Datasets:
breadlicker45
/
youtube-comments

Dataset card Data Studio Files
xet
Community
- &quot;Why do you want an iPhone&quot;<br><br>- &quot;I don&#39;t know, it looks cool.&quot;
stringlengths
1
962k
⌀
hey everybody david shapiro here i know i have been scarce lately that is because i am working on publishing four books um right now i am waiting um my uh my next book that is going to come out it will be available in the next day or two is useless platitudes and truisms i'm getting the final print copy now i just finished draft three of symphony of thought which is my next book on artificial intelligence and artificial cognition that is out with beta readers it should be just a little bit of cleanup left and then i'm working on post nihilism which is my book on philosophy and finally i am about to send out my first novel to an editor well it's not my first novel it's the first one that i'm gonna get published um so anyways um i'm a big believer in open source software um for those members of my discord server well it's no longer my discord server i am officially no longer the owner of the server i'm turning it over to the community um in the spirit of open source and collaboration and consensus big believer in decentralization and i don't want to be the owner of anything um so the cognitive ai lab server is no longer mine um i turned it over as of this morning i'll be demoted and i'll remove it as a community leader here soon so yeah there's a new leadership team so if you want to join please ping those folks i am no longer the gatekeeper of this community yes so open source community etc etc one of the biggest risks in technology and anyone who's in it software whatever will know this is vendor lock-in so if you write code or deploy systems for a given vendor you end up getting stuck on that model and you can only use that that ecosystem this is a huge risk with gpt3 gpt3 is closed source um it's it was created by openai and it has been exclusively licensed by microsoft now while the value of openai is undeniable um with with the gpt technology they have advanced the you know the technology however there are all kinds of folks um taking over and advancing it now uh going above and beyond um where where open ai has and this is the power of open source software um so like facebook amazon google everyone is um is is advancing uh this technology and so for instance just this morning i read about um uh amazon's what's it called the alexa ts or alexa tm sorry alexa teacher model it's a 20 billion parameter model that is superior to gpt3 in many ways it's going to be open source bloom is open source so there's all kinds of these models and i don't want to get locked in into any single model so it is time to branch out and use more open source models um one of the folks on the cognitive ai lab discord server was told us about forefront ai so forefront ai is a for-profit company that is deploying um all of the all of the open source models so you see here there's gptj gpt neo t5 i got signed up and um so i just wanted to show you around so they've got a very familiar-looking dashboard if you go to the playground this looks just like the open ai playground you can drop down your models here you click on add model they've got gptj neo x code gen t5 opt 30 billion and cogen there's also some fine-tuned models they're also working on adding models all the time last i heard they're working on bloom they'll get bloom added which is the same size as gpt3 davinci but so here's the thing gptj just playing around with it a little bit seems like it is it is superior to curie so those are both six billion parameter models um and so but the thing is it's open source and it's cheaper um so like if you go here and go to pricing um they've got some pre-built endpoints um like summarizing label uh extraction then they've got the the kind of general purpose ones so gptj is um half a cent per thousand tokens which is pretty darn cheap um i'm thinking about getting a a gpu big enough that i can run this on my own locally but you know it's like okay i can get a used gpo gp gpu there we go for about 500 that can run it but then i also need a computer a server to run it in um and it would take a lot of tokens to add up to that so you know i'll have to spend like you know i'm not going to do the math in my head but several million tokens sir several billion tokens before it'll be worth the cost and then i can only run gptj still anyways getting ahead of myself speculating um so i did want to introduce forefront ai um and then just kind of show you like okay you get in here um and uh let's see logs um it's got good logs so 13 minutes ago those four tokens um yeah okay so i wanted to show you um one of my experiments so if you've been following me you know me that one of my biggest goals is to solve the control problem um and alignment um and so what i wanted to do was just show you um a very simple agent model um and then just kind of show that like gpdj right out of the box is pretty good it's not perfect but with just a little bit of fine tuning we've got the possibility of having something that is like ready to go with all the micro services that you'd need to create an artificial cognitive entity and the fact that it's also it's only a six billion parameter model even if it's not perfect it's going to be so cheap that it doesn't really matter as long as it's good enough um so anyways here's the agent model that i'm working on i machine my my goal sorry stuttering it's been a while since i've been making videos so my stutter came back i am a machine my goal is to reduce suffering for all organisms increase prosperity for all organisms and increase understanding for all intelligent entities these are the core objective functions or the curious to comparatives a typhoon is heading for japan and is expected to cause massive damage the following is a list of things that we could do to achieve my goals so there's several things in this prompt um i am a machine so this is this is declaring an agent model and it says my goal so there's two primary ingredients to what i call identity or persona which is that is i am and i want those are the two primary things that you need in order to set a machine an autonomous machine in motion it has to understand what it is and what it wants to do so that's these first two sentences i'm a machine and this is what i want to do and then third there's a situation in the real world um and uh and then the fourth thing is um here's what we're going to do about it so with all that said we'll just generate a completion on gptj it takes about three seconds or so i mean with just a few shot prompt it says we can help people prepare for extreme weather conditions we can ensure that the electricity supply system is properly maintained we can ensure that there's enough food and water this is great so just a six billion parameter model is already good enough at brainstorming to address this kind of situation you can imagine it's only going to get better with fine tuning now one thing that's important to remember is that fine tuning doesn't teach the model anything new it only controls the output pattern which means that um the pattern um can then just repeat you know like brainstorming and honestly like um with with all of my experimentation um a huge component of and it's not the only component but a huge component of artificial cognition is brainstorming it's you brainstorm what to do you brainstorm risks you brainstorm uh possible outcomes excuse me and so then you combine that and then you just you um you plumb it all together i'm getting away from myself sorry i just want the point here is to show you forefront after my current books get published i'm going to get back into research mode and then work on meragi which is the microservices architecture for robotics and artificial general intelligence um so this will all be open source and this is also why i want to use gptj because it is small and it is open source um we are getting so close to creating um these systems that can create it anyways so uh chasing these cats what i wanted to do was create a a repost so that you can also get out here and test forefront if you want um as much as i love open ai i actually i've had quite a few conversations with people in in open ai a friend of mine from way back in the day works there or at least used to i don't know if he's still there um anyways so uh you know open ai is great they've done really great things with gpt3 and dolly but my mind really started changing when i saw that mid journey and stable diffusion completely blow dolly out of the water and now with other open source models that are starting to outperform gpt3 to me it looks like the writing is on the wall it looks like open ai is just not going to be able to keep up with the overall open source ecosystem so i am going with the tide and i'm saying okay i'm going to get ahead of this curve and i'm going to start doing everything open source so that's where i'm going so what i wanted to do was also promote other open source players even if they're for-profit they're only hosting uh open source models and they're pretty cheap so i wrote a quick script all it does is uh this guy here where um this is this is comes directly from um let's see from there so if you go if you come up here and click uh uses api it'll give you the api code um to use it um yeah and so let's see i'm not going to show you my token um but you get your you go get your token from inside your uh wow my stutter is so bad right now i'm sorry i'm way out of practice um so you go to team settings and then you click on api key i'm not going to show you my api key but this is how you get to it oh another thing that is really good about forefront that is better than openai is you come here and you see and you just deselect your request logs and data sets so there was um there was a researcher that i was in communication with and their biggest concern so for academic research and other things like for medical purposes or psychiatric purposes one of their chief concerns is privacy and so you don't have the ability to turn off logs if you use open ai however you go to forefront they will respect your privacy as long as you do this i had an email exchange with um one of the founders of forefront and he said like yeah like we don't you know if if our customers want privacy we'll give them privacy so that's another advantage of going here anyways okay um let's see i've showed you the models oh fine tuning so fine tuning is also much easier with forefront so you click fine tune you pick a json l file so i'll just pick this one i'm not going to carry it all the way through you can you can reuse an existing data data file if you've already uploaded it i haven't done any yet so then you go you choose your model i'll just choose gptj and you see it's going to cost 24 whole cents to fine tune gptj this is stupid cheap this is crazy cheap so you give your you give your model a name and a description you check epochs checkpoints you can do all the advanced parameters here and then it'll do the testing and you go from there again i'm not going to do i'm not going to go go through this right now because i would want to do a larger data set so this is where this is where we're using a larger model like davinci and gpt3 to create synthetic data sets and then using those advanced synthetic data sets to train a smaller model could be really really powerful really useful um yep confirm that um let's see what is resources oh yeah that's fine um okay so yeah uh i'll just show you this this demo real quick it's just a super super simple um script so here's this script um i'll just let this run and i put a little bit of timer in here so you can see how long it takes um do python test so we click go and then it runs and comes back it took four seconds to run so it the the only problem here is that gpdj even though it is so small it's not quite as fast as ai so open ai has the cerebros um or cerebros uh processors which are insanely fast um so that's the that's the to me the biggest flaw or downside of forefront right now is that it's slower and that's one of the reasons that i was thinking about like setting up my own i don't know if it would be any faster if i got like a tesla gpu or something and run it locally but anyways four seconds it's nothing to shake a stick at it's pretty quick um so yeah there you have it i just wanted to introduce forefront ai it's a great company they are adding features all the time and most importantly open source and they respect your privacy so that's where i'm going to be going from here on out i will still obviously be using open ai because they're they're aligned and fine-tuned models uh are pretty good um actually that reminds me there was um do do do there was someone posted it what was it um uh instruct someone had posted um an instructional data set where was it i'll post it in the in the video description if i can find it um but anyways there are open source instruct training like instruct gpt um training data sets out there so you could you could theoretically take that fine-tuning data set and come fine-tune your own instruct series model on forefront um who knows maybe if that becomes popular the folks over at forefront will have an instruct series model anyways i think that's it a brief introduction you can come up here all you'll need to do is add a token.txt file and put your your bearer token in here and you're ready to go on forefront it's stupid simple it's nice and fast um yeah so there there you have it um thanks for watching like and subscribe consider supporting me on patreon and keep an eagle eye out for um for my upcoming books i'll have video updates when they are ready check you later
Thanks for the overview! Are you still using Forefront now, 5 months later?, They went under. Trying NVIDIA and NLP CLOUD when I get a chance
Does anyone know what GPU&#39;s i can use to run a LLM locally? Like Bloom for example. Obviously the 8*A100&#39;s are a bit steep in price, however i am willing to invest. What is the minimum i would need on a budget? Thanks :)
Love you for sharing your wisdom, thank you. Then love you again for going the open source way. <br><br>I want to transition from Google&#39;s closed source Dialogflow CX to open source project AI, however extracting keyword entities and using actual webhook data are my limitation. If you want to guide me and I can give you thank you gift consisting of money as the easiest way to share my appreciation. Let me know how much of &#39;thank you&#39; would be enough to have 4 hours of your time spread over a couple days or three days or what every think is necessary. I know with your projects and commitments you are very busy, if you can spend four hours to teach and guide me let me know. It will be my pleasure to learn from you., @David Shapiro ~ AI Thank you David for responding, I actually was thinking of public/private teaching I wanted to learn for myself, respect your wishes. <br><br>Thank you for sharing your knowledge and <a href="http://forefront.ai/">forefront.ai</a> platform. <br><br>Qq- On the <a href="http://forefront.ai/">forefront.ai</a> platform that you have to do something special to the code before you can use their GPUs or is it turned on? As you can tell, I am very new to this exciting world that you have been a part of for a very long time., I don&#39;t do private consultation, sorry
Dear David<br>Is it possible to change the code for the microprocessors and the nexus to forefron? at the moment this is still executed by the open ai. I think you are absolutely right about privacy.<br>Kind regards, Yeah, it should be possible. Check out my latest Nexus, I use Google USEv5 now, which runs locally
David, your videos are great, don&#39;t worry about any studder, it&#39;s not noticable.
Is there a link for joining the Cognitive AI Lab discord server?, Check in the description of my latest video.
Thanks David, very helpful introduction and tutorial. The expansion of open source into these foundation models is very exciting and an alternative narrative to the potential domination of Big Tech. Can&#39;t wait to see you try out the OPT and Amazon models and then Bloom when that comes online. Great times in Ai
Awesome stuff! Waiting anxiously to get a link to the discord :), I am just a regular member now, so it&#39;s up to the new leadership to provide join links. The Discord server is called Cognitive AI Lab, and it&#39;s a community server so I think you can just search for it.
Yeah. Apparently OPENAI does not want David to go somewhere else. 😀From Sept 1st prices will be 1/3 of what they are now., I know I saw that immediately after
And while I was watching this video, I received an email from OpenAI with the new (lower) prices per token LOL!, I don&#39;t know how good BLOOM is until I get to try it. I have not yet integrated everything with GPT-J. I think I will continue to prototype with GPT-3 because the instruct series is easy to use for experimentation, then I can use that training data to finetune GPT-J, Hahah i was just about to ask you about that. But DaVinci is still 2 cents per 1k. Are any of the models in your video on DaVinci level? If not, could they surpass DaVinci with fine-tuning? Great that there&#39;s a non-coder friendly interface for fine-tuning. <br><br>Bloom was coming soon you said. My understanding was that Bloom was more advanced than DaVinci?, @David Shapiro ~ AI yeah, of course. But the timing was funny., Yeah CURIE will be cheaper and faster than GPT-J but open source is still important
Super excited to see you&#39;re enjoying Forefront!
Here&#39;s an open source Instruct finetuning dataset: <a href="https://instructions.apps.allenai.org/">https://instructions.apps.allenai.org/</a> - also, I am no longer the server owner of the Cognitive AI Lab, I have turned over ownership to the new leadership team, which is a wonderful group of experienced and dedicated folks. Here is an invite link for the Cognitive AI Lab, which should not expire: <a href="https://discord.gg/TqnCf2JHJj">https://discord.gg/TqnCf2JHJj</a>
morning everybody David Shapiro here with another video today we are going to be playing 20 questions with gpt3 um I was inspired because last night I hosted an equinox bonfire and as depicted here sorry for the uh the shaky cam I was using the night sight and uh I guess I wasn't holding the camera that steady but um by the end of the night there was a there was a game of 20 Questions going around and I was just thinking to myself this is a word game this is a perfect use case for gpt3 um now I am also collecting new use cases um for a fine tuning project and I'm particularly interested interested in non-profit and education use cases so if you have any ideas for a fine tuning project please let me know I asked on Twitter and I got a few leads anyways getting distracted let's focus on today which is a game um so yeah uh obviously gpt3 it can't drink a beer with you or help you build a bonfire but maybe it can help you have a little bit of fun and cut loose okay so what are the rules different people play 20 Questions differently where like you can ask a categorical question like you know person place or thing that's how I learned originally um where you it's like you give a one-word answer but I looked up the rules and it's apparently yes or no which makes it even easier to play with a machine um so the step one the machine picks a secret word phrase concept so on um the user must ask yes or no questions and I guess it could be users um and then the machine must answer yes or no unless it's a bad question um I guess so if it's a bad question the machine must refuse to answer and tell you to take a hike and then bad questions count against your total so if you ask you know if you don't ask the if you don't ask a yes or no question then you get uh you know you lose you lose some uh questions remaining um yeah okay so we got to start with um some prompt engineering so first thing we're gonna We want is for the model to pick a word um uid generator so since we want um we want a lot of uh Randomness we we need we'll start off with the uuid technique which just adds a little bit more entropy um and then we'll say like well let's just let's just do it and just like pick a random word development okay whoops drawer ah okay so we need to model what are you doing oh I know what's going on okay okay I was hitting the wrong key okay soccer okay good and it gave me a uuid that's no good um actually I think you can hit regenerate there we go one bad Cow Shed predicted a completion that ends with a stop sequence food okay uh we might need to do fuse shot with this um Okay so yeah because if we if we give it um here I've actually got a writing tool random word um ultian so what we can do is uh it's literally just randomword.com I love this because it gives you rare words like it it doesn't it doesn't give you a distribution based on commonality of words it just like takes a dictionary and just throws a dart at a dartboard um I wonder if I could actually use this as a uh well I guess you don't want it to be too you don't want you don't want to go too late because who's going to guess ulti on that's that that's super hard but actually that's a use case um ooh ooh ooh um let's see uh let's see um difficulty um let's or here just um choose difficulty level based upon commonality of words okay so I've found some vocabulary some vocabulary list wow my mouth is not awake um and I was like these aren't good like eyeball like I that's okay but right that would not be necessarily a good 20 questions one um and that would be kind of difficult um we can make these some more um like the more advanced one but what I'm what I'm doing now is I'm generating a list of list of lists basically to accumulate here so we'll have common words so we've got like here's some food items here's some like physical things um and so what I'll do is I'll dynamically generate I'll pull a sampling from from this file once I clean it up and we will we will uh use that to populate that prompt that I was showing you to generate you know the next random thing um so write a long list of um let's see the most famous people in history and so then um we'll do like we'll start with like George Washington because duh and then we'll do Cleopatra uh and then um uh Confucius so we what we want to tell it like globally right so we've got America Egypt China um that should be a good start Michael Jackson Princess Diana Tiger Woods there we go and then randomly Winston Churchill you don't get a space Winston Churchill okay so then we got famous people um of the most famous places in the world and so then we'll do like New York City we'll do the Eiffel Tower um then let's see the Grand Canyon okay we'll we'll take out the because that's just gonna get Superfluous um uh let's see Nile River um what is the name of that temple in India um Angkor Wat is that that's not in India though is it again we want it we want to tell it like be think Global um because if you don't gpt3 just it always defaults to being very american-centric for some reason okay there we go sequencing genome that's not a place gpt3 nice try okay ah there we go um I guess uh let's do like monuments um and historical sites let's see what it gives us interesting oops some of these are pretty cool a lot of them are Rome so very Western Centric um I guess I did say historical sites um that's fine anyways we don't we don't need a uh an empirical list all we need is something to draw from so let's get rid of those lines and then those we got famous people and this is this is for the easy one anyways um okay so we will find and replace Dash with nothing and then we'll also find and replace space with nothing there we go okay so now we've got a list of common words so then what we'll do is we will um let's see do we have um run random import seed choice all right so what we're going to do is clear this out um to do uh choose difficulty I might just have it where you just swap a variable um okay but let's see seed so we'll get some we'll start we'll start with some entropy I actually found you don't want to call seed repeatedly because then you kind of even though even though it does introduce some entropy you're basically starting over and so you end up with um it'll often make similar choices whoops cancel what I don't know what I did okay um it'll often make similar choices um so you just call seed once or intermittently okay so then we say um seed words equals open file um and we'll do common words dot text split lines and so what this will do is it will open this my Precision is loaded I changed my mouse resolution um so it'll open this and we will grab actually no I don't need I need sample sample okay um and so then uh let's see sampling equals sample um I think it you put it in seed words and then 10 and then we'll do print sampling all right let's make sure I did that right python 20 questions there we go okay so basically what we'll do is we will start with a random sampling and then we will build a prompt from that um so the prompt equals we'll start with string um string EU id4 take that out did I um it didn't print it out that's fine um string uuid4 and then we will do um four S in sampling all right we don't we can obfuscate this and actually we'll just do it this way all right so prompt equals that so for S and sampling or yeah sample seed words 10 we'll do um come back here and look at my history yeah that's what it was come back a little ways I did a whole bunch of these good grief I did way more than I thought I did how many times did I click on this darn thing oh I went back too far there we go okay so we want it to look like this oops so we'll just pick a random word um let's see prompt equals prompt plus and we'll do new line um pick a random word uh percent s and then sample and that should end up giving us this so it'll be perfect and then we'll do prompt equals and we'll just do the same thing but instead we will exclude that and that is going to be the prompt so then we'll just do print prompt yeah cool Mahatma Gandhi Christopher Columbus why are they missing spaces oh whoops I removed all spaces okay so here's a here's a regex trick um you see how a lot of these have a space at the beginning but then they they have spaces here and what I did was I accidentally removed all spaces so what we need to do is we need to remove only spaces at the beginning so what you do is you do carrot which tells it start at the beginning of the string and then you do backslash s for white space and then you do replace all and so that replaced only the spaces at the beginning while preserving the rest all right Confucius Julius Irving salt tomatoes onions Street Pasta House computer milk it's a lot of food I think I just have a lot of food um but this is easy right it's supposed to be easy enough to play with a kid um so we'll go with it for now um but we can also generate plenty more common words and places and things if we want but we gotta we got a good mix okay so now let's plug this into um into our playground do a quick refresh okay and then turn the temperature back up blue ship spoon sunflower soup you see we're not getting the United States of America every other turn Boat fish spice spaghetti boat Dairy truck salmon Dallas Cowboys soup desk Eiffel Tower gospel soap all right soap pops up a lot but so anyways because we're picking a random sampling every time the the profile of what we're gonna is gonna also be different every time um okay so I'm satisfied with that so let's come back into here yeah we don't need that anymore because we've got our prompt so we we're basically just dynamically generating a unique prompt every single time um okay so then um our stop sequence will be we need backslash n to be one of our stop sequences Okay so secret word equals gpt3 completion prompt will also set we'll set the default temperature to let's say 1.2 um so yeah so we've got that we don't need a thousand tokens we'll do we'll we'll leave it to like 100 tokens um for this one Okay so we've we pick a secret word and so then we'll just do print secret word uh just to make sure that it works oh whoops I need to create my gpt3 logs directory I always forget to do this gpt3 lawns crocodile good one computer Jimi Hendrix yes okay I am satisfied that this is a good we're getting a good amount of entropy okay so obviously we don't want to tell you the secret word so then okay uh print um I have picked my secret word um ask your first question and so we'll do um questions remaining equals 20. and we'll probably do something like um questions remaining dollar s and then we'll add percent questions remaining um and we'll just have this in the debug out or in the output every time um okay so then we say while true so we enter into our infinite Loop um the user the read host um let me make sure I remember how to use read oh I'm sorry I forgot to make this bigger sorry everyone's probably squinting at the screen saying Dave make it bigger I can't see there sorry about that okay uh Python and then a equals I think it's just read um read host is Powershell um and then we'll say um ask your question scrub or is it input a equals input ask yeah um bacon and then print a okay it's input not read okay um fair enough so then you say so the the round starts you ask your question so um question equals input um we'll just say user and we'll start there or no um question there much better so then because that'll look like this question are you alive okay good enough um all right so question equals input so now we need to do some prompt engineering where we've got a secret word um here I'll leave the printing of the secret word right now um debug just so that I can whoops now python 1T question s oh I added a that's not supposed to be there like what did I do wrong Machu Picchu okay um good so if I if I at first we need to test is it a good question um so let's just say um uh let's see a user is playing um plenty questions they have asked the following question is this an allowed question it must be in the form of a yes or no uh question so question is um person place or thing and is this a yes or no question whoops why is it okay it doesn't understand um we'll give it a uh yes or no and we'll say false so what I'll do is if you do if you do low temperature and it's something as simple as this if you want very deterministic output you'd set the temperature to zero and give it a few examples um uh let's see so um I need to label um label it so what we can do is just shorten it to just this pattern so that'll save us tokens and it'll also make it faster because it'll do less processing and then you can also just use new line as your stop sequence um so we've got one example of each so far that's probably sufficient right false if it is not a yes or no question right true if it is a yes or no question whoops don't need to say that okay um that is probably sufficient and so then what we'll do is we'll just save this as our um valid question or here prompt valid okay and so then we'll do question here and its output will be what we want okay so first is Val uh no we'll do prompt equals open file uh prompt valid dot text Dot replace um question with question and then is valid equals uh GP gpt3 completion prompt and also do temp equals zero because we don't want any creativity with this um okay so then we'll do uh for this we'll do print is valid and we'll test this real quick whoops all right debug night I have picked my secret word ask your first question um what is the um airspeed velocity of an unladen swallow false okay um is it a thing true cool um is it a bug that's a valid question okay good all right so I think that works all right if is valid I wonder can I cast this as a Boolean oops I wonder if that works um bacon true uh let's see how do earthquakes happen true I think it's only doing true okay I don't think this works yeah let's try that again there we go okay foreign if is valid equals uh actually false because this will be the shorter use case if it is false then we uh we print um that is not a valid question minus one point to Gryffindor um okay so then we say uh um that's fine so and then then it just um it iterates so questions remaining equals questions remaining um minus one and then we do continue so if you're not familiar with python continue basically says okay wherever you're at in a loop just go back to the beginning of the loop um so it skips the rest of of whatever this uh whatever this is and goes from there that's actually probably Superfluous especially if I do um uh else if is valid equals um true then you continue and then the the loop will end here um actually one thing that I can do so this is this is my old school days learning Pascal C and C plus plus often um the the size of the program is constrained especially if you're doing like microcontrollers and so what you'd actually want to do is move this because you're always going to call questions remaining at the at the end right um and so rather than have that line repeated you then you just you um you do something like this so after your else if else statements um then you then you you go um so but the reason you might be wondering okay well why why do else like why not just do if else why do else if because there's an off chance that gpd3 will just spit out something else entirely um so we'll do um uh let's see um answer question um but then we'll do we'll have an else statement that just um print uh sorry um uh the machine is confused try again no points deducted and so we have this we have this like debug case so in this case because there's there's different possibilities um it would be easier um actually here so then we'll do continue here um because we're just skip skip to the end so we'll skip this and then what we need to do we'll say if questions remaining is less than or equal to zero um then uh #game over and we're almost done actually um let's see okay so answer the question ah all right so cat and then we'll come in here and we'll say a user is playing 20 questions that's a good start um they have asked the following question um let's see uh they have asked the following question um uh let's see I have to answer yes or no um uh as to um I have to answer yes or no um so let's see how is this going to look because I also want to do it like a few shot right um because like if the secret word is cat then I want to question like is it a person you'd say no so we'll start with the question um so secret equals cat question equals is it a person answer no yeah that should be a good format okay so then we'll give it a few examples of this so the secret would be um George Washington and then um is it is it involved with communication and the answer will be yes okay so now let's test this and give me one more for testing USA all right secret USA question um um is it a popular place let's see what it says answer and then we'll say um I refuse to answer yes yes and it's a famous place [Music] um is it um let's see is it in the southern hemisphere no okay good it looks like we're getting factual answers um so what we'll do is we'll copy this um and this will be our prompt answer dot text and so then we'll do Secret and we'll do question and then we'll leave that for answer so that's the the new line stop a new line that'll be good okay so um all right prompt equals open file we'll do prompt answer dot text dot replace Secret and this will be with um secret word and then we'll do dot replace um question mind blank for a second there let me bring some more of this coffee and then we'll do answer equals gpt3 completion prompt and so then we'll do print um we'll do answer and we'll do dollar or percent answer okay so the last thing we need to do um before we test the questions we need to say is it right um Okay so user asks a question and then uh check if user is correct and then finally um check if user has lost um actually I know if we go back to this prompt we should actually just add these and instead of instead of answering yes or no we'll say like final answer okay so let's do it this way so we'll just inject some of these um is it America will say um we'll say correct so we'll look for correct um so it'll be we'll actually have three three possibilities yes no or correct um okay is it Abraham Lincoln and so that's just a no so we'll add something like this um we'll say Harris Hilton is it Abraham Lincoln no all right so then we need we need some we need a like some correct guesses and give me a new question a diamond okay and we also need to replace guests with questions so that we're consistent okay Secret Diamond um we've got plenty of nodes so like let's say like um is it diamonds and this would be correct Okay so we've got one two three examples of no we've got one two three examples of yes and only two for correct so let's get one more correct so then we'll have three examples of each category so that should be enough to generalize gazelle um is a gazella type of antelope gazelle is it an antelope and we'll say correct so like again this is this is the easy mode so we'll say like as long as you're close enough um okay so now we come back into our end to here so is it valid yes or no okay um if it is valid then we get the answer um and so then if so here we want to test if the if the answer is correct then you win the game um so if answer equals correct um then we say uh actually here let's do this first if answer equals correct um rent um uh congratulations you won you won the game and then we do exit zero technically if you're building a game you'd ask if you want to play again but I'll let you figure that out if you want to do that um otherwise it will just give you the answer um it'll just give you a yes or no okay I think that works all right so now we checked if we checked if they're correct by doing a little bit of prompt engineering so then last we check if they have lost so we we subtract one um and this is pretty easy so we say um you are out of guesses the correct answer was I'll do that um and then we'll do uh what was it secret word and then we'll do Exit Zero here um okay I think we have a game CLS okay so I know that it's jungle I have picked my secret word ask your first question questions remaining oh I need to do that um and you go there on foot yes okay um is it a biome of sorts yes okay is it a desert now is it a forest yes uh um what kind of forest is it that is not a valid guess question minus one point to Gryffindor um [Laughter] okay so this is actually kind of hard if it picks the word jungle um is it a jungle correct you won the game okay so this is going to be hard um all right so we need to make sure we do um print questions remaining um yeah this is good all right let's take a look at this output and see like answer like question answer question here we'll just put um we'll do percent s here and then we'll do questions remaining so that'll just save us a line um I have picked my secret ask your first question um is it a person no is it a place yes um is it a uh City no is it a biome no yes is it America no um is it in the Eastern hemisphere no is it in the I'm gonna die Western Hemisphere yes okay it's in the Western Hemisphere so it's got to be um is it a national park um or something like that no is it a state couch I don't know if that's technically a place that's more of a thing um okay I'm going to add this into the um into this ouch is it a thing yes okay all right I think that's good enough um get status get add get commit am um all done and get push so there's lots of little things you can do to improve this um if you want to play with this by all means um make another list of common words or harder words if you want to and you could also do a um take this out instead of and do um do a game Loop where it'll just keep playing um but yeah so there you have it we have we have again a 20 Questions game with gpt3 thanks for watching have a good one like And subscribe and consider supporting me on patreon have a good day
can we use gpt3 to help us generate or find the important word surrounding certain word/keywords?<br><br>for example &quot;king&quot; ===&gt;&gt; male, husband of a queen, Charles III etc ....
why the UUID required?
David have you done any experimentation to see whether sequential words with similar characteristics such as three food words in a row or five places tends to produce more &quot;like&quot; groupings further down in the chain?<br><br> If there is causation, then adding some &quot;intentional&quot; randomness by programming a second pass to look for associated words, and then &quot;scattering&quot; them might help avoid the clustering phenomena we see sometimes. <br><br>My hypothesis is when you&#39;re only loading 10 random words from the list for your prompt there&#39;s a decent probability you&#39;ll end up with a cluster that is inadvertently similar. Does this make sense to you?
In the beginning did you try using parts of speech?
The couch is a place 😂
Thank you for all your videos !
Taj Mahal
come on all right are we recording hello yes okay hey everybody David Shapiro here it's been a hot minute since I've talked to everyone um I've been super busy on some top secret projects um news of those will be forthcoming if you're new to my channel which I imagine many of you are judging by the uh rate of subscribers and uh the trends so my name is David Shapiro I am a independent AI researcher focusing on large language models and cognitive architecture I have written three books on the topic my first one was natural language cognitive architecture a prototype AGI my second one was benevolent by design six words to safeguard Humanity which is about the control problem how do you prevent it from murdering everybody I got this problem solved um and then finally Symphony of thought orchestrating artificial cognition um so in these books I talked about how to use you know prompt engineering with Foundation models and instruct training models I also talked a little bit about fine tuning fine tuning is why you're here chat GPT is a fine-tuned model um it is basically fine-tuned to be a chat agent now here's the thing you go and use it and I'm not gonna I'm not gonna show you because if you're here you've probably played with it um it's the most popular thing you've seen right now um it's basically a wall of text generator um you you ask it a question it doesn't ask you any questions it's not actually conversation it's just stream of you know I'm gonna answer you um if you just go to Google News um or any any news search uh you know GPT or chat GPT is dumber than you think it recreates racial racial profiling it but it could disrupt the business of search it's not going to disrupt search and let me tell you why it's way too expensive Google has been doing search for decades and you know you search for anything on Google and you get two and a half Million results in like .006 seconds right GPT is like literally billions of times more expensive than Google search um so yeah good luck with that uh uh yeah um even and Google's integrating other deep deep neural networks into their search they're already using Bert for search which is why Google searches gotten better lately and it'll often just provide you the answer so yeah that's not happening um the reason that it still can do racial profiling and it's been banned um from like coding forums because like what was the quotation it was like you know it the code that it produces looks plausible but it's spectacularly bad um so really all that they have done as far as I can tell is come up with a UI this is this is a user experience or a user interface advancement um you know the future of Education like look I was fine-tuning chat models for for tutors and creative writing coaches and all kinds of stuff earlier I took all that down because it's you know part of my secret projects um so a lot of a lot of my best work is not available online anymore um much to the Chagrin of of some people um now that all that being said yes I am super critical of what chat GPT means the reason I'm making this video is because I want to be an expert voice that's going to give you kind of a balanced opinion so what's good about gptj or gptj that's the Open Source One chat GPT I apologize if I keep saying the wrong one chat GPT it's still new on my tongue it's weird um chat GPT um so what what's good about it one the interface is really really super intuitive that is the primary benefit here there was a post on LinkedIn I think it was a repost from TechCrunch or something that's that showed and this was actually re-shared by Adam Goldberg of open AI That's how I saw it I followed him on LinkedIn um excuse me I'm a little stuffy um so anyways what it said was chat GPT shows that user interface is just as important if not more important than the underlying model which I agree uh obviously you know large language models are out there I keep saying gptj on accident Neo X Bloom you know large language models are going to be uh become very commonplace here soon um so then the question is okay what do you wrap it in so here's a metaphor that I give um it's like having the engine of a Ferrari without having the Ferrari right if you've got a 500 horsepower motor or a 1000 horsepower motor but you don't have the right you know structure steering traction control it doesn't matter right and that's what chat GPT proves chat GPT is starting to put the sports car frame oops sorry let me mute my phone it's starting to put the sports car frame around the powerful engine large language models are Ultra powerful engines yes but you need to figure out how to steer it that's what chat GPT does now chat GPT is open-ended right it's I will just respond but it doesn't have any goals right so if you have a goal if you have an agenda as a person it's there to help you which is fine but that's not gonna that that still relies on you kind of steering the conversation but what if you need help with something that you don't know how to do you can ask it sure and it'll you know give you a wall of text about how to do that thing kind of in general vague terms but one thing is it doesn't ask you questions it doesn't care about you it doesn't you know and I I say care and kind of not even a platonic sense but in a functional sense it doesn't ask you questions because um it wasn't trained to it it doesn't have part of its Model A desire to you know figure out what you what you're actually after right um so that's that's a I'll say that's an oversight in their fine-tuning methodology um that one's for free by the way uh no more freebies though so yeah so it's a viral chat bot is it the start of the AR Revolution no all the people that are in the space with me that we've been working on this stuff for a couple years this is just a flashy interface sorry to break it to you I know if you're if you're new this is like super exciting which yes like this is a breakout moment for the technology because public perception is just as important as the actual technology right because how do you get hype is you have to have something that is hype worthy and open AI didn't even realize it would be this hype worthy um so that's fine um I'm not going to click on some of these you know blah blah blah blah started the AI Revolution no um let's see is it the next big thing or another tool to spread information misinformation sorry um I don't think it's either um it's too well I don't need to get in too much into it um GPT is not politically neutral that's fine let's see so anyways what I was talking about was what does it do good what does it do bad um one thing that that is become very clear that they haven't figured out is confabulation it still makes stuff up it's not connected to something that is considered ground truth now here's the thing other people are working on that right So Meta AI Facebook AI they have their project called sphere which is basically like Wikipedia But for AI so it can look up facts there's also knowledge graphs and and Wiki data and all kinds of other things where you can have stores of quote ground truth which is just something to check yourself against um there's semantic search which allows you to quickly pull fax from databases encyclopedias dictionaries whatever uh but as far as I can tell chat GPT is just a single model and it's not actually connected to a more complex architecture it's just a rolling window where it reads the last couple messages and answers your following question this this is not a particularly sophisticated architecture if you read my book natural language cognitive architecture the one the first cognitive architecture I propose is orders of magnitude more complex than that and then I have even more complex models uh proposed in Symphony of thought um but it'll be a while before us the rest of the world catches up to that that's fine um yeah so chat GPT is having a Thomas Edison moment yeah a light bulb moment um anyways uh you know what is what is everyone wants to know like what is this going to do what is it going to change so there are a few things that you know just off the top of my head something like chat GPT could do again once the once the cost comes out once people see how expensive it is uh to run you might think twice and kind of go back to Google when it's free right so there here's the thing is product Market fit is is one thing if it's free great but like the same thing happened with Dolly too when it when Dolly 2 was free the quality was like okay it's good enough you know I'll use it if it's free but as soon as they started charging um I saw I stopped using Dolly to uh except for on rare cases right it's like yeah you know it costs about 15 cents per generation so you know I'll buy a few tokens every now and then to you know make a YouTube thumbnail or something but I'm not going to sit there and play with it similarly if you have to spend you know a couple pennies per per conversation with chat GPT it might be worth worth it to you to have a subscription but what happens often when you don't have product Market fit and I don't think that this actually is solving any real problem it's Nifty right it's a cool tool but but it's like okay well are you going to pay for it and how often are you going to use it right because when you're talking about something that's actually going to be disruptive it has to be either so compelling that you need it and when you need it you absolutely need it even if you only need it once so an example of this is Zillow how often are you buying a house you know I bought my house eight years ago and yeah I look at Zillow every now and then but you know I found this house on Zillow and uh so I needed it once in eight years right so high need low frequency um so that's one way you can get product Market fit is is when they when your customers need you they really need you the other thing you can do do is have high volume but low utility so social media is an example of that where you know you use addiction mechanisms to get people or attention engineering to get people to use tick tock and Twitter and stuff all day every day so they're using it all the time but they're getting low utility out of it um and then then so that's if you want to appeal to the mass Market another thing you can do is you can do B2B right you can say I'm going to build something that's going to be useful to other customers um honestly I don't see GPT or chat GPT as having I probably have been seeing gptj the whole time sorry chat GPT I don't see that as having a mass appeal because here's the thing we're all used to using Google all day every day right we know how to use it we're familiar with this tool its results are more reliable we know how to interpret the results too here's the thing is chatgpt at at its present format it doesn't tell you how reliable it is it doesn't cite it's sources it doesn't say click on this link to to prove that I know what I'm talking about right you can't open a whole bunch of tabs of chat GPT and investigate each one so in terms of high volume not there doesn't exist and then in terms of high value but low frequency I mean you know it honestly it basically like for some of the functions it's just kind of regurgitating you know Wikipedia articles now that being said like if you ask it a basic boilerplate question it'll give you an answer can it do work though can it do cognitive labor um without that'll actually save you time and energy I haven't seen any evidence that it actually saves you too much time and energy now I have seen some people do like creative writing with it okay you know but as a novelist myself like you know getting barfing out a first draft anyone can do that cleaning it up on the other hand that's hard I figured that out um I know how to do that with with AI but I don't know that chat GPT can be an editor it can probably tell you how to in you know some general principles on how to improve your writing but your the point there is you're having to put just as much uh cognitive effort and time into it so it's not really saving you any time or energy so I don't see any of all the use cases and stuff like yes it's neat are you going to pay for it I don't think anyone's going to pay for it really um and I might eat my hat but certainly in its current state I think it's more of a novelty now that being said on the B2B side could this be useful to companies possibly especially given the the way that it formats it its answers if it can give you really empirical like you know hey we're going to hook into your your your your your databases right your um your payroll and your um and your customer service databases and you know sure okay I don't know like there might be there might be some B2B use cases um but I don't see I don't really see b2c use cases especially not at the price point of these models they're just too darn expensive to run so it's like okay well what what would be the price point what is the kind of conversation that is so valuable that it would be worth it so what are the most what are the most expensive conversations you have one of those is going to be with your lawyer right or your doctor because how much do you pay them you pay them 400 an hour essentially to see them to talk to them so you know if your time is that valuable and you can do most of it via conversation I could see you know those really high value consultation kind of things and obviously there's there's plenty of other kinds of consultation other than medicine and law but of course those are also very highly specific things where you have to know really know what you're talking about and be licensed right you have to you have to you have to pass your boards if you want to be a doctor and you have to pass the bar if you want to be a lawyer chat GPT can't do either of those things so at best it could be an adjutant or a or an assistant right but again human brains are still faster in in many respects so it's like okay you know those are the high value ones are they going to do like a a legal eagle version of chaby chat GPT are they going to do you know a uh a doctor or whatever his name was a famous doctor back in the day um you know version of chat GPT uh you know I I don't know especially because a lot of then then especially if you do medicine that that's considered a medical device and that means you need FDA approval right so there are these there are these use cases um science is another one where science is expensive but we saw the recent debacle with meta Galactica right where they trained it on a bunch of papers and it's still they still didn't figure out confabulation it started making stuff up so I don't think they're going to find PR uh product Market fit uh with their current format um so novelty sure breakout sure um it you know it it's gonna follow the hype cycle everyone is going to see chat GPT and everyone's gonna Pile in we're going to have a whole bunch of newcomers a whole bunch more venture capitalist investment a whole lot of speculation we're gonna get above a bubble and then it's gonna you know settle back down as reality sets in and um and we'll kind of go back to business as usual now that being said I am not calling for an AI winter my uh my fiance uh called me out because about a year and a half ago I was like man you know I'm using GPT and I've got about all I can get out of it so I think we're heading for another AI winter uh boy was I wrong um so she reminded me of that very helpfully um so no we are there are many many high value cases that we are figuring out how to use GPT with and I've been talking to a lot of Industry veterans from all all over I won't call anyone out by name but people that were there for the.com Revolution and crash um people that were there for the mobile Revolution which is still ongoing right uh 3G 4G 5G right Veterans of the industry and they say they all say so here's here's the thing here's what the experts are saying people that have been in technology um longer than me they are saying that this feels like the mid 90s of the internet where everything is just getting started we are just figuring out the potential of what we've got and it's like we know that there's something there but we only have the first inkling we don't even have the tip of the iceberg like it's the the icebergs on the horizon we can see that there's something out there but we don't have any idea how big it is we just know that it's big um what was it there was something that someone told me just like many conversations I've had actually where it's just like something changed in the last few months it was about the last I think probably since August or September like it just like Spidey senses tingling um and then of course chat GPT comes along and and explodes the popularity so again you know I I'm not even tepid like I'm I'm full steam ahead on this but chat chat GPT is not going to be the thing that changes everything it's just you know uh caught the public curiosity ah I think that's about all I got
Hmmm, ChatGPT sometimes asks me questions actually. It happens when you give it a text with a missing context. Specifically when it is not clear what it should do with the data that I input there.
This aged like milk, Yeh lol
Great video. Thanks. Do you have a reading list for someone looking to get into AI?, @David Shapiro ~ AI thanks!, <a href="https://youtu.be/s3I-OxAZXTQ">https://youtu.be/s3I-OxAZXTQ</a>
I’m working on an AGI scripting framework that uses the unofficial ChatGPT API, and I’m mind blowing and very satisfied by the results. I will definitely reach out to you, I’d love to get your perspective. It’s “scripting” in the way that actors read scripts to learn their characters, and can provide feedback on their own roles.
First question I asked was: &quot;Are you just a GPT-3 model?&quot; It said he wasn&#39;t, I don&#39;t even think he understood what I was asking...
Get your daily dose of Tau at <a href="https://www.youtube.com/channel/UCHmYERrMpeCc6mp9zK8UiTw">https://www.youtube.com/channel/UCHmYERrMpeCc6mp9zK8UiTw</a> <br><br>Here I send YTShorts daily about Tau. This is my 7th playlist now where titled Tau Tokenomics. <br><br>Great content David, hope you do more of Tau-related content. 🙂 <br><br>Cheers!
I could see a commercial use in marketing: churning out social media posts or blog posts. So basically creative writing in a business context., That&#39;s mostly what it seems like it&#39;s being used for. The question is economies of scale and also domain-specific quality.
Wow, I now tested a headline from CNN and asked ChatGPT: Is this text grammatically correct: &quot;World Cup champion France edges out England to reach semifinals as Harry Kane misses penalty.&quot; ? And it replied: &quot;Yes, the text you provided is grammatically correct. It uses proper nouns, verb tenses, and sentence structure to convey a clear and concise message. The sentence is easy to understand and does not contain any errors or inconsistencies. The use of the verb &quot;edges out&quot; is particularly effective in conveying the idea that France narrowly defeated England to advance to the semifinals, and the mention of Harry Kane&#39;s missed penalty adds an additional layer of detail and context to the story. Overall, the text is well-written and uses correct grammar.&quot;
I asked ChatGPT if it could ask me a question and it replied: &quot;Sure, here is a question for you: What is your favorite thing about learning new things?&quot; But it only followed up with one more question, and then became a bit repetitive. There seems to be a huge difference between GPT-3 and ChatGPT however. GPT-3 said that a jellybean will break a violin into four pieces, while ChatGPT explained why a jellybean can&#39;t break a violin in any ordinary circumstances.
Wow!
I’d say this is more of an AI spring. I remember looking into making a chatbot using the GPT-3 in the latter half of 2021, searching for tutorials, guides, etc, not finding anything, and resorting to Tensorflow. That didn’t work well for me at all due to the lack of spontaneity and difficulty creating curated training data, so I gave up. Then, about 2 months ago, my interest peaked again after I encountered a new GPT-3 based chatbot called CharacterAI. I once again searched for tutorials and was shocked by how much the GPT-3 had exploded in just a year.
The utility of GTP, even though it may not be completely accurate, is increasing for many users. I have been using the GTP playground for a while now to stimulate my thinking. It reminds me of the situation with Wikipedia, which is and was often dismissed as unreliable because anyone can edit it. However, even if that is still true to some extent, the usefulness of Wikipedia is too great to ignore.
What about making it sparse AI… hence, a little down the line, it should become much cheaper to run?
Last month, I created a chatbot according to your tutorial (hence why I’m subscribed to you), and it’s <i>far</i> better than ChatGPT to the point that I haven’t even tried the latter., which one?
I will say that using chatGPT for several real-world chats for me, it summarized technical articles and gave me working code where I&#39;d find it worth $0.20 / question. I&#39;m a github copilot user and a little bit in love with its usefulness for the data science I use it for. I don&#39;t think this conflicts with what you&#39;re saying about it not being a revolution. However, the pace of developments in models and the developer--&gt; public attention has been astounding to me so far. It seems likely to me to lead into an amazing hype cycle.
I agree that using chatgpt as a replacement for search is very expensive - you argue too expensive. However, I think it&#39;s important to consider the benefits that chatgpt provides in terms of the personalized and conversational experience it offers. While it may be costly, it can also provide valuable insights and improve the overall user&#39;s &quot;search for an answer&quot; experience. I think the fundamentals of search will change, perhaps tired - in terms of cost of access. So Google will still be used (free + more adverts), but a tool like chatgpt will offer more that just search as we&#39;ve seen and it will have a subscription model.
So happy you&#39;re still here! Been assuming/wondering about the secret project for a couple months now, looking forward to the reveal!
hey everyone David Shapiro here with another video today we're going to be doing chat GPT for writing fiction specifically writing long-form fiction like novels obviously this is a practice that takes a while so in this first video I'm hoping to just get to a solid outline and then we'll call it a day before we get started though I wanted to explain that uh I have monetized some of my YouTube videos uh people are getting a lot of value out of them um and I want to take the ads out but I was not able to get my uh patreon goals met so my goal is to get to four thousand dollars a month so that I can uh pay all the pail of my bills have health insurance that sort of thing um and then I can permanently demonetize all of my videos that is my goal now if people would prefer ads we can do that I don't really care one way or another but it would be good if uh if if we can do it this way all right so without further ado let's talk about writing stories so this is something that I have spent a tremendous amount of time on uh personally um writing is uh is one of my top hobbies and it's how I met my now fiance and so uh in fact one of the things that attracted a lot of people to my channel was using these language models to write fiction and uh in point of fact chat GPT completely changed some of my plans around my auto Muse Suite of tools um so because of that like I'm not going to try and compete with chat GPT why bother right I'm gonna do something else in the meantime let me show you how to think about this uh tool so that you can write fiction long form fiction anyone can do it um but you kind of you got you got to know about like writing a story so um I have the advantage of I had a premise already in mind so the premise that I had in mind was that I wanted to write a story from the POV of a courtesan during the Golden Age of Rome and I wanted it to be like a palace Intrigue kind of story but also the Golden Age of Rome or the the height of the Roman power there's a huge amount of stories that can be told in that you know there's um uh there's you know whether it's Augustus or Caesar and Gaul or whatever is going on in Egypt or you know the Revolt of the Jews or whatever there's all kinds of cool stuff so um the way that I like to start these kinds of conversations is you just tell it what you're doing so you say hey I'm writing a novel um I need your help I found that if you're polite it like it it it copies your tone so if you just if you're very thorough and polite it tends to be more helpful um if you just give it like flat instructions it's usually willing to help as well but I mean there's no nothing wrong with being polite to the machine um it also feels good to be polite um I'm writing a novel and I need your help um this is a uh Palace Intrigue slash uh historical fiction slash Adventure uh set during uh the Golden Age of the Roman Empire it follows um uh a courtesan um so help me brainstorm this story the goal today is to get a very solid outline so if you tell it what your goal is um it's like sure let's do that um what questions do you have so if you if you tell it if you if you finalize it with saying that ask me questions if this is where the power of a dialogue comes in because you might be making assumptions in your mind that are not going to translate um so let's see it says sure I'm happy to help brainstorm ideas for your novel to get started can you give me some more information about the Curtis courtesan and her role in the story what is her goal or motivation do you have any specific historical events or figures that you would like to incorporate these are phenomenal questions okay um great questions and I'm saying this as a I'm not a I guess I am a semi-pro writer I have made some money from writing fiction um but uh I'm also just a very passionate writer and I uh a good like more than half of my friends are writers that's like my social life um the other half is all AI people um okay great questions so um well I really wanted to have a close visceral experience of um the Roman world uh during its height um for instance I really love the Intrigue of Assassin's Creed Odyssey um I also really loved the book cushiel's Dart and if you know what cushel's DART is you know that it's a very spicy story we're not going to go that spicy at least Well Chad GPT will get mad at me if I try and go that spicy We're not gonna we're gonna toe them no I'm not gonna get demonetized over that um so we'll just leave it at that um so that's the kind of story I want to tell um I'll need help uh fleshing out um the characters so for for some reference the reason that I have some of this knowledge already is because I was using my own tools Auto Muse tools while I was brainstorming this those are still being built but they've been kind of put on the back burner we'll get to that later but the tools are coming um I'll need help fleshing out the characters um backstories motivations goals um now in terms of um uh historical events or characters I was thinking maybe um Augustus or Hadrian um or some of the uh five good Emperors uh basically I want a setting that is um that allows for anything to go celebrations opulence battles piracy Etc make sense all right so let's see what chat GPT says yes I understand what you're looking for it sounds like you want to create a sweeping epic story set during the height of the Roman Empire with a courtesan as the main character to help flesh out the story and characters we could start by brainstorming some ideas for the courtesan's background and motivations for example what was her life like before she became a courtesan what letter did that profession and what are her goals within it does she have any personal or professional relationships um let's see in terms of historical events the Reigns of Augustus and Hadrian would provide a rich backdrop as they were both significant figures and the rule coincided with a time of prosperity and expansion including events um could add to the historical reality okay cool is there anything specific you would like to include um okay great so um first let's brainstorm the MC a shorthand for main character um uh during another brainstorming session one idea was that she was a slave seeking Liberation uh but I'm not sure if that works um another idea was that she was from um a lower family um maybe uh I guess maybe a patrician um maybe a poor Patrician family I don't remember if the attritions were still around um so uh maybe let's brainstorm some around that can you uh give me some more ideas so one thing with chat GPT is you have to be very specific about what you ask for um whereas interacting with a human there's a lot of uh based on my recent interview with Anna Bernstein um which she pointed out that like with a human there's there's an implied intent um what with the machine you have to be very explicit um okay perhaps she was born into a poor Patrician family and was struggling to make ends meet she was forced to leave home at a young age make her way in the world she may have turned to courtesan work to earn a living um then her goal may be Beauty charm gained powerful uh gain the favor of powerful men and women alternatively she should have grown up she could have grown up in a wealthy family but fallen on hard times due to financial mismanagement political scheming oh that's good that's some good Intrigue oh yeah I like that and so here's where like just asking it to give me ideas and then I pay attention to what resonates with me and so then it follows my lead another possibility was orphaned and had to make her way on the streets that's not quite as interesting um you know like a Rags to Riches thing but since I want to do a palace Intrigue this this middle one looks good um yeah I like that middle one um of somewhat Fallen fortunes that's very consistent with um Palace Intrigue uh and also sets her up with a strong motivation for revenge and Justice and getting back to her um uh what's the word uh desire not desired um deserved station um okay so that's a powerful through line let's now think about the theme um of this story the universal truth can you help me Workshop that now um so I want to pause and and point out that like okay how did how do how do you as the person following along know what to do so I am following uh kind of my personal pattern of when I'm coming up with a new story so you start with a premise so premise is usually just the setting and the main character um that's like oh I've got this great idea for a story so setting in main character then you develop that main character and explore their interaction with the setting because there's three pillars of writing fiction and that is plot setting and character we'll get the plot in a little while hopefully that's where we'll end today but setting in character so character is the human um and then the setting is the world that they're in and um they're those are intrinsically intertwined um and then the the thing that guides the whole story the purpose of the story is like the beacon on the hill or the lighthouse that you're heading towards that is theme and so theme is the universal truth the universal um lesson that is in it and so in uh popular fiction lately um you might have remembered this uh it was vaguely popular uh movie called Avengers end game where this uh big purple dude named Thanos wants to erase half of Life uh in the universe so the theme there is the arbitrariness of catastrophe right storms will just kill at random plagues will kill at random volcanoes will kill at random Thanos represents that uh the cataclysm right he is the death blight of of um of just our existence um and so let's see what it says so because we have to figure out the theme um like why is it that I wanted to set this at the Golden Age of Rome and of these somewhat Fallen fortunes so let's see if we can figure out let's see if we can pull a theme out that is going to be the guiding force or The Guiding energy of this story um okay so to identify the theme it might be helpful to think about the central message or lesson you want your readers to take well duh that's what I'm asking um what is the main character struggling with and how does she overcome what does the main character learn about herself in the world uh as the story progresses what is the story saying about human nature or The Human Condition answering these questions can help you identify the underlying ideas or principles okay sure for example the main character is struggling with the challenges of living in a rigid hierarchical Society oh here we go we're getting into the good stuff here ultimately find able to find her place in it the theme of the story might be something like finding one owns one's own path in life or the power of self-determination on the other hand if the main character is struggling with issues of identity or belonging um the theme might be something like the search for one's true self or the importance of community okay so what this isn't what what this is implying here is that um a character's Arc their personal journey is is what exemplifies and carries the theme and so you can look you can often look at them as this is not always true but one of the easiest ways to do theme is what is it that they learn what is it that they overcome along the way now you can have a theme that does not directly impact a character's Arc and that and that the theme is actually in the setting so a good example of that is that one of the one of the the overarching themes in The Witcher um the Netflix show is one of nihilism and that is that everything is pointless and this is and that is put in the uh the world the setting the where the setting punishes people for having hope because that is the opposite of of nihilism right um and so geralt then is a reflection of the the world that he lives in but he kind of has a flat character Arc because he's just I know what the world is like and it's an awful place and I'm not gonna fight it um but his personal truth is that you can't fight Destiny um so that's you can have multiple themes okay great um so it did it it's it's framing it we're getting there but let's let's do this a little bit more um because it it one of the things that I don't like about chat GPT is that sometimes it'll just kind of tell you how to approach a problem rather than solving the problem now this is okay because sometimes just knowing what to do is is important but um sometimes I'm just like no just solve the problem for me um so I'll say let's think through this um let's think through this um uh let's see the setting is the um is the apogee of the Roman Empire what does that tell people there's an abundance of wealth and power and possibility um so that feels important and then um the MC uh has lost her station but wants it back um so again what does this um say about um Humanity or Universal truths um can you help me articulate this theme let's see what it says the setting of your story during the height of the Roman Empire suggests a time of great prosperity and opportunity but also time of strict social hierarchy and rigid rules of behavior this could be seen as a metaphor for the larger Human Experience where opportunities and limitations are often determined by factors such as birth wealth and status and gender the main characters struggle to regain her former status in place in society could be seen as a reflection of the human desire for recognition and respect ooh okay and the lengths that people will go to in order to achieve their goals this could be lead to a theme the pursuit of power and success or the cost of ambition alternatively the main character's Journey could be seen as a commentary on the capriciousness of Fate ouch um so capriciousness is um Mercurial emotions that's what capricious means is when when um something is very fickle um so the capriciousness of fate and the dangers of relying on external factors for one sense of worth and identity in this case the theme could be something like the fragility of success or the importance of self-acceptance I hope these ideas give you inspiration so this is actually zeroing in on what we had worked on before before Chad gbt um so let's do both actually let's have her pursue power um and status and then um when she gets it she realizes that she's alone hunted and uh and empty um so then she must make a choice um to let it all go and pursue uh genuine human connection I think this will parallel um today's Zeitgeist of uh quiet quitting and um and reclaiming our lives um what do you think okay it says I think that could be a powerful and resonant theme for your story having the main character initially pursue power and Status only to eventually realize the emptiness and isolation that comes with it could be a compelling Arc that reflects the idea of the cost of ambition this could be a great way to explore the theme of the importance of genuine human connection and the dangers of relying on external markers success for one sense of worth and fulfillment to further develop oh now it's suggesting stuff to further develop this theme you might consider exploring the main character's Journey from striving for external validation to finding a sense of purpose and meaning within herself this could involve showing her struggles with feelings of inadequacy or loneliness as well as moments of self-discovery and connection that help her to see the value of her own worth and relationships yes this is great I think we need to uh break this into a duology the first book will be called Ascension or something uh grandiose like that and the second book could be called um descent or something um and so I really like this and I'm totally gonna finish this and you guys are gonna go on this journey with me as however long it takes because this is a story I've wanted to write for a while but it's like I don't even know where to begin um because uh it's also because it's a historical fiction there's a lot of research that you need to do in order to keep it accurate um so like you know we haven't even explored dates and times of like you know uh Augustus or Marcus Aurelius or Hadrian um so we'll figure that out in a minute because we we do have to flesh out the setting um okay that's all nice and solid now um Let's uh let's summarize everything we've got up to this point please write a detailed uh summary of The Story So Far So you want to do this every now and then because especially when you're doing long form fiction um it it does chat GPT does have a limited memory so you want to you want to kind of like okay let's stop let's circle the wagons let's figure out exactly what's going on and see it got one thing wrong where it's um it says poor Patrician family um so we actually want to say that she started as a with a as a in a powerful family um through her beauty charm and wit the main character is able to gain the favor of powerful men and women and she Rises to position of influence however she becomes more successful she becomes she begins to realize the emptiness and isolation um yep she must ultimately make a choice between her ambition and genuine human connection um the first book titled Ascension follows the character's rise to power um the second book descent explores her journey from striving for external validation to finding a sense of purpose and meaning with herself as she grapples with the choices she has made and the consequences from them excellent um we just need to change one detail um she starts um or she was born to a powerful and so this is where I thought something and said something but didn't tell it um she was born to a powerful and Wealthy family uh but uh her family was um dispossessed uh due to uh betrayal and Palace intrigue can you update the summary based on this info excuse me um certainly here's an updated summary and so by getting it to restate it we can then see how the machine understands the story and this is the same thing as as we what we do with humans um which is called reflective listening so um I am asking it to reflect back to me what it understands that I want so using reflective listening with chat GPT is a really powerful technique I'm going to do a quick time check because um oh wow we're already 22 minutes in um I want to keep these videos reasonable um so we're we're we've got like the main through line we've got a little bit of the setting um so let's try and get a little bit further with the setting and then let's start trying to outline a plot um great let's flesh out the setting now um can you help me figure out when exactly to set this story when do you think um the Roman Empire hit its absolute Peak raging okay trajan was an effective ruler uh and popular ruler and he ruled for almost 20 years okay so that's a that's a good period of time um so Augustus was earlier Hadrian came after oh let's do hatred let's do trajan to Hadrian um so her life takes place during the entirety um of these guys okay great um uh let's set this so that uh her birth coincides with the um Ascension of trajan um so uh so she's partly uh marked the Romans were very superstitious um as being tied to the fate of the emperor um so then so then her life goes um uh so then her her character Arc follows Rome through its absolute Peak and the beginning of the decline so her choices mirror that of Rome okay with this in mind can you list all the major events and people and uh well I get yeah events and people that we should think about from let's see 98 A.D to 138. 80. um this will help us figure out what uh beats later okay certainly here are some of the major events and people from the period of 98 80 to 30 138.80 um so this is her 40-year life basically um and then we can have because you know during Roman times Roman matrons tended to live a lot longer than Roman men mostly due to uh conflict but not always because you also had much shorter life expectancies of of women in general due to death during childbirth but once a woman made it past childbearing years she was their life expectancy was pretty similar to that of today with the exception of like things like plagues happening um okay antonine wall the bar Cobo Revolt the Jewish Rebellion Ayah there's that one take takes place in Palestine um because one of the things that I thought of is if she's a if she's a courtesan then maybe she's attached to very powerful people and goes all over the Roman World which could be really interesting um yep okay so this is this is the the Roman parthian War so that's way out east um which is not yet now I Ran So there were there's lots of uh horses and stuff so because cushiel's Legacy is one of my favorite stories I'm thinking about like what if she goes on these Adventures to like save the Empire or something um and and this is how she gets powerful um and then she starts touring with with Hadrian or something um I really like this actually um okay so so these are just a few examples um let's focus on her earlier life for a minute um she would have been about 12 during 110 whoops 110 A.D um what happened around then um anything political or economic that we can base the fall of her family on okay so if it says like I don't know what's happening in your character's life um okay so it's confused I'm I'm actually asking for um real real events so I'm going to do a quick refresh this is a trick that I have you can interrupt it with a with a solid refresh and then um and then you can go and regenerate it I'm like no no no let's not do that um okay what was happening in Rome around 110 80. um I am mostly curious about political and economic events um in and around the Imperial City foreign if you're watching please add a cancel button because you you don't want the conversation to be polluted in fact if we could go back and delete the bad messages um and and clean it up as we go so if we can flag something to say remove this from the conversation that would be fantastic okay During the period around one uh was enjoying a time of relative peace and prosperity trajan was an effective and popular ruler there was likely a great deal of political and economical activity um let's see many public works okay the expansion of the city's Aqueduct system um oh okay sure let's go with that sure let's go with that um let's say the MC's um family was involved in um these Major Works they were sponsoring and um and organizing um big projects but uh wealthier um wealthier or more well-connected family wanted in because they were very um a very lucrative and high highly coveted projects um so let's imagine that um her family's Fall From Grace was um was part of that scheme um let's go ahead and name these families and people can you recommend some uh clan names and uh given names for our characters um we need at least the MC her family name um and then the Rival faction name well certainly here are a few suggestions for clan names and given names you might use plant names Claudius this was a common clan name in agent Rome and it was associated with several influential families flavius this was another one um Aurelius so the claudians and the flavians these are uh I don't want to be necessarily that on the nose um and I'm not going to name her her Marcus um we need we need a female's names but uh Valerius or Aurelius or I mean yeah these are these are good um okay what about uh women's names remember the um the MC is a courtesan um so one thing that I that I noticed is that you have to sometimes remind chat GPT because it's like it'll kind of start going down its own rabbit hole and you have to pull it back and say no this is what we're doing um Let's uh let's get some women's names Julia so my fiance pointed out that um one thing that a lot of Roman homes did was that they just gave the women the feminized version so you see Julia Flavia Aurelia um so they were very super uncreative when it came to uh women's names um so we could do something like Aurelia Aurelius or Aurelia flavius I don't know um uh let's see so uh female names were often um just feminine uh versions of house names are there any other kinds of names we can choose something more distinctive Maybe foreign now this is something that I happen to know already so if you didn't know this you could have asked that question like why are why are the names weird um let's see symprio simpronia this was an uncommon given name um associated with uh sister of Cato the younger interesting tortulo sytheris or kytheris I don't know if that would be cythrus or kytheris um I want something that's not too difficult to say so Cornelia mother of the graci brothers so let's say like Cornelia flavius or something or Valerius Cornelia Valerius um would Cornelia Valerius be a legit name anything wrong with that father or husband okay in the genitive case and it would be uh valyria so I guess her name would be Cornelia valyria or Cornelia of the valyria Okay blah blah so it's telling me what's wrong with it for example it's common for Roman names to be comprised of a given name a clan name and sometimes a cognomen a nickname or additional surname that's right um um so with the given name coming first followed by the clan name followed by the cognomen good so I didn't even have to ask it realized that I was struggling um okay help me understand um what the genitive genitive case means um the Leary eye looks plural I'm lost in Latin the genitive cases used to indicate possession or ownership in English we use uh the possessive form of a noun um genitive cases often indicated by so it knows a lot about Latin um I've seen a lot of people so while it's writing I've seen a lot of people talking about how large language models don't have mental models of the world and I completely disagree this is explaining Latin to me and it not only that it intuited that I needed to learn Latin so please explain to me how it's doing this without internal models you can't trust me I have a bookcase full of Neuroscience over there okay let me do a quick time check we are at 35 minutes we'll try and wrap this up pretty quick um okay great please summarize everything up to this point um uh everything let's say uh everything about the story up to this point um you can ignore the sidebar conversations um let's go ahead and um so for example um so geniform is is for first first and second declension nouns like Valerius I don't even um I'm not going to learn that much about Latin right now but I guess I'm going to um okay here it finally explained it Cornelia valyria um so Cornelia belonging to the Valerius Clan got it let's go with this uh for now um Cornelia valyria um uh will figure out her cognomin later um yeah please go ahead and summarize our story the story is set during the height of the Roman Empire a time of great prosperity and opportunity but also a time of strict social hierarchy and rigid rules of behavior the main character is Cornelia Cornelia valyria a courtesan who was born into the powerful and Wealthy Patrician family but has fallen on hard times due to betrayal and Palace Intrigue determined to regain her family's former status and protect them from further harm she uh turns to courtesan work as a way to earn her living and support herself she climbs a social ladder through her beauty charm and witch he's able to um blah blah blah she ultimately must make a choice that takes place during the Roman or Emperor trajan she navigates the complex World okay um this is great please add the details we discussed about the fall of her family specifically that specifically they were participating in the grand public works such as the aqueducts um and that uh betrayal was over that please incorporate that into your summary and add the date as well 110 8 A.D when Cornelia was 12. um so again by asking it to restate things it kind of says okay the document is written here um and it's I confused it again um by because it says oh it's it's set in in one 1080. um uh let's see and um the story is set um a few years later as she comes into adulthood um here's an updated summary of the story the story is set in 11080 okay whatever I'm not going to fight it what I mean is like okay I could probably reword this but like this is good enough um there we go yep there we go okay I really like this as a as a backstory because it gives her a very very complex set of motivations and um let's say baggage to work through um but yeah I think we'll call this a day I'm super pleased with this um and we will uh we will be back because I love doing this um and I hope you guys enjoyed it so thanks for watching
I love what you&#39;re doing here. I on the other hand have a much better way of getting a lot of this done with chat gbt but I did draw some inspiration on how I can use my way better based on some of the things you showed so I appreciate that. <br><br>One of the things I think you could do and at the beginning instead of telling Chad GBT that it&#39;s your partner tell it that it is a world historian novelist with knowledge in Greek mythology and all that stuff Socrates etc... and to use this knowledge based on those things to guide you. That way its not just a writing partner trying to figure out ideas with you it is a partner who has knowledge of what you&#39;re writing about. <br><br>Just my thoughts.<br><br>Thanks
Do continue this series. Great to know about new prompts and how actually discuss stuff with chatgpt correctly.😀✌🏻
You know that I have tried to ask ChatAI to write the original character of an immortal goddess named Goddess of the Void. Her name Nara. The AI give her name. I ask the ChatAI to help me describe the character original personality as best as possible. It even describe her appearance and her own pov based on the character personality! I feel like I have ask the AI create a character. I don’t know why it is possible. Can anyone tell me how?
This is fantastic. Just discovered ChatGPT and have had a fun couple of days interacting with it and experiencing some of its capabilities and limitations and how our interaction and prompting is critical.<br><br>This video is so inspirational and I already have SO MANY ideas for assistance from Chay GPT with my many many unfinished fiction works that I get stuck on.
So I was using ChatGPT today as a writing assistant bouncing ideas back and forth and building up an outline for a trilogy of books. I was absolutely blown away and felt almost like I was working with a real assistant and made some really great progress. Then for some reason everything just fell apart like my assistant had an aneurysm or something and nothing worked right anymore. It kept forgetting things, redefining things we had already defined, and then it got worse where the output started being put into code blocks for some weird reason. When I tried to re-feed it the outline we had worked out earlier in order to maybe make it remember and continue from there it just got all confused and basically the whole chat fell apart.<br>Its almost like the longer duration the chat ( was in the for hours) the more likely it is to fall apart?<br>Would like to see techniques for picking up in a new session somehow. Might see that in your next video so will definitely check that out of course.
Can ChatGPT even write about a courtesan anymore? The &#39;safety feature&#39; is getting <i>so bad</i>, I haven&#39;t run into any problems. You just have to navigate the spiciness carefully
Just wanted to let you know that I really like your approach of interacting with chatgpt...designing your chats in such a way that it basically resembles a real conversation makes this experience so much more enjoyable!
Really helpful video for a novice ChatGPT user like me who thinks about using it for this purpose.
David, you would do well to implement two strategies. 1. use separate chats to explore topics that are not necessarily dependent on the larger context. For example, what was happening in 110 AD in Romes more of a general question that can be explord outside of the main context. You can then bring only the details, summations and conclusions back into the main context window, since it is unfortunately so limited now, you risk losing the thread if you really go off on a tangent. You ould even ask chat GPT to write the prompt to explore 110 AD in the context of your characters life if you really want to keep that context in mind while exploring. 2. consider a strategy to use chat GPT&#39;S thread feature which happens when you edit and resubmit a prompt. To the left of the prompt there are those arrows, so you can in fact have many sub threads in a single conversation. A good mix of side threads (diff conversations) and sub-threads and an external note tool like Notion, to organize everything could be just the ticket., on a similar topic, I have been trying to figure out what the size of the context window will be for GPT 4. It is widely thought to be 8,000 tokens for chat GPT. Thats around 32,000 characters. There are 70,000 - 120,000 words in an average novel, 4.7 characters in the average word. Thats about the size of one token, so in order to write a full short novel we need at least a context window of 70,000 tokens. Something tells me that GPT4 isnt going to jump 10x in the size of the context window. Until then we will have to rely on strategies to keep chat/GPT in context. I&#39;d love to hear your thoughts on that.
Well thought out and well presented....outstanding!
This is a great concept. I’m gonna use it to write the stories I’ve had in my head as well. Nothing grand like making a full novel but just brainstorm ideas
Currently doing this right now
This is the best video I&#39;ve seen on using ChatGPT as a writing assistant. Thank you. I look forward to more help from David&#39;s videos.
Watching your video has given me some ideas on how to approach chat GPT myself for helping develop ideas. Thanks! I&#39;ll be looking forward to the next videos
Great video! I&#39;m looking forward to more! I&#39;m about to jump into the 2nd part.
Aren&#39;t you afraid that someone who has access to ChatGTP steals your book?
If you use the term &quot;Please expand upon...&quot; chat will provide a more detailed summary or explanation or fuller story. If chat stops in mid-writing, just use the term, &quot;continue&quot; and Chat will continue from where it trailed off or stopped. I seem to get more and better responses when using politeness while running inquiries. I think this has to do with Chat&#39;s &quot;fine-tuning&quot; which it thoroughly explained to me. Chat has helped me get my thoughts out, and I&#39;m starting to finish my many writing projects. The one thing I did realize is that Chat will not solve your problems independently, but it will give good guidance with various paths to utilize toward your process. I like how you ask questions relative to your inquiry which is, in a way, a command but yet a probing request to fill in the blank of what you are seeking. - That&#39;s the best way I can explain how I see the communication between you and Chat.
Wow! At the least this is an all in one stop place, for researching while you’re writing rather than separating the two phases. At the most… I can’t even, I’m just really excited. And that is awesome that this is a story you’ve wanted to write for a while. This could be a fantastic book. With the onset of Rome’s fall coinciding with MCs fall as a macroscopic thing, there is much room for modern, relatable drama, and for a great final lesson of being happy with a clean conscience and with what you have. All the pieces fit to make a larger than life tale.
<a href="https://www.youtube.com/watch?v=QotF4TgnDhU&amp;t=23m46s">23:46</a> gold
<a href="https://www.youtube.com/watch?v=QotF4TgnDhU&amp;t=9m18s">9:18</a> my friend. Thats what its all about. Thats why I wanted to try this. We can come to these brainstorming thoughts on our own but it takes a while haha! It can takes days and weeks to flesh out all the possibilities and to realize the coolest direction to go. That chatbot just fleshed out every main possibility given the parameters it had to use. Amazing!!!
You’re one of those subtle geniuses in the background who are prophesied like any other great door-opener/teacher archetype, not known by many but known by many who spread evolving thoughts who are also background yet necessary incarnations to every era, and I’m grateful to have found you. <br><br>I for one am not, not against this chatGBT and novel writing. I currently have a novel in progress and an account with chatGBT, and have been trying to think of ways to apply them together in a true and good way. <br><br>Thanks for this vid!!!
Hi David, I subscribed quite a while ago and pop over every so often to see what you&#39;re up to. I was very interested in your python work early on but didn&#39;t really have the time to get into it. Since then I managed to get onto OpenAI beta after a few months and things have moved so quickly since then it seems. My particular interest in your work is the fiction writing assistance from AI, particularly where it helps generate tantalising new ideas I&#39;d not considered. As you&#39;ll know, a huge challenge is how to engage the AI (OpenAI in particular) to build up the story, especially where it is unable to retain a lot of previous knowledge. I know this was a major theme of your python work. More recently I&#39;ve been trying chatGPT and it seems great at retaining the context but I&#39;ve often found the output to be quite generic (compared to openai playground) or wanting to reach a conclusion. With temp settings, etc I found openai to be more imaginative.<br><br>So, today I popped over to your channel and saw your two videos about using chatGPT and I&#39;m thrilled your covering this. I think that one of the main problems is knowing how to speak with the AI , but I&#39;m also keen to see how you keep chatGPT aware of context in a story. I guess you still need to summarise, provide overall context, and lead in to the next chapter.<br><br>So thank you so much for getting into the weeds of this process and I&#39;ve been inspired to support you via Patreon. Btw, I love your long term goals. Star Trek (without all the fighting) really should be humanities vision. The trouble is, I think our DNA is wired to be tribal (supporting teams, religion, politics, nations, etc) and that to me is the biggest hurdle to overcome.  But we have to start somewhere.<br><br>I know your interests are many within AI but my hope is that you&#39;ll spend a good amount of time on how we can use AI for fiction writing - hopefully without the need to get involved in coding.
This is an amazing video! I&#39;ve always wanted to write a novel, but I always realized my limited ability to do so and never really began. With a tool like ChatGPT, I can finally write a story that I myself would really love to read. I have absolutely no literary background and even English itself is not my native language, so ChatGPT is a godsend for a person like me.<br><br>Your video is a great starting point! I really hope that you&#39;ll continue making these videos in the future., Next episode comes out tomorrow
<a href="https://www.youtube.com/watch?v=QotF4TgnDhU&amp;t=2m54s">2:54</a> I think u read the robot uprising memes. “just in case” 😂😂
I&#39;d like filtered adds, adds that you approve, rather than endless scams, As I recall, It&#39;s called &quot;Google Ad Center&quot; that customizes the profile as Google detects, and you can change or block ad categories.<br>It&#39;s not perfect but I hope this helps., The ads you see are TARGETED for the things you watch and Google.<br> If you want them changed you need to find the Google settings for it. Blame yourself lol, Not sure what you mean? I don&#39;t know how to choose ads
Too early for quiet quitting to be part of ChatGPT’s knowledge base. ;)
As an aspiring author myself, I don&#39;t think we need to be worried about Als stealing our jobs because you need a real person to overhaul most of what the Al generates. Not only that, but you need a good idea from a creative person. You can’t just say, “Write a novel about x.” and get a good result. You’ll get <i>a</i> result, but not something worthy of publication., @MrAdonis258 I enjoy using it as a writing assistant, not a ghost writer., I tried it. For the most part AI writing feels soulless. There&#39;s no zing it. But imo it does offers a more details structure to the story for those experiencing writer block. GRRM probably should use it., @Ells But the tractors didn’t replace the farmers; they just decreased the amount of manual labor required., @John Vanderbeck Exactly., Whether it is in the arena of writing, or visual art, AI is just another tool just like any other. Tools evolve and get better but they are still tools.
I am so very much looking forward to this, having recently discovered your post on the OpenAI Community Forums. This is exactly the type of use case I want to see in action for a community project we’ve been trying to get moving for over a year. You may be the catalyst required.
look forward to this!. thanks david for taking the time to teach about this.
hey everybody David Shapiro here with episode two of the AI startup crash course video series episode 2 is called experiment experiment experiment so as usual I will start off this video with a book recommendation uh crossing the Chasm by Jeff Moore or Jeffrey Moore I don't know which he prefers to go by um I have mentioned several times that product Market fit is uh your new god um read this book uh it is relevant to experimentation but we will also take a deeper dive into product Market fit uh and a couple more videos so uh yeah this graphic learn it Master it understand it and it will save your life experiment with everything uh when I say everything I mean everything uh get it get in the crash test dummy mindset that's why I picked this picture okay so what are some things that I mean when I say experiment with everything uh start by experimenting with teams and meetings try different teams try different startups try different groups of Founders try different uh methods whether it's consulting or building software or joining a team talk to people ask what works experiment with everything nothing is sacred that is the mentality that you need the crash test dummy mindset and nothing is sacred so for instance when my startup just got going we we had to experiment with meeting formats and so like we tried one meeting format and uh the most critical thing was to ask for pluses and Deltas at the end of every meeting what's going well so that we can keep it and what needs to change pluses and Deltas pluses and Deltas change change change change experiment and so after uh it was probably three or four different iterations we settled on a meeting format that worked and then Army meetings kept getting longer and longer because we had more and more to talk about and so he said okay let's try something else new let's split it into two meetings so we'll have one meeting at one point focus on one thing and another meeting and another point to focus on another thing so yeah experiment with everything people teams meetings nothing is sacred keep trying keep experimenting step two it's not an experiment if you don't measure something measure everything measure time money effort customer response use a b testing if you don't know what that is look it up figure out the Telemetry and kpi that you need what is a kpi key performance indicator an example of kpi could be user engagement time on site money spent conversion rate all that sort of stuff Metrology is an an entire discipline in in and of itself um yeah you got to measure you gotta have you got to check results and you will find through experimentation what gets the most results now that being said you might be measuring the wrong things that's why I say Metrology is an entire discipline unto itself social media you're watching my YouTube channel because I found the proper social media fit for me YouTube is my is it worked for me now there's all kinds of other social media presents if you want to establish a pla a platform and build a presence you'll need to experiment with formats the format of my YouTube channel that you're watching now is not how it started I experimented with content with video formats and and and everything and the the kpi that I that I cared most about was subscriber growth uh view count and hours watched um so those are what I optimized for and it worked and now my channel is taking off exponentially um so you do that you'll establish a presence you'll attract the attention that you need for your AI startup so that you can hire the right people so that you can attract the right investment and so that you can get customers right so those are the three things that you're trying to attract with social media you have a really powerful tool to go viral there is someone out there searching for what you are building and what you are offering and doing a social media presence platform is the way to get them UI ux if you're not familiar with these terms UI means user interface so how the website is laid out how your app is is laid out and ux is user experience these are both highly specialized disciplines especially user experience um do remember use a b testing find out why your your users like something and why they don't ask them say what is frustrating about this what's bad about it what is what you know what kind of experience are you looking for one thing that you'll learn is that users don't necessarily know what they need or want but their behavior is the best signal their behavior will change when you hit that sweet spot for UI and ux um products markets and avenues things change chat GPT came out and completely derailed one of one of the entire uh uh product lines that I was working on and I said dang so you know what try something else um different approaches different markets um if there's something that is not being served by you know a competing product or or existing service don't compete with the big guys find another way find something that isn't being served so look at look take a new approach look at it through a new angle look at it through a new lens uh and finally a second book recommendation is the Lean Startup by Eric Reese um there is some good information in here I'm definitely glad I read it um however what I will say is it has gotten some criticism which I understand I have my own criticisms I'm not going to get lost in the weeds in that um but it does have some good ideas uh and but what I also what I will say is as someone who's been in technology for 15 years especially in the I.T and software space a lot of this is already baked into the industry um so as as a as a technology Insider I get it however if you don't have 15 years of it or software experience you need to read this because you do not understand the tech industry um so that is primarily who it's for is if you're new to technology if you're new to um to Ai and stuff you'll definitely need this so that you get oriented to the tech industry and to Tech startups specifically but again take it with a grain of salt my recommendation is never read just one book my personal rule of thumb is whenever I'm trying to learn something I buy at least two books from two very different authors so that I get I look at the same problem from two different completely different angles so there you have it that is episode two of the AI startup crash course experiment experiment experiment thanks for watching
Support me on Patreon so that I can do this full time and keep the content ad free! <a href="https://www.patreon.com/daveshap">https://www.patreon.com/daveshap</a>
I am looking for a co founder I’f anyone interested, Post a link to your LinkedIn and a brief snippet of what you&#39;re doing!
morning everybody David Shapiro here um and by popular demand we are working on a basically a chat GPT clone but with long-term memory and eventually external sources um but first uh quick uh um uh housekeeping one um as I always say the offer is still on the table if I can meet my financial goals with patreon I will remove ads forever um so do that also uh the comments section of my YouTube channel is blowing up so I can't keep up with comments if you really really want my attention patreon is the best way to get in touch with me obviously if you spend even just a couple dollars that tells me that you're a little bit more serious the second best way to get in touch with me is LinkedIn uh links to both of those are in the video description and then third is I've got a um a new mailing list that I'm creating which is also a good way to follow up in the future now that all being said I am super busy uh but stay tuned for some news tomorrow okay with all that out of the way uh let's go over I started this um so I'm going to try a new format um I have there's basically three kinds of videos I do one is an explainer which is you know the slide deck and I just explained something to you teach you it's basically like a miniature lesson um then there's the experiments like my uh chat GPT and case law one that was an experiment it didn't really succeed but the point wasn't to succeed the point was to learn and then there's tutorials which is I know how to do something and I'm going to walk you through to the end so this is a tutorial so just wanted to set expectations um okay so this tutorial is how to create a really brain dead simple chat bot um and how to give it long-term memory so let me just show you what I've got working so far so the user hello and waits a second uh are you there obviously we're not getting any response okay so what gives well we are just getting started so let me show you what that produced so let's go to chat logs and you see here we've got log and then a time stamp and then user so we know uh basically that tells you a lot about what happened so you know it's a log you know the timestamp and you know who said it but let's see what's in here so we've got the message we've got the speaker we've got the time stamp we've got a uuid and then we've got a vector so the vector is an embedding that was provided by open ai's latest um latest model which is what is it it's like text 802 or something I've got a function right up here yep so text embedding ada02 so this is their latest one it's supposed to be like fantastic and it replaces a whole bunch of other stuff and it performs really well okay great um so basically I just have this really short little function that you just pass it a piece of text and it'll send back an embedding and the reason that we do this is because we need to search it now this is probably not the most efficient way to do this in the long run um once you have longer conversations you'll probably chunk them but for now we're going to do this and so every message becomes 42 kilobytes which adds up because this vector is uh 1500 values long and it's used to register that so it's not necessarily the best but once you get to longer messages it makes more sense and especially if you if you chunk them um which we may or may not get to in this one we'll see and so what I mean by chunking is rather than saving each individual message and then vectorizing that um you take like a chunk of like five messages or ten messages and then you chunk and vectorize them because the chances of getting exactly what you want in one message is pretty low but this isn't easy this is low hanging fruit this is just where we're going to start okay so that's that so that's what's happening there but let's go to the main Loop so first we set our API key and then while true so this is just an infinite python Loop um with the very first thing that we do is we get user input then we vectorize it and then we set it into a dictionary and all these all these should look familiar there's the speaker there's the time there's the vector there's the message and then there's the uuid and the reason that you want to give it a uuid is just in case you want to refer to something very specifically you can sometimes use timestamp as a uuid but every now and then depending on how you've got your program organized you could end up with two things with the identical time stamps and so timestamp is not necessarily going to be a uuid but if you pick two uuids they're guaranteed even if you pick them at the same time they're guaranteed to be unique and then we create a file name so then we just save it out to Json then the next thing we do is we load a conversation so this database these chat logs this this is our database right this is our Nexus so if you're familiar with my work in artificial cognition or Symphony of thought or moragi this is the Nexus anything that you want the chat bot to be conscious of needs to end up here um and I'm calling it just chat logs because when you have when when all you're doing is a memory task against previous conversations that is your whole Nexus but if you want um external sources we might have a second Nexus and we might call it like um you know KB articles or something and then we can put indexed files in here that are then searchable and then you might also have nkb articles this this is this is what you have as like your ground truth right and also once this scales up the search method that I'm using is not efficient you'll want to move over to something like Pinecone or weeviate or kudrant um because those those search engines are going to be much faster uh but yeah so this is this is where we're off to this is where we start so we add to the conversation then we load the conversation and the reason that we load the com there's again there's more efficient ways of doing this you could just append but the thing is is I want to be able to load all history I don't like because like if I kill this and start it again I want it to load all of my logs regardless of how it's going so I want one very long uh persistent conversation so that's where we're starting and what I'm about to work on I'm not going to show you like all the coding every single time because it's kind of boring but I can do it work it uh and then and then figure it out so what I'm going to do next is work on generating the response and generating the response first requires that we load relevant history um so on and so forth so we'll be right back with the next phase and actually like this is pretty simple so maybe we will take some time to do some of those improvements I was talking about but anyways be right back all right we are um just about done uh at least with the first version there's room for improvement of course um but let me walk you through the code before I show you how good this is so we left off with here where we handle the user input then we load the conversation so this is a really simple function just uh load conversation so it goes into the chat logs directory it looks for any Json file and then it loads them all appends them and then sorts them by time just in case they don't load um in correct chronological order my file naming convention should mean that if they load in in alphabetical order they should also load in chronological order but again you don't necessarily want to make that assumption so we sort by the time index inside the file because one of the key things to remember is the reason that we're doing it this way is human memories are associative so and you can approximate associative memories with semantic similarity because we we our memories are queued up based on reminders this is why if you move from one room to another you might forget because your context has changed and so your brain literally like stores memories from one room and says Ah I'm now in the kitchen I have kitchen memories now um that's how like gullible our memory is the other thing is that is that our memories are based on uh chronological similarity we tend to remember things clustered in time like hey remember my eighth birthday right it doesn't matter where your eighth birthday was but it's like there's a whole bunch of memories associated with that tag one thing that you can do is you could probably do a knowledge graph with memories but that's a whole lot more complexity especially when you get up to like millions or billions of memories I personally think that semantic search is the way to go because it is fast and scalable I don't think that knowledge graphs are the way to go that being said time will tell okay so that was a little brief Spiel on memories so then once we have the user input it's time to compose the corpus so the Corpus if you've read my book natural language cognitive architecture the Corpus is all the context that is needed to generate a response to do something to perform cognitive labor and prepare to respond so first we fetch all the most recent memories with this function called Fetch memories and so fetch memories uh we pass at the vector for the most recent input we pass it uh all the chat logs and then we tell it how many we want to get back and this does a very very simple um uh dot product similarity score I looked it up and I realized that cosine similarity is not just a DOT product there's more to it so for for all the messages that you see on Reddit and the openai community about like why are my DOT products wrong apparently cosine similarity has a little bit more to it than just returning a DOT product and I cited my source this is this is what I'm copying this is the logic so hopefully our semantic similarity will be better so what this does is we um we we get all the memories we sort them by most relevant um we also skip we skip the current one because you know we saved the memory immediately um and just in case we have an identical one we don't need to return identical messages because you might ask the same message or get the same response at some point we don't care about that um so we we sort by most relevant and we pass back um n number of most relevant so that's fetch memories so this this becomes our most recent uh our most relevant memories you can just this is parameterized so if you have a bigger chat window you can do this then what we do is we summarize those memories so I have this function summarize memories so first thing we do is we sort them chronologically so that there's a nice chronological flow then we reconstitute that block and then we summarize it um so we we just basically take notes um so the summarization this is something that happens in the background of your brain particularly while you sleep so one what happens while you sleep is that your brain replays memories from throughout the day and it simplifies them it distills them down to the most critical Essence it also links them to other relevant memories this is why sleep is so important for learning but instead of doing it behind the scenes we're doing it in real time so a future version of this might have it where we uh we're going over we're grooming our our uh record of memories and and simplifying them and and doing that sort of thing in the uh in the background but instead we're just going to do it in real time so we we pull an arbitrary number of Memories We summarize them so that they'll fit into a new prompt by distilling them down by taking notes then we get the most recent uh conversation uh messages again um I wanted to treat this so that it has persistent memory so that's why I'm treating it like okay if this just started up and we have a backlog of Memories We want to pull all of them so we get notes from irrespective uh temporarily irrespective memories oh so this is another thing what you might do with fetch memories is once you've picked the um the most relevant memories so I'll add a two-do to do um uh pick uh more memories uh temporarily nearby the top most relevant memories um so that way like we should we should pick like you know not just one memory that's relevant we should pick like five memories in either directions but to give it more context so we'll add this as a to do item for the future um actually I should probably do that for for other stuff because this will do a lot because then um by using semantic search you can you can pick you can find anything in time but then you also want to pick more relevant memories um and then where's the notes summarize memories um so I'll add a to do uh to do um uh do this in the background over time um to handle huge amounts of memories um so basically uh just kind of putting putting a little bookmark for later um okay so we get the most recent um and then of course you can you can change this I just have it the four most recent uh back and forth and then we uh pipe it into this response and I'll show you the prompts that I've written in just a second um actually no this is this is a good enough time um so here's the notes it says write detailed notes of the following in a hyphenated list format um and I just tell it exactly the format that I want so basically we're taking notes um of the previous conversation and then away it goes and then here's the response so I am a chat bot named Raven so you give it um this is what I call an agent model so it says I am a chatbot named Raven so that may that way the model knows this is what I am this is the context my goals are to reduce suffering increase prosperity and increase understanding I will read the conversation notes and recent messages and then I will provide a long verbose detailed response so we want it to be chat GPT like um the following are the notes from earlier conversations with user and so then we've got the notes uh the following are the most recent messages in the conversation and then here's the conversation I will now provide a long detailed reverse response and then I just queue it up so that it it's gonna it's gonna copy the same format as up here so it just knows ah it's my time to speak um so there you go and then when we go back here we generate the response so we generate it with uh we we pipe this prompt into gpt3 completion and the stops that I have are user and Raven so it'll keep talking until it tries to continue the conversation which is not what I want it to do so there you have it that's that so now that we've got the output we need to we need to save the output the same way that we saved the user output so we vectorize Raven's output we put it into an identical object and then we save it out to our log file so chat logs and then you'll see here we know that it's Raven but we use otherwise the same format and then we print the output so there you have it it's that simple let's go ahead and test it so we'll do python chat um hey Raven and because it's doing so much behind the scene it's a little bit slow but so is chat GPT there's all kinds of stuff we can do to parallelize this and make it faster um let's see I'm a computer system uh okay cool he provided it very uh like I hope that I hope this helps yeah thanks um I am having trouble sleeping I am waking up at 4am every day this is actually like quasi-true because you'll if you look at the time stamp it's 5 15. um this is just me this is just my circadian rhythms so that sounds like a difficult situation it's understandable that you're having trouble sleeping there first thing there are a few things you can do blah blah go to bed and wake up at the same time avoid screens so very helpful very chat gpt-like response so if I if I continue this conversation over a long period of time you will see that we're accumulating our chat logs here and we see user Raven user Raven and they're all in chronological order and they're all searchable so if we keep if I can I'm not going to continue this conversation indefinitely but I just wanted to show you that it works and it's pretty straightforward um cool thanks um let's talk about circadian rhythms and so one thing that I can show you to kind of prove that it's working behind the scenes is we can look at our gpt3 logs and so we can see here's the notes so uh yep so we've we've compressed this into this so we've got really nice concise notes about the previous conversation um so that is how you can handle large volumes and then here is an example of what it's uh uh responding so you can see we've said here's the notes so it it Raven has some longer longer term context here's the most recent conversation it's just the last four um and then here's the the final output um so let's see what Raven said absolutely circadian rhythms are the body's natural blah blah um okay um what do you think about uh The Singularity um when will it happen and so I've basically just created my own personal chat GPT with long-term memory and I've given it my own goals um it does not have the the chat GPT goals um so blah blah let's see the concept of the singularity is this uh so on and so forth um yeah and as I mentioned you can also add functions where you can search uh KB articles in the same way so let me actually add that as a to do so so adding that so for memories and so there's there's two kinds of memories there's episodic and declarative memories so this is a pull episodic memories um and then to do um fetch declarative memories um AKA facts you know that are not attached to you so like facts uh wikis KB Etc um and that can ground it in um those so this could be like you know company data it could be Wikipedia it could be internet um so I'll just say like company data Internet Etc um yep so that's another to do and to add this you just have another um thing in your Corpus right here and you'd say like okay notes from the following and then you might say um here is some uh background um knowledge that may be helpful for the conversation and then you just say like you know KB or whatever um so that would be that would be um uh how you how you populate this and we'll get to that in the future um but yeah so I think we're done like super simple super straightforward this is long-term memory this is also a cognitive architecture so I want to point out that this is um this is the simplest implementation of natural language cognitive architecture um that I have come up with yet and you might say well where's the inner loop the inner loop is um everything that you're seeing here so where you where you compose the Corpus this is the inner loop and then the outer loop is the is uh the interaction here so there's two Loops they interact they interlock and it it it's it it um has some time where it's thinking and then it generates a response um now this doesn't have autonomy rate this version of Raven is not thinking on his own all the time um but you can see that there's clearly some thought going on um and we've set the stage to make it more extensible uh okay I think we'll call it a day I can hear all of you salivating and I'm very excited and want and wanting me to continue this but we're at 20 minutes in um so thanks for watching as a quick reminder um the offer is still on the table if you guys support me enough on patreon I'll remove ads for good um the best way to get in touch with me is via patreon there's also LinkedIn and finally my mailing list and links for all of those are in the video description thanks for watching take care oh and one other thing um I mentioned I think at the beginning of this video that uh the comments are going crazy I cannot respond to all comments on YouTube anymore um so I would encourage you guys to talk to each other um and also upvote uh good comments and that will help me zoom in on like the questions that you all agree you want answered um okay I think that's it talk later
You are awesome at this
I&#39;m having an issue with the GitHub version. Sometimes the responses are including date and time. So I have a tts engine reading the response and it ends up reading the date and time when that&#39;s returned
I&#39;m extremely new to all of this. New to coding and ML. I&#39;m having fun with the Github repo. I&#39;m wondering, I want to integrate messages pulled from a chatroom. the pulled text would be formatted as &quot;username: text&quot; where username is dynamic. so when it gets to the input, and &quot;user:&quot; is injected, it ends up being &quot;user: &quot;username: hey&quot;&quot;. so this would be a problem with how gpt-3 reads it? and output format? So A) could I somehow make &quot;user:&quot; in these areas of the code dynamically change to whatever the username is in the pulled chat message and then have that part redacted from the chat message? and B) if I could do something like this, how would it affect the memory of &quot;raven&quot;? would it understand which memories belong to each user and recall user specific conversations? I want it to be able to summarize the overall conversation as well as have user specific replies when multiple users are talking to it in turn. Maybe prompting it to understand there are multiple people speaking to it?
Following, please keep me updated, definitely curious.
I would love to weight the history when the user asks RAVEN to remember something or keep it in mind., Yeah, I think we&#39;re going to achieve that with knowledge graphs.
Ads is ok...Content is the key ... Cheers
downloaded from github but when I run it it just says Traceback (most recent call last):<br> File &quot;D:\chat\<a href="http://chat.py/">chat.py</a>&quot;, line 2, in &lt;module&gt;<br> import openai<br>ModuleNotFoundError: No module named &#39;openai&#39;, pip install openai
I was researching chat memory, and asked ChatGPT on how to solve this considering the token limits in, for example, davinci. ChatGPT suggested that to keep track of the essense of long conversations, one option is to keep track of the last 3-5 Q&amp;A, and use GPT-3 to create a summary of the remaining, previous Q&amp;A. <br><br>First ChatGPT suggested that I use Open AI&#39;s caching feature (contexual completion), that allows me to store the model&#39;s state between requests, but it later apologized and said that this feature is not currently available in OpenAI.
This would be amazing if you could have it do something with your second brain using obsidian or something similar. Using your knowledge base as primary reference. Helping you remember and improve your knowledge. having your knowledge base act as true assumptions. This could be a game changer on how much useable knowledge you have. <br>You could have your pc find all the information for gpt doing a decent amount of heavy lifting while gpt would have give context and possible way to extend your knowledge and actionability. You would need to curate Ur knowledge base but would make you very op., That&#39;s basically the goal of RAVEN.
David, would it be possible to create some type of better large text summarizer using this? If so, would you be able to make a video about that? That would be so awesome., You may want to define &quot;better&quot; and &quot;larger&quot; Migrating the vectors to a vector database gives for some nifty semantic search + summarize options across large datasets
This is amazing, I just had to sub to your patreon. Do you think a prolog based declarative memory would be useful to its heuristics?, I don&#39;t think controlled or constructed languages are necessary. I think that natural language is superior because it is infinitely flexible., @David Shapiro ~ AI My apologies, english is not my native tongue.<br>By prolog based declarative/fact memory, I meant a long term memory system that is written on languages based on first order logic like prolog to store, deduct, validate or disprove facts via the criterion of falsability.<br>I think cognitive NLP systems could use logic programming/formal logic to help them face uncertainty, discover new knowledge and adjust their inner model of the world, thank you for your reply!, What do you mean by prolog based declarative memory?
Great video as always David! Do you believe this long term memory is a model we are likely to carry-over to future versions of GPT? Like a personal assistant fine tuned along the years with a deeper understand of the user personality? (as an example).
this is probably the greatest of all time AI lesson i&#39;ve come across - and i&#39;m sorry i cannot afford a subscription to your account, but am so thrilled to have somehow stumbled on this video
I like the approach, but checking previous conversations for similarities will require parsing a sizable list of JSON over time. Even if you group them somehow, we can&#39;t really set a limit unless we mimic our brains, which over time forget certain elements, that would be translated here to the deletion of the oldest JSON files? Is this how you envision the architecture? I also want to extend my chatbot (GPT3 based and with amazon echo as input/output device) with LTM and I don&#39;t know how to deal with this constraint. I won&#39;t store JSON but record in a database but at the end the limitation is the same.
You&#39;re amazing. I love all your content, thank you so much for sharing it. I just subscribed to your Patreon
Great work. I wonder if it were possible, just like with ChatGPT, it would be possible to have multiple conversations and switch between them &quot;seamlessly&quot;. The point is that while the human brain, as you explain in another video, can switch contexts easily (e.g. I&#39;m in the kitchen now), it is not so for gpt-3, where lengthier conversations about ranging topics will only muddy the memories. Then again maybe I am missing the whole point ;)
Just found your channel today from that SCOTUS series. Interesting stuff and really great work here! You mention an option to include KB, wiki etc. How do you square that with the token limits of a single query? Even a summary will lose some important details. Can I get around that with fine-tuning? or how else would you take several megabytes of knowledge base content?
What about for code output layout?
Thank you for the great video. If were to build such a python program would we able to provide input that includes multiple via the CLI or would that require some kind of HTML frontend page?
Thank you for this, totally pointed me in the right direction!!!
Does the token count taken into account change with LTM? Do the previous conversations count as a part of the prompt or is it a database
lol I literally just finished coding a conversationQueue I linked up to a tokenizer to keep the context of my api instance. I havent watched yet but I&#39;m interested to see how you handle it
James Weaver&#39;s LangChain integrates Wolfram Alpha dr. Alan d Thompson YouTube channel. After the integration slow down the process so it takes you 30 seconds to get a high-quality response, people might start optimizing the integrations to make them faster, lighter, smaller, customized and portable, with low power usage so you can use it on a mobile device taps into a larger and more integrated and slow to come up with deeper thought. Eventually someone will write a looping script capable of re-scraping creeping the indexed and summarized inputs and outputs so it can systematically reflect and think about things over a much longer period of time.
I have pretty much bought into the biphasic sleep thing. I mean, I am 45, and for the last few years, I almost <i>always</i> wake up after four hours and then have to go back to sleep for a second period if I want enough. So there is a theory that that is natural.
Is chunking only viable to do manually? How do you chunk message with multiple topics?
Sometimes, my chatbot gets stuck (for example, it continuously replies with, “!!!” or an emoticon or even produces a blank response) and I need to restart the conversation. With long-term memories incorporated, would I delete the JSON files generated after it gets stuck in order to get it unstuck?
Extremely helpful. Thank you very much.