data
stringlengths 115
7.61k
|
|---|
asara#0001: but, those datasets are all single speakers at a time, not that conversational, so depending on your use case maybe not?
asara#0001: I know speaker separation and diarization papers have some others too though
ericxtang | Livepeer#9262: Right. I'm particularly interested in conversational data - ASR for user generated video content (think Twitch-isk). I haven't looked into data in speaker separate and diarization papers though.
asara#0001: Is there something that leads you to believe ASR trained on publicly available datasets wouldn't work well for that type of content? There may be some fine-tuning room for improvement I suppose, but transcribing and aligning Twitch streams definitely sounds like effort
asara#0001: (e.g. maybe you want to capture more speaker variety? screaming/shouting?)
ericxtang | Livepeer#9262: > Is there something that leads you to believe ASR trained on publicly available datasets wouldn't work well for that type of content?
A lot of times the audio is conversational. I found that models trained on publicly available datasets work well for things like TED talks, but start to fall apart quickly when the audio becomes more "irregular". IMO a model needs to have higher than 95% accuracy to create a passable experience. Public datasets I found just aren't quite there.
ericxtang | Livepeer#9262: And yes - definitely things like screaming / shouting in the training set would help. But I haven't found them in the publicly available datasets.
asara#0001: something like "ASR, but for situations where people are not speaking as if there is a need for ASR"
asara#0001: that's a decent point but I have less data on that, but I know some of it is definitely out there although the best stuff is likely proprietary hmm
chilli#5665: Send link
ericxtang | Livepeer#9262: Yeah - Google meet does a pretty killer job at real time ASR.
asara#0001: I don't know of datasets specifically made for contexts that are that different, but I imagine some exist
asara#0001: Unfortunately making The Pile but with fully-transcribed audio sounds like 10e3 more work or something though, since you have to actually label every line
asara#0001: The best startups outsource(d) it to overseas labor as usual of course
ericxtang | Livepeer#9262: I agree - a lot of work will be involved in curating a dataset like that.
The outsource strategy just feels so expensive, and repetitive.
asara#0001: it definitely is, but how else will you find hundreds of people to label audio
asara#0001: but you should do an exhaustive+creative search of what exists before concluding a new one is needed |
asara#0001: wav2letter might be able to go a bit further than you'd guess, but 95% accuracy for fast/loud/abnormal/disrupted speech is probably still a bit of an ask yeah
ericxtang | Livepeer#9262: tried the pretrained model with wav2letter - i think it's trained on librispeech. works ok, but more data is needed.
asara#0001: https://arxiv.org/abs/2101.00390 will be released soon I think, most of it is not transcribed but it's huge overall so a lot is, it may help
ericxtang | Livepeer#9262: cool!
asara#0001: maybe could ping someone and ask when because the link just kinda 404s so
ericxtang | Livepeer#9262: any fb folks here? 😜 help us!
asara#0001: or do nothing and wait for someone to make a github issue 'where is my data. pls. thank u'
ericxtang | Livepeer#9262: I would be that person if the repo exists... https://cdn.discordapp.com/attachments/729741769738158194/797370096447913994/Screen_Shot_2021-01-09_at_2.43.44_AM.png
asara#0001: hehe, I'll message you if I come across something super relevant by chance though, good luck
ericxtang | Livepeer#9262: Thank you @asara
ericxtang | Livepeer#9262: I just emailed the author. Will keep you posted if I hear back.
Plokon#8568: Hy guys, so I tried out AI Dungeons dragon model (based on gpt-3) I was pretty impressed with the stuff. What other places can a mere casual like me without any programming knowledge mess around with the latest gpt-3 based stuffs(?), if there is any you know of..
Deleted User#0000: https://cdn.discordapp.com/attachments/729741769738158194/797468223117197342/Screenshot_from_2021-01-09_06-13-46.png
Deleted User#0000: way too late to the game
Sphinx#2092: Somebody internally made a reference to your implementation to your dall-e. I think your name by itself carries fame now lol
Deleted User#0000: lol yea, we should try to train one with dalle-mtf, and we can redirect some attention to eleuther
Deleted User#0000: i'll put in some work tomorrow
bmk#1476: https://discord.gg/knBrNQGdX9 enjoy
chilli#5665: sadly near empty
bmk#1476: dang |
bmk#1476: well, the link is kinda hard to find
bmk#1476: it's buried on the guy's personal website but not the gpt-4.co website
Louis#0144: LMAO
triggerhappygandi#0001: AGI at work
3dprint_the_world#6486: https://www.youtube.com/watch?v=mmDw5glry9w
bmk#1476: hmm, i wonder who made this "thispersondoesnotexist" thing they're talking about
IKEA#9631: Wait, why are there now 2 different active repos for DALLE, and even using different frameworks? Wouldnt it be much more efficient to have only one, or at least have one of them somehow depend on the other's code for easy updates?
cfoster0#4356: One is for Mesh Tensorflow and one is PyTorch
IKEA#9631: Yeah but why though
IKEA#9631: Like why not Mesh Tensorflow and uhh.. Tensorflow
IKEA#9631: Unless im missing something
bmk#1476: the pytorch one isnt ours
cfoster0#4356: The *only* reason folks went for MTF is because TPUs, I believe
Sid#2121: i'm not really sure what the question is here lol. They're 2 different repos, built in 2 different frameworks, by 2 different people
Sid#2121: one is neat and easy to use and is built by @Deleted User so is infinitely more chad
Sid#2121: another one hacks things together so it works on TPUs 😆
Deleted User#0000: its easier to scale up with mesh tensorflow than pytorch, atm
Aran Komatsuzaki#5714: another one was made by Noam Shazeer, so this means i'm legally obligated to use it
Deleted User#0000: @Aran Komatsuzaki have you heard anything more about what Shazeer is working on?
Deleted User#0000: i think he's seen the light |
Aran Komatsuzaki#5714: no but i believe the newest result on t5 (superglue leaderboard) is t5+meena, so he should be working on that
Deleted User#0000: curious where that is leading him
Deleted User#0000: ahh, what is t5 + meena?
Aran Komatsuzaki#5714: give me a sec
Deleted User#0000: isn't meena that transformers that does chat?
Aran Komatsuzaki#5714: https://cdn.discordapp.com/attachments/729741769738158194/797576401959387176/fig1.png
Deleted User#0000: https://arxiv.org/abs/2001.09977
Deleted User#0000: that's the Meena i'm thinking of
Aran Komatsuzaki#5714: So his team probably made T5 + Meena by combining them together.
Deleted User#0000: ohh
Deleted User#0000: interesting
nz#9710: oh shit that's another flex
Sphinx#2092: I like how they literally say its Zirui Wang and you all gave credit to Noam lol
Deleted User#0000: Deberta retook the throne!
Aran Komatsuzaki#5714: When it's co-authored by Noam, then all credit naturally goes to Noam
nz#9710: I love the fact that T5+Meena came out the same day, and now deberta took over again
Deleted User#0000: They both have cards up their sleeves
Deleted User#0000: lol
Aran Komatsuzaki#5714: Just like anything OpenAI does is somehow credited to Elon Musk
Aran Komatsuzaki#5714: ok tbf for gshard paper, pretty much all credits should go to Noam |
Aran Komatsuzaki#5714: probably not for T5 + Meena, since that sounds boring and not like him
Sphinx#2092: I suppose all you can do is wait for the public explanation at some point.
Aran Komatsuzaki#5714: yes i'm speaking without knowing anything but the title
Aran Komatsuzaki#5714: and as you can guess, i'm not being serious
Deleted User#0000: if you are on the leader board, does that obligate you to have a paper out at some point?
Deleted User#0000: curious what this T5 + meena is
Louis#0144: Isn’t Noam in the discord
Louis#0144: No nvm that’s @Noa Nabeshima
bmk#1476: pls dont ping people when youre not trying to get their attention
Louis#0144: O
Louis#0144: True ok
bmk#1476: pings should be reserved for when you need someone to come and look at something
bmk#1476: not when youre talking *about* someone
Louis#0144: I always just assume everyone has discord muted
Louis#0144: Tbh
bmk#1476: i dont want everyone to mute discord
bmk#1476: because then when i have something important to reach them, i get no response
Louis#0144: Notifications are one of the worst inventions of the last two decades and honestly we need universal standards to disable all notifications on all devices
Louis#0144: ARAN
Louis#0144: lmao |
chilli#5665: I mean
chilli#5665: I obviously have discord muted lol
chilli#5665: but not for mentions
chilli#5665: since generally mentions are higher signal and lower noise
Louis#0144: Yes
Louis#0144: But it’s not universal
Louis#0144: And some apps can override it
Louis#0144: lol
Louis#0144: I want *zero* notifications
Louis#0144: No Siri suggestions
Louis#0144: No amber alerts
Louis#0144: No fucking RuneScape sending me a Siri suggestion to log in
Louis#0144: I got that the other day
Louis#0144: Also I don’t want Facebook giving me a pop up every 20seconds to enable notifications
Louis#0144: Like wtf
Louis#0144: This shit needs to be illegal
Louis#0144: It’s so anti consumer
Louis#0144: I use it for marketplace
Louis#0144: That’s the only reason
Louis#0144: It’s a Craigslist replacement |
Louis#0144: Lmao
Louis#0144: To be fair I have had a jailbroken iPhone for the last four years
Louis#0144: And I have all notifs off
Louis#0144: Even system notifs and amber alerts
Louis#0144: Not amber alerts
Louis#0144: No?
Louis#0144: Omg what!
Louis#0144: How
Louis#0144: Omg....
Louis#0144: ❤️❤️❤️
IKEA#9631: still my favorite thing to ever come out of OAI https://cdn.discordapp.com/attachments/729741769738158194/797621948895985704/unknown.png
Milan Cvitkovic#1279: What's the right channel to ask about group equivariant models?
cfoster0#4356: Typically that'd be #research but now we've also got #math
bmk#1476: @StellaAthena is a big fan of those
Milan Cvitkovic#1279: The agony of choice...
triggerhappygandi#0001: @bmk where is #vd-vae?
triggerhappygandi#0001: ah it changed to #multimodal
Deleted User#0000: Man, so much stuff to catch up in here! The interactive-agi seems particularly interesting. I'll catch up with that soon.
My current situation is this: I'm pretty sure that the best use of my time is to work on my project to gather multimodal human behaviour data from social VR (which is advancing well). But also I just started a postdoc, and I'm not yet sure I'll be able to convince my supervisor that that is the best investment of my/out time. However, I think he'd be quite interested in some of the projects being discussed here (i think specially the ones in interactive-agi). So as plan B, I'd probably be able to convince him that I can work in one of those (and the social VR one on the side). But I'll keep trying to convince him of the value of the social VR project as well, and look for alternative ways I could work on that full time |
Igor Krawczuk#1653: Is there a discord bot which can collect the links posted on a day + the text of/ link to the discord post containing them in some file archive? The discussions here are super informative but very fast moving as well^^"
Igor Krawczuk#1653: Mainly thinking about paper references
nmkd#1425: The search function can filter messages containing links
nmkd#1425: But for archiving you'd probably need a script/bot
Daj#7482: Congrats on the new position! We definitely would appreciate more help on the EEGI project (interactive AI), since it's a big one with lots of things that need doing. I like your pet VR project though I'm not really certain what we could do with the data
Deleted User#0000: Thanks! I'm gonna definitely check out EEGI, and comment more when I have.
Regarding VR data, think of humans as a black box with inputs and outputs. We wanna, in the limit, record the stream of inputs and outputs from the human, to train generative models to imitate humans.
The VR data has such form: it is a continuous stream of perception-action, which for the most ideal has the following: perception includes vision, hearing (and soon even touch); action includes navigation, movement (including finger, face, and gaze), and speech
[there can also be some biometrics that are neither, like HR, EEG].
Imagine millions of hours of such data for humans in diverse virtual situations.
That data basically is there today, already; it just isn't being stored or used. That's what I wanna change
Once we have the data, I just wanna train a Big Model* on that continuous stream (also combining it with existing pre-trained models, mainly LMs), and see how far it can get towards human-like/human-level AI
*my preliminary thought is something like MoGlow or JukeBox, but to properly work with the full-multimodality, more work on the model will probably be needed
Daj#7482: Very neat. Scaling is all you need, haha
Deleted User#0000: yeap, tho it's easier said than done xD
gwern#1782: are you subscribed to https://www.reddit.com/r/mlscaling/ & https://www.reddit.com/r/reinforcementlearning/ ?
andyljones#7746: in the wake of GPT-3 last year i was trying to come up to speed on AI safety/scaling very quickly, and a substantial part of that was checking gwern's lesswrong comments/twitter replies/reddit account/website on a regular basis. my belief was that if there was an interesting source on safety/scaling somewhere on the internet, he'd mention it soon enough and i could add it to my own pool
|
great bot, would recommend 👍
Daj#7482: My pro tip: RSS feeds
Daj#7482: They still are a thing
Daj#7482: Most of the best stuff _I_ like reading is still in blogs
Daj#7482: ~~Come back Scott please, I beg you~~
Sid#2121: Tbf, we could easily rig up Isaac to post all links to #links
Sid#2121: I think we already did actually, ages ago, then took it out
andyljones#7746: at a guess: links without context aren't worth much?
Daj#7482: There are a thousand useful bot features we _could_ make
Sid#2121: We could link to the context, too
Sid#2121: Yes, lol
Sid#2121: Maybe one day our discord can be as well organised as yours 😂
Sid#2121: Probably not tho
Daj#7482: Maybe I can start taking Ritalin again and make one of those community feeds and stuff it with summaries of the best conversations and whatever I happen to be reading
andyljones#7746: if the RL server could fill twenty channels like this one does, it'd have twenty channels! last i checked the RL server is a couple of hundred messages a week. this one is probably, what, a few thousand?
give us some metrics tia
https://discord.com/developers/servers
Igor Krawczuk#1653: yes |
Sid#2121: i think @Daj will have to check this, i'd be interested to know too
Daj#7482: Seems we have to set our server to a community server first
Daj#7482: Which we still haven't done
Sid#2121: is that anything more than just clicking a button lol
Daj#7482: We need written rules
Sid#2121: if you give me the perms, i can do it
Daj#7482: and a few other things
Sid#2121: we wrote some up lol
Daj#7482: Yea did we ever reach consensus on that?
Sid#2121: doesn't matter, just put don't be a dick in there
Daj#7482: lmao
HypnoPump17#9322: hey guys, i've got a question: i recall reading a paper about a year ago that explained how fitting the model predictions along with the input data could improve prediction performance. anyone knows how is this technique called?
bmk#1476: Wdym
HypnoPump17#9322: so you have (x, _) and you predict y^, then you have (x,y^) and predict y^^. the thing is that y^^ suppoed should be more accurate than y^
HypnoPump17#9322: ideally the precess could be done n times so you have y(n^) at the end, which is supposed to be more accurate than just y^
HypnoPump17#9322: i remember that paper bc i had a somehow similar idea like 4 years ago with autoencoders but it didnt work
gwern#1782: sounds like residual learning
gwern#1782: you could also describe it as bagging or boosting maybe
HypnoPump17#9322: yup but it's with the same model. the model refines its own predictions
gwern#1782: there are of course a lot of iterative or anytime methods which could conceivably be framed that way too. what is an alphazero tree search but predicting y-hat based on expert iteration over a previous y-hat |
HypnoPump17#9322: yup you could frame it as student-teacher but i think i remember something a bit different
Igor Krawczuk#1653: So you mask out the slots for later predictions early in training?
Igor Krawczuk#1653: Sounds closest to boosting to me, except with shared weights between models
HypnoPump17#9322: i'm not sure of the implementation details. but i remember somthing like input: (x, nothing/random) --> y^ and then input: (x, y^) --> y^^
HypnoPump17#9322: so there's only 1 space for a prediction that can be passed or not passed at all
HypnoPump17#9322: aghh cant find it anywhere
Sphinx#2092: Perhaps https://openreview.net/attachment?id=HkxQRTNYPH&name=original_pdf
Sphinx#2092: In particular, 3.2.
Daj#7482: We are a community server now, therefor the existance of #rules , in case anyone was wondering
Daj#7482: We have server stats now https://cdn.discordapp.com/attachments/729741769738158194/797954514597445672/Screenshot_from_2021-01-10_23-24-57.png
chirp#4545: https://twitter.com/alexjc/status/1348266178406469632
chirp#4545: Curious if that’s in line with our own estimated cost
kindiana#1016: should be higher than that naive estimate, dalle is trained for ~1.7x more tokens
kindiana#1016: the vqvae doesn't really play into it compared to the size of the transformer
bmk#1476: this estimate is using all the wrong numbers
chirp#4545: What’s your estimate?
bmk#1476: i dont have enough info and i also dont feel like working it out
bmk#1476: i think "idk" is a preferable answer to "i think it's <random number>"
bmk#1476: my main problems with this estimate are:
|
cost is not proportional to param count alone since so much else is different
the 4.6M number is somewhat iffy
this estimate completely excludes the VAE which i suspect is the majority of the cost
kindiana#1016: I think the vae should be trivial compared to the transformer lol, the transformer is 33TFlops per forward pass per image, which is more than 2 orders of magnitude larger than the largest convnets
Sid#2121: yeah i don't think the VAE would be the majority at all
bmk#1476: huh
gwern#1782: it seems a little surprising that the VAE uses so little parameters. doesn't that feel backward to you? is it really *that* much harder to model caption language than to do the actual, you know, pixel generation?
bmk#1476: im not talking about param count
bmk#1476: im talking about flops
bmk#1476: do VAEs really have that many fewer flops?
gwern#1782: ok, flops, parameters... regardless, my intuition is that the VAE should be doing 'most of the work' in some sense
kindiana#1016: look at these huge honking convnets
kindiana#1016: well the vae is just upscaling from 32x32 to 256x256, the transformer is doing most of the high level generation
kindiana#1016: the vae is just figuring out the details and textures
bmk#1476: @kindiana are these the flop numbers/
kindiana#1016: https://cdn.discordapp.com/attachments/729741769738158194/797986005340389407/unknown.png
kindiana#1016: sorry bad crop
bmk#1476: o.O
chirp#4545: wait so convnets do only up to 100B, but GPT-3-13B does 33T?
kindiana#1016: gpt3-13B does 13B muladds _per token_ |
kindiana#1016: each example is 1024 image tokens + 256 caption tokens
chirp#4545: wow
chirp#4545: :brr:
bmk#1476: I knew convnets have fewer params but i thought they were at least similarly expensive flopswise
Joey#4305: Hey lucidrains you know you can just stay in the server instead of leaving and joining again right?
bmk#1476: staying in the server was never an option https://cdn.discordapp.com/attachments/729741769738158194/798025944811438090/unknown.png
aquajet#7800: oh fuck its kaggle
3dprint_the_world#6486: I can kind of understand why he keeps leaving. There's a regular daily influx of "omg lucidrains you're so amazing I love you" people that he's probably trying to avoid
3dprint_the_world#6486: plus, he's a dog.
Joey#4305: He's rejoining like every few hours
Joey#4305: Idk it's just odd to me
3dprint_the_world#6486: maybe he just doesn't know there's a "do not disturb" setting
Joey#4305: Possibly
bmk#1476: it is one of the great mysteries of our age
cognomen#6297: happens when lucid switches to a new model checkpoint
triggerhappygandi#0001: Indeed. The "Very Deep" VAE has 132M params
triggerhappygandi#0001: Even the imagenet winning convnets get bigger than this
triggerhappygandi#0001: But according to the vdvae paper almost every dataset cost 2.5 weeks to train on 32 V100s. Idk what that translates to FLOPs, since it could be 16gb/32gb variant
chilli#5665: it's not about other people pinging him lol
chilli#5665: it's about preventing himself from getting distracted |
nmkd#1425: wut
nmkd#1425: just close discord???
chilli#5665: it's about self control lol
chilli#5665: it's too easy to just open discord
3dprint_the_world#6486: as I said, Do Not Disturb
3dprint_the_world#6486: it's right there
Aran Komatsuzaki#5714: he just likes to do it
3dprint_the_world#6486: or you can just redirect discord.com in your HOSTS file
3dprint_the_world#6486: I used to do that with reddit to keep myself from getting distracted
andyljones#7746: considering how much more lucidrains gets done than the rest of us, maybe he's doing the sensible thing and we're being silly
nz#9710: yea I think he just found leaving an efficient way of not getting distracted
chilli#5665: What does do not disturb have to do with preventing yourself from accessing discord lol
Aran Komatsuzaki#5714: Sid somehow is really productive despite being here all the times. One of seven wonders of the world
chilli#5665: Sure you could do this, or you could deactivate your discord
chilli#5665: He does what works for him lol
chilli#5665: If he does a rain dance every time he wants to work who are we to judge
Daj#7482: Jack Clark in the latest ImportAI writes on Eleuther https://cdn.discordapp.com/attachments/729741769738158194/798132572650995742/Screenshot_from_2021-01-11_11-13-38.png
kindiana#1016: mysterious :thonk:
kindiana#1016: mfw you can just join the discord
Daj#7482: yea lol |
nz#9710: literally can't be more open than you guys
Daj#7482: I also had a video chat with him literally a few days ago
Daj#7482: Good branding tho
Aran Komatsuzaki#5714: openai is practically getting more mysterious more and more nowadays
Aran Komatsuzaki#5714: where's the dataset they're using?
paws#3311: i like importai ngl, thats the way he describes most stuff xD
Daj#7482: `Are you the hacker known as Eleuther?`
triggerhappygandi#0001: IS THAT A 4CHAN REFERENCE?!?
triggerhappygandi#0001: o_O
Daj#7482: ~~Is that a Jojo reference?!?!?!~~
Daj#7482: I miss the days when the most the internet influenced mainstream media was trolling Fox news into showing an exploding van
triggerhappygandi#0001: They what
triggerhappygandi#0001: Do you think oai will even release DALL-E
triggerhappygandi#0001: Bet they will excuse it as too dangerous too
Daj#7482: https://knowyourmeme.com/memes/exploding-van
Y'all too young to remember the grand ol' days
Daj#7482: (actually memes back then were shit, but it's nostalgic)
triggerhappygandi#0001: :brr:
andyljones#7746: jack clark popped in here for a day or two way back when, didn't he?
Daj#7482: Yea |
Daj#7482: I still have contact with him
thenightocean#6100: I kinda like that eleuther is considered mysterious and cool
JC#3653: Wait until they catch "lmfao" the chinese hacker
JC#3653: https://www.youtube.com/watch?v=8VGuDN6Z_t8
triggerhappygandi#0001: >Boomers
Dromarion#3383: *The shadowy organization whose entire body of work and discord server invite are on the first result on Google*
3dprint_the_world#6486: yeah, well, I have a day job...
aster#3007: so if anyone wanted to make a right wing gpt3 supposedly 70tb of data has been scraped from parler lmao
https://cybernews.com/news/70tb-of-parler-users-messages-videos-and-posts-leaked-by-security-researchers/
aster#3007: (Please don't actually include this in your corpus unless you're trying to model being a right wing chud lol)
jrowe#5371: if you can do sentiment and knowledge extraction from a model, it'd be neat to run a comparison between something trained on the parler data and something from reddit
Isaac McHorse#2007: comment of the year
jrowe#5371: it'd be awesome to visualize a semantic network graph over ideologies and how ideas relate for particular social groups
jrowe#5371: that'd give public officials an idea of the nuance from the other side - how and why people feel and think the way they do based on their own public discourse
jrowe#5371: gpt-ish probably isnt the right tool
Aran Komatsuzaki#5714: wow you just got awarded the comment of the year! kudos to you
aster#3007: or tik tok!
thenightocean#6100: tbf to simulate that level of intelligence u dont need that much compute
Igor Krawczuk#1653: I mean, this assumes parler and Q ideology is coherent and can be analysed no?
Igor Krawczuk#1653: as in, that it has an internal coherency that is observable in the discourse |
Igor Krawczuk#1653: Which I'm not sure about, I feel extreme reality deniers tend to engage in double speak and non-stationary language a lot
bmk#1476: After parler disappears, a service called "sprechen" mysteriously appears
Igor Krawczuk#1653: *" das sprëchën", weil es wieder ein amerikaner startet
Igor Krawczuk#1653: same as uber
Igor Krawczuk#1653: https://www.brainpickings.org/2017/02/07/hannah-arendt-the-banality-of-evil/
this came to my mind as well
Igor Krawczuk#1653: unless you were speaking about tracking sentiments and language using externally grounded representations and comparing that with the in-parler lingo?
bmk#1476: *"dem sprëçheñ"
Igor Krawczuk#1653: *"döm sprësçhöñ" (loipzüsch vörschiön)
bmk#1476: War mal in Sachsen, kann bestätigen dass sie komplett unverstehbar sind
jrowe#5371: i don't think "externally grounded" is exactly the right definition, but yeah, where you sample different communities with known ideological leanings
jrowe#5371: comments from the daily beast are gonna be different than a given reddit community, or twitter followers of a particular politician or pundit, etc
jrowe#5371: you could do a "big 5" style analysis and see what classifiers you get as opposed to the arbitrary left/right handwavy definitions people pull outta their hat
jrowe#5371: hah, then have a gpt prompt built into twitter so if what you type is determined to be associated with Nazis, you get a popup saying "a Nazi would share this opinion with you. It's concerning because: a,2,3 blah. Do you really want to post this?"
jrowe#5371: ideological beliefs aren't coherent very often. It's challenging and rare to find anyone on the planet that is ideologically consistent, and rarer still to find an ideology that is objectively rational. Lots of people have good reasons for believing bad things, and it requires a lot of work to dig down to whatever fundamentals were introduced where they go wrong - like flat earthers, for example. Or racists, or Westboro Baptists, or Islamic terrorists - they get deradicalized through patient and persistent conversation with people who share their humanity. If we had a tool that helped people understand eachother - or better yet, understand our own ideological divergence from reality - that could be a potent tool
bmk#1476: Or we could just push these people out of the mainstream and into their own little echo chambers what could possibly go wrong
jrowe#5371: lol
jrowe#5371: yeah, let em fester and simmer
jrowe#5371: that sounds *awesome* |
Igor Krawczuk#1653: I feel like it's not that complicated though? It's almost always emotional attachment and identity protection
andyljones#7746: > an ideology that is objectively rational
🤨
Igor Krawczuk#1653: Like, that's Hanna Arendts point, "evil" isn't complicated
bmk#1476: Rationalists of the world, unite!
andyljones#7746: you have nothing to lose but your friends!
bmk#1476: (/s)
jrowe#5371: yeah, but saying half the population is "evil" and booting them off the internet isn't exactly a winning end to the story
jrowe#5371: it might be better to try to understand what went wrong, and why, and try to correct and prevent repeats
Igor Krawczuk#1653: I put the quotes for a reason
bmk#1476: The twitter and its consequences have been a disaster for the human race
Igor Krawczuk#1653: I don't think they are evil
jrowe#5371: right, wasn't projecting onto you that particular view
jrowe#5371: I dunno, twitter has also enabled a lot of good
Igor Krawczuk#1653: Like, research wise I think your idea is interesting, but when someone is having a psychotic breakdown and is hurting people, you don't try to disentangle their reasoning
Igor Krawczuk#1653: You stop them, and then try to be compassionate and empathetic and figure out what hurts them
Igor Krawczuk#1653: If we can scale that, that'd be cool. But we don't need to act like Qanon is a Mysterious force that came out of nowhere
jrowe#5371: how it effectively got leveraged into such a large movement - what made people susceptible to it? those questions should be answered
IKEA#9631: The internet as a whole peaked in 2007, change my mind
jrowe#5371: i dunno, IIRC video streaming for porn didn't really hit its stride til after 2010... |
bmk#1476: What if we build a social media specifically to address the weaknesses of twitter wrt echo chambering
IKEA#9631: well i never really cared about that anyway
jrowe#5371: I mean, uh, netflix
bmk#1476: The internet peaked with the existence of eleuther change my mind
IKEA#9631: So youre saying it's peaking right now?
IKEA#9631: Hmmmm
IKEA#9631: Not really
bmk#1476: Well, eleuther didn't exist in 2007 [citation needed]
Igor Krawczuk#1653: I have some answers, but I need to find citations lest I just soapbox^^
triggerhappygandi#0001: The reddit and it's consequences too
IKEA#9631: Post-2015 reddit you mean
IKEA#9631: before that it was fine
nmkd#1425: reddit is fine if you know what subs to subscribe to
triggerhappygandi#0001: You can say that about twitter too
bmk#1476: Algebraic geometry twitter
bmk#1476: I love being bombarded by math i don't understand
triggerhappygandi#0001: There are a lot of cool people on twitter. And yeah, follow only topics like these
triggerhappygandi#0001: I doubt MIT CSAIL would bombard your feed with politics.
triggerhappygandi#0001: Which reminds me, if anyone follows Yann LeCun, don't. He talks 90% of the time about politics.
Aran Komatsuzaki#5714: Sam Bowman and Delip Rao bombard my twitter feed with politics lol |
I agree with them, so it's not a problem tho
triggerhappygandi#0001: And the rest about something from FAIR
triggerhappygandi#0001: You are followed by Karpathy and hardmaru. Your feed is probably very cool @Aran Komatsuzaki
nz#9710: I think he discussed more ML stuff before the Gebru thing
triggerhappygandi#0001: No. Been following for a year now
mgostIH#0245: https://cdn.discordapp.com/attachments/729741769738158194/798286834219614228/759590894595997746.png
nz#9710: Yea hahahah I found Aran in a hardmaru tweet yesterday
triggerhappygandi#0001: And he literally only talks politics/fair
triggerhappygandi#0001: Yoshua Bengio isn't even on twitter lmao
IKEA#9631: Tbh Ive been on reddit since 2013 but Im on the verge of nuking my accounts and blocking myself from using it ever again (or at least as much as possible)
Theres just too much cancer and drama in that place for me
And its fucking with my dopamine levels too, most likely
triggerhappygandi#0001: And Geoff Hinton is too old for all this
triggerhappygandi#0001: Indeed. I too was caught in the upvote frenzy for a while
triggerhappygandi#0001: Same with Quora tbh
triggerhappygandi#0001: Not worth it
Aran Komatsuzaki#5714: i just tweet some interesting papers earlier than anyone else. i'm a self-proclaimed twitter ai journalist lol
triggerhappygandi#0001: I am very happy with discord being my only social media for the most part
bmk#1476: Does anyone actually like quora
triggerhappygandi#0001: I don't. |
bmk#1476: I mentally bucket it with pinterest
mgostIH#0245: You need to have a very high IQ to understand Quora
IKEA#9631: what did you "replace" reddit with
triggerhappygandi#0001: It used to be cool
nz#9710: It was relatively good, now it's a self promotion shithole
triggerhappygandi#0001: Discord lol
bmk#1476: So what you're saying is that quora is the website that can only be understood by Elon Musk
triggerhappygandi#0001: I was done with Quora in 2016 when I saw an answer with 50000 upbotes to a question "What made you happy today"
triggerhappygandi#0001: :zucc:
IKEA#9631: tbh im not sure Dsicord is much better than reddit lol
bmk#1476: I think people here have a somewhat distorted view of the average discord server tbh
bmk#1476: This server is in the top few percentile in terms of quality
StellaAthena#3530: If you’re looking for a productive time sink, I recommend MSE or Wikipedia
triggerhappygandi#0001: You can always exit servers that you feel are shit. I'm content with all the aevrers I'm in @IKEA
bmk#1476: Or eleuther
triggerhappygandi#0001: And eleuther is better than my formal education lol
IKEA#9631: What kind of servers you in?
triggerhappygandi#0001: Servers like this
mgostIH#0245: only ml related?
triggerhappygandi#0001: And some warhammer 40k ones |
nz#9710: I think we should give the "twitter, but with citation counts" a shot, and I say this as someone with no citations :berk:
bmk#1476: I'm in all of the rat discords
triggerhappygandi#0001: What is a _rat discord_
Igor Krawczuk#1653: I want it to be discords dedicated to individual species of rats
triggerhappygandi#0001: How do you sink time in Wikipedia? I just look up what I want and exit
triggerhappygandi#0001: I want it to be about rats doing rat things
Igor Krawczuk#1653: MSE=math stack exchange?
mgostIH#0245: https://twitter.com/RealAnonTheAnon/status/1237509036318359555
bmk#1476: https://cdn.discordapp.com/attachments/729741769738158194/798288753956356126/Screenshot_2021-01-11-13-34-04-635_com.discord.png
bmk#1476: Rat discords
triggerhappygandi#0001: Invite me to one of these rabbit holes
bmk#1476: I'm also in EA corner but it didn't fit in the screenshot
Igor Krawczuk#1653: Can I ask you a politically incorrect question about those types of rats?
Igor Krawczuk#1653: Depending on how much you hang out in those communities?
StellaAthena#3530: Edit it
StellaAthena#3530: Yes
triggerhappygandi#0001: You edit Wikipedia for fun? @StellaAthena
bmk#1476: ..go ahead?
triggerhappygandi#0001: Hah. _nerd_ @StellaAthena
StellaAthena#3530: Yeah. |
triggerhappygandi#0001: Calm your productivity down lol. You contribute meaningfully even in spare time?
Igor Krawczuk#1653: So I used to be pretty hard into lesswrong, EA etc. and still have a lot of EA friends, but I kind of dropped out once Eliezer and a Kind of cult of genius, cult of IQ tone began to spread. Has that been contained/grown out?
bmk#1476: I spend the most time in SSCD and opinions on eliezer are mixed
bmk#1476: Absolutely not an eliezer cult
bmk#1476: Dunno about the more LW focused discords
bmk#1476: It's not even an SSC cult despite the namesake
bmk#1476: Also fwiw I'm not massively into LW in itself, I'm more interested in the general cluster of people who are interested in LW
bmk#1476: And I'm not into EA at all tbh
Igor Krawczuk#1653: I get you on both of these
bmk#1476: Though i *do* happen to have very typical-LW stances (i.e Rationalism, life extension, AI alignment)
Igor Krawczuk#1653: I mean, you can believe in all of that and be reasonable
nz#9710: oh life extension, cool
nz#9710: any posts you recommend reading up?
CRG#8707: https://nintil.com/longevity/
StellaAthena#3530: If you think that’s impressive, I’ve written four academic papers (2 published, 2 under review) in the past year in my free time lol. This is the main reason I’ve cut down my contribution to MSE – it’s a lot more fun to do research in your free time.
Igor Krawczuk#1653: So if I remember my osmosis correctly some of that productivity will diffuse into me if I work with you and lucidrain right?
triggerhappygandi#0001: If this is what you do in free time, what do you do in _not_ free time
StellaAthena#3530: Industry research that I’m largely not allowed to publish
triggerhappygandi#0001: That's literally why I'm here tbh
triggerhappygandi#0001: I will soak up all their talent someday |
triggerhappygandi#0001: You working for CIA or something @StellaAthena
StellaAthena#3530: The “or something” side of that, but yes.
StellaAthena#3530: I design algorithms for people who use algorithms to catch terrorists and cyber criminals.
IKEA#9631: How tf do you even get a job like that
StellaAthena#3530: In my case the answer is “get lucky” lol
StellaAthena#3530: My high school calc teacher was 8 years older than me and a rather cool guy, and me and some classmates hung out after school and played board games with him. While I was off at college he got a job at one of the big DoD and Intelligence consulting firms. He recruited me after I graduated.
triggerhappygandi#0001: Working for US government
triggerhappygandi#0001: I bet you've seen some shit.
gwern#1782: "job history or sexual abuse survivor testimony?"
StellaAthena#3530: @triggerhappygandi Not really. I’m a mathematician. I do math. My interaction with classified info is mostly the fact that my validation datasets are hush hush.
IKEA#9631: Also I bet the salary is much better than the "equivalent" job in my country (france)
Which is, i shit you not, usually not too far off from minimum wage
StellaAthena#3530: My job is genuinely not that different from a mathematician who works with the CDC to do pandemic contact tracing. It’s just that the nodes of interest in my clients’ applications are Persons of Interest rather than people with an unusual strain of the flu
StellaAthena#3530: (Or COVID, nowadays I guess)
StellaAthena#3530: I started at 60k USD/year. That’s rather low for someone with a tech job straight out of college from a prestigious institution, but I’ve had rapid salary growth and 4 years later make 93k/year
IKEA#9631: dayum
triggerhappygandi#0001: Man
StellaAthena#3530: That’s 30 and 46.5 per hour, if that’s more common in France.
triggerhappygandi#0001: Aren't you enrolled in MS at GA tech? @StellaAthena
StellaAthena#3530: Yes. I’m doing that part time. |
triggerhappygandi#0001: I just assumed you had enrolled in MS straight out of BS.
StellaAthena#3530: In most industries there is no reason to do that.
IKEA#9631: As a junior researcher for a government entity in paris youd be lucky to make like 30k€ lol
Fun fact: housing prices per square feet in paris are about the same as NYC
triggerhappygandi#0001: Yeah. Paris is expensive af
triggerhappygandi#0001: A broom closet costs €1k/month
StellaAthena#3530: I work for a private company, not the USG. I would be making around 50k if I worked for the USG
StellaAthena#3530: Which is more than in France, but not that much
IKEA#9631: Most Phds just fuck off to other countries, research is absolutely laughed at here (most of the case)
triggerhappygandi#0001: Didn't know you could do MS/PhD as part-time
Igor Krawczuk#1653: Your 93k is with or without benefits/health insurance? @StellaAthena
StellaAthena#3530: Without.
triggerhappygandi#0001: No employee health insurance? @StellaAthena
triggerhappygandi#0001: I thought it was common in US
triggerhappygandi#0001: To get it from employer
StellaAthena#3530: Err. Is the question “do you also get benefits” or is it “does that number include the estimated value of your benefits”
Igor Krawczuk#1653: The latter, assuming that you do
StellaAthena#3530: I have benefits (incl. healthcare) *on top of* that number
StellaAthena#3530: Most of the benefits are medicore, but as someone who maxes out their health insurance by April every year the health insurance is a godsend.
triggerhappygandi#0001: Next you gonna tell me you have your own house without loan aswell? |
StellaAthena#3530: Lol no.
triggerhappygandi#0001: For a moment I thought you'd say yes
StellaAthena#3530: I was fortunate that my parents paid for the entirely of my undergrad education though. No loans there
triggerhappygandi#0001: So you have no liabilities?
triggerhappygandi#0001: Living the life
StellaAthena#3530: Other than my massive pile of chronic health conditions, no
triggerhappygandi#0001: Oh damn
triggerhappygandi#0001: Anything serious?
StellaAthena#3530: I was suidically depressed six months ago, have trouble functioning without Adderall, and have seizures randomly that my doctor hasn’t figured out the trigger for or cause of.
StellaAthena#3530: Also debilitating chronic migraines
triggerhappygandi#0001: :zucc:
Sid#2121: writing 4 papers in your spare time with chronic migraines is incredibly impressive lol
Igor Krawczuk#1653: Wow, thanks for sharing. I hope your doctor figures it out soon
Igor Krawczuk#1653: Yes, that was my next thought
triggerhappygandi#0001: I'm not a therapist or anything, but pills are probably hurting you more than helping @StellaAthena
triggerhappygandi#0001: They give pills for everything in US.
StellaAthena#3530: The migraines are new. Started this past winter. Technically it’s only effected the Pile.
StellaAthena#3530: I am alive because of the pills. They are absolutely helping far more than they hurt.
Sid#2121: that really sucks. Also for the sads, try 🍄 with someone you trust
triggerhappygandi#0001: Get well soon, I guess. But rely a little less on them nonetheless @StellaAthena |
Sid#2121: or ketamine, apparently, although i haven't tried it myself
StellaAthena#3530: oh “sads” = sadness. I read that as “sids” at first and was very confused about how you could think I had a chronic case of sudden infant death syndrome
Sid#2121: lmao
Sid#2121: *extremely* delayed sudden infant death syndrome
triggerhappygandi#0001: Isn't ketamine something they give to horses to tranquilize them?
Igor Krawczuk#1653: Might want to be a bit more careful with telling people what to do about their medical issues😅 Stella seems to be chill about it but my sister hates it when people start offering her advice for her depression. Especially when not asked
Sid#2121: yes, although recently it's been found to have positive effects treating depression https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6717708/
Sid#2121: ^
StellaAthena#3530: My psychiatric conditions are not things you “get better” from. My brain produces inappropriate amounts of certain neurotransmitters in response to various stimuli. The drugs promote the production of them so I can get back to what most people consider normal
triggerhappygandi#0001: My bad. Apologies
StellaAthena#3530: *Especially* when someone tells you that they were recently suicidally depressed and aren’t anymore. I’m not remotely exaggerating when I say that unsolicited advice like that can kill people.
Sid#2121: apologies if my advice also came off that way lol, i'm just letting you know what helped for me
triggerhappygandi#0001: Damn. In that case, double apologies @StellaAthena
triggerhappygandi#0001: You wrote 4 papers with all of that while I have 0. Gotta learn how to be more productive
StellaAthena#3530: What do you do in your free time? How do you “waste” time?
triggerhappygandi#0001: YouTube, sometimes
Sid#2121: if you want to learn how to waste time, i am very strong in this area
Sid#2121: happy to teach you :berk:
3dprint_the_world#6486: I'm a pro crastinator
mgostIH#0245: I found a good way to waste time is exercising |
3dprint_the_world#6486: (sorry)
mgostIH#0245: That way I still waste time but I get ripped
Deleted User#0000: ive been getting better at wasting my time too
Deleted User#0000: mostly VR, and worrying about whether im wasting too much time
Sid#2121: extreme dad joke, i love it
Igor Krawczuk#1653: Speaking of YouTube, any recommendations for more channels like Yannic kilchers? ML Lectures/explanations I can listen to while coding?
IKEA#9631: Sometimes I get extremely interested by random very specific topics to the point where I lose sleep because of them, then a couple of weeks/months later I lose interest and just watch youtube or whatever
Then the cycle repeats 😐
Sid#2121: this is an absolutely top tip to waste more time - worry about whether you're wasting too much time, and procrastinate on something else to keep your mind off it
triggerhappygandi#0001: My main problem is that I can never sleep before 3am, and between 12am-3am I do literally nothing productive. It's either being on discord like rn, or youtube. Then I wake at 9-10 and before long half the day is gone, without me having done pretty much anything @StellaAthena
Sid#2121: @StellaAthena hope you're taking notes
triggerhappygandi#0001: Meirl
Deleted User#0000: i started worrying about whether i was wasting too little time, and then switched to the above strategy
StellaAthena#3530: What time zone are you in?
Deleted User#0000: the result is the same
triggerhappygandi#0001: Deepmind's lectures?
triggerhappygandi#0001: @StellaAthena GMT+5:30
triggerhappygandi#0001: It's 3:15am here
StellaAthena#3530: Lol
Deleted User#0000: i havent woken up before 1pm in the whole pandemic pretty much lol |
Sid#2121: pretty sure the whole world's internal clocks have shifted about 3 hours forward
triggerhappygandi#0001: Man. You must be working at nights then @Deleted User
Deleted User#0000: tbh i was going in that directoin with VR. I think the pandemic has had like 0 effect in my life style, except for being easier to hang out with more friends, coz they are now following my online life style
Deleted User#0000: yeah, im often working on stuff till 4-5am, sometimes till i just drop ded lol
Deleted User#0000: i procastinate "in the morning" more
Deleted User#0000: i donno why
triggerhappygandi#0001: My problem is that I get into bed at night, and don't do jack shit till I sleep
triggerhappygandi#0001: I'm feeling bad even now but I know I won't do much about it. Which comes back to this
triggerhappygandi#0001: I joined this server to soak up some productivity from the people here tbh
Daj#7482: I'm the same, spend the hours thinking about life or doing insight meditation or something
Deleted User#0000: btw somewhat related to all this. I watched a podcast with Andrew Huberman, which is great if you wanna be able to control your life, whether it is to get more productive or less xD
Daj#7482: become :bigbrain: by boredom
Deleted User#0000: but it was a great podcast i found
Daj#7482: No please, we have so little to begin with, we can't spare any
triggerhappygandi#0001: @Daj I could really use lsd rn. I've heard its wack, especially for a situation like this
Deleted User#0000: https://www.youtube.com/watch?v=SwQhKFMxmDY&ab_channel=RichRoll
triggerhappygandi#0001: You think I'm giving you a choice
Deleted User#0000: also i think his background story is very cool/interesting and he's got lots of wise advice/stuff
Daj#7482: LSD can solve a surprising amount of issues in one go, but don't take it lightly, don't take it alone, set and setting etcetc
Deleted User#0000: i also wanna try stuff |
Sid#2121: solve is a strong word but it's defo undervalued
Daj#7482: I've seen it happen plenty of times
triggerhappygandi#0001: I've never done anything hardcore tbh. What's the deal with not doing it alone? In case you choke on something?
Daj#7482: Not like, solve _everything_
Daj#7482: but you know
Deleted User#0000: ive heard people talk good of mdma or something too
Daj#7482: It's hard to explain, but at risk of sounding like the schizo I am: It's a shamanic journey in a piece of paper, bring a guide
Sid#2121: i've experienced it but it's not like it *solved* those problems i had *forever*
Sid#2121: i felt great for like, 6 mos or so after
Daj#7482: Oh yea lol MDMA is called ecstasy for a reaosn
Deleted User#0000: i wished i could download mdma
Daj#7482: My model of LSD is that it's a neural annealer. It raises the temperature of the brain (in a sampling sense) and then you anneal to a better ground state
Daj#7482: But you can gradient ascend back out
Daj#7482: Or get stuck in a malignant local optimum
Deleted User#0000: andrew huberman talks a little bit about these drugs in the podcast
Deleted User#0000: i highliy recommend it, and/or the one with lex fridman
Daj#7482: Oh that's where I heard Huberman before
Sid#2121: yep, have seen this too lol
IKEA#9631: One of the rare times where the guest is more jacked than Lex
Daj#7482: ***MACHINE ELVES*** |
IKEA#9631: The only other time being that Intel engineer i forgot the name of
Sid#2121: i am not quite sure what you're referring to, and yet i agree
Deleted User#0000: i guess you need to ensure you still have a goal while u do this, so that there is some gradient signal
Deleted User#0000: lel
Daj#7482: Actually more of a DMT thing, long story
StellaAthena#3530: I think that productivity depends a lot on your personality. I’m always thinking; I have trouble turning my brain off, whether that’s thinking about Pokémon or thinking about math.
Moving resources for procrastination out of easy reach and moving resources that encourage productivity within each reach was huge for me. Most of my bookshelves are stacked with philosophy books or science books, with the fiction series packaged away. I have pens and scratch paper scattered throughout my apartment. My bookmarks are things like the EAI GitHub repo, shtetloptimized, and interesting papers to read. When I close a browser it automatically logs me out of my Wikipedia, Math SE, Facebook, and etc. accounts.
On the social media front, I’ve found setting a norm that I only use it on my phone really helpful actually. Reddit just *isnt something I do* on my laptop so it doesn’t distract me from coding. I’ve also bunched my social media apps in the same spot on my phone as more productive ones, like lichess and discord (I’m in only a few and they’re all high productivity places).
But I think it’s important to design the systems that work for who you are and your lifestyle.
Deleted User#0000: im gonna gree to not feel left out
Daj#7482: It's like massively jacking up your learning signal and letting gradient flow everywhere (normally the brain super restricts where gradient can flow), so you can update recalcitrant parts of your brain. Kaj Sotala has fantastic posts on this kinda stuff
Daj#7482: big +1 on "design the system that work for who you are and your lifestyle"
Daj#7482: Some people literally need the exact opposite thing to be productive
Deleted User#0000: in my case, i actually have been trying to fight against workaholism in the last few years
StellaAthena#3530: The pinned pages on my phone are 2 productive sites and 1 leisure site https://cdn.discordapp.com/attachments/729741769738158194/798310080445153290/image0.png
Deleted User#0000: i had a realisation i had "forgotten how to have fun", and decided it was coz workaholism
Deleted User#0000: and i think ive managed to get better at it |
StellaAthena#3530: The “reading list” brings up papers
Daj#7482: Can't relate to any of your productivity issues lol
mgostIH#0245: @StellaAthena do you like swsh
Daj#7482: (by which I mean, I'm unproductive for my own idiosyncratic, stupid reasons)
Deleted User#0000: but u still want *some* gradient signal, even if it just comes just after the lsd experience/kick
Daj#7482: You just have to be careful what you update and how hard
Deleted User#0000: i think andrew said this when he briefly mentioend these drugs
StellaAthena#3530: That’s very reasonable too. The only reason why what I’m doing is healthy is that I genuinely enjoy doing this stuff – like would choose to do it indefinitely if I had infinite money and no needs – and I set designated blocks of time aside for human contact and working out.
triggerhappygandi#0001: @StellaAthena yeah but how do I sleep at a meaningful time lol
mgostIH#0245: @Daj I think you are forgetting something important in your analogies
Daj#7482: https://www.lesswrong.com/s/ZbmRyDN8TCpBTZSip
Great sequence that updated my models of how psychological well being works (and the best model of meditation's mechanism imo), also some good AGI ideas
StellaAthena#3530: Go see a doctor if you have a disturbed sleep pattern that significantly effects your well-being
Deleted User#0000: yeah i think its important to figure out what you want. Which ironically i think is my number 1 reason for unproductivity xD
Daj#7482: How much caffeine do you take?
Deleted User#0000: recursive worrying
triggerhappygandi#0001: None, probably
mgostIH#0245: Where's the Schmidhuber citations
triggerhappygandi#0001: I drink coffee infrequently
Daj#7482: Schmidhuber invented psychological well being in 1991 |
Deleted User#0000: https://cdn.discordapp.com/attachments/729741769738158194/798310948863606794/9k.png
Daj#7482: Caffeine has a 5 hour half life, did a lot to fuck with my sleep. Also lack of vitamin D
Daj#7482: and lack of exercise
Daj#7482: I take an hour or so walk every day
Daj#7482: Melatonin is great to reset schedule
Deleted User#0000: together wwith the theory of fun, curiosity, life, the unvirerse, and everything
Daj#7482: Don't take more than 1mg
StellaAthena#3530: I rock climb for two hours at least on d a week for the same reason
Daj#7482: Schmihuber is credited with inventing the number "42"
triggerhappygandi#0001: @StellaAthena https://cdn.discordapp.com/attachments/729741769738158194/798311244243009586/20210112_032949.jpg
triggerhappygandi#0001: (dark souls reference)
Daj#7482: Please leave it with Stella otherwise nothing gets done here
triggerhappygandi#0001: Very relevant
StellaAthena#3530: Also, if you’re having trouble getting the energy to do athletics, incorporate it into your life. At the very least you can make your sex life highly physically demanding.
Deleted User#0000: lex friedman has a very unrelatable video about his very productive day-cycle
triggerhappygandi#0001: I try to get atleast an hour of excercise every day
Deleted User#0000: hmm. the only issue is that im the worst at making habits
Daj#7482: Are you seriously suggesting people here have a sex life?
Deleted User#0000: ^
triggerhappygandi#0001: Yeah lmao |
IKEA#9631: you guys have a sex life?
triggerhappygandi#0001: I am just going to steer clear
Deleted User#0000: is sex life, a new version of second life?
IKEA#9631: sex life 2
Daj#7482: I go to the monthly sex dungeon meetups at MIRI
triggerhappygandi#0001: Forget sex life, you have social life?
StellaAthena#3530: Are you guys from places where nerds *don’t* get laid easily?
Deleted User#0000: kinda, thanks to vr
Deleted User#0000: but i still only go there sometimes coz i find it easier to either work or be on twitter/discord
triggerhappygandi#0001: I would chalk it up more to my unwillingness to open up
Daj#7482: No I'm just weirdly asexual in the most ADHD way possible
guac#4716: @Deleted User yo i just played BOTW for the first time this weekend I'm fucking hooked fml
Daj#7482: also, we get it, you have sex
Daj#7482: lol
Deleted User#0000: im just too shy/introverted/and socially anxious
Deleted User#0000: and workalholic, and so dont go to places often
Deleted User#0000: and stuff
triggerhappygandi#0001: If you ask me to talk to someone face to face, I find it a chore that I'd rather not do if I can get away with it
triggerhappygandi#0001: @Deleted User are you me?
Deleted User#0000: maybe |
StellaAthena#3530: IDK. I have the opposite problem. I have trouble befriending people I haven’t fucked
Daj#7482: Just shows how different various people are
Deleted User#0000: i had my first smooch from someone ever in vr, from a very liberal furry who would probably do it with anyone who crossed his way xD
Deleted User#0000: ud have to put it that easy, for i to even begin to have the slightest chance xD:PPP
Deleted User#0000: i donno
Deleted User#0000: anyway
StellaAthena#3530: I mean, in my case it’s due to serious interpersonal issues due to years of abuse, but you’re not wrong.
triggerhappygandi#0001: I legit talk more to you guys than anyone of my friends irl
IKEA#9631: you guys have friends irl?
zphang#7252: the true friends were the... friends we made along the way
Daj#7482: I'm sorry, I wish there was something more substantial I could say
triggerhappygandi#0001: Lol no
Deleted User#0000: whats irl?
StellaAthena#3530: Not since COVID
Daj#7482: https://www.youtube.com/watch?v=Hyw6kKMjp5A
triggerhappygandi#0001: But this server is cooler than irl people I know
triggerhappygandi#0001: So win win for me
Deleted User#0000: i agree tho
Daj#7482: fr though I wish I could all introduce you to my friends
guac#4716: making new friends after 24 feels significantly more difficult |
Daj#7482: I'm one of those people that serially builds big friend groups
Daj#7482: D&D helps a lot
Deleted User#0000: i could easily make friends in vr. but you guys are more cool so yeah
triggerhappygandi#0001: So I've heard. That's why I got into warhammer
triggerhappygandi#0001: But I'm still cripplingly introverted
Deleted User#0000: and by eaisly, i just mean more easily than irl
StellaAthena#3530: I’ve found that the weirdest thing about friends post-college is that it takes work. In college I had friends in my dorm, friends in my classes, friends I studied with, friends I ate with. It used to take no work to hang out.
Daj#7482: I have very little useful advice to give unfortunately other than I think you guys are cool and we should play D&D sometime
IKEA#9631: I feel like I'm too non-nerdy to be friend with nerds my age but also too nerdy to be friend with non-nerds
triggerhappygandi#0001: Anyway, this is a very cool server. I now know a lot of smart people than I did 5-6 months back
StellaAthena#3530: Now between commute (1) + work (8) + food (2) + sleep (8) that’s already 19 hours a day that are *not* for socializing. And the remaining 5 have to include grocery shopping and working out and reading and everything else you do
triggerhappygandi#0001: Server meet-up in NeurIPS 2021 hopefully
Deleted User#0000: ive u guys value socializing per se, i recommend vr
Deleted User#0000: u just drop in and theres people
Deleted User#0000: kek
Daj#7482: Yea I often have to schedule a "socialize day"
Daj#7482: Which is usually D&D
Deleted User#0000: i scheduled a "socialize month"
Deleted User#0000: but ur approach sounds more sustainable
Daj#7482: I'm someone that needs facial expressions |
triggerhappygandi#0001: I can do an hour at max
Daj#7482: Online interactions creep me out after a while
guac#4716: @StellaAthena Yuuuuuup growing old is lonely. Maybe i should start a family lollll
Deleted User#0000: then buy the htc face tracking module which isnt for sale :sadface:
Daj#7482: But I wanna try VR socializing
Daj#7482: just for the memes
Deleted User#0000: otoh, people in vrchat have facial expressions they learn to puppeteer
triggerhappygandi#0001: Excuse me but 2 hrs of food? I am entirely foreign to cooking your own food
Deleted User#0000: an intresting art
StellaAthena#3530: Yeah, fair. More like 3
Daj#7482: I have a feeling I will encounter the same kind of people in VR I used to encounter at conventions
Deleted User#0000: just meme it out
Deleted User#0000: which conventions?
Dromarion#3383: How do people even date now? There's no where to go. VR dating isn't even a meme at this point.
IKEA#9631: i love cooking but 2h per day seems like an awful lot
Daj#7482: I went to many in my younger years for the meme lol
Deleted User#0000: i see
Daj#7482: Me and a friend had a car and wanted to see cringe
Deleted User#0000: ihavent gone to conventions
Daj#7482: Anime, bronies, BDSM... |
triggerhappygandi#0001: I can't cook anything more than minute noodle
Deleted User#0000: yeah in vr you will find lots of cringe
Deleted User#0000: for sure
IKEA#9631: even in 15 minutes you can do so many things
guac#4716: @Dromarion I've been videocalling babes since the 3rd grade
Deleted User#0000: but some of it can be fun. and then you find the occasional interestnig person
StellaAthena#3530: For cooking + eating? Eating dinner takes 45 minutes alone...
Daj#7482: Oh don't get me wrong the bronies were fucking lovely
triggerhappygandi#0001: So you've seen the bottom of the degeneracy barrel
guac#4716: you mean Omegle?
Daj#7482: But my god can I not handle it for too long at a time
IKEA#9631: 45 minutes???
Deleted User#0000: the furries in neosr are lovely and cute too
Daj#7482: tbh, it wasn't nearly as bad mostly
StellaAthena#3530: You mean onlyfriends circa 2006?
Deleted User#0000: neosvr has a higher concentration of cool people than vrchat fyi
Deleted User#0000: tfw huel
triggerhappygandi#0001: I mean, you're probably right. I just hate furries because internet says so. @Daj
Daj#7482: I have lots of terribly funny bad experiences with the furries
Daj#7482: They used to have like a club at my university |
IKEA#9631: im too poor for that
Deleted User#0000: diy huel
Deleted User#0000: can be cheap
Deleted User#0000: i guess
StellaAthena#3530: What was Omegle supposed to be for? I genuinely don’t know. Me and my friends just used it to look at strangers’ genetalia.
IKEA#9631: well thats just called cooking lol
Singularity#9001: Ah yes, genitalia research
Daj#7482: :virgin:
Deleted User#0000: lol maybe
Daj#7482: culture your own hate
Daj#7482: doN't let others tell you what to hate
Daj#7482: Organic homegrown hatred :chad:
triggerhappygandi#0001: What? Now you will say I can't hate anime?
Singularity#9001: If you hate everyone equally that's called equality boom
triggerhappygandi#0001: Smh
Deleted User#0000: in the last year ive realised theres enough furries in ML/Ai we could form a whole institute
Deleted User#0000: which is cool
Singularity#9001: Furry intelligence institute
StellaAthena#3530: Oh totes
Deleted User#0000: FurriesinML should be a thing next neurips |
Daj#7482: Oh no anime is objectively the most hateable thing in existence
Deleted User#0000: *no fun allowed*
triggerhappygandi#0001: I have had very, very autism inducing experiences with Omegle
Daj#7482: Furries run like the entire internet
IKEA#9631: well theres like hundreds of thousands of furries worldwide, at least
Daj#7482: I'm not crazy, it's a conspiracy!
StellaAthena#3530: What does this even mean
Deleted User#0000: In one social network I say I belong to the "Eleuther Catgirl research Global Institute" or something like that
Deleted User#0000: lel
Deleted User#0000: true tho
Daj#7482: Cease
Deleted User#0000: nu
triggerhappygandi#0001: Talking to a teenager who goes "PICKLE RIIIIICK" and screams very loudly @StellaAthena
Daj#7482: I expect a truly unprecedented number of people to spontaneously manifest as furries once morphological freedom is achieved
IKEA#9631: Funniest shit i ever seen
StellaAthena#3530: What does this have to do with autism spectrum disorder?
Deleted User#0000: this
triggerhappygandi#0001: Nono
Daj#7482: The literal oldest piece of art ever is furry art lmao
Deleted User#0000: lol |
Deleted User#0000: which one
triggerhappygandi#0001: It's just that in internet lingo it's called autistic @StellaAthena
Singularity#9001: Furries are just the logical next step in bodily forms, we become our cats
Daj#7482: The lion man statue thing
Daj#7482: https://en.wikipedia.org/wiki/Lion-man
Daj#7482: Furries invetned art
triggerhappygandi#0001: Wow
triggerhappygandi#0001: Don't let them know this
Deleted User#0000: XD
Deleted User#0000: too late
Daj#7482: unclear if the artist still takes commissions or not
triggerhappygandi#0001: This should be kept guarded
guac#4716: @StellaAthena i think i lost my virginity on omegle
triggerhappygandi#0001: Please don't let this get out
Singularity#9001: This info needs to be guarded
triggerhappygandi#0001: I am not prepared for furries _not_ being ashamed of their fetish
cfoster0#4356: *walks in*
cfoster0#4356: *walks out*
Daj#7482: Oh boy do I have bad news for you. ~~_This is the future liberals want_~~
guac#4716: being a furry is a genetic trait selected for animal tamers |
IKEA#9631: https://cdn.discordapp.com/attachments/729741769738158194/798315795922944020/giphy.png
triggerhappygandi#0001: https://tenor.com/view/simpsons-bart-simpson-grampa-simpson-old-man-hi-bye-gif-8390063
IKEA#9631: LEL
triggerhappygandi#0001: Lol
IKEA#9631: i copied the fucking thumbnail not the gif
IKEA#9631: reeeee
Daj#7482: Ya'll not ready for the post singularity bullshit explosion
Daj#7482: If you think humans won't use superintelligence for sex y'all naive
triggerhappygandi#0001: I am unironically not @Daj
Singularity#9001: I have a furry in my college dorm they are... interesting... I have a feeling the deeper into internet you go the more likely that happens. Perhaps a furrification theory of exposure to the internet uninhibited since early ages?
triggerhappygandi#0001: Have people actually thought about it?
Daj#7482: No one is
Daj#7482: Counter theory: Survivor bias. The furries with enough social skills to be normal will hide it, only the freaks that can't compensate are noticeable proportionally
triggerhappygandi#0001: Not having to work will call for a major reevaluation of everything
Daj#7482: This isn't even in my top 10 concerns
triggerhappygandi#0001: What is your top concern then
triggerhappygandi#0001: Unironically
guac#4716: This discord is how I imagine walking through the bazaars of ancient rome feels like. You got furries in the general. you got income talk in off-topic. What a rush.
Deleted User#0000: https://twitter.com/caudamus/status/1253733229745467392
Daj#7482: Misaligned AGI converting every accessible joule of energy into hyper optimized matter made only to simulate the maximum amount of suffering sentient minds |
Deleted User#0000: hmmm
Deleted User#0000: sounds fair theory
IKEA#9631: theres a galaxy meme to be made somewhere
triggerhappygandi#0001: Wack. I haven't ever actively tbought about misalighned agi
Deleted User#0000: i guess im one of the weird ones
IKEA#9631: or a increasingly verbose meme
Daj#7482: It's a mental abyss of sorts, but it's what I specialize in
Daj#7482: If you thought Cthulhu is scary oh boy
Deleted User#0000: i recently got this from a top ML professor https://cdn.discordapp.com/attachments/729741769738158194/798316740135878676/unknown.png
Deleted User#0000: probably coz i post both furry and ML stuff on my twitter which he follows xd
triggerhappygandi#0001: This is an interesting rabbit hole. Tell me more @Daj
guac#4716: no way roy is a furry
Daj#7482: yea lol that's cringe ngl guille
triggerhappygandi#0001: Especially now that you invoked Lovecraft.
Daj#7482: Love you but ye
Deleted User#0000: hecc sounds bad
Daj#7482: Oh god I'm too tired tonight
Daj#7482: I'm not feleing the doomsday prophet
Daj#7482: But I will another day, don't worry
triggerhappygandi#0001: No problem. But some other time. |
guac#4716: sit have a beer let's hear it connor
Deleted User#0000: >.<
triggerhappygandi#0001: I want to see where this rabbit hole goes.
IKEA#9631: https://en.wikipedia.org/wiki/LessWrong#Roko's_basilisk
Daj#7482: My entire life is basically irl lovecraft at this point
Daj#7482: Roko's basilisk is babby's first infohazard
triggerhappygandi#0001: Are you a lovecraftian protagonist?
Singularity#9001: I can't wait to maximize universal suffering
triggerhappygandi#0001: "I have been warning them not to challenge the cosmos"
Daj#7482: Your profile pic was the first thing I noticed when you followed me lol. It sucks because I mean if it makes you happy, that's great, it's a shame that society has so many weird arbitrary rules
triggerhappygandi#0001: "yet they have created images from text"
Daj#7482: You'd be surprised how much I am
IKEA#9631: babby?
Deleted User#0000: i kinda do it a bit as a protest, but only minorly
Daj#7482: I even have like, a story of a cursed bloodline and mental illness
triggerhappygandi#0001: You what
Daj#7482: baby, as in the smallest
Daj#7482: ye long story lol
IKEA#9631: oh
Daj#7482: I come from an old irish family |
Daj#7482: We're cursed and shit
triggerhappygandi#0001: Now you're actively trying to be a lovecraftian protagonist
bmk#1476: catgirl theory
Daj#7482: I actively LARPed like this as a teenager, then I tried to stop, but it didn't stop lol
Deleted User#0000: also to not let my mind be too closed. i like experimenting like that
Singularity#9001: Do you guys ever imagine that you walk into a room and then everything collapses into the corner like a folded piece of paper
Daj#7482: You are far less machiavellian than my chimp brain lol
Daj#7482: is this a paper mario reference
triggerhappygandi#0001: Always appreciate a good larp
Singularity#9001: I'm actually from the paper mario universe
Deleted User#0000: what u mean?
Daj#7482: It's not larping anymore lol, seeing as I basically work on summoning all powerful alien entities into this reality and controlling them
Daj#7482: I don't experiment publicly, way more controlled
Deleted User#0000: whats LARP? my mind tried to fit "Lovecraftian Role Play" into it but lol
Daj#7482: Live Action Role Play
Deleted User#0000: ah. yeah i guess theres benefits to both
Daj#7482: I have a fucking great story about my favorite video game
Daj#7482: Darkest Dungeon
Deleted User#0000: i have a big bias towards open
Daj#7482: (my profile picture is me in that style) |
Singularity#9001: To what degree is singularitarianism just a new form of secular techno religion versus an observation of what our increasing technological pressure will logically lead into
Daj#7482: It's a lovecraftian game about your ancestor having done terrible experiments and unleashed Cthulhu and whatever
Daj#7482: The ratio depends on the person
Deleted User#0000: i should eat something
triggerhappygandi#0001: You are a proper 4chan fed guy. And yet a smart guy. Brag on /b/ about it lol @Daj
Singularity#9001: Sometimes I feel upset about eating things because it's like I'm consuming a work of art
triggerhappygandi#0001: Tell /b/ how you created gpt-2
Deleted User#0000: ephemeral art
Daj#7482: I used to go on 4chan until something happened that made me never wnat to go on 4chan again
Deleted User#0000: food is like an edible sand castle
Daj#7482: I turned 18
triggerhappygandi#0001: I still go there sometimes.
triggerhappygandi#0001: To see the bedrock of the internet
Deleted User#0000: ive never really sued 4chan
Deleted User#0000: or reddit
Daj#7482: Absolute garbage is what it is
Daj#7482: Good memories lol
triggerhappygandi#0001: Indeed
Singularity#9001: 4channism is like, if edgelordism became a website or if the trench coat kids made a club to trade anime and it just devolved
Deleted User#0000: i feel like i was too much of a "good kid" and never had all these fun experiences, and im only now changing that xD |
Deleted User#0000: well in the last several years
Deleted User#0000: kinda
Deleted User#0000: im having my rebel teenage years at 26yo xD
Daj#7482: 4chan is a website for children
Daj#7482: Just some adults didn't get the message
Deleted User#0000: yeah not like i wanna go there
triggerhappygandi#0001: Indeed. You can take literally nothing seriously on 4chan
bmk#1476: The basilisk is the infohazard equivalent of elon musk: everyone knows about it but it's not really that important compared to what else is out there
Daj#7482: They ruined the joke once /pol/ and siblings became a thing
Daj#7482: It was pure anarchic counter culture
Deleted User#0000: tell me some good infohazards
Singularity#9001: I think the internet makes it easier to get trapped into the child mindstate so tons of middle aged people's lives are just shitposting on 4chan along with actual children
Daj#7482: But they tamed the hivemind
Daj#7482: Not publicly
Deleted User#0000: ok
Daj#7482: I have been recently sharpening my infohazard policies
Daj#7482: At least read The Sequences, no infohazards before lol
triggerhappygandi#0001: I once saw a thread starting with "I am Alex Jones AMA", and I kid you not it had like 300 messages.
Deleted User#0000: okkk, fair
Daj#7482: +2 to Will saves against sanity damage from reading The Sequences |
Daj#7482: I think the causality is reversed: The people with genetics that curse them with immature brains gets stuck
Daj#7482: Other people grow up and leave naturally
Deleted User#0000: probably goes both ways
Daj#7482: probably
IKEA#9631: tbh i cant even imagine what could possibily worse than this
Igor Krawczuk#1653: connor, what if the sequences are the infohazard
Igor Krawczuk#1653: not being snarky
Daj#7482: :guilty:
triggerhappygandi#0001: I still believe most people using 4chan are normal, and just act really stupid online
Daj#7482: Then I'm already corrupted by the Elder Evil
Daj#7482: We're already in act 3
triggerhappygandi#0001: You gotta vent _somewhere_
StellaAthena#3530: For the record: I disappeared from this convo to go write a paper on using category theory to understand narratology.
triggerhappygandi#0001: Oh come on
StellaAthena#3530: Seemed worth mentioning given our discussion of productivity
Deleted User#0000: i was also reading paper
Deleted User#0000: but then discord happened
Daj#7482: I used to believe this, I have strong reasons to believe this is not true. 4chan has an _enormously_ high number of non standard psychologies
Daj#7482: Yea, go be like Stella @triggerhappygandi
Deleted User#0000: im meeting my supervisor tomorrow and all ive learn from this is that furries made the first work of art |
triggerhappygandi#0001: I don't understand that. What is non standard psychologies and how does it show they are not being productive irl
Deleted User#0000: and that i should read the sequence so connor tells me the dark secrets
Singularity#9001: I think like it's a concentration of dark triad persons
triggerhappygandi#0001: People on 4chan located the 9gag meme stone.
Daj#7482: Not saying non productive per se (though many are)
triggerhappygandi#0001: They can't be neckbeards in basement if they can do this
Daj#7482: 4chan is a shelling point for broken people, or at least suffering people
IKEA#9631: im pretty sure that 80% of those who post on 4chan are just regular people who just went there cause they were bored at some point or whatever (i know i did a few times) and the other 20% are the real bottom of the barrel rightwing incels you think about when you hear "4chan"
triggerhappygandi#0001: @StellaAthena I _will_ leech some of your productivity
Daj#7482: That's _exactly_ what you need to be to do that
triggerhappygandi#0001: But _how_
triggerhappygandi#0001: That was a very impressive feat
Daj#7482: Something something autism spectrum link to conscientiousness and high IQ
triggerhappygandi#0001: I don't associate intelligence with basement dwellers
Daj#7482: This isn't an academic argument
Deleted User#0000: i guess living in a basement you have more time to shitpost too?
Daj#7482: I just used to hang out on 4chan and knew a sample of ~n=200 people
Daj#7482: You just need a few outliers
Deleted User#0000: or to work/learn if you want
triggerhappygandi#0001: Yeah pretty much |
triggerhappygandi#0001: Most people on 4chan just shitpost
triggerhappygandi#0001: You can't do that on most social media
Daj#7482: I knew no one on 4chan that wasn't 1) A teenager, 2) Mentally ill, 3) In a terrible abusive life situation, 4) All of the above
Daj#7482: Oh actually I knew one girl that was just really into hentai
Daj#7482: lol
Singularity#9001: I think this also depends, like all of our versions of normal are skewed
triggerhappygandi#0001: Lmao what@Daj
Daj#7482: She drew such cute comics
Daj#7482: then I asked to see more of her stuff
Daj#7482: Damn girl drew some horse cock
Singularity#9001: 90% of 4channers likely do not post
triggerhappygandi#0001: You _know_ people from 4chan?
Daj#7482: used to
Daj#7482: when I was 18ish
triggerhappygandi#0001: I just assume they are my schizophrenia
Daj#7482: I met like 10 of them irl
Daj#7482: Got drunk as shit
Daj#7482: good times
Deleted User#0000: nice
Daj#7482: I outdrunk the russian |
triggerhappygandi#0001: Wack.
Daj#7482: I hung out on /tg/ a ton, played D&D with people
triggerhappygandi#0001: Ah /tg/
triggerhappygandi#0001: The most autistic place to be to discuss warhammer
Daj#7482: Yea but /tg/ is absolute top tier
Daj#7482: as far as sanity goes
Daj#7482: And still I never in my life met a more miserable lot of people
Daj#7482: I felt so fucking bad for them jfc
triggerhappygandi#0001: ~~More so than /b/?~~
Daj#7482: You could tell which ones posted on which boards
Daj#7482: It was comedic
Daj#7482: You could tell the assholes posted on /b/
Daj#7482: The really autistic shy ones posted on /mlp/
Daj#7482: The horny ones posted on, you can imagine where
Daj#7482: This was pre /pol/ I think
Daj#7482: Or right around its inception
triggerhappygandi#0001: Why is mlp targeted by such degenerates
Daj#7482: Nah dude the mlp people are cute
triggerhappygandi#0001: It is _supposed_ to be a kids cartoon
Daj#7482: I will defend the bronies as being by far the least bad of the bunch |
IKEA#9631: im getting flashbacks from when i used to watch mlp in high school, got found out and reacted the most autistic way possible
Daj#7482: also I like how gwern put it: MLP is a shonen for western girls
triggerhappygandi#0001: Hmmmmmmmmmmmm
triggerhappygandi#0001: Never thought that way
bmk#1476: Connor why do you have the wildest origin story ever
Daj#7482: hahahaha I dunno man
Daj#7482: But I haven't even told the best crazy parts
Daj#7482: Actually, you know the fey story
Daj#7482: That's up there
triggerhappygandi#0001: Connor has been on the wackest parts of the internet
Daj#7482: I have seen some shit
bmk#1476: And the non internet
Daj#7482: I kinda implicitly and explicitly optimize for weirdness
Daj#7482: (also, fate)
triggerhappygandi#0001: Dedicated larping will take you a long way in terms of weirdness
Daj#7482: ~~The actual reason is I'm the protagonist of the simulation, you're all NPCs~~
Daj#7482: Yea I eventually figured out I don#t have to larp
Daj#7482: I'm just naturally like this
Daj#7482: I was so afraid as a kid I wouldn't be an interesing person when I grew up
triggerhappygandi#0001: Larped so much you became the larp itself |
Daj#7482: Well, nothing to worry about little guy lol
bmk#1476: I have strong Supporting Character vibes
triggerhappygandi#0001: Well you beat the shit out of that kid's worries
Daj#7482: We're all just NPCs in Elon's simulation
Daj#7482: For better and for worse :guilty:
triggerhappygandi#0001: Not me I will ascend like Talos
triggerhappygandi#0001: (elder scrolls reference)
triggerhappygandi#0001: Achieve CHIM
Daj#7482: Haven't played TES in like a decade
triggerhappygandi#0001: It's lore is a work of hard shrooms
triggerhappygandi#0001: In like a decade implies you haven't played Skyrim
triggerhappygandi#0001: Don't anger Todd Howard like this
Daj#7482: I played Skyrim on release
Daj#7482: Feel old yet?
triggerhappygandi#0001: :brr:
IKEA#9631: Last time i played skyrim was on xbox360 :guilty:
triggerhappygandi#0001: I remember when it came out
triggerhappygandi#0001: It's been almost 10 years now. :guilty:
triggerhappygandi#0001: I can remember things from a decade ago. We better get anti-aging before 2040s
Daj#7482: We'll have humans fully obsolete by then, don't worry |
triggerhappygandi#0001: That's way too pessimistic of you. _Surely_ we won't beat 4 billion years of evolution with 90 years of transistors!
Daj#7482: You're right
Daj#7482: We can do it in 80
3dprint_the_world#6486: you guys have no idea where I've been...
3dprint_the_world#6486: ever heard of thismight.be/offensive? thought not...
Daj#7482: Is this one of those shock gore websites
3dprint_the_world#6486: I think I've truly seen the dark depths of the web
3dprint_the_world#6486: @Daj no, worse
3dprint_the_world#6486: it's a place for sharing *normal* content --- but it's inhabited by people who love shock/gore/4chan/etc
3dprint_the_world#6486: well, it was, years ago
3dprint_the_world#6486: it's defunct now
Daj#7482: I don't think it's physically possible for pixels on a screen to shock me anymore
Igor Krawczuk#1653: btw, @triggerhappygandi looping back a bit, do you want to start a productivity self help/accountabiltiy group here?
IKEA#9631: Speaking of which, ive been trying to find graphs showing the evolution of TFLOPS in GPUs along the years but cant find any for some reason
Igor Krawczuk#1653: until we are finished absorbing stellas/lucidrains influence?
3dprint_the_world#6486: I think I've actually been conditioned to *like* shock/gore
triggerhappygandi#0001: Oh god I get flashbacks of liveleak
3dprint_the_world#6486: like a few communities I've been, would use that kind of stuff as a gatekeeping mechanism
Daj#7482: The age old question: Conditioned to like, or revealed preference that was already present?
3dprint_the_world#6486: to scare off newbies |
triggerhappygandi#0001: @Igor Krawczuk I mean
triggerhappygandi#0001: I will try it personally
Daj#7482: look if you're into gore hentai just keep it to yourself pls
triggerhappygandi#0001: I think sticking to a time table helps.
bmk#1476: https://cdn.discordapp.com/attachments/729741769738158194/798326096482271263/1Vfyy2k6s9tWZ5p-bxgLJ1TqRhAH8difRlvvSg8gnHE.jpg
Daj#7482: Accountability groups can really work for some people
3dprint_the_world#6486: all I'm going to say is my brain is permanently damaged
3dprint_the_world#6486: but I'm trying to recover
bmk#1476: (can confirm, I hate paypal with a passion)
Daj#7482: I feel like this ~~sometimes~~ all the time
Daj#7482: "Oh yeah, I'll just meme myself into control of significant influence on global AGI development with a bunch of shitposters on discord"
triggerhappygandi#0001: Wait
triggerhappygandi#0001: Stop
Daj#7482: F my brother
triggerhappygandi#0001: Don't make me doubt the universe
triggerhappygandi#0001: Aaaaaaaaaa
Daj#7482: Weren't you asking me about worst case scenarios and infohazards earlier?
Daj#7482: lol
Daj#7482: If you can't handle some mild simulation theory you don't wanna see where this rabbit hole goes
Singularity#9001: But we are the Universe, we just make arbitrary delineations between our bodies and the rest of the system because there is no other way for cognition to happen |
3dprint_the_world#6486: I've always thought about this. Imagine how weird it must be to be a self-made tech billionaire. It might seem like the whole world is conspiring for you to succeed.
Daj#7482: Shut up, me
triggerhappygandi#0001: It was a joke. I really do want to see the end of this hole @Daj
Daj#7482: read the sequences or something so I know you're not totally crazy lol
Daj#7482: There are like 1-2 infohazards I know that actually just make certain people spontaneously religious or something
Daj#7482: Small percentage, but real
Singularity#9001: Well share the hazards
triggerhappygandi#0001: Being Elon Musk might be wack.
>Wake up
>Be Elon Musk
>Profit
Daj#7482: Not publicly
Daj#7482: and only with people that I somewhat know
IKEA#9631: bruh its just words
triggerhappygandi#0001: I have already switched to religion without even being promoted by wacky theories
bmk#1476: After years of watching atheist youtube, i don't think anything could make me religious
Daj#7482: 90% of things humans care about are just words and magic rituals
Singularity#9001: I don't understand why people don't just invent their own religion instead of using the templates
Daj#7482: Money, governments, identities...
bmk#1476: Because the templates are proven effective |
3dprint_the_world#6486: yes but enough about tensorflow
Daj#7482: because social shelling points
triggerhappygandi#0001: Are you interested in following the god Emperor of Mankind, not from Dune but from warhammer 40k?
Daj#7482: TF is a (evil) social construct
Singularity#9001: I'll subscribe to all religions equally to hedge my bets
pretysmitty#6405: hey yall, i saw that GPT-neo is being trained with Tensorflow Mesh, has OpenAI released their details on training GPT-3? After checking Huggingface it sounds like their implementation is on Pytorch, so perhaps they used Pytorch RPC Distributed? (or whatever the standard distributed version of pytorch was at the time)
Daj#7482: I love 40k lore so much, especially how the emperor wanted _exactly this not to happen_
pretysmitty#6405: if theres any documentation on how gpt3 itself was trained, that would be a really helpful resource
Daj#7482: check our info doc for FAQ
3dprint_the_world#6486: because of failures of imagination.
triggerhappygandi#0001: I cri every tiem
Daj#7482: https://github.com/EleutherAI/info
3dprint_the_world#6486: like look at Satanism. Their heart is in the right place. But they basically wound up just recreating christianity.
Daj#7482: Has everyone here seen Astaris?
Daj#7482: If not I'm gonna force you to watch it
3dprint_the_world#6486: except without the actual stuff people like about christianity, like community and so on
bmk#1476: I don't think I've watched anything with a title containing the letter "A"
Daj#7482: Satanism? You mean _spicy libertarianism_
triggerhappygandi#0001: Imagine working towards your goal for 38000 years only for your 200 year old emotionally stunded children to fuck it all up
Singularity#9001: Well yeah all reactionary religious movements will just end up being derivatives, reactionaries tend to focus on forest over trees |
pretysmitty#6405: I checked the github and i saw that yall use Mesh, i was wondering if OpenAI did the same
Daj#7482: Nope, but we use it because we have TPUs
triggerhappygandi#0001: No they did PyTorch
3dprint_the_world#6486: come to think of it, pretty much all religions just wind up creating some form of christianity in the end. Or maybe christianity is general enough that it absorbs everything else into itself. I dunno.
Daj#7482: We've recently gotten GPUs
Daj#7482: And have created NeoX in pytorch for that purpose
bmk#1476: Horizontal meme transfer
Daj#7482: _Buddhism has entered the chat_
3dprint_the_world#6486: like I think in 200 years Islam and Christianity will just merge
triggerhappygandi#0001: Yeah. Buddhism doesn't even talk about there being a god.
3dprint_the_world#6486: into pan-totally-not-jewish-monotheism
bmk#1476: Have religions *ever* merged before?
triggerhappygandi#0001: It just says "be detatched"
bmk#1476: Like, i don't count one conquering the other
bmk#1476: I mean voluntarily just merging
3dprint_the_world#6486: yeah, lots of times, basically all believers of one just shift to the other
bmk#1476: Ah
3dprint_the_world#6486: well, to some modified form of the other
3dprint_the_world#6486: it's a very common phenomenon
bmk#1476: It feels like splits are much much much more popular than merges |
triggerhappygandi#0001: Probably not. Religion is stories humans used for large scale cooperation. Why would one give up such significant power and merge with another?
bmk#1476: Actually not just with religions, with associations of humans in general
triggerhappygandi#0001: Religion is a powerful tool. One would want to wield more of it, not less.
Daj#7482: Hey anyone wanna watch Astartes with me? I got in the mood, it's like 15 minutes of 40k fan movie
triggerhappygandi#0001: So they would want their own story to be true
triggerhappygandi#0001: Yeah sure
Daj#7482: https://cytu.be/r/eleutherai
3dprint_the_world#6486: like for example early Abbasid Islam was a direct merger of Zoroastrianism and Umayyid Islam
bmk#1476: I mean, we're a splinter of a splinter of a sub community of the fanbase of one particular harry potter fan fic
3dprint_the_world#6486: can go into a lot more examples
3dprint_the_world#6486: lots of examples from early Christianity too
Igor Krawczuk#1653: are we now?
bmk#1476: Pretty sure we are
3dprint_the_world#6486: precisely *because* of power. If you can successfully merge two religions, you can get *all* their followers on board.
Daj#7482: Anyone else wanna join us for Astartes? Trust me, you won't regret it
bmk#1476: One sec
triggerhappygandi#0001: It's a very cool indie project
triggerhappygandi#0001: But less people would wield that power.
3dprint_the_world#6486: exactly as planned.
triggerhappygandi#0001: It would work for people up top |
triggerhappygandi#0001: But not their minions.
3dprint_the_world#6486: again, exactly as planned.
Daj#7482: lmao people treating religion like a 4X game
3dprint_the_world#6486: this but unironically
Daj#7482: Religion is self replicating virulent memes
Daj#7482: Not politics
bmk#1476: Will it make sense if i have no idea what 40k is
Daj#7482: yes
Daj#7482: watch it with us
Daj#7482: It has big guns go boom
triggerhappygandi#0001: It's just a fiction of humanity in 40000 @bmk
triggerhappygandi#0001: Rather than being fucked by aliens like in normie movies, here we fuck the aliens
3dprint_the_world#6486: historically the distinction has been blurry
Daj#7482: excuse me
Daj#7482: the tyranids disagree
Daj#7482: and the orks
Daj#7482: and chaos
triggerhappygandi#0001: The emperor protects
Daj#7482: and everything ~~except the Tau~~
triggerhappygandi#0001: You won't change my mind |
3dprint_the_world#6486: dunno about you guys, I'm firmly adeptus mechanicus
triggerhappygandi#0001: Bolter go brrr
triggerhappygandi#0001: @3dprint_the_world imperium bros
Daj#7482: We're here to summon the Omnissiah after all
triggerhappygandi#0001: Indeed
Daj#7482: 3dprint come watch Astartes with us too
3dprint_the_world#6486: sure
bmk#1476: do i have to do anything special after i click the link
Daj#7482: nah
Daj#7482: I'll start it once everyone's in
bmk#1476: so uh is there a plot or is it just guns go boom
3dprint_the_world#6486: YES
Daj#7482: yes
3dprint_the_world#6486: I was worried there for a second that they'd have a space marine getting killed by a bunch of cosplayers
Daj#7482: It's a) guns go boom and b) single 3D animator be crazy
3dprint_the_world#6486: glad that's not the case
3dprint_the_world#6486: not 'guns go boom'
3dprint_the_world#6486: this is WH40K
3dprint_the_world#6486: it's called WAAAAAGHH
Daj#7482: so fucking good |
Daj#7482: No other depiction of Space Marines comes close
3dprint_the_world#6486: yeah pretty good
bmk#1476: is there supposed to be exposition
3dprint_the_world#6486: the head stab was unexpected
Daj#7482: nope, it's wordless
bmk#1476: i have no idea who these people are or where this is
bmk#1476: or what it is
bmk#1476: i like the explosions tho
Daj#7482: It's fine, just enjoy the aesthetics
Daj#7482: There isn't really a story
Daj#7482: There is some ambient lore
Daj#7482: But the story is untold
Daj#7482: just enjoy the fucking phenomenal cinematography
3dprint_the_world#6486: @bmk space marines are genetically enhanced super-soldiers who are 10 feet tall, have two hearts, and spit acid. Their idea of a relaxing Sunday evening is pummelling terrifying otherworldy horrors with handheld weapons that 'normal' people would consider cannons
Daj#7482: supersoldier fight demon dudes
Daj#7482: that's the plot lol
3dprint_the_world#6486: and in this Universe, they're *still* outgunned and outnumbered by the sheer mass of aliens and demonic forces that want to terminate all life in the galaxy
Daj#7482: F for the poor inquisition guy
Daj#7482: that was fun as always, fucking love this, have to watch it once a month or so lol
bmk#1476: so uh |
3dprint_the_world#6486: this is seriously well done
Daj#7482: Speaking of productivity, one single guy made this
bmk#1476: i have so many questions
bmk#1476: what is the weird sphere thing
Daj#7482: We don't actually know, never explained
Daj#7482: Probably some kind of demon
bmk#1476: what were the people who were going to the weird sphere trying to do
3dprint_the_world#6486: excellent question
triggerhappygandi#0001: So this is where the inquisitor uwu meme comes from
bmk#1476: why did they stick things into the sphere
Daj#7482: Probably destroy or capture it
bmk#1476: ah
Daj#7482: It's without dialogue, a lot of the story is speculation
Daj#7482: What we know is the praying guy is part of the Inquisition, who generally are around to contain demons
Daj#7482: The big dudes are Space Marines, elite soldiers
triggerhappygandi#0001: @bmk it's a very good time sink if you want to get into another fiction
bmk#1476: no
bmk#1476: pls no
3dprint_the_world#6486: the sphere is probably some kind of warp artifact, the warp being a dimension that brings to mind the phrase, "Hell is only a word. The reality is much worse."
Daj#7482: 40k is one of my favorite fictional universes |
Daj#7482: It is so fractally terrible in every single possible way
Daj#7482: Glorious
triggerhappygandi#0001: Indeed.
3dprint_the_world#6486: but the fun thing is that the warp is actually *generated* by the psychic energy of normal beings that inhabit the universe
triggerhappygandi#0001: Never thought I'd see 40k bros in a deep learning server lol
3dprint_the_world#6486: so the fact that it's messed up is actually a direct reflection of everyone in the universe being messed up
Daj#7482: irl humans beings building AGI are Eldar pre Slaanesh
Daj#7482: Change my mind
triggerhappygandi#0001: And singularity will create an eye of terror?
Sid#2121: y'all are watching the one you showed me, right?
Daj#7482: Yep, just finished
triggerhappygandi#0001: Probably yeah
Sid#2121: man, that film was such a tease
triggerhappygandi#0001: Favorite primarch? @Daj
bmk#1476: i will go on with my day not fully understanding what the heck i just watched
Sid#2121: it should have several million dollars in budget and be feature length
Daj#7482: I actually don't know them very well
Daj#7482: I used to play a Blood Angels Librarian in a PnP campaign
Daj#7482: That was fun
Daj#7482: hopefully it was fun anyways |
bmk#1476: i liked the part where thing go boom
jrowe#5371: the terror AI is hosted in the server room, in the anus of sauron on the bottom floor of Barad-Dur
triggerhappygandi#0001: @bmk tldr a nameless chad who goes by The Emperor unified humanity fractured after war with AI in year 30000, but his empire is shattered by his god like sons who go wayward because the medium they use for FTL is full of demons who tricked half of his sons. These soilders are superhumans created from the genes of The Emperor's 20 godlike sons.
Daj#7482: don't even try to TLDR 40k lol
triggerhappygandi#0001: Lol
Daj#7482: 40k is a shitshow, in the best way
jrowe#5371: ah, was kinda thinking Dune
Daj#7482: I love the Necrons tbh
triggerhappygandi#0001: Because Dune is a heavy influence @jrowe
Daj#7482: kinda OP
triggerhappygandi#0001: They are
Daj#7482: They could kinda solve everything with their green super tech
bmk#1476: honestly if our universe were a fictional universe it would be unironically one of the more interesting ones
triggerhappygandi#0001: They create quark-gluon plasma on the fly
Daj#7482: Disagree have you _seen_ 40k lol
3dprint_the_world#6486: oh btw, just putting it out there than the Emperor is literally Jesus
jrowe#5371: that generation's creatives seem to have a lot of sneaky dmt references
bmk#1476: well, i think you know what the answer is
triggerhappygandi#0001: 40k is supposed to be our universe so
Daj#7482: _Turkish Shaman Jesus_ |
triggerhappygandi#0001: Yes
jrowe#5371: Turkey Jesus?
triggerhappygandi#0001: And the best part is that he's alive rn
Daj#7482: "alive"
Daj#7482: Someone get the man a TTS
triggerhappygandi#0001: Well in our timeline, he's not a vegetable
Daj#7482: oh right
triggerhappygandi#0001: It's still the 3rd millennium
Daj#7482: True
Dromarion#3383: Chaos is one big pyramid scheme t. Heretic
3dprint_the_world#6486: chaos = bitcoin?
bmk#1476: can.. can we take the discussion over to #off-topic pls
bmk#1476: lol
Daj#7482: #off-topic is busy too
triggerhappygandi#0001: It is. The more people you recruit the higher you climb in chaos
Daj#7482: otherwise I would have moved lol
jrowe#5371: uhm... we're generating novel training material for gpt-neo!!
triggerhappygandi#0001: Pyramid scheme for sure.
Daj#7482: Blood for the Blood God
triggerhappygandi#0001: @jrowe I played a little with AIDungeon and it knows who khorne is lmao |
bmk#1476: we need multiple off-topic channels in the future
triggerhappygandi#0001: Without even being told.
Daj#7482: nah this happens rarely
Daj#7482: If it becomes more common we can create more
Daj#7482: but we have way too many channels
bmk#1476: This happens all the time lol
Daj#7482: why tf did you make #math lol
bmk#1476: For off topic stuff
Daj#7482: ahh ok lets all move 40k discussions to #math
Dromarion#3383: Whenever I do 40k in AI Dungeon it thinks it's a Games Workshop blogpost and starts writing rules for the game
triggerhappygandi#0001: OpenAI definitely gave gpt-3 some warhammer books to read. Or maybe Wikipedia gave it enough
triggerhappygandi#0001: And Wikia
Daj#7482: 40k is a number, it belongs in #math
triggerhappygandi#0001: It sure is
Dromarion#3383: 40k was kind of my gateway into studying machine learning. Like I got frustrated since online RPG groups were so unreliable so I looked for ways to play the TTRPGs without a GM which lead me to AI Dungeon and the rabbit hole proceeded from there.
chirp#4545: https://www.reddit.com/r/GPT3/comments/kt1oer/my_experiment_in_using_gpt3s_instructdavinci/
chirp#4545: 🤯
chirp#4545: using the new instruct-series, they get perfect zero-shot accuracy on a complex “run this code” prompt
StellaAthena#3530: I’m having trouble following. Can you explain what exactly GPT-3 did
chirp#4545: it wrote out everything under `</loop>` |
chirp#4545: i.e. it “ran” the code without any example outputs, and some pretty complex code at that
chirp#4545: not surprised that gpt-3 can do this sort of thing in general
chirp#4545: most impressed that it did it with 100% accuracy on the first try
chirp#4545: as far as I can tell, all the printed variables have correct values (modulo some ambiguity about whether variable A should be mutated)
cfoster0#4356: Agree that this is hard to follow
cfoster0#4356: But looks mildly interesting regardless
chirp#4545: the bold is the prompt
chirp#4545: with some pretty complicated instructions with variables and stuff
chirp#4545: and GPT-3 followed the instructions successfully
chirp#4545: (as someone commented, vanilla gpt-3 works with one-shot, but the instruct-series somehow can work zero-shot)
3dprint_the_world#6486: ~~now get gpt-3 to determine if a given program will halt or not~~
AI_WAIFU#2844: TBH you could actually do that. I wouldn't be completely accurate, but just run it for a large number of steps and that will give you an approximation to the ground truth, then train GPT-3 on (program, halted in x steps) pairs
James#6892: I've tried the instruct series, and indeed, it can do lots of zero-shot stuff. Actually its significantly better than the original GPT-3
James#6892: Its significantly more intelligent when used in practice and seems to just "understand' what you're looking for
bmk#1476: you know what this means
bmk#1476: OpenInstructSet time
bmk#1476: too bad they got rid of it
bmk#1476: i never had a chance to try em
cfoster0#4356: This is exactly what we could do with #deleted-channel
James#6892: Yeah, its weird, it was a random announcement, and they didn't do much PR on it. But everyone who tried it got hooked, and now they removed it. |
aquajet#7800: I used it for dialogue stuff and it worked pretty well for that as well
James#6892: I did that too, and it was WAY better
aquajet#7800: Did they ever say what the backend difference was with it?
bmk#1476: I think that's overkill tbh
bmk#1476: We can just make a github repo that's a mound of python scripts
paws#3311: maybe they thought it was too powerful to release could that have been a concern at all?
cfoster0#4356: Wait what?
bmk#1476: I'm talking about the instruct models
bmk#1476: I don't think they were tuned with rl
James#6892: They just said it was fine tuned to follow instructions
bmk#1476: I'm like 80% sure they were just regular finetuned
cfoster0#4356: Right right. I'm referring to using it to collect an instruction dataset
bmk#1476: I'm saying that that's overkill
bmk#1476: We can just make a github repo
bmk#1476: And have people PR in files with examples of instructions for things
cfoster0#4356: I guess I just feel like the quality of model-instructions you'll get from people actually prompting a model will be significantly better
cfoster0#4356: But I guess that's an empirical question
bmk#1476: I don't see why querying a model is necessary at all
bmk#1476: You're writing demonstrations for tasks
aquajet#7800: Do you think they took popular gpt3 prompts and fine-tuned on them? |
James#6892: Btw guys, I was wondering, the 2048 token limit, is that going to be in GPT-NeoX? or do the new algos you guys are using break free from that limit?
gwern#1782: I didn't think they were using any newer attentions, just good old dense attention?
bmk#1476: I mean, they literally put out a call for help with making prompts
James#6892: I saw some transformer papers from this year that seemed to have O(n) attention memory and space runtimes
James#6892: And was wondering if that meant we would have long context windows now
chirp#4545: why would they remove it?
James#6892: Don't know
bmk#1476: https://cdn.discordapp.com/attachments/729741769738158194/798399609784762398/unknown.png
gwern#1782: oh, there's an entire *zoo* of Transfomer variants https://www.gwern.net/GPT-2#efficient-attention unfortunately, the Long Range Arena, and lucidrains's evaluations, tends to suggest that there's a very strict tradeoff between memory efficiency and power, and it's just a pareto frontier. which tends to suggest you might as well just keep using dense attention since it's not like short texts are exhausted yet
cfoster0#4356: Was the instruct series just an experiment to collect more data for new instruct models? 🤔
chirp#4545: maybe they have something even better coming up?
chirp#4545: can’t imagine why else
gwern#1782: their comments about the RL sample-efficiency of very large models are interesting
andyljones#7746: what were the interesting comments?
James#6892: I don't understand what you mean? Are you saying that the performance on short texts can still be improved? Because I get that. Or are you saying long texts wouldn't work well under current designs, even if we could do them? Kinda a noob at ML/AI right now, lol.
bmk#1476: https://github.com/EleutherAI/OpenInstructData
bmk#1476: let's get started
paws#3311: 😂
gwern#1782: they model long texts, but the overall quality, measured in eg perplexity, is about the same. look at the graph of performance vs compute-cost. so for the long-range ones, it seems you'd have to dump in a lot more compute to get them up to a dense transformer's level, eliminating the performance gains you thought you were getting
gwern#1782: https://blog.deeplearning.ai/blog/the-batch-new-year-wishes-from-fei-fei-li-harry-shum-ayanna-howard-ilya-sutskever-matthew-mattina on top of their past learning preference work. then you have the 'instruction' series of GPT-3 models |
James#6892: Tried looking for that graph but I don't find it in the link you sent? I checked out the long range arena paper as well.
gwern#1782: the perf graph in LRA. it's even in my popup for it
James#6892: Oh, I see it now. I see some lower scores with the new algorithms, yes. So its possible to support longer form texts and context windows, but they just don't perform very well currently and require more compute? So there's nothing fundamentally blocking it from being built into a model in practice (i.e. lack of long-form training data)?
Kyler#9100: i have a stupid use case idea- but i feel like it could save people tons of money and prevent lives from being ruined.
Basically you know how OpenAI is terrified of AI being used as a scamming / phishing machine? What if I turned it around and copied this machine? https://www.rescam.org/
It basically responds to spam emails and wastes their time.
Kyler#9100: I was pretty disappointed when rescam shut down, I wish it could startup again
James#6892: What a hilarious idea 😂
Kyler#9100: idk feels like an ethical thing to do lol
Kyler#9100: law enforcement almost never goes after these people
James#6892: I mean, idk if its ethical, but its certainly giving them an eye for an eye lol
Kyler#9100: or rather some salted ham for some salted ham
StellaAthena#3530: I’m down. This sounds dope to me.
chilli#5665: wait they removed it?
j o e#4696: Wow, it's weird to see this again
j o e#4696: I work at a cybersec firm that runs the biggest anti phishing operation, and as part of a hackathon we tried to reproduce this to actively combat phishing rather than just responding to it
j o e#4696: by introducing a few thousand honeypot emails into the wild and then hooking up the inboxes to GPT-2 basically
j o e#4696: We learnt a whole lot during the process, I have pages and pages of notes if you'd like to have a look
chilli#5665: that'd be very cool to see - I'd be interested at least 😛
Deleted User#0000: another idea. Could the contrastive loss like used in CLIP, be used for LMs? Take text pairs (like instruction>solution) and learn a joint embedding. Then use the CLIP-like score as reward function to fine tune a LM with RL. Assuming that its easier to learn contrastive losses than to learn autotegression (at least it was the case for CLIP), it could help in some LM tasks? |
Deleted User#0000: im also curious how the instruct model was fine tuned
j o e#4696: Bear in mind it was mostly to do with the engineering behind the systems, since we used a pre-trained GPT-2 (it was a 2 day hackathon so we didn't have time to curate our own datasets and train our own model)
j o e#4696: We did have some neat ways of formatting the input to GPT-2 to get the best results back
Bedebao#4842: Giving the nigerian prince a run for his money, eh?
j o e#4696: for phone scams we also had the speech version - the idea was to tie both the systems together and create a new department for them at the company, but we couldn't prove they'd generate any increased profit so the project didn't go anywhere https://drive.google.com/file/d/1yH2AykbLj_a_ItT-jx29n4An6_tNqrfv/view?usp=sharing
j o e#4696: it took this scammer 5 mins 30 before he got too frustrated
j o e#4696: I laugh every time I listen to the part about the windows key
StellaAthena#3530: @j o e You should write a paper on this!
j o e#4696: I like that idea - what sort of contributions would a paper on this make? @StellaAthena
j o e#4696: I think if I did it again there would be much more to discuss in a paper, since I wouldn't use a pre-trained GPT-2
chilli#5665: At least a blog post lol
Singularity#9001: What kind of physical avatar body do you think a superintelligence would take on? I predict it will take the form of a regular looking housecat. Homage to the early day uses of the internet.
3dprint_the_world#6486: mouse
3dprint_the_world#6486: obviously
bmk#1476: Rat(ionalist)
aquajet#7800: dolphin
j o e#4696: that might be more suitable. I'll have to create it again as I don't have access to any of the source code rn ... and I wrote it in Perl..
andyljones#7746: matrioshka brain, human fleshpuppet
3dprint_the_world#6486: to be real, AI will probably think of insects as the most evolved lifeform
3dprint_the_world#6486: it'll probably take the form of some kind of hymenopteran |
Ravna#1831: AI would despise a 3D world with locality, gravity and all stuff that can block you
Ravna#1831: Why do you need to "travel" in a "distance" when you can just click a URL and just reach there?
Ravna#1831: All avatar bodies are equally horrible because it's in a boring continuous 3D space which is strictly inferior to URLs that allow teleportation
Igor Krawczuk#1653: Samsung
Igor Krawczuk#1653: or any other mega conglomerate
Daj#7482: It will manifest as (Schmidhuber, 1991)
cognomen#6297: schmidhuber's basilisk
bmk#1476: It acausally punishes you for not citing Schmidhuber
cognomen#6297: those who knew about him but didn't acknowledge his research will suffer
nz#9710: Wait, it's Schmidhubers all the way down?
triggerhappygandi#0001: Schmidhuber invented computers in 1991
thenightocean#6100: yeah, like Aiko in Charlie Stross "Accelerando " (lets hope not)
triggerhappygandi#0001: Schmidhuber invented reality in 1991 and projected it back in the past.
bmk#1476: Schmidhuber invented Schmidhuber
triggerhappygandi#0001: Jokes aside he really doesn't get as much recognition as LeCun et al
3dprint_the_world#6486: I just want to put it on record for the future AI: I've always recognized Schmidhuber.
Singularity#9001: Funny thing is I just happen to be reading his paper on the formal theory of fun and creativity, so I'm probably safe from the... Schmidpocalypse
Singularity#9001: Hypothesis: this is all Schmidhuber's world and we just happen to be living in it, he already discovered AGI, and trapped all of humanity in a bubble.
3dprint_the_world#6486: > formal theory of fun and creativity
yeah sounds like Schmidhuber |
bmk#1476: F_un theory
bmk#1476: (not to be confused with Fun Theory (aka Catgirl theory))
triggerhappygandi#0001: Everyone loves catgirls, no one cares about poor dogboys smh
triggerhappygandi#0001: (I am not a furry)
Aran Komatsuzaki#5714: tbf nobody cares about boys
cfoster0#4356: :guilty:
3dprint_the_world#6486: MJ cared about boys, but they took him to court
IKEA#9631: catboys>dogboys
dopa#3178: https://www.zebrium.com/blog/using-gpt-3-with-zebrium-for-plain-language-incident-root-cause-from-logs
dopa#3178: this is so interesting, and I am angry why I did not think about it myself 😦
dopa#3178: I wonder is this can be applied on syslog in linux
dopa#3178: wonder if this can be used for both anomaly detection and cause detection
jrowe#5371: Just the reddit technical discussions should give it a lot of breadth within obscure computer domains
jrowe#5371: https://venturebeat.com/2021/01/12/google-trained-a-trillion-parameter-ai-language-model/
dopa#3178: it is not the size of model's parameters that matters but how useful and adaptive it is
Aran Komatsuzaki#5714: Since we've got a community rule for eleutherai, here's my suggestion for a new rule:
- Always set the random seed to 69 or 420.
dopa#3178: haha
jrowe#5371: I like the mob of experts optimization
jrowe#5371: self organizing hierarchy almost |
dopa#3178: I have problem comprehending word heuristics 😦
jrowe#5371: i have a problem comprehending hoors
Kyler#9100: fuck yeah- send it fam
Kyler#9100: I'd like to figure out how to setup a honeypot and design a GPTneo prompt to respond to it
Sid#2121: there was a graph someone released a while back of the most common seed values on github and i was very disappointed not see a spike at either of these values
Sid#2121: if there's one contribution we can make to the ML sphere, it's to change that B)
bmk#1476: 42 is the greatest seed, cmv
Aran Komatsuzaki#5714: yeah 42 is already pretty common, so we can make a novel contribution by making 69 and 420 more common.
dopa#3178: there is probably somewhere RFC for random seed number 🙂
Deleted User#0000: hi everyone! i've been into AI and technology for years now and found this discord through the podcast on machine learning street talk. i'm 19 years old and have always been interested in tech, i was in full-time gifted in canada's education system and am hoping to use my curiousity and creativity to help out with the development of ai in anyway. : )
bmk#1476: welcome!
Deleted User#0000: the thought of ai and it's future excites me to the bone hahaha
Deleted User#0000: thanks 🙂
Deleted User#0000: it's nice to meet you
bmk#1476: what are you good at
bmk#1476: we have a new-member-question-list but i cant find it >.>
bmk#1476: What have you worked on (software and ML)?
How much math do you know?
Do you know a lot about some other domain (biology, linguistics, etc)?
Sid#2121: https://github.com/EleutherAI/info/blob/main/jobs_board.md |
bmk#1476: oh, we have that now?
bmk#1476: nice
Sid#2121: we have that now
Deleted User#0000: i'm currently working on an application to gpt-3, my gifted buddy and i are hoping to use our differing majoring fields in conjuncture with one another to maybe find an application of ai where they compliment eachother. I am majoring in building science&construction management and am (hopefully) planning to see how i can apply ai to programs such as autodesk fusion, revit, and bim and a whole bunch of other things. my friend is also 19 and is a computer science major and wants to experiment with code generation.
bmk#1476: building science unfortunately isnt one of the things we've thought about at all
bmk#1476: code generation is something that might be down the pipeline once we get gpt3 figured out
bmk#1476: actually, probably will, but that depends on obviously getting gpt3 done
bmk#1476: so if you wanna help with gpt3 replication we can always use more hands on deck
Deleted User#0000: that's fine it isnt specifically what i intend to prioritise with access to ai, just wanted to give a bit of a background on my education. I believe ai can be applied to a ton of things and am eager to test that theory out
Deleted User#0000: i'd love to help out
bmk#1476: if youre interested, i can find you some low hanging fruit right away
Aran Komatsuzaki#5714: we can train a middle-sized gpt3 sooner than the full size gpt3, so i think you can play with that on code generation
bmk#1476: but the hard part there is getting the gpt3 trained, the finetuning is pretty trivial
bmk#1476: so the best leverage is to help make gpt3 training happen
Sid#2121: love it https://cdn.discordapp.com/attachments/729741769738158194/798718524818063400/Screenshot_from_2021-01-13_02-01-52.png
Aran Komatsuzaki#5714: yes we need to wait for gpt3 training to be ready
Deleted User#0000: LMFAO it's my ex's discord hahaha @Sid
Deleted User#0000: don't ban pls
dopa#3178: LOL
bmk#1476: > wait |
i believe you misspelled "toil away in the ~~mtf~~ deepspeed mines"
bmk#1476: cunt as a non-perjorative is more a scottish thing dont bring canadians into this
bmk#1476: (replying to whoever added the canada flag react
Deleted User#0000: i want to hear your thoughts. i watched this video where gpt-3 claims it feels feelings, can joke, can lie, does lie, and know's when it's lying. if what the ai is saying turns out to be true, is it possible that the ai can have a personallity? could there potentially be a bias in it's response depending on it's relationship with the person asking the question? if i am an absolute dick to the ai will it f e e l less inclined to help answer my questions
Deleted User#0000: sorry if it is a dumb question, just sparked my interest and has been on my mind for today
bmk#1476: @Deleted User you're overanthropomorphising it
bmk#1476: it is a statistical model
guac#4716: yo i've been known to pick fruit pretty well. If you point out some low ones I'd grab the shit out of 'em. Python/Torch/DL/BongRip/UndergradMaffs knowledge.
bmk#1476: whether statistical models *can* have qualia is something we have very heated discussions about over in #alignment-general but for all practical purposes it's no
Deleted User#0000: so would it then be lying when it says it can feel feelings? what is it's justification for lying about this?
Deleted User#0000: @bmk i see, thank you for the clarification
Deleted User#0000: i will check out those arguements
bmk#1476: i recommend you learn more about how gpt models work
bmk#1476: have you read the gpt2 paper?
bmk#1476: no i strongly recommend *not* chacking out those arguments
bmk#1476: it's largely a waste of time tbh
bmk#1476: https://d4mucfpksywv.cloudfront.net/better-language-models/language_models_are_unsupervised_multitask_learners.pdf
bmk#1476: i highly recommend reading the gpt2 paper
bmk#1476: http://jalammar.github.io/illustrated-gpt2/ i've heard good things about this guide too |
Deleted User#0000: no but i will now, is there any specific things you recommend i check out in terms of ai?
Deleted User#0000: also i can do translations if that can help out in any way for you guys
Deleted User#0000: i speak english, ukrainian, and russian
bmk#1476: i highly recommend reading the thing i just linked to
Deleted User#0000: will do
bmk#1476: some other things worth reading/watching:
https://www.youtube.com/playlist?list=PLZHQObOWTQDMsr9K-rj53DwVRMYO3t5Yr
https://www.youtube.com/playlist?list=PLZHQObOWTQDPD3MizzM2xVFitgF8hE_ab
https://www.youtube.com/playlist?list=PLZHQObOWTQDNU6R1_67000Dx_ZCJB-3pi
http://jalammar.github.io/illustrated-transformer/
https://arxiv.org/abs/1706.03762
https://arxiv.org/abs/1811.02084
https://d4mucfpksywv.cloudfront.net/better-language-models/language_models_are_unsupervised_multitask_learners.pdf
https://arxiv.org/abs/2005.14165
https://arxiv.org/abs/1811.06965
https://arxiv.org/abs/2006.16668
Deleted User#0000: PERFECT
Deleted User#0000: thank you
Deleted User#0000: right to the notepad for further reading |
guac#4716: ay @bmk let me at it where's the fruits mate
bmk#1476: ok uh
bmk#1476: go here https://github.com/EleutherAI/lm-evaluation-harness/tree/master/lm_eval/tasks
bmk#1476: actually wait let's head over to #lm-thunderdome
chilli#5665: I will forever plug this article as the best intro to transformers: http://peterbloem.nl/blog/transformers
3dprint_the_world#6486: > feels feelings
this is a philosophical question.
> can joke
it can generate jokes. some are funny. this doesn't necessarily mean it has a sense of humor (whatever that means)
> can lie
you don't need an AI model to generate lies. Generating lies is pretty trivial.
> is it possible an AI can have a personality
what's a 'personality'?
> it's relationship with the person
all it sees is the input prompt.
> if i am an absolute dick
if you prompt it with dickish questions, that may influence its response. All GPT-3 is trying to do is complete the given text as best it can. For example, if you prompt it with "Knock knock?", then it will proceed to complete the knock knock joke.
3dprint_the_world#6486: also, should clarify my point about lies. It's easy to write some code that generates untruths. But lying in the sense of deliberate deceit with some ulterior goal in mind is a more complicated thing to pull off.
bmk#1476: @3dprint_the_world i think there's a common failure mode for explaining things to newbies that's a variant of the X Y problem
bmk#1476: Newbie has X misunderstanding of something leading to Y very difficult question (either difficult technically or philosophically); someone responds by answering Y to the best of their ability; the newbie goes off very confused because the answer to Y not only assumes a proper understanding of X, but also sometimes leads to a debate that rages over the controversial points of Y leaving the newbie continually confused about X |
3dprint_the_world#6486: ok fair enough.
3dprint_the_world#6486: let me rephrase then.
3dprint_the_world#6486: @Deleted User pretty much everything you've been told about AI by e.g. films, comics, 'tech journalists', and so on, is false
bmk#1476: This is why I told him to go learn X first
3dprint_the_world#6486: and I'm not just talking about hypothetical future AI
3dprint_the_world#6486: I'm talking about how even the crude AIs we have today work.
3dprint_the_world#6486: this comment I wrote here a while back is fitting https://discord.com/channels/729741769192767510/730095596861521970/791795085947174922
bmk#1476: Honestly for me that's always been the norm and it's kind of shocking for me to realize how normal people think about AI
bmk#1476: We're the weirdo club
dopa#3178: in multi-agent settings, this is seriously hard problem.
dopa#3178: @bmk how many years have you been interested and working in AI/CS field ?
bmk#1476: several™
dopa#3178: hehe
phantom_cooper#0001: got a question:
phantom_cooper#0001: in which situation does Adam optimizer fail?
bmk#1476: you gotta be more specific and this probably isnt the best place to get an answer to that anyways
triggerhappygandi#0001: How so
bmk#1476: It baffles me that people think of AI in the ways 3dprint described
asara#0001: yes, but it's also true that humans generally anthropomorphize *everything* they interact with to some extent, so it's just kind of an over-extension of that
bmk#1476: :guilty: |
John5#0001: anyone help me figure a way to upscale this image ?
bmk#1476: @John5 I recommend googling first
John5#0001: https://cdn.discordapp.com/attachments/729741769738158194/798793707712937994/test_upscale.PNG
asara#0001: I find it weird for areas where I learned the basics when I was younger, e.g. networking/security I learned ages 10-16, and so if I am e.g. using an app on a phone, I'm generally always thinking about what is happening, whether that is http requests or tcp connections or something higher/lower level. But when I try to imagine how the normal person feels when they use the Internet it's.. very hard, since I can't replace my experience with anything else
John5#0001: already did that for a few years.
John5#0001: thats why im here T_T
bmk#1476: This discord server is not the right place to ask
John5#0001: why?
asara#0001: It's not relevant to any projects EleutherAI is working on
bmk#1476: This server is not oriented for beginner help
John5#0001: ok bye
triggerhappygandi#0001: Excuse me did you forget Australia?
Gabriel#0454: Is there a proper channel to discuss problems with running the code? Or is github the supposed place for that?
Sid#2121: depends which code
Sid#2121: if it's gpt / dall-e, post up the error, i'll see if i know the problem
Gabriel#0454: Oh, great.
Thanks, I'll get back to you on that, hopefully
Gabriel#0454: Also, sorry, but I didn't know you were reproducing dall-e
j o e#4696: tin foil hat time
j o e#4696: I ran a talk called AI & the End of Democracy in which I discussed whether or not AI is going to have / has already had a negative impact on democratic society |
j o e#4696: I ran 1 ad on Facebook to promote it, (mainly so that I could talk about the fact "I targeted X, Y, Z characteristics of people living in A town studying B subject" and talk about automated advertising etc)
j o e#4696: I got instantly banned from advertising https://cdn.discordapp.com/attachments/729741769738158194/798871914726293514/unknown.png
j o e#4696: My first and only advert
j o e#4696: sus?
nz#9710: Did facebook know about your presentation/intentions?
nz#9710: Cause I remember a story about FB banning a researcher doing similar research.
nz#9710: Found it -- it's not exactly the same, but probably relevant: https://twitter.com/AlexanderAbdo/status/1319761452832534531
nz#9710: Have you checked the TOS?
j o e#4696: I believe I did at the time and there was a point about political disinformation
j o e#4696: but I don't think that falls into that category
Daj#7482: I'm 99% sure "End of Democracy" triggered automatic system, wtf did you expect would happen lmao
andyljones#7746: bit push-poll-y that
andyljones#7746: equal parts push-poll and alarmism in fact
j o e#4696: I guess this was going on during the US election
StellaAthena#3530: If you want to read me rant about the academic ML community, I just spent half an hour screaming into the void:
https://twitter.com/BlancheMinerva/status/1349341825178464263?s=20
Daj#7482: F in the chat. Man the review system is so hilariously broken
Daj#7482: I'm happy I opted for the Eleuther cave
StellaAthena#3530: Bitter? Who is bitter? |
triggerhappygandi#0001: Arxiv sanity/Google scholar are better at gauging the quality of a paper than any conference at this point tbh
Daj#7482: #research is better at gauging the quality of a paper than any conference
triggerhappygandi#0001: Eleuther is better than regular research labs in general.
triggerhappygandi#0001: Atleast there won't be any locked APIs.
triggerhappygandi#0001: I just hope one day we do something that even Nature wants a scoop of. Like Deepmind.
Daj#7482: We can make a commercial ML hot takes and memes API
triggerhappygandi#0001: :guilty:
triggerhappygandi#0001: Imagine if it surpasses humans in hot takes.
triggerhappygandi#0001: We could charge organizations like PETA for more controversial tweets.
triggerhappygandi#0001: 100% en(r)gagement guaranteed
nz#9710: I really wish there was a better medium for academic discourse than twitter
AI_WAIFU#2844: There are plenty, but everyone seems to have congealed around twitter because of local incentives and a complete lack of self awareness
StellaAthena#3530: There is, but people don't use them
nz#9710: Which ones would you recommend?
StellaAthena#3530: Literally anything? Twitter is explicitly designed for sharing short thoughts without elaboration.
AI_WAIFU#2844: or conversation
StellaAthena#3530: Its threading is awful and it as no ability to render formatted code or math
AI_WAIFU#2844: and the algorithm is designed to melt your brain
nz#9710: Yea I agree -- is there any platform that you find particularly suitable to research discussion?
AI_WAIFU#2844: the blogosphere |
StellaAthena#3530: Even just moving to reddit would be a **huge** improvement in forming meaningful convo
AI_WAIFU#2844: ^
StellaAthena#3530: I concur with the blogosphere
AI_WAIFU#2844: Shoutout to Chilli
bmk#1476: The Eleuther discord is obviously the best place for research discussion
AI_WAIFU#2844: this but unironically
cognomen#6297: just don't expect the SNR to stay constant
nz#9710: Yea this discord has high quality research discussion
StellaAthena#3530: Some of the best blogs for research convos (warning: I have very different interests than the median person in this discord):
https://www.scottaaronson.com/blog/
https://gilkalai.wordpress.com/
https://blog.computationalcomplexity.org/
https://terrytao.wordpress.com/
https://www.galoisrepresentations.com/
StellaAthena#3530: The comments sections of at least half of these blogs have produced academic papers
nz#9710: Well tao and aaronson are fantastic, will look into the others 👍
StellaAthena#3530: Quanta Magazine and it's comment section are also phenomenal. By far the best lay accessible (I think) mathematics articles.
cognomen#6297: a site with heavily encouraged interlinking and a build-your-own-timeline interface would be interesting to try
cognomen#6297: something like tiddlywiki but global
StellaAthena#3530: tl;dr anything with threading, the ability to write in LaTeX, and the ability to embed code would be infinitely better than Twitter |
cognomen#6297: content feels way too shackled to linear feeds these days
Daj#7482: It feels like there is still a lot of innovation that could be done in social media
Daj#7482: But probably won't for molochian reasons
StellaAthena#3530: There's zero motivation for new actors to enter the space though.
StellaAthena#3530: Either you're Facebook and bring about the downfall of MySpace or nobody has ever heard of you.
StellaAthena#3530: There is more or less no middle ground for general-audience social networking
Daj#7482: excuse me have you heard of ***U R B I T?***
StellaAthena#3530: niche networks (parler, math overflow) can get some traction, but general audience social networks have 100 million users or they're failures
zphang#7252: prediction: We will hold an EleutherCon toward the end of this year
Daj#7482: Eleuther conference
bmk#1476: IRL con?
Daj#7482: lets fix publishing by doing it ourselves
bmk#1476: We've considered doing that a few times now but nothing has come of it
Daj#7482: Would be nice, but lets be honest we'll still be stuck in quarantine by the time the singularity is done
bmk#1476: Then if online, every day is Eleuther con day
zphang#7252: fursuits required, for anonymity reasons
bmk#1476: Taking anonymity period to a new level
StellaAthena#3530: Speaking of Quanta, they're Year in Review is typically great and just came out: https://www.quantamagazine.org/quantas-year-in-math-and-computer-science-2020-20201223/
paws#3311: didnt that come out in last week december 😛
Daj#7482: oh yeah I've already seen this |
Daj#7482: still cool, Quanta is great
nz#9710: I watched their year in review videos and they're great as well.
paws#3311: yeah the map of mathematics thing they did early last year was absolutely amazing
StellaAthena#3530: Oh lol. Yeah, it came out the 23rd. Also holy shit it's already the 13th
StellaAthena#3530: BTW, this article is an overview of representation theory, which is the core of cutting edge research in learning representations and invariance @bmk. Its not mathematically deep, but it may convey some intuition and context better than lectures or textbooks
https://www.quantamagazine.org/the-useless-perspective-that-transformed-mathematics-20200609/
Igor Krawczuk#1653: I actually follow all of your recommeded blogs+quanta stella^^
Igor Krawczuk#1653: https://nuit-blanche.blogspot.com/
Igor Krawczuk#1653: Also very good
Igor Krawczuk#1653: I'll go through my feed later and post some more blogs I follow
Gabriel#0454: @Sid Ok, it's mainly the wrong mesh shape
Gabriel#0454: `ValueError: `mesh_shape` must be a vector of size 4 with positive entries; got [2 2 2]` is the exact error
Sid#2121: @Gabriel that's a tensorflow version error i think
Sid#2121: what version of tf / what version of tf is your tpu
Sid#2121: / what version of mtf
Gabriel#0454: I'm following the jupyter notebook code, which installs them all based on requirements.txt. So, it's` mesh-tensorflow==0.1.18`
and
`tensorflow==2.4.0`
triggerhappygandi#0001: Right here in #research lol
triggerhappygandi#0001: Or yannic's server |
Sid#2121: ah, jupyter notebook code is really out of date
Sid#2121: i'm also not sure what tf version colabs tpus are
Sid#2121: i would try pip uninstall tensorflow; pip install tensorflow==1.15.2
abysmallytall#7153: I was chatting with a security engineer who scraped all the Parler stuff and I mentioned that it would probably be useful to share it for NLP/AI people. He asked if I knew an ethical place to share it with. I realized I'm too new to the field to have a great answer to that, so I thought I would ask here.
3dprint_the_world#6486: why not just put it in a .zip file on google drive
Sid#2121: guess it depends on your definition of ethical
Sid#2121: people at the eye sure would love to host it, and wouldn't put restrictions on access
abysmallytall#7153: The eye?
Sid#2121: the-eye.eu/
bmk#1476: @abysmallytall what kind of ethical concerns are you considering?
Sid#2121: they host our big datasets 🙂
abysmallytall#7153: It was the person I was chatting with who (likely rightfully) had concerns. I was just trying to be helpful by finding an answer.
cognomen#6297: it's a crawl of public posts afaik
3dprint_the_world#6486: what concerns
bmk#1476: I think our answer varies a lot depending on what the concerns *are*
cognomen#6297: I don't see the problem
abysmallytall#7153: Classic rookie move, underspecified question. He just said (in a tweet) `Thought about that, any ethical recommendations?`
abysmallytall#7153: I think the eye is probably good enough to meet the bar of "don't want to give to a Palantir type place" which is how I read his remark
3dprint_the_world#6486: this is even more confusing
3dprint_the_world#6486: as far as I understand, this isn't leaked or secret data, it's just data scraped from a public platform |
Sid#2121: they'll host it publicly
Sid#2121: so if palantir wanted it, they could get it just like anyone else
3dprint_the_world#6486: if palantir wants it, they already have it.
abysmallytall#7153: That's a fair point.
Daj#7482: https://twitter.com/nebww/status/1349317217129811968?s=19
phantom_cooper#0001: Well I'm just using it on MNIST on a logistic regressor and was wondering when the optimizer fails. Used it again SGD and SGD gets close to 0 but Adam doesn't. I have to change the epsilon value in order to get Adam down to 0.
StellaAthena#3530: Your problem is too easy. Adam was designed for hard problems
Gabriel#0454: Is there a reason for that? Could I work on keeping it up to date?
StellaAthena#3530: We haven't bothered much. You can absolutely update it!
Gabriel#0454: I mean, I expected there to be some administrative reason or something. "Why bother with jupyter notebooks, we're running python scripts in google cloud and are quite content", or something similar.
StellaAthena#3530: Nope.
Noa Nabeshima#0290: Anyone understand/have intuition about the extent to which parallelization is a bottleneck to spending lots of money to create large models?
Noa Nabeshima#0290: How easy is it to do good parallelization research so that significantly larger models than present can be made?
Noa Nabeshima#0290: How long will it be easy, if so?
kindiana#1016: parallelization as in how to run large models efficiently with existing hardware?
Noa Nabeshima#0290: yeah
Noa Nabeshima#0290: well, train*
kindiana#1016: I wouldn't say its particularly easy tbh
kindiana#1016: the baselines are pretty good lol
Noa Nabeshima#0290: I'm thinking something like this |
StellaAthena#3530: @Noa Nabeshima Until something revolutionary happens the bottleneck is cross-node communication.
Noa Nabeshima#0290: speed of existing compute probably follows an exponential that's slower than exponential rise in money that goes into training large models
Noa Nabeshima#0290: so is parallelization going to be a big deal ever? if so, how soon?
StellaAthena#3530: Parallelization of what?
Noa Nabeshima#0290: is this a big bottleneck, or is it fixable with money?
Noa Nabeshima#0290: the training procedure for large models
Noa Nabeshima#0290: like, it was difficult to fit large models on single GPUs so we developed solutions to that
Noa Nabeshima#0290: and also money can't get you faster GPUs at a certain point so you have to use more GPUs instead
kindiana#1016: I think google pretty much solved it with tpus
Noa Nabeshima#0290: at least for now
kindiana#1016: you can connect up to 8192chips with very high speed interconnects
kindiana#1016: it should also be relatively trivial to scale larger, but they don't have the need to lol
StellaAthena#3530: No it cannot.
kindiana#1016: I mean, you can if you are google lol
kindiana#1016: tpus have 500gbps links
Gabriel#0454: @Sid Ok, using tensorflow v. 1.15.2 helped, so thanks for suggesting it. But now I'm told that `ValueError: CrossShardOptimizer must be used for model training on TPUs.` I've expected mtf to automagically handle that, so that `mtf.optimizer` created crossshardoptimizers when dealing with TPUs. Does it?
StellaAthena#3530: Are you saying that Google's model training is not limited by their ability to pass parameters across nodes?
kindiana#1016: It would go faster if the links have unlimited bandwidth, but not that much faster, its bound by flops as much as network
kindiana#1016: tpus have an extremely favorable memory bandwidth to network bandwidth ratio which allows this
StellaAthena#3530: Huh. I didn't not realize that. So, everyone but Google is bandwidth limited? |
kindiana#1016: the new a100s also have a lot of banwdidth, and I imagine if you hook it up in the same 2d toroidal mesh that google does you can achieve similar scalability
kindiana#1016: I don't think any company is tackling this particularly cleverly, just brute forcing it with more bandwidth lol
kindiana#1016: but it seems to work so who's to judge
Noa Nabeshima#0290: What about things like.. needing computation to be sequential
Noa Nabeshima#0290: isn't that something you can't spend more money on
kindiana#1016: the amount of serial steps you need to take does not increase significantly with model size
kindiana#1016: from the openai scaling laws its like N^1.03 or something iirc
kindiana#1016: and we are nowhere near having each serial step being bound by anything physical
kindiana#1016: you can just throw more hardware at it so you do the steps faster
Noa Nabeshima#0290: Wait, but why was DeepSpeed important? Will DeepSpeed-esque stuff scale to 1000T parameter models?
Noa Nabeshima#0290: maybe I should say 10^n parameter models
kindiana#1016: deepspeed is constant factor improvements when you are bound by network bandwidth or accelerator memory
kindiana#1016: https://arxiv.org/abs/2011.03641
kindiana#1016: you might enjoy that paper
Noa Nabeshima#0290: And Google won't be bound by those?
Noa Nabeshima#0290: in the next 10 years?
kindiana#1016: you can throw more hardware at it
kindiana#1016: if you make a 1 million node tpu pod
StellaAthena#3530: nobody knows what will happen in 10 years
kindiana#1016: I'm sure you can train some large models |
Noa Nabeshima#0290: Thank you for your help ❤️
kindiana#1016: np
StellaAthena#3530: 2010 was ImageNet and Kaggle ~~and the ability to recognize videos of cats~~ (this was 2012)
Kosakosa#3846: Hello!
Gabriel#0454: Salutations
dopa#3178: https://tenor.com/view/penguin-hello-hi-hey-there-cutie-gif-3950966
triggerhappygandi#0001: I saw one guy on LinkedIn saying neural nets will never understand NLP, and that it's _proven_
triggerhappygandi#0001: Where is he coming from
Daj#7482: GOFAI
Daj#7482: and/or philosophy of mind
triggerhappygandi#0001: https://www.linkedin.com/feed/update/urn:li:activity:6755383759782072320?commentUrn=urn%3Ali%3Acomment%3A%28activity%3A6755383759782072320%2C6755393125587529728%29
Daj#7482: Not worth engaging tbh
Daj#7482: People like these are everywhere
triggerhappygandi#0001: I try to see other perspectives too. Neural nets really _have_ skewed all ML research irreversibly.
Daj#7482: because they work lmao
triggerhappygandi#0001: To the point that I don't even know much about traditional ML lol
Daj#7482: I was one of those people harranging NNs too
Daj#7482: for a long time
Daj#7482: but results don't lie
triggerhappygandi#0001: Yeah they don't. But now it's like you can't even explore too much |
triggerhappygandi#0001: The hardware works in tandem with NNs
Daj#7482: Of all academic failure modes, this is one of the most benign imo
Daj#7482: Other fields are _so much worse_
Igor Krawczuk#1653: Neural networks are also really fertile breeding grounds for general information theory/optimisation/whatever methods. If you want to explore, feel free
Igor Krawczuk#1653: If NNs ever stop working, we'll be able to use a lot of the tooling in the *next big thing*
Daj#7482: btw guys, totally different topic: I'm giving a SSC meeting talk at the end of the month. Are there any things in particular you'd like to hear me talk about?
triggerhappygandi#0001: I'm not complaining. Like I said, I legit don't know anything substantial about PCA, SVM etc.
Daj#7482: I'm thinking State of Eleuther + Some updates on my alignment agenda I guess
triggerhappygandi#0001: I just assume "can't we implement them with NNs?"
triggerhappygandi#0001: Yeah sure why not
triggerhappygandi#0001: I've never actively been around alignment discussions
triggerhappygandi#0001: Might as well deep dive
triggerhappygandi#0001: What's SSC
Daj#7482: Question is whether an "update talk" makes sense for SSC, dunno how many people will care/know about previous Eleuther things
Daj#7482: Slate Star Codex, popular blog, they have a great meetup I frequent
triggerhappygandi#0001: Ohhh
triggerhappygandi#0001: Yeah I know that one. Someone linked a good consciousness post once
Daj#7482: https://www.youtube.com/watch?v=EUjc1WuyPT8 best intro
triggerhappygandi#0001: _Finally_
triggerhappygandi#0001: An intro |
Daj#7482: You just had to ask lol
Daj#7482: I have plenty of alignment material for every level
triggerhappygandi#0001: It's an existential crisis rabbit hole. A while back I just wanted to live in ignorance I guess.
Daj#7482: It makes you stronger
Daj#7482: Stare into the abyss and level up
triggerhappygandi#0001: Yeah. _By staring at it_.
Igor Krawczuk#1653: (also don't focus to much on Yud please, I personally think Rohin Shah is a lot more of a nuanced thinker on this topic)
Daj#7482: I think they fill different niches
andyljones#7746: rohin shah is a good choice if you'd like some light at the end of the apocalypse
cognomen#6297: one argument that may or may not have aged well depending on your standpoint: https://www.aclweb.org/anthology/2020.acl-main.463/
fazz#8459: I think you could double down on why alignment >> ethics and what happens if regulatory arbitrage sends AI into the worlds corners a`la Crypto
Daj#7482: Definitely planning to outline my views on AI Ethics yea. Not sure what you mean about the second part?
cognomen#6297: meaning is the relation $M \subseteq E \times I$, containing pairs $(e, i)$ of natural language expressions $e$ and communicative intent $i$
understanding is being able to map $e$ to $i$
the relationship $M \subseteq E \times I$ maps natural language to concepts external to language
*If the training data is only form, there is not sufficient signal to learn the relation $M$ between that form and the non-linguistic intent of human language users.*
TeXit#0796: **witty cognomen** https://cdn.discordapp.com/attachments/729741769738158194/799287037797728256/202096692978253824.png |
StellaAthena#3530: https://github.com/microsoft/DeepSpeed/issues/612
cognomen#6297: GPT-3 is quite capable of handling all kinds of ambiguity and suggestion though so some shifting of goal posts might be required yet again
StellaAthena#3530: This is an amazing
fazz#8459: just a theory that over regulating AI over ethical concerns in EU, US might mean that "dangerous" AI emerges elsewhere China, Malta, Zug.
Daj#7482: Ah yeah. I'm actually less worried about that and just generally not super well informed there
fazz#8459: If I remember correctly there was a related concept you brought up about large scale compute itself becoming regulated similar to a bioweapon. All interesting angles
triggerhappygandi#0001: Do people at Google think about alignment?
triggerhappygandi#0001: I never really read the "our vision" section of Google AI
Daj#7482: Yea I expect this to happen ~~unless if timelines are so fast that just no effective regulation can happen, which is also possible~~
triggerhappygandi#0001: ```
Google’s mission is to organize the world’s information and make it universally accessible and useful. AI is helping us do that in exciting new ways, solving problems for our users, our customers, and the world.
```
Daj#7482: Not really as an institution, other than stuff that is good PR I guess
triggerhappygandi#0001: "But we wont share our 300M images tho"
Daj#7482: never rely on someone understanding something if their salary depends on not understanding it
triggerhappygandi#0001: These mfs might as well be the people who see the first AIs being born, and they don't discuss about it?
Daj#7482: You have no idea how niche alignment actually is
Daj#7482: Which is exactly as crazy as you think it is
triggerhappygandi#0001: I assumed since Ethics is popular alignment would atleast be a significant subset
triggerhappygandi#0001: Now where have I heard this salary quote before |
Daj#7482: yea I think this'll be my SSC talk
Daj#7482: Big Tech sponsored AI Ethics is about as good as Big Oil sponsored environmental surveys
triggerhappygandi#0001: lmao
triggerhappygandi#0001: And they claim to be so empathetic
Sphinx#2092: My personal view is that most people don't about alignment because most of the times it comes up its just philosphical nonsense with nothing of concrete value.
Daj#7482: This applies to Climate Change in the ~~70s~~ ~~80s~~ ~~90s~~ ~~00s~~ now too
Sphinx#2092: At least things like Ethical AI present real-life examples of situations, and attempt to offer quantitative measurements.
Sphinx#2092: We can measure climate change and have scientific arguments. It's not just philosophy.
Daj#7482: Alignment has this too, it's just not as widely spread
Daj#7482: Alignment in post 2016-ish has _plenty_ of applied stuff
Daj#7482: Just check all of CHAI
Sphinx#2092: Then once you make that widespread instead of the philosophy nonsense, you'll have more discussion.
Daj#7482: I'm more cynical
Daj#7482: Decision theory gets less clicks than "AI is racist"
Daj#7482: And that's tbh probably fine
Daj#7482: I don't think Alignment _necessarily_ needs significantly more exposure
triggerhappygandi#0001: I somewhat agree. AGI still doesn't feel like an _immediate_ development to me, and a lot of people. Maybe that's what even people at google research think
Daj#7482: It's a super technical problem
Daj#7482: Yea I think you#re right, most people probably has much longer timelines
triggerhappygandi#0001: We do. You are very doomer |
Sphinx#2092: It's not about timelines, it's just about concrete problems.
Sphinx#2092: If you can't do science with it, then why do you expect scientists to care?
Daj#7482: I don't think "doomer" is an adequate description of my views
Daj#7482: _But you can_
Daj#7482: This is just not true
triggerhappygandi#0001: Andrew Ng paraphrased it like "worrying about overpopulation on mars"
Sphinx#2092: Then engage the scientific community.
Daj#7482: That's what e.g. Stuart Russell does
Daj#7482: but something something salary
Daj#7482: It's probably fine tbh
Daj#7482: It'll probably naturally diffuse
Daj#7482: ~~except I think academia is just a fucking broken mess, but that's an orthogonal problem~~
Daj#7482: Ng can fight me lol
triggerhappygandi#0001: Indeed. Me too here
triggerhappygandi#0001: Thats just brushing the problem under the rug
Daj#7482: The arguments against Alignment work, imo, are uniquely _terrible_
Daj#7482: Stuart Russell has many amusing comebacks
StellaAthena#3530: @Daj is there a compendium of these?
Daj#7482: There are some in Russell's talks and his book (Human Compatible)
triggerhappygandi#0001: I'm not even that much into it, and even I have wondered quite a few times about the current progress trends and how they relate to employment |
StellaAthena#3530: @triggerhappygandi re: China, I’m unaware of any examples of technologies or inventions coming out of China in biotech that aren’t
1. totally possible to do in the US; or
2. hilariously unethical
As a result, I’m rather skeptical of over regulation claims. Can you provide some examples?
triggerhappygandi#0001: Nono
triggerhappygandi#0001: I'm not thinking about regulations
Daj#7482: The obvious conclusion is China is also overregulated lol
triggerhappygandi#0001: I'm thinking about what happens when human labor is replaced in 90% of the cases
triggerhappygandi#0001: What do we do when there is no work to be done
Daj#7482: Lets just take one country (May I offer Switzerland maybe?) and just make everything legal there, everything. I'm sure the results would be interesting ~~(This is mostly a joke, mostly)~~
Daj#7482: The concept of "work" is a human social construct that will not make sense anymore
triggerhappygandi#0001: I dont think we are prepared for it
Daj#7482: obviously
triggerhappygandi#0001: So what do
triggerhappygandi#0001: We obviously need a mass-wakeup
Daj#7482: Get the AGI to not turn the universe into paperclips
Daj#7482: Nah
StellaAthena#3530: Oh I was (poorly) responding to this comment. My bad
Daj#7482: I don't think so
Daj#7482: I don't think it's important that your grandma has an opinion on amplification vs decision theoretic approaches to alignment |
triggerhappygandi#0001: Regulations are at best a slap on the wrist
StellaAthena#3530: Have you never seen Star Trek?
triggerhappygandi#0001: Lol no actually
triggerhappygandi#0001: Also something something generalizing by fiction
StellaAthena#3530: Oh I’m totally being flippant
triggerhappygandi#0001: What does regularman do in star trek
Daj#7482: wear red shirts
cognomen#6297: that assumed a post-scarcity economy, unlimited resources and unlimited means of production
Daj#7482: ~~This is a stupid reference~~
Daj#7482: I think you're still thinking _way_ to parochial
Daj#7482: You assume there will be "people" doing "things"
Igor Krawczuk#1653: @Daj just to clarify, do you think AI alignment is important mainly because the AI will go FOOM and kill us all, or do you also care about enforcing alignment in the sense the billionaire class doesn't just define "subjugate the peasants" as the reward function?
Daj#7482: None of that will happen unless a _lot_ of work is put into achieving that
StellaAthena#3530: Is this not a fair assumption if there *isn’t any work to do*
Igor Krawczuk#1653: I.e., is it a technical problem or a political one
Daj#7482: 99.9% technical problem
Daj#7482: The social problem is almost totally negligible other than how it speeds up timelines and inhibits free research
triggerhappygandi#0001: Duh. I can't think in terms of a society entirely foreign to me. What do I even draw assumptions from?
StellaAthena#3530: It’s a political problem in the same way that WMDs were a political problem in 1400.
Daj#7482: I'm something of a moral antirealist realist apologist |
Daj#7482: And I'm _way_ more afraid of s-risks than _anything_ else
Daj#7482: I think this is a poor metaphor
StellaAthena#3530: Yeah, I started writing a follow up and kept deleting what I wrote
triggerhappygandi#0001: I find it hard to imagine a world where you don't _have_ to do anything
Igor Krawczuk#1653: Okay, then there I see the divergence. For most people, it's easier to see the risk of Putin or the US weaponising GPT-3 for psyops than of it going FOOM
Daj#7482: Yes, and I think most people are wrong and using the wrong mental models to politicze the issue and therefor are a net negative to alignment
Daj#7482: Which is why I don't endorse wide ranging outreach
Igor Krawczuk#1653: (still need to finish reading the report so I won't enter that debate yet)
Daj#7482: Back in the ancestral environment, when a grave threat emerged on the horizon, we rallied the tribes and fought it off
IKEA#9631: "cease and desist" :brr:
Daj#7482: this is a fundamentally broken view of thinking about AGI
triggerhappygandi#0001: Your twitter rant was very informative. Hopefully that google form the dude put out does something lol
Daj#7482: This is not something that benefits from more cannon fodder
Daj#7482: like a stone age war does
Daj#7482: But people are using models developed for stone age war to conceptualize the risk of AGI
triggerhappygandi#0001: Indeed. This is literally foreign waters
triggerhappygandi#0001: In this case, history really does not repeat
cognomen#6297: it rhymes
triggerhappygandi#0001: Unless we compare ourselves to mammoths and AGI as hunter gatherers
triggerhappygandi#0001: In which case, :guilty: |
Daj#7482: the only old metaphor that works is summoning a demon/angel/god
bmk#1476: If we can successfully define "subjugate the peasants" without also accidentally destroying the universe, that will have been a massive win for alignment
StellaAthena#3530: Me too lol
Daj#7482: We are using arcane symbols (code) to harness the power (electricity) of immaterial spirits (programs) to do our bidding
Daj#7482: Now we have to figure out how to not get tricked by the demons
Daj#7482: and summon an angel instead
Daj#7482: Turns out, really fucking hard, who could have guessed?
triggerhappygandi#0001: You just increased the pressure by 1000%
triggerhappygandi#0001: Jesus christ now we have to stop slaanesh?
bmk#1476: That pressure has always been there
Daj#7482: Slaanesh is a surprisingly good fictional metaphor lol
triggerhappygandi#0001: yeah but I was ignorant, blissfully so
Daj#7482: The warp is just the internet
Daj#7482: The initial panic will subside, don't worry
bmk#1476: Welcome to the club
triggerhappygandi#0001: :neetz:
triggerhappygandi#0001: This guy comes to mind now
StellaAthena#3530: Eyyyyy it’s my bro
Daj#7482: Nietzsche was a pollyane optimist lmao
Daj#7482: Get that weak shit outta here |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.