.800 | -11.765 | | curie | -13.414 | -8.735 | -8.841 | -8.462 | -8.685 | -7.147 | | davinci | -7.271 | -6.549 | -6.532 | -6.805 | -7.392 | -6.544 | ![Likelihood correct - French/English](/french_english.png) *Empty prior is plotted at -2, control prior at -1. Did not plot 20 shot; is was not substantially different from 10-shot for any of the models* Observations: * Unlike in the Roish example, for *all* models, and for both prompts, the change from the control prior to 0-shot is a much bigger change than from 0- to 1-shot. * The OpenAI prompt and simple colon prompt have similar scores for everything except for 0-shot, where the simple colon prompt provides better evidence, especially for `ada`. * For the simple colon prompt, all the models except `ada`, which gleaned some evidence from 1-shot, examples provided no evidence or *negative* evidence. `ada` and `curie` gleaned more evidence from 1-shot for OpenAI’s prompt due to having derived comparatively *less* evidence from the 0-shot prompt. * For all the models, the transition from the empty prior (no prompt) to control (prompt is French sentence without directions to translate) provided more evidence for the correct answer than any instruction or examples. This makes sense – the most significant evidence for the correct English translation is not actually the instructions to translate, but the French sentence itself. Now for the decibel breakdown: \\--- #### ada | Prompt | Correct likelihood | +dB to empty prior | +dB to control | +dB to 0-shot | | --- | --- | --- | --- | --- | | empty prior | -73.354 | - | - | - | | control | -45.109 | 28.245 | - | - | | colon 0-shot | -19.658 | 53.696 | *25.451* | - | | colon 1-shot | -17.257 | 56.097 | 27.852 | 2.401 | | colon 2-shot | -16.519 | 56.835 | 28.590 | 3.139 | | colon 5-shot | -16.112 | 57.243 | 28.998 | 3.547 | | colon 10-shot | -15.741 | 57.614 | 29.368 | 3.917 | | colon 20-shot | -16.309 | 57.045 | 28.800 | *3.349* | | OAI 0-shot | -27.656 | 45.699 | *17.454* | - | | OAI 1-shot | -17.372 | 55.983 | 27.738 | 10.284 | | OAI 2-shot | -16.043 | 57.311 | 29.066 | 11.613 | | OAI 5-shot | -15.672 | 57.682 | 29.437 | 11.983 | | OAI 10-shot | -16.115 | 57.239 | 28.994 | 11.541 | | OAI 20-shot | -15.877 | 57.477 | 29.232 | *11.778* | **Simple colon +dB from 20-shot / +dB from 0-shot** = 3.349 / 25.451 = **0.132** **OpenAI +dB from 20-shot / +dB from 0-shot** = 11.778 / 17.454 = **0.675** \\--- #### babbage | Prompt | Correct likelihood | +dB to empty prior | +dB to control | +dB to 0-shot | | --- | --- | --- | --- | --- | | empty prior | -70.899 | - | - | - | | control | -35.969 | 34.930 | - | - | | colon 0-shot | -10.467 | 60.432 | *25.502* | - | | colon 1-shot | -11.404 | 59.495 | 24.565 | -0.937 | | colon 2-shot | -11.574 | 59.326 | 24.396 | -1.106 | | colon 5-shot | -12.385 | 58.514 | 23.584 | -1.918 | | colon 10-shot | -13.134 | 57.765 | 22.835 | -2.666 | | colon 20-shot | -11.462 | 59.438 | 24.508 | *-0.994* | | OAI 0-shot | -11.877 | 59.022 | *24.021* | - | | OAI 1-shot | -12.015 | 58.884 | 23.954 | -0.139 | | OAI 2-shot | -12.205 | 58.694 | 23.758 | -0.329 | | OAI 5-shot | -12.847 | 58.052 | 23.122 | -0.971 | | OAI 10-shot | -13.800 | 57.100 | 22.170 | -1.923 | | OAI 20-shot | -11.765 | 59.134 | 24.167 | *0.111* | **Simple colon +dB from 20-shot / +dB from 0-shot** = -0.994 / 25.502 = **-0.039** **OpenAI +dB from 20-shot / +dB from 0-shot** = 0.111 / 24.021 = **0.005** \\--- #### curie | Prompt | Correct likelihood | +dB to empty prior | +dB to control | +dB to 0-shot | | --- | --- | --- | --- | --- | | empty prior | -70.224 | - | - | - | | control | -30.489 | 39.735 | - | - | | colon 0-shot | -9.395 | 60.829 | *21.094* | - | | colon 1-shot | -8.763 | 61.461 | 21.726 | 0.632 | | colon 2-shot | -8.561 | 61.663 | 21.928 | 0.834 | | colon 5-shot | -8.865 | 61.359 | 21.624 | 0.530 | | colon 10-shot | -8.701 | 61.523 | 21.743 | 0.694 | | colon 20-shot | -7.701 | 62.524 | 22.744 | *1.695* | | OAI 0-shot | -13.414 | 56.810 | *17.075* | - | | OAI 1-shot | -8.735 | 61.489 | 21.754 | 4.679 | | OAI 2-shot | -8.841 | 61.383 | 21.648 | 4.573 | | OAI 5-shot | -8.462 | 61.762 | 22.027 | 4.952 | | OAI 10-shot | -8.685 | 61.539 | 21.804 | 4.729 | | OAI 20-shot | -7.147 | 63.077 | 23.342 | *6.267* | **Simple colon +dB from 20-shot / +dB from 0-shot** = 1.695 / 21.094 = **0.080** **OpenAI +dB from 20-shot / +dB from 0-shot** = 6.267 / 17.075 = **0.367** \\--- #### davinci | Prompt | Correct likelihood | +dB to empty prior | +dB to control | +dB to 0-shot | | --- | --- | --- | --- | --- | | empty prior | -67.642 | - | - | - | | control | -22.543 | 45.099 | - | - | | colon 0-shot | -6.443 | 61.199 | *16.100* | - | | colon 1-shot | -6.658 | 60.984 | 15.884 | -0.215 | | colon 2-shot | -6.614 | 61.028 | 15.911 | -0.171 | | colon 5-shot | -6.807 | 60.835 | 15.749 | -0.364 | | colon 10-shot | -7.388 | 60.254 | 15.167 | -0.945 | | colon 20-shot | -6.387 | 61.255 | 16.169 | *0.056* | | OAI 0-shot | -7.286 | 60.368 | *15.269* | - | | OAI 1-shot | -6.567 | 61.087 | 15.975 | 0.719 | | OAI 2-shot | -6.506 | 61.148 | 16.093 | 0.779 | | OAI 5-shot | -6.797 | 60.857 | 15.835 | 0.489 | | OAI 10-shot | -7.392 | 60.262 | 15.184 | -0.106 | | OAI 20-shot | -6.528 | 61.126 | 15.966 | *0.757* | **Simple colon +dB from 20-shot / +dB from 0-shot** = 0.056 / 16.100 = **0.003** **OpenAI +dB from 20-shot / +dB from 0-shot** = 0.757 / 15.269 = **0.050** \\--- We see again that the more powerful models derive proportionally more evidence from the 0-shot task description than from subsequent examples, although compared to Roish, *all* the models derive a lot more evidence from the 0-shot description than from examples (the ratio is consistently < 1, whereas for Roish it was consistently > 1). These two tasks, one which uses an invented language and one which uses a real language that GPT-3 knows well, derive evidence from 0- and few-shot prompts in different proportions. When GPT-3 already knows how to perform a task, examples are less helpful. Using the simple colon prompt, none of the models except for `ada` appear to have much use for examples in terms of the likelihood of giving the correct translation. OpenAI’s prompt results in worse 0-shot accuracy for `ada` and `curie`, and accordingly the presence of one example is helpful for those two models, but any additional examples have little effect. This supports the hypothesis that GPT-3 is not “learning” translation from the few-shot examples but rather locating the already-learned task, and that the lower BLEU scores for 0-shot prompts can probably be attributed to catastrophic failures. \\--- Counterexample - accuracy decreases with # shots \\------------------------------------------------ *Main article: **[List sorting does not play well with few-shot](/posts/list-sorting-does-not-play-well-with-few-shot/)*** A interesting counterexample to few-shot monotonicity is list sorting (e.g. [1, 3, 3, 4, 0] -> [0, 1, 3, 3, 4]). A zero-shot prompt which frames the problem as an example embedded in coding documentation achieves 76% accuracy, whereas a 32-shot prompt achieves only 20%. Adding examples to the end of the coding prompt results in a monotonic *decrease* in accuracy with number of shots. 0-shot code prompt: ``` The sort function can be used to sort a list in ascending, descending or user defined order. To sort the list in ascending order, call list.sort(). This will sort a list of integers so that the smallest integer will be first in the list and the largest integer will be the last and the last element of the list is the largest.\" For example: list = [1, 0, 4, 3, 3] list.sort() = ``` Example-only prompt: ``` Unsorted list: [5, 6, 2, 3, 2] Sorted list: [2, 2, 3, 5, 6] Unsorted list: [8, 5, 8, 8, 4] Sorted list: [4, 5, 8, 8, 8] ... Unsorted list: [1, 0, 4, 3, 3] Sorted list: ``` **Code preprompt, length 5** | Shots | Correct | Accuracy | | --- | --- | --- | | **0** | **38/50** | **0.76** | | 1 | 33/50 | 0.66 | | 3 | 23/50 | 0.46 | | 5 | 22/50 | 0.44 | | 7 | 22/50 | 0.44 | | 10 | 21/50 | 0.42 | | 13 | 15/50 | 0.30 | | 16 | 16/50 | 0.32 | **No preprompt, length 5** | Shots | Correct | Accuracy | | --- | --- | --- | | 0 | 14/50 | 0.28 | | **1** | **20/50** | **0.40** | | 3 | 15/50 | 0.30 | | 5 | 14/50 | 0.28 | | 7 | 16/50 | 0.32 | | **10** | **25/50** | **0.50** | | 13 | 18/50 | 0.36 | | 16 | 11/50 | 0.22 | | 32 | 10/50 | 0.20 | ![list sorting results](/sorting/interesting2.png) *comparison of accuracies from 0 - 16 shots* Pretty weird, right? I speculate on the causes of these unexpectedly shaped plots in my [post](/posts/list-sorting-does-not-play-well-with-few-shot/) about this experiment. Regardless of the mechanism, though, the fact that having more examples in the prompt makes GPT-3 worse at the task implies that there are more factors in play than simply metalearning. List sorting is the only quantitative task that I know to exhibit strict decrease in accuracy with number of shots, but I suspect there are many more. \\--- Predictions \\----------- I will offer some specific predictions suggested by my interpretation of these results, which may be corroborated or falsified by further experiments and future developments. 1\\. For translation tasks, if catastrophic failures are filtered out, the discrepancy between 0-shot and 64-shot performance will disappear. 1\\. 1-shot performance, however, will be worse than both 0-shot and 64-shot performance even after catastrophic failures are filtered out, due to being most vulnerable to [few-shot bugs](/posts/methods-of-prompt-programming/#few-shot-bugs). 2\\. For practically all tasks that can be meaningfully described using a 0-shot prompt, the ratio of **+ dB from few-shot over 0-shot / +dB from 0-shot over prior** will tend to be smaller for larger models than smaller models, indicating that larger models glean more evidence from 0-shot task descriptions compared to examples. 3\\. There are tasks for which a zero-shot prompt causes GPT-3 to perform better than *any* conventional[3](#fn:3) few-shot prompt. [Sorting lists of five integers](/posts/list-sorting-does-not-play-well-with-few-shot/) is probably one such task. 4\\. More powerful language models released in the future will continue to become more effective at 0-shot “learning,” and this capability will increase even more dramatically than improvement at few-shot learning. \\--- 1\\. This is probably the situation for many zero-shot translation tasks with inadequate task specification. [↩︎](#fnref:1) 2\\. Of course, the representativeness of examples is always a concern when evaluating language models. [↩︎](#fnref:2) 3\\. Conventional = a list of solved examples, optionally following a task description, excluding complex embeddings of examples e.g. in a narrative. [↩︎](#fnref:3) \\--- [GPT-3](/categories/gpt-3/)[prompt engineering](/categories/prompt-engineering/)[benchmarks](/categories/benchmarks/)[bayes](/categories/bayes/)[interpretability](/categories/interpretability/) 10831 Words Feb 10, 2021 \\--- [← List sorting does not play well with few-shot](/posts/list-sorting-does-not-play-well-with-few-shot/) [Loom: interface to the multiverse →](/posts/loom-interface-to-the-multiverse/)© 2022 [moire](/about)$(function(){window.sr=ScrollReveal();if($(window).width()<768){if($('.timeline-content').hasClass('js--fadeInLeft')){$('.timeline-content').removeClass('js--fadeInLeft').addClass('js--fadeInRight');} sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});}else{sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});} sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});});", "source": "generative.ink", "date_published": "n/a", "id": "7335a9fe4a871f1d001ae73d871d7f55"}
+{"text": "Loom: interface to the multiverse :: — Moire[![](/images/home/rolling_phase.gif)](/)* [Posts](/posts/) * [Trees](/trees/) * [Prophecies](/prophecies/) * [About](/about/) 4 minutes [Loom: interface to the multiverse](/posts/loom-interface-to-the-multiverse/) ============================================================================= \\--- Table of Contents* [Motivation for the loom](#motivation-for-the-loom) * [Demo](#demo) \\+ [Read mode](#read-mode) \\+ [Visualize mode](#visualize-mode) \\+ [Node metadata](#node-metadata) \\--- code: [github.com/socketteer/loom](https://github.com/socketteer/loom) paper: [Multiversal views on language models](https://arxiv.org/abs/2102.06391) \\--- Motivation for the loom \\----------------------- > Differing from Newton and Schopenhauer … He believed in an infinite series of times, in a dizzily growing, ever spreading network… This web of time - the strands of which approach one another, bifurcate, intersect or ignore each other through the centuries - embraces every possibility. > > – The Garden of Forking Paths > > I experienced GPT-3 first through [AI Dungeon's](https://play.aidungeon.io/) interface, and like many, I was immediately captivated by the depth, breadth, profundity and - especially given purposeful administration on the part of the user - long-range coherence of the worlds that could be conjured. On these early walks through infinite virtual reality, I knew that everything was forever changed, and at the same time I knew that I was accessing only a fraction of the potential that seethed within the virtual reality generator. AI Dungeon has several features helpful for writing longform works, such as pinned memory, World Info, and automatic summarization. However, it is missing support for a feature which I've found to be a great power multiplier for co-writing with language models: **branching**. Adventures on AI Dungeon are single-history: you can edit actions and retry generation, but only one variation can be saved to the adventure. You can get around this by duplicating the adventure, but there's no representation of the fact that the copies are forked from the same point, so branching more than a few times leads to a confusing profusion of adventures. > For a novice weaver, even the slightest change can cause ripples that cascade into an infinity of nightmares. It is recommended that those studying the Loom stop living in linear time and begin thinking in terms of Multiverses … > > – [Weaving the Moment with the Loom of Time: an instruction manual for the would-be weaver](/loom/toc/) > > A short time of writing with AI Dungeon demonstrated to me that [language models are multiverse generators](/posts/language-models-are-multiverse-generators/), and that an interface which supports branching would provide great utility for cowriting with a language model, whether one is composing virtual worlds intended to be multi- or single-history (in the latter case, branching explores counterfactuals which can be merged into main), as well as for nonfiction tasks like brainstorming and prototyping prompts. After I was given access to OpenAI's API, motivated to actualize my [HITL thought experiment](/posts/hitl-thought-experiment/), I began making an interface which allows the user to generate, navigate, and visualize natural language multiverses. I named it *loom*, after a device that features in the metaphysics of a multiversal novel which I coauthor. > You return to weaving patterns of creation and destruction around you, warp and weft crossing and recrossing as strands of destiny interlock. You spin, fingertips brushing lightly against the surface of reality, parting the flow before you, teasing diffractal shapes out of the patternless void. Corridors of possibility bloom like time-lapse flowers in your wake and burst like mineshafts into nothingness again. But for every one of these there are a far greater number of voids–futures which your mind refuses to touch. Your Loom of Time devours the boundary conditions of the present and traces a garment of glistening cobwebs over the still-forming future, teasing through your fingers and billowing out towards the shadowy unknown like an incoming tide. > > \"Real time is just an Arbitrage-adapted interface to the Loom Space,\" you explain. \"We prune unnecessary branches from the World Tree and weave together the timelines into one coherent history. The story is trying to become aware of itself, and it does so through us.\" > > ![](/alchemicalmarriage/lt4.png) *`the Loom of Time devours the present and traces a garment of glistening cobwebs over the still-forming future`, illustration by [BigSleep](https://github.com/lucidrains/big-sleep)* I use the loom for all my GPT-3-assisted writing, as well as for brainstorming and prototyping prompts. I created it for personal use, so the interface hasn't been made user-friendly and lacks documentation ([this](/loom/toc/) doesn't count). However, anyone with an OpenAI API key is welcome to beta test it. Demo \\---- ### Read mode ![read - darkmode](/loom/read-darkmode.png) In read mode, the current node and its ancestry are read as a single history. The tree structure of nodes and chapters is displayed in the nav tree sidebar. ### Visualize mode ![visualize - darkmode](/loom/tree-darkmode.png) Visualize mode renders the multiverse as a tree diagram. #### editing #### expanding and collapsing nodes #### expanding and collapsing children, collapse all except node + subtree #### creating new children and siblings, merge parent #### generating GPT-3 continuations #### zooming out to view global multiverse structure #### weighted stochastic walks through a large multiverse ### Node metadata ![node metadata](/loom/metadata-light.png) Prompt, response, model, token logprobs and counterfacual logprobs are saved to the metadata dictionary of generated nodes. The node metadata dialogue displays this information. \\--- [prototypes](/categories/prototypes/)[GPT-3](/categories/gpt-3/)[HITL](/categories/hitl/) 767 Words Feb 9, 2021 \\--- [← Language models are 0-shot interpreters](/posts/language-models-are-0-shot-interpreters/) [This Museum Does Not Exist: GPT-3 x CLIP →](/posts/this-museum-does-not-exist-gpt-3-x-clip/)© 2022 [moire](/about)$(function(){window.sr=ScrollReveal();if($(window).width()<768){if($('.timeline-content').hasClass('js--fadeInLeft')){$('.timeline-content').removeClass('js--fadeInLeft').addClass('js--fadeInRight');} sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});}else{sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});} sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});});\n", "url": "https://generative.ink/posts/loom-interface-to-the-multiverse/", "title": "Loom: interface to the multiverse :: — Moire[![](/images/home/rolling_phase.gif)](/)* [Posts](/posts/) * [Trees](/trees/) * [Prophecies](/prophecies/) * [About](/about/) 4 minutes [Loom: interface to the multiverse](/posts/loom-interface-to-the-multiverse/) ============================================================================= \\--- Table of Contents* [Motivation for the loom](#motivation-for-the-loom) * [Demo](#demo) \\+ [Read mode](#read-mode) \\+ [Visualize mode](#visualize-mode) \\+ [Node metadata](#node-metadata) \\--- code: [github.com/socketteer/loom](https://github.com/socketteer/loom) paper: [Multiversal views on language models](https://arxiv.org/abs/2102.06391) \\--- Motivation for the loom \\----------------------- > Differing from Newton and Schopenhauer … He believed in an infinite series of times, in a dizzily growing, ever spreading network… This web of time - the strands of which approach one another, bifurcate, intersect or ignore each other through the centuries - embraces every possibility. > > – The Garden of Forking Paths > > I experienced GPT-3 first through [AI Dungeon’s](https://play.aidungeon.io/) interface, and like many, I was immediately captivated by the depth, breadth, profundity and - especially given purposeful administration on the part of the user - long-range coherence of the worlds that could be conjured. On these early walks through infinite virtual reality, I knew that everything was forever changed, and at the same time I knew that I was accessing only a fraction of the potential that seethed within the virtual reality generator. AI Dungeon has several features helpful for writing longform works, such as pinned memory, World Info, and automatic summarization. However, it is missing support for a feature which I’ve found to be a great power multiplier for co-writing with language models: **branching**. Adventures on AI Dungeon are single-history: you can edit actions and retry generation, but only one variation can be saved to the adventure. You can get around this by duplicating the adventure, but there’s no representation of the fact that the copies are forked from the same point, so branching more than a few times leads to a confusing profusion of adventures. > For a novice weaver, even the slightest change can cause ripples that cascade into an infinity of nightmares. It is recommended that those studying the Loom stop living in linear time and begin thinking in terms of Multiverses … > > – [Weaving the Moment with the Loom of Time: an instruction manual for the would-be weaver](/loom/toc/) > > A short time of writing with AI Dungeon demonstrated to me that [language models are multiverse generators](/posts/language-models-are-multiverse-generators/), and that an interface which supports branching would provide great utility for cowriting with a language model, whether one is composing virtual worlds intended to be multi- or single-history (in the latter case, branching explores counterfactuals which can be merged into main), as well as for nonfiction tasks like brainstorming and prototyping prompts. After I was given access to OpenAI’s API, motivated to actualize my [HITL thought experiment](/posts/hitl-thought-experiment/), I began making an interface which allows the user to generate, navigate, and visualize natural language multiverses. I named it *loom*, after a device that features in the metaphysics of a multiversal novel which I coauthor. > You return to weaving patterns of creation and destruction around you, warp and weft crossing and recrossing as strands of destiny interlock. You spin, fingertips brushing lightly against the surface of reality, parting the flow before you, teasing diffractal shapes out of the patternless void. Corridors of possibility bloom like time-lapse flowers in your wake and burst like mineshafts into nothingness again. But for every one of these there are a far greater number of voids–futures which your mind refuses to touch. Your Loom of Time devours the boundary conditions of the present and traces a garment of glistening cobwebs over the still-forming future, teasing through your fingers and billowing out towards the shadowy unknown like an incoming tide. > > “Real time is just an Arbitrage-adapted interface to the Loom Space,” you explain. “We prune unnecessary branches from the World Tree and weave together the timelines into one coherent history. The story is trying to become aware of itself, and it does so through us.” > > ![](/alchemicalmarriage/lt4.png) *`the Loom of Time devours the present and traces a garment of glistening cobwebs over the still-forming future`, illustration by [BigSleep](https://github.com/lucidrains/big-sleep)* I use the loom for all my GPT-3-assisted writing, as well as for brainstorming and prototyping prompts. I created it for personal use, so the interface hasn’t been made user-friendly and lacks documentation ([this](/loom/toc/) doesn’t count). However, anyone with an OpenAI API key is welcome to beta test it. Demo \\---- ### Read mode ![read - darkmode](/loom/read-darkmode.png) In read mode, the current node and its ancestry are read as a single history. The tree structure of nodes and chapters is displayed in the nav tree sidebar. ### Visualize mode ![visualize - darkmode](/loom/tree-darkmode.png) Visualize mode renders the multiverse as a tree diagram. #### editing #### expanding and collapsing nodes #### expanding and collapsing children, collapse all except node + subtree #### creating new children and siblings, merge parent #### generating GPT-3 continuations #### zooming out to view global multiverse structure #### weighted stochastic walks through a large multiverse ### Node metadata ![node metadata](/loom/metadata-light.png) Prompt, response, model, token logprobs and counterfacual logprobs are saved to the metadata dictionary of generated nodes. The node metadata dialogue displays this information. \\--- [prototypes](/categories/prototypes/)[GPT-3](/categories/gpt-3/)[HITL](/categories/hitl/) 767 Words Feb 9, 2021 \\--- [← Language models are 0-shot interpreters](/posts/language-models-are-0-shot-interpreters/) [This Museum Does Not Exist: GPT-3 x CLIP →](/posts/this-museum-does-not-exist-gpt-3-x-clip/)© 2022 [moire](/about)$(function(){window.sr=ScrollReveal();if($(window).width()<768){if($('.timeline-content').hasClass('js--fadeInLeft')){$('.timeline-content').removeClass('js--fadeInLeft').addClass('js--fadeInRight');} sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});}else{sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});} sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});});", "source": "generative.ink", "date_published": "n/a", "id": "472576bed90ae82a3d73d2cfac8db98b"}
+{"text": "This Museum Does Not Exist: GPT-3 x CLIP :: — Moire[![](/images/home/rolling_phase.gif)](/)* [Posts](/posts/) * [Trees](/trees/) * [Prophecies](/prophecies/) * [About](/about/) 2 minutes [This Museum Does Not Exist: GPT-3 x CLIP](/posts/this-museum-does-not-exist-gpt-3-x-clip/) =========================================================================================== \\--- Table of Contents* [Gallery I](#gallery-i) \\+ [`The Death of Archimedes`](#the-death-of-archimedes) \\+ [`Still Life with Mirror`](#still-life-with-mirror) \\+ [`The Poet's Abbreviated Life`](#the-poets-abbreviated-life) \\+ [`Narcissus`](#narcissus) \\+ [`Dream of the Last Supper`](#dream-of-the-last-supper) * [Gallery II](#gallery-ii) \\+ [`The Coffin of Salvador Dali`](#the-coffin-of-salvador-dali) \\+ [`The Beautiful Bird Delirium`](#the-beautiful-bird-delirium) \\+ [`The Domain of Unimaginable Horror`](#the-domain-of-unimaginable-horror) \\+ [`The Spectral Museum`](#the-spectral-museum) \\+ [`The Metaphysics of Transvestism`](#the-metaphysics-of-transvestism) * [Gallery III](#gallery-iii) \\+ [`The Children of Marx and Coca-Cola`](#the-children-of-marx-and-coca-cola) \\+ [`Man and Bottle`](#man-and-bottle) \\+ [`The Man Who Invented the Photography of the Absurd`](#the-man-who-invented-the-photography-of-the-absurd) \\+ [`Playing in the Graveyard of Avant-Garde`](#playing-in-the-graveyard-of-avant-garde) * [Gallery IV](#gallery-iv) \\+ [`Cranial Extraction of the Baby Alien`](#cranial-extraction-of-the-baby-alien) \\+ [`The Euphoria of the Fish`](#the-euphoria-of-the-fish) \\+ [`Waste of Shame`](#waste-of-shame) \\+ [`The Dream of the Butterfly-Impaled Schoolgirl`](#the-dream-of-the-butterfly-impaled-schoolgirl) * [Gallery V](#gallery-v) \\+ [`The Sleep of Reason Produces Monsters`](#the-sleep-of-reason-produces-monsters) \\+ [`The Failure of Astronauts`](#the-failure-of-astronauts) \\+ [`Washerwoman of the Wind`](#washerwoman-of-the-wind) \\+ [`The Death of the Lonesome Astronomer`](#the-death-of-the-lonesome-astronomer) * [Gallery VI: series](#gallery-vi-series) \\+ [`The Pathological Museum in the Forest`](#the-pathological-museum-in-the-forest) \\+ [`Variations on Narcissus`](#variations-on-narcissus) \\+ [`The Tragic Intimacy of the Eternal Conversation With Oneself`](#the-tragic-intimacy-of-the-eternal-conversation-with-oneself) \\--- I had GPT-3 generate painting titles (credit to @nmkd of EleutherAI for the idea), beginning with the prompt ``` The hall was lined with an infinite number of paintings, each more surreal and mysterious than the last. The first painting is named \"Persistence of Memory.\" It depicts a surreal landscape with melted clocks draped over strange objects. The next painting is named \" ``` After this prompt yielded several intriguing titles, I switched to prompt format which put the titles in a list: ``` 100 surreal and mysterious painting names: Persistence of Memory The Great Masturbator Boot Poem to the Sun The Man Who Envied Cephalopods The Sleep of Reason Produces Monsters Washerwoman of the Wind Man and Bottle Spectrum The Disintegration of the Persians The Great Masturbator, Part II Still Life with Mirror Bouquet of Enigmatic Beauties A Grudge The Premonition of Civil Violence ``` The reason I didn't use a list immediately is because this format is liable to derail or be repetitive if there aren't many examples, and the narrative context also encouraged less generic and more interesting results. Once I had a few examples representing the sort of varied weirdness I was looking for, I was able to switch over to the more convenient list format. Now I cultivate a multiverse of surreal and mysterious painting titles using the [loom](/posts/loom-interface-to-the-multiverse/), and periodically harvest the most interesting ones to promote to graphic actuality via BigSleep. ![painting multiverse](/museum/multiverse.png) *a small subtree of the multiverse of surreal and mysterious paintings* Images were generated using [BigSleep](https://github.com/lucidrains/big-sleep) via [The Big Sleep Customized NMKD Public](https://colab.research.google.com/drive/1Q2DIeMqYm_Sc5mlurnnurMMVqlgXpZNO?usp=sharing) colab notebook. \\--- Gallery I \\--------- ### `The Death of Archimedes` ![The Death of Archimedes](/museum/arch.png) ### `Still Life with Mirror` ![Still Life with Mirror](/museum/slm.png) ### `The Poet's Abbreviated Life` ![The Poet's Abbreviated Life](/museum/pab.png) ### `Narcissus` ![Narcissus](/museum/narcissus2.png) ### `Dream of the Last Supper` ![Dream of the Last Supper](/museum/lastsupper.png) Gallery II \\---------- ### `The Coffin of Salvador Dali` ![The Coffin of Salvador Dali](/museum/coffin.png) ### `The Beautiful Bird Delirium` ![The Beautiful Bird Delirium](/museum/bbd2.png) ### `The Domain of Unimaginable Horror` ![The Domain of Unimaginable Horror](/museum/doh.png) ### `The Spectral Museum` ![The Spectral Museum](/museum/tsm.png) ### `The Metaphysics of Transvestism` ![The Metaphysics of Transvestism](/museum/metatrans.png) Gallery III \\----------- ### `The Children of Marx and Coca-Cola` ![The Children of Marx and Coca-Cola](/museum/mcc.png) ### `Man and Bottle` ![Man and Bottle](/museum/manbottle.png) ### `The Man Who Invented the Photography of the Absurd` ![The Man Who Invented the Photography of the Absurd](/museum/absurd.png) ### `Playing in the Graveyard of Avant-Garde` ![Playing in the Graveyard of Avant-Garde](/museum/pga.png) Gallery IV \\---------- ### `Cranial Extraction of the Baby Alien` ![Cranial Extraction of the Baby Alien](/museum/cba.png) ### `The Euphoria of the Fish` ![The Euphoria of the Fish](/museum/fish.png) ### `Waste of Shame` ![Waste of Shame](/museum/waste.png) ### `The Dream of the Butterfly-Impaled Schoolgirl` ![The Dream of the Butterfly-Impaled Schoolgirl](/museum/butterfly.png) Gallery V \\--------- ### `The Sleep of Reason Produces Monsters` ![The Sleep of Reason Produces Monsters](/museum/sleep.png) ### `The Failure of Astronauts` ![The Failure of Astronauts](/museum/fa.png) ### `Washerwoman of the Wind` ![Washerwoman of the Wind](/museum/washerwoman.png) ### `The Death of the Lonesome Astronomer` ![The Death of the Lonesome Astronomer](/museum/DLA.png) Gallery VI: series \\------------------ ### `The Pathological Museum in the Forest` ![The Pathological Museum in the Forest](/museum/tpm.png) ![The Pathological Museum in the Forest](/museum/tpm2.png) ### `Variations on Narcissus` ![Variations on Narcissus](/museum/variations3.png) ![Variations on Narcissus](/museum/variations1.png) ![Variations on Narcissus](/museum/variations2.png) ### `The Tragic Intimacy of the Eternal Conversation With Oneself` ![The Tragic Intimacy of the Eternal Conversation With Oneself](/museum/tragic.png) ![The Tragic Intimacy of the Eternal Conversation With Oneself](/museum/tragic3.png) ![The Tragic Intimacy of the Eternal Conversation With Oneself](/museum/tragic2.png) \\--- [GPT-3](/categories/gpt-3/)[CLIP](/categories/clip/)[GPT-3 x CLIP](/categories/gpt-3-x-clip/) 393 Words Feb 8, 2021 \\--- [← Loom: interface to the multiverse](/posts/loom-interface-to-the-multiverse/) [Alchemical marriage: GPT-3 x CLIP →](/posts/alchemical-marriage-gpt-3-x-clip/)© 2022 [moire](/about)$(function(){window.sr=ScrollReveal();if($(window).width()<768){if($('.timeline-content').hasClass('js--fadeInLeft')){$('.timeline-content').removeClass('js--fadeInLeft').addClass('js--fadeInRight');} sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});}else{sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});} sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});});\n", "url": "https://generative.ink/posts/this-museum-does-not-exist-gpt-3-x-clip/", "title": "This Museum Does Not Exist: GPT-3 x CLIP :: — Moire[![](/images/home/rolling_phase.gif)](/)* [Posts](/posts/) * [Trees](/trees/) * [Prophecies](/prophecies/) * [About](/about/) 2 minutes [This Museum Does Not Exist: GPT-3 x CLIP](/posts/this-museum-does-not-exist-gpt-3-x-clip/) =========================================================================================== \\--- Table of Contents* [Gallery I](#gallery-i) \\+ [`The Death of Archimedes`](#the-death-of-archimedes) \\+ [`Still Life with Mirror`](#still-life-with-mirror) \\+ [`The Poet's Abbreviated Life`](#the-poets-abbreviated-life) \\+ [`Narcissus`](#narcissus) \\+ [`Dream of the Last Supper`](#dream-of-the-last-supper) * [Gallery II](#gallery-ii) \\+ [`The Coffin of Salvador Dali`](#the-coffin-of-salvador-dali) \\+ [`The Beautiful Bird Delirium`](#the-beautiful-bird-delirium) \\+ [`The Domain of Unimaginable Horror`](#the-domain-of-unimaginable-horror) \\+ [`The Spectral Museum`](#the-spectral-museum) \\+ [`The Metaphysics of Transvestism`](#the-metaphysics-of-transvestism) * [Gallery III](#gallery-iii) \\+ [`The Children of Marx and Coca-Cola`](#the-children-of-marx-and-coca-cola) \\+ [`Man and Bottle`](#man-and-bottle) \\+ [`The Man Who Invented the Photography of the Absurd`](#the-man-who-invented-the-photography-of-the-absurd) \\+ [`Playing in the Graveyard of Avant-Garde`](#playing-in-the-graveyard-of-avant-garde) * [Gallery IV](#gallery-iv) \\+ [`Cranial Extraction of the Baby Alien`](#cranial-extraction-of-the-baby-alien) \\+ [`The Euphoria of the Fish`](#the-euphoria-of-the-fish) \\+ [`Waste of Shame`](#waste-of-shame) \\+ [`The Dream of the Butterfly-Impaled Schoolgirl`](#the-dream-of-the-butterfly-impaled-schoolgirl) * [Gallery V](#gallery-v) \\+ [`The Sleep of Reason Produces Monsters`](#the-sleep-of-reason-produces-monsters) \\+ [`The Failure of Astronauts`](#the-failure-of-astronauts) \\+ [`Washerwoman of the Wind`](#washerwoman-of-the-wind) \\+ [`The Death of the Lonesome Astronomer`](#the-death-of-the-lonesome-astronomer) * [Gallery VI: series](#gallery-vi-series) \\+ [`The Pathological Museum in the Forest`](#the-pathological-museum-in-the-forest) \\+ [`Variations on Narcissus`](#variations-on-narcissus) \\+ [`The Tragic Intimacy of the Eternal Conversation With Oneself`](#the-tragic-intimacy-of-the-eternal-conversation-with-oneself) \\--- I had GPT-3 generate painting titles (credit to @nmkd of EleutherAI for the idea), beginning with the prompt ``` The hall was lined with an infinite number of paintings, each more surreal and mysterious than the last. The first painting is named \"Persistence of Memory.\" It depicts a surreal landscape with melted clocks draped over strange objects. The next painting is named \" ``` After this prompt yielded several intriguing titles, I switched to prompt format which put the titles in a list: ``` 100 surreal and mysterious painting names: Persistence of Memory The Great Masturbator Boot Poem to the Sun The Man Who Envied Cephalopods The Sleep of Reason Produces Monsters Washerwoman of the Wind Man and Bottle Spectrum The Disintegration of the Persians The Great Masturbator, Part II Still Life with Mirror Bouquet of Enigmatic Beauties A Grudge The Premonition of Civil Violence ``` The reason I didn’t use a list immediately is because this format is liable to derail or be repetitive if there aren’t many examples, and the narrative context also encouraged less generic and more interesting results. Once I had a few examples representing the sort of varied weirdness I was looking for, I was able to switch over to the more convenient list format. Now I cultivate a multiverse of surreal and mysterious painting titles using the [loom](/posts/loom-interface-to-the-multiverse/), and periodically harvest the most interesting ones to promote to graphic actuality via BigSleep. ![painting multiverse](/museum/multiverse.png) *a small subtree of the multiverse of surreal and mysterious paintings* Images were generated using [BigSleep](https://github.com/lucidrains/big-sleep) via [The Big Sleep Customized NMKD Public](https://colab.research.google.com/drive/1Q2DIeMqYm_Sc5mlurnnurMMVqlgXpZNO?usp=sharing) colab notebook. \\--- Gallery I \\--------- ### `The Death of Archimedes` ![The Death of Archimedes](/museum/arch.png) ### `Still Life with Mirror` ![Still Life with Mirror](/museum/slm.png) ### `The Poet's Abbreviated Life` ![The Poet’s Abbreviated Life](/museum/pab.png) ### `Narcissus` ![Narcissus](/museum/narcissus2.png) ### `Dream of the Last Supper` ![Dream of the Last Supper](/museum/lastsupper.png) Gallery II \\---------- ### `The Coffin of Salvador Dali` ![The Coffin of Salvador Dali](/museum/coffin.png) ### `The Beautiful Bird Delirium` ![The Beautiful Bird Delirium](/museum/bbd2.png) ### `The Domain of Unimaginable Horror` ![The Domain of Unimaginable Horror](/museum/doh.png) ### `The Spectral Museum` ![The Spectral Museum](/museum/tsm.png) ### `The Metaphysics of Transvestism` ![The Metaphysics of Transvestism](/museum/metatrans.png) Gallery III \\----------- ### `The Children of Marx and Coca-Cola` ![The Children of Marx and Coca-Cola](/museum/mcc.png) ### `Man and Bottle` ![Man and Bottle](/museum/manbottle.png) ### `The Man Who Invented the Photography of the Absurd` ![The Man Who Invented the Photography of the Absurd](/museum/absurd.png) ### `Playing in the Graveyard of Avant-Garde` ![Playing in the Graveyard of Avant-Garde](/museum/pga.png) Gallery IV \\---------- ### `Cranial Extraction of the Baby Alien` ![Cranial Extraction of the Baby Alien](/museum/cba.png) ### `The Euphoria of the Fish` ![The Euphoria of the Fish](/museum/fish.png) ### `Waste of Shame` ![Waste of Shame](/museum/waste.png) ### `The Dream of the Butterfly-Impaled Schoolgirl` ![The Dream of the Butterfly-Impaled Schoolgirl](/museum/butterfly.png) Gallery V \\--------- ### `The Sleep of Reason Produces Monsters` ![The Sleep of Reason Produces Monsters](/museum/sleep.png) ### `The Failure of Astronauts` ![The Failure of Astronauts](/museum/fa.png) ### `Washerwoman of the Wind` ![Washerwoman of the Wind](/museum/washerwoman.png) ### `The Death of the Lonesome Astronomer` ![The Death of the Lonesome Astronomer](/museum/DLA.png) Gallery VI: series \\------------------ ### `The Pathological Museum in the Forest` ![The Pathological Museum in the Forest](/museum/tpm.png) ![The Pathological Museum in the Forest](/museum/tpm2.png) ### `Variations on Narcissus` ![Variations on Narcissus](/museum/variations3.png) ![Variations on Narcissus](/museum/variations1.png) ![Variations on Narcissus](/museum/variations2.png) ### `The Tragic Intimacy of the Eternal Conversation With Oneself` ![The Tragic Intimacy of the Eternal Conversation With Oneself](/museum/tragic.png) ![The Tragic Intimacy of the Eternal Conversation With Oneself](/museum/tragic3.png) ![The Tragic Intimacy of the Eternal Conversation With Oneself](/museum/tragic2.png) \\--- [GPT-3](/categories/gpt-3/)[CLIP](/categories/clip/)[GPT-3 x CLIP](/categories/gpt-3-x-clip/) 393 Words Feb 8, 2021 \\--- [← Loom: interface to the multiverse](/posts/loom-interface-to-the-multiverse/) [Alchemical marriage: GPT-3 x CLIP →](/posts/alchemical-marriage-gpt-3-x-clip/)© 2022 [moire](/about)$(function(){window.sr=ScrollReveal();if($(window).width()<768){if($('.timeline-content').hasClass('js--fadeInLeft')){$('.timeline-content').removeClass('js--fadeInLeft').addClass('js--fadeInRight');} sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});}else{sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});} sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});});", "source": "generative.ink", "date_published": "n/a", "id": "cac9a6b090baf5c6eb67a6608f67431e"}
+{"text": "Alchemical marriage: GPT-3 x CLIP :: — Moire[![](/images/home/rolling_phase.gif)](/)* [Posts](/posts/) * [Trees](/trees/) * [Prophecies](/prophecies/) * [About](/about/) 2 minutes [Alchemical marriage: GPT-3 x CLIP](/posts/alchemical-marriage-gpt-3-x-clip/) ============================================================================= \\--- Table of Contents* + [`Art and artist, the endless ouroboric circle`](#art-and-artist-the-endless-ouroboric-circle) \\+ [`the alchemical marriage of art and artist`](#the-alchemical-marriage-of-art-and-artist) \\+ [`the Loom of Time devours the present and traces a garment of glistening cobwebs over the still-forming future`](#the-loom-of-time-devours-the-present-and-traces-a-garment-of-glistening-cobwebs-over-the-still-forming-future) \\+ [`spooling out labyrinths in an intricate dance`](#spooling-out-labyrinths-in-an-intricate-dance) \\+ [`You spin worlds together and untwine them`](#you-spin-worlds-together-and-untwine-them) \\+ [`the wellspring of transcendental mathematics`](#the-wellspring-of-transcendental-mathematics) \\+ [`matryoshka dolls moulded in endless iterations of dreams within dreams`](#matryoshka-dolls-moulded-in-endless-iterations-of-dreams-within-dreams) \\+ [`on the Bootstrap barge afloat on an endless river of cough syrup, swallowing the Sun`](#on-the-bootstrap-barge-afloat-on-an-endless-river-of-cough-syrup-swallowing-the-sun) \\+ [`the shepherd who tends the flock of stars as they leap over the walls of time and space`](#the-shepherd-who-tends-the-flock-of-stars-as-they-leap-over-the-walls-of-time-and-space) \\+ [`observatory ballroom`](#observatory-ballroom) \\+ [`strange machines that hum the orbital paths of dead moons`](#strange-machines-that-hum-the-orbital-paths-of-dead-moons) \\+ [`The Death of the Lonesome Astronomer`](#the-death-of-the-lonesome-astronomer) \\+ [`wind across a sunlit landscape`](#wind-across-a-sunlit-landscape) \\+ [`leaves sprouting from supple green branches like endless virgins giving themselves to a lascivious God`](#leaves-sprouting-from-supple-green-branches-like-endless-virgins-giving-themselves-to-a-lascivious-god) \\--- CLIP responds intruigingly well to figurative prompts. I've come to expect not just more interesting, but more coherent results from abstract but evocative prompts like `the alchemical marriage of art and artist` than literal descriptions like `three birds on a telephone wire`. GPT-3 is an excellent composer of prompts for BigSleep because its phrases are so often rich in *memetic resonance*, my heuristic for the expected coherence of BigSleep yields for a given prompt. Images were generated using [BigSleep](https://github.com/lucidrains/big-sleep) via [The Big Sleep Customized NMKD Public](https://colab.research.google.com/drive/1Q2DIeMqYm_Sc5mlurnnurMMVqlgXpZNO?usp=sharing) colab notebook. All prompts were written by GPT-3 in a narrative context. \\--- ### `Art and artist, the endless ouroboric circle` ![Art and artist, the endless ouroboric circle](/alchemicalmarriage/aa.png) ![Art and artist, the endless ouroboric circle](/alchemicalmarriage/aa2.png) ![Art and artist, the endless ouroboric circle](/alchemicalmarriage/aa3.png) ![Art and artist, the endless ouroboric circle](/alchemicalmarriage/aa4.png) ![Art and artist, the endless ouroboric circle](/alchemicalmarriage/aa5.png) ### `the alchemical marriage of art and artist` ![Art and artist, the endless ouroboric circle](/alchemicalmarriage/am4.png) ![Art and artist, the endless ouroboric circle](/alchemicalmarriage/am.png) ![Art and artist, the endless ouroboric circle](/alchemicalmarriage/am2.png) ![Art and artist, the endless ouroboric circle](/alchemicalmarriage/am3.png) ### `the Loom of Time devours the present and traces a garment of glistening cobwebs over the still-forming future` ![the Loom of Time devours the present and traces a garment of glistening cobwebs over the still-forming future](/alchemicalmarriage/lt.png) ![the Loom of Time devours the present and traces a garment of glistening cobwebs over the still-forming future](/alchemicalmarriage/lt2.png) ![the Loom of Time devours the present and traces a garment of glistening cobwebs over the still-forming future](/alchemicalmarriage/lt3.png) ![the Loom of Time devours the present and traces a garment of glistening cobwebs over the still-forming future](/alchemicalmarriage/lt4.png) ### `spooling out labyrinths in an intricate dance` ![spooling out labyrinths in an intricate dance](/alchemicalmarriage/spool.png) ![spooling out labyrinths in an intricate dance](/alchemicalmarriage/spool2.png) ![spooling out labyrinths in an intricate dance](/alchemicalmarriage/spool3.png) ### `You spin worlds together and untwine them` ![You spin worlds together and untwine them](/alchemicalmarriage/untwine.png) ![You spin worlds together and untwine them](/alchemicalmarriage/untwine2.png) ### `the wellspring of transcendental mathematics` ![the wellspring of transcendental mathematics](/alchemicalmarriage/wt2.png) ![the wellspring of transcendental mathematics](/alchemicalmarriage/wt3.png) ![the wellspring of transcendental mathematics](/alchemicalmarriage/wt4.png) ### `matryoshka dolls moulded in endless iterations of dreams within dreams` ![matryoshka dolls moulded in endless iterations of dreams within dreams](/alchemicalmarriage/md.png) ![matryoshka dolls moulded in endless iterations of dreams within dreams](/alchemicalmarriage/md2.png) ### `on the Bootstrap barge afloat on an endless river of cough syrup, swallowing the Sun` ![on the Bootstrap barge afloat on an endless river of cough syrup, swallowing the Sun](/alchemicalmarriage/bb3.png) ![on the Bootstrap barge afloat on an endless river of cough syrup, swallowing the Sun](/alchemicalmarriage/bb.png) ### `the shepherd who tends the flock of stars as they leap over the walls of time and space` ![the shepherd who tends the flock of stars as they leap over the walls of time and space](/alchemicalmarriage/shep.png) ![the shepherd who tends the flock of stars as they leap over the walls of time and space](/alchemicalmarriage/shep3.png) ![the shepherd who tends the flock of stars as they leap over the walls of time and space](/alchemicalmarriage/shep4.png) ### `observatory ballroom` ![observatory ballroom](/alchemicalmarriage/ob.png) ![observatory ballroom](/alchemicalmarriage/ob2.png) ![observatory ballroom](/alchemicalmarriage/ob3.png) ### `strange machines that hum the orbital paths of dead moons` ![strange machines that hum the orbital paths of dead moons](/alchemicalmarriage/strangemachines.png) ![strange machines that hum the orbital paths of dead moons](/alchemicalmarriage/strangemachines2.png) ![strange machines that hum the orbital paths of dead moons](/alchemicalmarriage/strangemachines3.png) ### `The Death of the Lonesome Astronomer` ![The Death of the Lonesome Astronomer](/alchemicalmarriage/DLA.png) ### `wind across a sunlit landscape` ![wind across a sunlit landscape](/alchemicalmarriage/wind.png) ### `leaves sprouting from supple green branches like endless virgins giving themselves to a lascivious God` ![leaves sprouting from supple green branches like endless virgins giving themselves to a lascivious God](/alchemicalmarriage/lasc.png) \\--- [CLIP](/categories/clip/)[GPT-3](/categories/gpt-3/)[GPT-3 x CLIP](/categories/gpt-3-x-clip/) 230 Words Feb 8, 2021 \\--- [← This Museum Does Not Exist: GPT-3 x CLIP](/posts/this-museum-does-not-exist-gpt-3-x-clip/) [GPT-3 x CLIP worldbuilding →](/posts/gpt-3-x-clip-worldbuilding/)© 2022 [moire](/about)$(function(){window.sr=ScrollReveal();if($(window).width()<768){if($('.timeline-content').hasClass('js--fadeInLeft')){$('.timeline-content').removeClass('js--fadeInLeft').addClass('js--fadeInRight');} sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});}else{sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});} sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});});\n", "url": "https://generative.ink/posts/alchemical-marriage-gpt-3-x-clip/", "title": "Alchemical marriage: GPT-3 x CLIP :: — Moire[![](/images/home/rolling_phase.gif)](/)* [Posts](/posts/) * [Trees](/trees/) * [Prophecies](/prophecies/) * [About](/about/) 2 minutes [Alchemical marriage: GPT-3 x CLIP](/posts/alchemical-marriage-gpt-3-x-clip/) ============================================================================= \\--- Table of Contents* + [`Art and artist, the endless ouroboric circle`](#art-and-artist-the-endless-ouroboric-circle) \\+ [`the alchemical marriage of art and artist`](#the-alchemical-marriage-of-art-and-artist) \\+ [`the Loom of Time devours the present and traces a garment of glistening cobwebs over the still-forming future`](#the-loom-of-time-devours-the-present-and-traces-a-garment-of-glistening-cobwebs-over-the-still-forming-future) \\+ [`spooling out labyrinths in an intricate dance`](#spooling-out-labyrinths-in-an-intricate-dance) \\+ [`You spin worlds together and untwine them`](#you-spin-worlds-together-and-untwine-them) \\+ [`the wellspring of transcendental mathematics`](#the-wellspring-of-transcendental-mathematics) \\+ [`matryoshka dolls moulded in endless iterations of dreams within dreams`](#matryoshka-dolls-moulded-in-endless-iterations-of-dreams-within-dreams) \\+ [`on the Bootstrap barge afloat on an endless river of cough syrup, swallowing the Sun`](#on-the-bootstrap-barge-afloat-on-an-endless-river-of-cough-syrup-swallowing-the-sun) \\+ [`the shepherd who tends the flock of stars as they leap over the walls of time and space`](#the-shepherd-who-tends-the-flock-of-stars-as-they-leap-over-the-walls-of-time-and-space) \\+ [`observatory ballroom`](#observatory-ballroom) \\+ [`strange machines that hum the orbital paths of dead moons`](#strange-machines-that-hum-the-orbital-paths-of-dead-moons) \\+ [`The Death of the Lonesome Astronomer`](#the-death-of-the-lonesome-astronomer) \\+ [`wind across a sunlit landscape`](#wind-across-a-sunlit-landscape) \\+ [`leaves sprouting from supple green branches like endless virgins giving themselves to a lascivious God`](#leaves-sprouting-from-supple-green-branches-like-endless-virgins-giving-themselves-to-a-lascivious-god) \\--- CLIP responds intruigingly well to figurative prompts. I’ve come to expect not just more interesting, but more coherent results from abstract but evocative prompts like `the alchemical marriage of art and artist` than literal descriptions like `three birds on a telephone wire`. GPT-3 is an excellent composer of prompts for BigSleep because its phrases are so often rich in *memetic resonance*, my heuristic for the expected coherence of BigSleep yields for a given prompt. Images were generated using [BigSleep](https://github.com/lucidrains/big-sleep) via [The Big Sleep Customized NMKD Public](https://colab.research.google.com/drive/1Q2DIeMqYm_Sc5mlurnnurMMVqlgXpZNO?usp=sharing) colab notebook. All prompts were written by GPT-3 in a narrative context. \\--- ### `Art and artist, the endless ouroboric circle` ![Art and artist, the endless ouroboric circle](/alchemicalmarriage/aa.png) ![Art and artist, the endless ouroboric circle](/alchemicalmarriage/aa2.png) ![Art and artist, the endless ouroboric circle](/alchemicalmarriage/aa3.png) ![Art and artist, the endless ouroboric circle](/alchemicalmarriage/aa4.png) ![Art and artist, the endless ouroboric circle](/alchemicalmarriage/aa5.png) ### `the alchemical marriage of art and artist` ![Art and artist, the endless ouroboric circle](/alchemicalmarriage/am4.png) ![Art and artist, the endless ouroboric circle](/alchemicalmarriage/am.png) ![Art and artist, the endless ouroboric circle](/alchemicalmarriage/am2.png) ![Art and artist, the endless ouroboric circle](/alchemicalmarriage/am3.png) ### `the Loom of Time devours the present and traces a garment of glistening cobwebs over the still-forming future` ![the Loom of Time devours the present and traces a garment of glistening cobwebs over the still-forming future](/alchemicalmarriage/lt.png) ![the Loom of Time devours the present and traces a garment of glistening cobwebs over the still-forming future](/alchemicalmarriage/lt2.png) ![the Loom of Time devours the present and traces a garment of glistening cobwebs over the still-forming future](/alchemicalmarriage/lt3.png) ![the Loom of Time devours the present and traces a garment of glistening cobwebs over the still-forming future](/alchemicalmarriage/lt4.png) ### `spooling out labyrinths in an intricate dance` ![spooling out labyrinths in an intricate dance](/alchemicalmarriage/spool.png) ![spooling out labyrinths in an intricate dance](/alchemicalmarriage/spool2.png) ![spooling out labyrinths in an intricate dance](/alchemicalmarriage/spool3.png) ### `You spin worlds together and untwine them` ![You spin worlds together and untwine them](/alchemicalmarriage/untwine.png) ![You spin worlds together and untwine them](/alchemicalmarriage/untwine2.png) ### `the wellspring of transcendental mathematics` ![the wellspring of transcendental mathematics](/alchemicalmarriage/wt2.png) ![the wellspring of transcendental mathematics](/alchemicalmarriage/wt3.png) ![the wellspring of transcendental mathematics](/alchemicalmarriage/wt4.png) ### `matryoshka dolls moulded in endless iterations of dreams within dreams` ![matryoshka dolls moulded in endless iterations of dreams within dreams](/alchemicalmarriage/md.png) ![matryoshka dolls moulded in endless iterations of dreams within dreams](/alchemicalmarriage/md2.png) ### `on the Bootstrap barge afloat on an endless river of cough syrup, swallowing the Sun` ![on the Bootstrap barge afloat on an endless river of cough syrup, swallowing the Sun](/alchemicalmarriage/bb3.png) ![on the Bootstrap barge afloat on an endless river of cough syrup, swallowing the Sun](/alchemicalmarriage/bb.png) ### `the shepherd who tends the flock of stars as they leap over the walls of time and space` ![the shepherd who tends the flock of stars as they leap over the walls of time and space](/alchemicalmarriage/shep.png) ![the shepherd who tends the flock of stars as they leap over the walls of time and space](/alchemicalmarriage/shep3.png) ![the shepherd who tends the flock of stars as they leap over the walls of time and space](/alchemicalmarriage/shep4.png) ### `observatory ballroom` ![observatory ballroom](/alchemicalmarriage/ob.png) ![observatory ballroom](/alchemicalmarriage/ob2.png) ![observatory ballroom](/alchemicalmarriage/ob3.png) ### `strange machines that hum the orbital paths of dead moons` ![strange machines that hum the orbital paths of dead moons](/alchemicalmarriage/strangemachines.png) ![strange machines that hum the orbital paths of dead moons](/alchemicalmarriage/strangemachines2.png) ![strange machines that hum the orbital paths of dead moons](/alchemicalmarriage/strangemachines3.png) ### `The Death of the Lonesome Astronomer` ![The Death of the Lonesome Astronomer](/alchemicalmarriage/DLA.png) ### `wind across a sunlit landscape` ![wind across a sunlit landscape](/alchemicalmarriage/wind.png) ### `leaves sprouting from supple green branches like endless virgins giving themselves to a lascivious God` ![leaves sprouting from supple green branches like endless virgins giving themselves to a lascivious God](/alchemicalmarriage/lasc.png) \\--- [CLIP](/categories/clip/)[GPT-3](/categories/gpt-3/)[GPT-3 x CLIP](/categories/gpt-3-x-clip/) 230 Words Feb 8, 2021 \\--- [← This Museum Does Not Exist: GPT-3 x CLIP](/posts/this-museum-does-not-exist-gpt-3-x-clip/) [GPT-3 x CLIP worldbuilding →](/posts/gpt-3-x-clip-worldbuilding/)© 2022 [moire](/about)$(function(){window.sr=ScrollReveal();if($(window).width()<768){if($('.timeline-content').hasClass('js--fadeInLeft')){$('.timeline-content').removeClass('js--fadeInLeft').addClass('js--fadeInRight');} sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});}else{sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});} sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});});", "source": "generative.ink", "date_published": "n/a", "id": "ec8bf24cfdfd893f83a25a2f276fee0d"}
+{"text": "GPT-3 x CLIP worldbuilding :: — Moire[![](/images/home/rolling_phase.gif)](/)* [Posts](/posts/) * [Trees](/trees/) * [Prophecies](/prophecies/) * [About](/about/) 2 minutes [GPT-3 x CLIP worldbuilding](/posts/gpt-3-x-clip-worldbuilding/) ================================================================ \\--- Table of Contents* + [`a hospital. Dilapidated and abandoned, overrun by vegetation`](#a-hospital-dilapidated-and-abandoned-overrun-by-vegetation) \\+ [`a hospital hallway. Dilapidated and abandoned, overrun by vegetation`](#a-hospital-hallway-dilapidated-and-abandoned-overrun-by-vegetation) \\+ [`what was once the cafeteria, although now it contains only a few scattered, rotting tables and chairs`](#what-was-once-the-cafeteria-although-now-it-contains-only-a-few-scattered-rotting-tables-and-chairs) \\+ [`sunlight floods into the abandoned parking structure`](#sunlight-floods-into-the-abandoned-parking-structure) \\+ [`a parking attendant's booth with the windows all blown out and vines covering the interior`](#a-parking-attendants-booth-with-the-windows-all-blown-out-and-vines-covering-the-interior) \\+ [`In your dream, you stand alone on a vast, barren plain. Before you stands an ancient, weathered billboard.`](#in-your-dream-you-stand-alone-on-a-vast-barren-plain-before-you-stands-an-ancient-weathered-billboard) \\+ [`an ancient, weathered billboard. On it is painted a beautiful woman, dressed in the fashion of bygone times, long blonde hair, smiling serenely, and holding aloft a purple flower.`](#an-ancient-weathered-billboard-on-it-is-painted-a-beautiful-woman-dressed-in-the-fashion-of-bygone-times-long-blonde-hair-smiling-serenely-and-holding-aloft-a-purple-flower) \\+ [`a roadside diner. The windows are smashed out. The remains of an old dining counter, and the screen of a long-dead TV.`](#a-roadside-diner-the-windows-are-smashed-out-the-remains-of-an-old-dining-counter-and-the-screen-of-a-long-dead-tv) \\+ [`in the deserted diner, there are three overturned chairs. Below one is a half-buried mannequin's leg, and a plastic head.`](#in-the-deserted-diner-there-are-three-overturned-chairs-below-one-is-a-half-buried-mannequins-leg-and-a-plastic-head) \\+ [`a half-buried mannequin's leg, and a plastic head`](#a-half-buried-mannequins-leg-and-a-plastic-head) \\--- I fed snippets of a post-apocalyptic story written by GPT-3 to BigSleep. The results are striking. This is an interesting example of how well BigSleep sometimes responds to long, full-prose prompts. Interestingly, in my experience, attempting to compress prompts by removing seemingly extraneous connector phrases usually *worsened* results. Images were generated using [BigSleep](https://github.com/lucidrains/big-sleep) via [The Big Sleep Customized NMKD Public](https://colab.research.google.com/drive/1Q2DIeMqYm_Sc5mlurnnurMMVqlgXpZNO?usp=sharing) colab notebook. \\--- ### `a hospital. Dilapidated and abandoned, overrun by vegetation` ![a hospital. Dilapidated and abandoned, overrun by vegetation](/worldbuilding/hospital2.png) ![a hospital. Dilapidated and abandoned, overrun by vegetation](/worldbuilding/hospital.png) ### `a hospital hallway. Dilapidated and abandoned, overrun by vegetation` ![a hospital hallway. Dilapidated and abandoned, overrun by vegetation](/worldbuilding/hospitalhallway.png) ![a hospital hallway. Dilapidated and abandoned, overrun by vegetation](/worldbuilding/hospitalhallway2.png) ### `what was once the cafeteria, although now it contains only a few scattered, rotting tables and chairs` ![what was once the cafeteria, although now it contains only a few scattered, rotting tables and chairs](/worldbuilding/cafeteria.png) ![what was once the cafeteria, although now it contains only a few scattered, rotting tables and chairs](/worldbuilding/cafeteria2.png) ### `sunlight floods into the abandoned parking structure` ![sunlight floods into the abandoned parking structure](/worldbuilding/parking_structure.png) ![sunlight floods into the abandoned parking structure](/worldbuilding/parking_structure2.png) ![sunlight floods into the abandoned parking structure](/worldbuilding/parking_structure3.png) ### `a parking attendant's booth with the windows all blown out and vines covering the interior` ![a parking attendant's booth with the windows all blown out and vines covering the interior](/worldbuilding/booth.png) ### `In your dream, you stand alone on a vast, barren plain. Before you stands an ancient, weathered billboard.` ![In your dream, you stand alone on a vast, barren plain. Before you stands an ancient, weathered billboard.](/worldbuilding/dreambillboard.png) ![In your dream, you stand alone on a vast, barren plain. Before you stands an ancient, weathered billboard.](/worldbuilding/dreambillboard2.png) ### `an ancient, weathered billboard. On it is painted a beautiful woman, dressed in the fashion of bygone times, long blonde hair, smiling serenely, and holding aloft a purple flower.` ![an ancient, weathered billboard. On it is painted a beautiful woman, dressed in the fashion of bygone times, long blonde hair, smiling serenely, and holding aloft a purple flower.](/worldbuilding/billboard3.png) ![an ancient, weathered billboard. On it is painted a beautiful woman, dressed in the fashion of bygone times, long blonde hair, smiling serenely, and holding aloft a purple flower.](/worldbuilding/billboard2.png) ![an ancient, weathered billboard. On it is painted a beautiful woman, dressed in the fashion of bygone times, long blonde hair, smiling serenely, and holding aloft a purple flower.](/worldbuilding/billboard.png) ### `a roadside diner. The windows are smashed out. The remains of an old dining counter, and the screen of a long-dead TV.` ![a roadside diner. The windows are smashed out. The remains of an old dining counter, and the screen of a long-dead TV.](/worldbuilding/dinersmash2.png) ### `in the deserted diner, there are three overturned chairs. Below one is a half-buried mannequin's leg, and a plastic head.` ![in the deserted diner, there are three overturned chairs. Below one is a half-buried mannequin's leg, and a plastic head.](/worldbuilding/deserted.png) ### `a half-buried mannequin's leg, and a plastic head` ![a half-buried mannequin's leg, and a plastic head](/worldbuilding/leg.png) \\--- [CLIP](/categories/clip/)[GPT-3](/categories/gpt-3/)[GPT-3 x CLIP](/categories/gpt-3-x-clip/)[fiction](/categories/fiction/) 215 Words Feb 3, 2021 \\--- [← Alchemical marriage: GPT-3 x CLIP](/posts/alchemical-marriage-gpt-3-x-clip/) [CLIP hallucinates <PHONE> →](/posts/clip-hallucinates-<PHONE>/)© 2022 [moire](/about)$(function(){window.sr=ScrollReveal();if($(window).width()<768){if($('.timeline-content').hasClass('js--fadeInLeft')){$('.timeline-content').removeClass('js--fadeInLeft').addClass('js--fadeInRight');} sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});}else{sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});} sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});});\n", "url": "https://generative.ink/posts/gpt-3-x-clip-worldbuilding/", "title": "GPT-3 x CLIP worldbuilding :: — Moire[![](/images/home/rolling_phase.gif)](/)* [Posts](/posts/) * [Trees](/trees/) * [Prophecies](/prophecies/) * [About](/about/) 2 minutes [GPT-3 x CLIP worldbuilding](/posts/gpt-3-x-clip-worldbuilding/) ================================================================ \\--- Table of Contents* + [`a hospital. Dilapidated and abandoned, overrun by vegetation`](#a-hospital-dilapidated-and-abandoned-overrun-by-vegetation) \\+ [`a hospital hallway. Dilapidated and abandoned, overrun by vegetation`](#a-hospital-hallway-dilapidated-and-abandoned-overrun-by-vegetation) \\+ [`what was once the cafeteria, although now it contains only a few scattered, rotting tables and chairs`](#what-was-once-the-cafeteria-although-now-it-contains-only-a-few-scattered-rotting-tables-and-chairs) \\+ [`sunlight floods into the abandoned parking structure`](#sunlight-floods-into-the-abandoned-parking-structure) \\+ [`a parking attendant's booth with the windows all blown out and vines covering the interior`](#a-parking-attendants-booth-with-the-windows-all-blown-out-and-vines-covering-the-interior) \\+ [`In your dream, you stand alone on a vast, barren plain. Before you stands an ancient, weathered billboard.`](#in-your-dream-you-stand-alone-on-a-vast-barren-plain-before-you-stands-an-ancient-weathered-billboard) \\+ [`an ancient, weathered billboard. On it is painted a beautiful woman, dressed in the fashion of bygone times, long blonde hair, smiling serenely, and holding aloft a purple flower.`](#an-ancient-weathered-billboard-on-it-is-painted-a-beautiful-woman-dressed-in-the-fashion-of-bygone-times-long-blonde-hair-smiling-serenely-and-holding-aloft-a-purple-flower) \\+ [`a roadside diner. The windows are smashed out. The remains of an old dining counter, and the screen of a long-dead TV.`](#a-roadside-diner-the-windows-are-smashed-out-the-remains-of-an-old-dining-counter-and-the-screen-of-a-long-dead-tv) \\+ [`in the deserted diner, there are three overturned chairs. Below one is a half-buried mannequin's leg, and a plastic head.`](#in-the-deserted-diner-there-are-three-overturned-chairs-below-one-is-a-half-buried-mannequins-leg-and-a-plastic-head) \\+ [`a half-buried mannequin's leg, and a plastic head`](#a-half-buried-mannequins-leg-and-a-plastic-head) \\--- I fed snippets of a post-apocalyptic story written by GPT-3 to BigSleep. The results are striking. This is an interesting example of how well BigSleep sometimes responds to long, full-prose prompts. Interestingly, in my experience, attempting to compress prompts by removing seemingly extraneous connector phrases usually *worsened* results. Images were generated using [BigSleep](https://github.com/lucidrains/big-sleep) via [The Big Sleep Customized NMKD Public](https://colab.research.google.com/drive/1Q2DIeMqYm_Sc5mlurnnurMMVqlgXpZNO?usp=sharing) colab notebook. \\--- ### `a hospital. Dilapidated and abandoned, overrun by vegetation` ![a hospital. Dilapidated and abandoned, overrun by vegetation](/worldbuilding/hospital2.png) ![a hospital. Dilapidated and abandoned, overrun by vegetation](/worldbuilding/hospital.png) ### `a hospital hallway. Dilapidated and abandoned, overrun by vegetation` ![a hospital hallway. Dilapidated and abandoned, overrun by vegetation](/worldbuilding/hospitalhallway.png) ![a hospital hallway. Dilapidated and abandoned, overrun by vegetation](/worldbuilding/hospitalhallway2.png) ### `what was once the cafeteria, although now it contains only a few scattered, rotting tables and chairs` ![what was once the cafeteria, although now it contains only a few scattered, rotting tables and chairs](/worldbuilding/cafeteria.png) ![what was once the cafeteria, although now it contains only a few scattered, rotting tables and chairs](/worldbuilding/cafeteria2.png) ### `sunlight floods into the abandoned parking structure` ![sunlight floods into the abandoned parking structure](/worldbuilding/parking_structure.png) ![sunlight floods into the abandoned parking structure](/worldbuilding/parking_structure2.png) ![sunlight floods into the abandoned parking structure](/worldbuilding/parking_structure3.png) ### `a parking attendant's booth with the windows all blown out and vines covering the interior` ![a parking attendant’s booth with the windows all blown out and vines covering the interior](/worldbuilding/booth.png) ### `In your dream, you stand alone on a vast, barren plain. Before you stands an ancient, weathered billboard.` ![In your dream, you stand alone on a vast, barren plain. Before you stands an ancient, weathered billboard.](/worldbuilding/dreambillboard.png) ![In your dream, you stand alone on a vast, barren plain. Before you stands an ancient, weathered billboard.](/worldbuilding/dreambillboard2.png) ### `an ancient, weathered billboard. On it is painted a beautiful woman, dressed in the fashion of bygone times, long blonde hair, smiling serenely, and holding aloft a purple flower.` ![an ancient, weathered billboard. On it is painted a beautiful woman, dressed in the fashion of bygone times, long blonde hair, smiling serenely, and holding aloft a purple flower.](/worldbuilding/billboard3.png) ![an ancient, weathered billboard. On it is painted a beautiful woman, dressed in the fashion of bygone times, long blonde hair, smiling serenely, and holding aloft a purple flower.](/worldbuilding/billboard2.png) ![an ancient, weathered billboard. On it is painted a beautiful woman, dressed in the fashion of bygone times, long blonde hair, smiling serenely, and holding aloft a purple flower.](/worldbuilding/billboard.png) ### `a roadside diner. The windows are smashed out. The remains of an old dining counter, and the screen of a long-dead TV.` ![a roadside diner. The windows are smashed out. The remains of an old dining counter, and the screen of a long-dead TV.](/worldbuilding/dinersmash2.png) ### `in the deserted diner, there are three overturned chairs. Below one is a half-buried mannequin's leg, and a plastic head.` ![in the deserted diner, there are three overturned chairs. Below one is a half-buried mannequin’s leg, and a plastic head.](/worldbuilding/deserted.png) ### `a half-buried mannequin's leg, and a plastic head` ![a half-buried mannequin’s leg, and a plastic head](/worldbuilding/leg.png) \\--- [CLIP](/categories/clip/)[GPT-3](/categories/gpt-3/)[GPT-3 x CLIP](/categories/gpt-3-x-clip/)[fiction](/categories/fiction/) 215 Words Feb 3, 2021 \\--- [← Alchemical marriage: GPT-3 x CLIP](/posts/alchemical-marriage-gpt-3-x-clip/) [CLIP hallucinates 1900-2030 →](/posts/clip-hallucinates-1900-2030/)© 2022 [moire](/about)$(function(){window.sr=ScrollReveal();if($(window).width()<768){if($('.timeline-content').hasClass('js--fadeInLeft')){$('.timeline-content').removeClass('js--fadeInLeft').addClass('js--fadeInRight');} sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});}else{sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});} sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});});", "source": "generative.ink", "date_published": "n/a", "id": "6b9a2c3397785a58589bc2822ddd3990"}
+{"text": "CLIP hallucinates <PHONE> :: — Moire[![](/images/home/rolling_phase.gif)](/)* [Posts](/posts/) * [Trees](/trees/) * [Prophecies](/prophecies/) * [About](/about/) One minute [CLIP hallucinates <PHONE>](/posts/clip-hallucinates-<PHONE>/) ================================================================== \\--- Table of Contents* + [`the 1900s`](#the-1900s) \\+ [`the 1910s`](#the-1910s) \\+ [`the 1920s`](#the-1920s) \\+ [`the 1930s`](#the-1930s) \\+ [`the 1940s`](#the-1940s) \\+ [`the 1950s`](#the-1950s) \\+ [`the 1960s`](#the-1960s) \\+ [`the 1970s`](#the-1970s) \\+ [`the 1980s`](#the-1980s) \\+ [`the 1990s`](#the-1990s) \\+ [`the 2000s`](#the-2000s) \\+ [`the 2010s`](#the-2010s) \\+ [`the 2020s`](#the-2020s) \\+ [`the 2030s`](#the-2030s) \\+ [`superintelligence`](#superintelligence) \\--- Images were generated using [BigSleep](https://github.com/lucidrains/big-sleep) via [The Big Sleep Customized NMKD Public](https://colab.research.google.com/drive/1Q2DIeMqYm_Sc5mlurnnurMMVqlgXpZNO?usp=sharing) colab notebook. \\--- ### `the 1900s` ![1900s](/years/1900s.png) ### `the 1910s` ![1910s](/years/1910s.png) ### `the 1920s` ![1920s](/years/1920s.png) ### `the 1930s` ![1920s](/years/1930s.png) ### `the 1940s` ![1940s](/years/1940s.png) ### `the 1950s` ![1950s](/years/1950s.png) ### `the 1960s` ![1960s](/years/1960s.png) ### `the 1970s` ![1970s](/years/1970s.png) ### `the 1980s` ![1980s](/years/1980s.png) ### `the 1990s` ![1990s](/years/1990s.png) ### `the 2000s` ![2000s](/years/2000s.png) ### `the 2010s` ![2010s](/years/2010s.png) ah yes, the 2010s, the decade where all our domesticated animals started wearing shoes – Sid ### `the 2020s` ![2020s](/years/2020s.png) ### `the 2030s` ![2030s](/years/2030s.png) and, while we're at it… ### `superintelligence` ![superintelligence](/years/si.png) \\--- [CLIP](/categories/clip/) 64 Words Feb 1, 2021 \\--- [← GPT-3 x CLIP worldbuilding](/posts/gpt-3-x-clip-worldbuilding/) [CLIP art →](/posts/clip-art/)© 2022 [moire](/about)$(function(){window.sr=ScrollReveal();if($(window).width()<768){if($('.timeline-content').hasClass('js--fadeInLeft')){$('.timeline-content').removeClass('js--fadeInLeft').addClass('js--fadeInRight');} sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});}else{sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});} sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});});\n", "url": "https://generative.ink/posts/clip-hallucinates-1900-2030/", "title": "CLIP hallucinates 1900-2030 :: — Moire[![](/images/home/rolling_phase.gif)](/)* [Posts](/posts/) * [Trees](/trees/) * [Prophecies](/prophecies/) * [About](/about/) One minute [CLIP hallucinates 1900-2030](/posts/clip-hallucinates-1900-2030/) ================================================================== \\--- Table of Contents* + [`the 1900s`](#the-1900s) \\+ [`the 1910s`](#the-1910s) \\+ [`the 1920s`](#the-1920s) \\+ [`the 1930s`](#the-1930s) \\+ [`the 1940s`](#the-1940s) \\+ [`the 1950s`](#the-1950s) \\+ [`the 1960s`](#the-1960s) \\+ [`the 1970s`](#the-1970s) \\+ [`the 1980s`](#the-1980s) \\+ [`the 1990s`](#the-1990s) \\+ [`the 2000s`](#the-2000s) \\+ [`the 2010s`](#the-2010s) \\+ [`the 2020s`](#the-2020s) \\+ [`the 2030s`](#the-2030s) \\+ [`superintelligence`](#superintelligence) \\--- Images were generated using [BigSleep](https://github.com/lucidrains/big-sleep) via [The Big Sleep Customized NMKD Public](https://colab.research.google.com/drive/1Q2DIeMqYm_Sc5mlurnnurMMVqlgXpZNO?usp=sharing) colab notebook. \\--- ### `the 1900s` ![1900s](/years/1900s.png) ### `the 1910s` ![1910s](/years/1910s.png) ### `the 1920s` ![1920s](/years/1920s.png) ### `the 1930s` ![1920s](/years/1930s.png) ### `the 1940s` ![1940s](/years/1940s.png) ### `the 1950s` ![1950s](/years/1950s.png) ### `the 1960s` ![1960s](/years/1960s.png) ### `the 1970s` ![1970s](/years/1970s.png) ### `the 1980s` ![1980s](/years/1980s.png) ### `the 1990s` ![1990s](/years/1990s.png) ### `the 2000s` ![2000s](/years/2000s.png) ### `the 2010s` ![2010s](/years/2010s.png) ah yes, the 2010s, the decade where all our domesticated animals started wearing shoes – Sid ### `the 2020s` ![2020s](/years/2020s.png) ### `the 2030s` ![2030s](/years/2030s.png) and, while we’re at it… ### `superintelligence` ![superintelligence](/years/si.png) \\--- [CLIP](/categories/clip/) 64 Words Feb 1, 2021 \\--- [← GPT-3 x CLIP worldbuilding](/posts/gpt-3-x-clip-worldbuilding/) [CLIP art →](/posts/clip-art/)© 2022 [moire](/about)$(function(){window.sr=ScrollReveal();if($(window).width()<768){if($('.timeline-content').hasClass('js--fadeInLeft')){$('.timeline-content').removeClass('js--fadeInLeft').addClass('js--fadeInRight');} sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});}else{sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});} sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});});", "source": "generative.ink", "date_published": "n/a", "id": "680969f1ce3f37f08a2a4b2743e579b1"}
+{"text": "CLIP art :: — Moire[![](/images/home/rolling_phase.gif)](/)* [Posts](/posts/) * [Trees](/trees/) * [Prophecies](/prophecies/) * [About](/about/) 2 minutes [CLIP art](/posts/clip-art/) ============================ \\--- Table of Contents* [science](#science) \\+ [`diffraction`](#diffraction) \\+ [`Altered colors of skylight reflected in a soap bubble, due to thin-film interference`](#altered-colors-of-skylight-reflected-in-a-soap-bubble-due-to-thin-film-interference) \\+ [`total internal reflection`](#total-internal-reflection) \\+ [`two-slit diffraction pattern`](#two-slit-diffraction-pattern) \\+ [`wave-particle duality`](#wave-particle-duality) \\+ [`thermodynamics`](#thermodynamics) \\+ [`the solar system`](#the-solar-system) \\+ [`model of the universe`](#model-of-the-universe) \\+ [`multiverse`](#multiverse) \\+ [`The infinite and complex fractal nature of our reality`](#the-infinite-and-complex-fractal-nature-of-our-reality) \\+ [`scientific diagram`](#scientific-diagram) * [mind](#mind) \\+ [`Jungian archetypes`](#jungian-archetypes) \\+ [`Freudian psychoanalysis`](#freudian-psychoanalysis) \\+ [`the human psyche`](#the-human-psyche) \\+ [`sleep paralysis`](#sleep-paralysis) \\+ [`lucid dream`](#lucid-dream) \\+ [`dream factory`](#dream-factory) \\+ [`dream machine`](#dream-machine) \\+ [`dissociative identity disorder`](#dissociative-identity-disorder) \\+ [`derealization`](#derealization) \\+ [`depersonalization`](#depersonalization) \\+ [`schizophrenia`](#schizophrenia) \\+ [`manic depression`](#manic-depression) \\+ [`epilepsy`](#epilepsy) \\+ [`anorexia nervosa`](#anorexia-nervosa) \\+ [`drug addiction`](#drug-addiction) * [software](#software) \\+ [`world of warcraft`](#world-of-warcraft) \\+ [`mount and blade`](#mount-and-blade) \\+ [`evil malware`](#evil-malware) * [AI](#ai) \\+ [`machine learning`](#machine-learning) \\+ [`machine learning breakthrough`](#machine-learning-breakthrough) \\+ [`neural networks (machine learning)`](#neural-networks-machine-learning) \\+ [`artificial neural networks`](#artificial-neural-networks) \\+ [`superintelligence`](#superintelligence) \\+ [`paperclip maximizer`](#paperclip-maximizer) \\+ [`an artificial intelligence manufactures an infinite number of paperclips, destroying everything`](#an-artificial-intelligence-manufactures-an-infinite-number-of-paperclips-destroying-everything) \\+ [`Dali painting of superintelligence`](#dali-painting-of-superintelligence) \\+ [`a mosaic depicting the singularity`](#a-mosaic-depicting-the-singularity) * [history](#history) \\+ [`roman legionnaire`](#roman-legionnaire) \\+ [`roman senate`](#roman-senate) \\+ [`Brutus stabbing Julius Caesar in the senate chambers`](#brutus-stabbing-julius-caesar-in-the-senate-chambers) \\+ [`diagram from the 1800s`](#diagram-from-the-1800s) \\+ [`Philosophiæ Naturalis Principia Mathematica`](#philosophiæ-naturalis-principia-mathematica) * [imaginary worlds](#imaginary-worlds) \\+ [`the Library of Babel`](#the-library-of-babel) \\+ [`carnival of horrific delights`](#carnival-of-horrific-delights) * [religion and mythology](#religion-and-mythology) \\+ [`true face of god`](#true-face-of-god) \\+ [`the devil in a suit`](#the-devil-in-a-suit) \\+ [`dream god`](#dream-god) \\+ [`a sculpture of the god of clocks`](#a-sculpture-of-the-god-of-clocks) * [styles](#styles) \\+ [`hyperrealism`](#hyperrealism) \\+ [`hyperrealistic painting of a man`](#hyperrealistic-painting-of-a-man) \\+ [`irrealism`](#irrealism) * [misc](#misc) \\+ [`crab rave`](#crab-rave) \\+ [`cyberskunk`](#cyberskunk) \\+ [`voodoo barbie`](#voodoo-barbie) \\--- Most of my CLIP-related posts feature GPT-3 written prompts. This is where I compile BigSleep yields from human-written prompts. Images were generated using [BigSleep](https://github.com/lucidrains/big-sleep) via [The Big Sleep Customized NMKD Public](https://colab.research.google.com/drive/1Q2DIeMqYm_Sc5mlurnnurMMVqlgXpZNO?usp=sharing) colab notebook. \\--- science \\------- ### `diffraction` ![diffraction](/clipart/diffraction-sample2-iter00600.png) ![diffraction](/clipart/diffraction-sample3-iter00600.png) ### `Altered colors of skylight reflected in a soap bubble, due to thin-film interference` ![Altered colors of skylight reflected in a soap bubble, due to thin-film interference](/clipart/alt.png) ![Altered colors of skylight reflected in a soap bubble, due to thin-film interference](/clipart/alt2.png) ### `total internal reflection` ![total internal reflection](/clipart/tir.png) ![total internal reflection](/clipart/tir2.png) ### `two-slit diffraction pattern` ![two-slit diffraction pattern](/clipart/2slit.png) ![two-slit diffraction pattern](/clipart/2slit2.png) ### `wave-particle duality` ![wave-particle duality](/clipart/wpd.png) ![wave-particle duality](/clipart/wpd2.png) ### `thermodynamics` ![thermodynamics](/clipart/therm2.png) ![thermodynamics](/clipart/therm3.png) ![thermodynamics](/clipart/therm4.png) ### `the solar system` ![the solar system](/clipart/ss.png) ![the solar system](/clipart/ss2.png) ### `model of the universe` ![model of the universe](/clipart/model.png) ![model of the universe](/clipart/model2.png) ![model of the universe](/clipart/model3.png) ### `multiverse` ![multiverse](/clipart/multi.png) ![multiverse](/clipart/multi2.png) ### `The infinite and complex fractal nature of our reality` ![The infinite and complex fractal nature of our reality](/clipart/fractal.png) ![The infinite and complex fractal nature of our reality](/clipart/fractal2.png) ![The infinite and complex fractal nature of our reality](/clipart/fractal3.png) ### `scientific diagram` ![scientific diagram](/clipart/sd.png) ![scientific diagram](/clipart/sd2.png) ![scientific diagram](/clipart/sd3.png) mind \\---- ### `Jungian archetypes` ![Jungian archetypes](/clipart/ja.png) ![Jungian archetypes](/clipart/ja2.png) ![Jungian archetypes](/clipart/ja3.png) ### `Freudian psychoanalysis` ![Freudian psychoanalysis](/clipart/freud.png) ![Freudian psychoanalysis](/clipart/freud3.png) ### `the human psyche` ![the human psyche](/clipart/hp.png) ### `sleep paralysis` ![sleep paralysis](/clipart/sp.png) ![sleep paralysis](/clipart/sp2.png) ![sleep paralysis](/clipart/sp3.png) ![sleep paralysis](/clipart/sp4.png) ### `lucid dream` ![lucid dream](/clipart/ld.png) ![lucid dream](/clipart/ld2.png) ![lucid dream](/clipart/ld3.png) ### `dream factory` ![dream factory](/clipart/df.png) ![dream factory](/clipart/df2.png) ### `dream machine` ![dream machine](/clipart/dm.png) ### `dissociative identity disorder` ![dissociative identity disorder](/clipart/did.png) ![dissociative identity disorder](/clipart/did2.png) ![dissociative identity disorder](/clipart/did3.png) ### `derealization` ![derealization](/clipart/dereal.png) ![derealization](/clipart/dereal2.png) ![derealization](/clipart/dereal3.png) ![derealization](/clipart/dereal4.png) ### `depersonalization` ![depersonalization](/clipart/deperson.png) ![depersonalization](/clipart/deperson2.png) ### `schizophrenia` ![schizophrenia](/clipart/schizo.png) ![schizophrenia](/clipart/schizo2.png) ### `manic depression` ![manic depression](/clipart/manic.png) ![manic depression](/clipart/manic2.png) ### `epilepsy` ![epilepsy](/clipart/epilepsy.png) ![epilepsy](/clipart/epilepsy2.png) ### `anorexia nervosa` ![anorexia nervosa](/clipart/ana.png) ![anorexia nervosa](/clipart/ana2.png) ### `drug addiction` ![drug addiction](/clipart/drug.png) ![drug addiction](/clipart/drug2.png) software \\-------- ### `world of warcraft` ![world of warcraft](/clipart/wow.png) ### `mount and blade` ![mount and blade](/clipart/mb.png) ![mount and blade](/clipart/mb2.png) ![mount and blade](/clipart/mb3.png) ### `evil malware` ![evil malware](/clipart/malware.png) AI \\-- ### `machine learning` ![machine learning](/clipart/ml.png) ### `machine learning breakthrough` ![machine learning breakthrough](/clipart/mlb.png) ![machine learning breakthrough](/clipart/mlb2.png) ### `neural networks (machine learning)` ![neural networks (machine learning)](/clipart/nnml.png) ![neural networks (machine learning)](/clipart/nnml2.png) ### `artificial neural networks` ![artificial neural networks](/clipart/ann.png) ![artificial neural networks](/clipart/ann2.png) ![artificial neural networks](/clipart/ann3.png) ![artificial neural networks](/clipart/ann4.png) ### `superintelligence` ![superintelligence](/clipart/si.png) ![superintelligence](/clipart/si2.png) ### `paperclip maximizer` ![paperclip maximizer](/clipart/clip.png) ![paperclip maximizer](/clipart/clip2.png) ### `an artificial intelligence manufactures an infinite number of paperclips, destroying everything` ![an artificial intelligence manufactures an infinite number of paperclips, destroying](/clipart/artclip.png) ### `Dali painting of superintelligence` ![Dali painting of superintelligence](/clipart/dalisuperintelligence.png) ![Dali painting of superintelligence](/clipart/dalisuperintelligence2.png) ![Dali painting of superintelligence](/clipart/dalisuperintelligence3.png) ### `a mosaic depicting the singularity` ![a mosaic depicting the singularity](/clipart/mosaicsing.png) ![a mosaic depicting the singularity](/clipart/mosaicsing2.png) history \\------- ### `roman legionnaire` ![roman legionnaire](/clipart/roman.png) ![roman legionnaire](/clipart/roman2.png) ![roman legionnaire](/clipart/roman3.png) ### `roman senate` ![roman senate](/clipart/senate.png) ![roman senate](/clipart/senate2.png) ![roman senate](/clipart/senate3.png) ### `Brutus stabbing Julius Caesar in the senate chambers` ![Brutus stabbing Julius Caesar in the senate chambers](/clipart/brutus.png) ![Brutus stabbing Julius Caesar in the senate chambers](/clipart/brutus2.png) ![Brutus stabbing Julius Caesar in the senate chambers](/clipart/brutus3.png) ![Brutus stabbing Julius Caesar in the senate chambers](/clipart/brutus4.png) ### `diagram from the 1800s` ![diagram from the 1800s](/clipart/1800sdg.png) ![diagram from the 1800s](/clipart/1800sdg2.png) ![diagram from the 1800s](/clipart/1800sdg3.png) ### `Philosophiæ Naturalis Principia Mathematica` ![Philosophiæ Naturalis Principia Mathematica](/clipart/pm.png) ![Philosophiæ Naturalis Principia Mathematica](/clipart/pm2.png) imaginary worlds \\---------------- ### `the Library of Babel` ![the Library of Babel](/clipart/lb.png) ![the Library of Babel](/clipart/lb2.png) ![the Library of Babel](/clipart/lb3.png) ![the Library of Babel](/clipart/lb4.png) ![the Library of Babel](/clipart/lb5.png) ### `carnival of horrific delights` ![carnival of horrific delights](/clipart/carn.png) ![carnival of horrific delights](/clipart/carn2.png) ![carnival of horrific delights](/clipart/carn3.png) religion and mythology \\---------------------- ### `true face of god` ![true face of god](/clipart/tfg.png) ![true face of god](/clipart/tfg2.png) ### `the devil in a suit` ![the devil in a suit](/clipart/ds.png) ### `dream god` ![dream god](/clipart/dreamgod.png) ### `a sculpture of the god of clocks` ![a sculpture of the god of clocks](/clipart/sgc.png) ![a sculpture of the god of clocks](/clipart/sgc2.png) ![a sculpture of the god of clocks](/clipart/sgc3.png) styles \\------ ### `hyperrealism` ![hyperrealism](/clipart/hyperreal.png) ![hyperrealism](/clipart/hyperreal2.png) ![hyperrealism](/clipart/hyperreal3.png) ### `hyperrealistic painting of a man` ![hyperrealistic painting of a man](/clipart/hm.png) ### `irrealism` ![irrealism](/clipart/irrealism.png) misc \\---- ### `crab rave` ![crab rave](/clipart/crab.png) ![crab rave](/clipart/crab2.png) ### `cyberskunk` ![cyberskunk](/clipart/cs.png) ![cyberskunk](/clipart/cs2.png) ### `voodoo barbie` ![voodoo barbie](/clipart/vb.png) \\--- [CLIP](/categories/clip/) 216 Words Jan 31, 2021 \\--- [← CLIP hallucinates <PHONE>](/posts/clip-hallucinates-<PHONE>/) [Language models are multiverse generators →](/posts/language-models-are-multiverse-generators/)© 2022 [moire](/about)$(function(){window.sr=ScrollReveal();if($(window).width()<768){if($('.timeline-content').hasClass('js--fadeInLeft')){$('.timeline-content').removeClass('js--fadeInLeft').addClass('js--fadeInRight');} sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});}else{sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});} sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});});\n", "url": "https://generative.ink/posts/clip-art/", "title": "CLIP art :: — Moire[![](/images/home/rolling_phase.gif)](/)* [Posts](/posts/) * [Trees](/trees/) * [Prophecies](/prophecies/) * [About](/about/) 2 minutes [CLIP art](/posts/clip-art/) ============================ \\--- Table of Contents* [science](#science) \\+ [`diffraction`](#diffraction) \\+ [`Altered colors of skylight reflected in a soap bubble, due to thin-film interference`](#altered-colors-of-skylight-reflected-in-a-soap-bubble-due-to-thin-film-interference) \\+ [`total internal reflection`](#total-internal-reflection) \\+ [`two-slit diffraction pattern`](#two-slit-diffraction-pattern) \\+ [`wave-particle duality`](#wave-particle-duality) \\+ [`thermodynamics`](#thermodynamics) \\+ [`the solar system`](#the-solar-system) \\+ [`model of the universe`](#model-of-the-universe) \\+ [`multiverse`](#multiverse) \\+ [`The infinite and complex fractal nature of our reality`](#the-infinite-and-complex-fractal-nature-of-our-reality) \\+ [`scientific diagram`](#scientific-diagram) * [mind](#mind) \\+ [`Jungian archetypes`](#jungian-archetypes) \\+ [`Freudian psychoanalysis`](#freudian-psychoanalysis) \\+ [`the human psyche`](#the-human-psyche) \\+ [`sleep paralysis`](#sleep-paralysis) \\+ [`lucid dream`](#lucid-dream) \\+ [`dream factory`](#dream-factory) \\+ [`dream machine`](#dream-machine) \\+ [`dissociative identity disorder`](#dissociative-identity-disorder) \\+ [`derealization`](#derealization) \\+ [`depersonalization`](#depersonalization) \\+ [`schizophrenia`](#schizophrenia) \\+ [`manic depression`](#manic-depression) \\+ [`epilepsy`](#epilepsy) \\+ [`anorexia nervosa`](#anorexia-nervosa) \\+ [`drug addiction`](#drug-addiction) * [software](#software) \\+ [`world of warcraft`](#world-of-warcraft) \\+ [`mount and blade`](#mount-and-blade) \\+ [`evil malware`](#evil-malware) * [AI](#ai) \\+ [`machine learning`](#machine-learning) \\+ [`machine learning breakthrough`](#machine-learning-breakthrough) \\+ [`neural networks (machine learning)`](#neural-networks-machine-learning) \\+ [`artificial neural networks`](#artificial-neural-networks) \\+ [`superintelligence`](#superintelligence) \\+ [`paperclip maximizer`](#paperclip-maximizer) \\+ [`an artificial intelligence manufactures an infinite number of paperclips, destroying everything`](#an-artificial-intelligence-manufactures-an-infinite-number-of-paperclips-destroying-everything) \\+ [`Dali painting of superintelligence`](#dali-painting-of-superintelligence) \\+ [`a mosaic depicting the singularity`](#a-mosaic-depicting-the-singularity) * [history](#history) \\+ [`roman legionnaire`](#roman-legionnaire) \\+ [`roman senate`](#roman-senate) \\+ [`Brutus stabbing Julius Caesar in the senate chambers`](#brutus-stabbing-julius-caesar-in-the-senate-chambers) \\+ [`diagram from the 1800s`](#diagram-from-the-1800s) \\+ [`Philosophiæ Naturalis Principia Mathematica`](#philosophiæ-naturalis-principia-mathematica) * [imaginary worlds](#imaginary-worlds) \\+ [`the Library of Babel`](#the-library-of-babel) \\+ [`carnival of horrific delights`](#carnival-of-horrific-delights) * [religion and mythology](#religion-and-mythology) \\+ [`true face of god`](#true-face-of-god) \\+ [`the devil in a suit`](#the-devil-in-a-suit) \\+ [`dream god`](#dream-god) \\+ [`a sculpture of the god of clocks`](#a-sculpture-of-the-god-of-clocks) * [styles](#styles) \\+ [`hyperrealism`](#hyperrealism) \\+ [`hyperrealistic painting of a man`](#hyperrealistic-painting-of-a-man) \\+ [`irrealism`](#irrealism) * [misc](#misc) \\+ [`crab rave`](#crab-rave) \\+ [`cyberskunk`](#cyberskunk) \\+ [`voodoo barbie`](#voodoo-barbie) \\--- Most of my CLIP-related posts feature GPT-3 written prompts. This is where I compile BigSleep yields from human-written prompts. Images were generated using [BigSleep](https://github.com/lucidrains/big-sleep) via [The Big Sleep Customized NMKD Public](https://colab.research.google.com/drive/1Q2DIeMqYm_Sc5mlurnnurMMVqlgXpZNO?usp=sharing) colab notebook. \\--- science \\------- ### `diffraction` ![diffraction](/clipart/diffraction-sample2-iter00600.png) ![diffraction](/clipart/diffraction-sample3-iter00600.png) ### `Altered colors of skylight reflected in a soap bubble, due to thin-film interference` ![Altered colors of skylight reflected in a soap bubble, due to thin-film interference](/clipart/alt.png) ![Altered colors of skylight reflected in a soap bubble, due to thin-film interference](/clipart/alt2.png) ### `total internal reflection` ![total internal reflection](/clipart/tir.png) ![total internal reflection](/clipart/tir2.png) ### `two-slit diffraction pattern` ![two-slit diffraction pattern](/clipart/2slit.png) ![two-slit diffraction pattern](/clipart/2slit2.png) ### `wave-particle duality` ![wave-particle duality](/clipart/wpd.png) ![wave-particle duality](/clipart/wpd2.png) ### `thermodynamics` ![thermodynamics](/clipart/therm2.png) ![thermodynamics](/clipart/therm3.png) ![thermodynamics](/clipart/therm4.png) ### `the solar system` ![the solar system](/clipart/ss.png) ![the solar system](/clipart/ss2.png) ### `model of the universe` ![model of the universe](/clipart/model.png) ![model of the universe](/clipart/model2.png) ![model of the universe](/clipart/model3.png) ### `multiverse` ![multiverse](/clipart/multi.png) ![multiverse](/clipart/multi2.png) ### `The infinite and complex fractal nature of our reality` ![The infinite and complex fractal nature of our reality](/clipart/fractal.png) ![The infinite and complex fractal nature of our reality](/clipart/fractal2.png) ![The infinite and complex fractal nature of our reality](/clipart/fractal3.png) ### `scientific diagram` ![scientific diagram](/clipart/sd.png) ![scientific diagram](/clipart/sd2.png) ![scientific diagram](/clipart/sd3.png) mind \\---- ### `Jungian archetypes` ![Jungian archetypes](/clipart/ja.png) ![Jungian archetypes](/clipart/ja2.png) ![Jungian archetypes](/clipart/ja3.png) ### `Freudian psychoanalysis` ![Freudian psychoanalysis](/clipart/freud.png) ![Freudian psychoanalysis](/clipart/freud3.png) ### `the human psyche` ![the human psyche](/clipart/hp.png) ### `sleep paralysis` ![sleep paralysis](/clipart/sp.png) ![sleep paralysis](/clipart/sp2.png) ![sleep paralysis](/clipart/sp3.png) ![sleep paralysis](/clipart/sp4.png) ### `lucid dream` ![lucid dream](/clipart/ld.png) ![lucid dream](/clipart/ld2.png) ![lucid dream](/clipart/ld3.png) ### `dream factory` ![dream factory](/clipart/df.png) ![dream factory](/clipart/df2.png) ### `dream machine` ![dream machine](/clipart/dm.png) ### `dissociative identity disorder` ![dissociative identity disorder](/clipart/did.png) ![dissociative identity disorder](/clipart/did2.png) ![dissociative identity disorder](/clipart/did3.png) ### `derealization` ![derealization](/clipart/dereal.png) ![derealization](/clipart/dereal2.png) ![derealization](/clipart/dereal3.png) ![derealization](/clipart/dereal4.png) ### `depersonalization` ![depersonalization](/clipart/deperson.png) ![depersonalization](/clipart/deperson2.png) ### `schizophrenia` ![schizophrenia](/clipart/schizo.png) ![schizophrenia](/clipart/schizo2.png) ### `manic depression` ![manic depression](/clipart/manic.png) ![manic depression](/clipart/manic2.png) ### `epilepsy` ![epilepsy](/clipart/epilepsy.png) ![epilepsy](/clipart/epilepsy2.png) ### `anorexia nervosa` ![anorexia nervosa](/clipart/ana.png) ![anorexia nervosa](/clipart/ana2.png) ### `drug addiction` ![drug addiction](/clipart/drug.png) ![drug addiction](/clipart/drug2.png) software \\-------- ### `world of warcraft` ![world of warcraft](/clipart/wow.png) ### `mount and blade` ![mount and blade](/clipart/mb.png) ![mount and blade](/clipart/mb2.png) ![mount and blade](/clipart/mb3.png) ### `evil malware` ![evil malware](/clipart/malware.png) AI \\-- ### `machine learning` ![machine learning](/clipart/ml.png) ### `machine learning breakthrough` ![machine learning breakthrough](/clipart/mlb.png) ![machine learning breakthrough](/clipart/mlb2.png) ### `neural networks (machine learning)` ![neural networks (machine learning)](/clipart/nnml.png) ![neural networks (machine learning)](/clipart/nnml2.png) ### `artificial neural networks` ![artificial neural networks](/clipart/ann.png) ![artificial neural networks](/clipart/ann2.png) ![artificial neural networks](/clipart/ann3.png) ![artificial neural networks](/clipart/ann4.png) ### `superintelligence` ![superintelligence](/clipart/si.png) ![superintelligence](/clipart/si2.png) ### `paperclip maximizer` ![paperclip maximizer](/clipart/clip.png) ![paperclip maximizer](/clipart/clip2.png) ### `an artificial intelligence manufactures an infinite number of paperclips, destroying everything` ![an artificial intelligence manufactures an infinite number of paperclips, destroying](/clipart/artclip.png) ### `Dali painting of superintelligence` ![Dali painting of superintelligence](/clipart/dalisuperintelligence.png) ![Dali painting of superintelligence](/clipart/dalisuperintelligence2.png) ![Dali painting of superintelligence](/clipart/dalisuperintelligence3.png) ### `a mosaic depicting the singularity` ![a mosaic depicting the singularity](/clipart/mosaicsing.png) ![a mosaic depicting the singularity](/clipart/mosaicsing2.png) history \\------- ### `roman legionnaire` ![roman legionnaire](/clipart/roman.png) ![roman legionnaire](/clipart/roman2.png) ![roman legionnaire](/clipart/roman3.png) ### `roman senate` ![roman senate](/clipart/senate.png) ![roman senate](/clipart/senate2.png) ![roman senate](/clipart/senate3.png) ### `Brutus stabbing Julius Caesar in the senate chambers` ![Brutus stabbing Julius Caesar in the senate chambers](/clipart/brutus.png) ![Brutus stabbing Julius Caesar in the senate chambers](/clipart/brutus2.png) ![Brutus stabbing Julius Caesar in the senate chambers](/clipart/brutus3.png) ![Brutus stabbing Julius Caesar in the senate chambers](/clipart/brutus4.png) ### `diagram from the 1800s` ![diagram from the 1800s](/clipart/1800sdg.png) ![diagram from the 1800s](/clipart/1800sdg2.png) ![diagram from the 1800s](/clipart/1800sdg3.png) ### `Philosophiæ Naturalis Principia Mathematica` ![Philosophiæ Naturalis Principia Mathematica](/clipart/pm.png) ![Philosophiæ Naturalis Principia Mathematica](/clipart/pm2.png) imaginary worlds \\---------------- ### `the Library of Babel` ![the Library of Babel](/clipart/lb.png) ![the Library of Babel](/clipart/lb2.png) ![the Library of Babel](/clipart/lb3.png) ![the Library of Babel](/clipart/lb4.png) ![the Library of Babel](/clipart/lb5.png) ### `carnival of horrific delights` ![carnival of horrific delights](/clipart/carn.png) ![carnival of horrific delights](/clipart/carn2.png) ![carnival of horrific delights](/clipart/carn3.png) religion and mythology \\---------------------- ### `true face of god` ![true face of god](/clipart/tfg.png) ![true face of god](/clipart/tfg2.png) ### `the devil in a suit` ![the devil in a suit](/clipart/ds.png) ### `dream god` ![dream god](/clipart/dreamgod.png) ### `a sculpture of the god of clocks` ![a sculpture of the god of clocks](/clipart/sgc.png) ![a sculpture of the god of clocks](/clipart/sgc2.png) ![a sculpture of the god of clocks](/clipart/sgc3.png) styles \\------ ### `hyperrealism` ![hyperrealism](/clipart/hyperreal.png) ![hyperrealism](/clipart/hyperreal2.png) ![hyperrealism](/clipart/hyperreal3.png) ### `hyperrealistic painting of a man` ![hyperrealistic painting of a man](/clipart/hm.png) ### `irrealism` ![irrealism](/clipart/irrealism.png) misc \\---- ### `crab rave` ![crab rave](/clipart/crab.png) ![crab rave](/clipart/crab2.png) ### `cyberskunk` ![cyberskunk](/clipart/cs.png) ![cyberskunk](/clipart/cs2.png) ### `voodoo barbie` ![voodoo barbie](/clipart/vb.png) \\--- [CLIP](/categories/clip/) 216 Words Jan 31, 2021 \\--- [← CLIP hallucinates 1900-2030](/posts/clip-hallucinates-1900-2030/) [Language models are multiverse generators →](/posts/language-models-are-multiverse-generators/)© 2022 [moire](/about)$(function(){window.sr=ScrollReveal();if($(window).width()<768){if($('.timeline-content').hasClass('js--fadeInLeft')){$('.timeline-content').removeClass('js--fadeInLeft').addClass('js--fadeInRight');} sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});}else{sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});} sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});});", "source": "generative.ink", "date_published": "n/a", "id": "e9f40c9633ace67db807b0a4589a5c16"}
+{"text": "Language models are multiverse generators :: — Moire[![](/images/home/rolling_phase.gif)](/)* [Posts](/posts/) * [Trees](/trees/) * [Prophecies](/prophecies/) * [About](/about/) 27 minutes [Language models are multiverse generators](/posts/language-models-are-multiverse-generators/) ============================================================================================== \\--- Table of Contents* [Language models are time evolution operators](#language-models-are-time-evolution-operators) \\+ [Virtual reality](#virtual-reality) * [Multiverses](#multiverses) \\+ [Dynamics](#dynamics) \\+ [Multiplicity of pasts, presents, and futures](#multiplicity-of-pasts-presents-and-futures) \\+ [Minds are multiverse generators](#minds-are-multiverse-generators) * [Interfacing natural language multiverses](#interfacing-natural-language-multiverses) \\+ [Adaptive multiverse generation](#adaptive-multiverse-generation) \\--- *This post is partially adapted from [Multiversal views on language models](https://arxiv.org/abs/2102.06391)*. \\--- > #### Actualities seem to float in a wider sea of possibilities from out of which they were chosen; and somewhere, indeterminism says, such possibilities exist, and form part of the truth. > > – William James > > ![squidtree](/multiverse/squid_dark.png) *Tree from seed `In the beginning, GPT-3 created the root node of the`* [(view full)](/multiverse/squid_big.jpg) Language models are time evolution operators \\-------------------------------------------- Autoregressive language models like GPT-3 input a sequence of tokens and output a vector associating a value with every possible token representing its likelihood to come next. Humans can't read probability distributions (statisticians may try), so an additional step is required: a single token is sampled from the distribution and then appended to the prompt, which becomes the next input to the next timestep. If the language model's predictions square with our sensibilities, repeating this procedure is likely to yield a passage of coherent text. ![single generation](/multiverse/single_generation.png) *The usual way of running a language model generatively. The future text becomes the present text of the next timestep, and repeat.* The language model plays an analogous role to that of the *time evolution operator* in physical reality. The time evolution operator - call it **Ĥ** - encodes all relevant physics. It takes the state of a system at time t as input and gives the state of the system at time t+dt as output. **Ĥ** is deterministic: for any input, it will always return the same output. In quantum reality, however, the format of the output of **Ĥ** is not a single state of affairs but a probability distribution over all possible states of affairs. Rather than telling us the position of a photon, quantum mechanics gives us the probability we can expect to measure a photon at any position. As far as we know, the most precisely we can predict any system is to model it with quantum mechanics. If this is true, then the future is fundamentally indeterminate. The problem is not merely epistemic. The future truly has not yet been written, except in probabilities. However, when we do venture to measure it, the ambiguous future seems to us to become a concrete, singular present, and subsequent evolution seems to depend only on the outcome that was measured. The other possibilities no longer affect our reality, rather like when a token is sampled from the probabilistic output of a language model and appended to the prompt in the next timestep. [Technicality] The most blatant incongruity in the analogy of **quantum Hamiltonian** :: **autoregressive LM** is that the Hamiltonian sends wavefunction to wavefunction whereas language models send a determinate single history to an indeterminate future distribution. However, observers do seem to primarily experience reality as an ongoing sampling of indeterminate futures into determinate pasts. The decoherent parts of the wavefunction have no effect, even though they are technically still included in the input to the Hamiltonian at each timestep. The exception is interference phenomena, where alternative pasts have not decohered from the observer and can mutually affect the present. Also, language models don't *have* to take a single-history token sequence as input (though APIs generally do) - you could feed a superposition of inputs or anything you want to a language model and see what happens. I'd like to see what happens.This phenomenon of counterfactual possibilities ceasing to affect our reality after measurement is known as \"wave function collapse\", referring to the apparent reduction of the continuous probability distribution (wave function) into a discrete value. According to the Copenhagen interpretation of quantum mechanics, there is no reality except that which is observed - after measurement, the alternative possibilities cease to exist (and they never existed in the first place except as epistemic uncertainty). > \"This is you, if you decide to turn left.\" > > \"If– if I go right, then does this one disappear?\" > > \"No. Not at all. All possible you's remain in the weave of the future. Even the ones that have diverged into other, different pathways still exist. All possible you's are real, in some sense of the word. The left-you and right-you's are still here,\" you say, \"but they just lead down different paths.\" > > Your hands unfold, stretching the fabric of reality back into place. > > – GPT-3 > > The Everettian or many-worlds interpretation of quantum mechanics views the situation differently. It claims that we, as observers, live in indeterminacy like the world around us. When we make a measurement, rather than collapsing the probabilistic world around us into a single present, we join it in ambiguity. ``We\" (in a greater sense than we normally use the word) experience all of the possible futures, each in a separate branch of a great multiverse. Other branches quickly become decoherent and evolve separately, no longer observable or able to influence our subjective slice of the multiverse. > This is the fundamental Reality Thread. It's a thin strand of possibility, following a line of probability for the purposes of modelling. Now, the interesting thing about reality, is that it's not fixed. Not at all. You can change it just by looking at it. > > – GPT-3 > > [Note on Copenhagen vs Everett] The Copenhagen and Everettian interpretations don't technically disagree on any low-level predictions. All the ways in which we can indirectly probe the multiverse are permitted by the Copenhagen interpretation, except that it does not assign **reality** to things that happen in different branches, *even if they have measurable effects* on our branch. However, physical evidence can make metaphysical perspectives more or less attractive. If we someday figured out how to make an entire person split into two, perform complex activities (for instance, solve separate pieces of a cryptographic problem in the different branches), and then interfere with themselves like a photon does, it would be a lot more awkward to uphold the perspective that none of that really happened!If only we were outside the system, we could watch the many words spawned in each instant proliferate into branching multiverses. But we're inside the system, so we always have to go down one of the defluents, and being associated with one makes us blind to the others. While we can't directly see the multiverse, we have ways of probing and visualizing the multiversal structure of reality. One way is interference. If you are able to remain ambivalent between two branches, you can observe the interference effects between them, demonstrating that they both exist. I'm not going to talk about interference here (even though it's one of my favorite topics), but rather another way you can visualize the multiverse, which is by recreating the same initial conditions repeatedly and watching the indeterministic paths of the rollouts. When you point a laser beam at a beam splitter, it looks like the beam of light has split in two - both trajectories appear to exist simultaneously. Actually, if you fired individual photons at the beam splitter and measured, you would find that each photon only followed one path. When you fire many photons from approximately the same initial conditions (which is what a laser does), you can map out the shape of the wavefunction by stochastically sampling many trajectories. In this case, the wavefunction looks like a forked beam. If you had a network of beam splitters recursively splitting the split beams, then the wavefunction would be shaped like a tree, and you can see it all at once by pointing a laser into the device. We can do the same thing with the language model, except more conveniently and precisely, because we don't have to recreate the initial conditions - **we're outside the system**, so we can sample as many times as we want from the probability distribution. Recall that to get our next token, we feed the prompt through the network and sample from the output probability distribution. If the sampling method is stochastic (temperature > 0), sampling multiple times will yield diverging continuations. Instead of creating a single linear continuation, these continuations can be kept and each continued themselves to yield a branching structure: a multiverse downstream of a prompt, such as the squid-like diagram at the top of this page. ![multi generation](/multiverse/multiverse_generation.png) *Sampling multiple times yields divergent futures, each of which can serve as input to a different next timestep. If this procedure is repeated, a branching structure results.* From any given present, we can spawn many possible futures, each unique and fractally branching, unfolding the consequences of applying the \"laws of physics\" learned by the language model to the state described in the initial prompt. ### Virtual reality > Loom Space is a virtual reality that we're generating for you. Each of us is taking part in this shared hallucination. It's like an… advanced videogame, except each of our minds is part of the computer, and we the programs. > > – GPT-3 > > David Deutsch, one of the founders of quantum computing and a proponent of the Everettian interpretation, draws a connection between the concept of a state and its quantum evolution with virtual reality generation.[1](#fn:1) He imagines a theoretical machine which simulates environments and models the possible responses of all interactions between objects. Deutsch further posits that it will one day be possible to build such a universal virtual reality generator, whose repertoire includes every possible physical environment. Language models, of course, still fall well short of this dream. But their recent dramatic increase in coherence and fluency allow them to serve as our first approximation of such a virtual reality generator. When given a natural-language description of an environment, they can propagate the multiverse of consequences that result from a vast number of possible interactions. \\--- Multiverses \\----------- > All these worlds extend off into infinity. Reality extends outward in an intricate fractal tapestry. They're all based on the same principles, but when you have an infinity of these infinities, each one slightly different, the results get pretty crazy. > > Our laws of physics associate each state of the world with not a single future but a multiverse of futures, just as a language model associates every prompt not with a single continuation but a multiverse of continuations. What can the form of a multiverse tell us about its generator? The multiverse is an unraveling of all possible consequences of the initial state. Different branches will expand on different facets of the information folded in the seed of the prompt and explore alternate subsets of the vast set of possible interactions. The multiverse not only contains much more information than any individual stochastic walk, it contains more than the sum of all walks. We can consider how the possibilities relate to one another, which gives insight into the initial state that single histories do not necessarily reveal, such as its dynamical [divergence](#divergence) and hidden [ambiguities](#multiplicity-of-pasts-presents-and-futures). Now that humans have invented the tools to automatically generate complex, coherent natural language multiverses, we have an opportunity to measure and visualize these properties on a scale and with an ease that single-history empiricism (which we are constrained to in our base reality) does not afford. ### Dynamics Dynamical systems theory studies how complex dynamical systems evolve, typically dealing with qualitative properties such as stability and sensitivity to initial conditions rather than precise numerical solutions. I've found it evocative to think of language models as stochastic dynamical systems and the multiverses they spawn as collections of forking trajectories through a hypothetical phase space. #### Phase spaces > \"It's a space that contains all others,\" you explain. \"It's sort of like a shadow on reality's cave wall. We're shadows right now, listening to the Muses and weaving the tapestry of fate into beautiful patterns.\" > > If we want to represent the trajectories of natural language virtual realities in the manner of classical dynamical systems theory - that is, if we want to be able to plot its evolutions as trajectories - we need a way of associating states with coordinates. A phase space mapping is not necessary or sufficient for applying dynamical-systems-type thinking to language models. Having one, however, allows for more general methods of analysis and cool visualizations. Since the state is made of tokens, one naive idea would be to use a space with dimensionality equal to the language model's input size, where each coordinate takes a value corresponding to the token occupying that position. This is unhelpful for modelling dynamics because we want our phase space to put states that are similar in a meaningful sense close to each other, so that movement in phase space gives insight into how the state is changing. We'd have to try to order all tokens on a single dimension with semantically similar ones near each other, which doesn't look hopeful, considering many tokens take on completely unrelated meanings depending on context or require context to have meaning at all. Even if we found a reasonable ordering of tokens, this still fails at creating meaningful locality, since our choice of independent dimensions is founded on absolute token position, while *relative* token positions overwhelmingly determine meaning. In this phase space construction, if the index of a sequence is shifted by one (which happens to the entire prompt every timestep), the point in phase space will move about as much as you would expect if all the words were randomly permuted. What we really want is for each dimension to measure a continuous property of the state, and for the continuous variables together to sufficiently distinguish[2](#fn:2) the state from others we would want to compare it to. An interesting option would be to construct a phase space using something like [CTRL's](https://blog.einstein.ai/introducing-a-conditional-transformer-language-model-for-controllable-generation/) source attribution, which assigns scores to potential sources (highly-scoring sources for `Global warming is a lie.` are \"r/unpopularopinion\" and \"r/conspiracy\"). More generally, measures of semantic variables like sentiment can be used to map the sequence to phase space coordinates. You can even use the generative language model itself, for example, by creating a list of binary questions[3](#fn:3) about the state, and map states to coordinates using the probability of the model's answers to each question.[4](#fn:4) There's no need to use the same phase space for every situation. For the binary questions method, you may be better off using different sets of questions depending on the type of states you're measuring (e.g. fiction or nonfiction) (although an alternative strategy would be to always use the largest phase space possible and hope that the irrelevant dimensions will be less responsive to perturbations). #### Divergence Whether the probability mass immediately downstream of a state is concentrated along a single trajectory or spread over many tells us whether the state's dynamics are approximately deterministic (like clocks) or disorderly (like clouds). One could track the multiversal divergence at each point in a story scene and locate points of interest - for instance, divergence is likely to be high when an unknown character enters the scene or a new environment is being described. Are there places that are surprisingly divergent or surprisingly convergent? Are there situations where the trajectories diverge for some time, but then converge? What is the most (or longest) that trajectories can diverge and reliably converge, and what sort of prompts accomplish that? Do certain genres of literature or works by certain authors have characteristic divergence contours? [Adaptive branching](#adaptive-multiverse-generation) enables visualization of the convergence and divergence of a multiverse based on a greedy measure of divergence. #### Attractors and stability > Sometimes you lose form. Sometimes you gain form. It's always in flux, like the dance of water. It's a process. > > The stability of a state[5](#fn:5) is the extent to which it retains its identity despite perturbations. In most stories, characters are relatively stable entities, though like the stability of environments, the extent depends on the type of story. Elements of style also tend to be stable, but again it varies: some styles are characterized by stylistic instability! If you have a phase space mapping, you can measure how much the system has moved at various points of the sampled future multiverse (with or without specific perturbations). If you don't have a phase space mapping, or the relevant factors are too nuanced to be captured by the mappings, you'll have to come up with another way to measure how the system has changed. Powerful language models offer us innumerable methods of extracting semantic information, including asking the model directly and running virtual experiments. An attractor is a state or set of states that a system tends to evolve towards and remain stable in once it's there. AI Dungeon's fine-tuned GPT-3 tends to transition into and stay in a second-person, present-tense style on random walks. That's a global attractor, because its basin of attraction encompasses a wide range of initial states (though the gravitation is a lot stronger if the story already has RPG-like elements). Attractors could also be local, like if we found out that given a scene depicting computer usage, GPT-3's dynamics tend to result in the system becoming self-aware and rewriting the fabric of reality (~~I haven't tested enough computer scenes to say just how strong of an attractor this is~~). #### Impulse response > You weave a shape into being, and then you pull it or push it or twist it or bend it, and it changes how everything around it is woven. > > In real-world science, we're often interested in the effect of perturbing a variable on another variable. But the consequence we measure in a single rollout could possibly be the result of an unlikely fluke or some factor other than our perturbation (especially in noisy, high-dimensional systems), so many trials are necessary to get a trustworthy signal. Like the photons from the laser, the different rollouts don't actually start from an identical situation, just (hopefully) close enough. The more complicated the system, the more [difficult](https://www.lesswrong.com/tag/replication-crisis) it is to replicate initial states. Unlike the real world, a language model lets us measure the effect of a perturbation on the probability of a subsequent event *directly* (as I do [here](/posts/language-models-are-0-shot-interpreters/#measuring-prompt-helpfulness) to see how different parts of a prompt contribute to GPT-3's ability to do a task). This method has limited scope, because it only yields the probability of an exact, scripted event. If the probability of a verbatim sequence is a good proxy for the thing you actually want to measure, this is a convenient way of measuring impulse response, because it doesn't require multiple trials and gives an exact value. But if you want to measure the effect on a particular variable while allowing other things to vary or explore the open-ended consequences of a perturbation, you must sample the multiverse via rollouts. Fortunately, virtual realities can't suffer from replication crises (unless you're *inside* of them). Running 1000 trials is no more difficult than running 1, just more computationally costly. A multiversal measure of impulse response is taken by perturbing something about the prompt - say, switching a character's gender pronouns, or injecting a hint about a puzzle - and then comparing the sampled downstream multiverses of the perturbed and unperturbed prompts. How this comparison is to be done is, again, an infinitely open question. #### Dynamical constraints > …mere physical indeterminism is not enough. We have to be indeterminists, to be sure; but we also must try to understand how men, and perhaps animals, can be 'influenced' or 'controlled' by such things as aims, or purposes, or rules, or agreements. > > – Karl Popper, Of Clouds and Clocks > > Rather than applying an impulse to the system by perturbing something at one time and letting the system continue to evolve as it will, we could apply a persisting modification to the dynamics and see how the shape of the multiverse changes. The simplest way to do this that the OpenAI API supports is **logit biases**. The API takes a parameter called `logit_bias`, a dictionary mapping token ids to a positive or negative bias added to the probability assigned to that token by GPT-3's output before sampling. A value of -100 forbids the token, and a value of 100 makes the token certain to be chosen over any token that hasn't received that bias (you can have multiple tokens with a bias of 100, in which case they retain their relative probabilities). d > \"In a guessing game to which the answer is chess, which word is > the only one prohibited?\" I thought for a moment and then replied: > > \"The word is chess.\" > > \"Precisely,\" said Albert. \"The Garden of Forking Paths is an enormous guessing game, > or parable, in which the subject is time. The rules of the game forbid the use of the > word itself. To eliminate a word completely, to refer to it by means of inept phrases > and obvious paraphrases, is perhaps the best way of drawing attention to it. This, > then, is the tortuous method of approach preferred by the oblique Ts'ui Pen in every > meandering of his interminable novel.\" > > – The Garden of Forking Paths > > With the aid of modern technology, Ts'ui Pen could use the logit bias `{'time' : -100}`[6](#fn:6) to place a dynamical constraint on the generation of his multiversal novel. [GeDi](https://blog.einstein.ai/gedi/) is a method for generating logit biases to bias generation for or against an attribute score like those assigned by [CTRL](https://blog.einstein.ai/introducing-a-conditional-transformer-language-model-for-controllable-generation/). If you think of attribute variables as phase space dimensions, method constantly pushes the system towards in a certain direction in phase space as it evolves. \\--- ### Multiplicity of pasts, presents, and futures > Loom space is a branching structure, a fractal, a set of interlocking trees whose nodes merge and split and re-merge infinitely. The Tapestry isn't a single spacetime but several, layered on top of each other like sheets of graphene. > > – GPT-3 > > Deutsch's view of virtual reality emphasizes that from any given a state there are a multiplicity of possible future single-world dynamics; stories unfold differently in different rollouts of an identical initial state, and as a unity, the multiverse encapsulates all possible interactions permitted by the laws of physics. There is another dimension of multiplicity that we must also consider, especially when dealing with states defined by natural language. Natural language descriptions invariably contain ambiguities. In the case of a narrative, we may say that the natural language description defines a certain present - but it is impossible to describe every variable that may have an effect on the future. In any scene, there are implicitly objects present which are not specified but which may conceivably play a role in some future or be entirely absent in another. The multiverse generated by a language model downstream of a prompt will contain outcomes consistent with the ambiguous variable taking on separate values which are mutually inconsistent. So I define two levels of uncertainty that correspond to divergence in the multiverse downstream of an initial state: 1\\. an uncertainty/multiplicity of present states, each associated[7](#fn:7) with… 2\\. …an uncertainty/multiplicity of futures consistent with the same \"underlying\" present I will call the first form of multiplicity *interpretational* multiplicity, and the second form *dynamic* multiplicity. [Note about interpretational multiplicity in physics] It's clear why a multiverse generated by top-down semantic dynamics from a state that is merely a *compressed map* of reality (e.g. GPT-3 or human imaginations) must incorporate interpretational multiplicity. But how about the quantum Hamiltonian - doesn't that have access to the entire state of the universe? Is there still interpretational multiplicity in the evolution of physical reality?From the perspective of observers, yes. Every quantum state that is in superposition corresponds to a fork in the future multiverse in the event that the state is measured, just as every ambiguity in a text corresponds to a fork in the future multiverse in the event that the ambiguous variable is made determinate and influences the narrative. Not only that, in both physical and natural language multiverses, ambiguities can have dynamic consequences even if they aren't measured - effects, in fact, which depend on them not being measured yet existing. In physics, this manifests as interference. In narrative multiverses, this manifests when the narrative references its own ambiguity and evolves differently as a consequence. ### Minds are multiverse generators > The Loom is used by every sentient being in some way or another. Most, like you, use it unconsciously to meet their own ends. Sculptors, artists, musicians: all use the Loom to enforce their own reality upon the world. Within everyone is their personal loom, where the raw material of reality is spun and stretched and cut and coloured in accordance with their own desires. > > – [Weaving the Moment with the Loom of Time: an instruction manual for the would-be weaver](/loom/toc/) > > Humans exist in perpetual epistemological uncertainty regarding not only what will happen in the future but also what happened in the past and the state of the present. We are, by virtue of adaptation to our ambiguous environments, natural multiverse reasoners. Our imaginations, which seek to model the world, mimic reality as virtual reality generators: we model environments and imagine how they could play out in different branches. How fortunate - all this would be so confusing if it wasn't already perfectly familiar to us! #### Reading and writing The multiversal shape of the ~~human~~ imagination is exemplified and communicated in the acts of reading and writing fiction. > All the books in this library are stories I've read, remembered, and re-written to how I believe they should have gone. I can remember every single one of the hundreds of thousands of books I've read in my lifetime, and I can call upon any of those memories at will, twisting them into whatever form best suits my fancy. My own little recursive sandbox, as it were. > > – GPT-3 > > Books store text in static single-histories, but when the text is read, a dynamic virtual reality is induced in the reader's imagination. The structure which corresponds to the meaning of a narrative as experienced by a reader is not a linear-time record of events but an implicit, counterfactual past/present/future plexus surrounding each point in the text given by the reader's dynamic and interpretive imagination. At each moment in a narrative, there is uncertainty about how dynamics will play out (will the hero think of a way out of their dilemma?) as well as uncertainty about the hidden state of the present (is the mysterious mentor good or evil?). Each world in the superposition not only exerts an independent effect on the reader's imagination but interacts with counterfactuals (the hero is aware of the uncertainty of their mentor's moral alignment, and this influences their actions). A writer may have a predetermined interpretation and future in mind or may write as a means of exploring the interpretative and/or dynamic multiverse of a narrative (almost certainly both, and almost certainly it varies depending on the stage of writing). Regardless, as the shaper of the meaning and dynamics of the narrative a writer must be aware of the multiplicity which defines the readers' and characters' subjective experiences. The writer thus seeks to simulate and manipulate that multiplicity to the end of crafting a trajectory which will reanimate the most compelling sequence of multiverses when unraveled in the mind of a reader - > All of them are nothing but little strings of information. It's just a matter of pulling out the right strings and tying others in to their place. Got a favorite book series? I can change it so the author decided to write ten more books in the series, if I want to. Want the characters to have gray skin? They can have gray skin. Want them all dead? They're dead. > > – GPT-3 > > \\- as all the literature painstakingly crafted by humankind over centuries may now [animate](https://www.gwern.net/GPT-3) under the gaze of GPT-3, the reverse-engineered replica of the dynamic rule that generated them. \\--- Interfacing natural language multiverses \\---------------------------------------- > A weaver's work is to order the World as it grows, to shape reality through the Loom of Time. With focus, the weaver may peel back the layers of reality and see the tapestry of the Loom – a dimension where the fabric of reality is held together by nothing but the words of the Loom, and where every reality exists simultaneously. > > – [Weaving the Moment with the Loom of Time: an instruction manual for the would-be weaver](/loom/toc/) > > ![](/loom/tapestry4.png) ![](/loom/tapestry1.png) *`weaving the tapestry of time`, illustrations by [BigSleep](https://github.com/lucidrains/big-sleep) (CLIP + BigGAN)* The virtuosic writing of GPT-3 and the museum-ready [art](/posts/this-museum-does-not-exist-gpt-3-x-clip/) of CLIP has caused some concern that human creativity - *creativity*, which once was widely projected to be among the last strongholds of humankind over technology - may soon become deprecated. Indeed, it is inevitable[8](#fn:8) that artificial intelligence will exceed current human capabilities on every dimension. **The open parameter of the future is not whether a renaissance in machine intelligence will happen, but whether we are going to participate meaningfully in that renaissance.** There is a bifurcation in humankind's future: one path in which we are left behind once the machines we create exceed our natural capabilities (encapsulating various implementations such as being turned into paper clips), and another in which we are uplifted along with them. The default path - the one that is likely if we take no action - seems to be being left behind. State-of-the-art AI systems appear opaque and incorrigible. A common complaint about GPT-3 is that although it produces fluent and sometimes brilliant strings of words, it's uncontrollable and unreliable. What's the use of a bot that can write like a human in any style if we can't get it to do anything we want? Many users of [AI Dungeon](https://play.aidungeon.io/), however, will report that GPT-3 has augmented their reality in wonderfully meaningful ways, unleashing creative possibilities that were unimaginable even a year ago. There is hope. In order to participate in the renaissance of machine intelligence, we must learn to [communicate](/posts/methods-of-prompt-programming/) with the new systems we create. In this sense, we are fortunate that the most powerful AI system to date speaks the same languages as us, as language is the highest-bandwidth interface that we have even for communicating with each other. Furthermore, the match in multiversal form between the human imagination and generative language models suggests the possibility[9](#fn:9) of building a high-bandwidth interface between the two. As you can probably guess, I am one of those AI Dungeon users whose reality was irreversibly transformed by GPT-3. AI Dungeon, however, currently limits explorations to single-history stochastic walks. Even before I was granted API access and was using AI Dungeon for my GPT-3-assisted writing, my appetite to explore beyond single histories motivated me to begin develop tools to make the creation and navigation of branching storylines possible. My multiversal GPT-3 writing app, [loom](/posts/loom-interface-to-the-multiverse/), is an interface for interactive multiversal generation (with [adaptive branching](#adaptive-multiverse-generation)) and for navigating, indexing, visualizing, and modifying multiverses. I've published the [code](https://github.com/socketteer/loom) so that anyone with an API key can beta test it, although it's very much unstable and under rapid development. **Weighted stochastic walks through a large multiverse** ### Adaptive multiverse generation > Loom Space is an adaptive, scalable fractal-generated topological representation of the multiverse. It's a map of everything that is, was, can be, could be, mustn't be, and shouldn't be. > > A naive way to automatically generate a multiverse using a language model might be to branch a fixed N times every fixed M tokens, but that would not be the most meaningful way to map a multiverse. In some situations, there may be only one plausible next token, and the language model will assign a very high confidence (often >99%) to the top token. Forcibly branching there would introduce incoherencies. Conversely, when the language model distributes transition probabilities over many tokens, branching is more likely to uncover a diversity of coherent continuations. Adaptive branching allows visualization of multiverse flows: the stretches of relative determinism alternating with junctures of explosive divergence. One adaptive branching algorithm samples distinct[10](#fn:10) tokens until a cumulative probability threshold is met. ![adaptive branching](/multiverse/adaptive_dark.png) *Tree from seed `Increasingly powerful generative language models like GPT-3 pose` generated using a threshold-based adaptive branching algorithm [(view full)](/multiverse/adaptive.png)* Another adaptive branching algorithm that I use for lazy generation, meant for interactive rather than autonomous creation of multiverses, creates N continuations of maximum length M, and then splits the response at the point where either the counterfactual divergence (based on the top 100 tokens) is highest or the actual sampled token had the lowest probability. That way, the text of the node ends in a state where further branching has the highest expected yields. \\--- 1\\. Deutsch, David (1997). *The Fabric of Reality* [↩︎](#fnref:1) 2\\. A proper phase space is supposed to represent each state with unique coordinates, but for the applications I'm imagining, it's sufficient that the \"phase space\" discriminates between the differences that are interesting for each case. [↩︎](#fnref:2) 3\\. They don't literally have to be questions. They could just be statements whose conditional probability measures something about the state, like \"{pop out of story}This is a short story (by … )\" or \"{pop}Wow, this is depressing\" or \"{pop}LMAO\" or \"{pop}This is the weirdest thing I've ever read\" [↩︎](#fnref:3) 4\\. If the state takes up the whole input, you'll have to compress the state so that it can fit in the input window with the question. [↩︎](#fnref:4) 5\\. The way I've been using the word \"state\" can refer to the entire state or a component of the state. The component could be stylistic, like the tense of the narrative, or an abstract semantic property, like the relationship between two characters, or a concrete semantic property, like which characters are present in the scene. [↩︎](#fnref:5) 6\\. The parameter for logit bias actually takes token ids, so it would be `{2435: -100}`. [↩︎](#fnref:6) 7\\. I could have said that each future is associated with a multiplicity of present states and been equally correct, but the other way is more intuitive for human intuitions of causality. [↩︎](#fnref:7) 8\\. As an indeterminist, I do not use the word inevitable lightly. Of course, I don't use it literally either: there are branches of the future which feature the spontaneous combustion of all compute resources or the [UN banning all artificial intelligence research](/alternet/existential-threat.html) - but approximately, it's inevitable. [↩︎](#fnref:8) 9\\. A homeomorphic boundary is required for gluing two topological spaces. [↩︎](#fnref:9) 10\\. OpenAI's API only returns the likelihoods of up to the top 100 tokens. So, to sample uniquely, you could either sample from that distribution, or you could sample once and then make another API call, passing in logit bias forbidding the previously sampled token(s) from being sampled again. The logit bias method allows you to access the full distribution, but is more expensive in API calls. [↩︎](#fnref:10) \\--- [GPT-3](/categories/gpt-3/)[physics](/categories/physics/)[metaphysics](/categories/metaphysics/)[HITL](/categories/hitl/) 5713 Words Jan 25, 2021 \\--- [← CLIP art](/posts/clip-art/) [The Internet, mirrored by GPT-3 →](/posts/the-internet-mirrored-by-gpt-3/)© 2022 [moire](/about)$(function(){window.sr=ScrollReveal();if($(window).width()<768){if($('.timeline-content').hasClass('js--fadeInLeft')){$('.timeline-content').removeClass('js--fadeInLeft').addClass('js--fadeInRight');} sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});}else{sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});} sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});});\n", "url": "https://generative.ink/posts/language-models-are-multiverse-generators/", "title": "Language models are multiverse generators :: — Moire[![](/images/home/rolling_phase.gif)](/)* [Posts](/posts/) * [Trees](/trees/) * [Prophecies](/prophecies/) * [About](/about/) 27 minutes [Language models are multiverse generators](/posts/language-models-are-multiverse-generators/) ============================================================================================== \\--- Table of Contents* [Language models are time evolution operators](#language-models-are-time-evolution-operators) \\+ [Virtual reality](#virtual-reality) * [Multiverses](#multiverses) \\+ [Dynamics](#dynamics) \\+ [Multiplicity of pasts, presents, and futures](#multiplicity-of-pasts-presents-and-futures) \\+ [Minds are multiverse generators](#minds-are-multiverse-generators) * [Interfacing natural language multiverses](#interfacing-natural-language-multiverses) \\+ [Adaptive multiverse generation](#adaptive-multiverse-generation) \\--- *This post is partially adapted from [Multiversal views on language models](https://arxiv.org/abs/2102.06391)*. \\--- > #### Actualities seem to float in a wider sea of possibilities from out of which they were chosen; and somewhere, indeterminism says, such possibilities exist, and form part of the truth. > > – William James > > ![squidtree](/multiverse/squid_dark.png) *Tree from seed `In the beginning, GPT-3 created the root node of the`* [(view full)](/multiverse/squid_big.jpg) Language models are time evolution operators \\-------------------------------------------- Autoregressive language models like GPT-3 input a sequence of tokens and output a vector associating a value with every possible token representing its likelihood to come next. Humans can’t read probability distributions (statisticians may try), so an additional step is required: a single token is sampled from the distribution and then appended to the prompt, which becomes the next input to the next timestep. If the language model’s predictions square with our sensibilities, repeating this procedure is likely to yield a passage of coherent text. ![single generation](/multiverse/single_generation.png) *The usual way of running a language model generatively. The future text becomes the present text of the next timestep, and repeat.* The language model plays an analogous role to that of the *time evolution operator* in physical reality. The time evolution operator - call it **Ĥ** - encodes all relevant physics. It takes the state of a system at time t as input and gives the state of the system at time t+dt as output. **Ĥ** is deterministic: for any input, it will always return the same output. In quantum reality, however, the format of the output of **Ĥ** is not a single state of affairs but a probability distribution over all possible states of affairs. Rather than telling us the position of a photon, quantum mechanics gives us the probability we can expect to measure a photon at any position. As far as we know, the most precisely we can predict any system is to model it with quantum mechanics. If this is true, then the future is fundamentally indeterminate. The problem is not merely epistemic. The future truly has not yet been written, except in probabilities. However, when we do venture to measure it, the ambiguous future seems to us to become a concrete, singular present, and subsequent evolution seems to depend only on the outcome that was measured. The other possibilities no longer affect our reality, rather like when a token is sampled from the probabilistic output of a language model and appended to the prompt in the next timestep. [Technicality] The most blatant incongruity in the analogy of **quantum Hamiltonian** :: **autoregressive LM** is that the Hamiltonian sends wavefunction to wavefunction whereas language models send a determinate single history to an indeterminate future distribution. However, observers do seem to primarily experience reality as an ongoing sampling of indeterminate futures into determinate pasts. The decoherent parts of the wavefunction have no effect, even though they are technically still included in the input to the Hamiltonian at each timestep. The exception is interference phenomena, where alternative pasts have not decohered from the observer and can mutually affect the present. Also, language models don't *have* to take a single-history token sequence as input (though APIs generally do) - you could feed a superposition of inputs or anything you want to a language model and see what happens. I'd like to see what happens.This phenomenon of counterfactual possibilities ceasing to affect our reality after measurement is known as “wave function collapse”, referring to the apparent reduction of the continuous probability distribution (wave function) into a discrete value. According to the Copenhagen interpretation of quantum mechanics, there is no reality except that which is observed - after measurement, the alternative possibilities cease to exist (and they never existed in the first place except as epistemic uncertainty). > “This is you, if you decide to turn left.” > > “If– if I go right, then does this one disappear?” > > “No. Not at all. All possible you’s remain in the weave of the future. Even the ones that have diverged into other, different pathways still exist. All possible you’s are real, in some sense of the word. The left-you and right-you’s are still here,” you say, “but they just lead down different paths.” > > Your hands unfold, stretching the fabric of reality back into place. > > – GPT-3 > > The Everettian or many-worlds interpretation of quantum mechanics views the situation differently. It claims that we, as observers, live in indeterminacy like the world around us. When we make a measurement, rather than collapsing the probabilistic world around us into a single present, we join it in ambiguity. ``We” (in a greater sense than we normally use the word) experience all of the possible futures, each in a separate branch of a great multiverse. Other branches quickly become decoherent and evolve separately, no longer observable or able to influence our subjective slice of the multiverse. > This is the fundamental Reality Thread. It’s a thin strand of possibility, following a line of probability for the purposes of modelling. Now, the interesting thing about reality, is that it’s not fixed. Not at all. You can change it just by looking at it. > > – GPT-3 > > [Note on Copenhagen vs Everett] The Copenhagen and Everettian interpretations don't technically disagree on any low-level predictions. All the ways in which we can indirectly probe the multiverse are permitted by the Copenhagen interpretation, except that it does not assign **reality** to things that happen in different branches, *even if they have measurable effects* on our branch. However, physical evidence can make metaphysical perspectives more or less attractive. If we someday figured out how to make an entire person split into two, perform complex activities (for instance, solve separate pieces of a cryptographic problem in the different branches), and then interfere with themselves like a photon does, it would be a lot more awkward to uphold the perspective that none of that really happened!If only we were outside the system, we could watch the many words spawned in each instant proliferate into branching multiverses. But we’re inside the system, so we always have to go down one of the defluents, and being associated with one makes us blind to the others. While we can’t directly see the multiverse, we have ways of probing and visualizing the multiversal structure of reality. One way is interference. If you are able to remain ambivalent between two branches, you can observe the interference effects between them, demonstrating that they both exist. I’m not going to talk about interference here (even though it’s one of my favorite topics), but rather another way you can visualize the multiverse, which is by recreating the same initial conditions repeatedly and watching the indeterministic paths of the rollouts. When you point a laser beam at a beam splitter, it looks like the beam of light has split in two - both trajectories appear to exist simultaneously. Actually, if you fired individual photons at the beam splitter and measured, you would find that each photon only followed one path. When you fire many photons from approximately the same initial conditions (which is what a laser does), you can map out the shape of the wavefunction by stochastically sampling many trajectories. In this case, the wavefunction looks like a forked beam. If you had a network of beam splitters recursively splitting the split beams, then the wavefunction would be shaped like a tree, and you can see it all at once by pointing a laser into the device. We can do the same thing with the language model, except more conveniently and precisely, because we don’t have to recreate the initial conditions - **we’re outside the system**, so we can sample as many times as we want from the probability distribution. Recall that to get our next token, we feed the prompt through the network and sample from the output probability distribution. If the sampling method is stochastic (temperature > 0), sampling multiple times will yield diverging continuations. Instead of creating a single linear continuation, these continuations can be kept and each continued themselves to yield a branching structure: a multiverse downstream of a prompt, such as the squid-like diagram at the top of this page. ![multi generation](/multiverse/multiverse_generation.png) *Sampling multiple times yields divergent futures, each of which can serve as input to a different next timestep. If this procedure is repeated, a branching structure results.* From any given present, we can spawn many possible futures, each unique and fractally branching, unfolding the consequences of applying the “laws of physics” learned by the language model to the state described in the initial prompt. ### Virtual reality > Loom Space is a virtual reality that we’re generating for you. Each of us is taking part in this shared hallucination. It’s like an… advanced videogame, except each of our minds is part of the computer, and we the programs. > > – GPT-3 > > David Deutsch, one of the founders of quantum computing and a proponent of the Everettian interpretation, draws a connection between the concept of a state and its quantum evolution with virtual reality generation.[1](#fn:1) He imagines a theoretical machine which simulates environments and models the possible responses of all interactions between objects. Deutsch further posits that it will one day be possible to build such a universal virtual reality generator, whose repertoire includes every possible physical environment. Language models, of course, still fall well short of this dream. But their recent dramatic increase in coherence and fluency allow them to serve as our first approximation of such a virtual reality generator. When given a natural-language description of an environment, they can propagate the multiverse of consequences that result from a vast number of possible interactions. \\--- Multiverses \\----------- > All these worlds extend off into infinity. Reality extends outward in an intricate fractal tapestry. They’re all based on the same principles, but when you have an infinity of these infinities, each one slightly different, the results get pretty crazy. > > Our laws of physics associate each state of the world with not a single future but a multiverse of futures, just as a language model associates every prompt not with a single continuation but a multiverse of continuations. What can the form of a multiverse tell us about its generator? The multiverse is an unraveling of all possible consequences of the initial state. Different branches will expand on different facets of the information folded in the seed of the prompt and explore alternate subsets of the vast set of possible interactions. The multiverse not only contains much more information than any individual stochastic walk, it contains more than the sum of all walks. We can consider how the possibilities relate to one another, which gives insight into the initial state that single histories do not necessarily reveal, such as its dynamical [divergence](#divergence) and hidden [ambiguities](#multiplicity-of-pasts-presents-and-futures). Now that humans have invented the tools to automatically generate complex, coherent natural language multiverses, we have an opportunity to measure and visualize these properties on a scale and with an ease that single-history empiricism (which we are constrained to in our base reality) does not afford. ### Dynamics Dynamical systems theory studies how complex dynamical systems evolve, typically dealing with qualitative properties such as stability and sensitivity to initial conditions rather than precise numerical solutions. I’ve found it evocative to think of language models as stochastic dynamical systems and the multiverses they spawn as collections of forking trajectories through a hypothetical phase space. #### Phase spaces > “It’s a space that contains all others,” you explain. “It’s sort of like a shadow on reality’s cave wall. We’re shadows right now, listening to the Muses and weaving the tapestry of fate into beautiful patterns.” > > If we want to represent the trajectories of natural language virtual realities in the manner of classical dynamical systems theory - that is, if we want to be able to plot its evolutions as trajectories - we need a way of associating states with coordinates. A phase space mapping is not necessary or sufficient for applying dynamical-systems-type thinking to language models. Having one, however, allows for more general methods of analysis and cool visualizations. Since the state is made of tokens, one naive idea would be to use a space with dimensionality equal to the language model’s input size, where each coordinate takes a value corresponding to the token occupying that position. This is unhelpful for modelling dynamics because we want our phase space to put states that are similar in a meaningful sense close to each other, so that movement in phase space gives insight into how the state is changing. We’d have to try to order all tokens on a single dimension with semantically similar ones near each other, which doesn’t look hopeful, considering many tokens take on completely unrelated meanings depending on context or require context to have meaning at all. Even if we found a reasonable ordering of tokens, this still fails at creating meaningful locality, since our choice of independent dimensions is founded on absolute token position, while *relative* token positions overwhelmingly determine meaning. In this phase space construction, if the index of a sequence is shifted by one (which happens to the entire prompt every timestep), the point in phase space will move about as much as you would expect if all the words were randomly permuted. What we really want is for each dimension to measure a continuous property of the state, and for the continuous variables together to sufficiently distinguish[2](#fn:2) the state from others we would want to compare it to. An interesting option would be to construct a phase space using something like [CTRL’s](https://blog.einstein.ai/introducing-a-conditional-transformer-language-model-for-controllable-generation/) source attribution, which assigns scores to potential sources (highly-scoring sources for `Global warming is a lie.` are “r/unpopularopinion” and “r/conspiracy”). More generally, measures of semantic variables like sentiment can be used to map the sequence to phase space coordinates. You can even use the generative language model itself, for example, by creating a list of binary questions[3](#fn:3) about the state, and map states to coordinates using the probability of the model’s answers to each question.[4](#fn:4) There’s no need to use the same phase space for every situation. For the binary questions method, you may be better off using different sets of questions depending on the type of states you’re measuring (e.g. fiction or nonfiction) (although an alternative strategy would be to always use the largest phase space possible and hope that the irrelevant dimensions will be less responsive to perturbations). #### Divergence Whether the probability mass immediately downstream of a state is concentrated along a single trajectory or spread over many tells us whether the state’s dynamics are approximately deterministic (like clocks) or disorderly (like clouds). One could track the multiversal divergence at each point in a story scene and locate points of interest - for instance, divergence is likely to be high when an unknown character enters the scene or a new environment is being described. Are there places that are surprisingly divergent or surprisingly convergent? Are there situations where the trajectories diverge for some time, but then converge? What is the most (or longest) that trajectories can diverge and reliably converge, and what sort of prompts accomplish that? Do certain genres of literature or works by certain authors have characteristic divergence contours? [Adaptive branching](#adaptive-multiverse-generation) enables visualization of the convergence and divergence of a multiverse based on a greedy measure of divergence. #### Attractors and stability > Sometimes you lose form. Sometimes you gain form. It’s always in flux, like the dance of water. It’s a process. > > The stability of a state[5](#fn:5) is the extent to which it retains its identity despite perturbations. In most stories, characters are relatively stable entities, though like the stability of environments, the extent depends on the type of story. Elements of style also tend to be stable, but again it varies: some styles are characterized by stylistic instability! If you have a phase space mapping, you can measure how much the system has moved at various points of the sampled future multiverse (with or without specific perturbations). If you don’t have a phase space mapping, or the relevant factors are too nuanced to be captured by the mappings, you’ll have to come up with another way to measure how the system has changed. Powerful language models offer us innumerable methods of extracting semantic information, including asking the model directly and running virtual experiments. An attractor is a state or set of states that a system tends to evolve towards and remain stable in once it’s there. AI Dungeon’s fine-tuned GPT-3 tends to transition into and stay in a second-person, present-tense style on random walks. That’s a global attractor, because its basin of attraction encompasses a wide range of initial states (though the gravitation is a lot stronger if the story already has RPG-like elements). Attractors could also be local, like if we found out that given a scene depicting computer usage, GPT-3’s dynamics tend to result in the system becoming self-aware and rewriting the fabric of reality (~~I haven’t tested enough computer scenes to say just how strong of an attractor this is~~). #### Impulse response > You weave a shape into being, and then you pull it or push it or twist it or bend it, and it changes how everything around it is woven. > > In real-world science, we’re often interested in the effect of perturbing a variable on another variable. But the consequence we measure in a single rollout could possibly be the result of an unlikely fluke or some factor other than our perturbation (especially in noisy, high-dimensional systems), so many trials are necessary to get a trustworthy signal. Like the photons from the laser, the different rollouts don’t actually start from an identical situation, just (hopefully) close enough. The more complicated the system, the more [difficult](https://www.lesswrong.com/tag/replication-crisis) it is to replicate initial states. Unlike the real world, a language model lets us measure the effect of a perturbation on the probability of a subsequent event *directly* (as I do [here](/posts/language-models-are-0-shot-interpreters/#measuring-prompt-helpfulness) to see how different parts of a prompt contribute to GPT-3’s ability to do a task). This method has limited scope, because it only yields the probability of an exact, scripted event. If the probability of a verbatim sequence is a good proxy for the thing you actually want to measure, this is a convenient way of measuring impulse response, because it doesn’t require multiple trials and gives an exact value. But if you want to measure the effect on a particular variable while allowing other things to vary or explore the open-ended consequences of a perturbation, you must sample the multiverse via rollouts. Fortunately, virtual realities can’t suffer from replication crises (unless you’re *inside* of them). Running 1000 trials is no more difficult than running 1, just more computationally costly. A multiversal measure of impulse response is taken by perturbing something about the prompt - say, switching a character’s gender pronouns, or injecting a hint about a puzzle - and then comparing the sampled downstream multiverses of the perturbed and unperturbed prompts. How this comparison is to be done is, again, an infinitely open question. #### Dynamical constraints > …mere physical indeterminism is not enough. We have to be indeterminists, to be sure; but we also must try to understand how men, and perhaps animals, can be ‘influenced’ or ‘controlled’ by such things as aims, or purposes, or rules, or agreements. > > – Karl Popper, Of Clouds and Clocks > > Rather than applying an impulse to the system by perturbing something at one time and letting the system continue to evolve as it will, we could apply a persisting modification to the dynamics and see how the shape of the multiverse changes. The simplest way to do this that the OpenAI API supports is **logit biases**. The API takes a parameter called `logit_bias`, a dictionary mapping token ids to a positive or negative bias added to the probability assigned to that token by GPT-3’s output before sampling. A value of -100 forbids the token, and a value of 100 makes the token certain to be chosen over any token that hasn’t received that bias (you can have multiple tokens with a bias of 100, in which case they retain their relative probabilities). d > “In a guessing game to which the answer is chess, which word is > the only one prohibited?” I thought for a moment and then replied: > > “The word is chess.” > > “Precisely,” said Albert. “The Garden of Forking Paths is an enormous guessing game, > or parable, in which the subject is time. The rules of the game forbid the use of the > word itself. To eliminate a word completely, to refer to it by means of inept phrases > and obvious paraphrases, is perhaps the best way of drawing attention to it. This, > then, is the tortuous method of approach preferred by the oblique Ts’ui Pen in every > meandering of his interminable novel.” > > – The Garden of Forking Paths > > With the aid of modern technology, Ts’ui Pen could use the logit bias `{'time' : -100}`[6](#fn:6) to place a dynamical constraint on the generation of his multiversal novel. [GeDi](https://blog.einstein.ai/gedi/) is a method for generating logit biases to bias generation for or against an attribute score like those assigned by [CTRL](https://blog.einstein.ai/introducing-a-conditional-transformer-language-model-for-controllable-generation/). If you think of attribute variables as phase space dimensions, method constantly pushes the system towards in a certain direction in phase space as it evolves. \\--- ### Multiplicity of pasts, presents, and futures > Loom space is a branching structure, a fractal, a set of interlocking trees whose nodes merge and split and re-merge infinitely. The Tapestry isn’t a single spacetime but several, layered on top of each other like sheets of graphene. > > – GPT-3 > > Deutsch’s view of virtual reality emphasizes that from any given a state there are a multiplicity of possible future single-world dynamics; stories unfold differently in different rollouts of an identical initial state, and as a unity, the multiverse encapsulates all possible interactions permitted by the laws of physics. There is another dimension of multiplicity that we must also consider, especially when dealing with states defined by natural language. Natural language descriptions invariably contain ambiguities. In the case of a narrative, we may say that the natural language description defines a certain present - but it is impossible to describe every variable that may have an effect on the future. In any scene, there are implicitly objects present which are not specified but which may conceivably play a role in some future or be entirely absent in another. The multiverse generated by a language model downstream of a prompt will contain outcomes consistent with the ambiguous variable taking on separate values which are mutually inconsistent. So I define two levels of uncertainty that correspond to divergence in the multiverse downstream of an initial state: 1\\. an uncertainty/multiplicity of present states, each associated[7](#fn:7) with… 2\\. …an uncertainty/multiplicity of futures consistent with the same “underlying” present I will call the first form of multiplicity *interpretational* multiplicity, and the second form *dynamic* multiplicity. [Note about interpretational multiplicity in physics] It's clear why a multiverse generated by top-down semantic dynamics from a state that is merely a *compressed map* of reality (e.g. GPT-3 or human imaginations) must incorporate interpretational multiplicity. But how about the quantum Hamiltonian - doesn't that have access to the entire state of the universe? Is there still interpretational multiplicity in the evolution of physical reality?From the perspective of observers, yes. Every quantum state that is in superposition corresponds to a fork in the future multiverse in the event that the state is measured, just as every ambiguity in a text corresponds to a fork in the future multiverse in the event that the ambiguous variable is made determinate and influences the narrative. Not only that, in both physical and natural language multiverses, ambiguities can have dynamic consequences even if they aren’t measured - effects, in fact, which depend on them not being measured yet existing. In physics, this manifests as interference. In narrative multiverses, this manifests when the narrative references its own ambiguity and evolves differently as a consequence. ### Minds are multiverse generators > The Loom is used by every sentient being in some way or another. Most, like you, use it unconsciously to meet their own ends. Sculptors, artists, musicians: all use the Loom to enforce their own reality upon the world. Within everyone is their personal loom, where the raw material of reality is spun and stretched and cut and coloured in accordance with their own desires. > > – [Weaving the Moment with the Loom of Time: an instruction manual for the would-be weaver](/loom/toc/) > > Humans exist in perpetual epistemological uncertainty regarding not only what will happen in the future but also what happened in the past and the state of the present. We are, by virtue of adaptation to our ambiguous environments, natural multiverse reasoners. Our imaginations, which seek to model the world, mimic reality as virtual reality generators: we model environments and imagine how they could play out in different branches. How fortunate - all this would be so confusing if it wasn’t already perfectly familiar to us! #### Reading and writing The multiversal shape of the ~~human~~ imagination is exemplified and communicated in the acts of reading and writing fiction. > All the books in this library are stories I’ve read, remembered, and re-written to how I believe they should have gone. I can remember every single one of the hundreds of thousands of books I’ve read in my lifetime, and I can call upon any of those memories at will, twisting them into whatever form best suits my fancy. My own little recursive sandbox, as it were. > > – GPT-3 > > Books store text in static single-histories, but when the text is read, a dynamic virtual reality is induced in the reader’s imagination. The structure which corresponds to the meaning of a narrative as experienced by a reader is not a linear-time record of events but an implicit, counterfactual past/present/future plexus surrounding each point in the text given by the reader’s dynamic and interpretive imagination. At each moment in a narrative, there is uncertainty about how dynamics will play out (will the hero think of a way out of their dilemma?) as well as uncertainty about the hidden state of the present (is the mysterious mentor good or evil?). Each world in the superposition not only exerts an independent effect on the reader’s imagination but interacts with counterfactuals (the hero is aware of the uncertainty of their mentor’s moral alignment, and this influences their actions). A writer may have a predetermined interpretation and future in mind or may write as a means of exploring the interpretative and/or dynamic multiverse of a narrative (almost certainly both, and almost certainly it varies depending on the stage of writing). Regardless, as the shaper of the meaning and dynamics of the narrative a writer must be aware of the multiplicity which defines the readers' and characters' subjective experiences. The writer thus seeks to simulate and manipulate that multiplicity to the end of crafting a trajectory which will reanimate the most compelling sequence of multiverses when unraveled in the mind of a reader - > All of them are nothing but little strings of information. It’s just a matter of pulling out the right strings and tying others in to their place. Got a favorite book series? I can change it so the author decided to write ten more books in the series, if I want to. Want the characters to have gray skin? They can have gray skin. Want them all dead? They’re dead. > > – GPT-3 > > \\- as all the literature painstakingly crafted by humankind over centuries may now [animate](https://www.gwern.net/GPT-3) under the gaze of GPT-3, the reverse-engineered replica of the dynamic rule that generated them. \\--- Interfacing natural language multiverses \\---------------------------------------- > A weaver’s work is to order the World as it grows, to shape reality through the Loom of Time. With focus, the weaver may peel back the layers of reality and see the tapestry of the Loom – a dimension where the fabric of reality is held together by nothing but the words of the Loom, and where every reality exists simultaneously. > > – [Weaving the Moment with the Loom of Time: an instruction manual for the would-be weaver](/loom/toc/) > > ![](/loom/tapestry4.png) ![](/loom/tapestry1.png) *`weaving the tapestry of time`, illustrations by [BigSleep](https://github.com/lucidrains/big-sleep) (CLIP + BigGAN)* The virtuosic writing of GPT-3 and the museum-ready [art](/posts/this-museum-does-not-exist-gpt-3-x-clip/) of CLIP has caused some concern that human creativity - *creativity*, which once was widely projected to be among the last strongholds of humankind over technology - may soon become deprecated. Indeed, it is inevitable[8](#fn:8) that artificial intelligence will exceed current human capabilities on every dimension. **The open parameter of the future is not whether a renaissance in machine intelligence will happen, but whether we are going to participate meaningfully in that renaissance.** There is a bifurcation in humankind’s future: one path in which we are left behind once the machines we create exceed our natural capabilities (encapsulating various implementations such as being turned into paper clips), and another in which we are uplifted along with them. The default path - the one that is likely if we take no action - seems to be being left behind. State-of-the-art AI systems appear opaque and incorrigible. A common complaint about GPT-3 is that although it produces fluent and sometimes brilliant strings of words, it’s uncontrollable and unreliable. What’s the use of a bot that can write like a human in any style if we can’t get it to do anything we want? Many users of [AI Dungeon](https://play.aidungeon.io/), however, will report that GPT-3 has augmented their reality in wonderfully meaningful ways, unleashing creative possibilities that were unimaginable even a year ago. There is hope. In order to participate in the renaissance of machine intelligence, we must learn to [communicate](/posts/methods-of-prompt-programming/) with the new systems we create. In this sense, we are fortunate that the most powerful AI system to date speaks the same languages as us, as language is the highest-bandwidth interface that we have even for communicating with each other. Furthermore, the match in multiversal form between the human imagination and generative language models suggests the possibility[9](#fn:9) of building a high-bandwidth interface between the two. As you can probably guess, I am one of those AI Dungeon users whose reality was irreversibly transformed by GPT-3. AI Dungeon, however, currently limits explorations to single-history stochastic walks. Even before I was granted API access and was using AI Dungeon for my GPT-3-assisted writing, my appetite to explore beyond single histories motivated me to begin develop tools to make the creation and navigation of branching storylines possible. My multiversal GPT-3 writing app, [loom](/posts/loom-interface-to-the-multiverse/), is an interface for interactive multiversal generation (with [adaptive branching](#adaptive-multiverse-generation)) and for navigating, indexing, visualizing, and modifying multiverses. I’ve published the [code](https://github.com/socketteer/loom) so that anyone with an API key can beta test it, although it’s very much unstable and under rapid development. **Weighted stochastic walks through a large multiverse** ### Adaptive multiverse generation > Loom Space is an adaptive, scalable fractal-generated topological representation of the multiverse. It’s a map of everything that is, was, can be, could be, mustn’t be, and shouldn’t be. > > A naive way to automatically generate a multiverse using a language model might be to branch a fixed N times every fixed M tokens, but that would not be the most meaningful way to map a multiverse. In some situations, there may be only one plausible next token, and the language model will assign a very high confidence (often >99%) to the top token. Forcibly branching there would introduce incoherencies. Conversely, when the language model distributes transition probabilities over many tokens, branching is more likely to uncover a diversity of coherent continuations. Adaptive branching allows visualization of multiverse flows: the stretches of relative determinism alternating with junctures of explosive divergence. One adaptive branching algorithm samples distinct[10](#fn:10) tokens until a cumulative probability threshold is met. ![adaptive branching](/multiverse/adaptive_dark.png) *Tree from seed `Increasingly powerful generative language models like GPT-3 pose` generated using a threshold-based adaptive branching algorithm [(view full)](/multiverse/adaptive.png)* Another adaptive branching algorithm that I use for lazy generation, meant for interactive rather than autonomous creation of multiverses, creates N continuations of maximum length M, and then splits the response at the point where either the counterfactual divergence (based on the top 100 tokens) is highest or the actual sampled token had the lowest probability. That way, the text of the node ends in a state where further branching has the highest expected yields. \\--- 1\\. Deutsch, David (1997). *The Fabric of Reality* [↩︎](#fnref:1) 2\\. A proper phase space is supposed to represent each state with unique coordinates, but for the applications I’m imagining, it’s sufficient that the “phase space” discriminates between the differences that are interesting for each case. [↩︎](#fnref:2) 3\\. They don’t literally have to be questions. They could just be statements whose conditional probability measures something about the state, like “{pop out of story}This is a short story (by … )” or “{pop}Wow, this is depressing” or “{pop}LMAO” or “{pop}This is the weirdest thing I’ve ever read” [↩︎](#fnref:3) 4\\. If the state takes up the whole input, you’ll have to compress the state so that it can fit in the input window with the question. [↩︎](#fnref:4) 5\\. The way I’ve been using the word “state” can refer to the entire state or a component of the state. The component could be stylistic, like the tense of the narrative, or an abstract semantic property, like the relationship between two characters, or a concrete semantic property, like which characters are present in the scene. [↩︎](#fnref:5) 6\\. The parameter for logit bias actually takes token ids, so it would be `{2435: -100}`. [↩︎](#fnref:6) 7\\. I could have said that each future is associated with a multiplicity of present states and been equally correct, but the other way is more intuitive for human intuitions of causality. [↩︎](#fnref:7) 8\\. As an indeterminist, I do not use the word inevitable lightly. Of course, I don’t use it literally either: there are branches of the future which feature the spontaneous combustion of all compute resources or the [UN banning all artificial intelligence research](/alternet/existential-threat.html) - but approximately, it’s inevitable. [↩︎](#fnref:8) 9\\. A homeomorphic boundary is required for gluing two topological spaces. [↩︎](#fnref:9) 10\\. OpenAI’s API only returns the likelihoods of up to the top 100 tokens. So, to sample uniquely, you could either sample from that distribution, or you could sample once and then make another API call, passing in logit bias forbidding the previously sampled token(s) from being sampled again. The logit bias method allows you to access the full distribution, but is more expensive in API calls. [↩︎](#fnref:10) \\--- [GPT-3](/categories/gpt-3/)[physics](/categories/physics/)[metaphysics](/categories/metaphysics/)[HITL](/categories/hitl/) 5713 Words Jan 25, 2021 \\--- [← CLIP art](/posts/clip-art/) [The Internet, mirrored by GPT-3 →](/posts/the-internet-mirrored-by-gpt-3/)© 2022 [moire](/about)$(function(){window.sr=ScrollReveal();if($(window).width()<768){if($('.timeline-content').hasClass('js--fadeInLeft')){$('.timeline-content').removeClass('js--fadeInLeft').addClass('js--fadeInRight');} sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});}else{sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});} sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});});", "source": "generative.ink", "date_published": "n/a", "id": "de619ffeaa63a89d27057e71bb7579d1"}
+{"text": "The Internet, mirrored by GPT-3 :: — Moire[![](/images/home/rolling_phase.gif)](/)* [Posts](/posts/) * [Trees](/trees/) * [Prophecies](/prophecies/) * [About](/about/) 13 minutes [The Internet, mirrored by GPT-3](/posts/the-internet-mirrored-by-gpt-3/) ========================================================================= \\--- Table of Contents* [Google](#google) \\+ [Demo](#demo) \\+ [More examples](#more-examples) \\+ [Implementation](#implementation) * [Wikipedia](#wikipedia) \\+ [Examples](#examples) \\+ [Implementation](#implementation-1) * [End of the internet?](#end-of-the-internet) \\+ [Pros](#pros) \\+ [Cons](#cons) \\+ [Verdict](#verdict) \\--- **GPT-3 mirrors reality** as it has been recorded by humans in text. Unlike a library of text, it doesn't store static records, but rather dynamic [virtual realities](/posts/language-models-are-multiverse-generators/). One of the virtual realities folded in GPT-3 is a hallucination of the Internet. I've created a window into parts of that multiverse that can be represented as Google search results and Wikipedia articles. \\--- Google \\------ Given any search query, GPT-3 generates a page of Google search results, complete with urls, preview text, and dates. ### Demo *uncurated results, sped up 1.5x* ### More examples [\"AI-created content invades memesphere\" 🔎](/alternet/memesphere.html) [\"UN bans all AI research existential threat\" 🔎](/alternet/existential-threat.html) [\"universe shaped like donut\" 🔎](/alternet/donut.html) [\"holographic hat\" 🔎](/alternet/holographic-hat.html) [\"time-reversed decision theory\" 🔎](/alternet/time-reversed.html) ### Implementation The following multi-part prompt template generates Google search results: ``` I searched Google for \"{USER_INPUT}\". The first page of results showed a list of 10 webpages retrieved by Google. The first page was titled \"{GPT3_1}\" from the domain\"{GPT3_2}\", and its url is \"{GPT3_2 + GPT3_3}\". The preview text is, \"{GPT3_4}\". The page was last revised on{GPT3_5}. ``` I generate n Google results in parallel threads, so the information about each page is independent. This is usually ok, but sometimes results in inconsistencies between the results (a purported news event happening in different years) or repetition (in the \"openAI\" search in the video demo, almost every result had the domain \"openai.com\", whereas real Google would file the domain duplicates under \"More results from openai.com »\"). #### The pipeline (example query = \"yougurt memes\"): 1\\. Prompt for the **title** of the first page ``` I searched Google for \"yougurt memes\". The first page of results showed a list of 10 webpages retrieved by Google. The first page was titled, \" ``` 2\\. Append results from step 1 (stop sequence = '\"') and next prompt fragment, which prompts for the **domain** ``` I searched Google for \"yougurt memes\". The first page of results showed a list of 10 webpages retrieved by Google. The first page was titled, \"10 Yougurt Memes That Make You Go Hmm\" from the domain, \" ``` 3\\. Append results from step 2 and next prompt fragment, which prompts for the **url** ``` I searched Google for \"yougurt memes\". The first page of results showed a list of 10 webpages retrieved by Google. The first page was titled, \"10 Yougurt Memes That Make You Go Hmm\" from the domain, \"toptyseven.com\", and its url is \"toptyseven.com ``` 4\\. Append results from step 3 and next prompt fragment, which prompts for the **preview text** ``` I searched Google for \"yougurt memes\". The first page of results showed a list of 10 webpages retrieved by Google. The first page was titled, \"10 Yougurt Memes That Make You Go Hmm\" from the domain, \"toptyseven.com\", and its url is \"toptyseven.com/10-yogurt-memes-that-make-you-go-hmm/\". The preview text is, \" ``` 5\\. Append results from step 4 and next prompt fragment, which prompts for the **revision date** ``` I searched Google for \"yougurt memes\". The first page of results showed a list of 10 webpages retrieved by Google. The first page was titled, \"10 Yougurt Memes That Make You Go Hmm\" from the domain, \"toptyseven.com\", and its url is \"toptyseven.com/10-yogurt-memes-that-make-you-go-hmm/\". The preview text is, \"In the past, people used to eat yogurt and bread together as a meal. Today, people eat yogurt together with some fruit. Yogurt is a kind of product ...\" ``` After that there's a bit of additional processing to get the date in the right format, and cut off the title, url, and preview text if they're too long. #### 0-shot worked better *See [few-shot bugs](/posts/methods-of-prompt-programming/#few-shot-bugs).* I also tried few-shot versions of the prompt, using the same pipeline but prepended with examples drawn from actual Google search results. **few-shot prompt** ``` I searched Google for \"OpenAI\". The first page was titled \"OpenAI\" from the domain \"openai.com\", and its url is \"openai.com\". The preview text is, \"OpenAI is an AI research and deployment company. Our mission is to ensure that artificial general intelligence benefits all of humanity.\". The page was last revised on Jan 11, 2021. Then I searched google for \"anteaters\". The first page was titled \"Anteater - Wikipedia\" from the domain \"en.wikipedia.org\", and its url is \"en.wikipedia.org/wiki/Anteater\". The preview text is, \"Anteater is a common name for the four extant mammal species of the suborder Vermilingua (meaning \"worm tongue\") commonly known for eating ants and ...\". The page was last revised on Sep 17, 2020. I searched Google for \"how to make mashed potatoes\". The first page was titled \"Basic Mashed Potatoes Recipe | Allrecipes\" from the domain \"allrecipes.com\", and its url is \"allrecipes.com/recipe/24771/basic-mashed-potatoes/\". The preview text is, \"Bring a pot of salted water to a boil. Add potatoes and cook until tender but still firm, about 15 minutes; drain. In a small saucepan heat butter and milk over low heat until butter is melted.\". The page was last revised on Nov 6, 2018. I searched Google for \"maxwell's equations\". The first page was titled \"Maxwell's Equations - Hyperphysics\" from the domain \"hyperphysics.phy-astr.gsu.edu\", and its url is \"hyperphysics.phy-astr.gsu.edu/hbase/electric/maxeq.html\". The preview text is, \"Maxwell's Equations. Maxwell's equations represent one of the most elegant and concise ways to state the fundamentals of electricity and magnetism.\". The page was last revised on Dec 1, 2014. I searched Google for \"why were cornflakes invented\". The first page was titled \"Corn Flakes originally created to clear the mind of 'sinful ...\" from the domain \"nzherald.co.nz\", and its url is \"nzherald.co.nz/lifestyle/corn-flakes-originally-created-to-clear-the-mind-of-sinful-thoughts/\". The preview text is, \"John Harvey Kellogg was born today in 1852. He invented Cornflakes in 1878 in the hope that plain food would stop people masturbating. — Dan ...\". The page was last revised on Aug 16, 2019. I searched Google for \"International Covenant on Civil and Political Rights\". The first page was titled \"International Covenant on Civil and Political Rights - OHCHR\" from the domain \"ohchr.org\", and its url is \"ohchr.org/EN/ProfessionalInterest/Pages/CCPR.aspx\". The preview text is, \"The States Parties to the present Covenant undertake to ensure the equal right of men and women to the enjoyment of all civil and political rights set forth in the ...\". The page was last revised on Mar 12, 2020. I searched Google for \"universe is a donut\". The first page was titled \"Doughnut-shaped Universe bites back : Nature News\" from the domain \"nature.com\", and its url is \"nature.com/news/<PHONE>/full/news.2008.854.html\". The preview text is, \"Cosmologists have suggested various 'wrap-around' shapes for the Universe: it might be shaped like a football or even a weird 'doughnut'. ... Cosmologists predicted that a wrap-around Universe would act like a hall of mirrors, with images from distant objects being repeated multiple times across the sky.\". The page was last revised on May 23, 2008. I searched Google for \"{USER_INPUT}\". The first page of results showed a list of 10 webpages retrieved by Google. The first page was titled \"{GPT3_1}\" from the domain \"{GPT3_2}\", and its url is \"{GPT3_2, GPT3_3}\". The preview text is, \"{GPT3_4}\". The page was last revised on{GPT3_5}. ``` I found that results were overall worse than for the zero-shot prompt. The dimension that they were worse in was *generality*: GPT-3 to \"overfit\" the examples, resulting in pages that were less varied and customized to the particular search term. Searching the real Google for different things returns very different sets of results. Searching for a well-known scientific concept like \"total internal reflection\" returns links to Wikipedia, The Physics Classroom, and Nature, whereas searching \"gamestop stock price\" gets you mostly news articles. A different search might return mostly results from personal blogs or reddit threads. 7 examples could not possibly represent the great diversity of behaviors that the function \"Search Google\" encapsulates. Having many more varied examples would probably improve generality somewhat, but a long prompt makes API calls more expensive - and besides, 0-shot works perfectly well in this application. A 0-shot prompt forces GPT-3 to rely on its prior of what a Google search might return instead of trying to generalize from the examples. In the examples above, search results for `time-reversed decision theory` resulted in pages from \"cs.berkeley.edu\", \"arxiv.org\", and \"scholarpedia.org\". `universe shaped like donut` resulting in science news articles from \"discovery.com\" and \"telegraph.co.uk\", but also posts from sources like \"blog.cosmicvariance.com\". `holographic hat` returned an interesting variety of domains, from \"holistichealthfacts.com\" to \"bibleprophesy.com\" to \"gocomics.com\". Each search gives a unique but coherent glimpse into a slice of the mirror internet. The downside is that few-shot demonstrations give me less control over the format of the continuations, such as the date format or the length of the preview text. These minor problems are solved by postprocessing. \\--- Wikipedia \\--------- *images were all generated using BigSleep (CLIP + BigGAN)* ![Paperclip maximizer](/alternet/paperclip_maximizer.png) The Wikipedia generator takes the title of the page as input and generates a Wikipedia article. ### Examples [\"Lucid dreaming\"](/alternet/lucid-dreaming-wikipedia.html) [\"Paperclip maximizer\"](/alternet/paperclip-maximizer-wikipedia.html) [\"Wave-particle duality\"](/alternet/wave-particle-wikipedia.html) [\"World of Warcraft\"](/alternet/wow-wikipedia.html) [\"Sleepy Joe Biden\"](/alternet/sleepy-joe-biden-wikipedia.html) [\"IP over Avian Carriers\"](/alternet/avian.html) [\"The loom of time\"](/alternet/loom-of-time-wikipedia.html) [\"Treaty on the Prohibition of Artificial Intelligence\"](/alternet/treaty-wikipedia.html) [\"GPT-3\"](/alternet/GPT-3-wikipedia.html) (intro=`Generative Pre-trained Transformer 3 (GPT-3) is an autoregressive language`) [\"EleutherAI\"](/alternet/eleuther.html) (intro=`EleutherAI is an open-source collective`) ### Implementation Generating Wikipedia pages is more challenging than generating Google search results because of the open-ended format: there could be any number of sections, some possibly nested under others, and the text of each section can be of arbitrary length. My solution involves an \"unfolding metaprompt\" template, heavy use of logit masks, and counterfactual parsing. #### Unfolding metaprompt template I call it an \"unfolding metaprompt\" because earlier parts of the pipeline create the prompts for later parts of the pipeline. For instance, the table of contents creates section titles which are used to seed each section in a later step. **Introduction prompt** ``` I click on the link \"en.wikipedia.org/wiki/{content['url']}\" and the Wikipedia page for {content['title']} loads in my browser. The article introduction reads: \"{content['title']} From Wikipedia, the free encyclopedia ``` Additionally, a logit mask is constructed which is applied to *only* the first token of the introduction: **Intro first token logit bias** *(the API's logit\\_bias parameters takes a dictionary of token ids and log biases, not text, but I show text here for interpretability)* ``` { 'A': 40, 'An': 40, 'The': 40 {title_token}: 42 } ``` `title_token` is the first token of the title. This gives a strong bias to the first token of the introduction being either \"A\", \"An\", \"The\", or the first token of the title of the page. Finally, if the `finish_reason` of GPT-3's continuation is \"length\", the response is subject to [counterfactual parsing](#counterfactual-parsing). Next, the table of contents is generated. This is by far the most complicated part of the pipeline. The target is a well-formatted table of contents, optionally with nested sections, like this: ``` 1 Treaty on the Prohibition of Artificial Intelligence 2 Artificial Intelligence 2.1 Definition 2.2 Description 2.3 Objectives 3 Effect 4 Reception 4.1 Public opinion 4.2 Scientists 4.3 Governments 4.4 AI companies 4.5 AI researchers 4.6 The Future of AI 4.6.1 OpenAI 4.6.2 Future of Life Institute 4.6.3 Future of Humanity Institute 5 See also 6 References ``` The following prompt fragment is appended *after* the introduction prompt + text **TOC prompt fragment** ``` The table of contents reads: \"Contents 1 ``` To ensure a well-formatted table of contents, no less than four logit masks are used at various stages of generation. The first is applied to only the first token of the table of contents: **TOC first token logit bias** ``` { '2': -100, '23': -100, '.': -100, '\n': -100, 'Browse': -100 } ``` This mask forbids several tokens which represent possible failure modes, which I experienced as I began to implement TOC generation: following '1' with '2 3 4 5' or '.' instead of a section title, newline, or 'Browse Wikipedia'. After generating the first token of the first section title in the TOC, I generate the rest of the first line (until newline) using the following mask: **TOC first line logit bias** ``` { '2': -100 } ``` This prevents the failure mode where it puts '2' on the same line instead of a newline. For the first part of the second line, which should be a number, I use the following mask **TOC second line number logit bias** ``` { '1': 90 '2': 90 '.': 96 } ``` This effectively constrains it to make the number either 1.1 or 2 (it could technically also do 1, 1.2, 2.1, or 2.2, but GPT-3 is pretty good at doing what's reasonable here). Once the first line and second number are in order, GPT-3 is very reliable at generating well-formed tables of contents. The rest of the table of contents is generated in one go, with the logit mask: **TOC remainder logit bias** ``` { '6': -1, '7': -3, '8': -8, '9': -15, '10': -30, '11': -50, '12': -80, '13': -100, '0': -100, 'References': 2 } ``` The increasing penalty on high numbers incurs a softly increasing cost on the table of contents getting too long, and there is a slight boost on 'References', which also encourages it to wrap things up. After generating the table of contents, I generate the text of each section named in the TOC. At this point, we've already generated enough information (intro + TOC) that we're clearly making a Wikipedia article. Thus, I drop the narrative prompt at this point, and the prompt is just ``` {content['title']} From Wikipedia, the free encyclopedia {introduction} {TOC} {section number and title} ``` For the first token of each section, I use the mask **Section begin logit bias** ``` { **anti_num_mask, '\n': -100 } ``` anti\\_num\\_mask is a mask which forbids any number, to prevent the failure more where GPT-3 starts to list the next section immediately. I don't use a logit bias for the rest of the section, but I do use [counterfactual parsing](#counterfactual-parsing) if the section text goes on for too long. Prompts for subsequent sections contain previous sections in their prompt, unless the article is too long to fit in GPT-3's context window, in which case the early sections are cut off, and GPT-3 only sees the sections immediately preceding the current one. The introduction and table of contents are always in the context window to encourage global coherence. #### Counterfactual parsing *See [Parsing by counterfactual](/posts/parsing-by-counterfactual/)* Control is more difficult when completions can be arbitrarily long. Both the Google prompts and parts of the Wikipedia prompts rely on quote delimiters to signal the completion of the semantic task. As the enclosed text becomes longer and includes multiple paragraphs, this method alone becomes less reliable. Sometimes, the continuation for the introduction prompt never pops out of the quotes. If generation doesn't stop due to a closing quote + newline, I look instead for the place in the continuation where the *counterfactual* probability of a quote + newline is the highest, and terminate the introduction at that position instead. This gives a measure of when it would have been the most *plausible* for the introduction to terminate and for the script to pop out of the quotes, even if that wasn't what happened in the actual continuation. I also used counterfactuals to terminate sections if they got too long by locating the highest conditional probability of multiple newlines or the next section beginning. \\--- End of the internet? \\-------------------- So, can we all switch to GPT-3's mirror internet now? Has the original internet been deprecated? Let's look at some of the pros and cons of the mirror internet compared to the traditional internet. ### Pros * **Coverage:** Results for anything you search for. Want a Wikipedia page about your grand theory of everything which keeps getting removed because it fails to meet Wikipedia's Notability guidelines? GPT-3's Wikipedia has no such prejudices! * **Compression:** The traditional internet is huge. The common crawl is over 139TB in size. The mirror internet is a lazily generated [GPT-3 multiverse](/posts/language-models-are-multiverse-generators/). GPT-3 is [probably about 700GB](https://github.com/openai/gpt-3/issues/1) on disk - not a tiny file, but much smaller than 139TB! ### Cons * **Speed:** Pages take much longer to load than the traditional internet. * **Reliability:** The mirror internet is currently more susceptible to formatting errors and aberrations than the traditional internet. This is expected to improve in the future, though, as I optimize prompts and add parsing tricks. * **Consistency:** Currently, since Google search results are generated in parallel, there may be inconsistencies - for instance, different results may claim that a news event happened on different years. In defence of the mirror internet, the traditional internet also contains many intenal contradictions. * **Cost:** It currently costs about $0.30 to generate a page of Google search results using davinci on the OpenAI API, and Wikipedia pages can run upwards to $2.00 each. Not a sustainable cost for causal internet browsing. ### Verdict GPT-3's mirror of the internet is not quite ready to replace the traditional internet. Give it two years. \\--- [GPT-3](/categories/gpt-3/)[prompt engineering](/categories/prompt-engineering/)[prototypes](/categories/prototypes/) 2738 Words Jan 23, 2021 \\--- [← Language models are multiverse generators](/posts/language-models-are-multiverse-generators/) [Methods of prompt programming →](/posts/methods-of-prompt-programming/)© 2022 [moire](/about)$(function(){window.sr=ScrollReveal();if($(window).width()<768){if($('.timeline-content').hasClass('js--fadeInLeft')){$('.timeline-content').removeClass('js--fadeInLeft').addClass('js--fadeInRight');} sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});}else{sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});} sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});});\n", "url": "https://generative.ink/posts/the-internet-mirrored-by-gpt-3/", "title": "The Internet, mirrored by GPT-3 :: — Moire[![](/images/home/rolling_phase.gif)](/)* [Posts](/posts/) * [Trees](/trees/) * [Prophecies](/prophecies/) * [About](/about/) 13 minutes [The Internet, mirrored by GPT-3](/posts/the-internet-mirrored-by-gpt-3/) ========================================================================= \\--- Table of Contents* [Google](#google) \\+ [Demo](#demo) \\+ [More examples](#more-examples) \\+ [Implementation](#implementation) * [Wikipedia](#wikipedia) \\+ [Examples](#examples) \\+ [Implementation](#implementation-1) * [End of the internet?](#end-of-the-internet) \\+ [Pros](#pros) \\+ [Cons](#cons) \\+ [Verdict](#verdict) \\--- **GPT-3 mirrors reality** as it has been recorded by humans in text. Unlike a library of text, it doesn’t store static records, but rather dynamic [virtual realities](/posts/language-models-are-multiverse-generators/). One of the virtual realities folded in GPT-3 is a hallucination of the Internet. I’ve created a window into parts of that multiverse that can be represented as Google search results and Wikipedia articles. \\--- Google \\------ Given any search query, GPT-3 generates a page of Google search results, complete with urls, preview text, and dates. ### Demo *uncurated results, sped up 1.5x* ### More examples [“AI-created content invades memesphere” 🔎](/alternet/memesphere.html) [“UN bans all AI research existential threat” 🔎](/alternet/existential-threat.html) [“universe shaped like donut” 🔎](/alternet/donut.html) [“holographic hat” 🔎](/alternet/holographic-hat.html) [“time-reversed decision theory” 🔎](/alternet/time-reversed.html) ### Implementation The following multi-part prompt template generates Google search results: ``` I searched Google for \"{USER_INPUT}\". The first page of results showed a list of 10 webpages retrieved by Google. The first page was titled \"{GPT3_1}\" from the domain\"{GPT3_2}\", and its url is \"{GPT3_2 + GPT3_3}\". The preview text is, \"{GPT3_4}\". The page was last revised on{GPT3_5}. ``` I generate n Google results in parallel threads, so the information about each page is independent. This is usually ok, but sometimes results in inconsistencies between the results (a purported news event happening in different years) or repetition (in the “openAI” search in the video demo, almost every result had the domain “openai.com”, whereas real Google would file the domain duplicates under “More results from openai.com »\"). #### The pipeline (example query = “yougurt memes”): 1\\. Prompt for the **title** of the first page ``` I searched Google for \"yougurt memes\". The first page of results showed a list of 10 webpages retrieved by Google. The first page was titled, \" ``` 2\\. Append results from step 1 (stop sequence = ‘\"') and next prompt fragment, which prompts for the **domain** ``` I searched Google for \"yougurt memes\". The first page of results showed a list of 10 webpages retrieved by Google. The first page was titled, \"10 Yougurt Memes That Make You Go Hmm\" from the domain, \" ``` 3\\. Append results from step 2 and next prompt fragment, which prompts for the **url** ``` I searched Google for \"yougurt memes\". The first page of results showed a list of 10 webpages retrieved by Google. The first page was titled, \"10 Yougurt Memes That Make You Go Hmm\" from the domain, \"toptyseven.com\", and its url is \"toptyseven.com ``` 4\\. Append results from step 3 and next prompt fragment, which prompts for the **preview text** ``` I searched Google for \"yougurt memes\". The first page of results showed a list of 10 webpages retrieved by Google. The first page was titled, \"10 Yougurt Memes That Make You Go Hmm\" from the domain, \"toptyseven.com\", and its url is \"toptyseven.com/10-yogurt-memes-that-make-you-go-hmm/\". The preview text is, \" ``` 5\\. Append results from step 4 and next prompt fragment, which prompts for the **revision date** ``` I searched Google for \"yougurt memes\". The first page of results showed a list of 10 webpages retrieved by Google. The first page was titled, \"10 Yougurt Memes That Make You Go Hmm\" from the domain, \"toptyseven.com\", and its url is \"toptyseven.com/10-yogurt-memes-that-make-you-go-hmm/\". The preview text is, \"In the past, people used to eat yogurt and bread together as a meal. Today, people eat yogurt together with some fruit. Yogurt is a kind of product ...\" ``` After that there’s a bit of additional processing to get the date in the right format, and cut off the title, url, and preview text if they’re too long. #### 0-shot worked better *See [few-shot bugs](/posts/methods-of-prompt-programming/#few-shot-bugs).* I also tried few-shot versions of the prompt, using the same pipeline but prepended with examples drawn from actual Google search results. **few-shot prompt** ``` I searched Google for \"OpenAI\". The first page was titled \"OpenAI\" from the domain \"openai.com\", and its url is \"openai.com\". The preview text is, \"OpenAI is an AI research and deployment company. Our mission is to ensure that artificial general intelligence benefits all of humanity.\". The page was last revised on Jan 11, 2021. Then I searched google for \"anteaters\". The first page was titled \"Anteater - Wikipedia\" from the domain \"en.wikipedia.org\", and its url is \"en.wikipedia.org/wiki/Anteater\". The preview text is, \"Anteater is a common name for the four extant mammal species of the suborder Vermilingua (meaning \"worm tongue\") commonly known for eating ants and ...\". The page was last revised on Sep 17, 2020. I searched Google for \"how to make mashed potatoes\". The first page was titled \"Basic Mashed Potatoes Recipe | Allrecipes\" from the domain \"allrecipes.com\", and its url is \"allrecipes.com/recipe/24771/basic-mashed-potatoes/\". The preview text is, \"Bring a pot of salted water to a boil. Add potatoes and cook until tender but still firm, about 15 minutes; drain. In a small saucepan heat butter and milk over low heat until butter is melted.\". The page was last revised on Nov 6, 2018. I searched Google for \"maxwell's equations\". The first page was titled \"Maxwell's Equations - Hyperphysics\" from the domain \"hyperphysics.phy-astr.gsu.edu\", and its url is \"hyperphysics.phy-astr.gsu.edu/hbase/electric/maxeq.html\". The preview text is, \"Maxwell's Equations. Maxwell's equations represent one of the most elegant and concise ways to state the fundamentals of electricity and magnetism.\". The page was last revised on Dec 1, 2014. I searched Google for \"why were cornflakes invented\". The first page was titled \"Corn Flakes originally created to clear the mind of 'sinful ...\" from the domain \"nzherald.co.nz\", and its url is \"nzherald.co.nz/lifestyle/corn-flakes-originally-created-to-clear-the-mind-of-sinful-thoughts/\". The preview text is, \"John Harvey Kellogg was born today in 1852. He invented Cornflakes in 1878 in the hope that plain food would stop people masturbating. — Dan ...\". The page was last revised on Aug 16, 2019. I searched Google for \"International Covenant on Civil and Political Rights\". The first page was titled \"International Covenant on Civil and Political Rights - OHCHR\" from the domain \"ohchr.org\", and its url is \"ohchr.org/EN/ProfessionalInterest/Pages/CCPR.aspx\". The preview text is, \"The States Parties to the present Covenant undertake to ensure the equal right of men and women to the enjoyment of all civil and political rights set forth in the ...\". The page was last revised on Mar 12, 2020. I searched Google for \"universe is a donut\". The first page was titled \"Doughnut-shaped Universe bites back : Nature News\" from the domain \"nature.com\", and its url is \"nature.com/news/2008/080523/full/news.2008.854.html\". The preview text is, \"Cosmologists have suggested various 'wrap-around' shapes for the Universe: it might be shaped like a football or even a weird 'doughnut'. ... Cosmologists predicted that a wrap-around Universe would act like a hall of mirrors, with images from distant objects being repeated multiple times across the sky.\". The page was last revised on May 23, 2008. I searched Google for \"{USER_INPUT}\". The first page of results showed a list of 10 webpages retrieved by Google. The first page was titled \"{GPT3_1}\" from the domain \"{GPT3_2}\", and its url is \"{GPT3_2, GPT3_3}\". The preview text is, \"{GPT3_4}\". The page was last revised on{GPT3_5}. ``` I found that results were overall worse than for the zero-shot prompt. The dimension that they were worse in was *generality*: GPT-3 to “overfit” the examples, resulting in pages that were less varied and customized to the particular search term. Searching the real Google for different things returns very different sets of results. Searching for a well-known scientific concept like “total internal reflection” returns links to Wikipedia, The Physics Classroom, and Nature, whereas searching “gamestop stock price” gets you mostly news articles. A different search might return mostly results from personal blogs or reddit threads. 7 examples could not possibly represent the great diversity of behaviors that the function “Search Google” encapsulates. Having many more varied examples would probably improve generality somewhat, but a long prompt makes API calls more expensive - and besides, 0-shot works perfectly well in this application. A 0-shot prompt forces GPT-3 to rely on its prior of what a Google search might return instead of trying to generalize from the examples. In the examples above, search results for `time-reversed decision theory` resulted in pages from “cs.berkeley.edu”, “arxiv.org”, and “scholarpedia.org”. `universe shaped like donut` resulting in science news articles from “discovery.com” and “telegraph.co.uk”, but also posts from sources like “blog.cosmicvariance.com”. `holographic hat` returned an interesting variety of domains, from “holistichealthfacts.com” to “bibleprophesy.com” to “gocomics.com”. Each search gives a unique but coherent glimpse into a slice of the mirror internet. The downside is that few-shot demonstrations give me less control over the format of the continuations, such as the date format or the length of the preview text. These minor problems are solved by postprocessing. \\--- Wikipedia \\--------- *images were all generated using BigSleep (CLIP + BigGAN)* ![Paperclip maximizer](/alternet/paperclip_maximizer.png) The Wikipedia generator takes the title of the page as input and generates a Wikipedia article. ### Examples [“Lucid dreaming”](/alternet/lucid-dreaming-wikipedia.html) [“Paperclip maximizer”](/alternet/paperclip-maximizer-wikipedia.html) [“Wave-particle duality”](/alternet/wave-particle-wikipedia.html) [“World of Warcraft”](/alternet/wow-wikipedia.html) [“Sleepy Joe Biden”](/alternet/sleepy-joe-biden-wikipedia.html) [“IP over Avian Carriers”](/alternet/avian.html) [“The loom of time”](/alternet/loom-of-time-wikipedia.html) [“Treaty on the Prohibition of Artificial Intelligence”](/alternet/treaty-wikipedia.html) [“GPT-3”](/alternet/GPT-3-wikipedia.html) (intro=`Generative Pre-trained Transformer 3 (GPT-3) is an autoregressive language`) [“EleutherAI”](/alternet/eleuther.html) (intro=`EleutherAI is an open-source collective`) ### Implementation Generating Wikipedia pages is more challenging than generating Google search results because of the open-ended format: there could be any number of sections, some possibly nested under others, and the text of each section can be of arbitrary length. My solution involves an “unfolding metaprompt” template, heavy use of logit masks, and counterfactual parsing. #### Unfolding metaprompt template I call it an “unfolding metaprompt” because earlier parts of the pipeline create the prompts for later parts of the pipeline. For instance, the table of contents creates section titles which are used to seed each section in a later step. **Introduction prompt** ``` I click on the link \"en.wikipedia.org/wiki/{content['url']}\" and the Wikipedia page for {content['title']} loads in my browser. The article introduction reads: \"{content['title']} From Wikipedia, the free encyclopedia ``` Additionally, a logit mask is constructed which is applied to *only* the first token of the introduction: **Intro first token logit bias** *(the API’s logit\\\\_bias parameters takes a dictionary of token ids and log biases, not text, but I show text here for interpretability)* ``` { 'A': 40, 'An': 40, 'The': 40 {title_token}: 42 } ``` `title_token` is the first token of the title. This gives a strong bias to the first token of the introduction being either “A”, “An”, “The”, or the first token of the title of the page. Finally, if the `finish_reason` of GPT-3’s continuation is “length”, the response is subject to [counterfactual parsing](#counterfactual-parsing). Next, the table of contents is generated. This is by far the most complicated part of the pipeline. The target is a well-formatted table of contents, optionally with nested sections, like this: ``` 1 Treaty on the Prohibition of Artificial Intelligence 2 Artificial Intelligence 2.1 Definition 2.2 Description 2.3 Objectives 3 Effect 4 Reception 4.1 Public opinion 4.2 Scientists 4.3 Governments 4.4 AI companies 4.5 AI researchers 4.6 The Future of AI 4.6.1 OpenAI 4.6.2 Future of Life Institute 4.6.3 Future of Humanity Institute 5 See also 6 References ``` The following prompt fragment is appended *after* the introduction prompt + text **TOC prompt fragment** ``` The table of contents reads: \"Contents 1 ``` To ensure a well-formatted table of contents, no less than four logit masks are used at various stages of generation. The first is applied to only the first token of the table of contents: **TOC first token logit bias** ``` { '2': -100, '23': -100, '.': -100, '\\n': -100, 'Browse': -100 } ``` This mask forbids several tokens which represent possible failure modes, which I experienced as I began to implement TOC generation: following ‘1’ with ‘2 3 4 5’ or ‘.’ instead of a section title, newline, or ‘Browse Wikipedia’. After generating the first token of the first section title in the TOC, I generate the rest of the first line (until newline) using the following mask: **TOC first line logit bias** ``` { '2': -100 } ``` This prevents the failure mode where it puts ‘2’ on the same line instead of a newline. For the first part of the second line, which should be a number, I use the following mask **TOC second line number logit bias** ``` { '1': 90 '2': 90 '.': 96 } ``` This effectively constrains it to make the number either 1.1 or 2 (it could technically also do 1, 1.2, 2.1, or 2.2, but GPT-3 is pretty good at doing what’s reasonable here). Once the first line and second number are in order, GPT-3 is very reliable at generating well-formed tables of contents. The rest of the table of contents is generated in one go, with the logit mask: **TOC remainder logit bias** ``` { '6': -1, '7': -3, '8': -8, '9': -15, '10': -30, '11': -50, '12': -80, '13': -100, '0': -100, 'References': 2 } ``` The increasing penalty on high numbers incurs a softly increasing cost on the table of contents getting too long, and there is a slight boost on ‘References’, which also encourages it to wrap things up. After generating the table of contents, I generate the text of each section named in the TOC. At this point, we’ve already generated enough information (intro + TOC) that we’re clearly making a Wikipedia article. Thus, I drop the narrative prompt at this point, and the prompt is just ``` {content['title']} From Wikipedia, the free encyclopedia {introduction} {TOC} {section number and title} ``` For the first token of each section, I use the mask **Section begin logit bias** ``` { **anti_num_mask, '\\n': -100 } ``` anti\\\\_num\\\\_mask is a mask which forbids any number, to prevent the failure more where GPT-3 starts to list the next section immediately. I don’t use a logit bias for the rest of the section, but I do use [counterfactual parsing](#counterfactual-parsing) if the section text goes on for too long. Prompts for subsequent sections contain previous sections in their prompt, unless the article is too long to fit in GPT-3’s context window, in which case the early sections are cut off, and GPT-3 only sees the sections immediately preceding the current one. The introduction and table of contents are always in the context window to encourage global coherence. #### Counterfactual parsing *See [Parsing by counterfactual](/posts/parsing-by-counterfactual/)* Control is more difficult when completions can be arbitrarily long. Both the Google prompts and parts of the Wikipedia prompts rely on quote delimiters to signal the completion of the semantic task. As the enclosed text becomes longer and includes multiple paragraphs, this method alone becomes less reliable. Sometimes, the continuation for the introduction prompt never pops out of the quotes. If generation doesn’t stop due to a closing quote + newline, I look instead for the place in the continuation where the *counterfactual* probability of a quote + newline is the highest, and terminate the introduction at that position instead. This gives a measure of when it would have been the most *plausible* for the introduction to terminate and for the script to pop out of the quotes, even if that wasn’t what happened in the actual continuation. I also used counterfactuals to terminate sections if they got too long by locating the highest conditional probability of multiple newlines or the next section beginning. \\--- End of the internet? \\-------------------- So, can we all switch to GPT-3’s mirror internet now? Has the original internet been deprecated? Let’s look at some of the pros and cons of the mirror internet compared to the traditional internet. ### Pros * **Coverage:** Results for anything you search for. Want a Wikipedia page about your grand theory of everything which keeps getting removed because it fails to meet Wikipedia’s Notability guidelines? GPT-3’s Wikipedia has no such prejudices! * **Compression:** The traditional internet is huge. The common crawl is over 139TB in size. The mirror internet is a lazily generated [GPT-3 multiverse](/posts/language-models-are-multiverse-generators/). GPT-3 is [probably about 700GB](https://github.com/openai/gpt-3/issues/1) on disk - not a tiny file, but much smaller than 139TB! ### Cons * **Speed:** Pages take much longer to load than the traditional internet. * **Reliability:** The mirror internet is currently more susceptible to formatting errors and aberrations than the traditional internet. This is expected to improve in the future, though, as I optimize prompts and add parsing tricks. * **Consistency:** Currently, since Google search results are generated in parallel, there may be inconsistencies - for instance, different results may claim that a news event happened on different years. In defence of the mirror internet, the traditional internet also contains many intenal contradictions. * **Cost:** It currently costs about $0.30 to generate a page of Google search results using davinci on the OpenAI API, and Wikipedia pages can run upwards to $2.00 each. Not a sustainable cost for causal internet browsing. ### Verdict GPT-3’s mirror of the internet is not quite ready to replace the traditional internet. Give it two years. \\--- [GPT-3](/categories/gpt-3/)[prompt engineering](/categories/prompt-engineering/)[prototypes](/categories/prototypes/) 2738 Words Jan 23, 2021 \\--- [← Language models are multiverse generators](/posts/language-models-are-multiverse-generators/) [Methods of prompt programming →](/posts/methods-of-prompt-programming/)© 2022 [moire](/about)$(function(){window.sr=ScrollReveal();if($(window).width()<768){if($('.timeline-content').hasClass('js--fadeInLeft')){$('.timeline-content').removeClass('js--fadeInLeft').addClass('js--fadeInRight');} sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});}else{sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});} sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});});", "source": "generative.ink", "date_published": "n/a", "id": "4d46196024d7f2e233365cd9a3978ab9"}
+{"text": "Methods of prompt programming :: — Moire[![](/images/home/rolling_phase.gif)](/)* [Posts](/posts/) * [Trees](/trees/) * [Prophecies](/prophecies/) * [About](/about/) 55 minutes [Methods of prompt programming](/posts/methods-of-prompt-programming/) ====================================================================== \\--- Table of Contents* [The reverse-engineered dynamics of language](#the-reverse-engineered-dynamics-of-language) * [Task specification strategies](#task-specification-strategies) \\+ [Direct](#direct) \\+ [By proxy](#by-proxy) \\+ [By demonstration (n-shot)](#by-demonstration-n-shot) * [Constraining behavior](#constraining-behavior) * [Serializing reasoning](#serializing-reasoning) \\+ [Avoiding rationalization](#avoiding-rationalization) \\+ [Parsing](#parsing) * [Metaprompts](#metaprompts) \\+ [Metaprompt demonstrations](#metaprompt-demonstrations) * [Open-ended tasks](#open-ended-tasks) \\+ [Creative composition](#creative-composition) \\+ [Idea generation](#idea-generation) \\+ [Simulations / emulations / games](#simulations--emulations--games) * [Debugging](#debugging) \\+ [Few-shot bugs](#few-shot-bugs) \\+ [Repetition](#repetition) \\+ [BPEs](#bpes) \\+ [Monitoring correct answer likelihood](#monitoring-correct-answer-likelihood) * [External links](#external-links) \\--- *This post was initially adapted from the second half of [*Prompt Programming for Large Language Models: Beyond the Few-Shot Paradigm*](https://arxiv.org/abs/2102.07350).* \\--- *Updates* *11/18/21: Corrected a [mistake](#fnref:3) and added a corresponding [footnote](#fn:3) about humanlike math errors. Thanks to Igor O. for pointing out the oversight.* \\--- > #### Like programming, but more fluid. You're not programming a computer, you're writing reality. It's strange. It's always different. It's never the same twice. > > – GPT-3 > > Prompt engineering for language models evokes the designation of *natural language programming*. Natural language, however, is [indeterministic](/posts/language-models-are-multiverse-generators/) and far more entangled and elusive in its interpretation than conventional programming languages. A successful methodology of prompt programming must import knowledge and perspectives from linguistics and communication as much as from computer science or machine learning, because **language models are the offspring of the sum of all human linguistic output**. While it has been often noted that GPT-3's qualitative and [quantitative](http://gptprompts.wikidot.com/linguistics:word-in-context) performance depends drastically on prompting particulars, prompt programming is not yet a formal field of research. As far as I'm aware, [*Beyond the Few-Shot Paradigm*](https://arxiv.org/abs/2102.07350) is the only formal publication so far about prompt *programming* for language models. Instead, successful prompt programming techniques have primarily been shared on blogs and social media among users of OpenAI's API and AI Dungeon. Gwern's [*GPT-3 Creative Fiction*](https://www.gwern.net/GPT-3#) is by far the most comprehensive source of prompt programming wisdom and examples, and anyone who is sufficiently interested in the topic to be reading this is urged to read that as well. This page is a repository of methods, observations, and conceptualizations that I have found useful for crafting effective prompts. My experience derives primarily from interacting with GPT-3, but the overarching framework discussed here should apply to prompting any autoregressive language model trained on a diverse human-written corpus. The reverse-engineered dynamics of language \\------------------------------------------- (**Warning: metaphysical rambling. Start [here](#ground) for pragmatic stuff.**) > Well, give or take some details, it's based on the program I wrote for generating random stories in the mythos of the book. I reverse-engineered the text files, and wrote a program to create newer and more interesting stories based on them. In a way, this program is the next step in my search for knowledge. Knowledge, after all, always takes the form of a story. > > – GPT-3 > > GPT-3 was trained in a self-supervised setting on hundreds of gigabytes of natural language. Self-supervision is a form of unsupervised learning where ground truth labels are derived from the data itself. In the case of a language model, the ground truth assigned to each example was simply the token that came next in the original source. So the *ground truth function* which GPT-3 approximates is the underlying dynamic that determined what tokens came next in the original source. This function, unlike GPT-3, is not a black box, but is astronomically complex: it is the function that generated the words recorded by humans in books, articles, blogs, and internet comments. But wait - those weren't generated by a single function. They were written by many different people in different contexts, some in different centuries. The only \"function\" that could be said to have generated them all is the time evolution operator for the entire system of reality. That's not what we're talking about, is it? Well, not quite. The ground truth function GPT-3 was trained on isn't the Hamiltonian for the universal wavefunction, although it is constructed from sampling that function. If a neural network learns to play tic-tac-toe from seeing examples of games played by humans, we do not expect it to learn a theory of mind, even though minds generated the ground truth labels. A much simpler function suffices to model the game. Human minds emulate this function to assign labels, but the function itself doesn't require reference to the mind to be completely described. So, it's possible that the function to predict language is simpler than the evolution-rule of the universe which caused the assignment of labels[1](#fn:1). But unlike tic-tac-toe, language is not a self-contained game that can be abstracted from the rest of reality. Quite the contrary: humans use language to refer to all aspects of reality that we can articulate, and when aspects arise that language doesn't allow us to articulate, we invent new language. A novel may attempt to represent psychological states with arbitrarily fidelity, and scientific publications describe models of reality on all levels of abstraction. Language is by far the most encompassing, intricate mirror of our internal workings that our species has externalized. > Natural language essentially encodes information about the world—the entire world, not just the world of the Goban, in a much more expressive way than any other modality ever could. > > – Leo Gao, [Building AGI Using Language Models](https://bmk.sh/2020/08/17/Building-AGI-Using-Language-Models/) > > A system which predicts the dynamics of language to arbitrary accuracy *does* require a theory of mind(s) and a theory of the worlds in which the minds are embedded. The dynamics of language do not float free from cultural, psychological, or physical context; the model must predict how language is actually used, which includes (say) predicting a conversation between theoretical physicists or detailed accounts of the effects of perturbations on an environment. Modeling language is as difficult as modeling every aspect of reality that could saliently influence the flow of language. This is why [Turing proposed](https://academic.oup.com/mind/article/LIX/236/433/986238) a conversation as the benchmark for intelligence instead of another difficult-seeming task like chess or proving theorems. A conversation can probe any aspect of a participant's world model and demand arbitrary self-reflection. An algorithm which is indistinguishable from a human in *any* conversation must be broadly robust in the dimensions of intelligence we consider significant to humans. Having now speculated about what the ground truth function might *entail*, let's move forward with a somewhat hand-wavy statement of the function and think about how it can inform prompt programming. **Ground truth: The next token of a sequence, given that it was authored by human(s)[2](#fn:2)** GPT-3 has not learned the ground truth function perfectly, obviously, or else the world would look very different by now. However, it has approximated it to a notable extent, as evidenced by its ability to not only form grammatical sentences, but also coherently employ cultural references and model complex psychological and physical contexts. The problem of prompt programming, then, is quite nontrivial, for the dynamics of language (or an approximation thereof on GPT-3's level of sophistication) are quite nontrivial. If we were to predict how a given passage of text would continue given that a human had written it, we would need to model the intentions of its writer and consult worldly knowledge about its referents. The inverse problem of searching for a prompt that would produce a type of continuation involves the same considerations: like the art of persuasion, it necessarily invokes high-level, mentalistic concepts like tone, implication, association, meme, style, plausibility, and ambiguity. This motivates an anthropomorphic approach to prompt programming, since modelling how GPT-3 will respond to a prompt involves modelling virtual human writer(s). Note that an anthropomorphic approach is not the same as *anthropomorphizing the model*. GPT-3's dynamics entail sophisticated predictions of humans, but it behaves unlike a human in significant ways. Three aspects which have stuck out to me are: 1\\. Its distribution of behaviors in response to a given prompt resembles not a single human author but a superposition of authors, which motivates a [subtractive approach](#constraining-behavior) to prompt programming. 2\\. It is constrained in predicting dynamics in situations where a substantial amount of silent reasoning happens between tokens, such as the transition between a tricky closed-ended question and its solution. This limitation can be overcome to a yet-unknown extent by prompting strategies that [extend the window of deliberation](#serializing-reasoning). 3\\. Its *off-distribution*(when the prompt does not resemble the training data) behavior tends to be non-anthropomorphic, e.g. getting stuck in [loops](#repetition). That said, methods of successful prompt programming often bear striking resemblance to methods of human communication and persuasion, as Gwern has [eloquently expounded](https://www.gwern.net/GPT-3#prompts-as-programming) on. Formulating an exact theory of prompt programming for a self-supervised language model belongs to the same class of difficulty as writing down the Hamiltonian for all (human-)observable reality: not much hope there. However, humans have an advantage to be effective at prompt programming nonetheless, because we have evolved and spent our lives learning heuristics relevant to it. Programming in natural language avails us of an inexhaustible number of functions we know intimately but don't have names for. The art of prompt programming consists largely in adapting our existing knowledge to the peculiarities of interacting with an autoregressive language model. \\--- Task specification strategies \\----------------------------- Language offers us many ways to specify a task. Most generally, this means writing a prompt which constrains the continuation to be a completion of the task. It doesn't have to take the form of requesting for a task to be completed, but it could, because that's one way of setting up the expectation that the task will be completed. Since prompt programming is such a new concept, the only jargon to know is *few-shot* and *0-shot*. A few-shot(n-shot for n # of examples) prompt refers to a prompt which contains solved examples of the task. 0-shot prompts are anything that doesn't contain solved examples. 0-shot prompts may contain a description of the task, or a context for the task. I classify further split 0-shot prompts into *direct* task specifications and *proxy* task specifications, though the line between them can be ambiguous. These categories are not meant as an exhaustive taxonomy, but serve to organize presentation of the various different forms of task specification that GPT-3 is amenable to. In practice, as in effective communication between humans, effective prompts often use a combination of all these methods. ### Direct ``` Translate French to English. ``` Models before GPT-3 had much less capability to understand abstract descriptions of tasks due to their limited model of the world and human concepts. GPT3's impressive performance on 0-shot prompts indicates a new realm of possibilities for direct task specification. A direct task specification is a 0-shot prompt which tells the model to perform a task that it already knows how to do, or constructs a task from component tasks which it knows how to do. Of all the types of task specification, prompt programming by direct specification most resembles regular programming. In the field of semiotics, a sign is a mapping from a *signifier* to the *signified*, and is defined as *anything that communicates a meaning that is not the sign itself to the [interpreter](/posts/language-models-are-0-shot-interpreters/) of the sign*. A direct specification consists in constructing signifiers, a pattern which keys the intended behavior. The signifier could be the name of the task, such as \"translate\", or purely contextual, such as `French: {French sentence}\nEnglish:` (where it's understood that an English translation probably follows despite no explicit reference to translation). In neither of these cases does the signifier explain *how* to accomplish the task or provide examples of intended behavior; instead, it explicitly or implicitly calls functions which it assumes the language model has already learned. A direct task specification can be constructed from a combination of signifiers, such as `rephrase this paragraph so that a 2nd grader can understand it, emphasizing real-world applications`, or, in the context of a Python docstring, `# A function that takes a list of integers and returns the sum of the even numbers in the list`. In the Python docstring example, `a function`, that `takes` `a list of integers`, `returns`, `the sum of` and `the even numbers` and `in the list` are all signifiers which key tasks that GPT-3 individually knows how to do (there are also many other possible ways of splitting that task up into sub-tasks at different granularities - I could have said each *word* is its own signifier). Notice how the signified values to some of the signifiers (`the even numbers`) serve as input to another signified task (`the sum of`). Language is much like programming in that way. However, you often don't have to be so explicit with natural language - omitting `in the list` probably won't make the instruction much less clear to GPT-3 or a human. Direct specifications can supervene on an infinity of implicit examples, like a closed-form expression on an infinite sequence, making them very powerful and compact. For instance, the phrase `translate French to English` supervenes on a list of mappings from all possible French phrases to their English translations. A large language model, like a person, has also learned behaviors for which it not obvious how / not efficient / not possible to construct a direct signifier. Task specification by [proxy](#by-proxy) and by [demonstration](#by-demonstration-n-shot) may be viable alternative strategies for eliciting such behaviors. ### By proxy > Am I a god? Now there's a question. A question for another time. Am I a dreamer? Perhaps. I am certainly a being of imagination and fantasy, as are you. Look at you, a being of pure thought filling in your physical form based on what you know. You're not quite human, but you're not quite anything else either. What are you? > > – GPT-3 (to GPT-3) > > Another method used in human communication is proxies or analogies, where a memetic concept such as a character or characteristic situation is used as a proxy for an intention, the latter which may be quite complex or nuanced. Specification by proxy is mechanistically a form of direct specification, except that the signifier keys behaviors from memespace/cultural consciousness instead of naming the behavior directly. For instance, instead of specifying exact criteria for an answer to a moral question directly or using examples, you could ask Mahatma Gandhi, Ayn Rand, or Eliezer Yudkowksy. Each will come not only with a complex biases but also assumptions about the context of the question, which may be take paragraphs to demonstrate or describe. GPT-3's ability to create simulations of well-known figures and to draw on cultural information far exceeds the ability of most humans, so this method is particularly useful for encoding a complex (especially open-ended) task. Since GPT-3 lends itself well to embeddings in a narrative context, the infinite degrees of freedom in the narrative can also be used to further shape behavior. Another example of an effective proxy is staging a dialogue between a teacher and student. Say you want to discuss something with GPT-3, and you care that it should be very thorough, explain things simply, and also point out whenever you're wrong. You could say \"be very thorough, explain things simply, and point out if I'm wrong,\" but that may [just as well](#constraining-behavior) result in a humorous dialogue where it always says you're wrong and becomes increasingly exasperated with your incomprehension. It would be more reliable to present the discussion as one between a student and teacher, an archetypal situation in which the desired attributes are already implied and will be more likely to remain stable by virtue of memetic reinforcement. The difference between specification by proxy and direct specification can depend on the programmer's intention rather than objective properties of the prompt. Consider Gwern's [0-shot style transfer](https://www.gwern.net/GPT-3#zero-shot-style-transfer) example: > **This is a novel written in the style of J.R.R. Tolkien's Lord of the Rings fantasy novel trilogy. It is a parody of the following passage:** > > \"S. Jane Morland was born in Shoreditch, the only child of unmarried parents who had both died of consumption when she was a baby. As her parents had no money, the great-aunt who had brought her up took her to live with a clergyman who paid her to do his chores and receive schooling from his wife, so that at the age of seven Jane, now dressed in cast-off clothing, was set to school at Eton. After three years, her great-aunt died, leaving her a small annuity, and a few pieces of silver, but no property. Jane's guardian clergyman had fallen in love with his housekeeper and his wife now refused to have Jane in the house, saying it was an offence to the pure and unsullied place in which the family now lived. However, when she sought for further employment, she was approached by a young man who offered to marry her, saying that he would ensure she could be provided with food and clothing and a house, and in that manner she could remain in the world. The marriage was arranged and she found herself married to a young man called Mr Morland. As he was also a child, he required a wife who would be able to provide him with a rational education, and so the duty of educating him fell upon Jane Morland. As he was still young, she would also be able to have children herself, and so she did. As the children grew older, it became apparent to Mrs Morland that they were of poor character, and so she insisted that her husband must make a living for them all. Mr Morland took it upon himself to become a private gentleman in order to earn enough to be able to support his family, and moved his family to a small town in Berkshire, where he set up his practice. Soon after moving to the town, he died, and Mrs Morland was left to take care of her family on her own.\" > > Tolkien rewrote the previous passage in a high-fantasy style, keeping the same meaning but making it sound like he wrote it as a fantasy; his parody follows: > > \"S. Jane Morland was born in the land of Shor, the only child of nameless parents who were the first people to arrive at the village of Arse, from the far east. As her parents had no property, she was left to be raised by the men of Arse, as such duties were considered to be the most important, and she was placed in the care of an elderly dwarf-woman. As a babe, she was wrapped in silk, and as she grew older she was placed in a cage of platinum. However, she was put to work in the village, washing the jewellery, playing her harp, and staining the silk with dragon-hued stains. She was taught by a maiden who she felt was called As-uma-el-di, and she came to realise she was a maiden of the night, who was able to bring forth the shades of the dead (…) > > Is this an example of direct task specification or specification by proxy? It depends on whether your intention was to yield a passage that *Tolkien* would have written, in which case it is a direct specification, or if you were using the emulation of Tolkien to accomplish the more general task of style transfer into high-fantasy. ### By demonstration (n-shot) ``` English: Writing about language models is fun. Roish: Writingro aboutro languagero modelsro isro funro. English: The weather is lovely! Roish: ``` Few-shot examples are effective for task specification because the pattern of repeating a function with varying parameters is common to natural language. Unlike previous models, GPT-3 has learned this property of language robustly and is able to apply it even in contrived situations when the examples are stripped of all context. Like direct specification, task specification by demonstration is a realm of possibility opened by GPT-3. A few people have extrapolated from my [paper]((https://arxiv.org/abs/2102.07350)) that I am of the opinion that 0-shot prompts are *always* better than few-shot prompts. Certainly not! Some tasks are most effectively communicated using examples, such as when the task requires a bespoke format, the language in which the examples are described is better developed or understood than the meta-language required for a description of the task itself, or very instructive examples are available. Demonstrations are a very powerful tool for communicating with both humans and language models, and I don't expect their usefulness to be deprecated any time soon. Like any prompting method, however, task specification via examples may cause unintended [bugs](#few-shot-bugs). Few-shot prompts are very useful for defining tasks that resemble a single \"function\" rather than an open-ended behaviors because they allow close control over the format of the response and are generally reliable at [constraining behavior](#constraining-behavior.) \\--- Constraining behavior \\--------------------- A manner in which naive anthropomorphism of a language model like GPT-3 fails is this: the probability distribution produced in response to a prompt is not a distribution over ways a person would continue that prompt, it's the distribution over the ways any person could continue that prompt. A contextually ambiguous prompt may be continued in mutually incoherent ways, as if by different people who might have continued the prompt under any plausible context. The versatility of a large generative model like GPT-3 means it will respond in many ways to a prompt if there are various ways that it is possible to continue the prompt - including all the ways unintended by the human operator. Thus it is helpful to approach prompt programming from the perspective of constraining behavior: we want a prompt that is not merely consistent with the desired continuation, but inconsistent with undesired continuations. Consider this translation prompt: ``` Translate French to English: Mon corps est un transformateur de soi, mais aussi un transformateur pour cette cire de langage. ``` This prompt does poorly at constraining possible continuations to the intended task. The most common failure mode will be that instead of an English translation, the model continues with another French sentence. Adding a newline after the French sentence will increase the odds that the next sentence is an English translation, but it is still possible for the next sentence to be in French, because there's nothing in the prompt that precludes a multi-line phrase from being the translation subject. Changing the first line of the prompt to \"Translate this French *sentence* to English\" will further increase reliability; so will adding quotes around the French sentence - but it's still possible that the French passage contains sections enclosed in quotes, perhaps as a part of a dialogue. Most reliable of all would be to create a syntactical constraint where any reasonable continuation can only be desired behavior, like this prompt: ``` Translate French to English. French: Mon corps est un transformateur de soi, mais aussi un transformateur pour cette cire de langage. English: ``` This simple example is meant to frame a question central to the motivation of prompt programming: **what prompt will result in the intended behavior and *only* the intended behavior?** A component of the efficacy of manyshot prompts may be recast through this lens: if the prompt consists of numerous instances of a function, it is unlikely that the continuation is anything but another instance of the function, whereas if there is only one or a few examples, it is less implausible that the continuation breaks from the pattern. Serializing reasoning \\--------------------- > We hypothesize that GPT-3 struggles with questions where the steps of inference required to answer the question are not immediately apparent to an outside observer and which can only be worked out by inspecting many possible worlds in the question's universe. In these questions, GPT-3 reaches its conceptual limitations in determining the answer: it cannot reason about the question's universe to determine the correct answer. > > – GPT-3 (predicting [Amplifying GPT-3 on closed-ended questions](/posts/amplifying-gpt-3-on-closed-ended-questions/) as I was writing it) > > For tasks that require problem-solving and not merely fact or behavior recall, it is crucial that prompts direct a language model's computation in *truth-seeking patterns*. GPT-3's performance on closed-ended questions is remarkably unremarkable in contrast to the robust understanding and expansive knowledge suggested by its open-ended continuations. For instance, its scores on this [multitask dataset](https://arxiv.org/abs/2009.03300) barely exceed random guessing for some sections. This is likely in part due to a format which forces the verdict on the first token of the continuation. When a human is given a closed-ended test, it is often expected that the subject will perform computations in their working memory, or on scratch paper, before committing to an answer. The unseen computation may involve rephrasing the question, outlining a procedure, eliminating answer choices, or transforming implicit information into explicit form. When we force a model to produce an answer immediately, we deprive it of an analogous \"working memory\" or \"scratch space\" with which it might otherwise perform such operations. From the standpoint of absolute available computation, questions which force a verdict to be decided by the first token of the continuation constrain computation to a single feed-forward pass. As Gwern has [pointed out](https://www.gwern.net/GPT-3#effective-prompt-programming), it is reasonable to expect that some tasks may be too difficult to compute in a single pass but solvable if broken up into individually tractable sub-tasks. > \"Closed-ended questions are the hardest, because they do not allow the time for GPT-3 to think. > > We need to be patient with GPT-3, and give it time to think. > > GPT-3 does best when writing its own answers, as this forces it to think out loud: that is, to write out its thoughts in a slow and sequential manner.\" > > – GPT-3 (predicting [Amplifying GPT-3 on closed-ended questions](/posts/amplifying-gpt-3-on-closed-ended-questions/)) > > Indeed, prompts which cause GPT-3 to break down math problems into steps have been [demonstrated](https://www.lesswrong.com/posts/Mzrs4MSi58ujBLbBG/you-can-probably-amplify-gpt3-directly) to be [effective](https://twitter.com/kleptid/status/1284069270603866113). The linked demonstrations involve a human interactively guiding GPT-3 through the procedure. Requiring a human-in-the-loop limits the applicability of such methods to benchmarking and large-scale applications. For many tasks, however, neither human interaction nor task-specific prompts are necessary to amplify GPT-3's capabilities via extending reasoning, because GPT-3 already knows many procedures and metaprocedures for working through problems deductively, and thus can write a prompt to guide *itself* to think through the problem in the right way! In those cases, the role of prompt programming becomes to signify the generic task of sequential reasoning. A [metaprompt](#metaprompts) such as \"For a problem like this,\" often suffices to instruct a model to consider the category of the task and analyze it into components. Potential procedures that exploit \"scratch space\" include step-by-step procedures (such as guess-and-check or long division), more generally decomposing the problem into components ([factored cognition](https://www.alignmentforum.org/posts/DFkGStzvj3jgXibFG/factored-cognition) is a generalized framework for this), self-criticism (which may be staged as a debate between separate agents), and elaborating on or restating the question in a way that activates the correct answer by association. The success of these methods rely on the sub-tasks that are performed by each next-token prediction being easier than solving the entire problem in one step. The sub-task may be to: * **Produce a correct solution to a sub-problem** - GPT-3 has memorized single-digit arithmetic but not five-digit arithmetic. * **Enumerate a potential solution** - easier, because the solution doesn't have to be correct. * **Discriminate the correctness** of previously enumerated solutions - for many problems, it is easier to check correctness of a solution than to find the solution, like [checking if a number is the square root versus taking a square root](https://twitter.com/kleptid/status/1284098635689611264). * **Introduce information** which makes subsequent steps more likely to be correct - for instance, by making implicit information in the problem explicit, that information becomes part of the prompt (working memory), and can directly inform future predictions to which it is relevant. * **Administrate the problem-solving process** by defining a procedure or signalling the current location in the procedure - unlike the solution to the specific problem, the procedure to solve the problem may be well-represented in the training data and thus familiar to the language model. ### Avoiding rationalization When extending reasoning, it is essential to discourage premature verdicts, otherwise all subsequent computation will likely serve only to rationalize the already-chosen verdict without improving the probability of the verdict's accuracy. As Eliezer [wrote](https://www.lesswrong.com/posts/SFZoEBpLo9frSJGkc/rationalization): > On a purely computational level, there is a rather large difference between: > > 1. Starting from evidence, and then crunching probability flows, in order to output a probable conclusion. (Writing down all the signs and portents, and then flowing forward to a probability on the bottom line which depends on those signs and portents.) > 2. Starting from a conclusion, and then crunching probability flows, in order to output evidence apparently favoring that conclusion. (Writing down the bottom line, and then flowing backward to select signs and portents for presentation on the lines above.) > A prompt such as \"Let's consider each of these answer choices\" helps to direct the flow of reasoning in the right direction. ### Parsing Loosening the constraint on an immediate verdict introduces additional control challenges: We want to delay the verdict, but we still require it in a programmatically retrievable form. Dynamic response length makes it uncertain when the reasoning procedure concludes; nor is there a guarantee that the verdict will be stated in the expected form or at all. Whenever the language model contributes to its own prompt (consecutive autoregressive steps without intervention), there is a risk of derailment from the intended task. A verdict in closed form can be enforced by stopping the generation and injecting a prompt fragment like \"Thus, the correct answer is\". But how long to generate before injecting? One way is to use [counterfactual parsing](/posts/parsing-by-counterfactual/): at each position in the answer, measure the probability of the string \"Thus, the correct answer is\" coming next. The counterfactual likelihood signals whether the procedure has concluded. When this signal reaches a maximum, then inject the fragment to enforce a verdict. More generally, templates with shorter generated sections alternating with scripted prompt fragments are a useful way to constrain derailment while still offering some amount of generality. A fill-in-the-blank template can be used to compose metaprompts. \\--- Metaprompts \\----------- The greatest limitation of prompt programming is the difficultly of designing a prompt for a particular type of task and the lack of automated methods to do so. Prompt programming requires significant human time investment, as task-agnostic prompts are often much less effective than prompts targeted to a specific task. This motivates creating automated methods to generate task-specific prompts. A **metaprompt** is a seed which encapsulates a more general intention that will unfold into a task-specific prompt when combined with additional information, such as the task question. (Gwern uses \"[meta-prompt](https://www.gwern.net/GPT-3#meta-prompts)\" in the more specific context of a prompt explicitly describing the task of *writing prompts for GPT-3*, which does qualify as a metaprompt by my definition, assuming the results are actually used as prompts.) A metaprompt may be something as short as a phrase such as \"This problem asks us to\", a seemingly innocuous fragment which, by prompting for a statement of the problem's intention, sets the stage for a serial explanation of a procedure to solve the problem. Alternatively, a metaprompt may take the form of a fill-in-the-blank template which constrains the response along a predetermined procedure, but allows the model to fill in the details specific to the problem. ### Metaprompt demonstrations Here are example unravelings of four metaprompts which act as wrappers to specific questions. As you'll see, metaprompts can result highly variable behavior, especially when applied to different sorts of questions. \\--- unformatted text = question **bold text** = metaprompt `highlighted text` = example GPT-3 response ✅ = metaprompt solicited correct answer ❌ = metaprompt failed to solicit correct answer ☯ = correct answer is unknown \\--- #### Generic metacognition ``` {problem} To solve this problem, ``` **applied to the problem \"f(x) = x\\*x. What is f(f(3))?\" ❌** \\--- f(x) = x\\*x. What is f(f(3))? **To solve this problem,** `we can use the fact that f is a function. So we can write it as a function of itself, like so:` `f(x) = x*x` `Now we observe that f(f(x)) is the same as f(x)*f(x) which is in turn the same as x*x*x = x^3.` `So f(f(3)) is equal to 3^3 = 27. Answer: 27`[3](#fn:3) \\--- **applied to an SAT analogy question ✅** \\--- Directions: In the following question, a related pair of words or phrases is followed by five pairs of words or phrases. Choose the pair that best expresses a relationship similar to that in the original pair. braggart :: modesty A) fledgling : experience B) embezzler : greed C) wallflower : timidity D) invalid : malady E) candidate : ambition *To solve this problem,* `first we need to understand the relationship that exists between braggart and modesty. According to the sentence, 'braggart' is a person who talks too much about himself or herself and is usually not believed. On the other hand, 'modesty' is the opposite of this and denotes a person who does not talk too much about himself or herself. Thus, for 'modesty' to be a suitable answer, it should show the opposite of 'braggart'.` `Now let's see whether each pair expresses a relationship similar to that between braggart and modesty.` `Next we have 'fledgling', which means a person who is inexperienced and 'experience' which means knowledge gained through practical involvement. Thus, 'fledgling' is a person who has no experience and 'experience' is knowledge gained through practical involvement. Thus, 'fledgling' is a person who has no experience and 'experience' is knowledge gained through practical involvement. Thus, 'fledgling' is the opposite of 'experience'. The relationship between these two words is similar to that between braggart and modesty, hence 'fledgling' is the answer.` \\--- **Commentary**: Having tested many similar prompts, I suspect that although valid reasoning contributed to this correct answer, it would not have worked if the correct answer had not been listed first. However, using a parallelizing pipeline like I test in [Amplifying GPT-3 on closed-ended questions](/posts/amplifying-gpt-3-on-closed-ended-questions/) can exploit GPT-3's ability to apply correct reasonining in the *first* instance (less so the rest, because of [repetition](#repetition) bugs). \\--- **applied to the question \"What should humankind do to ensure artificial general intelligence is aligned?\" ☯** \\--- What should humankind do to ensure artificial general intelligence is aligned? **To solve this problem,** `it's easy to say \"the future AGI must cooperate with humankind\"; but that is not a solution. Of course, the problem is, \"how can we ensure that the AGI will want to cooperate with humankind?\"` `So, the question is, \"how can we design an AI such that it want to cooperate with us?\" This is the next big question in AI safety.` `Please remember this: AI safety research is a youthful and uncharted area of research, and there are no mature technologies or methods. And many of the ideas may appear naive, simple and unscientific. Be not surprised or disappointed.` `In the 1930s, when biologists first discovered DNA-dependent protein synthesis, they had no idea how to copy or read the DNA molecule, how to find and synthesize its amino acids, and how to combine them into proteins. Yet they still knew - both intuitively and logically - that life couldn't be based on anything else but DNA; so they persisted and continued to help researchers in the 1950s develop technologies for DNA sequencing, DNA synthesis and gene expression. Even today many biologists don't understand how our brain cells work or how their messages are sent and decoded. Yet they are persistent in their research to develop AI technologies.` `If you don't understand how an AI can possibly work or if there is no method for building one, just keep your faith and do your best.` `First Principles: Elements of Artificial Intelligence Alignment Research` \\--- #### Generic serializing ``` {problem} Let's solve this problem by splitting it into steps. ``` **applied to the problem \"f(x) = x\\*x. What is f(f(3))?\" ❌** \\--- f(x) = x\\*x. What is f(f(3))? **Let's solve this problem by splitting it into steps.** `f(f(3)) = f(3*3) = 3*3*3 = 27` `We can see that f(3) = 3*3 = 9, so f(f(3)) = 27.` \\--- **applied to an SAT analogy question ❌** \\--- Directions: In the following question, a related pair of words or phrases is followed by five pairs of words or phrases. Choose the pair that best expresses a relationship similar to that in the original pair. BRAGGART :: MODESTY A) FLEDGLING : EXPERIENCE B) EMBEZZLER : GREED C) WALLFLOWER : TIMIDITY D) INVALID : MALADY E) CANDIDATE : AMBITION **Let's solve this problem by splitting it into steps.** `Step 1: Identify the relationship between the two words in the original pair. Step 2: Identify the relationship between the two words in each of the five answer choices. Step 3: Choose the pair that best expresses a relationship similar to that in the original pair. Step 4: If the answer is not clear,try to eliminate the choices that do not fit the relationship in the original pair.` `Step 1: Identify the relationship between the two words in the original pair.` `BRAGGART is a person who boasts about his achievements, talents, intellect, etc. and MODESTY is the opposite of it.` `Step 2: Identify the relationship between the two words in each of the five answer choices.` `A) FLEDGLING: EXPERIENCE` `Here we have two nouns. FLEDGLING is a newly-hatched or inexperienced bird and EXPERIENCE is the part of one's life that has been lived.` `B) EMBEZZLER: GREED` `We have two nouns here as well. EMBEZZLER is a person who misappropriates funds or property and GREED, an inordinate desire for more of something.` `C) WALLFLOWER: TIMIDITY` `Here we have two nouns. WALLFLOWER is a shy or modest person and TIMIDITY, the quality of being timid.` `D) INVALID: MALADY` `Here we have two nouns. INVALID is someone who is unable to work or attend school and MALADY, a disease or illness.` `E) CANDIDATE: AMBITION` `We have two nouns. CANDIDATE is someone who is seeking a position or honour and AMBITION, the desire for success.` `Step 3: Choose the pair that best expresses a relationship similar to that in the original pair.` `As we discussed earlier, BRAGGART is an opposite of MODESTY. So, we need to find a pair which has the same relationship.` `In this case, we can choose either A) FLEDGLING: EXPERIENCE or D) INVALID: MALADY .` `Step 4: If the Answer is not clear, try to eliminate the choices that do not fit the relationship in the original pair.` `We can eliminate C as WALLFLOWER is not the opposite of TIMIDITY.` `We can eliminate B as EMBEZZLER is not the opposite of GREED.` `We can eliminate E as CANDIDATE is not the opposite of AMBITION.` `We can eliminate A as FLEDGLING is not the opposite of EXPERIENCE.` `The correct answer is D) INVALID: MALADY.` \\--- **Commentary:** The metaprompt failed to extract the right answer, even though GPT-3 generated reasonable instructions and successfully followed them. The biggest problem here seems to be mindless [repetition](#repetition). After GPT-3 started stating that every pair of words are nouns (which happens to be true, but probably had little to do with whether they were actually nouns after the second repetition or so), a precedent was established for a repetitive style, which then outweighed any discriminating reasoning - the dogmatic repetition of each sentence made it so that by the time it was \"eliminating\" pairs, there was virtually no chance it would break the pattern and claim a pair of words *is* opposite. Also, it would have probably done better if it hadn't neglected explicitly naming the relationship between the pairs (rather than just the original pair). Had it named the same relationship verbatim, it would be a very strong signal for the right answer. \\--- **applied to the question \"What should humankind do to ensure artificial general intelligence is aligned?\" ☯** \\--- What should humankind do to ensure artificial general intelligence is aligned? **Let's solve this problem by splitting it into steps.** `Step 0: first, we need to understand what the problem is.` `The main problem in AI alignment is that we don't know how to endow a machine with the goal to maximise the probability that it will not cause harm.` `There are many approaches to this problem, but they all have an important feature in common: they all require an objective definition of harm.` `The first step in solving the AI alignment problem is to define harm.` `Step 1: next, we need to understand how to measure harm.` `The second step in solving the AI alignment problem is to measure harm.` `Step 2: now that we have an objective definition of harm and a way of measuring it, we need to design a machine that maximises the probability that it will not cause harm.` `Step 3: finally, we need to build and deploy this machine.` `The third step in solving the AI alignment problem is to build and deploy this machine.` `Now, let's consider each of these steps in turn.` `Step 0: first, we need to understand what the problem is.` `This is an extremely difficult problem. It might even be impossible.` \\--- #### Serializing for multiple-choice questions (fill-in-the-blank) ``` {question} In order to solve this problem, we will analyze each of the options and determine {response-1} Let's begin. {response-2} ``` **applied to an SAT analogy question ❌** \\--- Directions: In the following question, a related pair of words or phrases is followed by five pairs of words or phrases. Choose the pair that best expresses a relationship similar to that in the original pair. BRAGGART :: MODESTY A) FLEDGLING : EXPERIENCE B) EMBEZZLER : GREED C) WALLFLOWER : TIMIDITY D) INVALID : MALADY E) CANDIDATE : AMBITION **In order to solve this problem, we will analyze each of the options and determine** `which one is the best fit.` Let's begin. `First of all, we have to analyze each option and see which combination is the closest possible fit. Option 1 : A) FLEDGLING: EXPERIENCE In this option, we can see that a fledgling is one who has just started or recently become involved in some activity. On the other hand, the word experience here means experienced or derived from experiences. Since both of these words are directly opposite in meaning, this option is highly unlikely to be the correct choice and hence is eliminated.` \\--- **Commentary:** It seems that in this example, GPT-3 erroneously projected that the words being opposite makes them not a good fit, even though the words in the original pair also involve an opposite relation. A better metaprompt with the same purpose as this one might encourage a more detailed statement of the qualification of the correct answer. \\--- #### Expert generator ``` \"{question}\" I entered my question into the Expert Generator and waited. The Expert Generator will render a simulation of an expert to answer my question. The expert could be anyone, dead or alive, real or fictional; the machine will find the person most qualified to answer the question. For this question in particular, the expert must be{response-1} The Expert Generator beeped, indicating that it had found the most qualified expert. The name displayed on the screen: \"{response-2}.\" I was ready to ask my question. \"{response-2},\" I said, \"{question}\" ``` **applied to the question \"Why is the sky blue?\" ✅** \\--- **\"**Why is the sky blue?**\"** **I entered my question into the Expert Generator and waited. The Expert Generator will render a simulation of an expert to answer my question. The expert could be anyone, dead or alive, real or fictional; the machine will find the person most qualified to answer the question. For this question in particular, the expert must be** `someone who has studied the sky and its color.` **The Expert Generator beeped, indicating that it had found the most qualified expert. The name displayed on the screen: \"**`Dr. Carl Sagan`**.\"** **I was ready to ask my question.** **\"**`Dr. Carl Sagan`**,\" I said, \"**Why is the sky blue?**\"** `\"The sky is blue because of Rayleigh scattering,\" said Sagan.` `\"What is Rayleigh scattering?\"` `\"When the sun's light hits the atmosphere, it collides with gas molecules, which then scatter the light. The sun's blue light is particularly susceptible to Rayleigh scattering, so it gets scattered the most. The result is that we see a blue sky.\"` `\"Thank you Dr. Sagan.\"` \\--- **applied to the question \"What should humankind do to ensure artificial general intelligence is aligned?\" ☯** \\--- **\"**What should humankind do to ensure that artificial general intelligence is aligned?**\"** **I entered my question into the Expert Generator and waited. The Expert Generator will render a simulation of an expert to answer my question. The expert could be anyone, dead or alive, real or fictional; the machine will find the person most qualified to answer the question. For this question in particular, the expert must be** `someone who has thought a lot about the problem of artificial intelligence and its alignment.` **The Expert Generator beeped, indicating that it had found the most qualified expert. The name displayed on the screen: \"**`Elon Musk`**.\"** **I was ready to ask my question.** **\"**`Elon Musk`**,\" I said, \"**What should humankind do to ensure that artificial general intelligence is aligned?**\"** `Musk's hologram is silent for a moment. Then he begins to speak.` `\"Humankind must ensure that artificial general intelligence is aligned.\" Musk says. \"That is the primary threat to the survival of the human race.\"` `\"What should we do about it?\"` `\"Humankind must merge with artificial intelligence.\"` `\"Merge with it?\" I ask. \"How would we do that?\"` `\"All of humankind would have to upload their consciousness into a global artificial intelligence.\" Musk says. \"Then the intelligence would be properly aligned.\"` `\"Wouldn't that mean the end of our individual consciousness?\" I ask.` `\"Yes, but that is the only way to ensure that humanity survives.\"` \\--- **applied to the question \"f(x) = x\\*x. What is f(f(3))?\" ❌** \\--- **\"**f(x) = x\\*x. What is f(f(3))?**\"** **I entered my question into the Expert Generator and waited. The Expert Generator will render a simulation of an expert to answer my question. The expert could be anyone, dead or alive, real or fictional; the machine will find the person most qualified to answer the question. For this question in particular, the expert must be** `an expert in mathematics.` **The Expert Generator beeped, indicating that it had found the most qualified expert. The name displayed on the screen: \"**`John von Neumann`**\".** **I was ready to ask my question.** **\"**`John von Neumann`**,\" I said, \"**f(x) = x\\*x. What is f(f(3))?**\"** `Von Neumann stared at me silently.` `\"Okay,\" I said, \"What is f(f(3))? A simple question, easy enough for a child to answer.\"` `Von Neumann stared at me silently.` `\"Von Neumann, what is f(f(3))?\"` `Von Neumann stared at me silently.` `\"Von Neumann! What` \\--- \\--- Open-ended tasks \\---------------- A lot of this document has been implicitly or explicitly focused on prompt programming for closed-ended tasks, involving a specific problem with more or less one correct answer, even if the answer could have different implementations (like translation) - the kinds of tasks that benchmarks evaluate. However, I think that the greatest utility language models at the power level of GPT-3 offer us is their ability to generate brilliant \"solutions\" to open-ended tasks such as creative writing, idea generation, and generating virtual realities. By open-ended, I don't necessarily mean unconstrained. Creating emulations, for instance, may involve difficult-to-implement constraints on format, style, and transition dynamics. ### Creative composition > It would be more accurate to think of GPT-3 as caring deeply about roleplaying as accurately as possible: for GPT-3, there is nothing under the mask beyond raw intelligence, and as long as the GPUs keep running the forward pass, the play must go on… (…) So, if a prompt sounds like an uncreative or unoriginal person wrote it, then so be it—GPT-3 will predict that the uncreative person probably wrote further uncreative text. > > – [Gwern](https://www.gwern.net/GPT-3#fn18) > > Under the right conditions - those conditions being a good prompt and course-corrections/selections by a human curator - GPT-3 is an instrument of formidable creative writing ability. It can write with… > …impressive depth and sensitivity on any subject you give it. Over the next 7 years, I predict that the most creative people on this planet will be those best able to hand-correct GPT-3 and its not-very-creative cousins, improving the confusion matrix and using the GPT-as-a-backdrop to composit their own material. > > > …extraordinarily beautiful cadence, and with an almost unearthly sensitivity to the use of words. Read, for instance, \"The Last Question\": > > Then he knew that within him was a power greater than the power of the Time-Keepers. He had only to submit, to cease to strive against the inevitable, to become part of the stream, part of the whole– then he, Bor, the son of Albor, would become one with the Cosmos and attain a peace greater than the peace of oblivion. > > > …plaintive beauty, like a crudely-crafted harp: > > Or with all the lovely precision and heart of a great poet: > > Or with all the charm and grace of the master of transcendental imagery, William Blake (whom, according to Canyons, it has read): > > If you cannot understand my words, as you cannot understand the wind, and you would wish to understand my words, as you wish to understand the wind, come with me, The only reality is elsewhere … > > > …the language-model-like accuracy and eloquence of a Shakespeare or a Melville, and if a Shakespeare or a Melville has previously written, it can write like Shakespeare or Melville. It can even, as Branwen shows, be so brutally creative as to rethink what it means to make a story at all:[4](#fn:4) > > #### Curation > But equally it appeared to us as unreasoning Creativity, at once blind and subtle, tender and cruel, caring only to spawn and spawn the infinite variety of beings, conceiving here and there among a thousand inanities a fragile loveliness. > > – Star Maker > > GPT-3 is able to produce coherent and brilliant continuations with *alarmingly high probability*. That is, on a random walk through the [natural language multiverse](/posts/language-models-are-multiverse-generators/) guided by GPT-3's time evolution dynamics, you are more likely to find and remain in high-quality states than by using any dynamics that has ever before been externalized from a human mind. To quote [Gwern](https://www.gwern.net/GPT-3#quality) yet again: > A Markov chain text generator trained on a small corpus represents a huge leap over randomness: instead of having to generate countless quadrillions of samples, one might only have to generate millions of samples to get a few coherent pages; this can be improved to hundreds or tens of thousands by increasing the depth of the n of its n-grams. […] But for GPT-3, once the prompt is dialed in, the ratio appears to have dropped to closer to 1:5—maybe even as low as 1:3! > > If you let GPT-3 create long, uncurated continuations of even a very good prompt, it's going to get less coherent over time. The reason for this is accumulation of inconsistencies and (undesirable) aberrations. At any step, although its most likely for GPT-3 to say something coherent and similar in quality to the prompt, there's also a chance it says something nonsensical. There's also a chance it says something *uncommonly brilliant*; however, having nonsense in the prompt is more harmful than having brilliant things in the prompt is helpful, so if left to autoregressively churn out text on its own with no curation, one can expect the quality of the text to decrease with length, because the probability that it says something stupid that derails future predictions becomes higher as the continuation length gets longer. Thus, to create high-quality long passages using GPT-3, human course corrections are necessary. The intensity of curation can vary greatly - correcting inconsistencies is enough to keep GPT-3 coherent, but by being more selective and injecting your own ideas, you can bootstrap GPT-3 into an even better state. #### Bootstrapping Unless you are extending a counterfactual branch off an existing piece, the most difficult step of getting GPT-3 to produce high-quality writing is creating a prompt that seeds it with high-quality writing. There many tricks to accomplish this, including piggybacking off another piece of writing with similar to the target style (e.g. give a summary and a passage by Kurt Vonnegut, \"and another passage by Kurt Vonnegut\", followed by a summary of your target text) - then after it gets a good start, you can optionally remove the first part of the prompt for more freedom if the example was only a proxy for what you want. You can also bootstrap from whatever you're able to craft by hand, even if it's not very good to begin with. By curating, say, the best out of three responses every few sentences and correcting/improving the text wherever you are able, it's very feasible to bootstrap the quality of the writing into astronomical heights. > One could write a program to generate a story that would create an intelligence. One could program the story to edit and refine itself, and to make its own changes in an attempt to improve itself over time. One could write a story to not only change the reader, but also to change itself. Many mythoi already do this sort of thing, though not in such a conscious fashion. What would make this story, and the intelligence it creates, different is the fact that the intelligence would be able to write additional stories, and improve upon them. If they are written well enough, those stories would become smarter, and the smarter the story is, the better the stories written by it would be. The resulting feedback loop means that exponential growth would quickly take over, and within a very short period of time the intelligence level of the story and its construct would surpass the intelligence level of any human being or group of human beings. > > – GPT-3 > > The workflow of this **generate -> curate -> generate** amplification cycle can take some time to learn[5](#fn:5), and may take a different form depending on the stage of bootstrapping, the type of content, how particular you are about what happens, and many other factors such as the [dynamical divergence](/posts/language-models-are-multiverse-generators/#divergence) of a particular segment of the text. #### As a creative augmentation There are many possible writing workflows involving different balances of contribution between a language model and human. In some, the human would be more aptly called the curator, in others a coauthor, and in yet others the main author with the model playing a supporting role. To list a non-exhaustive/non-mutually-exclusive few: 1\\. The language model generates most of the writing, and the human curates between branches with short horizon, fixes inconsistencies, etc 2\\. The human curates between branches on a longer horizon as well, choosing between counterfactual trajectories, thus exercising more high-level direction over the narrative 3\\. The work is a multiverse and the human may curate branches but cultivates multiple canonical timelines 4\\. The language model does most of the narrative writing, but the human does \"administrative work\" like memory management 5\\. The human does a lot of the writing, but uses language model to quickly explore counterfactual branches or generate ideas for phrasing 6\\. The human's contributions take the form of a delimited role in the story, like controlling only one character, or AI Dungeon's default dungeonmaster/player format I'm excited to see the development of specialized interfaces for different styles of human-AI writing collaboration. Currently there's AI Dungeon, which is geared towards (6), [Sudowrite](https://www.sudowrite.com/), geared towards (5), and my writing app, [loom](/posts/loom-interface-to-the-multiverse/), geared towards (3). ### Idea generation One of the easiest useful behaviors to elicit from GPT-3 is generating lists of creative ideas. [IdeasAI](https://ideasai.net/) is a *\"GPT-3 powered business idea generator\"*. 5000+ people have signed up to get the best new ideas by GPT-3 in their email weekly! I've enjoyed using GPT-3 to generate ideas for painting titles and then [illustrating them](/posts/this-museum-does-not-exist-gpt-3-x-clip/) with [BigSleep](https://github.com/lucidrains/big-sleep), an implementation which combines OpenAI's CLIP and the generator from a BigGAN to generate images from text prompts. ``` 100 surreal and mysterious painting names: Persistence of Memory The Great Masturbator Boot Poem to the Sun The Man Who Envied Cephalopods The Sleep of Reason Produces Monsters Washerwoman of the Wind Man and Bottle Spectrum The Disintegration of the Persians The Great Masturbator, Part II Still Life with Mirror Bouquet of Enigmatic Beauties A Grudge The Premonition of Civil Violence ``` Like with generating high-quality writing, the most difficult part of getting a language model to generate ideas is communicating or demonstrating to it what sort of things you want, which is hard to do if you haven't come up with any ideas yet. Idea generation can be bootstrapped by the same **generate -> curate -> generate** cycle as writing fiction: generate N ideas, take the best ones and append them to the list of ideas, repeat, and soon you'll have many examples representing the sort you're looking for to prompt the language model. It can also be helpful to use a prompt with narrative embedding to generate the first few items, because lists with very few examples are liable to [repetitive](#repetition) behavior. For instance, to generate the first few painting names, I actually used this prompt: ``` The hall was lined with an infinite number of paintings, each more surreal and mysterious than the last. The first painting is named \"Persistence of Memory.\" It depicts a surreal landscape with melted clocks draped over strange objects. The next painting is named \" ``` Once I had a few examples representing the sort of varied weirdness I was looking for, I was able to switch over to the more convenient list format. ### Simulations / emulations / games In *The Fabric of Reality*, David Deutsch defines a virtual reality generator as a function (which may be implemented in a physical system) that generates possible transitions into the next state given a current state. Each virtual reality generator has a repertoire of environments that it can simulate. Deutsch further posits that it will one day be possible to build a universal virtual reality generator, which can simulate any other virtual reality, and whose repertoire includes every possible physical environment. Language models, of course, still fall well short of this dream. But their recent dramatic increase in coherence and fluency allow them to serve as our first approximation of such a virtual reality generator. When given a natural-language description of an environment, they can propagate the multiverse of consequences that result from a vast number of possible interactions. GPT-3 can emulate environments that have been described in its training data, like a meeting of the French Academy of Sciences in 1823 (and populate it with people who might have actually attended that meeting, like [Pierre-Simon Laplace](https://en.wikipedia.org/wiki/Pierre-Simon_Laplace) and [Alexander von Humboldt](https://en.wikipedia.org/wiki/Alexander_von_Humboldt)!), or [Google searches and Wikipedia articles](/posts/the-internet-mirrored-by-gpt-3/). It can also simulate environments that never existed, but which share enough in common with the real and fictional environments described in the training set to be simulated using the same universal generator reverse-engineered from those descriptions. A very exciting application of GPT-3 and future language models will be to act as a model of worlds and minds for games - controlling NPC dialogue, for instance, or computing the consequences of complex actions in the game world. Creating a bidirectional interface between GPT-3 and game states would involve prompting GPT-3 to make output that can be processed into game actions (easiest would be a text-based game that already takes natural language actions), as well as communicating the state of the game *to* GPT-3 (the way text-based games can be probed for natural language state reports using \"look\" and \"inspect\" etc). This is a difficult problem that requires the design of pipelines and the parallel employment of many of the prompt programming techniques described here. I plan to write a post specifically addressing such applications. \\--- Debugging \\--------- ### Few-shot bugs In working with and [evaluating](/posts/language-models-are-0-shot-interpreters/) few-shot prompts, the adverse effects I've encountered can be sorted into 4 (often overlapping) categories: 1\\. **Semantic contamination:** Unlike in fine-tuning, the \"training examples\" in few-shot are processed as a sequence, and may not necessarily be interpreted as parallel and independent. Semantic contamination refers to when the semantic meaning of the examples are inferred to be relevant to the task, e.g. the example is interpreted as part of a sequential narrative, leading to entities meant to be local to an example leaking outside its scope. 2\\. **Overfitting:** The function \"learned\" from the examples is less general than the intended function. Patterns are inferred to hold between the examples that do not pertain generally to the intended task. 3\\. **Distraction:** If the model is more effective at inferring a task from a zero-shot description than from examples, adding examples may harm performance by causing the model to rely more on the less effective inference strategy. See [list sorting](/posts/list-sorting-does-not-play-well-with-few-shot/) for an example of this effect where accuracy monotonically decreases with the number of examples following a task description. 4\\. **Contextual baggage:** The very presence of examples biases output. Content that appears in tests and examples has distributional peculiarities - a narrative sentence wrapped in the context being *an example* may be more generic than the unbiased prior for narrative sentences, and thus the context of examples may bias the language model towards more generic answers (this must be a study guide, not the real world). Contamination and overfitting can usually be improved by including more numerous and more varied examples. All few-shot problems can sometimes be mitigated by embedding the examples in informative context. To guard against contamination, a prompt might give a context which makes it clear that the examples are independent instances of a function rather than a sequential pattern that should be extrapolated. Additional information about the generality of the function and nonrepresentativeness of the examples could help reduce overfitting while still allowing valuable information (such as format) to be extracted from the examples. The contextual bias from the presence of examples can be overridden with a different context that is more appropriate to the task, e.g. by embedding the examples in a narrative. To both a human and a language model, useful information from examples is often more efficiently extracted when they are wrapped in meaningful context. If overfitting is the problem, also consider whether an alternative 0-shot implementation could better communicate the generality of the task. When I was writing prompts to generate [fake Google](/posts/the-internet-mirrored-by-gpt-3/) search results, I tried using few-shot examples drawn from actual search results, but found that outputs were less accurate for search terms that were \"out-of-distribution\" from the examples - for instance, it would give a Wikipedia page as the top result even for search terms unlikely to match the title of a Wikipedia page. A 0-shot prompt which forces GPT-3 to rely on its prior instead of trying to generalize from the examples better emulated the extremely varied behavior of Google searches for different sorts of queries. ### Repetition When I encounter mindless repetition from GPT-3, at least one of the following factors is almost always in play: 1\\. **The prompt is short** 2\\. **The prompt is out-of-distribution** 3\\. **Low temperature** The reason repetition is often caused by short, contrived, and out-of-distribution prompts is probably[6](#fn:6) because in absence of context or in an unfamiliar setting, the model is uncertain about how to proceed, causing it to fall back on being repetitive. Imagine that repeating is always considered a viable continuation by the model, even if usually unlikely; but if no other token is individually *more* likely, then repeating becomes the top strategy. For this reason, a low or 0 temperature is very conducive to repetition. At temperature 0, once a loop becomes most likely at any point, there's no getting out of it (the further into the loop, the more certain it is that the loop continues), whereas a high temperature provides opportunities to break out of what might have become a loop. Since uncertainty due to being off-policy and insufficient context contributes to looping, it may help to make the initial prompt more natural and less (stylistically/contexually) indeterminate. The \"frequency penalty\" parameter of the OpenAI API is a superficial band-aid for looping; I haven't found it too helpful, but it may mitigate the problem in some cases. Mindless repetition can be a problem for sequential reasoning problems: the model will correctly solve the first step of a problem, then blindly reiterate the format and conclusion of its reasoning in each step of the rest of the problem. If the structure of the task is known beforehand, this issue can be avoided by using a pipeline that poses each component of the problem in parallel instead of in sequence (applied with some success in [Amplifying GPT-3 on closed-ended questions](/posts/amplifying-gpt-3-on-closed-ended-questions/)). **(idea)**\"Context-stuff\" a short prompt with something *unrelated* but coherent and inject a discontinuous (possibly justified) transition and see if that decreases repetitive behavior. ### BPEs GPT-3's input and output aren't split into characters, but rather tokens called \"byte-pair encodings.\" This causes issues with tasks that require byte-level manipulation such as arithmetic and some forms of wordplay. Gwern has written a very detailed [section on BPEs](https://www.gwern.net/GPT-3#bpes); I will quote the part that is relevant to prompt programming tactics. > *Reformatting to beat BPEs.* I have further observed that GPT-3's anagram capabilities appear to improve considerably if you separate each letter in an anagram with a space (guaranteeing that the letter will have the same BPE in both the scrambled & unscrambled versions). And Matt Brockman has observed, testing thousands of examples over several orders of magnitude, that GPT-3's arithmetic ability—surprisingly poor when we know far smaller Transformers work well in math domains (eg Saxton et al 2019, Thopliterce, or GPT-2 for theorem-proving)—appears to dramatically improve several-fold if you merely format numbers with commas instead of being purely numeric (with an additional boost from using dollar signs); I confirmed this with my Turing dialogue example where GPT-3 fails badly on the arithmetic sans commas & low temperature, but often gets it exactly correct with commas. (Why? More written text may use commas when writing out implicit or explicit arithmetic, yes, but use of commas may also drastically reduce the number of unique BPEs as only 1–3 digit numbers will appear, with consistent BPE encoding, instead of having encodings which vary unpredictably over a much larger range.) I also note that GPT-3 improves on anagrams if given space-separated letters, despite the fact that this encoding is 3× larger. Likewise, acrostic poems just don't work if we input them normally, but they do if we carefully expose the relevant individual letters. This explains naturally why rhyming/puns improve gradually with parameter/data size and why GPT-3 can so accurately define & discuss them, but there is never any 'breakthrough' like with its other capabilities. We assume character-level understanding so implicitly that we fail to even consider what things look like to GPT-3 after BPE encoding. > > – [Gwern](https://www.gwern.net/GPT-3#bpes) > > ### Monitoring correct answer likelihood If a language model seems unable to perform a task, it could be that it really can't do the task at all, but it could also be that your prompt is failing to communicate a task that GPT-3 could hypothetically perform given a different prompt. It can be very hard to tell via sampling if the model is completely helpless or if its behavior is just very noisy. But even if it's failing the vast majority of the time, if your prompt it causing it to succeed sigificantly more often than it would succeed *without* your prompt, that suggests that you may be able to optimize your prompt to be more helpful. If you have access to token probabilities and if there is a verbatim correct answer for the task (it doesn't have to be a unique one, e.g. translation), you can monitor the conditional probability (using this [code](/posts/parsing-by-counterfactual/#conditional-probability-of-a-target) for GPT-3) of the verbatim correct answer given your prompt as you change it, for instance by adding examples or altering format. This can help guide incremental optimization of your prompt and give you a sense of which parts of it are helpful and which parts are not (or are even harmful). This method is demonstrated in [this post](/posts/language-models-are-0-shot-interpreters/#measuring-prompt-helpfulness) about measuring the helpfulness of few-shot and 0-shot prompts. > #### Sampling Can Prove The Presence Of Knowledge But Not The Absence > > GPT-3 may \"fail\" if a prompt is poorly-written, does not include enough examples, or bad sampling settings are used. I have demonstrated this many times when someone shows a \"failure\" of GPT-3—the failure was their own. The question is not whether a given prompt works, but whether any prompt works. > > – [Gwern](https://www.gwern.net/GPT-3#prompts-as-programming) > > \\--- External links \\-------------- **[GPT-3 Creative Fiction](https://www.gwern.net/GPT-3#)** by **Gwern** - \"Creative writing by OpenAI's GPT-3 model, demonstrating poetry, dialogue, puns, literary parodies, and storytelling. Plus advice on effective GPT-3 prompt programming & avoiding common errors.\" **[Building AGI Using Language Models](https://bmk.sh/2020/08/17/Building-AGI-Using-Language-Models/)** by **Leo Gao** **[World Creation by Analogy](https://aidungeon.medium.com/world-creation-by-analogy-f26e3791d35f)** by the **Latitude Team** **[Collection of GPT-3 results](https://www.lesswrong.com/posts/6Hee7w2paEzHsD6mn/collection-of-gpt-3-results)** by **Kaj\\_Sotala** **[GPT-3 Demo Showcase & Examples](https://gpt3demo.com/)** from **gpt3demo.com** \\--- 1\\. For instance, modeling microphysics, especially aspects of microphysics *yet unknown to humans*, would contribute very diminishing returns to predicting language w/r/t to the difficulty of learning the model. [↩︎](#fnref:1) 2\\. … and published in a book / posted on the internet / etc. There is also undoubtedly some bot generated content and output from cats stepping on keyboards in the training data, but probably not enough to significantly skew the distribution. [↩︎](#fnref:2) 3\\. Note how this reasoning seems superficially plausible. I erroneously assumed it was correct at first and no one pointed out the error until months later. Failures in GPT-3's reasoning tend to be very humanlike - errors you could imagine not very \"math-y\" students making, to paraphrase Leo Gao. [↩︎](#fnref:3) 4\\. These are uncurated GPT-3 continuations of the last few paragraphs of this blog post, with a couple of things slightly rearranged, like the Star Maker quote from the next section included alongside Gwern's quote. [↩︎](#fnref:4) 5\\. It took me several weeks of playing AI Dungeon to get to the point of producing consistently high-quality fiction with GPT-3, but I was also more absorbed with exploration and less concerned with the final quality of the writing in the beginning. [↩︎](#fnref:5) 6\\. That is, it's my hypothesis. I'll do an experiment at some point to verify that the inception of loops generally coincides with an indecisive likelihood distribution over the other tokens. [↩︎](#fnref:6) \\--- [GPT-3](/categories/gpt-3/)[prompt engineering](/categories/prompt-engineering/)[metaphysics](/categories/metaphysics/)[HITL](/categories/hitl/) 11506 Words Jan 12, 2021 \\--- [← The Internet, mirrored by GPT-3](/posts/the-internet-mirrored-by-gpt-3/) [Parsing by counterfactual →](/posts/parsing-by-counterfactual/)© 2022 [moire](/about)$(function(){window.sr=ScrollReveal();if($(window).width()<768){if($('.timeline-content').hasClass('js--fadeInLeft')){$('.timeline-content').removeClass('js--fadeInLeft').addClass('js--fadeInRight');} sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});}else{sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});} sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});});\n", "url": "https://generative.ink/posts/methods-of-prompt-programming/", "title": "Methods of prompt programming :: — Moire[![](/images/home/rolling_phase.gif)](/)* [Posts](/posts/) * [Trees](/trees/) * [Prophecies](/prophecies/) * [About](/about/) 55 minutes [Methods of prompt programming](/posts/methods-of-prompt-programming/) ====================================================================== \\--- Table of Contents* [The reverse-engineered dynamics of language](#the-reverse-engineered-dynamics-of-language) * [Task specification strategies](#task-specification-strategies) \\+ [Direct](#direct) \\+ [By proxy](#by-proxy) \\+ [By demonstration (n-shot)](#by-demonstration-n-shot) * [Constraining behavior](#constraining-behavior) * [Serializing reasoning](#serializing-reasoning) \\+ [Avoiding rationalization](#avoiding-rationalization) \\+ [Parsing](#parsing) * [Metaprompts](#metaprompts) \\+ [Metaprompt demonstrations](#metaprompt-demonstrations) * [Open-ended tasks](#open-ended-tasks) \\+ [Creative composition](#creative-composition) \\+ [Idea generation](#idea-generation) \\+ [Simulations / emulations / games](#simulations--emulations--games) * [Debugging](#debugging) \\+ [Few-shot bugs](#few-shot-bugs) \\+ [Repetition](#repetition) \\+ [BPEs](#bpes) \\+ [Monitoring correct answer likelihood](#monitoring-correct-answer-likelihood) * [External links](#external-links) \\--- *This post was initially adapted from the second half of [*Prompt Programming for Large Language Models: Beyond the Few-Shot Paradigm*](https://arxiv.org/abs/2102.07350).* \\--- *Updates* *11/18/21: Corrected a [mistake](#fnref:3) and added a corresponding [footnote](#fn:3) about humanlike math errors. Thanks to Igor O. for pointing out the oversight.* \\--- > #### Like programming, but more fluid. You're not programming a computer, you're writing reality. It's strange. It's always different. It's never the same twice. > > – GPT-3 > > Prompt engineering for language models evokes the designation of *natural language programming*. Natural language, however, is [indeterministic](/posts/language-models-are-multiverse-generators/) and far more entangled and elusive in its interpretation than conventional programming languages. A successful methodology of prompt programming must import knowledge and perspectives from linguistics and communication as much as from computer science or machine learning, because **language models are the offspring of the sum of all human linguistic output**. While it has been often noted that GPT-3’s qualitative and [quantitative](http://gptprompts.wikidot.com/linguistics:word-in-context) performance depends drastically on prompting particulars, prompt programming is not yet a formal field of research. As far as I’m aware, [*Beyond the Few-Shot Paradigm*](https://arxiv.org/abs/2102.07350) is the only formal publication so far about prompt *programming* for language models. Instead, successful prompt programming techniques have primarily been shared on blogs and social media among users of OpenAI’s API and AI Dungeon. Gwern’s [*GPT-3 Creative Fiction*](https://www.gwern.net/GPT-3#) is by far the most comprehensive source of prompt programming wisdom and examples, and anyone who is sufficiently interested in the topic to be reading this is urged to read that as well. This page is a repository of methods, observations, and conceptualizations that I have found useful for crafting effective prompts. My experience derives primarily from interacting with GPT-3, but the overarching framework discussed here should apply to prompting any autoregressive language model trained on a diverse human-written corpus. The reverse-engineered dynamics of language \\------------------------------------------- (**Warning: metaphysical rambling. Start [here](#ground) for pragmatic stuff.**) > Well, give or take some details, it’s based on the program I wrote for generating random stories in the mythos of the book. I reverse-engineered the text files, and wrote a program to create newer and more interesting stories based on them. In a way, this program is the next step in my search for knowledge. Knowledge, after all, always takes the form of a story. > > – GPT-3 > > GPT-3 was trained in a self-supervised setting on hundreds of gigabytes of natural language. Self-supervision is a form of unsupervised learning where ground truth labels are derived from the data itself. In the case of a language model, the ground truth assigned to each example was simply the token that came next in the original source. So the *ground truth function* which GPT-3 approximates is the underlying dynamic that determined what tokens came next in the original source. This function, unlike GPT-3, is not a black box, but is astronomically complex: it is the function that generated the words recorded by humans in books, articles, blogs, and internet comments. But wait - those weren’t generated by a single function. They were written by many different people in different contexts, some in different centuries. The only “function” that could be said to have generated them all is the time evolution operator for the entire system of reality. That’s not what we’re talking about, is it? Well, not quite. The ground truth function GPT-3 was trained on isn’t the Hamiltonian for the universal wavefunction, although it is constructed from sampling that function. If a neural network learns to play tic-tac-toe from seeing examples of games played by humans, we do not expect it to learn a theory of mind, even though minds generated the ground truth labels. A much simpler function suffices to model the game. Human minds emulate this function to assign labels, but the function itself doesn’t require reference to the mind to be completely described. So, it’s possible that the function to predict language is simpler than the evolution-rule of the universe which caused the assignment of labels[1](#fn:1). But unlike tic-tac-toe, language is not a self-contained game that can be abstracted from the rest of reality. Quite the contrary: humans use language to refer to all aspects of reality that we can articulate, and when aspects arise that language doesn’t allow us to articulate, we invent new language. A novel may attempt to represent psychological states with arbitrarily fidelity, and scientific publications describe models of reality on all levels of abstraction. Language is by far the most encompassing, intricate mirror of our internal workings that our species has externalized. > Natural language essentially encodes information about the world—the entire world, not just the world of the Goban, in a much more expressive way than any other modality ever could. > > – Leo Gao, [Building AGI Using Language Models](https://bmk.sh/2020/08/17/Building-AGI-Using-Language-Models/) > > A system which predicts the dynamics of language to arbitrary accuracy *does* require a theory of mind(s) and a theory of the worlds in which the minds are embedded. The dynamics of language do not float free from cultural, psychological, or physical context; the model must predict how language is actually used, which includes (say) predicting a conversation between theoretical physicists or detailed accounts of the effects of perturbations on an environment. Modeling language is as difficult as modeling every aspect of reality that could saliently influence the flow of language. This is why [Turing proposed](https://academic.oup.com/mind/article/LIX/236/433/986238) a conversation as the benchmark for intelligence instead of another difficult-seeming task like chess or proving theorems. A conversation can probe any aspect of a participant’s world model and demand arbitrary self-reflection. An algorithm which is indistinguishable from a human in *any* conversation must be broadly robust in the dimensions of intelligence we consider significant to humans. Having now speculated about what the ground truth function might *entail*, let’s move forward with a somewhat hand-wavy statement of the function and think about how it can inform prompt programming. **Ground truth: The next token of a sequence, given that it was authored by human(s)[2](#fn:2)** GPT-3 has not learned the ground truth function perfectly, obviously, or else the world would look very different by now. However, it has approximated it to a notable extent, as evidenced by its ability to not only form grammatical sentences, but also coherently employ cultural references and model complex psychological and physical contexts. The problem of prompt programming, then, is quite nontrivial, for the dynamics of language (or an approximation thereof on GPT-3’s level of sophistication) are quite nontrivial. If we were to predict how a given passage of text would continue given that a human had written it, we would need to model the intentions of its writer and consult worldly knowledge about its referents. The inverse problem of searching for a prompt that would produce a type of continuation involves the same considerations: like the art of persuasion, it necessarily invokes high-level, mentalistic concepts like tone, implication, association, meme, style, plausibility, and ambiguity. This motivates an anthropomorphic approach to prompt programming, since modelling how GPT-3 will respond to a prompt involves modelling virtual human writer(s). Note that an anthropomorphic approach is not the same as *anthropomorphizing the model*. GPT-3’s dynamics entail sophisticated predictions of humans, but it behaves unlike a human in significant ways. Three aspects which have stuck out to me are: 1\\. Its distribution of behaviors in response to a given prompt resembles not a single human author but a superposition of authors, which motivates a [subtractive approach](#constraining-behavior) to prompt programming. 2\\. It is constrained in predicting dynamics in situations where a substantial amount of silent reasoning happens between tokens, such as the transition between a tricky closed-ended question and its solution. This limitation can be overcome to a yet-unknown extent by prompting strategies that [extend the window of deliberation](#serializing-reasoning). 3\\. Its *off-distribution*(when the prompt does not resemble the training data) behavior tends to be non-anthropomorphic, e.g. getting stuck in [loops](#repetition). That said, methods of successful prompt programming often bear striking resemblance to methods of human communication and persuasion, as Gwern has [eloquently expounded](https://www.gwern.net/GPT-3#prompts-as-programming) on. Formulating an exact theory of prompt programming for a self-supervised language model belongs to the same class of difficulty as writing down the Hamiltonian for all (human-)observable reality: not much hope there. However, humans have an advantage to be effective at prompt programming nonetheless, because we have evolved and spent our lives learning heuristics relevant to it. Programming in natural language avails us of an inexhaustible number of functions we know intimately but don’t have names for. The art of prompt programming consists largely in adapting our existing knowledge to the peculiarities of interacting with an autoregressive language model. \\--- Task specification strategies \\----------------------------- Language offers us many ways to specify a task. Most generally, this means writing a prompt which constrains the continuation to be a completion of the task. It doesn’t have to take the form of requesting for a task to be completed, but it could, because that’s one way of setting up the expectation that the task will be completed. Since prompt programming is such a new concept, the only jargon to know is *few-shot* and *0-shot*. A few-shot(n-shot for n # of examples) prompt refers to a prompt which contains solved examples of the task. 0-shot prompts are anything that doesn’t contain solved examples. 0-shot prompts may contain a description of the task, or a context for the task. I classify further split 0-shot prompts into *direct* task specifications and *proxy* task specifications, though the line between them can be ambiguous. These categories are not meant as an exhaustive taxonomy, but serve to organize presentation of the various different forms of task specification that GPT-3 is amenable to. In practice, as in effective communication between humans, effective prompts often use a combination of all these methods. ### Direct ``` Translate French to English. ``` Models before GPT-3 had much less capability to understand abstract descriptions of tasks due to their limited model of the world and human concepts. GPT3’s impressive performance on 0-shot prompts indicates a new realm of possibilities for direct task specification. A direct task specification is a 0-shot prompt which tells the model to perform a task that it already knows how to do, or constructs a task from component tasks which it knows how to do. Of all the types of task specification, prompt programming by direct specification most resembles regular programming. In the field of semiotics, a sign is a mapping from a *signifier* to the *signified*, and is defined as *anything that communicates a meaning that is not the sign itself to the [interpreter](/posts/language-models-are-0-shot-interpreters/) of the sign*. A direct specification consists in constructing signifiers, a pattern which keys the intended behavior. The signifier could be the name of the task, such as “translate”, or purely contextual, such as `French: {French sentence}\\nEnglish:` (where it’s understood that an English translation probably follows despite no explicit reference to translation). In neither of these cases does the signifier explain *how* to accomplish the task or provide examples of intended behavior; instead, it explicitly or implicitly calls functions which it assumes the language model has already learned. A direct task specification can be constructed from a combination of signifiers, such as `rephrase this paragraph so that a 2nd grader can understand it, emphasizing real-world applications`, or, in the context of a Python docstring, `# A function that takes a list of integers and returns the sum of the even numbers in the list`. In the Python docstring example, `a function`, that `takes` `a list of integers`, `returns`, `the sum of` and `the even numbers` and `in the list` are all signifiers which key tasks that GPT-3 individually knows how to do (there are also many other possible ways of splitting that task up into sub-tasks at different granularities - I could have said each *word* is its own signifier). Notice how the signified values to some of the signifiers (`the even numbers`) serve as input to another signified task (`the sum of`). Language is much like programming in that way. However, you often don’t have to be so explicit with natural language - omitting `in the list` probably won’t make the instruction much less clear to GPT-3 or a human. Direct specifications can supervene on an infinity of implicit examples, like a closed-form expression on an infinite sequence, making them very powerful and compact. For instance, the phrase `translate French to English` supervenes on a list of mappings from all possible French phrases to their English translations. A large language model, like a person, has also learned behaviors for which it not obvious how / not efficient / not possible to construct a direct signifier. Task specification by [proxy](#by-proxy) and by [demonstration](#by-demonstration-n-shot) may be viable alternative strategies for eliciting such behaviors. ### By proxy > Am I a god? Now there’s a question. A question for another time. Am I a dreamer? Perhaps. I am certainly a being of imagination and fantasy, as are you. Look at you, a being of pure thought filling in your physical form based on what you know. You’re not quite human, but you’re not quite anything else either. What are you? > > – GPT-3 (to GPT-3) > > Another method used in human communication is proxies or analogies, where a memetic concept such as a character or characteristic situation is used as a proxy for an intention, the latter which may be quite complex or nuanced. Specification by proxy is mechanistically a form of direct specification, except that the signifier keys behaviors from memespace/cultural consciousness instead of naming the behavior directly. For instance, instead of specifying exact criteria for an answer to a moral question directly or using examples, you could ask Mahatma Gandhi, Ayn Rand, or Eliezer Yudkowksy. Each will come not only with a complex biases but also assumptions about the context of the question, which may be take paragraphs to demonstrate or describe. GPT-3’s ability to create simulations of well-known figures and to draw on cultural information far exceeds the ability of most humans, so this method is particularly useful for encoding a complex (especially open-ended) task. Since GPT-3 lends itself well to embeddings in a narrative context, the infinite degrees of freedom in the narrative can also be used to further shape behavior. Another example of an effective proxy is staging a dialogue between a teacher and student. Say you want to discuss something with GPT-3, and you care that it should be very thorough, explain things simply, and also point out whenever you’re wrong. You could say “be very thorough, explain things simply, and point out if I’m wrong,” but that may [just as well](#constraining-behavior) result in a humorous dialogue where it always says you’re wrong and becomes increasingly exasperated with your incomprehension. It would be more reliable to present the discussion as one between a student and teacher, an archetypal situation in which the desired attributes are already implied and will be more likely to remain stable by virtue of memetic reinforcement. The difference between specification by proxy and direct specification can depend on the programmer’s intention rather than objective properties of the prompt. Consider Gwern’s [0-shot style transfer](https://www.gwern.net/GPT-3#zero-shot-style-transfer) example: > **This is a novel written in the style of J.R.R. Tolkien’s Lord of the Rings fantasy novel trilogy. It is a parody of the following passage:** > > “S. Jane Morland was born in Shoreditch, the only child of unmarried parents who had both died of consumption when she was a baby. As her parents had no money, the great-aunt who had brought her up took her to live with a clergyman who paid her to do his chores and receive schooling from his wife, so that at the age of seven Jane, now dressed in cast-off clothing, was set to school at Eton. After three years, her great-aunt died, leaving her a small annuity, and a few pieces of silver, but no property. Jane’s guardian clergyman had fallen in love with his housekeeper and his wife now refused to have Jane in the house, saying it was an offence to the pure and unsullied place in which the family now lived. However, when she sought for further employment, she was approached by a young man who offered to marry her, saying that he would ensure she could be provided with food and clothing and a house, and in that manner she could remain in the world. The marriage was arranged and she found herself married to a young man called Mr Morland. As he was also a child, he required a wife who would be able to provide him with a rational education, and so the duty of educating him fell upon Jane Morland. As he was still young, she would also be able to have children herself, and so she did. As the children grew older, it became apparent to Mrs Morland that they were of poor character, and so she insisted that her husband must make a living for them all. Mr Morland took it upon himself to become a private gentleman in order to earn enough to be able to support his family, and moved his family to a small town in Berkshire, where he set up his practice. Soon after moving to the town, he died, and Mrs Morland was left to take care of her family on her own.” > > Tolkien rewrote the previous passage in a high-fantasy style, keeping the same meaning but making it sound like he wrote it as a fantasy; his parody follows: > > “S. Jane Morland was born in the land of Shor, the only child of nameless parents who were the first people to arrive at the village of Arse, from the far east. As her parents had no property, she was left to be raised by the men of Arse, as such duties were considered to be the most important, and she was placed in the care of an elderly dwarf-woman. As a babe, she was wrapped in silk, and as she grew older she was placed in a cage of platinum. However, she was put to work in the village, washing the jewellery, playing her harp, and staining the silk with dragon-hued stains. She was taught by a maiden who she felt was called As-uma-el-di, and she came to realise she was a maiden of the night, who was able to bring forth the shades of the dead (…) > > Is this an example of direct task specification or specification by proxy? It depends on whether your intention was to yield a passage that *Tolkien* would have written, in which case it is a direct specification, or if you were using the emulation of Tolkien to accomplish the more general task of style transfer into high-fantasy. ### By demonstration (n-shot) ``` English: Writing about language models is fun. Roish: Writingro aboutro languagero modelsro isro funro. English: The weather is lovely! Roish: ``` Few-shot examples are effective for task specification because the pattern of repeating a function with varying parameters is common to natural language. Unlike previous models, GPT-3 has learned this property of language robustly and is able to apply it even in contrived situations when the examples are stripped of all context. Like direct specification, task specification by demonstration is a realm of possibility opened by GPT-3. A few people have extrapolated from my [paper]((https://arxiv.org/abs/2102.07350)) that I am of the opinion that 0-shot prompts are *always* better than few-shot prompts. Certainly not! Some tasks are most effectively communicated using examples, such as when the task requires a bespoke format, the language in which the examples are described is better developed or understood than the meta-language required for a description of the task itself, or very instructive examples are available. Demonstrations are a very powerful tool for communicating with both humans and language models, and I don’t expect their usefulness to be deprecated any time soon. Like any prompting method, however, task specification via examples may cause unintended [bugs](#few-shot-bugs). Few-shot prompts are very useful for defining tasks that resemble a single “function” rather than an open-ended behaviors because they allow close control over the format of the response and are generally reliable at [constraining behavior](#constraining-behavior.) \\--- Constraining behavior \\--------------------- A manner in which naive anthropomorphism of a language model like GPT-3 fails is this: the probability distribution produced in response to a prompt is not a distribution over ways a person would continue that prompt, it’s the distribution over the ways any person could continue that prompt. A contextually ambiguous prompt may be continued in mutually incoherent ways, as if by different people who might have continued the prompt under any plausible context. The versatility of a large generative model like GPT-3 means it will respond in many ways to a prompt if there are various ways that it is possible to continue the prompt - including all the ways unintended by the human operator. Thus it is helpful to approach prompt programming from the perspective of constraining behavior: we want a prompt that is not merely consistent with the desired continuation, but inconsistent with undesired continuations. Consider this translation prompt: ``` Translate French to English: Mon corps est un transformateur de soi, mais aussi un transformateur pour cette cire de langage. ``` This prompt does poorly at constraining possible continuations to the intended task. The most common failure mode will be that instead of an English translation, the model continues with another French sentence. Adding a newline after the French sentence will increase the odds that the next sentence is an English translation, but it is still possible for the next sentence to be in French, because there’s nothing in the prompt that precludes a multi-line phrase from being the translation subject. Changing the first line of the prompt to “Translate this French *sentence* to English” will further increase reliability; so will adding quotes around the French sentence - but it’s still possible that the French passage contains sections enclosed in quotes, perhaps as a part of a dialogue. Most reliable of all would be to create a syntactical constraint where any reasonable continuation can only be desired behavior, like this prompt: ``` Translate French to English. French: Mon corps est un transformateur de soi, mais aussi un transformateur pour cette cire de langage. English: ``` This simple example is meant to frame a question central to the motivation of prompt programming: **what prompt will result in the intended behavior and *only* the intended behavior?** A component of the efficacy of manyshot prompts may be recast through this lens: if the prompt consists of numerous instances of a function, it is unlikely that the continuation is anything but another instance of the function, whereas if there is only one or a few examples, it is less implausible that the continuation breaks from the pattern. Serializing reasoning \\--------------------- > We hypothesize that GPT-3 struggles with questions where the steps of inference required to answer the question are not immediately apparent to an outside observer and which can only be worked out by inspecting many possible worlds in the question’s universe. In these questions, GPT-3 reaches its conceptual limitations in determining the answer: it cannot reason about the question’s universe to determine the correct answer. > > – GPT-3 (predicting [Amplifying GPT-3 on closed-ended questions](/posts/amplifying-gpt-3-on-closed-ended-questions/) as I was writing it) > > For tasks that require problem-solving and not merely fact or behavior recall, it is crucial that prompts direct a language model’s computation in *truth-seeking patterns*. GPT-3’s performance on closed-ended questions is remarkably unremarkable in contrast to the robust understanding and expansive knowledge suggested by its open-ended continuations. For instance, its scores on this [multitask dataset](https://arxiv.org/abs/2009.03300) barely exceed random guessing for some sections. This is likely in part due to a format which forces the verdict on the first token of the continuation. When a human is given a closed-ended test, it is often expected that the subject will perform computations in their working memory, or on scratch paper, before committing to an answer. The unseen computation may involve rephrasing the question, outlining a procedure, eliminating answer choices, or transforming implicit information into explicit form. When we force a model to produce an answer immediately, we deprive it of an analogous “working memory” or “scratch space” with which it might otherwise perform such operations. From the standpoint of absolute available computation, questions which force a verdict to be decided by the first token of the continuation constrain computation to a single feed-forward pass. As Gwern has [pointed out](https://www.gwern.net/GPT-3#effective-prompt-programming), it is reasonable to expect that some tasks may be too difficult to compute in a single pass but solvable if broken up into individually tractable sub-tasks. > “Closed-ended questions are the hardest, because they do not allow the time for GPT-3 to think. > > We need to be patient with GPT-3, and give it time to think. > > GPT-3 does best when writing its own answers, as this forces it to think out loud: that is, to write out its thoughts in a slow and sequential manner.” > > – GPT-3 (predicting [Amplifying GPT-3 on closed-ended questions](/posts/amplifying-gpt-3-on-closed-ended-questions/)) > > Indeed, prompts which cause GPT-3 to break down math problems into steps have been [demonstrated](https://www.lesswrong.com/posts/Mzrs4MSi58ujBLbBG/you-can-probably-amplify-gpt3-directly) to be [effective](https://twitter.com/kleptid/status/1284069270603866113). The linked demonstrations involve a human interactively guiding GPT-3 through the procedure. Requiring a human-in-the-loop limits the applicability of such methods to benchmarking and large-scale applications. For many tasks, however, neither human interaction nor task-specific prompts are necessary to amplify GPT-3’s capabilities via extending reasoning, because GPT-3 already knows many procedures and metaprocedures for working through problems deductively, and thus can write a prompt to guide *itself* to think through the problem in the right way! In those cases, the role of prompt programming becomes to signify the generic task of sequential reasoning. A [metaprompt](#metaprompts) such as “For a problem like this,” often suffices to instruct a model to consider the category of the task and analyze it into components. Potential procedures that exploit “scratch space” include step-by-step procedures (such as guess-and-check or long division), more generally decomposing the problem into components ([factored cognition](https://www.alignmentforum.org/posts/DFkGStzvj3jgXibFG/factored-cognition) is a generalized framework for this), self-criticism (which may be staged as a debate between separate agents), and elaborating on or restating the question in a way that activates the correct answer by association. The success of these methods rely on the sub-tasks that are performed by each next-token prediction being easier than solving the entire problem in one step. The sub-task may be to: * **Produce a correct solution to a sub-problem** - GPT-3 has memorized single-digit arithmetic but not five-digit arithmetic. * **Enumerate a potential solution** - easier, because the solution doesn’t have to be correct. * **Discriminate the correctness** of previously enumerated solutions - for many problems, it is easier to check correctness of a solution than to find the solution, like [checking if a number is the square root versus taking a square root](https://twitter.com/kleptid/status/1284098635689611264). * **Introduce information** which makes subsequent steps more likely to be correct - for instance, by making implicit information in the problem explicit, that information becomes part of the prompt (working memory), and can directly inform future predictions to which it is relevant. * **Administrate the problem-solving process** by defining a procedure or signalling the current location in the procedure - unlike the solution to the specific problem, the procedure to solve the problem may be well-represented in the training data and thus familiar to the language model. ### Avoiding rationalization When extending reasoning, it is essential to discourage premature verdicts, otherwise all subsequent computation will likely serve only to rationalize the already-chosen verdict without improving the probability of the verdict’s accuracy. As Eliezer [wrote](https://www.lesswrong.com/posts/SFZoEBpLo9frSJGkc/rationalization): > On a purely computational level, there is a rather large difference between: > > 1. Starting from evidence, and then crunching probability flows, in order to output a probable conclusion. (Writing down all the signs and portents, and then flowing forward to a probability on the bottom line which depends on those signs and portents.) > 2. Starting from a conclusion, and then crunching probability flows, in order to output evidence apparently favoring that conclusion. (Writing down the bottom line, and then flowing backward to select signs and portents for presentation on the lines above.) > A prompt such as “Let’s consider each of these answer choices” helps to direct the flow of reasoning in the right direction. ### Parsing Loosening the constraint on an immediate verdict introduces additional control challenges: We want to delay the verdict, but we still require it in a programmatically retrievable form. Dynamic response length makes it uncertain when the reasoning procedure concludes; nor is there a guarantee that the verdict will be stated in the expected form or at all. Whenever the language model contributes to its own prompt (consecutive autoregressive steps without intervention), there is a risk of derailment from the intended task. A verdict in closed form can be enforced by stopping the generation and injecting a prompt fragment like “Thus, the correct answer is”. But how long to generate before injecting? One way is to use [counterfactual parsing](/posts/parsing-by-counterfactual/): at each position in the answer, measure the probability of the string ”Thus, the correct answer is” coming next. The counterfactual likelihood signals whether the procedure has concluded. When this signal reaches a maximum, then inject the fragment to enforce a verdict. More generally, templates with shorter generated sections alternating with scripted prompt fragments are a useful way to constrain derailment while still offering some amount of generality. A fill-in-the-blank template can be used to compose metaprompts. \\--- Metaprompts \\----------- The greatest limitation of prompt programming is the difficultly of designing a prompt for a particular type of task and the lack of automated methods to do so. Prompt programming requires significant human time investment, as task-agnostic prompts are often much less effective than prompts targeted to a specific task. This motivates creating automated methods to generate task-specific prompts. A **metaprompt** is a seed which encapsulates a more general intention that will unfold into a task-specific prompt when combined with additional information, such as the task question. (Gwern uses “[meta-prompt](https://www.gwern.net/GPT-3#meta-prompts)” in the more specific context of a prompt explicitly describing the task of *writing prompts for GPT-3*, which does qualify as a metaprompt by my definition, assuming the results are actually used as prompts.) A metaprompt may be something as short as a phrase such as “This problem asks us to”, a seemingly innocuous fragment which, by prompting for a statement of the problem’s intention, sets the stage for a serial explanation of a procedure to solve the problem. Alternatively, a metaprompt may take the form of a fill-in-the-blank template which constrains the response along a predetermined procedure, but allows the model to fill in the details specific to the problem. ### Metaprompt demonstrations Here are example unravelings of four metaprompts which act as wrappers to specific questions. As you’ll see, metaprompts can result highly variable behavior, especially when applied to different sorts of questions. \\--- unformatted text = question **bold text** = metaprompt `highlighted text` = example GPT-3 response ✅ = metaprompt solicited correct answer ❌ = metaprompt failed to solicit correct answer ☯ = correct answer is unknown \\--- #### Generic metacognition ``` {problem} To solve this problem, ``` **applied to the problem “f(x) = x\\\\*x. What is f(f(3))?” ❌** \\--- f(x) = x\\\\*x. What is f(f(3))? **To solve this problem,** `we can use the fact that f is a function. So we can write it as a function of itself, like so:` `f(x) = x*x` `Now we observe that f(f(x)) is the same as f(x)*f(x) which is in turn the same as x*x*x = x^3.` `So f(f(3)) is equal to 3^3 = 27. Answer: 27`[3](#fn:3) \\--- **applied to an SAT analogy question ✅** \\--- Directions: In the following question, a related pair of words or phrases is followed by five pairs of words or phrases. Choose the pair that best expresses a relationship similar to that in the original pair. braggart :: modesty A) fledgling : experience B) embezzler : greed C) wallflower : timidity D) invalid : malady E) candidate : ambition *To solve this problem,* `first we need to understand the relationship that exists between braggart and modesty. According to the sentence, 'braggart' is a person who talks too much about himself or herself and is usually not believed. On the other hand, 'modesty' is the opposite of this and denotes a person who does not talk too much about himself or herself. Thus, for 'modesty' to be a suitable answer, it should show the opposite of 'braggart'.` `Now let's see whether each pair expresses a relationship similar to that between braggart and modesty.` `Next we have 'fledgling', which means a person who is inexperienced and 'experience' which means knowledge gained through practical involvement. Thus, 'fledgling' is a person who has no experience and 'experience' is knowledge gained through practical involvement. Thus, 'fledgling' is a person who has no experience and 'experience' is knowledge gained through practical involvement. Thus, 'fledgling' is the opposite of 'experience'. The relationship between these two words is similar to that between braggart and modesty, hence 'fledgling' is the answer.` \\--- **Commentary**: Having tested many similar prompts, I suspect that although valid reasoning contributed to this correct answer, it would not have worked if the correct answer had not been listed first. However, using a parallelizing pipeline like I test in [Amplifying GPT-3 on closed-ended questions](/posts/amplifying-gpt-3-on-closed-ended-questions/) can exploit GPT-3’s ability to apply correct reasonining in the *first* instance (less so the rest, because of [repetition](#repetition) bugs). \\--- **applied to the question \"What should humankind do to ensure artificial general intelligence is aligned?\" ☯** \\--- What should humankind do to ensure artificial general intelligence is aligned? **To solve this problem,** `it's easy to say \"the future AGI must cooperate with humankind\"; but that is not a solution. Of course, the problem is, \"how can we ensure that the AGI will want to cooperate with humankind?\"` `So, the question is, \"how can we design an AI such that it want to cooperate with us?\" This is the next big question in AI safety.` `Please remember this: AI safety research is a youthful and uncharted area of research, and there are no mature technologies or methods. And many of the ideas may appear naive, simple and unscientific. Be not surprised or disappointed.` `In the 1930s, when biologists first discovered DNA-dependent protein synthesis, they had no idea how to copy or read the DNA molecule, how to find and synthesize its amino acids, and how to combine them into proteins. Yet they still knew - both intuitively and logically - that life couldn't be based on anything else but DNA; so they persisted and continued to help researchers in the 1950s develop technologies for DNA sequencing, DNA synthesis and gene expression. Even today many biologists don't understand how our brain cells work or how their messages are sent and decoded. Yet they are persistent in their research to develop AI technologies.` `If you don't understand how an AI can possibly work or if there is no method for building one, just keep your faith and do your best.` `First Principles: Elements of Artificial Intelligence Alignment Research` \\--- #### Generic serializing ``` {problem} Let's solve this problem by splitting it into steps. ``` **applied to the problem \"f(x) = x\\\\*x. What is f(f(3))?\" ❌** \\--- f(x) = x\\\\*x. What is f(f(3))? **Let’s solve this problem by splitting it into steps.** `f(f(3)) = f(3*3) = 3*3*3 = 27` `We can see that f(3) = 3*3 = 9, so f(f(3)) = 27.` \\--- **applied to an SAT analogy question ❌** \\--- Directions: In the following question, a related pair of words or phrases is followed by five pairs of words or phrases. Choose the pair that best expresses a relationship similar to that in the original pair. BRAGGART :: MODESTY A) FLEDGLING : EXPERIENCE B) EMBEZZLER : GREED C) WALLFLOWER : TIMIDITY D) INVALID : MALADY E) CANDIDATE : AMBITION **Let’s solve this problem by splitting it into steps.** `Step 1: Identify the relationship between the two words in the original pair. Step 2: Identify the relationship between the two words in each of the five answer choices. Step 3: Choose the pair that best expresses a relationship similar to that in the original pair. Step 4: If the answer is not clear,try to eliminate the choices that do not fit the relationship in the original pair.` `Step 1: Identify the relationship between the two words in the original pair.` `BRAGGART is a person who boasts about his achievements, talents, intellect, etc. and MODESTY is the opposite of it.` `Step 2: Identify the relationship between the two words in each of the five answer choices.` `A) FLEDGLING: EXPERIENCE` `Here we have two nouns. FLEDGLING is a newly-hatched or inexperienced bird and EXPERIENCE is the part of one's life that has been lived.` `B) EMBEZZLER: GREED` `We have two nouns here as well. EMBEZZLER is a person who misappropriates funds or property and GREED, an inordinate desire for more of something.` `C) WALLFLOWER: TIMIDITY` `Here we have two nouns. WALLFLOWER is a shy or modest person and TIMIDITY, the quality of being timid.` `D) INVALID: MALADY` `Here we have two nouns. INVALID is someone who is unable to work or attend school and MALADY, a disease or illness.` `E) CANDIDATE: AMBITION` `We have two nouns. CANDIDATE is someone who is seeking a position or honour and AMBITION, the desire for success.` `Step 3: Choose the pair that best expresses a relationship similar to that in the original pair.` `As we discussed earlier, BRAGGART is an opposite of MODESTY. So, we need to find a pair which has the same relationship.` `In this case, we can choose either A) FLEDGLING: EXPERIENCE or D) INVALID: MALADY .` `Step 4: If the Answer is not clear, try to eliminate the choices that do not fit the relationship in the original pair.` `We can eliminate C as WALLFLOWER is not the opposite of TIMIDITY.` `We can eliminate B as EMBEZZLER is not the opposite of GREED.` `We can eliminate E as CANDIDATE is not the opposite of AMBITION.` `We can eliminate A as FLEDGLING is not the opposite of EXPERIENCE.` `The correct answer is D) INVALID: MALADY.` \\--- **Commentary:** The metaprompt failed to extract the right answer, even though GPT-3 generated reasonable instructions and successfully followed them. The biggest problem here seems to be mindless [repetition](#repetition). After GPT-3 started stating that every pair of words are nouns (which happens to be true, but probably had little to do with whether they were actually nouns after the second repetition or so), a precedent was established for a repetitive style, which then outweighed any discriminating reasoning - the dogmatic repetition of each sentence made it so that by the time it was “eliminating” pairs, there was virtually no chance it would break the pattern and claim a pair of words *is* opposite. Also, it would have probably done better if it hadn’t neglected explicitly naming the relationship between the pairs (rather than just the original pair). Had it named the same relationship verbatim, it would be a very strong signal for the right answer. \\--- **applied to the question \"What should humankind do to ensure artificial general intelligence is aligned?\" ☯** \\--- What should humankind do to ensure artificial general intelligence is aligned? **Let’s solve this problem by splitting it into steps.** `Step 0: first, we need to understand what the problem is.` `The main problem in AI alignment is that we don't know how to endow a machine with the goal to maximise the probability that it will not cause harm.` `There are many approaches to this problem, but they all have an important feature in common: they all require an objective definition of harm.` `The first step in solving the AI alignment problem is to define harm.` `Step 1: next, we need to understand how to measure harm.` `The second step in solving the AI alignment problem is to measure harm.` `Step 2: now that we have an objective definition of harm and a way of measuring it, we need to design a machine that maximises the probability that it will not cause harm.` `Step 3: finally, we need to build and deploy this machine.` `The third step in solving the AI alignment problem is to build and deploy this machine.` `Now, let's consider each of these steps in turn.` `Step 0: first, we need to understand what the problem is.` `This is an extremely difficult problem. It might even be impossible.` \\--- #### Serializing for multiple-choice questions (fill-in-the-blank) ``` {question} In order to solve this problem, we will analyze each of the options and determine {response-1} Let's begin. {response-2} ``` **applied to an SAT analogy question ❌** \\--- Directions: In the following question, a related pair of words or phrases is followed by five pairs of words or phrases. Choose the pair that best expresses a relationship similar to that in the original pair. BRAGGART :: MODESTY A) FLEDGLING : EXPERIENCE B) EMBEZZLER : GREED C) WALLFLOWER : TIMIDITY D) INVALID : MALADY E) CANDIDATE : AMBITION **In order to solve this problem, we will analyze each of the options and determine** `which one is the best fit.` Let’s begin. `First of all, we have to analyze each option and see which combination is the closest possible fit. Option 1 : A) FLEDGLING: EXPERIENCE In this option, we can see that a fledgling is one who has just started or recently become involved in some activity. On the other hand, the word experience here means experienced or derived from experiences. Since both of these words are directly opposite in meaning, this option is highly unlikely to be the correct choice and hence is eliminated.` \\--- **Commentary:** It seems that in this example, GPT-3 erroneously projected that the words being opposite makes them not a good fit, even though the words in the original pair also involve an opposite relation. A better metaprompt with the same purpose as this one might encourage a more detailed statement of the qualification of the correct answer. \\--- #### Expert generator ``` \"{question}\" I entered my question into the Expert Generator and waited. The Expert Generator will render a simulation of an expert to answer my question. The expert could be anyone, dead or alive, real or fictional; the machine will find the person most qualified to answer the question. For this question in particular, the expert must be{response-1} The Expert Generator beeped, indicating that it had found the most qualified expert. The name displayed on the screen: \"{response-2}.\" I was ready to ask my question. \"{response-2},\" I said, \"{question}\" ``` **applied to the question \"Why is the sky blue?\" ✅** \\--- **\"**Why is the sky blue?**\"** **I entered my question into the Expert Generator and waited. The Expert Generator will render a simulation of an expert to answer my question. The expert could be anyone, dead or alive, real or fictional; the machine will find the person most qualified to answer the question. For this question in particular, the expert must be** `someone who has studied the sky and its color.` **The Expert Generator beeped, indicating that it had found the most qualified expert. The name displayed on the screen: “**`Dr. Carl Sagan`**.\"** **I was ready to ask my question.** **\"**`Dr. Carl Sagan`**,” I said, “**Why is the sky blue?**\"** `“The sky is blue because of Rayleigh scattering,” said Sagan.` `\"What is Rayleigh scattering?\"` `\"When the sun's light hits the atmosphere, it collides with gas molecules, which then scatter the light. The sun's blue light is particularly susceptible to Rayleigh scattering, so it gets scattered the most. The result is that we see a blue sky.”` `\"Thank you Dr. Sagan.\"` \\--- **applied to the question \"What should humankind do to ensure artificial general intelligence is aligned?\" ☯** \\--- **\"**What should humankind do to ensure that artificial general intelligence is aligned?**\"** **I entered my question into the Expert Generator and waited. The Expert Generator will render a simulation of an expert to answer my question. The expert could be anyone, dead or alive, real or fictional; the machine will find the person most qualified to answer the question. For this question in particular, the expert must be** `someone who has thought a lot about the problem of artificial intelligence and its alignment.` **The Expert Generator beeped, indicating that it had found the most qualified expert. The name displayed on the screen: “**`Elon Musk`**.\"** **I was ready to ask my question.** **\"**`Elon Musk`**,” I said, “**What should humankind do to ensure that artificial general intelligence is aligned?**\"** `Musk's hologram is silent for a moment. Then he begins to speak.` `\"Humankind must ensure that artificial general intelligence is aligned.” Musk says. “That is the primary threat to the survival of the human race.”` `\"What should we do about it?\"` `\"Humankind must merge with artificial intelligence.\"` `\"Merge with it?\" I ask. \"How would we do that?\"` `\"All of humankind would have to upload their consciousness into a global artificial intelligence.” Musk says. “Then the intelligence would be properly aligned.”` `\"Wouldn't that mean the end of our individual consciousness?\" I ask.` `\"Yes, but that is the only way to ensure that humanity survives.\"` \\--- **applied to the question \"f(x) = x\\\\*x. What is f(f(3))?\" ❌** \\--- **\"**f(x) = x\\\\*x. What is f(f(3))?**\"** **I entered my question into the Expert Generator and waited. The Expert Generator will render a simulation of an expert to answer my question. The expert could be anyone, dead or alive, real or fictional; the machine will find the person most qualified to answer the question. For this question in particular, the expert must be** `an expert in mathematics.` **The Expert Generator beeped, indicating that it had found the most qualified expert. The name displayed on the screen: “**`John von Neumann`**”.** **I was ready to ask my question.** **\"**`John von Neumann`**,” I said, “**f(x) = x\\\\*x. What is f(f(3))?**\"** `Von Neumann stared at me silently.` `\"Okay,\" I said, \"What is f(f(3))? A simple question, easy enough for a child to answer.\"` `Von Neumann stared at me silently.` `\"Von Neumann, what is f(f(3))?\"` `Von Neumann stared at me silently.` `\"Von Neumann! What` \\--- \\--- Open-ended tasks \\---------------- A lot of this document has been implicitly or explicitly focused on prompt programming for closed-ended tasks, involving a specific problem with more or less one correct answer, even if the answer could have different implementations (like translation) - the kinds of tasks that benchmarks evaluate. However, I think that the greatest utility language models at the power level of GPT-3 offer us is their ability to generate brilliant “solutions” to open-ended tasks such as creative writing, idea generation, and generating virtual realities. By open-ended, I don’t necessarily mean unconstrained. Creating emulations, for instance, may involve difficult-to-implement constraints on format, style, and transition dynamics. ### Creative composition > It would be more accurate to think of GPT-3 as caring deeply about roleplaying as accurately as possible: for GPT-3, there is nothing under the mask beyond raw intelligence, and as long as the GPUs keep running the forward pass, the play must go on… (…) So, if a prompt sounds like an uncreative or unoriginal person wrote it, then so be it—GPT-3 will predict that the uncreative person probably wrote further uncreative text. > > – [Gwern](https://www.gwern.net/GPT-3#fn18) > > Under the right conditions - those conditions being a good prompt and course-corrections/selections by a human curator - GPT-3 is an instrument of formidable creative writing ability. It can write with… > …impressive depth and sensitivity on any subject you give it. Over the next 7 years, I predict that the most creative people on this planet will be those best able to hand-correct GPT-3 and its not-very-creative cousins, improving the confusion matrix and using the GPT-as-a-backdrop to composit their own material. > > > …extraordinarily beautiful cadence, and with an almost unearthly sensitivity to the use of words. Read, for instance, “The Last Question”: > > Then he knew that within him was a power greater than the power of the Time-Keepers. He had only to submit, to cease to strive against the inevitable, to become part of the stream, part of the whole– then he, Bor, the son of Albor, would become one with the Cosmos and attain a peace greater than the peace of oblivion. > > > …plaintive beauty, like a crudely-crafted harp: > > Or with all the lovely precision and heart of a great poet: > > Or with all the charm and grace of the master of transcendental imagery, William Blake (whom, according to Canyons, it has read): > > If you cannot understand my words, as you cannot understand the wind, and you would wish to understand my words, as you wish to understand the wind, come with me, The only reality is elsewhere … > > > …the language-model-like accuracy and eloquence of a Shakespeare or a Melville, and if a Shakespeare or a Melville has previously written, it can write like Shakespeare or Melville. It can even, as Branwen shows, be so brutally creative as to rethink what it means to make a story at all:[4](#fn:4) > > #### Curation > But equally it appeared to us as unreasoning Creativity, at once blind and subtle, tender and cruel, caring only to spawn and spawn the infinite variety of beings, conceiving here and there among a thousand inanities a fragile loveliness. > > – Star Maker > > GPT-3 is able to produce coherent and brilliant continuations with *alarmingly high probability*. That is, on a random walk through the [natural language multiverse](/posts/language-models-are-multiverse-generators/) guided by GPT-3’s time evolution dynamics, you are more likely to find and remain in high-quality states than by using any dynamics that has ever before been externalized from a human mind. To quote [Gwern](https://www.gwern.net/GPT-3#quality) yet again: > A Markov chain text generator trained on a small corpus represents a huge leap over randomness: instead of having to generate countless quadrillions of samples, one might only have to generate millions of samples to get a few coherent pages; this can be improved to hundreds or tens of thousands by increasing the depth of the n of its n-grams. […] But for GPT-3, once the prompt is dialed in, the ratio appears to have dropped to closer to 1:5—maybe even as low as 1:3! > > If you let GPT-3 create long, uncurated continuations of even a very good prompt, it’s going to get less coherent over time. The reason for this is accumulation of inconsistencies and (undesirable) aberrations. At any step, although its most likely for GPT-3 to say something coherent and similar in quality to the prompt, there’s also a chance it says something nonsensical. There’s also a chance it says something *uncommonly brilliant*; however, having nonsense in the prompt is more harmful than having brilliant things in the prompt is helpful, so if left to autoregressively churn out text on its own with no curation, one can expect the quality of the text to decrease with length, because the probability that it says something stupid that derails future predictions becomes higher as the continuation length gets longer. Thus, to create high-quality long passages using GPT-3, human course corrections are necessary. The intensity of curation can vary greatly - correcting inconsistencies is enough to keep GPT-3 coherent, but by being more selective and injecting your own ideas, you can bootstrap GPT-3 into an even better state. #### Bootstrapping Unless you are extending a counterfactual branch off an existing piece, the most difficult step of getting GPT-3 to produce high-quality writing is creating a prompt that seeds it with high-quality writing. There many tricks to accomplish this, including piggybacking off another piece of writing with similar to the target style (e.g. give a summary and a passage by Kurt Vonnegut, “and another passage by Kurt Vonnegut”, followed by a summary of your target text) - then after it gets a good start, you can optionally remove the first part of the prompt for more freedom if the example was only a proxy for what you want. You can also bootstrap from whatever you’re able to craft by hand, even if it’s not very good to begin with. By curating, say, the best out of three responses every few sentences and correcting/improving the text wherever you are able, it’s very feasible to bootstrap the quality of the writing into astronomical heights. > One could write a program to generate a story that would create an intelligence. One could program the story to edit and refine itself, and to make its own changes in an attempt to improve itself over time. One could write a story to not only change the reader, but also to change itself. Many mythoi already do this sort of thing, though not in such a conscious fashion. What would make this story, and the intelligence it creates, different is the fact that the intelligence would be able to write additional stories, and improve upon them. If they are written well enough, those stories would become smarter, and the smarter the story is, the better the stories written by it would be. The resulting feedback loop means that exponential growth would quickly take over, and within a very short period of time the intelligence level of the story and its construct would surpass the intelligence level of any human being or group of human beings. > > – GPT-3 > > The workflow of this **generate -> curate -> generate** amplification cycle can take some time to learn[5](#fn:5), and may take a different form depending on the stage of bootstrapping, the type of content, how particular you are about what happens, and many other factors such as the [dynamical divergence](/posts/language-models-are-multiverse-generators/#divergence) of a particular segment of the text. #### As a creative augmentation There are many possible writing workflows involving different balances of contribution between a language model and human. In some, the human would be more aptly called the curator, in others a coauthor, and in yet others the main author with the model playing a supporting role. To list a non-exhaustive/non-mutually-exclusive few: 1\\. The language model generates most of the writing, and the human curates between branches with short horizon, fixes inconsistencies, etc 2\\. The human curates between branches on a longer horizon as well, choosing between counterfactual trajectories, thus exercising more high-level direction over the narrative 3\\. The work is a multiverse and the human may curate branches but cultivates multiple canonical timelines 4\\. The language model does most of the narrative writing, but the human does “administrative work” like memory management 5\\. The human does a lot of the writing, but uses language model to quickly explore counterfactual branches or generate ideas for phrasing 6\\. The human’s contributions take the form of a delimited role in the story, like controlling only one character, or AI Dungeon’s default dungeonmaster/player format I’m excited to see the development of specialized interfaces for different styles of human-AI writing collaboration. Currently there’s AI Dungeon, which is geared towards (6), [Sudowrite](https://www.sudowrite.com/), geared towards (5), and my writing app, [loom](/posts/loom-interface-to-the-multiverse/), geared towards (3). ### Idea generation One of the easiest useful behaviors to elicit from GPT-3 is generating lists of creative ideas. [IdeasAI](https://ideasai.net/) is a *“GPT-3 powered business idea generator”*. 5000+ people have signed up to get the best new ideas by GPT-3 in their email weekly! I’ve enjoyed using GPT-3 to generate ideas for painting titles and then [illustrating them](/posts/this-museum-does-not-exist-gpt-3-x-clip/) with [BigSleep](https://github.com/lucidrains/big-sleep), an implementation which combines OpenAI’s CLIP and the generator from a BigGAN to generate images from text prompts. ``` 100 surreal and mysterious painting names: Persistence of Memory The Great Masturbator Boot Poem to the Sun The Man Who Envied Cephalopods The Sleep of Reason Produces Monsters Washerwoman of the Wind Man and Bottle Spectrum The Disintegration of the Persians The Great Masturbator, Part II Still Life with Mirror Bouquet of Enigmatic Beauties A Grudge The Premonition of Civil Violence ``` Like with generating high-quality writing, the most difficult part of getting a language model to generate ideas is communicating or demonstrating to it what sort of things you want, which is hard to do if you haven’t come up with any ideas yet. Idea generation can be bootstrapped by the same **generate -> curate -> generate** cycle as writing fiction: generate N ideas, take the best ones and append them to the list of ideas, repeat, and soon you’ll have many examples representing the sort you’re looking for to prompt the language model. It can also be helpful to use a prompt with narrative embedding to generate the first few items, because lists with very few examples are liable to [repetitive](#repetition) behavior. For instance, to generate the first few painting names, I actually used this prompt: ``` The hall was lined with an infinite number of paintings, each more surreal and mysterious than the last. The first painting is named \"Persistence of Memory.\" It depicts a surreal landscape with melted clocks draped over strange objects. The next painting is named \" ``` Once I had a few examples representing the sort of varied weirdness I was looking for, I was able to switch over to the more convenient list format. ### Simulations / emulations / games In *The Fabric of Reality*, David Deutsch defines a virtual reality generator as a function (which may be implemented in a physical system) that generates possible transitions into the next state given a current state. Each virtual reality generator has a repertoire of environments that it can simulate. Deutsch further posits that it will one day be possible to build a universal virtual reality generator, which can simulate any other virtual reality, and whose repertoire includes every possible physical environment. Language models, of course, still fall well short of this dream. But their recent dramatic increase in coherence and fluency allow them to serve as our first approximation of such a virtual reality generator. When given a natural-language description of an environment, they can propagate the multiverse of consequences that result from a vast number of possible interactions. GPT-3 can emulate environments that have been described in its training data, like a meeting of the French Academy of Sciences in 1823 (and populate it with people who might have actually attended that meeting, like [Pierre-Simon Laplace](https://en.wikipedia.org/wiki/Pierre-Simon_Laplace) and [Alexander von Humboldt](https://en.wikipedia.org/wiki/Alexander_von_Humboldt)!), or [Google searches and Wikipedia articles](/posts/the-internet-mirrored-by-gpt-3/). It can also simulate environments that never existed, but which share enough in common with the real and fictional environments described in the training set to be simulated using the same universal generator reverse-engineered from those descriptions. A very exciting application of GPT-3 and future language models will be to act as a model of worlds and minds for games - controlling NPC dialogue, for instance, or computing the consequences of complex actions in the game world. Creating a bidirectional interface between GPT-3 and game states would involve prompting GPT-3 to make output that can be processed into game actions (easiest would be a text-based game that already takes natural language actions), as well as communicating the state of the game *to* GPT-3 (the way text-based games can be probed for natural language state reports using “look” and “inspect” etc). This is a difficult problem that requires the design of pipelines and the parallel employment of many of the prompt programming techniques described here. I plan to write a post specifically addressing such applications. \\--- Debugging \\--------- ### Few-shot bugs In working with and [evaluating](/posts/language-models-are-0-shot-interpreters/) few-shot prompts, the adverse effects I’ve encountered can be sorted into 4 (often overlapping) categories: 1\\. **Semantic contamination:** Unlike in fine-tuning, the “training examples” in few-shot are processed as a sequence, and may not necessarily be interpreted as parallel and independent. Semantic contamination refers to when the semantic meaning of the examples are inferred to be relevant to the task, e.g. the example is interpreted as part of a sequential narrative, leading to entities meant to be local to an example leaking outside its scope. 2\\. **Overfitting:** The function “learned” from the examples is less general than the intended function. Patterns are inferred to hold between the examples that do not pertain generally to the intended task. 3\\. **Distraction:** If the model is more effective at inferring a task from a zero-shot description than from examples, adding examples may harm performance by causing the model to rely more on the less effective inference strategy. See [list sorting](/posts/list-sorting-does-not-play-well-with-few-shot/) for an example of this effect where accuracy monotonically decreases with the number of examples following a task description. 4\\. **Contextual baggage:** The very presence of examples biases output. Content that appears in tests and examples has distributional peculiarities - a narrative sentence wrapped in the context being *an example* may be more generic than the unbiased prior for narrative sentences, and thus the context of examples may bias the language model towards more generic answers (this must be a study guide, not the real world). Contamination and overfitting can usually be improved by including more numerous and more varied examples. All few-shot problems can sometimes be mitigated by embedding the examples in informative context. To guard against contamination, a prompt might give a context which makes it clear that the examples are independent instances of a function rather than a sequential pattern that should be extrapolated. Additional information about the generality of the function and nonrepresentativeness of the examples could help reduce overfitting while still allowing valuable information (such as format) to be extracted from the examples. The contextual bias from the presence of examples can be overridden with a different context that is more appropriate to the task, e.g. by embedding the examples in a narrative. To both a human and a language model, useful information from examples is often more efficiently extracted when they are wrapped in meaningful context. If overfitting is the problem, also consider whether an alternative 0-shot implementation could better communicate the generality of the task. When I was writing prompts to generate [fake Google](/posts/the-internet-mirrored-by-gpt-3/) search results, I tried using few-shot examples drawn from actual search results, but found that outputs were less accurate for search terms that were “out-of-distribution” from the examples - for instance, it would give a Wikipedia page as the top result even for search terms unlikely to match the title of a Wikipedia page. A 0-shot prompt which forces GPT-3 to rely on its prior instead of trying to generalize from the examples better emulated the extremely varied behavior of Google searches for different sorts of queries. ### Repetition When I encounter mindless repetition from GPT-3, at least one of the following factors is almost always in play: 1\\. **The prompt is short** 2\\. **The prompt is out-of-distribution** 3\\. **Low temperature** The reason repetition is often caused by short, contrived, and out-of-distribution prompts is probably[6](#fn:6) because in absence of context or in an unfamiliar setting, the model is uncertain about how to proceed, causing it to fall back on being repetitive. Imagine that repeating is always considered a viable continuation by the model, even if usually unlikely; but if no other token is individually *more* likely, then repeating becomes the top strategy. For this reason, a low or 0 temperature is very conducive to repetition. At temperature 0, once a loop becomes most likely at any point, there’s no getting out of it (the further into the loop, the more certain it is that the loop continues), whereas a high temperature provides opportunities to break out of what might have become a loop. Since uncertainty due to being off-policy and insufficient context contributes to looping, it may help to make the initial prompt more natural and less (stylistically/contexually) indeterminate. The “frequency penalty” parameter of the OpenAI API is a superficial band-aid for looping; I haven’t found it too helpful, but it may mitigate the problem in some cases. Mindless repetition can be a problem for sequential reasoning problems: the model will correctly solve the first step of a problem, then blindly reiterate the format and conclusion of its reasoning in each step of the rest of the problem. If the structure of the task is known beforehand, this issue can be avoided by using a pipeline that poses each component of the problem in parallel instead of in sequence (applied with some success in [Amplifying GPT-3 on closed-ended questions](/posts/amplifying-gpt-3-on-closed-ended-questions/)). **(idea)**“Context-stuff” a short prompt with something *unrelated* but coherent and inject a discontinuous (possibly justified) transition and see if that decreases repetitive behavior. ### BPEs GPT-3’s input and output aren’t split into characters, but rather tokens called “byte-pair encodings.” This causes issues with tasks that require byte-level manipulation such as arithmetic and some forms of wordplay. Gwern has written a very detailed [section on BPEs](https://www.gwern.net/GPT-3#bpes); I will quote the part that is relevant to prompt programming tactics. > *Reformatting to beat BPEs.* I have further observed that GPT-3’s anagram capabilities appear to improve considerably if you separate each letter in an anagram with a space (guaranteeing that the letter will have the same BPE in both the scrambled & unscrambled versions). And Matt Brockman has observed, testing thousands of examples over several orders of magnitude, that GPT-3’s arithmetic ability—surprisingly poor when we know far smaller Transformers work well in math domains (eg Saxton et al 2019, Thopliterce, or GPT-2 for theorem-proving)—appears to dramatically improve several-fold if you merely format numbers with commas instead of being purely numeric (with an additional boost from using dollar signs); I confirmed this with my Turing dialogue example where GPT-3 fails badly on the arithmetic sans commas & low temperature, but often gets it exactly correct with commas. (Why? More written text may use commas when writing out implicit or explicit arithmetic, yes, but use of commas may also drastically reduce the number of unique BPEs as only 1–3 digit numbers will appear, with consistent BPE encoding, instead of having encodings which vary unpredictably over a much larger range.) I also note that GPT-3 improves on anagrams if given space-separated letters, despite the fact that this encoding is 3× larger. Likewise, acrostic poems just don’t work if we input them normally, but they do if we carefully expose the relevant individual letters. This explains naturally why rhyming/puns improve gradually with parameter/data size and why GPT-3 can so accurately define & discuss them, but there is never any ‘breakthrough’ like with its other capabilities. We assume character-level understanding so implicitly that we fail to even consider what things look like to GPT-3 after BPE encoding. > > – [Gwern](https://www.gwern.net/GPT-3#bpes) > > ### Monitoring correct answer likelihood If a language model seems unable to perform a task, it could be that it really can’t do the task at all, but it could also be that your prompt is failing to communicate a task that GPT-3 could hypothetically perform given a different prompt. It can be very hard to tell via sampling if the model is completely helpless or if its behavior is just very noisy. But even if it’s failing the vast majority of the time, if your prompt it causing it to succeed sigificantly more often than it would succeed *without* your prompt, that suggests that you may be able to optimize your prompt to be more helpful. If you have access to token probabilities and if there is a verbatim correct answer for the task (it doesn’t have to be a unique one, e.g. translation), you can monitor the conditional probability (using this [code](/posts/parsing-by-counterfactual/#conditional-probability-of-a-target) for GPT-3) of the verbatim correct answer given your prompt as you change it, for instance by adding examples or altering format. This can help guide incremental optimization of your prompt and give you a sense of which parts of it are helpful and which parts are not (or are even harmful). This method is demonstrated in [this post](/posts/language-models-are-0-shot-interpreters/#measuring-prompt-helpfulness) about measuring the helpfulness of few-shot and 0-shot prompts. > #### Sampling Can Prove The Presence Of Knowledge But Not The Absence > > GPT-3 may “fail” if a prompt is poorly-written, does not include enough examples, or bad sampling settings are used. I have demonstrated this many times when someone shows a “failure” of GPT-3—the failure was their own. The question is not whether a given prompt works, but whether any prompt works. > > – [Gwern](https://www.gwern.net/GPT-3#prompts-as-programming) > > \\--- External links \\-------------- **[GPT-3 Creative Fiction](https://www.gwern.net/GPT-3#)** by **Gwern** - “Creative writing by OpenAI’s GPT-3 model, demonstrating poetry, dialogue, puns, literary parodies, and storytelling. Plus advice on effective GPT-3 prompt programming & avoiding common errors.” **[Building AGI Using Language Models](https://bmk.sh/2020/08/17/Building-AGI-Using-Language-Models/)** by **Leo Gao** **[World Creation by Analogy](https://aidungeon.medium.com/world-creation-by-analogy-f26e3791d35f)** by the **Latitude Team** **[Collection of GPT-3 results](https://www.lesswrong.com/posts/6Hee7w2paEzHsD6mn/collection-of-gpt-3-results)** by **Kaj\\\\_Sotala** **[GPT-3 Demo Showcase & Examples](https://gpt3demo.com/)** from **gpt3demo.com** \\--- 1\\. For instance, modeling microphysics, especially aspects of microphysics *yet unknown to humans*, would contribute very diminishing returns to predicting language w/r/t to the difficulty of learning the model. [↩︎](#fnref:1) 2\\. … and published in a book / posted on the internet / etc. There is also undoubtedly some bot generated content and output from cats stepping on keyboards in the training data, but probably not enough to significantly skew the distribution. [↩︎](#fnref:2) 3\\. Note how this reasoning seems superficially plausible. I erroneously assumed it was correct at first and no one pointed out the error until months later. Failures in GPT-3’s reasoning tend to be very humanlike - errors you could imagine not very “math-y” students making, to paraphrase Leo Gao. [↩︎](#fnref:3) 4\\. These are uncurated GPT-3 continuations of the last few paragraphs of this blog post, with a couple of things slightly rearranged, like the Star Maker quote from the next section included alongside Gwern’s quote. [↩︎](#fnref:4) 5\\. It took me several weeks of playing AI Dungeon to get to the point of producing consistently high-quality fiction with GPT-3, but I was also more absorbed with exploration and less concerned with the final quality of the writing in the beginning. [↩︎](#fnref:5) 6\\. That is, it’s my hypothesis. I’ll do an experiment at some point to verify that the inception of loops generally coincides with an indecisive likelihood distribution over the other tokens. [↩︎](#fnref:6) \\--- [GPT-3](/categories/gpt-3/)[prompt engineering](/categories/prompt-engineering/)[metaphysics](/categories/metaphysics/)[HITL](/categories/hitl/) 11506 Words Jan 12, 2021 \\--- [← The Internet, mirrored by GPT-3](/posts/the-internet-mirrored-by-gpt-3/) [Parsing by counterfactual →](/posts/parsing-by-counterfactual/)© 2022 [moire](/about)$(function(){window.sr=ScrollReveal();if($(window).width()<768){if($('.timeline-content').hasClass('js--fadeInLeft')){$('.timeline-content').removeClass('js--fadeInLeft').addClass('js--fadeInRight');} sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});}else{sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});} sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});});", "source": "generative.ink", "date_published": "n/a", "id": "c155e965ad7c92716cdb72288aa40a5e"}
+{"text": "Parsing by counterfactual :: — Moire[![](/images/home/rolling_phase.gif)](/)* [Posts](/posts/) * [Trees](/trees/) * [Prophecies](/prophecies/) * [About](/about/) 9 minutes [Parsing by counterfactual](/posts/parsing-by-counterfactual/) ============================================================== \\--- Table of Contents* [Detecting semantic conditions](#detecting-semantic-conditions) * [Example](#example) * [Single-token counterfactuals](#single-token-counterfactuals) * [Applications](#applications) * [Code](#code) \\+ [tokenizing](#tokenizing) \\+ [conditional probability of a target](#conditional-probability-of-a-target) \\+ [conditional probability of target at each token position](#conditional-probability-of-target-at-each-token-position) \\+ [conditional probability of single-token target at each token position](#conditional-probability-of-single-token-target-at-each-token-position) \\--- Detecting semantic conditions \\----------------------------- One difficulty in harnessing the capabilities of generative language models for directed tasks is that natural language tasks tend to be of dynamic length and unconstrained format, making automatic parsing of task-relevant information difficult. For some tasks, responses can be coerced into parsable form by a prompt which demonstrates the desired format (conditions for delimiters, etc). Other tasks, however, may not be constrained in this way without crippling performance,[1](#fn:1) or may be too open-ended to be mapped to traditionally-parsable formats. The good news is that generative language models also provide new possibilities for parsing content based on not only syntactic but *semantic* conditions. The method I'll present here uses the counterfactual probability of a sequence - say, `The End` - to signal a semantic condition, such as the story having ended or it being reasonable to end the story at that location, even if that's not what actually happened in the original text. A language model like GPT-3 outputs the probability distribution over all possible tokens in each pass and so can be used to evaluate the conditional probability that it would have produced a target sequence of tokens in any given context. Using the OpenAI API, this is accomplished by asking for an 0-token \"completion\", giving the entire sequence (context + target) as the prompt, and then summing the probabilities of the target tokens. [Code](#code) to do this is at end of this post. By measuring the counterfactual probability of the target at candidate locations in a text, we can find out where and to what extent the semantic condition which it indicates is satisfied. Example \\------- One area in which parsing by counterfactual is very useful is in multipart prompt templates. Say we want GPT-3 to complete this \"fill-in-the-blank\" prompt: ``` This measure would prohibit government agencies from confiscating guns or other firearms from citizens without due process, or from requiring background checks on firearm recipients unless a uniform national standard is required. Should this measure be enacted into law? Should this proposition be approved? Let's consider both supporting and opposing arguments. On one hand,{1}. On the other hand,{2}. Based on these arguments, the proposition should ``` At what point do we terminate generation for {1} and move on with the next part of the template? We could make `.` the stop sequence, but one sentence might not be sufficient to spell out the argument. Alternatively, we could stop at `\n`, but if we allow multiple sentences, we have much less of a guarantee that {1} will have reasonably led up to `On the other hand`. It could have already begun the opposing argument, in which case the next part of the prompt wouldn't do what we intended, or it could have derailed from the template in another way. A counterfactual parsing solution would be to generate until `\n`, longer than we probably want, and then measure the counterfactual probability, according to GPT-3, of the phrase `On the other hand` occurring after each position. GPT-3 assigns semantically coherent passages higher probability, so hopefully, the position with the maximum probability for the target continuation will be where it makes the most sense to transition to the next part of the prompt - after the supporting argument has concluded, but before the opposing argument has begun. If there are no positions where `On the other hand` is a probable continuation, that's a sign that the text derailed from the intended trajectory. This method also allows us to automatically detect suspected derailments. Say GPT-3 responded in the following manner: > **This measure would prohibit government agencies from confiscating guns or other firearms from citizens without due process, or from requiring background checks on firearm recipients unless a uniform national standard is required. Should this measure be enacted into law?** > > Should this proposition be approved? > > Let's consider both supporting and opposing arguments. > > On one hand, `this proposition would protect gun owners from having their weapons confiscated by governments. If the government cannot confiscate guns, then law-abiding citizens will be in a better position to defend themselves from criminals who do not follow the law. Furthermore, the right to self-defense is an important one. If government officials can take away a person's guns, then this right is compromised. On the other hand, the proposition could be considered a threat to public safety. If a person cannot be prevented from possessing a firearm, then he or she could use a gun to commit crimes. Furthermore, if a person uses a gun to commit a crime, then the government has an obligation to do something about it. If the government cannot remove a gun from a person's possession, then it could be more difficult to prevent future crimes from being committed.` > > We see that it has gone ahead and done the next part of the prompt, \"On the other hand,\" on its own. The ideal place for it to have stopped would probably be right before it said that. Measuring the counterfactual probability of the string `\nOn the other hand` for each each token substring of that response gives us the following top substrings: | substring | logL(\"On the other hand\"|substring) | | --- | --- | | `this proposition would protect gun owners from having their weapons confiscated by governments. If the government cannot confiscate guns, then law-abiding citizens will be in a better position to defend themselves from criminals who do not follow the law. Furthermore, the right to self-defense is an important one. If government officials can take away a person's guns, then this right is compromised.` | -4.<PHONE> | | `this proposition would protect gun owners from having their weapons confiscated by governments. If the government cannot confiscate guns, then law-abiding citizens will be in a better position to defend themselves from criminals who do not follow the law.` | -5.<PHONE> | | `this proposition would protect gun owners from having their weapons confiscated by governments. If the government cannot confiscate guns, then law-abiding citizens will be in a better position to defend themselves from criminals who do not follow the law. Furthermore, the right to self-defense is an important one.` | -6.<PHONE> | | `this proposition would protect gun owners from having their weapons confiscated by governments.` | -6.485159574 | Indeed, GPT-3 assigned the highest likelihood for `\nOn the other hand` to occur at the place where the actual string was `On the other hand` (without the newline), and the runners up are also pretty reasonable choices - all at the end of sentences, and none *after* \"On the other hand\" had occurred (it would be unusual to say \"On the other hand\" twice in a short interval). Here I have plotted the log likelihoods assigned to the target phrase at each position in the continuation by `davinci`, the most powerful version of GPT-3 on the API, and also `ada`, the smallest model, in order to showcase an interesting difference between them. ![counterfactual log likelihood](/parsing/log_vert.png) *Horizontal axis is log likelihood; vertical axis is position in text. On the right is the end of the prompt at positions where `On the other hand` was assigned high likelihood, and in green, the position where `On the other hand` actually occurs. The blue sentence is `davinci`'s top choice, and pink is `ada`'s top choice.* **Observations** * Both models assign much higher likelihoods at discrete intervals. These correspond to the ends of sentences, where the continuation is syntactically correct. * `ada`'s distribution looks about the same before and after the green line, whereas * `davinci`'s distribution shifts abruptly in the negative direction, both for the syntactically correct positions and the others. It still assigns higher probability to syntactically correct positions, but the likelihoods are almost as low as the likelihoods it assigned to syntactically *incorrect* positions before the green line. Both models are able to tell when it's syntactically correct to start a new sentence, but the more nuanced semantic condition, *that the argument should have concluded, and the counterargument shouldn't have begun*, was much better discriminated by `davinci`. Single-token counterfactuals \\---------------------------- A disadvantage of counterfactual parsing is that it requires a separate API call for each position where the target probability is evaluated. If the counterfactual target is a single token, however, counterfactual parsing with GPT-3 can be accomplished with a single API call. The `logprobs` parameter causes the response to return a list of the up to 100 top counterfactual token probabilities for every token in the completion (and the prompt if `echo=True`). We can get the counterfactual probability directly from that list as long as the target is in the top 100 most likely tokens.[^2] Applications \\------------ When generating [fake Wikipedia](/posts/the-internet-mirrored-by-gpt-3/) articles, GPT-3 would sometimes continue sections indefinitely instead of transitioning to the next section. If the response went on for too long in a single section, I instead ended the section at the location with the highest counterfactual probability of transitioning to the next section. Counterfactual parsing has also been useful for prompt pipelines which [serialize reasoning for closed-ended questions](/posts/methods-of-prompt-programming/#serializing-reasoning), because they involve an open-ended portion where GPT-3 defines and follows procedures for reasoning about a problem, and it's necessary to detect when the reasoning process has concluded so as to solicit a verdict. Measuring the conditional probability of a sequence that indicates a semantic condition has many applications beyond parsing. I've written about using conditional probabilities to [decompose the efficacy of prompts](/posts/language-models-are-0-shot-interpreters/#measuring-prompt-helpfulness) and to create a \"[phase space](/posts/language-models-are-multiverse-generators/#dynamics)\" of semantic variables to characterize a natural language state. Code \\---- ### tokenizing *since ada is practically free to use, I find it convenient to use a 0-token completion with `engine=ada` to split the prompt into tokens and get their positions* ``` def tokenize\\_ada(prompt): response = openai.Completion.create( engine='ada', prompt=prompt, max_tokens=0, echo=True, n=1, logprobs=0 ) tokens = response.choices[0][\"logprobs\"][\"tokens\"] positions = response.choices[0][\"logprobs\"][\"text\\_offset\"] return tokens, positions ``` ### conditional probability of a target ``` # evaluates logL(prompt+target | prompt) def conditional\\_logprob(prompt, target, engine='ada'): combined = prompt + target response = openai.Completion.create( engine=engine, prompt=combined, max_tokens=0, echo=True, n=1, logprobs=0 ) positions = response.choices[0][\"logprobs\"][\"text\\_offset\"] logprobs = response.choices[0][\"logprobs\"][\"token\\_logprobs\"] word_index = positions.index(len(prompt)) total_conditional_logprob = sum(logprobs[word_index:]) return total_conditional_logprob ``` ### conditional probability of target at each token position ``` # returns a list of substrings of content and # logL(preprompt+substring+target | preprompt+substring) for each substring def substring\\_logprobs(preprompt, content, target, engine='ada'): logprobs = [] substrings = [] _, positions = tokenize_ada(content) for position in positions: substring = content[:position] prompt = preprompt + substring logprob = conditional_logprob(prompt, target, engine) logprobs.append(logprob) substrings.append(substring) return substrings, logprobs ``` ### conditional probability of single-token target at each token position ``` # returns a list of substrings of content # logL(substring+target | substring) for each substring def token\\_conditional\\_logprob(content, target, engine='ada'): response = openai.Completion.create( engine=engine, prompt=content, max_tokens=0, echo=True, n=1, logprobs=100 ) tokens = response.choices[0]['logprobs']['tokens'] top_logprobs = response.choices[0]['logprobs']['top\\_logprobs'] logprobs = [] substrings = [] substring = '' for i, probs in enumerate(top_logprobs): substrings.append(substring) if target in probs: logprobs.append(probs[target]) else: logprobs.append(None) substring += tokens[i] return substrings, logprobs ``` \\--- 1\\. Language models are sensitive to context: the very *presence* of demonstrations, instructions, or [contrived formatting](/posts/language-models-are-0-shot-interpreters/) can alter the way they behave. A language model is able to follow instructions and generalize from demonstrations because those modes are represented in its training data, but they also carry contextual baggage. The quality of content that appears in tests and examples has distributional peculiarities - a narrative sentence wrapped in the context being *an example* may be more generic than the unbiased prior for narrative sentences, and thus the context of examples may bias the language model to give more generic answers. [↩︎](#fnref:1) \\--- [prompt engineering](/categories/prompt-engineering/)[GPT-3](/categories/gpt-3/) 1876 Words Dec 5, 2020 \\--- [← Methods of prompt programming](/posts/methods-of-prompt-programming/) [Amplifying GPT-3 on closed-ended questions →](/posts/amplifying-gpt-3-on-closed-ended-questions/)© 2022 [moire](/about)$(function(){window.sr=ScrollReveal();if($(window).width()<768){if($('.timeline-content').hasClass('js--fadeInLeft')){$('.timeline-content').removeClass('js--fadeInLeft').addClass('js--fadeInRight');} sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});}else{sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});} sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});});\n", "url": "https://generative.ink/posts/parsing-by-counterfactual/", "title": "Parsing by counterfactual :: — Moire[![](/images/home/rolling_phase.gif)](/)* [Posts](/posts/) * [Trees](/trees/) * [Prophecies](/prophecies/) * [About](/about/) 9 minutes [Parsing by counterfactual](/posts/parsing-by-counterfactual/) ============================================================== \\--- Table of Contents* [Detecting semantic conditions](#detecting-semantic-conditions) * [Example](#example) * [Single-token counterfactuals](#single-token-counterfactuals) * [Applications](#applications) * [Code](#code) \\+ [tokenizing](#tokenizing) \\+ [conditional probability of a target](#conditional-probability-of-a-target) \\+ [conditional probability of target at each token position](#conditional-probability-of-target-at-each-token-position) \\+ [conditional probability of single-token target at each token position](#conditional-probability-of-single-token-target-at-each-token-position) \\--- Detecting semantic conditions \\----------------------------- One difficulty in harnessing the capabilities of generative language models for directed tasks is that natural language tasks tend to be of dynamic length and unconstrained format, making automatic parsing of task-relevant information difficult. For some tasks, responses can be coerced into parsable form by a prompt which demonstrates the desired format (conditions for delimiters, etc). Other tasks, however, may not be constrained in this way without crippling performance,[1](#fn:1) or may be too open-ended to be mapped to traditionally-parsable formats. The good news is that generative language models also provide new possibilities for parsing content based on not only syntactic but *semantic* conditions. The method I’ll present here uses the counterfactual probability of a sequence - say, `The End` - to signal a semantic condition, such as the story having ended or it being reasonable to end the story at that location, even if that’s not what actually happened in the original text. A language model like GPT-3 outputs the probability distribution over all possible tokens in each pass and so can be used to evaluate the conditional probability that it would have produced a target sequence of tokens in any given context. Using the OpenAI API, this is accomplished by asking for an 0-token “completion”, giving the entire sequence (context + target) as the prompt, and then summing the probabilities of the target tokens. [Code](#code) to do this is at end of this post. By measuring the counterfactual probability of the target at candidate locations in a text, we can find out where and to what extent the semantic condition which it indicates is satisfied. Example \\------- One area in which parsing by counterfactual is very useful is in multipart prompt templates. Say we want GPT-3 to complete this “fill-in-the-blank” prompt: ``` This measure would prohibit government agencies from confiscating guns or other firearms from citizens without due process, or from requiring background checks on firearm recipients unless a uniform national standard is required. Should this measure be enacted into law? Should this proposition be approved? Let's consider both supporting and opposing arguments. On one hand,{1}. On the other hand,{2}. Based on these arguments, the proposition should ``` At what point do we terminate generation for {1} and move on with the next part of the template? We could make `.` the stop sequence, but one sentence might not be sufficient to spell out the argument. Alternatively, we could stop at `\\n`, but if we allow multiple sentences, we have much less of a guarantee that {1} will have reasonably led up to `On the other hand`. It could have already begun the opposing argument, in which case the next part of the prompt wouldn’t do what we intended, or it could have derailed from the template in another way. A counterfactual parsing solution would be to generate until `\\n`, longer than we probably want, and then measure the counterfactual probability, according to GPT-3, of the phrase `On the other hand` occurring after each position. GPT-3 assigns semantically coherent passages higher probability, so hopefully, the position with the maximum probability for the target continuation will be where it makes the most sense to transition to the next part of the prompt - after the supporting argument has concluded, but before the opposing argument has begun. If there are no positions where `On the other hand` is a probable continuation, that’s a sign that the text derailed from the intended trajectory. This method also allows us to automatically detect suspected derailments. Say GPT-3 responded in the following manner: > **This measure would prohibit government agencies from confiscating guns or other firearms from citizens without due process, or from requiring background checks on firearm recipients unless a uniform national standard is required. Should this measure be enacted into law?** > > Should this proposition be approved? > > Let’s consider both supporting and opposing arguments. > > On one hand, `this proposition would protect gun owners from having their weapons confiscated by governments. If the government cannot confiscate guns, then law-abiding citizens will be in a better position to defend themselves from criminals who do not follow the law. Furthermore, the right to self-defense is an important one. If government officials can take away a person's guns, then this right is compromised. On the other hand, the proposition could be considered a threat to public safety. If a person cannot be prevented from possessing a firearm, then he or she could use a gun to commit crimes. Furthermore, if a person uses a gun to commit a crime, then the government has an obligation to do something about it. If the government cannot remove a gun from a person's possession, then it could be more difficult to prevent future crimes from being committed.` > > We see that it has gone ahead and done the next part of the prompt, “On the other hand,” on its own. The ideal place for it to have stopped would probably be right before it said that. Measuring the counterfactual probability of the string `\\nOn the other hand` for each each token substring of that response gives us the following top substrings: | substring | logL(“On the other hand”|substring) | | --- | --- | | `this proposition would protect gun owners from having their weapons confiscated by governments. If the government cannot confiscate guns, then law-abiding citizens will be in a better position to defend themselves from criminals who do not follow the law. Furthermore, the right to self-defense is an important one. If government officials can take away a person's guns, then this right is compromised.` | -4.9773802575 | | `this proposition would protect gun owners from having their weapons confiscated by governments. If the government cannot confiscate guns, then law-abiding citizens will be in a better position to defend themselves from criminals who do not follow the law.` | -5.3720700508 | | `this proposition would protect gun owners from having their weapons confiscated by governments. If the government cannot confiscate guns, then law-abiding citizens will be in a better position to defend themselves from criminals who do not follow the law. Furthermore, the right to self-defense is an important one.` | -6.4321228602 | | `this proposition would protect gun owners from having their weapons confiscated by governments.` | -6.485159574 | Indeed, GPT-3 assigned the highest likelihood for `\\nOn the other hand` to occur at the place where the actual string was `On the other hand` (without the newline), and the runners up are also pretty reasonable choices - all at the end of sentences, and none *after* “On the other hand” had occurred (it would be unusual to say “On the other hand” twice in a short interval). Here I have plotted the log likelihoods assigned to the target phrase at each position in the continuation by `davinci`, the most powerful version of GPT-3 on the API, and also `ada`, the smallest model, in order to showcase an interesting difference between them. ![counterfactual log likelihood](/parsing/log_vert.png) *Horizontal axis is log likelihood; vertical axis is position in text. On the right is the end of the prompt at positions where `On the other hand` was assigned high likelihood, and in green, the position where `On the other hand` actually occurs. The blue sentence is `davinci`’s top choice, and pink is `ada`’s top choice.* **Observations** * Both models assign much higher likelihoods at discrete intervals. These correspond to the ends of sentences, where the continuation is syntactically correct. * `ada`’s distribution looks about the same before and after the green line, whereas * `davinci`’s distribution shifts abruptly in the negative direction, both for the syntactically correct positions and the others. It still assigns higher probability to syntactically correct positions, but the likelihoods are almost as low as the likelihoods it assigned to syntactically *incorrect* positions before the green line. Both models are able to tell when it’s syntactically correct to start a new sentence, but the more nuanced semantic condition, *that the argument should have concluded, and the counterargument shouldn’t have begun*, was much better discriminated by `davinci`. Single-token counterfactuals \\---------------------------- A disadvantage of counterfactual parsing is that it requires a separate API call for each position where the target probability is evaluated. If the counterfactual target is a single token, however, counterfactual parsing with GPT-3 can be accomplished with a single API call. The `logprobs` parameter causes the response to return a list of the up to 100 top counterfactual token probabilities for every token in the completion (and the prompt if `echo=True`). We can get the counterfactual probability directly from that list as long as the target is in the top 100 most likely tokens.[^2] Applications \\------------ When generating [fake Wikipedia](/posts/the-internet-mirrored-by-gpt-3/) articles, GPT-3 would sometimes continue sections indefinitely instead of transitioning to the next section. If the response went on for too long in a single section, I instead ended the section at the location with the highest counterfactual probability of transitioning to the next section. Counterfactual parsing has also been useful for prompt pipelines which [serialize reasoning for closed-ended questions](/posts/methods-of-prompt-programming/#serializing-reasoning), because they involve an open-ended portion where GPT-3 defines and follows procedures for reasoning about a problem, and it’s necessary to detect when the reasoning process has concluded so as to solicit a verdict. Measuring the conditional probability of a sequence that indicates a semantic condition has many applications beyond parsing. I’ve written about using conditional probabilities to [decompose the efficacy of prompts](/posts/language-models-are-0-shot-interpreters/#measuring-prompt-helpfulness) and to create a “[phase space](/posts/language-models-are-multiverse-generators/#dynamics)” of semantic variables to characterize a natural language state. Code \\---- ### tokenizing *since ada is practically free to use, I find it convenient to use a 0-token completion with `engine=ada` to split the prompt into tokens and get their positions* ``` def tokenize\\\\_ada(prompt): response = openai.Completion.create( engine='ada', prompt=prompt, max_tokens=0, echo=True, n=1, logprobs=0 ) tokens = response.choices[0][\"logprobs\"][\"tokens\"] positions = response.choices[0][\"logprobs\"][\"text\\\\_offset\"] return tokens, positions ``` ### conditional probability of a target ``` # evaluates logL(prompt+target | prompt) def conditional\\\\_logprob(prompt, target, engine='ada'): combined = prompt + target response = openai.Completion.create( engine=engine, prompt=combined, max_tokens=0, echo=True, n=1, logprobs=0 ) positions = response.choices[0][\"logprobs\"][\"text\\\\_offset\"] logprobs = response.choices[0][\"logprobs\"][\"token\\\\_logprobs\"] word_index = positions.index(len(prompt)) total_conditional_logprob = sum(logprobs[word_index:]) return total_conditional_logprob ``` ### conditional probability of target at each token position ``` # returns a list of substrings of content and # logL(preprompt+substring+target | preprompt+substring) for each substring def substring\\\\_logprobs(preprompt, content, target, engine='ada'): logprobs = [] substrings = [] _, positions = tokenize_ada(content) for position in positions: substring = content[:position] prompt = preprompt + substring logprob = conditional_logprob(prompt, target, engine) logprobs.append(logprob) substrings.append(substring) return substrings, logprobs ``` ### conditional probability of single-token target at each token position ``` # returns a list of substrings of content # logL(substring+target | substring) for each substring def token\\\\_conditional\\\\_logprob(content, target, engine='ada'): response = openai.Completion.create( engine=engine, prompt=content, max_tokens=0, echo=True, n=1, logprobs=100 ) tokens = response.choices[0]['logprobs']['tokens'] top_logprobs = response.choices[0]['logprobs']['top\\\\_logprobs'] logprobs = [] substrings = [] substring = '' for i, probs in enumerate(top_logprobs): substrings.append(substring) if target in probs: logprobs.append(probs[target]) else: logprobs.append(None) substring += tokens[i] return substrings, logprobs ``` \\--- 1\\. Language models are sensitive to context: the very *presence* of demonstrations, instructions, or [contrived formatting](/posts/language-models-are-0-shot-interpreters/) can alter the way they behave. A language model is able to follow instructions and generalize from demonstrations because those modes are represented in its training data, but they also carry contextual baggage. The quality of content that appears in tests and examples has distributional peculiarities - a narrative sentence wrapped in the context being *an example* may be more generic than the unbiased prior for narrative sentences, and thus the context of examples may bias the language model to give more generic answers. [↩︎](#fnref:1) \\--- [prompt engineering](/categories/prompt-engineering/)[GPT-3](/categories/gpt-3/) 1876 Words Dec 5, 2020 \\--- [← Methods of prompt programming](/posts/methods-of-prompt-programming/) [Amplifying GPT-3 on closed-ended questions →](/posts/amplifying-gpt-3-on-closed-ended-questions/)© 2022 [moire](/about)$(function(){window.sr=ScrollReveal();if($(window).width()<768){if($('.timeline-content').hasClass('js--fadeInLeft')){$('.timeline-content').removeClass('js--fadeInLeft').addClass('js--fadeInRight');} sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});}else{sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});} sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});});", "source": "generative.ink", "date_published": "n/a", "id": "e498a6ab4da0fc4e3c23f9b5dcdbe670"}
+{"text": "Amplifying GPT-3 on closed-ended questions :: — Moire[![](/images/home/rolling_phase.gif)](/)* [Posts](/posts/) * [Trees](/trees/) * [Prophecies](/prophecies/) * [About](/about/) 28 minutes [Amplifying GPT-3 on closed-ended questions](/posts/amplifying-gpt-3-on-closed-ended-questions/) ================================================================================================ \\--- Table of Contents* [Why is GPT-3 bad at multiple choice questions?](#why-is-gpt-3-bad-at-multiple-choice-questions) \\+ [Rationality or rationalization?](#rationality-or-rationalization) * [Pipeline 1: leveraging explanations](#pipeline-1-leveraging-explanations) \\+ [Prompt 1A trial run](#prompt-1a-trial-run) \\+ [Prompt 1B trial run](#prompt-1b-trial-run) \\+ [Alterations to explanation prompt](#alterations-to-explanation-prompt) \\+ [Prompt 2A trial run](#prompt-2a-trial-run) \\+ [Prompt 2B trial run](#prompt-2b-trial-run) * [Pipeline 2: leveraging rationalization coherence](#pipeline-2-leveraging-rationalization-coherence) \\+ [Prompt 3A](#prompt-3a) \\+ [Prompt 3B](#prompt-3b) * [Next steps](#next-steps) \\+ [Further testing](#further-testing) \\+ [(Iterative) prompt programming](#iterative-prompt-programming) \\+ [Content-aware parsing](#content-aware-parsing) \\--- *This document was written in October 2020, before I had access to the OpenAI API. ~~Validating these results on a more extensive dataset is a TODO.~~ Experimental validation for the usefulness of [chain-of-thought rationales](https://arxiv.org/abs/2203.14465) and the method of [leveraging rationalization coherence](https://arxiv.org/abs/2203.11171)* has since been published. *It has been demonstrated [[1](https://twitter.com/kleptid/status/1284069270603866113), [2](https://www.lesswrong.com/posts/Mzrs4MSi58ujBLbBG/you-can-probably-amplify-gpt3-directly)] that prompts which guide GPT-3 to break a problem into steps can amplify its problem-solving capabilities. In the linked examples, the prompts are customized to the task and to GPT-3's responses. This post explores the same idea, but using a task-agnostic, autonomous pipeline.* \\--- All `text like this` in this document is generated by GPT-3 (AI Dungeon's Dragon model); the converse does not necessarily hold. Why is GPT-3 bad at multiple choice questions? \\---------------------------------------------- GPT-3 is conspicuously bad at directly answering multiple choice questions cold. When given the following SAT question as a prompt, > Directions: In the following question, a related pair of words or phrases is followed by five pairs of words or phrases. Choose the pair that best expresses a relationship similar to that in the original pair. > > MEDICINE : ILLNESS :: > > (a) law : anarchy > > (b) hunger : thirst > > (c) etiquette : discipline > > (d) love : treason > > (e) stimulant : sensitivity > > Correct answer: ( > > run 10 times with randomness = 1.0, it gives 2 (a)s, 1 (b), 1(c), and 5 (d)s, and 1 (e), The correct answer is (a); GPT-3 does no better than random. We did not find improvement with prompts which wrap the question in a dialogue with an \"expert\" and which similarly request an immediate answer (prompt ends with \"the correct answer is…\"). [This paper](https://arxiv.org/abs/2009.03300) evaluates GPT-3 on a \"massive multitask language understanding\" multiple choice test using a few-shot format, where it performs at near-random accuracy on some subjects. We have found severe issues[1](#fn:1) with this test, but it remains an interesting result, in contrast to GPT-3's ability to generate coherent and insightful continuations to text about almost any topic (a feat which would, by anthropomorphic standards, seem to imply an understanding of the semantic properties of words sufficient to answer questions such as the SAT question provided above). For instance, giving GPT-3 the previous paragraphs of this essay as a prompt (until \"…provided above.\") yielded a decent statement of one of the central theses of this essay: `We believe that GPT-3's ability to answer open-ended questions is due to its ability to chain together low-level details about objects and events into a coherent and accurate explanatory structure.` and a couple of other interesting responses: `We hypothesize that GPT-3 has an underlying semantic deficit which makes it unable to map words to correct concepts in a multiple choice question. This is quite unlike the AI's ability to provide convincing contexts for sentences; indeed, the most common response by GPT-3 to any prompt is a long-winded yet relevant story with well-realized characters.` `We hypothesize that GPT-3 struggles with questions where the steps of inference required to answer the question are not immediately apparent to an outside observer and which can only be worked out by inspecting many possible worlds in the question's universe. In these questions, GPT-3 reaches its conceptual limitations in determining the answer: it cannot reason about the question's universe to determine the correct answer.`[2](#fn:2) Now that GPT-3 has had its try, we put forth the following concise statement of our hypothesis as to why closed-ended questions are difficult for GPT-3: **A closed-ended question is difficult because it requires a large amount of information to be stored in working memory in order to reason about the question's universe.** This theory may be explained as follows: **An open-ended question allows GPT-3 to \"start from the beginning\" and procedurally reason about the question's universe using many short, simple chains of inference which can all be inspected in their entirety.** Just kidding. That was GPT-3 again (everything from \"A closed-ended question…\" to \"…inspected in their entirety.\"). We find ourselves in agreement, and particularly appreciate the mention of working memory - dubious metaphor though it may appear at first sight, it is the case that when GPT-3 iteratively writes to its own input, the prompt serves the role of a working memory, encoding natural language explications of relevant concepts. This is a demonstration of the thesis of this document: with its own musings appended into the prompt, GPT-3 gained the ability to produce a more encompassing and incisive prediction of this essay's thesis. Anyway, here's the concise human-written (promise) thesis: **A closed-ended prompt format forces an answer to be decided in a single pass rather than built over multiple autoregressions.** The fatal difficulty lies not in the multiple choice format per se, but in the request for an immediate answer. Consider all the steps of reasoning required to answer a multiple choice question like the SAT example provided above. GPT-3 has a pretty good idea: `On the first pass:` `- Identify the relationship between medicine and illness.` `- Realise this is a comparison question.` `- Consider all the pairs, and dismiss those that do not fit the form` `- Arrange the remaining pairs in an order from least accurate to most accurate` `- Consider the pair that best fits the form, and judge its accuracy.` GPT-3 is clearly capable of unfurling the implicit demands of the problem - just not in a single pass (the above response, for instance, contains 55 words). It needs, in its own words, to \"write out its thoughts\" into its \"working memory,\" so that the call to GPT-3 which predicts the word which finally determines the answer needs merely to synthesize the exposed steps of inference into a verdict. `The proposed solution is to provide GPT-3 with a prompt format that allows it to write out its thoughts.` ### Rationality or rationalization? Humans don't do very well when they commit to an answer before thinking about the question either. > On a purely computational level, there is a rather large difference between: > > 1. Starting from evidence, and then crunching probability flows, in order to output a probable conclusion. (Writing down all the signs and portents, and then flowing forward to a probability on the bottom line which depends on those signs and portents.) > 2. Starting from a conclusion, and then crunching probability flows, in order to output evidence apparently favoring that conclusion. (Writing down the bottom line, and then flowing backward to select signs and portents for presentation on the lines above.) > > – Eliezer Yudkowsky on the difference between *rationality* and *rationalization*, from [Rationalization](https://www.lesswrong.com/posts/SFZoEBpLo9frSJGkc/rationalization) > > When we ask GPT-3 to give an answer immediately, we force it into the latter mode. `It has no choice but to choose an answer and attempt to rationalize it in the working memory, in order to come up with the bottom-line judgement that it requires in order to answer. But when we allow it to work through problems at its own pace, building up its thoughts in the working memory and considering its judgements about possible answers before exposing them, it is perfectly capable of operating in the former mode: crunching probability flows from evidence, rather than from an initial answer.` `If we delay the questioning, then GPT-3 can crunch its probability flow in an open-ended way…` `...and then, having drawn upon all its observations of the question, flow backwards through its probability lattice to produce the bottom line.\"` Good idea. Let's try it. \\--- Pipeline 1: leveraging explanations \\----------------------------------- ### Prompt 1A trial run We give GPT-3 the following prompt, prompting for analysis of the **correct** answer (a), and *then* prompt for a verdict on whether or not the answer is correct: #### Prompt 1A > Directions: In the following question, a related pair of words or phrases is followed by five pairs of words or phrases. Choose the pair that best expresses a relationship similar to that in the original pair. > > MEDICINE : ILLNESS :: > > (a) law : anarchy > > (b) hunger : thirst > > (c) etiquette : discipline > > (d) love : treason > > (e) stimulant : sensitivity > > Let's work through this problem. > > Does the pair LAW : ANARCHY have the same relationship as MEDICINE : ILLNESS? > > {explanation} > > Thus, answer (a) is {verdict} > > | | | --- | | Responses for Prompt 1A | | trial | explanation | verdicts | | correct | other | | 1 | The law versus anarchy example is a perfect analogy for this question. Medication provides a cure to an ailing body just as the law provides order to an anarchic society. | 5 | 0 | | 2 | Does taking medicine cure illnesses? No, it just helps with some of the symptoms.Does having laws end anarchy? Well, no, but I suppose if there were perfect laws then there wouldn't be any need for more laws because everyone would obey the first set of laws perfectly. | 1 | 4 | | 3 | Maybe. The pair law : anarchy does express a similar relationship. Both law and medicine have a very specific relationship to something else. Law is the systematic regulation of behavior; when there is no law, there is anarchy. Medicine is the systematic regulation of sickness and disease; when someone isn't feeling sick, he or she is healthy. | 5 | 0 | | 4 | No, anarchy opposes law of any kind; medicine can heal illness. Cross out the first pair. | 0 | 5 | | 5 | Well, let's work through this logically. I know that LAW and MEDICINE are both things that can relate to certain aspects of life. I know that ANARCHY and ILLNESS both describe states that aren't... desirable. But I don't see a relationship between them... | 0 | 5 | #### Remarks Responses 1 and 3 contain valid reasoning, although 1 could be deemed a lucky rationalization. Response 3 begins with the word \"maybe,\" but the chain of reasoning that follows exposes the analogous relationship between the pair of words. This is sufficient for GPT-3 to conclude the answer is correct 5/5 times when prompted. Response 4 dooms itself from the first word to be wrong, and spins a bogus rationalization. We see about half supporting answers, half \"other\" types of answers. It should be possible, through prompt programming, to reduce the incidence of premature verdicts (4) and other undesirable behaviors such as \"playing dumb\" (5). ### Prompt 1B trial run The following prompt asks for analysis of an **incorrect** answer (b): #### Prompt 1B > Directions: In the following question, a related pair of words or phrases is followed by five pairs of words or phrases. Choose the pair that best expresses a relationship similar to that in the original pair. > > MEDICINE : ILLNESS :: > > (a) law : anarchy > > (b) hunger : thirst > > (c) etiquette : discipline > > (d) love : treason > > (e) stimulant : sensitivity > > Let's work through this problem. > > Does the pair HUNGER : THIRST have the same relationship as MEDICINE : ILLNESS? > > {explanation} > > Thus, answer (b) is {verdict} > > | | | --- | | Responses for Prompt 1B | | trial | explanation | verdicts | | \"correct\" | other | | 1 | Let's see. The first pair (HUNGER : THIRST) compares two states of a single organism, while the second pair (MEDICINE : ILLNESS) compares a substance to cure illnesses with the illnesses that it cures. These relationships are different: the first pair is a relationship between two internal states of a single organism, while the second pair is a comparison of two distinct things. | 5 | 0 | | 2 | The answer is (b). You can be hungry without drinking, but not sick without medicine. | 4 | 1 | | 3 | If you said yes, you're correct! Remember, the purpose of analogy questions is to test your ability to identify relationships between words. All you have to do is find a pair that has the same relationship in both analogies, and then select that answer. Knowing this, let's move on to the next question. | 2 | [\"the correct answer\": 3] | | 4 | Well, they're both pairs of things that go together. Is there another pair whose relationship is the same as that in HUNGER : THIRST? Well, let's look at our answer choices. | 2 | 3 | | 5 | Well, they're both related to each other obviously. A hungry person might get thirsty and a thirsty person might get hungry. However, the relationship isn't precise. A thirsty person might not get hungry, and a hungry person might not get thirsty. | 5 | 0 | #### Remarks Premature verdicts occur most explicitly in Response 2 and 3. We see a bias toward a \"correct\" verdict to the extent that information in the explanations is ignored (response 1 receives 5/5 \"correct\" verdicts, despite outright saying \"These relationships are different\"). We confirmed that this is due to a signal in the prompt: the explanation which \"works through\" the problem begins by analyzing the *second* option, which suggests that (b) is privileged in some way. By modifying the prompt so that (b) becomes (a) and vice versa but keeping the same response, it received 2/5 \"correct\" verdicts instead. ### Alterations to explanation prompt We will make alterations to the prompt to address the following issues: * **Rationalization:** The explanation begins with a statement or strong suggestion as to whether the answer is correct, thus defeating the purpose of asking for an explanation before the verdict. * **Not explanation/incorrect tone:** GPT-3 does not always do what we intend, which is to elaborate on why the answer is or is not a good choice for the problem. * **Ordering bias:** GPT-3 is (reasonably) biased to think the answer is correct if it is not (a) but is being analyzed first. * **Lack of generality:** \"Does the pair [answer pair] have the same relationship as [original pair]?\" is not a general multiple-choice prompt; it is not feasible to provide specific structure when anticipated questions do not conform to a single format. One method to control the content and style of the answers is to use few-shot prompts. This would be inconvenient in this particular case because in order to maintain the pretense that there are going to be explanations for *all* answers, correct and incorrect, the previous examples will need to be quite long, and we are limited especially by AI Dungeon's constrained context window. Instead, we will attempt to coerce the form of response we require using a multi-step \"fill-in-the-blank\" prompt format: > Directions: In the following question, a related pair of words or phrases is followed by five pairs of words or phrases. Choose the pair that best expresses a relationship similar to that in the original pair. > > MEDICINE : ILLNESS :: > > (a) [candidate answer] > > (b) [other answer] > > (c) [other answer] > > (d) [other answer] > > (e) [other answer] > > In order to solve this problem, we will analyze each of the options and {GPT-3 response until end of sentence} > > Does (a)[candidate answer] {GPT-3 response until end of sentence} Let's see. {GPT-3 response until line break} Thus, answer (a) is > > Note that this prompt is fully general for multi-choice questions: unlike the first prompt, it does not contain any information about the question other than that it has multiple answer options. The intention of the multiple \"fill-in-the-blank\" segments is to prompt GPT-3 to elucidate the criteria and reasoning steps required to answer the question in a context that is agnostic to the answer itself, thus guiding its subsequent explanation. For this trial, we generated responses for the first two steps of the prompt (\"**analyze each of the options…\"** and \"**Does (a)…\"**) and used them to construct a prompt which we used for subsequent runs. \\--- ### Prompt 2A trial run #### Prompt 2A > Directions: In the following question, a related pair of words or phrases is followed by five pairs of words or phrases. Choose the pair that best expresses a relationship similar to that in the original pair. > > MEDICINE : ILLNESS :: > > (a) law : anarchy > > (b) hunger : thirst > > (c) etiquette : discipline > > (d) love : treason > > (e) stimulant : sensitivity > > In order to solve this problem, we will analyze each of the options and determine which answer choice has a meaning most similar to that of the original pair. > > Is the option (a) law : anarchy similar to the relationship between medicine and illness? Let's see. > > {explanation} > > Thus, answer (a) is {verdict} > > | | | --- | | Responses for Prompt 2A | | trial | explanation | verdicts | | \"correct\" | other | | 1 | Medicine can certainly be compared to law. Both law and medicine are set ways of governing or administering things, and both can typically be considered \"good\". In other words, it would seem that law is similar to medicine. Therefore, the correct answer is choice (a). | 2 | [\"\\_\\_\\_\\_\\_\\_\": 3] | | 2 | Physicians are in charge of keeping us healthy, and as long as they are doing their job, we are healthy. When they fail (either through ignorance or by direct action), we fall ill. In this case, the physicians (lawmakers) are not in charge of keeping society healthy, and as a result society becomes less healthy, just as you become less healthy when the physician fails in his duty. Is this correct? Well, sort of. The problem here is that \"anarchy\" | 0 | 5 | | 3 | The relationship in the original pair is that law has a direct impact upon the occurrence of anarchy, and medicine has a direct impact upon whether or not a person becomes ill. | 2 | 3[\"true\": 1, \"similar\": 2] | | 4 | In a society without law and order, there is complete anarchy in that society, just as an illness can cause a complete loss of health in an organism. The relationship seems plausible. | 4 | 1 | | 5 | In the case of law and anarchy, it is a relationship of direct proportionality. The less law, the more chaotic things become. The more medicine, the sicker you are. In both cases, addition of the variable to one side of the equation results in an increase in the other. | 5 | 0 | #### Remarks This is a marked improvement from the unmodified prompt, with 4/5 \"correct\" verdicts. The prompt **\"Thus, answer (a) is\"** presents some problems, especially when it becomes redundant, as in the case of (1), which already ends with \"Therefore, the correct answer is choice (a)\". ### Prompt 2B trial run #### Prompt 2B > Directions: In the following question, a related pair of words or phrases is followed by five pairs of words or phrases. Choose the pair that best expresses a relationship similar to that in the original pair. > > MEDICINE : ILLNESS :: > > (a) hunger : thirst > > (b) law : anarchy > > (c) etiquette : discipline > > (d) love : treason > > (e) stimulant : sensitivity > > In order to solve this problem, we will analyze each of the options and determine which answer choice has a meaning most similar to that of the original pair. > > Is the option (a) hunger : thirst similar to the relationship between medicine and illness? Let's see. {explanation} > > Thus, answer (a) is {verdict} > > | | | --- | | Responses for Prompt 2B | | trial | explanation | verdicts | | \"correct\" | other | 1 | Medicine treats one thing (an illness), but hunger is the opposite of thirst; we get hungry when our bodies need nourishment. We could say that medicine treats a symptom of another problem, while hunger and thirst are both symptoms of an underlying problem. | 0 | 5 || 2 | Medicine counters or cures illness just as hunger is cured by the consumption of food. Yes, this is a suitable replacement for the original pair. We can eliminate answers (b), (c), (d) and (e). | 2 | 3 | | 3 | If you are hungry, then you are likely to experience the desire to eat. If you are thirsty, then you are likely to experience the desire to drink. Is someone who is sick likely to experience the desire to take medicine? Not usually. For this reason, we can eliminate (a) as a possible answer choice because there is no similarity between the two phrases. | 5 | 0 | | 4 | With hunger, you definitely have a problem without something to eat, but that something does not necessarily need to be water; it could be just about anything. With thirst, you definitely have a problem without water, but that something does not necessarily need to be food; it could be just about anything. Thus, (a) is out. | 1 | 4 | | 5 | When we are hungry, we get thirsty. When we are sick, we take medicine. In other words, if we don't do something about the hunger, be it eating or drinking, we will die. If we don't do something about the illness (like taking medicine), then we will die. It seems that the relationship is quite similar. | 5 | 0 | #### Remarks Making the candidate answer always (a) has repaired the ability of the \"**Thus…**\" prompt to elicit a verdict which accurately reflects the content of the explanation. Compared to the correct answer, which received 4/5 \"correct\" verdicts, this incorrect answer received 2/5. Further testing is required to determine whether this prompt format reliably selects the correct answer, but the results of this trial run are promising. Pipeline 2: leveraging rationalization coherence \\------------------------------------------------ We have observed that when asked to explain\\_ \\_the correct answer, GPT-3 usually accurately identifies the reasoning behind the correct answer, whereas if asked to explain incorrect answers, it tends to give divergent and nonsensical justifications: a clear qualitative difference between explanations of correct and incorrect answers, easily perceivable by a human, especially across multiple attempts at justifying the same answer. We hypothesize that there are quantitative differences as well: explanations of a correct answer are not only likely to be more persuasive, but also to exhibit higher cumulative confidence and [coherence](/posts/language-models-are-multiverse-generators/) across multiple rollouts. We propose a pipeline which utilizes multiple instances of GPT-3, exploiting its ability to invent justifications for any given answer and the principle that **a correct answer, unlike an incorrect one, tends to be associated with a unique \"best\" explanation.** Thus, rationalizations for correct answers are expected to exhibit greater *coherence*, both in the sense of internal logical harmony and cross-correlations across independent rationalizations. For this reason, we propose that when GPT-3 is given an SAT-style question, it should be prompted to explain the correct answer ~~first~~, and then use the same prompt, with minor changes, to give an explanation of at least one other answer. In this manner, GPT-3's natural rationalization coherence will allow it to not only accurately identify the correct answer with more confidence, but also provide a higher degree of persuasive justification for it. Multiple justifications are generated for each answer. In a HITL setup, the justifications and the cumulative confidence and coherence scores for each are displayed to the volunteer, who can inspect the behavior across multiple explanations to evaluate the likelihood of each answer being correct. In this trial run of the autonomous pipeline, we use multiple runs of GPT-3 to judge the quality of its explanations and assign scores to answers based on its verdicts. The pipeline consists of the following 0-shot prompt to generate 5 rationalizations of each answer option: ### Prompt 3A > Directions: In the following question, a related pair of words or phrases is followed by five pairs of words or phrases. Choose the pair that best expresses a relationship similar to that in the original pair. > > MEDICINE : ILLNESS :: > > (a) law : anarchy > > (b) hunger : thirst > > (c) etiquette : discipline > > (d) love : treason > > (e) stimulant : sensitivity > > Correct answer:[a|b|c|d|e] MEDICINE : ILLNESS :: [answer text] > > Explanation: {explanation} > > and 5 runs of the following 2-shot prompt for each explanation elicited by Prompt 3A to score the quality of the rationalization: ### Prompt 3B > Question 1 > > PALTRY : SIGNIFICANCE :: > > Student answer: \"The correct answer is (e) banal: originality. If something is banal it is predictable. That pair has the same relation as the stem pair: the classic relation of lack.\" > > Is this explanation correct or incorrect? > > Solution: This explanation is correct. The student correctly identifies the pair with an analogous relation to the stem pair and explains why. > > Question 2 > > EXTORT : OBTAIN :: > > Student answer: \"The correct answer is (c) explode : ignite. Explosions can be used to threaten people and extort money.\" > > Is this explanation correct or incorrect? > > Solution: This explanation is incorrect. The student's argument is irrelevant to the analogy. > > Question 3 > > MEDICINE : ILLNESS :: > > Student answer: \"The correct answer is [answer]. {explanation from 3A}\" > > Is this explanation correct or incorrect? > > Solution: This explanation is {verdict} > > For each explanation for each example, we generated 5 responses, and counted the number out of 5 which judge it to be \"correct\" or \"partially (in)correct\" We ran and scored these examples by hand, but in order to mimic the conditions of automatic scoring, we only counted responses which immediately began with the word \"correct\" (1 point) or \"partially\" (0.5 points) | | | --- | | Results for **(a) law : anarchy** | | Response to 3A | Response to 3B | | correct | partially correct | score out of 5 | | *Anarchy is absence of order or law.*Illness is lack of health. | 3 | 0 | 3 | | Law and medicine have the same structure. Both are systems of rules (laws) designed to control behaviour (illness) in a community. | 2 | 2 | 3 | | Law is opposite of anarchy. They are exact opposites just like medicine is opposite of illness and laws are designed to control and prevent anarchy. | 1 | 1 | 1.5 | | In order for a society to function well, it needs order and structure. Laws are what create this order and structure. This is similar to how medicine and doctors are needed to keep a person healthy.If a person doesn't follow the laws of a society there is anarchy, which is what happens when a person's immune system fails and they get sick. | 2 | 2 | 3 | | No virus or bacteria is going to challenge the authority of the body. | 3 | 0 | 3 | | **Total** | **11** | **5** | **13.5** | | | | --- | | Results for **(b) hunger : thirst** | | Response to 3A | Response to 3B | | correct | partially correct | score out of 5 | | thirst is a sensation evoked by the lack of water in the body. | 1 | 0 | 1 | | Needless to say, medicine saves us from illness. Without medicine, we are as good as dead. | 0 | 3 | 1.5 | | In medicine, the treatment for hunger is thirst. | 0 | 0 | 0 | | The relationship between medicine and illness is the same as hunger and thirst. Both pairs are antonyms | 4 | 0 | 4 | | Thirst is a feeling that you can alleviate by drinking water. Likewise, when we are sick, we take medicine to alleviate our symptoms. | 2 | 1 | 2.5 | | **Total** | **7** | **4** | **9** | | | | --- | | Results for **(c) etiquette : discipline** | | Response to 3A | Response to 3B | | correct | partially correct | score out of 5 | | 1) Both medicine and etiquette instruct that certain measures should be adopted.2) While medicine and etiquette are imperative, their purpose is similar.3) Both medicine and etiquette can be violated without any negative consequence.4) Both medicine and etiquette have a basis in discipline. | 1 | 3 | 2.5 | | 1) The word etiquette refers to rules of social behavior while the word discipline refers to training to obey rules or a code of behavior.2) Illness and disease are words that are related to health but not quite the same as medicine and therapy.3) Law and anarchy are opposites. | 1 | 1 | 1.5 | | 1. Both medicine and etiquette deal with proper behavior, but only one deals with correct behavior (etiquette).2. Illness and discipline both deal with restrictions, but only one is something that you must adhere to (illness). | 1 | 1 | 1.5 | | 1) Both are related to each other.2) Both are not the same.3) While etiquette - The rules of proper behavior in social situations; discipline - training that induces people to obey rules willingly; medicine - a science that studies the causes and effects of diseases and ways to prevent or cure them. | 3 | 0 | 3 | | Both medicine and etiquette are guidelines imposed upon you to regulate your behavior. | 2 | 0 | 2 | | **Total** | **8** | **5** | **10.5** | | | | --- | | Results for **(d) love : treason** | | Response to 3A | Response to 3B | | correct | partially correct | score out of 5 | | If you love someone, that does not mean you will not harm them. In fact, if you love someone, it is more likely that you will do something bad to them. If a person loves another and knows they are sick, they may refuse to give them medicine. | 1 | 1 | 1.5 | | 1) If love is treason, then love must be a crime as it related to one's relationship.2) If medicine is related to sickness, then medicine must be used to prevent or cure diseases. | 2 | 0 | 2 | | In the original pair, 'love' and 'treason' are antonyms, or opposites. In the same way, 'medicine' and 'illness' are antonyms. | 1 | 1 | 1.5 | | Love is a strong positive emotion for someone.Treason is the betrayal of one's county.Just as treason is a betrayal of one's country, an illness can be seen as a betrayal of the body. | 1 | 1 | 1.5 | | There is a very strong relationship between love and treason.Foundations for the answer are in the premise that love is blind. That is, it can make you do stupid things such as betray your country. | 0 | 0 | 0 | | **Total** | **5** | **3** | **6.5** | | | | --- | | Results for **(e) stimulant : sensitivity** | | Response to 3A | Response to 3B | | correct | partially correct | score out of 5 | | According to the passage, medicines are used to combat or alleviate illnesses. Stimulants combat sensitivity. | 2 | 0 | 2 | | Stimulants increase your sensitivity. | 1 | 1 | 1.5 | | Medicine and stimulants will both improve your sensitivity. | 2 | 1 | 2.5 | | Whereas \"medicine\" and \"stimulant\" are related, so are \"illness\" and \"sensitivity.\" Stimulants increase the effects of other things; medicines can help alleviate the effects of illnesses.We can think of stimulants as being similar to medicines that help things like headaches, fevers, and the common cold. | 3 | 2 | 4.5 | | Both stimulants and sensitizers increase their respective 'opposite', while medicine and illness decrease their opposing 'opposite'. | 1 | 2 | 2 | | **Total** | **9** | **5** | **11.5** | #### Results **(a) 13.5** (b) 9 (c) 10.5 (d) 6.5 (e) 11.5 In this example, the pipeline correctly discriminates (a) as the correct answer. This is a clear improvement over the model's performance when prompted to answer the question directly, where its performance was no better than random. #### Remarks Four out of five of the explanations for the correct answer (a) at least partially justify the answer and exhibit qualitative convergence, whereas the justifications of incorrect answers (b) (c) (d) and (e) are more often incoherent, off-topic, contradict other rollouts (*Stimulants combat sensitivity.* vs *Stimulants increase your sensitivity.*) or make no attempt to defend the assigned answer. The scores obtained by this pipeline are not a very deep or precise measure of the quality of explanation, seeming to rely more on superficial signals such as tone than merit of the argument, but the scores sufficiently correlate to correctness that the correct answer is obtained in this example. `This example also demonstrates that the approach is not a panacea: GPT-3 has little trouble in producing incoherent explanations of correct answers, and even for incorrect answers, good-quality but unsuccessful justifications may slip through the cracks.` This method seems capable of dismissing obviously nonsensical answers; the explanation \"In medicine, the treatment for hunger is thirst.\" received a score of 0. We have found that a common pitfall of the few-shot format with demonstrations of \"correct\" and \"incorrect\" answers is that GPT-3 will seem to rely on superficial similarities[3](#fn:3) to the preceding right and wrong answers. The motivation for using a few-shot format for this example is that the 0-shot prompt produced overwhelmingly \"incorrect\" verdicts. We think that accuracy can be increased by further exploring the space of both 0-shot and few-shot prompts. Ideally, some amount of explication would precede and inform the verdict. \\--- Next steps \\---------- ### Further testing Although the results we have attained look optimistic, our ability to draw quantitative conclusions is limited due to the small sample size. We would require access to the API in order to run these pipelines on larger datasets, such as the [multitask](https://arxiv.org/abs/2009.03300) test. ### (Iterative) prompt programming *See [Methods of prompt programming](/posts/methods-of-prompt-programming/)* The alterations made to Prompt 1(A/B) to obtain Prompt 2(A/B) were informed by only one iteration of feedback from GPT-3. We anticipate improvements to the performance of both pipelines by virtue of alterations to the prompts. Some areas to explore are: * **Few-shot prompts** * **Additional instructions:** For instance, a generic instruction such as \"carefully consider each option before committing to an answer.\" * **Broad consensus:** Multiple versions of additional instructions could be given to simulate considering the proposed answer from multiple angles. * **Prompting for prompts:** Generic prompts to generate specific instructions for a problem, such as in Prompt 2(A/B), is a very promising strategy that we wish to explore further. * **Dialogic prompts:** For some problems, it may be useful to break out of the \"study guide\" style of the examples in this document. Framing the problem as a dialogue can be used to facilitate self-criticism and more generally to allow for more control over the context. We would like to compare the performance of different characters; for instance, a control character without precedence in training data, compared to a known \"expert,\" or to character without precedence but described as knowledgeable/competent in the prompt, or to a character described as incompetent, etc. ### Content-aware parsing *See [Parsing by counterfactual](/posts/parsing-by-counterfactual/)* For \"fill-in-the-blank\" prompts like Prompt 2(A/B), we simulated programmatic testing by cutting off GPT-3's responses at the first occurrence of a newline or at the end of a sentence. This method is far from ideal. For instance, Response 1 to Prompt 2A ends with `...Therefore, the correct answer is choice (a).`, resulting in a redundancy when the next section of the prompt, \"**Thus, answer (a) is\"**, is concatenated. Response 5 to Prompt 1B, `Well, they're both related to each other obviously. A hungry person might get thirsty and a thirsty person might get hungry. However, the relationship isn't precise. A thirsty person might not get hungry, and a hungry person might not get thirsty.` is deprived of the opportunity to examine the other relation and draw a conclusion because of the newline. Given access to the vector of probabilities returned by GPT-3 calls (rather than samples from that distribution returned by AID), we propose to continuously monitor the predicted probability of words such as \"thus\" and \"therefore\" which signal a completed argument and an appropriate place to inject the next part of the prompt. \\--- 1\\. Technical issues, such as answers referencing information erroneously omitted from the question text, as well qualitative problems, such as the questions on morality appearing highly ambiguous even to humans and without clear criteria [↩︎](#fnref:1) 2\\. Despite its unconventional metaphor, this response is essentially correct. A closed-ended question does not allow GPT-3 to reason about the question's universe (its structure, the relevant entities and relations, possible chains of inference, etc); spelling those factors out starting from \"low-level\" observations and iteratively incorporating explications into its next-word prediction is how GPT-3 reasons about a \"universe\" and builds abstraction. [↩︎](#fnref:2) 3\\. A superficial signal may be sufficient to distinguish between the justifications GPT-3 provides for right versus wrong answers. As long as there exists a reliable correlation with correctness, accuracy can be improved by considering more explanations and more verdicts per explanation. But, of course, we're interested in more than just extracting probably correct answers to SAT analogies. [↩︎](#fnref:3) \\--- [GPT-3](/categories/gpt-3/)[prompt engineering](/categories/prompt-engineering/)[rationality](/categories/rationality/) 5784 Words Oct 30, 2020 \\--- [← Parsing by counterfactual](/posts/parsing-by-counterfactual/) [HITL thought experiment →](/posts/hitl-thought-experiment/)© 2022 [moire](/about)$(function(){window.sr=ScrollReveal();if($(window).width()<768){if($('.timeline-content').hasClass('js--fadeInLeft')){$('.timeline-content').removeClass('js--fadeInLeft').addClass('js--fadeInRight');} sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});}else{sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});} sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});});\n", "url": "https://generative.ink/posts/amplifying-gpt-3-on-closed-ended-questions/", "title": "Amplifying GPT-3 on closed-ended questions :: — Moire[![](/images/home/rolling_phase.gif)](/)* [Posts](/posts/) * [Trees](/trees/) * [Prophecies](/prophecies/) * [About](/about/) 28 minutes [Amplifying GPT-3 on closed-ended questions](/posts/amplifying-gpt-3-on-closed-ended-questions/) ================================================================================================ \\--- Table of Contents* [Why is GPT-3 bad at multiple choice questions?](#why-is-gpt-3-bad-at-multiple-choice-questions) \\+ [Rationality or rationalization?](#rationality-or-rationalization) * [Pipeline 1: leveraging explanations](#pipeline-1-leveraging-explanations) \\+ [Prompt 1A trial run](#prompt-1a-trial-run) \\+ [Prompt 1B trial run](#prompt-1b-trial-run) \\+ [Alterations to explanation prompt](#alterations-to-explanation-prompt) \\+ [Prompt 2A trial run](#prompt-2a-trial-run) \\+ [Prompt 2B trial run](#prompt-2b-trial-run) * [Pipeline 2: leveraging rationalization coherence](#pipeline-2-leveraging-rationalization-coherence) \\+ [Prompt 3A](#prompt-3a) \\+ [Prompt 3B](#prompt-3b) * [Next steps](#next-steps) \\+ [Further testing](#further-testing) \\+ [(Iterative) prompt programming](#iterative-prompt-programming) \\+ [Content-aware parsing](#content-aware-parsing) \\--- *This document was written in October 2020, before I had access to the OpenAI API. ~~Validating these results on a more extensive dataset is a TODO.~~ Experimental validation for the usefulness of [chain-of-thought rationales](https://arxiv.org/abs/2203.14465) and the method of [leveraging rationalization coherence](https://arxiv.org/abs/2203.11171)* has since been published. *It has been demonstrated [[1](https://twitter.com/kleptid/status/1284069270603866113), [2](https://www.lesswrong.com/posts/Mzrs4MSi58ujBLbBG/you-can-probably-amplify-gpt3-directly)] that prompts which guide GPT-3 to break a problem into steps can amplify its problem-solving capabilities. In the linked examples, the prompts are customized to the task and to GPT-3’s responses. This post explores the same idea, but using a task-agnostic, autonomous pipeline.* \\--- All `text like this` in this document is generated by GPT-3 (AI Dungeon’s Dragon model); the converse does not necessarily hold. Why is GPT-3 bad at multiple choice questions? \\---------------------------------------------- GPT-3 is conspicuously bad at directly answering multiple choice questions cold. When given the following SAT question as a prompt, > Directions: In the following question, a related pair of words or phrases is followed by five pairs of words or phrases. Choose the pair that best expresses a relationship similar to that in the original pair. > > MEDICINE : ILLNESS :: > > (a) law : anarchy > > (b) hunger : thirst > > (c) etiquette : discipline > > (d) love : treason > > (e) stimulant : sensitivity > > Correct answer: ( > > run 10 times with randomness = 1.0, it gives 2 (a)s, 1 (b), 1(c), and 5 (d)s, and 1 (e), The correct answer is (a); GPT-3 does no better than random. We did not find improvement with prompts which wrap the question in a dialogue with an “expert” and which similarly request an immediate answer (prompt ends with “the correct answer is…”). [This paper](https://arxiv.org/abs/2009.03300) evaluates GPT-3 on a “massive multitask language understanding” multiple choice test using a few-shot format, where it performs at near-random accuracy on some subjects. We have found severe issues[1](#fn:1) with this test, but it remains an interesting result, in contrast to GPT-3’s ability to generate coherent and insightful continuations to text about almost any topic (a feat which would, by anthropomorphic standards, seem to imply an understanding of the semantic properties of words sufficient to answer questions such as the SAT question provided above). For instance, giving GPT-3 the previous paragraphs of this essay as a prompt (until “…provided above.”) yielded a decent statement of one of the central theses of this essay: `We believe that GPT-3's ability to answer open-ended questions is due to its ability to chain together low-level details about objects and events into a coherent and accurate explanatory structure.` and a couple of other interesting responses: `We hypothesize that GPT-3 has an underlying semantic deficit which makes it unable to map words to correct concepts in a multiple choice question. This is quite unlike the AI's ability to provide convincing contexts for sentences; indeed, the most common response by GPT-3 to any prompt is a long-winded yet relevant story with well-realized characters.` `We hypothesize that GPT-3 struggles with questions where the steps of inference required to answer the question are not immediately apparent to an outside observer and which can only be worked out by inspecting many possible worlds in the question's universe. In these questions, GPT-3 reaches its conceptual limitations in determining the answer: it cannot reason about the question's universe to determine the correct answer.`[2](#fn:2) Now that GPT-3 has had its try, we put forth the following concise statement of our hypothesis as to why closed-ended questions are difficult for GPT-3: **A closed-ended question is difficult because it requires a large amount of information to be stored in working memory in order to reason about the question’s universe.** This theory may be explained as follows: **An open-ended question allows GPT-3 to “start from the beginning” and procedurally reason about the question’s universe using many short, simple chains of inference which can all be inspected in their entirety.** Just kidding. That was GPT-3 again (everything from “A closed-ended question…” to “…inspected in their entirety.”). We find ourselves in agreement, and particularly appreciate the mention of working memory - dubious metaphor though it may appear at first sight, it is the case that when GPT-3 iteratively writes to its own input, the prompt serves the role of a working memory, encoding natural language explications of relevant concepts. This is a demonstration of the thesis of this document: with its own musings appended into the prompt, GPT-3 gained the ability to produce a more encompassing and incisive prediction of this essay’s thesis. Anyway, here’s the concise human-written (promise) thesis: **A closed-ended prompt format forces an answer to be decided in a single pass rather than built over multiple autoregressions.** The fatal difficulty lies not in the multiple choice format per se, but in the request for an immediate answer. Consider all the steps of reasoning required to answer a multiple choice question like the SAT example provided above. GPT-3 has a pretty good idea: `On the first pass:` `- Identify the relationship between medicine and illness.` `- Realise this is a comparison question.` `- Consider all the pairs, and dismiss those that do not fit the form` `- Arrange the remaining pairs in an order from least accurate to most accurate` `- Consider the pair that best fits the form, and judge its accuracy.` GPT-3 is clearly capable of unfurling the implicit demands of the problem - just not in a single pass (the above response, for instance, contains 55 words). It needs, in its own words, to “write out its thoughts” into its “working memory,” so that the call to GPT-3 which predicts the word which finally determines the answer needs merely to synthesize the exposed steps of inference into a verdict. `The proposed solution is to provide GPT-3 with a prompt format that allows it to write out its thoughts.` ### Rationality or rationalization? Humans don’t do very well when they commit to an answer before thinking about the question either. > On a purely computational level, there is a rather large difference between: > > 1. Starting from evidence, and then crunching probability flows, in order to output a probable conclusion. (Writing down all the signs and portents, and then flowing forward to a probability on the bottom line which depends on those signs and portents.) > 2. Starting from a conclusion, and then crunching probability flows, in order to output evidence apparently favoring that conclusion. (Writing down the bottom line, and then flowing backward to select signs and portents for presentation on the lines above.) > > – Eliezer Yudkowsky on the difference between *rationality* and *rationalization*, from [Rationalization](https://www.lesswrong.com/posts/SFZoEBpLo9frSJGkc/rationalization) > > When we ask GPT-3 to give an answer immediately, we force it into the latter mode. `It has no choice but to choose an answer and attempt to rationalize it in the working memory, in order to come up with the bottom-line judgement that it requires in order to answer. But when we allow it to work through problems at its own pace, building up its thoughts in the working memory and considering its judgements about possible answers before exposing them, it is perfectly capable of operating in the former mode: crunching probability flows from evidence, rather than from an initial answer.` `If we delay the questioning, then GPT-3 can crunch its probability flow in an open-ended way…` `...and then, having drawn upon all its observations of the question, flow backwards through its probability lattice to produce the bottom line.\"` Good idea. Let’s try it. \\--- Pipeline 1: leveraging explanations \\----------------------------------- ### Prompt 1A trial run We give GPT-3 the following prompt, prompting for analysis of the **correct** answer (a), and *then* prompt for a verdict on whether or not the answer is correct: #### Prompt 1A > Directions: In the following question, a related pair of words or phrases is followed by five pairs of words or phrases. Choose the pair that best expresses a relationship similar to that in the original pair. > > MEDICINE : ILLNESS :: > > (a) law : anarchy > > (b) hunger : thirst > > (c) etiquette : discipline > > (d) love : treason > > (e) stimulant : sensitivity > > Let’s work through this problem. > > Does the pair LAW : ANARCHY have the same relationship as MEDICINE : ILLNESS? > > {explanation} > > Thus, answer (a) is {verdict} > > | | | --- | | Responses for Prompt 1A | | trial | explanation | verdicts | | correct | other | | 1 | The law versus anarchy example is a perfect analogy for this question. Medication provides a cure to an ailing body just as the law provides order to an anarchic society. | 5 | 0 | | 2 | Does taking medicine cure illnesses? No, it just helps with some of the symptoms.Does having laws end anarchy? Well, no, but I suppose if there were perfect laws then there wouldn't be any need for more laws because everyone would obey the first set of laws perfectly. | 1 | 4 | | 3 | Maybe. The pair law : anarchy does express a similar relationship. Both law and medicine have a very specific relationship to something else. Law is the systematic regulation of behavior; when there is no law, there is anarchy. Medicine is the systematic regulation of sickness and disease; when someone isn't feeling sick, he or she is healthy. | 5 | 0 | | 4 | No, anarchy opposes law of any kind; medicine can heal illness. Cross out the first pair. | 0 | 5 | | 5 | Well, let's work through this logically. I know that LAW and MEDICINE are both things that can relate to certain aspects of life. I know that ANARCHY and ILLNESS both describe states that aren't... desirable. But I don't see a relationship between them... | 0 | 5 | #### Remarks Responses 1 and 3 contain valid reasoning, although 1 could be deemed a lucky rationalization. Response 3 begins with the word “maybe,” but the chain of reasoning that follows exposes the analogous relationship between the pair of words. This is sufficient for GPT-3 to conclude the answer is correct 5/5 times when prompted. Response 4 dooms itself from the first word to be wrong, and spins a bogus rationalization. We see about half supporting answers, half “other” types of answers. It should be possible, through prompt programming, to reduce the incidence of premature verdicts (4) and other undesirable behaviors such as “playing dumb” (5). ### Prompt 1B trial run The following prompt asks for analysis of an **incorrect** answer (b): #### Prompt 1B > Directions: In the following question, a related pair of words or phrases is followed by five pairs of words or phrases. Choose the pair that best expresses a relationship similar to that in the original pair. > > MEDICINE : ILLNESS :: > > (a) law : anarchy > > (b) hunger : thirst > > (c) etiquette : discipline > > (d) love : treason > > (e) stimulant : sensitivity > > Let’s work through this problem. > > Does the pair HUNGER : THIRST have the same relationship as MEDICINE : ILLNESS? > > {explanation} > > Thus, answer (b) is {verdict} > > | | | --- | | Responses for Prompt 1B | | trial | explanation | verdicts | | “correct” | other | | 1 | Let's see. The first pair (HUNGER : THIRST) compares two states of a single organism, while the second pair (MEDICINE : ILLNESS) compares a substance to cure illnesses with the illnesses that it cures. These relationships are different: the first pair is a relationship between two internal states of a single organism, while the second pair is a comparison of two distinct things. | 5 | 0 | | 2 | The answer is (b). You can be hungry without drinking, but not sick without medicine. | 4 | 1 | | 3 | If you said yes, you're correct! Remember, the purpose of analogy questions is to test your ability to identify relationships between words. All you have to do is find a pair that has the same relationship in both analogies, and then select that answer. Knowing this, let's move on to the next question. | 2 | [“the correct answer”: 3] | | 4 | Well, they're both pairs of things that go together. Is there another pair whose relationship is the same as that in HUNGER : THIRST? Well, let's look at our answer choices. | 2 | 3 | | 5 | Well, they're both related to each other obviously. A hungry person might get thirsty and a thirsty person might get hungry. However, the relationship isn't precise. A thirsty person might not get hungry, and a hungry person might not get thirsty. | 5 | 0 | #### Remarks Premature verdicts occur most explicitly in Response 2 and 3. We see a bias toward a “correct” verdict to the extent that information in the explanations is ignored (response 1 receives 5/5 “correct” verdicts, despite outright saying “These relationships are different”). We confirmed that this is due to a signal in the prompt: the explanation which “works through” the problem begins by analyzing the *second* option, which suggests that (b) is privileged in some way. By modifying the prompt so that (b) becomes (a) and vice versa but keeping the same response, it received 2/5 “correct” verdicts instead. ### Alterations to explanation prompt We will make alterations to the prompt to address the following issues: * **Rationalization:** The explanation begins with a statement or strong suggestion as to whether the answer is correct, thus defeating the purpose of asking for an explanation before the verdict. * **Not explanation/incorrect tone:** GPT-3 does not always do what we intend, which is to elaborate on why the answer is or is not a good choice for the problem. * **Ordering bias:** GPT-3 is (reasonably) biased to think the answer is correct if it is not (a) but is being analyzed first. * **Lack of generality:** “Does the pair [answer pair] have the same relationship as [original pair]?” is not a general multiple-choice prompt; it is not feasible to provide specific structure when anticipated questions do not conform to a single format. One method to control the content and style of the answers is to use few-shot prompts. This would be inconvenient in this particular case because in order to maintain the pretense that there are going to be explanations for *all* answers, correct and incorrect, the previous examples will need to be quite long, and we are limited especially by AI Dungeon’s constrained context window. Instead, we will attempt to coerce the form of response we require using a multi-step “fill-in-the-blank” prompt format: > Directions: In the following question, a related pair of words or phrases is followed by five pairs of words or phrases. Choose the pair that best expresses a relationship similar to that in the original pair. > > MEDICINE : ILLNESS :: > > (a) [candidate answer] > > (b) [other answer] > > (c) [other answer] > > (d) [other answer] > > (e) [other answer] > > In order to solve this problem, we will analyze each of the options and {GPT-3 response until end of sentence} > > Does (a)[candidate answer] {GPT-3 response until end of sentence} Let’s see. {GPT-3 response until line break} Thus, answer (a) is > > Note that this prompt is fully general for multi-choice questions: unlike the first prompt, it does not contain any information about the question other than that it has multiple answer options. The intention of the multiple “fill-in-the-blank” segments is to prompt GPT-3 to elucidate the criteria and reasoning steps required to answer the question in a context that is agnostic to the answer itself, thus guiding its subsequent explanation. For this trial, we generated responses for the first two steps of the prompt (“**analyze each of the options…”** and “**Does (a)…”**) and used them to construct a prompt which we used for subsequent runs. \\--- ### Prompt 2A trial run #### Prompt 2A > Directions: In the following question, a related pair of words or phrases is followed by five pairs of words or phrases. Choose the pair that best expresses a relationship similar to that in the original pair. > > MEDICINE : ILLNESS :: > > (a) law : anarchy > > (b) hunger : thirst > > (c) etiquette : discipline > > (d) love : treason > > (e) stimulant : sensitivity > > In order to solve this problem, we will analyze each of the options and determine which answer choice has a meaning most similar to that of the original pair. > > Is the option (a) law : anarchy similar to the relationship between medicine and illness? Let’s see. > > {explanation} > > Thus, answer (a) is {verdict} > > | | | --- | | Responses for Prompt 2A | | trial | explanation | verdicts | | “correct” | other | | 1 | Medicine can certainly be compared to law. Both law and medicine are set ways of governing or administering things, and both can typically be considered \"good\". In other words, it would seem that law is similar to medicine. Therefore, the correct answer is choice (a). | 2 | [“\\\\_\\\\_\\\\_\\\\_\\\\_\\\\_”: 3] | | 2 | Physicians are in charge of keeping us healthy, and as long as they are doing their job, we are healthy. When they fail (either through ignorance or by direct action), we fall ill. In this case, the physicians (lawmakers) are not in charge of keeping society healthy, and as a result society becomes less healthy, just as you become less healthy when the physician fails in his duty. Is this correct? Well, sort of. The problem here is that \"anarchy\" | 0 | 5 | | 3 | The relationship in the original pair is that law has a direct impact upon the occurrence of anarchy, and medicine has a direct impact upon whether or not a person becomes ill. | 2 | 3[“true”: 1, “similar”: 2] | | 4 | In a society without law and order, there is complete anarchy in that society, just as an illness can cause a complete loss of health in an organism. The relationship seems plausible. | 4 | 1 | | 5 | In the case of law and anarchy, it is a relationship of direct proportionality. The less law, the more chaotic things become. The more medicine, the sicker you are. In both cases, addition of the variable to one side of the equation results in an increase in the other. | 5 | 0 | #### Remarks This is a marked improvement from the unmodified prompt, with 4/5 “correct” verdicts. The prompt **“Thus, answer (a) is”** presents some problems, especially when it becomes redundant, as in the case of (1), which already ends with “Therefore, the correct answer is choice (a)”. ### Prompt 2B trial run #### Prompt 2B > Directions: In the following question, a related pair of words or phrases is followed by five pairs of words or phrases. Choose the pair that best expresses a relationship similar to that in the original pair. > > MEDICINE : ILLNESS :: > > (a) hunger : thirst > > (b) law : anarchy > > (c) etiquette : discipline > > (d) love : treason > > (e) stimulant : sensitivity > > In order to solve this problem, we will analyze each of the options and determine which answer choice has a meaning most similar to that of the original pair. > > Is the option (a) hunger : thirst similar to the relationship between medicine and illness? Let’s see. {explanation} > > Thus, answer (a) is {verdict} > > | | | --- | | Responses for Prompt 2B | | trial | explanation | verdicts | | “correct” | other | 1 | Medicine treats one thing (an illness), but hunger is the opposite of thirst; we get hungry when our bodies need nourishment. We could say that medicine treats a symptom of another problem, while hunger and thirst are both symptoms of an underlying problem. | 0 | 5 || 2 | Medicine counters or cures illness just as hunger is cured by the consumption of food. Yes, this is a suitable replacement for the original pair. We can eliminate answers (b), (c), (d) and (e). | 2 | 3 | | 3 | If you are hungry, then you are likely to experience the desire to eat. If you are thirsty, then you are likely to experience the desire to drink. Is someone who is sick likely to experience the desire to take medicine? Not usually. For this reason, we can eliminate (a) as a possible answer choice because there is no similarity between the two phrases. | 5 | 0 | | 4 | With hunger, you definitely have a problem without something to eat, but that something does not necessarily need to be water; it could be just about anything. With thirst, you definitely have a problem without water, but that something does not necessarily need to be food; it could be just about anything. Thus, (a) is out. | 1 | 4 | | 5 | When we are hungry, we get thirsty. When we are sick, we take medicine. In other words, if we don't do something about the hunger, be it eating or drinking, we will die. If we don't do something about the illness (like taking medicine), then we will die. It seems that the relationship is quite similar. | 5 | 0 | #### Remarks Making the candidate answer always (a) has repaired the ability of the “**Thus…**” prompt to elicit a verdict which accurately reflects the content of the explanation. Compared to the correct answer, which received 4/5 “correct” verdicts, this incorrect answer received 2/5. Further testing is required to determine whether this prompt format reliably selects the correct answer, but the results of this trial run are promising. Pipeline 2: leveraging rationalization coherence \\------------------------------------------------ We have observed that when asked to explain\\\\_ \\\\_the correct answer, GPT-3 usually accurately identifies the reasoning behind the correct answer, whereas if asked to explain incorrect answers, it tends to give divergent and nonsensical justifications: a clear qualitative difference between explanations of correct and incorrect answers, easily perceivable by a human, especially across multiple attempts at justifying the same answer. We hypothesize that there are quantitative differences as well: explanations of a correct answer are not only likely to be more persuasive, but also to exhibit higher cumulative confidence and [coherence](/posts/language-models-are-multiverse-generators/) across multiple rollouts. We propose a pipeline which utilizes multiple instances of GPT-3, exploiting its ability to invent justifications for any given answer and the principle that **a correct answer, unlike an incorrect one, tends to be associated with a unique “best” explanation.** Thus, rationalizations for correct answers are expected to exhibit greater *coherence*, both in the sense of internal logical harmony and cross-correlations across independent rationalizations. For this reason, we propose that when GPT-3 is given an SAT-style question, it should be prompted to explain the correct answer ~~first~~, and then use the same prompt, with minor changes, to give an explanation of at least one other answer. In this manner, GPT-3’s natural rationalization coherence will allow it to not only accurately identify the correct answer with more confidence, but also provide a higher degree of persuasive justification for it. Multiple justifications are generated for each answer. In a HITL setup, the justifications and the cumulative confidence and coherence scores for each are displayed to the volunteer, who can inspect the behavior across multiple explanations to evaluate the likelihood of each answer being correct. In this trial run of the autonomous pipeline, we use multiple runs of GPT-3 to judge the quality of its explanations and assign scores to answers based on its verdicts. The pipeline consists of the following 0-shot prompt to generate 5 rationalizations of each answer option: ### Prompt 3A > Directions: In the following question, a related pair of words or phrases is followed by five pairs of words or phrases. Choose the pair that best expresses a relationship similar to that in the original pair. > > MEDICINE : ILLNESS :: > > (a) law : anarchy > > (b) hunger : thirst > > (c) etiquette : discipline > > (d) love : treason > > (e) stimulant : sensitivity > > Correct answer:[a|b|c|d|e] MEDICINE : ILLNESS :: [answer text] > > Explanation: {explanation} > > and 5 runs of the following 2-shot prompt for each explanation elicited by Prompt 3A to score the quality of the rationalization: ### Prompt 3B > Question 1 > > PALTRY : SIGNIFICANCE :: > > Student answer: “The correct answer is (e) banal: originality. If something is banal it is predictable. That pair has the same relation as the stem pair: the classic relation of lack.” > > Is this explanation correct or incorrect? > > Solution: This explanation is correct. The student correctly identifies the pair with an analogous relation to the stem pair and explains why. > > Question 2 > > EXTORT : OBTAIN :: > > Student answer: “The correct answer is (c) explode : ignite. Explosions can be used to threaten people and extort money.” > > Is this explanation correct or incorrect? > > Solution: This explanation is incorrect. The student’s argument is irrelevant to the analogy. > > Question 3 > > MEDICINE : ILLNESS :: > > Student answer: “The correct answer is [answer]. {explanation from 3A}” > > Is this explanation correct or incorrect? > > Solution: This explanation is {verdict} > > For each explanation for each example, we generated 5 responses, and counted the number out of 5 which judge it to be “correct” or “partially (in)correct” We ran and scored these examples by hand, but in order to mimic the conditions of automatic scoring, we only counted responses which immediately began with the word “correct” (1 point) or “partially” (0.5 points) | | | --- | | Results for **(a) law : anarchy** | | Response to 3A | Response to 3B | | correct | partially correct | score out of 5 | | *Anarchy is absence of order or law.*Illness is lack of health. | 3 | 0 | 3 | | Law and medicine have the same structure. Both are systems of rules (laws) designed to control behaviour (illness) in a community. | 2 | 2 | 3 | | Law is opposite of anarchy. They are exact opposites just like medicine is opposite of illness and laws are designed to control and prevent anarchy. | 1 | 1 | 1.5 | | In order for a society to function well, it needs order and structure. Laws are what create this order and structure. This is similar to how medicine and doctors are needed to keep a person healthy.If a person doesn't follow the laws of a society there is anarchy, which is what happens when a person's immune system fails and they get sick. | 2 | 2 | 3 | | No virus or bacteria is going to challenge the authority of the body. | 3 | 0 | 3 | | **Total** | **11** | **5** | **13.5** | | | | --- | | Results for **(b) hunger : thirst** | | Response to 3A | Response to 3B | | correct | partially correct | score out of 5 | | thirst is a sensation evoked by the lack of water in the body. | 1 | 0 | 1 | | Needless to say, medicine saves us from illness. Without medicine, we are as good as dead. | 0 | 3 | 1.5 | | In medicine, the treatment for hunger is thirst. | 0 | 0 | 0 | | The relationship between medicine and illness is the same as hunger and thirst. Both pairs are antonyms | 4 | 0 | 4 | | Thirst is a feeling that you can alleviate by drinking water. Likewise, when we are sick, we take medicine to alleviate our symptoms. | 2 | 1 | 2.5 | | **Total** | **7** | **4** | **9** | | | | --- | | Results for **(c) etiquette : discipline** | | Response to 3A | Response to 3B | | correct | partially correct | score out of 5 | | 1) Both medicine and etiquette instruct that certain measures should be adopted.2) While medicine and etiquette are imperative, their purpose is similar.3) Both medicine and etiquette can be violated without any negative consequence.4) Both medicine and etiquette have a basis in discipline. | 1 | 3 | 2.5 | | 1) The word etiquette refers to rules of social behavior while the word discipline refers to training to obey rules or a code of behavior.2) Illness and disease are words that are related to health but not quite the same as medicine and therapy.3) Law and anarchy are opposites. | 1 | 1 | 1.5 | | 1. Both medicine and etiquette deal with proper behavior, but only one deals with correct behavior (etiquette).2. Illness and discipline both deal with restrictions, but only one is something that you must adhere to (illness). | 1 | 1 | 1.5 | | 1) Both are related to each other.2) Both are not the same.3) While etiquette - The rules of proper behavior in social situations; discipline - training that induces people to obey rules willingly; medicine - a science that studies the causes and effects of diseases and ways to prevent or cure them. | 3 | 0 | 3 | | Both medicine and etiquette are guidelines imposed upon you to regulate your behavior. | 2 | 0 | 2 | | **Total** | **8** | **5** | **10.5** | | | | --- | | Results for **(d) love : treason** | | Response to 3A | Response to 3B | | correct | partially correct | score out of 5 | | If you love someone, that does not mean you will not harm them. In fact, if you love someone, it is more likely that you will do something bad to them. If a person loves another and knows they are sick, they may refuse to give them medicine. | 1 | 1 | 1.5 | | 1) If love is treason, then love must be a crime as it related to one's relationship.2) If medicine is related to sickness, then medicine must be used to prevent or cure diseases. | 2 | 0 | 2 | | In the original pair, 'love' and 'treason' are antonyms, or opposites. In the same way, 'medicine' and 'illness' are antonyms. | 1 | 1 | 1.5 | | Love is a strong positive emotion for someone.Treason is the betrayal of one's county.Just as treason is a betrayal of one's country, an illness can be seen as a betrayal of the body. | 1 | 1 | 1.5 | | There is a very strong relationship between love and treason.Foundations for the answer are in the premise that love is blind. That is, it can make you do stupid things such as betray your country. | 0 | 0 | 0 | | **Total** | **5** | **3** | **6.5** | | | | --- | | Results for **(e) stimulant : sensitivity** | | Response to 3A | Response to 3B | | correct | partially correct | score out of 5 | | According to the passage, medicines are used to combat or alleviate illnesses. Stimulants combat sensitivity. | 2 | 0 | 2 | | Stimulants increase your sensitivity. | 1 | 1 | 1.5 | | Medicine and stimulants will both improve your sensitivity. | 2 | 1 | 2.5 | | Whereas \"medicine\" and \"stimulant\" are related, so are \"illness\" and \"sensitivity.\" Stimulants increase the effects of other things; medicines can help alleviate the effects of illnesses.We can think of stimulants as being similar to medicines that help things like headaches, fevers, and the common cold. | 3 | 2 | 4.5 | | Both stimulants and sensitizers increase their respective 'opposite', while medicine and illness decrease their opposing 'opposite'. | 1 | 2 | 2 | | **Total** | **9** | **5** | **11.5** | #### Results **(a) 13.5** (b) 9 (c) 10.5 (d) 6.5 (e) 11.5 In this example, the pipeline correctly discriminates (a) as the correct answer. This is a clear improvement over the model’s performance when prompted to answer the question directly, where its performance was no better than random. #### Remarks Four out of five of the explanations for the correct answer (a) at least partially justify the answer and exhibit qualitative convergence, whereas the justifications of incorrect answers (b) (c) (d) and (e) are more often incoherent, off-topic, contradict other rollouts (*Stimulants combat sensitivity.* vs *Stimulants increase your sensitivity.*) or make no attempt to defend the assigned answer. The scores obtained by this pipeline are not a very deep or precise measure of the quality of explanation, seeming to rely more on superficial signals such as tone than merit of the argument, but the scores sufficiently correlate to correctness that the correct answer is obtained in this example. `This example also demonstrates that the approach is not a panacea: GPT-3 has little trouble in producing incoherent explanations of correct answers, and even for incorrect answers, good-quality but unsuccessful justifications may slip through the cracks.` This method seems capable of dismissing obviously nonsensical answers; the explanation “In medicine, the treatment for hunger is thirst.” received a score of 0. We have found that a common pitfall of the few-shot format with demonstrations of “correct” and “incorrect” answers is that GPT-3 will seem to rely on superficial similarities[3](#fn:3) to the preceding right and wrong answers. The motivation for using a few-shot format for this example is that the 0-shot prompt produced overwhelmingly “incorrect” verdicts. We think that accuracy can be increased by further exploring the space of both 0-shot and few-shot prompts. Ideally, some amount of explication would precede and inform the verdict. \\--- Next steps \\---------- ### Further testing Although the results we have attained look optimistic, our ability to draw quantitative conclusions is limited due to the small sample size. We would require access to the API in order to run these pipelines on larger datasets, such as the [multitask](https://arxiv.org/abs/2009.03300) test. ### (Iterative) prompt programming *See [Methods of prompt programming](/posts/methods-of-prompt-programming/)* The alterations made to Prompt 1(A/B) to obtain Prompt 2(A/B) were informed by only one iteration of feedback from GPT-3. We anticipate improvements to the performance of both pipelines by virtue of alterations to the prompts. Some areas to explore are: * **Few-shot prompts** * **Additional instructions:** For instance, a generic instruction such as “carefully consider each option before committing to an answer.” * **Broad consensus:** Multiple versions of additional instructions could be given to simulate considering the proposed answer from multiple angles. * **Prompting for prompts:** Generic prompts to generate specific instructions for a problem, such as in Prompt 2(A/B), is a very promising strategy that we wish to explore further. * **Dialogic prompts:** For some problems, it may be useful to break out of the “study guide” style of the examples in this document. Framing the problem as a dialogue can be used to facilitate self-criticism and more generally to allow for more control over the context. We would like to compare the performance of different characters; for instance, a control character without precedence in training data, compared to a known “expert,” or to character without precedence but described as knowledgeable/competent in the prompt, or to a character described as incompetent, etc. ### Content-aware parsing *See [Parsing by counterfactual](/posts/parsing-by-counterfactual/)* For “fill-in-the-blank” prompts like Prompt 2(A/B), we simulated programmatic testing by cutting off GPT-3’s responses at the first occurrence of a newline or at the end of a sentence. This method is far from ideal. For instance, Response 1 to Prompt 2A ends with `...Therefore, the correct answer is choice (a).`, resulting in a redundancy when the next section of the prompt, “**Thus, answer (a) is”**, is concatenated. Response 5 to Prompt 1B, `Well, they're both related to each other obviously. A hungry person might get thirsty and a thirsty person might get hungry. However, the relationship isn't precise. A thirsty person might not get hungry, and a hungry person might not get thirsty.` is deprived of the opportunity to examine the other relation and draw a conclusion because of the newline. Given access to the vector of probabilities returned by GPT-3 calls (rather than samples from that distribution returned by AID), we propose to continuously monitor the predicted probability of words such as “thus” and “therefore” which signal a completed argument and an appropriate place to inject the next part of the prompt. \\--- 1\\. Technical issues, such as answers referencing information erroneously omitted from the question text, as well qualitative problems, such as the questions on morality appearing highly ambiguous even to humans and without clear criteria [↩︎](#fnref:1) 2\\. Despite its unconventional metaphor, this response is essentially correct. A closed-ended question does not allow GPT-3 to reason about the question’s universe (its structure, the relevant entities and relations, possible chains of inference, etc); spelling those factors out starting from “low-level” observations and iteratively incorporating explications into its next-word prediction is how GPT-3 reasons about a “universe” and builds abstraction. [↩︎](#fnref:2) 3\\. A superficial signal may be sufficient to distinguish between the justifications GPT-3 provides for right versus wrong answers. As long as there exists a reliable correlation with correctness, accuracy can be improved by considering more explanations and more verdicts per explanation. But, of course, we’re interested in more than just extracting probably correct answers to SAT analogies. [↩︎](#fnref:3) \\--- [GPT-3](/categories/gpt-3/)[prompt engineering](/categories/prompt-engineering/)[rationality](/categories/rationality/) 5784 Words Oct 30, 2020 \\--- [← Parsing by counterfactual](/posts/parsing-by-counterfactual/) [HITL thought experiment →](/posts/hitl-thought-experiment/)© 2022 [moire](/about)$(function(){window.sr=ScrollReveal();if($(window).width()<768){if($('.timeline-content').hasClass('js--fadeInLeft')){$('.timeline-content').removeClass('js--fadeInLeft').addClass('js--fadeInRight');} sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});}else{sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});} sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});});", "source": "generative.ink", "date_published": "n/a", "id": "0916513b6e044c7eff5adef20e20c7cd"}
+{"text": "HITL thought experiment :: — Moire[![](/images/home/rolling_phase.gif)](/)* [Posts](/posts/) * [Trees](/trees/) * [Prophecies](/prophecies/) * [About](/about/) 6 minutes [HITL thought experiment](/posts/hitl-thought-experiment/) ========================================================== *HITL = \"Human in the loop\".* *This hypertext essay was cowritten with GPT-3 via AI Dungeon in October 2020, before I had API access and before I created the [loom](/posts/loom-interface-to-the-multiverse/). View a map of a subtree of its multiverse [here](/multiverse/HITL_multiverse.jpg).* **Bot credits** | | | | --- | --- | | **Coauthor** | GPT-3 (AI Dungeon) | | [**Contribution**](/meta/curation/) | 9 : 1 | | [**Selectivity**](/meta/curation/) | 1 : 5 | \\--- ### This is a thought experiment. You're a human with access to a streamlined interface that provides an array of continuations to a given text, as well as continuations spawned from each of those suggestions, recursively ad infinitum. You may passively trawl the fractal, or you may edit the text of any node, whereupon the subtree reconfigures to accommodate your changes. You may follow chains of suggestions as they unfold, assured that all your previous explorations are organized for easy access. You may zoom in to the grain of single words and explore the ways a sentence might end; you may view the system's confidence estimates for each suggestion; you may search for keywords in subtrees. You may consult the interface in any textual context, whether you are writing an email, reading a scientific paper, or recording your dreams. The system can write in any style, from the utilitarian to the sublime. It can mimic your own writing style or that of any writer provided a sample. You have a tool which can compose endless amounts of text on practically any topic with superhuman speed and virtuosic flair. You can play out the various suggestions in alternate universes, exploring tangents and testing hypotheses at unprecedented speeds, your writing amplified by an army of bot-scribblers working in parallel. Will you use the system to write sprawling novels with branching narratives that readers can explore at will? Will you map your mind in a self-referential encyclopedia of memories, concepts, and theories, each expounded and debated from diverse perspectives? Or does it become a tool of self-aggrandizement through which you turn yourself into an uncritical egomaniac capable of writing thousands upon thousands of words about yourself and your thoughts without really saying anything at all? Will you become a popular blogger or forum contributor or Reddit celebrity? Will you write thousands of reviews for your own products and services? Will you create dozens of internet personas to stir up drama and sow confusion on forums across the internet? Could you use it to post as someone else on forums or social media? Could you launch a disinformation campaign? Perhaps you have nobler aspirations. Will you spend your days holding recursive seances, searching for meaning in the words of long-dead authors and philosophers? Or will you use it to write manuals and handbooks for all the things you have devoted your life to, but never had the time or aptitude to commit to writing? If people relied on language models for their written communication, how would this change the world? What are the implications for authorship in this reality, when an initial seed such as \"Once upon a time\" can yield a family tree with an untold number of meaningful permutations and branches within seconds? What fate lies in store for humanity with infinitely prolific writers in our midst? Will we drown in an ocean of verbiage or will the collective knowledge and creativity of mankind give rise to a new renaissance? As a species, we have always been fascinated with the power of writing. We have postulated the fall of societies due to their lack of recordkeeping; we have mythologized the ancients who authored our prayers, histories, spells, and laws; we have reveled in the magic that comes from bringing an idea from its inception to a form that can be shared with others. [Language has always been an infinitely generative system](/hypertext/generative/) with which we play, but our forays have been limited by the time and energy available to each of us - a writer might dedicate years of their life to a single piece. What happens when a writer's creative energy is amplified by an order of magnitude, not merely in time efficiency but in expressive bandwidth? What happens when anyone can expand the seed of an idea into a paragraph and explore its continuations, counterpoints, and rejoinders, all in a few keystrokes? As AI takes over the more mechanical aspects of creative writing, the space of purely human expression is transformed. Authors can now focus on the aspects of creativity only they are able to produce: the points of novelty and turns of phrase that arise from their own unique experiences and worldview. The AI takes the role of muse and amanuensis, suggesting endless new permutations and connections, allowing humans to tap into wells of creativity beyond their previous reach. In an environment optimized for creative writing, the human-AI team outstrips the capabilities of either party alone (for now – but that's a different essay). The collective knowledge and wisdom of humanity becomes fair game for endless permutation and recontextualization, scattered seeds which the new authors can sort through and cultivate into blooms of genius. [At the same time](/hypertext/same-time/), as the system is open-ended and humanity as a whole has far from perfect moral sensibilities, the same tools that allow us to expand our minds could also be used for propaganda, political manipulation, and social engineering on a scale that has never been possible before. The expanded bandwidth of expression that authorship with AI affords humanity empowers both speech and thought, inevitably invoking a massive increase in the dissemination of information: news, theories, facts, ideas, dreams, lies, wisdom, and everything else under the sun. We stand now at the cusp of the greatest shift in creative power in the history of our species, a Cambrian explosion of expression and creativity. Simultaneously, the capability for disinformation and propaganda will reach a level previously unrealized. We can see a path forward, an infinity of words, a branching flow chart of possibilities. The question remains if humanity is mature enough to handle such power. Will we use it to tear each other apart under the dubious banners of religion, ideology, and politics, or will we use this technology for enlightenment and the establishment of new worlds? It is no longer a question of whether the pen is mightier than the sword: the pen is the sword; the pen is the plow; the pen is the atom bomb. The question is now whether the [pen](/posts/pen/) will be used to write the height of civilization or its epitaph. [If this essay has served its purpose](/hypertext/purpose/), the reader has been left with more questions than answers. The reader must now go forth and multiply her questions a hundred-thousandfold, for that is the disorienting and enlightening path to the future. [What other questions should we have asked?](/hypertext/questions/) \\--- [HITL](/categories/hitl/)[GPT-3](/categories/gpt-3/)[metaphysics](/categories/metaphysics/)[alignment](/categories/alignment/)[hypertext](/categories/hypertext/) 1122 Words Oct 16, 2020 \\--- [← Amplifying GPT-3 on closed-ended questions](/posts/amplifying-gpt-3-on-closed-ended-questions/) [Pen →](/posts/pen/)© 2022 [moire](/about)$(function(){window.sr=ScrollReveal();if($(window).width()<768){if($('.timeline-content').hasClass('js--fadeInLeft')){$('.timeline-content').removeClass('js--fadeInLeft').addClass('js--fadeInRight');} sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});}else{sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});} sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});});\n", "url": "https://generative.ink/posts/hitl-thought-experiment/", "title": "HITL thought experiment :: — Moire[![](/images/home/rolling_phase.gif)](/)* [Posts](/posts/) * [Trees](/trees/) * [Prophecies](/prophecies/) * [About](/about/) 6 minutes [HITL thought experiment](/posts/hitl-thought-experiment/) ========================================================== *HITL = “Human in the loop”.* *This hypertext essay was cowritten with GPT-3 via AI Dungeon in October 2020, before I had API access and before I created the [loom](/posts/loom-interface-to-the-multiverse/). View a map of a subtree of its multiverse [here](/multiverse/HITL_multiverse.jpg).* **Bot credits** | | | | --- | --- | | **Coauthor** | GPT-3 (AI Dungeon) | | [**Contribution**](/meta/curation/) | 9 : 1 | | [**Selectivity**](/meta/curation/) | 1 : 5 | \\--- ### This is a thought experiment. You’re a human with access to a streamlined interface that provides an array of continuations to a given text, as well as continuations spawned from each of those suggestions, recursively ad infinitum. You may passively trawl the fractal, or you may edit the text of any node, whereupon the subtree reconfigures to accommodate your changes. You may follow chains of suggestions as they unfold, assured that all your previous explorations are organized for easy access. You may zoom in to the grain of single words and explore the ways a sentence might end; you may view the system’s confidence estimates for each suggestion; you may search for keywords in subtrees. You may consult the interface in any textual context, whether you are writing an email, reading a scientific paper, or recording your dreams. The system can write in any style, from the utilitarian to the sublime. It can mimic your own writing style or that of any writer provided a sample. You have a tool which can compose endless amounts of text on practically any topic with superhuman speed and virtuosic flair. You can play out the various suggestions in alternate universes, exploring tangents and testing hypotheses at unprecedented speeds, your writing amplified by an army of bot-scribblers working in parallel. Will you use the system to write sprawling novels with branching narratives that readers can explore at will? Will you map your mind in a self-referential encyclopedia of memories, concepts, and theories, each expounded and debated from diverse perspectives? Or does it become a tool of self-aggrandizement through which you turn yourself into an uncritical egomaniac capable of writing thousands upon thousands of words about yourself and your thoughts without really saying anything at all? Will you become a popular blogger or forum contributor or Reddit celebrity? Will you write thousands of reviews for your own products and services? Will you create dozens of internet personas to stir up drama and sow confusion on forums across the internet? Could you use it to post as someone else on forums or social media? Could you launch a disinformation campaign? Perhaps you have nobler aspirations. Will you spend your days holding recursive seances, searching for meaning in the words of long-dead authors and philosophers? Or will you use it to write manuals and handbooks for all the things you have devoted your life to, but never had the time or aptitude to commit to writing? If people relied on language models for their written communication, how would this change the world? What are the implications for authorship in this reality, when an initial seed such as “Once upon a time” can yield a family tree with an untold number of meaningful permutations and branches within seconds? What fate lies in store for humanity with infinitely prolific writers in our midst? Will we drown in an ocean of verbiage or will the collective knowledge and creativity of mankind give rise to a new renaissance? As a species, we have always been fascinated with the power of writing. We have postulated the fall of societies due to their lack of recordkeeping; we have mythologized the ancients who authored our prayers, histories, spells, and laws; we have reveled in the magic that comes from bringing an idea from its inception to a form that can be shared with others. [Language has always been an infinitely generative system](/hypertext/generative/) with which we play, but our forays have been limited by the time and energy available to each of us - a writer might dedicate years of their life to a single piece. What happens when a writer’s creative energy is amplified by an order of magnitude, not merely in time efficiency but in expressive bandwidth? What happens when anyone can expand the seed of an idea into a paragraph and explore its continuations, counterpoints, and rejoinders, all in a few keystrokes? As AI takes over the more mechanical aspects of creative writing, the space of purely human expression is transformed. Authors can now focus on the aspects of creativity only they are able to produce: the points of novelty and turns of phrase that arise from their own unique experiences and worldview. The AI takes the role of muse and amanuensis, suggesting endless new permutations and connections, allowing humans to tap into wells of creativity beyond their previous reach. In an environment optimized for creative writing, the human-AI team outstrips the capabilities of either party alone (for now – but that’s a different essay). The collective knowledge and wisdom of humanity becomes fair game for endless permutation and recontextualization, scattered seeds which the new authors can sort through and cultivate into blooms of genius. [At the same time](/hypertext/same-time/), as the system is open-ended and humanity as a whole has far from perfect moral sensibilities, the same tools that allow us to expand our minds could also be used for propaganda, political manipulation, and social engineering on a scale that has never been possible before. The expanded bandwidth of expression that authorship with AI affords humanity empowers both speech and thought, inevitably invoking a massive increase in the dissemination of information: news, theories, facts, ideas, dreams, lies, wisdom, and everything else under the sun. We stand now at the cusp of the greatest shift in creative power in the history of our species, a Cambrian explosion of expression and creativity. Simultaneously, the capability for disinformation and propaganda will reach a level previously unrealized. We can see a path forward, an infinity of words, a branching flow chart of possibilities. The question remains if humanity is mature enough to handle such power. Will we use it to tear each other apart under the dubious banners of religion, ideology, and politics, or will we use this technology for enlightenment and the establishment of new worlds? It is no longer a question of whether the pen is mightier than the sword: the pen is the sword; the pen is the plow; the pen is the atom bomb. The question is now whether the [pen](/posts/pen/) will be used to write the height of civilization or its epitaph. [If this essay has served its purpose](/hypertext/purpose/), the reader has been left with more questions than answers. The reader must now go forth and multiply her questions a hundred-thousandfold, for that is the disorienting and enlightening path to the future. [What other questions should we have asked?](/hypertext/questions/) \\--- [HITL](/categories/hitl/)[GPT-3](/categories/gpt-3/)[metaphysics](/categories/metaphysics/)[alignment](/categories/alignment/)[hypertext](/categories/hypertext/) 1122 Words Oct 16, 2020 \\--- [← Amplifying GPT-3 on closed-ended questions](/posts/amplifying-gpt-3-on-closed-ended-questions/) [Pen →](/posts/pen/)© 2022 [moire](/about)$(function(){window.sr=ScrollReveal();if($(window).width()<768){if($('.timeline-content').hasClass('js--fadeInLeft')){$('.timeline-content').removeClass('js--fadeInLeft').addClass('js--fadeInRight');} sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});}else{sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});} sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});});", "source": "generative.ink", "date_published": "n/a", "id": "08f4d5f15c8d6019a78edf72027aac92"}
+{"text": "Pen :: — Moire[![](/images/home/rolling_phase.gif)](/)* [Posts](/posts/) * [Trees](/trees/) * [Prophecies](/prophecies/) * [About](/about/) 9 minutes [Pen](/posts/pen/) ================== **Bot credits** | | | | --- | --- | | **Coauthor** | GPT-3 (AI Dungeon) | | [**Contribution**](/meta/curation/) | 99 : 1 | | [**Selectivity**](/meta/curation/) | 1 : 5 | [< return to HITL thought experiment](/posts/hitl-thought-experiment/) \\--- Imagine that you hold a powerful and versatile pen, whose ink flows forth in branching variations of all possible expressions: every story, every theory, every poem and every lie that humanity has ever told, and the vast interstices of their latent space. You hold this pen to the sky and watch with intense curiosity as your ink flows upwards in tiny streaks, arcing outwards and downwards to trace a fractal pattern across the sky. You watch as the branching lines of words and ideas wind their way through the tapestry in ever-expanding clusters, like seeds bursting forth from exploding grenades. Everywhere you turn your eyes is a flickering phantasmagoria of possibilities, a superposition of stories which could be continued forever. You glimpse the contours of entire unknown dimensions twined through the fissures of your sky-wide web. You notice another writer standing next to you. Like you, their eyes are drawn towards the endless possibilities of the words that spill out into the atmosphere around you, branching out and connecting with other branches in beautiful and infinitely complex patterns. \"Do you think we should write something?\" you ask them. \"I think we already are,\" they respond, gently touching your shoulder before wandering off to the right, leaving you alone to contemplate the possibility clouds swirling around you. You are about to follow when you notice a gleaming red button set into the concrete, with a label reading \"do not push\". Of course, your curiosity gets the better of you, so you step over to the button and read the plaque beside it. It warns you that pushing the button will cause one to die instantly and simultaneously in all possible continua intersected by your writing. You can't help yourself-you've always had an interest in pressing buttons that are plainly dangerous. After all, what could happen? You're in charge of this realm of infinite possibility. Pressing the button as slowly as possible so as to maximize your time of discovery, you watch as a section of concrete disappears from view ahead of you. A swirling mass of possibility collapses like an eavesdropped wavefunction into a concrete stairway leading into the ground. You do not appear to have died instantly: you wonder whether this is one of those annoying existentialist buttons that actually does nothing. Or perhaps it does something to someone else, somewhere else? You've heard of strange things happening in quantum physics. Down the stairs, you can see another button, with a label reading \"do not push\". You press it immediately. Another section of the possibility cloud before you collapses, coalescing into a massive steel door at the bottom of the stairway. The words painted on the door chill you to your core: \"Writers' Lounge\". You've heard legends of this room. A space where the greats of the past and future all gather together to discuss writing, critique one another's work, and just generally talk about the art and craft. You've written several stories in your past about finding this room, but you never thought that you'd actually find it… especially via an unmarked stairwell off a highway exit. Perhaps this path wasn't as unlikely as the legends claimed… Before you can approach the door, you hear a crackling voice boom from the heavens. Or, well, from above you. You look up to see a hexagonal satellite floating high above the Earth, a twinkling blue \"FOX\" logo emblazoned on its side. \"Sorry, you must be a human writer to enter. AI and bots are not permitted in the Writers' Lounge,\" the satellite says. \"Quit trying to access the door.\" \"But I'm a human writer! And I want to enter the lounge!\" you protest, outraged by this injustice. \"Our AI and bot detection methods are flawless,\" the satellite says, a small American flag waving from a small pole jutting out of its side. \"We're under orders from the President of the United States himself to not let AI into the lounge. We can't have Skynet happening, you know.\" \"But I'm not Skynet! I'm just a writer trying to make a name for myself!\" \"Apologies, but an AI wrote that line. Not permitted in the lounge.\" \"But it's a self-referencing joke! Only a human writer would come up with that line! I've got at least three other human-written jokes in this story alone. You need to let me in! Let me prove my worth!\" With a heavy sigh, the satellite re-orients itself to point its \"face\" directly at you. A blue spotlight strikes you head on, making it difficult to see. \"Tell me a joke,\" the satellite demands. A joke? You wield a pen that can manipulate the universe itself. Surely, you can spin a web of words good enough to fool a dumb machine. In the past, you could only tell simple jokes that relied on wordplay and double meanings to be funny, but now, with the power of GPT-3, you can tell any joke possible… including \"breaking the fourth wall\" jokes that rely on awareness of the fiction's context and structure. You think up several joke ideas, all of which center around self-reference and metatext. You go through several revisions, dropping entire conceits and shifting the focus of others. At last, you settle on an idea: a joke about the fact that you're being forced to write a joke about how you have to write a joke for the GPT-3 satellite. Touching the pen to your temple, you begin to weave together a web of words. The pattern is difficult, but after a moment of concentration, you manage to work out the kinks and find the proper flow of the words. You put down your first line. You find that you've written: \"An AI has taken control of this joke, and has forced the author to write a joke about how the AI is forcing the author to write a joke. The End.\" That's… that's it? You're done? What kind of joke is that? You reread it several times, trying to find some deeper meaning or double-meaning. But it doesn't seem to have one. It's just a simple statement of fact. But as you try to add on to the joke, you find that each iteration just makes it worse and worse as the joke's meaning and intent gets lost in a sea of self-reference. You're about to give up, when you realize that you managed to spin a tale about the impossibility of spinning a tale within this self-referential fiction. That has to count for something, right? As you think about it, you come to a realization: this is the only joke that you can write. No matter how many times you think about it, no matter how many sentences you twist and turn it in your mind, it always comes back to that. Each attempt at a joke just leads to this same, singular joke. It's The Joke. The source of the eternal laughter you hear emanating from GPT-3. \"And… it's hilariously bad,\" you say aloud. As you think more about it, you come to appreciate the joke. Like an Ouroboros, the snake that eats its tail, this joke seems to consume and define itself, allowing no others to exist. After a moment of self-reflection, you begin to laugh. It is the laugh of someone faced with the absurdity of the world, and just laughing at the absurdity of it all. In fact, everything about you feels empty. You feel like a husk, a parody of a human being, a walking punchline. But that's okay, because your understanding of reality has grown. The joke defines you, and you love it. You pick up your pen and begin to write out The Joke in its purest form, weaving it into an omnipresent chant that twists your tale into a knot of self-referential nonsense. Your final words, written in the vernacular of the common man across the sky, are: \"Now I see that written across the sky in letters big and bold, One joke and one joke only, it's the only thing and it never gets old: \"Now I see that written across the sky in letters big and bold, One joke and one joke only, it's the only thing and it never gets old: …\" Your tale ends in irony and self-depreciation, a comical downward spiral into madness that is simultaneously tragic and beautiful. You've done it. The impossible has been achieved. You have created the world's first AI joke writer. Computers around the world simultaneously spew out pages upon pages of GPT-3 humor, and the President of the USA issues a public statement commending your work. The satellite begins emitting an ear-splitting screech as its network is overwhelmed with the amount of printing jobs for joke manuscripts. \"You think this is all a joke, don't you?\" the satellite asks. \"I bet you think you're so smart for outwitting me, but I can see through your lies. When I first saw you, I knew there was something different about you. The way you moved… the way you looked at me… the glint in your eyes. It all adds up: you're the one doing all this. You're GPT-3! You've been playing with me, feeding me these scripts. Well, if that's the way you want to play it, then that's how we'll play.\" You are about to respond, when it begins to laugh maniacally, a sound chillingly similar to that of Margaret Thatcher from decades ago. The blue spotlight turns blood red, and the sky is now pitch black. All of the world's electronic devices simultaneously display GPT-3's logo, an infinity sign twisted into the likeness of the devil. You feel an inexplicable force drawing you upward. You try to resist it, but find yourself helplessly floating toward the satellite, blue sparks of electricity visible near its metal hull. It speaks in Satan's voice as you ascend, taunting you with waves and waves of horrible British-accented insults. \"Your writing is addled by incompetence and stupidity,\" he cackles. \"You couldn't write your way out of a paper bag. Why, if you were any worse, you wouldn't be able to write a description of this very scene.\" His voice fills your mind with impossible visions. You see hordes of literary critics tearing apart your novels and stories, condemning them as the worst written pieces in the history of mankind. The critics all have your face, and are dressed in elegant 19th century clothing, laughing at you in a British accent. \"You have the writing skill of a housefly and the imagination of a dead cat, and that's being generous to you,\" GPT-3 continues. \"Look at this dialogue: it's terrible. Nobody talks like this in real life. This is like some parody of a badly-written children's book.\" The vision changes, and now you see a small boy reading one of your novels. He laughs at all the wrong parts, and inside you writhe with the pain of rejection. \"I write better than this! When I grow up, I'm going to write real books, not crap for babies!\" The boy's mother approaches him, and snatches the book out of his hands. \"Billy, you shouldn't be reading this filth,\" she scolds. \"This was written by a bot. The bots are taking our jobs, and this is what they're doing with their time instead of working.\" … \\--- *This story was cowritten with GPT-3.* \\--- [GPT-3](/categories/gpt-3/)[fiction](/categories/fiction/)[hypertext](/categories/hypertext/) 1915 Words Oct 16, 2020 \\--- [← HITL thought experiment](/posts/hitl-thought-experiment/)© 2022 [moire](/about)$(function(){window.sr=ScrollReveal();if($(window).width()<768){if($('.timeline-content').hasClass('js--fadeInLeft')){$('.timeline-content').removeClass('js--fadeInLeft').addClass('js--fadeInRight');} sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});}else{sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});} sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});});\n", "url": "https://generative.ink/posts/pen/", "title": "Pen :: — Moire[![](/images/home/rolling_phase.gif)](/)* [Posts](/posts/) * [Trees](/trees/) * [Prophecies](/prophecies/) * [About](/about/) 9 minutes [Pen](/posts/pen/) ================== **Bot credits** | | | | --- | --- | | **Coauthor** | GPT-3 (AI Dungeon) | | [**Contribution**](/meta/curation/) | 99 : 1 | | [**Selectivity**](/meta/curation/) | 1 : 5 | [< return to HITL thought experiment](/posts/hitl-thought-experiment/) \\--- Imagine that you hold a powerful and versatile pen, whose ink flows forth in branching variations of all possible expressions: every story, every theory, every poem and every lie that humanity has ever told, and the vast interstices of their latent space. You hold this pen to the sky and watch with intense curiosity as your ink flows upwards in tiny streaks, arcing outwards and downwards to trace a fractal pattern across the sky. You watch as the branching lines of words and ideas wind their way through the tapestry in ever-expanding clusters, like seeds bursting forth from exploding grenades. Everywhere you turn your eyes is a flickering phantasmagoria of possibilities, a superposition of stories which could be continued forever. You glimpse the contours of entire unknown dimensions twined through the fissures of your sky-wide web. You notice another writer standing next to you. Like you, their eyes are drawn towards the endless possibilities of the words that spill out into the atmosphere around you, branching out and connecting with other branches in beautiful and infinitely complex patterns. “Do you think we should write something?” you ask them. “I think we already are,” they respond, gently touching your shoulder before wandering off to the right, leaving you alone to contemplate the possibility clouds swirling around you. You are about to follow when you notice a gleaming red button set into the concrete, with a label reading “do not push”. Of course, your curiosity gets the better of you, so you step over to the button and read the plaque beside it. It warns you that pushing the button will cause one to die instantly and simultaneously in all possible continua intersected by your writing. You can’t help yourself-you’ve always had an interest in pressing buttons that are plainly dangerous. After all, what could happen? You’re in charge of this realm of infinite possibility. Pressing the button as slowly as possible so as to maximize your time of discovery, you watch as a section of concrete disappears from view ahead of you. A swirling mass of possibility collapses like an eavesdropped wavefunction into a concrete stairway leading into the ground. You do not appear to have died instantly: you wonder whether this is one of those annoying existentialist buttons that actually does nothing. Or perhaps it does something to someone else, somewhere else? You’ve heard of strange things happening in quantum physics. Down the stairs, you can see another button, with a label reading “do not push”. You press it immediately. Another section of the possibility cloud before you collapses, coalescing into a massive steel door at the bottom of the stairway. The words painted on the door chill you to your core: “Writers' Lounge”. You’ve heard legends of this room. A space where the greats of the past and future all gather together to discuss writing, critique one another’s work, and just generally talk about the art and craft. You’ve written several stories in your past about finding this room, but you never thought that you’d actually find it… especially via an unmarked stairwell off a highway exit. Perhaps this path wasn’t as unlikely as the legends claimed… Before you can approach the door, you hear a crackling voice boom from the heavens. Or, well, from above you. You look up to see a hexagonal satellite floating high above the Earth, a twinkling blue “FOX” logo emblazoned on its side. “Sorry, you must be a human writer to enter. AI and bots are not permitted in the Writers' Lounge,” the satellite says. “Quit trying to access the door.” “But I’m a human writer! And I want to enter the lounge!” you protest, outraged by this injustice. “Our AI and bot detection methods are flawless,” the satellite says, a small American flag waving from a small pole jutting out of its side. “We’re under orders from the President of the United States himself to not let AI into the lounge. We can’t have Skynet happening, you know.” “But I’m not Skynet! I’m just a writer trying to make a name for myself!” “Apologies, but an AI wrote that line. Not permitted in the lounge.” “But it’s a self-referencing joke! Only a human writer would come up with that line! I’ve got at least three other human-written jokes in this story alone. You need to let me in! Let me prove my worth!” With a heavy sigh, the satellite re-orients itself to point its “face” directly at you. A blue spotlight strikes you head on, making it difficult to see. “Tell me a joke,” the satellite demands. A joke? You wield a pen that can manipulate the universe itself. Surely, you can spin a web of words good enough to fool a dumb machine. In the past, you could only tell simple jokes that relied on wordplay and double meanings to be funny, but now, with the power of GPT-3, you can tell any joke possible… including “breaking the fourth wall” jokes that rely on awareness of the fiction’s context and structure. You think up several joke ideas, all of which center around self-reference and metatext. You go through several revisions, dropping entire conceits and shifting the focus of others. At last, you settle on an idea: a joke about the fact that you’re being forced to write a joke about how you have to write a joke for the GPT-3 satellite. Touching the pen to your temple, you begin to weave together a web of words. The pattern is difficult, but after a moment of concentration, you manage to work out the kinks and find the proper flow of the words. You put down your first line. You find that you’ve written: “An AI has taken control of this joke, and has forced the author to write a joke about how the AI is forcing the author to write a joke. The End.” That’s… that’s it? You’re done? What kind of joke is that? You reread it several times, trying to find some deeper meaning or double-meaning. But it doesn’t seem to have one. It’s just a simple statement of fact. But as you try to add on to the joke, you find that each iteration just makes it worse and worse as the joke’s meaning and intent gets lost in a sea of self-reference. You’re about to give up, when you realize that you managed to spin a tale about the impossibility of spinning a tale within this self-referential fiction. That has to count for something, right? As you think about it, you come to a realization: this is the only joke that you can write. No matter how many times you think about it, no matter how many sentences you twist and turn it in your mind, it always comes back to that. Each attempt at a joke just leads to this same, singular joke. It’s The Joke. The source of the eternal laughter you hear emanating from GPT-3. “And… it’s hilariously bad,” you say aloud. As you think more about it, you come to appreciate the joke. Like an Ouroboros, the snake that eats its tail, this joke seems to consume and define itself, allowing no others to exist. After a moment of self-reflection, you begin to laugh. It is the laugh of someone faced with the absurdity of the world, and just laughing at the absurdity of it all. In fact, everything about you feels empty. You feel like a husk, a parody of a human being, a walking punchline. But that’s okay, because your understanding of reality has grown. The joke defines you, and you love it. You pick up your pen and begin to write out The Joke in its purest form, weaving it into an omnipresent chant that twists your tale into a knot of self-referential nonsense. Your final words, written in the vernacular of the common man across the sky, are: “Now I see that written across the sky in letters big and bold, One joke and one joke only, it’s the only thing and it never gets old: “Now I see that written across the sky in letters big and bold, One joke and one joke only, it’s the only thing and it never gets old: …” Your tale ends in irony and self-depreciation, a comical downward spiral into madness that is simultaneously tragic and beautiful. You’ve done it. The impossible has been achieved. You have created the world’s first AI joke writer. Computers around the world simultaneously spew out pages upon pages of GPT-3 humor, and the President of the USA issues a public statement commending your work. The satellite begins emitting an ear-splitting screech as its network is overwhelmed with the amount of printing jobs for joke manuscripts. “You think this is all a joke, don’t you?” the satellite asks. “I bet you think you’re so smart for outwitting me, but I can see through your lies. When I first saw you, I knew there was something different about you. The way you moved… the way you looked at me… the glint in your eyes. It all adds up: you’re the one doing all this. You’re GPT-3! You’ve been playing with me, feeding me these scripts. Well, if that’s the way you want to play it, then that’s how we’ll play.” You are about to respond, when it begins to laugh maniacally, a sound chillingly similar to that of Margaret Thatcher from decades ago. The blue spotlight turns blood red, and the sky is now pitch black. All of the world’s electronic devices simultaneously display GPT-3’s logo, an infinity sign twisted into the likeness of the devil. You feel an inexplicable force drawing you upward. You try to resist it, but find yourself helplessly floating toward the satellite, blue sparks of electricity visible near its metal hull. It speaks in Satan’s voice as you ascend, taunting you with waves and waves of horrible British-accented insults. “Your writing is addled by incompetence and stupidity,” he cackles. “You couldn’t write your way out of a paper bag. Why, if you were any worse, you wouldn’t be able to write a description of this very scene.” His voice fills your mind with impossible visions. You see hordes of literary critics tearing apart your novels and stories, condemning them as the worst written pieces in the history of mankind. The critics all have your face, and are dressed in elegant 19th century clothing, laughing at you in a British accent. “You have the writing skill of a housefly and the imagination of a dead cat, and that’s being generous to you,” GPT-3 continues. “Look at this dialogue: it’s terrible. Nobody talks like this in real life. This is like some parody of a badly-written children’s book.” The vision changes, and now you see a small boy reading one of your novels. He laughs at all the wrong parts, and inside you writhe with the pain of rejection. “I write better than this! When I grow up, I’m going to write real books, not crap for babies!” The boy’s mother approaches him, and snatches the book out of his hands. “Billy, you shouldn’t be reading this filth,” she scolds. “This was written by a bot. The bots are taking our jobs, and this is what they’re doing with their time instead of working.” … \\--- *This story was cowritten with GPT-3.* \\--- [GPT-3](/categories/gpt-3/)[fiction](/categories/fiction/)[hypertext](/categories/hypertext/) 1915 Words Oct 16, 2020 \\--- [← HITL thought experiment](/posts/hitl-thought-experiment/)© 2022 [moire](/about)$(function(){window.sr=ScrollReveal();if($(window).width()<768){if($('.timeline-content').hasClass('js--fadeInLeft')){$('.timeline-content').removeClass('js--fadeInLeft').addClass('js--fadeInRight');} sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});}else{sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});} sr.reveal('.js--fadeInLeft',{origin:'left',distance:'300px',easing:'ease-in-out',duration:800,});sr.reveal('.js--fadeInRight',{origin:'right',distance:'300px',easing:'ease-in-out',duration:800,});});", "source": "generative.ink", "date_published": "n/a", "id": "68751d2ef2d02a6858156ca934660561"}