{"id":"BbrAOE8_BR4","title":"NEAR AI Office Hours #8 whiteboard session with Illia and Alex","description":"Building Smarter AI: Overcoming Challenges in Language Model Development\n\nIllia & Alex discuss the latest AI research topics and unpack some NEAR.AI focus areas\n\nNEAR AI: https://near.ai/\nIllia Polosukhin https://x.com/ilblackdragon\nAlex Skidanov https://x.com/AlexSkidanov\n\nIn this NEAR AI office hours the team delves into the intricacies of developing AI agents capable of building user-defined applications using language models. They share insights on using open-source models, the importance of good prompts, and the challenges faced in summarizing and compressing large contexts. Listeners will gain a deeper understanding of the current state and future potential of AI in automating complex tasks.\n\nHigh-Level Takeaways:\n• Good Prompts Matter: Effective prompt engineering significantly enhances the performance of language models.\n• Open Source vs. Proprietary Models: While open-source models like Llama have potential, proprietary models like GPT-4 currently perform better for complex tasks.\n• Context Compression Challenges: Managing and summarizing large amounts of context is a major hurdle in developing efficient AI agents.\n• Human-AI Collaboration: Human intervention is crucial for refining AI-generated summaries and ensuring accurate task execution.\n• Future of AI in Development: The ultimate goal is to enable AI to build sophisticated applications, including smart contracts and decentralized web applications.\n\n\nJoin NEAR's ecosystem:\nWebsite: https://near.org/\nTwitter: https://x.com/NEARProtocol\nBlog: https://near.org/blog/","channel_name":"NEAR Protocol","published_at":"2024-07-02T00:00:00.000Z","thumbnail_url":"https://img.youtube.com/vi/BbrAOE8_BR4/hqdefault.jpg","transcript_text":"blockchain flows near the one we're close to won't stop us now together we'll red Define the game with near we'll find the flame empowered Minds rise and store use your own AI we explore hold the future hand in hand in this teex driven Wonderland okay I guess we're live uh great to have everyone joining us at the episode number 8ish I guess uh we're here from signing pretty hot marach uh at the near AI offsite and brainstorming and kind of c-d developing and so we're going to dive in into some of the stuff we've been working on and some of the problems we're trying to solve um as usual not prepared by Design and so you're going to be learning with us uh what are we doing so yeah let's dive in kind of the problem statement we started with was um given a user who wants to build something and a language model how do you actually get it to build what user wants and Alex build a s yeah I guess we started by having everybody build a version of Devon uh and um none of them worked very well uh we have certain constraints which uh uh not everybody has one of them is we only using open source models uh and uh particularly we're mostly using llamas we tried to do mistal and mistal wasn't as good for this particular purpose uh and we narrow down this scope right now quite a bit to only building console applications because they easier to test you can once you nail that you can add browsers and whatnot right and what interesting it we learn well one thing is uh good prompts matter uh and models you know work much better if you give them examples of what you want from them um second thing we learned is that uh if you do not restrict yourself to open source models and do replace Lama with gp4 it does perform much better so there is something to be done about access to user owned AI for everybody so I guess some effort to build larger Community owned models um is in order um couple other things one is that as you develop uh it takes quite a bit of the model is like talking to itself right and it takes a lot of context to do that and even if you have longish context you run out of it so you need some way of compressing it that's actually quite a fun problem um that we've been playing around with uh again uh the only thing we really tried is trying to different trying different prompts that say Hey look at this big context and make it small uh but again uh good PRS matter um and something we are about to launch and uh we will we will announce it somewhere but uh as a as a live Watcher you get early early access yeah well I guess yeah yeah yeah we will have some sort of code to get your access but we uh one interesting hypothesis we have is that uh if you keep so the way the way kind of the we call it zela agent um uh and the way it works is it begins with the task specification uh and then it thinks out loud uh and as it thinks out loud it has access to certain tools uh and right now the set of tools is very restricted uh it's reading files in the file system writing files and executing commands in the console uh and uh in this period of uh living life only once I'm running it on the same on the same machine on which I have keys to all my crypto uh and so it already made its own home somewhere uh yes yes Escape uh and as it does that at some point it runs out of context and when we see that it's almost almost out of context uh this is where we invoke uh the summary module that says hey look at all of that uh and write enough text so that you forget of all of that including the initial task uh but you still have access to your environment you can continue working right and yeah so there're kind of some kind of environment that it writes reads you know execute commands run tests and so this environment kind of persists yeah the environment is still there but the history is gone uh and when you play with it it's generally not great but it's not bad either right it's it's actually making quite a bit of progress and it can write some very cool stuff um uh but summaries is the biggest failure mode right so like if it messes up here it usually is capable of recovering not always but most of the time but if the summary is messed up there's no recovery right you don't have access to history anymore so the hypothesis is the following the hypothesis states that Lama 7B or code Lama 7B uh is good enough at doing this uh like let's say it has infinite context and it's very good at using it the the hypothesis is that it will be able to recover from any failure uh by you know like repeatedly reflecting what on what it's done uh and like interacting with environment but the summaries need fixing and so we will make humans fix them for us right so we will uh we will launch very soon on near uh a project where everyone can earn some near by making better summaries given the um uh the the actions of the agent in the environment uh and then yeah like the the idea here is that yeah it's going to be a not that great model thinking out loud but a very good model um doing the snapshots and then we'll see what happens like maybe in a couple of office hours uh we will have maybe the way to kind of formalize it a little bit right like before you go just like the ultimate goal is not to be building console applications the ultimate goal is to be building smart contracts and front ends and and you name it right it's just console application is a very good self-contained task like until we can write a python script that you know implements a game of tic tac toe we definitely not there to build a contract sorry interrupting mostly can but yeah but yeah I mean because this requires also writing tests and making sure that uh it runs correctly and uh in a way a lot of the even smart contracts are end up being kind of resembling command line tools but yeah maybe to formalize it right like we have uh you know some tasks and it is able to you know run for a while interact with environment and in a way like this is just another session like this as well right it gets an environment which you know here it's also can be not empty yeah yeah maybe maybe there was something maybe it's on the first one yeah yeah and so in a in a general way right it's like you dropped into some you know file system folders with some already files written and uh you're given some summary of what's going on and what you should do and you should be able to do it within you know some period of time or be able to kind of summarize it such that somebody can follow from there and uh yeah I mean the the benefit here is like then you can even open any project right that already exists and say like yeah well here's the the summary here's the summary and here's what you need to do and you know you can send it to actually try to fix a bug or Implement a feature in existing codebase so it is a kind of General General enough framework General enough framework for being able to operate on uh kind of a code base and and yeah I guess we don't necessarily believe that this is the best approach to doing that uh the idea here is to build infrastructure we can very quickly iterate on those ideas right so like as we right now work on it is it open source yet not yet right it will be open source very soon but like we we have like building such an agent in our infrastructure right now is extremely simple you like drop a small python file you say hey uh like environment start with that agent uh and it's like doing all the stuff and the same for data notation so like we uh the idea is that if if we think that a particular set of data would help us uh the idea is that you should you should go from idea to uh you know humans already annotating it in like in couple hours uh which is uh not quite the case yet but we getting there well and it's a combination of right you have so you have environment as a kind of place you can run agents right and this this doesn't need to be one agent it can be a whole host of them then this can be run in few modes right interactive mode where as a user you also can you know either help with this or even help it with kind of its own process and things so user you can run it in so-called kind of task where you just you know tell it to go and do it and you can run it in benchmarking mode where you actually have a data set with set of tests where you know you give an environment and a task and you know it runs it maybe multiple of the sessions and then when it completes you can run on on set of tests it hasn't seen to validate that it's actually done the correct job and so this is important because this allows to now when this is open source anybody to go and kind of play with it propose different agent framework kind of agent setups as well as different ways of this interactions happen as well as you can you know inject here obviously an LM or multiple LMS and this is where you can swap to you know a fine tune model from some data to actually have uh uh for example if like summaries are written by a dedicated model that is fine tun on specific examples so that's kind of the framework that been designed obviously there's a you know sessions recorded and and uh uh then can be used as well for training and fine truning now the interesting question is you know looking a little bit ahead uh and anal analyzing some of the failures in general with LMS uh both in open source on but even in the close Source from what we've seen uh so maybe we can look at that as well as there's been some new research that has been published um around kind of various search techniques and uh some of the reinforcement learning both from human and from synthetic data um which also can be interesting to explore any initial yeah so so maybe just like as a as a simple right right now you know you give a model a task right like if you give to LM and you can you know it can just be able to answer it uh so this kind of direct mode you can say you know thing step by step and you know provide me an answer after it's called Chain of Thought now the problem with Chain of Thought is that sometimes it kind of gets into a stuck yeah it gets stuck itself and and because it keeps attending to what it just thought about it keeps kind of in that Loop and so one of the ideas was proposed it's called tree of thought where you know as you kind of unrolling tokens right let's say you know I you can you know unroll M or will for example they the second most probable or third like and then you're kind of unrolling like few different options uh of of this thought process and then you're evaluating you know few of these paths for actually getting giving an answer or you kind of getting all answers from here and then scoring them in some way as well now um doing that on a token level is pretty expensive uh it's Al uh kind of the challenge I see with this is that it probably still will just select same semantic meaning just like a different token that's kind of close close enough to that right like it will not actually try to think in a different um in a different way I think it's also interesting so like one idea when I was thinking about anything related to trees with the ranking I was always thinking that there would be a separate Network separately trained for it that tells you how good it is right but like as we were reading I I would not recall the name of the paper they say you just ask the model to rank it right and in the example here one of the that we just erased but it wasn't that High Fidelity anyway um one of the things that we observe for example is you say build me game of checkers right and as it works uh it can confuse itself uh it can confuse itself in a very interesting fashion is which is in the game of checkers sometimes the promoted piece is called a king or a knight King right kinging yeah king or nighting anyway let's say king and once it names it a king it gets confused and starts building chess uh and we wanted to see we're like well that's a great example right we see a failure the model was asked to build a game of checkers it didn't get to to the summary yet it has the whole history in the context but it it converts to building chess right and we're like this is a great example to test the ranking function we give it the implementation of Chess and say is this Checkers you want your task is to build Checkers give us the rank and it's like this is not an implementation of checkers it's more of an implementation of with chess but looks pretty good six out of 10 right so it's a yeah so it's uh I guess it's probably a signal but it's a very noisy signal which is not U and again that was lama lama 7B right yeah Lama 7B but but importantly like the way this models are trained right is they're trained on predicting next token and so kind of a lot of the a lot of the times they kind of dropped in the middle of the document I mean like in the beginning of document but docu document may have some context of like being located on some website but they kind of need to like start predicting everything without knowing what it is and so similarly here it's like hey you know I see chess so I can G to continue doing chess even though you know one token was saying Checkers right in the beginning of the document and so so kind of relevant relevant the relative importance of tokens in in the prompt and context right like if you think of it there's a checker here it's one token like you know build checkers game and then you know there's chess king queen yeah King gang yeah yeah like all all of this you know and so like when it's predicting ex toen it's like just by the fact that um like it is fine tune that this is like a you know system message and you should pay more attention to it but uh just naturally there's a lot more semantic meaning so it's kind of more of a systematic you know ne ne toen predictor um question and so I mean you can keep fine tuning it right you can keep kind of giving it more examples where this is I mean and this is where the synthetic data can be pretty interesting because you can unroll indeed it building something else and then say you know this should not be a some you know some other application and so you can train it on you know lots and lots of different routes where it's not yep to be able to learn to kind of not diverge um as part of you know just general training procedure and so in general maybe yeah the the kind of um there's one one of this process called kind of review so like you know after after model done responding you say like okay now review the response of this model as as a Critic and tell what's wrong with it and uh then you use the output of that model to feed it back like okay now that you have this review and this original context you know refine your response um and again the the question here so one of the challenges this is usually fine tuned with so-called instruct data right you collect examples of like you know either answering questions or maybe doing some like outputting something this is where things are complicated because we don't have like a thought process of a of a people usually so worth mentioning instruct is probably also human anotated you canate Chain of Thought yes but that's not readily available but but that that's a good example is like having some Chain of Thought especially the ones that includes doing something wrong and then backtracking and saying like actually I was wrong let me go back and you know fix it because again this model is rarely seen that like oh actually I was supposed to do check Checkers let me continue with that and so you know train it to kind of MH critically review its own thought process yeah and then there's a little bit more sophisticated than tree of thought is some kind of Monte Carlo uh Tre search where you are pretty much deciding what node to explore based on kind of uh some value function that may or may not use a different language model and indeed you can train also like another discriminator that is trying to if you have tests that's outstanding right because at the end of the day like oh this is a this passes test yes send the gradients yeah and so I mean all of these methods don't require actually training models unless you want to you know F tun it to be better at doing that process um but in the model itself usually is not trained on on this processes so I think the the the interesting question I've been thinking about is how do we how do we go like how do we reduce the importance of the prompt right because the challenge is right now like you need to be very you know almost like AB testing different problems to find what is working and then uh there's some small things that actually can go and mislead the model in the prompt itself and uh and then the other question is you know assuming something like this the challenges like when when you're enrolling in different alternative like you learn something here it's but that knowledge that oh this didn't work is not actually available at the different rooll so uh so in the way maybe restructuring this not like on it is available in the sub right like if something broke here then all the subs know it doesn't work well I mean let's say let's make it not a token but like you know let's say it unroll it got stuck on doing chess right and let's say we even are able to you know score that well you're going here you're starting to unroll a different sub tree and you know getting again into chess because like this I mean statistically right you just sample a different token for example right well hopefully well the the better like yeah that that's not that big of a problem right so even if so effectiv we're saying there's a partic there's an initial State we know the particular task and let's say that the model because of some biases is very likely to do something wrong repeatedly right so chess is more popular game than Checkers so it is repeatedly going to be doing checkers right or like this idea with like hey you need to move the Cabbage to another M you know Shore but because cabbage and nothing else but because you you've been seeing that problem you have very high chance but as long as you have any way to test right uh and you enroll sufficiently many right you like unroll it 100 times and it was saying hey go with cabbage go with cabbage go with cabbage but once it was like oh wait a second go with cabbage and you're done mhm right and that is the correct one as long as you can reinforce it as long as it has any chance of getting there Tre search will help you but this this is you're saying in at this is at training time I'm talking about inference time yeah at inference time you like you just wasted 100 Cycles on doing the same thing so the question is like at inference time when you like go with cabbage go with cabbage like how do you actually force it out of the because most probably it will you know go with a like in general this models are like you know here's here's an average of Reddit right an average Reddit goes in this direction so like how do you actually sample like some more uh you know different op like semantically different options so and I mean there's obviously methods to do that but it's specifically important in this kind of approach is where if you want to go get some some different ideas like you're not going to you know keep hammering which again this is what we see the behavior it does it's like it keeps hammering on the same idea thinking that you know okay well if I do it just you know a little bit differently it's going to work but yeah I mean the the other approach is indeed like training it to give sufficiently different approaches in right so in general uh the experience of of using a model to critique something without specifically training it for critiquing that particular aspect are like at best okay right so so it will it will tell you something meaningful and it will spot the most obvious things uh but it's uh that's the same model yeah yeah it's uh so like I would say I mean these two are specifically benefiting from critic yeah from a separate critic it would be much better if the critic is trained yeah true I mean the you know you can also then train CH a Chain of Thought you can train it in this way right that that's kind of the difference this is like is this the inference time or this is a training time where you use it to improve your direction of thought because obviously as a person we kind of we're doing some kind of search but it's all in linear fashion it's like well let me think about that actually that's wrong let me go back well yeah like you can think of like if this is your tree yeah that that is continuation obviously tree of thought is convertible to Chain of Thought with like this right yeah so like so you have all of the all of these things like this and then like this and then maybe this this and then yeah hey here's a problem I will use dynamic programming to solve it and you like think about it for a while it doesn't work then you try something else and like maybe dynamic programming was not a bad idea yeah but yeah obviously I'm not literally backtracking my brain to the previous checkpoint that would be cool but it's but it's kind of what we I mean what generally is happening right now when you're like redoing the prompts right imagine you can unroll your brain to before you saw the debates that's a separate line um yeah okay um what else we can dive in here yeah I think I mean a lot a lot of it is then end up being in kind of just having very convenient infrastructure to run fine tuning being able to run these experiments so so part of it so so one interesting thing here is that um having a way to measure um progress is helpful for both training and at the end of day at the inference time right so if we have a data set uh where we have a you know definitive way of saying that is the correct result right like let's say we're trying to solve competitive programming right for every problem we have tests we we we know exactly when it's solved it is much easier to experiment to experiment right first of all you can measure progress you just run you know you have a hold out set of 1,000 problems um which hopefully is not in the training Set uh and you can test right and second of all things like three of thought suddenly are way more easily trainable because now you have feedback you can say oh I solved the problem I can backtrack uh and give reward so sort of like you know when you do game of Go you know that you won you can unroll positive work but like if you have if you want to build um when you work on general purpose development right you say well I want a website I want a website that is like notion but different right but Buton chain yeah you can yeah button chain you can have some tests uh to to test sanity right but it's uh uh it's much harder to have a comprehensive Suite of tests and it will be at the end the user will be interacting with it explaining what exactly they actually mean and making some decisions so I think here it sort of goes uh here I think one one thing is that it helps to to have a more generic agent solving both things that like adjacent but like let's say our ultimate goal uh is to build uh applications right not just let's say it is actually you want to be building like web and I think smart contracts in this sense are somewhat easier with smart contracts you can um you can like generally you have a very comprehensive set of tests for smart contracts that's that's the whole goal right but then you have a web UI for it uh and you mess something up in the UI and people say hey I want to swap I don't know uh 10 near for a uh for whatever for 10 W near uh but in reality conversion goes wrong in you swap you know a th000 and yeah so um so the idea what would help is if you have a combined data set where a big part of it has is testable and has good tests right so the the model can continuously improve and another one is sort of adjacent um and improves together right uh it also would be uh there's like a meta level you can teach the um ideally you would have some sort of system that interacts with the um with a result uh and provides feedback so that's something that um especially as you move to the websites uh gp4 might be decently well good at it right now but if if you think of Open Source models there isn't that many good if any um visual open source models right and and when they will come up I'm pretty sure they will be trained like on Layon or Json data sets right so it will not have that much of interacting with uis so that's something that I think would be uh pretty cool if we if we can have again that's this is something we're probably going to do in the foreseeable future but like uh a good data set and a Model F tun on it of people interacting with the applications providing feedback on their experience that that can then be like you know like you have this you have like a good uh critique model that is actually properly trained to criticize right if you have an app you like plug it in then and it's like interacting with it saying hey it's not what yeah yeah doesn't work or yeah like this flow doesn't make sense yeah so pretty much like what I mean what you're saying there's a whole yeah like training a critics SL summarization MH are the steps where okay and where you can thoroughly test test you should thoroughly test that's a I mean that generally should be the case yeah yes um I guess thinking more though how do we how can we upscale some of the existing open source models to to have kind of a more robust thinking well better data never hurts better data more yeah more reliable data I mean we discussed last whiteboard right that like currently the models trained on kind of random set of data at all times so a more curriculum where you actually find where it's not y getting it and uh fixing that y but on near people have been experimenting with the annotating data for a while there's multiple good ecosystem projects right now working towards it uh so so we will have uh a pretty large effort towards annotating a bunch of data to open source it later um and generally I think data notation is generally a very interesting uh area because uh it requires a lot of interaction between people like ideally if you if you want to annotate you know millions of examples of data uh and it requires people doing it you have to you have to like like build a scalable system where there are going to be a lot of malicious actors uh without an ability it's almost like designing a blockchain with a difference between blockchain at least certain things are verifiable like you know like I I can say that this message is signed by that person while in annotating data everything is subjective right so you have thousands of people doing subjective work verifying subjective work and obviously like for a particular you know like there's a um you there have there have been some large scale efforts and uh like you will never be able to document every single possible interaction right so people need to convert to some uh to some level of uh uh agreement between themselves right so it's a uh it's it's a very problem which is very um similar to what we design in blockchain so we have like I would like to believe that we can do it better than centralized um centralized companies better cheaper and more accessible right because like yeah it can it works at like yeah fraction of a cost yeah because all of the marshalling of kind of money happens on chain yeah and like a lot of orchestration can be done on chain as well so it's a yeah like I'm a big believer that blockchain based data notation tools will win eventually against the centralized uh companies with like armies of salese and uh um yep um another what else was okay so another interesting idea we were discussing today more in an AI space than blockchain we will be jumping between the two uh at the end of the day uh there's no blockchain without a uh is that's something El was mentioning he was like hey how can we escape from the uh from The Prompt hell uh from The Prompt hell uh and uh uh one of the engineers here with us in Morocco he suggested an interesting idea where um so so this is your prompt and uh uh so prompt engineering is a big thing right like you you you try different prompts and uh some of them work better some of them work worse anyway you you come up with a prompt which is somewhat good um but then you have some other way of measuring uh the sucess of your model right so like let's say you actually have a data set which has a lot of tests and ideally we should have a bunch of those uh so then you start training and when you fit the pr into the Transformer by Design right it will uh it will go through the layers uh and uh it will like populate those U keys and values uh that like future things will attend to would be nice like to select and clim it think it also made it bigger for oh yeah probably it's probably like a tiny yeah I guess anyway um and an interesting thing here is that for a particular prompt which presumably will be fixed throughout your usage of the model those keys and values are always the same right so one obvious thing that I'm pretty sure almost everybody does today uh is you can just cach them right you compute them once you drop them into memory uh and then whenever you actually use your model in practice uh you do not actually unroll that you just start from the next token um however um right and so in inad of having a full Transformer here you can imagine now that you have just a bunch of keys and values and the actual layer is not executed right so like let's Transformer has four layers total right so this is uh this is the actual roll right so this is first layer second layer third layer and so they are outputting qkv qkv Etc but here before the The Prompt is you have those Q KS and vs uh but this doesn't exist anymore right you don't not actually compute them they're just constants right KV KV for each layer so now the cool thing is that um like you know when you dra in the model the gradients flow through these things right uh but the gradients update uh this block uh like MLP and attention that compute them but here instead now that this block is going on K and Via constants you can f tune the con con themselves like the actual vectors so for a specific problem for a specific problem right so with a downstream task right like you unroll it and you're like oh it passed all my tests you send the you know the gradients back they update uh your weights probably luras um well or you don't that's or or or even or yeah maybe you you fix you fix the model so the weights are the same you're just saying actually there is no actual like effective prompt it's just some set of vectors attend to maybe initially you computed them by enrolling transformer for a prompt or maybe they're completely random oh I mean if if you wanted to find answer to increase the chance of Ever Getting to successful trajectory right yeah but then you can find tune the things themselves so at that point assuming your prom was at least somewhat good to the beginning to get to any success the goodness of your prom doesn't matter anymore because the promp just becomes the perfect prompt for this particular problem right yeah uh and not just the at the input level right at every yeah it's like what whatever the key value is that whenever you need to Output the next token would be the most useful yeah which is you know contains instructions and whatever else it can yep and if the model if you don't actually send gradients right then you just like drop one prefix at another I guess probably as expensive as plugging in lower on top of it but yeah so that's that's something interesting relatively easy to try right so like as we like I guess the closest example where we can play with it is uh with the summaries that I was showing at the beginning once we have them we um there's a good chance we will try this yeah we can one we can low which I mean one of the challenges with in general fine tuning right that it gets better on the set of data that you're showing to but it gets you know worse in some other places you don't even know where speaking of if we try this and it doesn't work you can be sure we will say it unlike everybody else who tries fails and does not like that's something that should become a norm a norm yeah like if you try something it doesn't work yeah like it makes sense to say hey we tried this it didn't work this is how we tried it like maybe maybe we messed it up yeah yeah I mean one of the going back to the infrastructure one of the things we're trying to do is like how do we keep track of experiments including failed ones and then you are able indeed to have like you know hey what what what was this okay well this was actually this code was this data and you can see kind of the outcome and maybe indeed maybe this outcome is an Insight on this doesn't work maybe it's you know has a bug in the code or some other failure or maybe data you know is too noisy and so ideally you should be able to have this Logan be able to easily navigate all of them um and then you know have this tooling open for everyone so everybody can collaborate yep also would are we talking about the router I mean not yet but yeah that's going to be part part of the I didn't I didn't mention anything you unroll back and the context of everybody watching yeah but uh yeah few weeks um yeah so I think I mean the the interesting question here in general right like because is there is a prom that kind of like structures thing like the process of the further kind of processing data and generating and uh the the the other interesting question I mean the question I've been always trying to think through is how do we actually like as as a people as people will not just you know unroll unroll unroll and then maybe collect the data and fine tune but you know we kind of trying to update our weights somewhat on the F and so if you can imagine you know we have this models running with lots of people actually hammering at them even if it's just like okay well let's take all of the sun rolls and kind of fine tune the problem is the signal is not always there but uh how do you how do you actually like extract the signal and update the model and uh going back to the problem of fine tuning right again if you like you know there's some some distribution of knowledge it it has right and then like you you wanted to solve some specific problem that maybe it's not that good at it may like you know lose some of the yeah oh by the way this is something we were discussing today right that when you do Laura you inevitably forget something right like naturally if you mess up with the weights in which your knowledge is stored eventually some of them will forget something yeah yeah but this adds new weights but if you add new weights uh it doesn't necessarily mean that the performance will not be regressed right so like it is very likely and very possible that certain values in these keys and values will inhibit particular behaviors which you might want want want to preserve they just weren't in the data set but I think it is fundamentally the case that uh if this is what you find tune and not the weights then fundamentally model is as capable As It Ever Was because the model hasn't changed right it's uh it's extra you know strictly new data that you fine tune that naturally fits into the model so it's a yeah yeah Ilia yesterday Ilia out of nowhere stood up and philosophically said I want a function G such that if you fed into it a model and a batch of data it it yields you a new model such that the new model knows everything that or model knew plus the information from the batch right and feeling very smart I was like well this is obviously impossible because then by doing it repeatedly you can learn infinite amount of data into the finite size of the model but I guess obviously yeah there's no nothing in this problem statement was requiring the model to remain of the same size right and this is sort of that's a version of the function G yeah yeah so feeling smart and being smart is not quite quite the same thing yeah but the point of this is then you know you're able indeed to maintain the Knowledge and Skills but also potentially grow them but I guess I mean the other question is like how to how to you know how to evaluate this that it preserved well then like assuming you know what you want not to forget just have a benchmark yeah have like a universal benchmark which T the whole set of things you think are fundamentally important to know then if it regresses you're like well too far yeah except I mean if if you want to train this function G then that's a expensive process to evaluate every time at every step so um well you can hypothesize you can hypothesize that uh as you train especially like let's say we specifically train Laura right so we're not saying it doesn't forget everything M knows it just doesn't catastrophically forget important things then you can hypothesize that to random perturbations to the to the weights uh the level of forgetting is the same um for a particular distance of the new model right so like you you use some sort of distance like a k Divergence or what not and you like you know you do some experiments and observe how if kale Divergence is up to some value the model doesn't catastrophically forget stuff otherwise does then you know then you then you can cheaply say uh moreover like as you train you can like push it like like they do with poo yeah you add you add a loss pretty much I guess that's exactly the problem they were fighting right when you DOL yeah then were simp so small you you just yeah yeah yeah and this goes back I mean this can you know you can express it as a real problem where you know you have a policy I mean it literally is an real problem yeah and uh you know it's almost like there's some kind of AAR algorithm some ke learning um yeah but yeah I think like to summarize the kind of we we employing well we're building a set of infrastructure that is going to be open source and available to everyone to use so you can track experiments understand how experiments went um potentially run some on you know some shared compute um that's provided by some decentralized providers then you can use that inside this kind of agent framework to you know either actually try to solve your problem or generate unrolled data and that goes to crowd sourcing and then from there you can you know hypothesize some experiment around new data that would solve for specific capabilities while also benchmarking for everybody everything else and uh and test it out yep what ises general what does near do I guess when we didn't quite I I mean we have a blog post but it's I think it's worth mentioning what we do in shortterm right so um I guess the shortterm goal uh is we do want to get as far as possible in helping people build U web three apps right so so there's a certain uh there's kind of a chicken the neck problem where the number of web apps is not that large and so when the users come into the web ecosystem they're like yeah it's you know fun place but not much for me to do here uh and then therefore the like hypothesis is if we had a bigger ecosystem more users would come uh but you know developers like hey the number of users is not yet that high so um right and so web three remains relatively nich right but but if anyone can come and say Hey I want to build like the point here is you know having a notion which is decentralized would be outstanding right so today uh like notion uh you know you use notion five years later that will that will be gone right there's there's no chance of uh I mean there is a chance but there's definitely a chance of notion not existing five years from now or 10 years from now right so like it's like short-term ephemeral um stuff uh and like it would be so cool if there was a version of notion that snaps us to chain or or like to some sort of decentral storage solution which I know is going to survive uh but building it is very involved right like building notion in general is not easy uh that's why U the competition is relatively uh lacking even in the web two space right especially like I think the first 80% is probably relatively simple but like making it polished and whatnot right but if anyone can come and say Hey I want to build notion version on chain and you know some sort of AI agent will get you 90% there with guidance from you uh that will be a very different world right so suddenly uh you can build much better well you can General faster like go to market as well right because like oh I have an idea you know tap it in boom you have a first version of the app and now you can test the idea with an app yeah and and every day it looks less and less like science fiction um yeah we started on this problem in 17 when it when it was more science fiction yes but but like now even just like observing you know the this example I gave at the beginning with an agent tracking code like it's uh it's un canany close to being very good right it just like it messes up every now and then slightly but you know the the Delta to be close is very small at this point and uh it will happen um I mean even on the on the simpler programming competition problems it's already doing like 30% yeah on and so again compared to when we started in 17 it was doing zero yeah so so so so like like generally every day uh what was it the website we were playing with today web Sim yeah right so like there's this website I will write it down I yeah we we the only reason we know about it is because it's uh it's using what open router yeah yeah so there's a website called open router and uh I guess webs was one of the bigger users we were like well what is it uh and the first my first several interactions with web. it feels like you you have a portal to another Universe from which they ship stuff to you because it just I still have no idea how this works right so like I will give you an example uh we had access to Deon like to the actual one from cognition and I wanted to build a game of rock paper scissors with spaceships and it took me uh it built something eventually uh like it wasn't a complete flop but uh it took me a good part of a day uh you know following up and then in websam I just said you know rock paper scissor with spaceships and it just loaded in 15 seconds a fully functional game with much better animations it just or or like one thing given the example of of U chess and checkers I said build me a game of checkers where every piece is a chess Knight and again 15 seconds later did exactly that I didn't try with Devon but it's it would be an interesting uh interesting thing to see so yeah uh but the example is every now and then in the modern world you wake up to the news of something that yesterday would have felt like science fiction like Sora when it came out came out right it was like I mean it still didn't come out technically but right like yeah something completely impossible so I think the the world in which you just go and say build me this smart contract on near with the web UI and middleware is pretty close uh and there's a good chance it will happen before General P was developer just because the scope is so contained and smart contracts are much smaller way easier to test right to like some sort of research uh is pretty likely to to to yield good results and because we have access to such a you know uh like the data wheel where we can you know very quickly spin up projects where people who know near and like near and want near can very quickly you know put together a bunch of good examples so yeah I'm quite optimistic that that one will work out relatively short term um but longer term as the song as this beautiful song Don't suggested longer term that's like you know just a stepping step towards having a whole ecosystem like ideally yeah everything data sets uh inference um servers training servers um data I said data right yeah still data sets uh pretty much yeah like I mean experiments dat yeah it should be all I mean there's kind of a data layer kind of model experiments infra layer and then there's an application layer right and so in data layer you have you know crowd sourcing you have you know data acquisition yeah data contribution Network yeah you have some synthetic data probably and uh some form of curation as well and there's more stuff here right obviously and then here you have compute kind of broadly you have inference training the models themselves kind of customized models and then application layer you kind of have all all of the fun stuff that people are building and so you want kind of the full stack right to not depend on you know this guy is going to just like whatever their API was down for example and like all the application stack wasn't working mhm and similarly you want you know be able to access data um or both cheaper and kind of in a different way where you know creators potentially getting rewarded versus it's all you know free freefor all um but again this is kind of more of a you know this should be a better stack to build on not just hey this is like you know because of values we want this it's more this is a better way to do it that's why we're building it and it's going to be you know cheaper more accessible there can be multiple versions of different uh pieces because they're actually you know having a healthy competition in incidentally the the existing stack is not that easy to use right I don't know what's the reason for that uh but like you know uh setting everything up even even even installing python on the Mac on the Mac yeah is not is not that trivial right so so it's like a rare situation where you can build a decentralized stack uh and you will not be fighting an uphill battle against a very you know like mac level you know everything yeah very streamlined process of the centralized uh Solutions so so I think that's a well because in centralized it's also all like kind of different right like you're using like computes from AWS you know yeah the influence is here the like the data the data in general you know you need to talk to salespeople you need to get API for you know $66,000 a month Etc right so you generally don't have like a good setup unless you're in one of the big companies where this is all kind of given as a as a one thing yeah so yeah I mean one of the kind of you know this experiment tracking and in general being able to ideally spun experiments in this setup is something that like you know internally some of the organizations have because it was worse for them building when you have like you know thousands of people together working but it doesn't really exist in open source because every every team is themselves kind of are having some version of that and so you can have like a shared version that everybody's integrating with if you work in open way in open source way yeah but uh I mean the the in the blog post right um kind of the idea that the I developer is also leads to yeah researcher yeah researcher and this is kind of I think there's a lot of the conversation right now about this I I think I just saw another thread uh by Jeremy Howard about this that's like like the current LMS are just you know just giving you statistically most probable thing and so yeah clearly it's not very good at like exploring different options um yet and so but you need you for sure will need some kind of logical reasoning and this is where the formal systems come in yeah um I mean you will need some way of yeah kind getting un stock and figuring out Alternatives like sufficiently different Alternatives um so like this kind of Search exploration which again we already see we kind of need even on the simpler examples worth mentioning this is brainstorming we have no idea how to build the a yeah but this is like I mean we we see this things are needed even fori developer right and they're clearly need for researcher um and so the I think generally like um making agents creative do agents in general um some influential figures on Twitter say uh it will not be LMS uh U whether it will be LMS or not hard to say right but but like we need we need creativity um as a yeah someone someone said that you know like FMA spent 20 years of his life learning math and then I don't know how old he was anyway the rest of his life creating math uh so it's something that uh uh no idea yet how to go in that direction but that's like a North star uh to make sure that you know the eventually the the systems we create are capable of advancing science advancing science and Humanity yeah but you know starting to focus on these two sides to make sure sure we actually making progress how we doing on time it says 32 seconds oh 1 hour 34 okay yeah yeah all right well hopefully this was good and educational um as always you know we're just kind of brainstormed and discussed some of the things we're working on and we're going to continue doing that especially when we are in person um and as well posting more stuff on the blog and uh releasing some of the code and experiments from this so stay tuned and uh thanks for joining yeah hope it was fun yeah","transcript_synced_at":"2026-11-03T05:46:35.609Z","segments":[{"start_seconds":3,"text":"blockchain","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3s"},{"start_seconds":9,"text":"flows near the one we're close to won't stop us","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=9s"},{"start_seconds":16,"text":"now together we'll red Define the game with near we'll find the flame","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=16s"},{"start_seconds":25,"text":"empowered Minds rise and store use your own AI we","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=25s"},{"start_seconds":38,"text":"explore hold the future hand in hand in this teex driven","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=38s"},{"start_seconds":47,"text":"Wonderland okay I guess we're live uh great to have everyone joining us at the","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=47s"},{"start_seconds":55,"text":"episode number 8ish I guess uh we're here from signing pretty hot marach uh","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=55s"},{"start_seconds":62,"text":"at the near AI offsite and brainstorming and kind of c-d developing and so we're going to dive in into some of the stuff","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=62s"},{"start_seconds":70,"text":"we've been working on and some of the problems we're trying to solve um as usual not prepared by Design and so","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=70s"},{"start_seconds":77,"text":"you're going to be learning with us uh what are we doing so yeah let's dive in kind of the problem statement we started","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=77s"},{"start_seconds":85,"text":"with was um given a user who wants to build something and a language model how do","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=85s"},{"start_seconds":94,"text":"you actually get it to build what user wants and Alex build a s yeah I guess we","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=94s"},{"start_seconds":105,"text":"started by having everybody build a version of Devon uh and","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=105s"},{"start_seconds":112,"text":"um none of them worked very well uh we have certain constraints which uh uh not everybody has one of them is we only","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=112s"},{"start_seconds":119,"text":"using open source models uh and uh particularly we're mostly using llamas we tried to do mistal and mistal wasn't","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=119s"},{"start_seconds":127,"text":"as good for this particular purpose uh and we narrow down this scope right now","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=127s"},{"start_seconds":134,"text":"quite a bit to only building console applications because they easier to test you can once you nail that you can add browsers and","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=134s"},{"start_seconds":142,"text":"whatnot right and what interesting it we learn well one thing is uh good prompts matter uh","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=142s"},{"start_seconds":150,"text":"and models you know work much better if you give them examples of what you want from them","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=150s"},{"start_seconds":158,"text":"um second thing we learned is that uh if you do not restrict yourself to open source models and do replace Lama with gp4 it does perform much better so there","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=158s"},{"start_seconds":166,"text":"is something to be done about access to user owned AI for everybody so I guess some effort to build larger Community","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=166s"},{"start_seconds":174,"text":"owned models um is in order um","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=174s"},{"start_seconds":180,"text":"couple other things one is that as you develop uh it takes quite a bit","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=180s"},{"start_seconds":187,"text":"of the model is like talking to itself right and it takes a lot of context to do that and even if you have longish context you run out of it so you need some way of compressing it that's","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=187s"},{"start_seconds":195,"text":"actually quite a fun problem um that we've been playing around with uh again uh the only thing we really tried","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=195s"},{"start_seconds":203,"text":"is trying to different trying different prompts that say Hey look at this big context and make it small uh but again uh good PRS","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=203s"},{"start_seconds":211,"text":"matter um and something we are about to launch and uh we will we will announce it somewhere but uh as a as a live","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=211s"},{"start_seconds":221,"text":"Watcher you get early early access yeah well I guess yeah yeah yeah we will have some sort of code to get your access but we uh one","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=221s"},{"start_seconds":230,"text":"interesting hypothesis we have is that uh if you keep so the way the way kind of the we call it zela agent um","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=230s"},{"start_seconds":240,"text":"uh and the way it works is it begins with the task specification uh and then it thinks out loud uh and as it thinks out loud it has access to certain tools","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=240s"},{"start_seconds":248,"text":"uh and right now the set of tools is very restricted uh it's reading files in the file system writing files and","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=248s"},{"start_seconds":254,"text":"executing commands in the console uh and uh in this period of uh living life only","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=254s"},{"start_seconds":261,"text":"once I'm running it on the same on the same machine on which I have keys to all my","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=261s"},{"start_seconds":268,"text":"crypto uh and so it already made its own home somewhere uh yes yes","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=268s"},{"start_seconds":276,"text":"Escape uh and as it does that at some point it runs out of context and when we see that it's almost almost out of","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=276s"},{"start_seconds":283,"text":"context uh this is where we invoke uh the summary module that says hey look at all of that uh and write enough text so","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=283s"},{"start_seconds":291,"text":"that you forget of all of that including the initial task uh but you still have access to your environment you can continue working right and yeah so","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=291s"},{"start_seconds":300,"text":"there're kind of some kind of environment that it writes reads you know execute commands run tests and so this environment kind of persists yeah","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=300s"},{"start_seconds":308,"text":"the environment is still there but the history is gone uh and when you play with it it's generally not great but it's not bad either right it's it's","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=308s"},{"start_seconds":317,"text":"actually making quite a bit of progress and it can write some very cool stuff um uh but summaries is the biggest failure mode right so like if it messes up here","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=317s"},{"start_seconds":326,"text":"it usually is capable of recovering not always but most of the time but if the summary is messed up there's no recovery right you don't have access to history","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=326s"},{"start_seconds":333,"text":"anymore so the hypothesis is the following the hypothesis states that Lama 7B or code Lama","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=333s"},{"start_seconds":340,"text":"7B uh is good enough at doing this uh like let's say it has infinite context and it's very good at using it","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=340s"},{"start_seconds":349,"text":"the the hypothesis is that it will be able to recover from any failure uh by you know like repeatedly reflecting what on what it's done uh and like","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=349s"},{"start_seconds":358,"text":"interacting with environment but the summaries need fixing and so we will make humans fix them for us right","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=358s"},{"start_seconds":366,"text":"so we will uh we will launch very soon on near uh a project where everyone can earn some near by making better","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=366s"},{"start_seconds":376,"text":"summaries given the um uh the the actions of the agent in the environment","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=376s"},{"start_seconds":384,"text":"uh and then yeah like the the idea here is that yeah it's going to be a not that great model thinking out loud but a very good model","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=384s"},{"start_seconds":393,"text":"um doing the snapshots and then we'll see what happens like maybe in a couple of office hours uh we will","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=393s"},{"start_seconds":401,"text":"have maybe the way to kind of formalize it a little bit right like before you go just like the ultimate goal is not to be building console applications the","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=401s"},{"start_seconds":409,"text":"ultimate goal is to be building smart contracts and front ends and and you name it right it's just console application is a very good self-contained task like until we can","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=409s"},{"start_seconds":418,"text":"write a python script that you know implements a game of tic tac toe we definitely not there to build a","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=418s"},{"start_seconds":426,"text":"contract sorry interrupting mostly can but yeah but yeah I mean because this requires also writing tests and making","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=426s"},{"start_seconds":433,"text":"sure that uh it runs correctly and uh in a way a lot of the even smart contracts are end up being kind of resembling","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=433s"},{"start_seconds":442,"text":"command line tools but yeah maybe to formalize it right like we have uh you know some tasks","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=442s"},{"start_seconds":450,"text":"and it is able to you know run for a while interact with environment and in a way like this is just another session","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=450s"},{"start_seconds":456,"text":"like this as well right it gets an environment which you know here it's also can be not empty yeah yeah maybe maybe there was something maybe it's on","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=456s"},{"start_seconds":465,"text":"the first one yeah yeah and so in a in a general way right it's like you dropped into some you know file system folders","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=465s"},{"start_seconds":473,"text":"with some already files written and uh you're given some summary of what's going on and what you should do and you","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=473s"},{"start_seconds":480,"text":"should be able to do it within you know some period of time or be able to kind of summarize it such that somebody can","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=480s"},{"start_seconds":487,"text":"follow from there and uh yeah I mean the the benefit here is like then you can","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=487s"},{"start_seconds":496,"text":"even open any project right that already exists and say like yeah well here's the the summary here's the summary and","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=496s"},{"start_seconds":503,"text":"here's what you need to do and you know you can send it to actually try to fix a bug or Implement a feature in existing","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=503s"},{"start_seconds":510,"text":"codebase so it is a kind of General General enough framework General enough framework for being able to","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=510s"},{"start_seconds":518,"text":"operate on uh kind of a code base and and yeah I guess we don't","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=518s"},{"start_seconds":527,"text":"necessarily believe that this is the best approach to doing that uh the idea here is to build infrastructure we can very quickly iterate on those ideas","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=527s"},{"start_seconds":534,"text":"right so like as we right now work on it is it open source yet not yet right it will be open source very soon but like we we have like building such an agent","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=534s"},{"start_seconds":542,"text":"in our infrastructure right now is extremely simple you like drop a small python file you say hey uh like environment start with that agent uh and","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=542s"},{"start_seconds":551,"text":"it's like doing all the stuff and the same for data notation so like we uh the idea is that if if we think that a","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=551s"},{"start_seconds":559,"text":"particular set of data would help us uh the idea is that you should you should go from idea to uh you know humans","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=559s"},{"start_seconds":565,"text":"already annotating it in like in couple hours uh which is uh not quite the case yet","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=565s"},{"start_seconds":573,"text":"but we getting there well and it's a combination of right you have so you have environment as a kind of place you can run agents right","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=573s"},{"start_seconds":582,"text":"and this this doesn't need to be one agent it can be a whole host of them then this can be run in few modes right","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=582s"},{"start_seconds":591,"text":"interactive mode where as a user you also can you know either help with this or even help it with kind of its own","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=591s"},{"start_seconds":598,"text":"process and things so user you can run it in so-called kind of task where you just you know tell it","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=598s"},{"start_seconds":607,"text":"to go and do it and you can run it in benchmarking mode where you actually have a data set","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=607s"},{"start_seconds":615,"text":"with set of tests where you know you give an environment and a task and you know it runs it maybe multiple of the sessions","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=615s"},{"start_seconds":624,"text":"and then when it completes you can run on on set of tests it hasn't seen to validate that it's actually done the","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=624s"},{"start_seconds":631,"text":"correct job and so this is important because this allows to now when this is open source anybody to go and kind of","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=631s"},{"start_seconds":639,"text":"play with it propose different agent framework kind of agent setups as well as different ways of this interactions","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=639s"},{"start_seconds":646,"text":"happen as well as you can you know inject here obviously an LM or multiple","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=646s"},{"start_seconds":652,"text":"LMS and this is where you can swap to you know a fine tune model from some data","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=652s"},{"start_seconds":660,"text":"to actually have uh uh for example if like summaries are written by a dedicated model that is fine tun on","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=660s"},{"start_seconds":667,"text":"specific examples so that's kind of the framework that been designed obviously there's a","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=667s"},{"start_seconds":675,"text":"you know sessions recorded and and uh uh then can be used as well for","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=675s"},{"start_seconds":682,"text":"training and fine truning now the interesting question is you know looking a little bit","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=682s"},{"start_seconds":690,"text":"ahead uh and anal analyzing some of the failures in general with LMS uh both in open source on but even","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=690s"},{"start_seconds":698,"text":"in the close Source from what we've seen uh so maybe we can look at that as well as there's been some new research that","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=698s"},{"start_seconds":706,"text":"has been published um around kind of various search techniques and uh some of the","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=706s"},{"start_seconds":715,"text":"reinforcement learning both from human and from synthetic data um which also can be interesting to","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=715s"},{"start_seconds":725,"text":"explore any initial yeah so so maybe just like as a as a simple right right","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=725s"},{"start_seconds":732,"text":"now you know you give a model a task right like if you give to","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=732s"},{"start_seconds":737,"text":"LM and you can you know it can just be able to answer it uh so this kind of","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=737s"},{"start_seconds":745,"text":"direct mode you can say you know thing step by step and you know provide me an answer","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=745s"},{"start_seconds":752,"text":"after it's called Chain of Thought now the problem with Chain of","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=752s"},{"start_seconds":761,"text":"Thought is that sometimes it kind of gets into a stuck yeah it gets stuck itself and and because it keeps","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=761s"},{"start_seconds":768,"text":"attending to what it just thought about it keeps kind of in that Loop and so one of the ideas was proposed it's called","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=768s"},{"start_seconds":775,"text":"tree of thought where you know as you kind of unrolling","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=775s"},{"start_seconds":782,"text":"tokens right let's say you know I you can you know unroll M or","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=782s"},{"start_seconds":790,"text":"will for example they the second most probable or third like and then you're kind of unrolling like few different","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=790s"},{"start_seconds":799,"text":"options uh of of this thought process and then you're evaluating you know few of these paths for actually getting","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=799s"},{"start_seconds":808,"text":"giving an answer or you kind of getting all answers from here and then scoring","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=808s"},{"start_seconds":814,"text":"them in some way as well now um doing that on a token level is pretty","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=814s"},{"start_seconds":822,"text":"expensive uh it's Al uh kind of the challenge I see with this is that it probably still will just","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=822s"},{"start_seconds":832,"text":"select same semantic meaning just like a different token that's kind of close close enough to that right","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=832s"},{"start_seconds":839,"text":"like it will not actually try to think in a different um in a different way I","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=839s"},{"start_seconds":846,"text":"think it's also interesting so like one idea when I was thinking about anything related to trees with the ranking I was","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=846s"},{"start_seconds":854,"text":"always thinking that there would be a separate Network separately trained for it that tells you how good it is right but like as we were reading I I would","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=854s"},{"start_seconds":861,"text":"not recall the name of the paper they say you just ask the model to rank it right and in the example here one of the","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=861s"},{"start_seconds":868,"text":"that we just erased but it wasn't that High Fidelity anyway um one of the things that we observe for example is","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=868s"},{"start_seconds":876,"text":"you say build me game of checkers right and as it works uh it can confuse itself uh it can confuse itself in a very interesting fashion is which is in the","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=876s"},{"start_seconds":885,"text":"game of checkers sometimes the promoted piece is called a king or a knight King right kinging yeah king or nighting anyway","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=885s"},{"start_seconds":894,"text":"let's say king and once it names it a king it gets confused and starts building chess uh and we wanted to see we're like","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=894s"},{"start_seconds":902,"text":"well that's a great example right we see a failure the model was asked to build a game of checkers it didn't get to to the summary yet it has the whole history in","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=902s"},{"start_seconds":909,"text":"the context but it it converts to building chess right and we're like this is a great example to test the ranking function we give it the implementation","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=909s"},{"start_seconds":917,"text":"of Chess and say is this Checkers you want your task is to build Checkers give us the rank and it's","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=917s"},{"start_seconds":925,"text":"like this is not an implementation of checkers it's more of an implementation of with chess but looks pretty good six out of 10 right so it's","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=925s"},{"start_seconds":933,"text":"a yeah so it's uh I guess it's probably a signal but it's a very noisy signal which is not U and again that was lama","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=933s"},{"start_seconds":942,"text":"lama 7B right yeah Lama 7B but but importantly like the way this models are trained right is they're trained on predicting next token and so kind","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=942s"},{"start_seconds":950,"text":"of a lot of the a lot of the times they kind of dropped in the middle of the document I mean like in the beginning of","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=950s"},{"start_seconds":957,"text":"document but docu document may have some context of like being located on some website but they kind of need to like start predicting everything without","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=957s"},{"start_seconds":966,"text":"knowing what it is and so similarly here it's like hey you know I see chess so I can G to continue doing chess even","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=966s"},{"start_seconds":973,"text":"though you know one token was saying Checkers right in the beginning of the document and so so kind of relevant","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=973s"},{"start_seconds":981,"text":"relevant the relative importance of tokens in in the prompt and context right like if you think of it there's a","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=981s"},{"start_seconds":988,"text":"checker here it's one token like you know build checkers game and then you know there's","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=988s"},{"start_seconds":996,"text":"chess king queen yeah King gang yeah yeah like all all of this you know and so like","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=996s"},{"start_seconds":1003,"text":"when it's predicting ex toen it's like just by the fact that um like it is fine tune that this is like a you know system","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1003s"},{"start_seconds":1012,"text":"message and you should pay more attention to it but uh just naturally there's a lot more semantic meaning so","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1012s"},{"start_seconds":1020,"text":"it's kind of more of a systematic you know ne ne toen","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1020s"},{"start_seconds":1025,"text":"predictor um question and so I mean you can keep fine tuning it right you can keep kind of giving it more examples","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1025s"},{"start_seconds":1033,"text":"where this is I mean and this is where the synthetic data can be pretty interesting because you can unroll indeed it building something else and","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1033s"},{"start_seconds":1042,"text":"then say you know this should not be a some you know some other application","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1042s"},{"start_seconds":1050,"text":"and so you can train it on you know lots and lots of different routes where it's not yep to be able to learn to kind of","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1050s"},{"start_seconds":1059,"text":"not diverge um as part of you know just general training procedure and so in general","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1059s"},{"start_seconds":1066,"text":"maybe yeah the the kind of um there's one one of this process called kind of","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1066s"},{"start_seconds":1075,"text":"review so like you know after after model done responding you say like okay now review the response of this model as","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1075s"},{"start_seconds":1083,"text":"as a Critic and tell what's wrong with it and uh then you use the output of that model to feed it back like okay now","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1083s"},{"start_seconds":1091,"text":"that you have this review and this original context you know refine your response um and again the the question","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1091s"},{"start_seconds":1100,"text":"here so one of the challenges this is usually fine tuned with so-called instruct data right you collect examples","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1100s"},{"start_seconds":1108,"text":"of like you know either answering questions or maybe doing some like outputting something this is where things are complicated because we don't","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1108s"},{"start_seconds":1117,"text":"have like a thought process of a of a people usually so worth mentioning instruct is probably also human anotated","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1117s"},{"start_seconds":1125,"text":"you canate Chain of Thought yes but that's not readily available but but that that's a good example is","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1125s"},{"start_seconds":1134,"text":"like having some Chain of Thought especially the ones that includes doing something wrong and then backtracking and saying like actually I was wrong let","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1134s"},{"start_seconds":1142,"text":"me go back and you know fix it because again this model is rarely seen that like oh actually I was supposed to do","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1142s"},{"start_seconds":1150,"text":"check Checkers let me continue with that and so you know train it to kind of MH","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1150s"},{"start_seconds":1159,"text":"critically review its own thought process yeah and then there's a little","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1159s"},{"start_seconds":1167,"text":"bit more sophisticated than tree of thought is some kind of Monte Carlo uh Tre search where you are pretty much","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1167s"},{"start_seconds":1175,"text":"deciding what node to explore based on kind of uh some value function that may or may not use a","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1175s"},{"start_seconds":1183,"text":"different language model and indeed you can train also like another discriminator that is trying","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1183s"},{"start_seconds":1190,"text":"to if you have tests that's outstanding right because at the end of the day like oh this is a this passes test","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1190s"},{"start_seconds":1199,"text":"yes send the gradients yeah and so I mean all of these methods don't require actually","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1199s"},{"start_seconds":1206,"text":"training models unless you want to you know F tun it to be better at doing that","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1206s"},{"start_seconds":1213,"text":"process um but in the model itself usually is not trained on on this","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1213s"},{"start_seconds":1223,"text":"processes so I think the the the interesting question I've been thinking about is how do we","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1223s"},{"start_seconds":1230,"text":"how do we go like how do we reduce the importance of the prompt right because the challenge","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1230s"},{"start_seconds":1236,"text":"is right now like you need to be very you know almost like AB testing","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1236s"},{"start_seconds":1243,"text":"different problems to find what is working and then uh there's some small things that","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1243s"},{"start_seconds":1250,"text":"actually can go and mislead the model in the prompt itself","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1250s"},{"start_seconds":1258,"text":"and uh and then the other question is you know assuming something like this the challenges like when when you're","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1258s"},{"start_seconds":1265,"text":"enrolling in different alternative like you learn something here it's but that knowledge that oh this didn't work is not actually","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1265s"},{"start_seconds":1274,"text":"available at the different rooll so uh so in the way maybe restructuring","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1274s"},{"start_seconds":1283,"text":"this not like on it is available in the sub right like if something broke here","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1283s"},{"start_seconds":1290,"text":"then all the subs know it doesn't work well I mean let's say let's make it not a token but like you know let's say","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1290s"},{"start_seconds":1298,"text":"it unroll it got stuck on doing chess right and let's say we even are able to you know score that well you're going","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1298s"},{"start_seconds":1307,"text":"here you're starting to unroll a different sub tree and you know getting again into","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1307s"},{"start_seconds":1314,"text":"chess because like this I mean statistically right you just sample a different token for example","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1314s"},{"start_seconds":1322,"text":"right well hopefully well the the better like yeah that that's not that big of a problem","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1322s"},{"start_seconds":1330,"text":"right so even if so effectiv we're saying there's a partic there's an initial State we know the particular task and let's say that","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1330s"},{"start_seconds":1337,"text":"the model because of some biases is very likely to do something wrong repeatedly right so chess is more popular game than","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1337s"},{"start_seconds":1345,"text":"Checkers so it is repeatedly going to be doing checkers right or like this idea with like hey you need to move the Cabbage to another M you know Shore but","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1345s"},{"start_seconds":1355,"text":"because cabbage and nothing else but because you you've been seeing that problem you have very high chance but as long as you have any way to test right","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1355s"},{"start_seconds":1363,"text":"uh and you enroll sufficiently many right you like unroll it 100 times and it was saying hey go with cabbage go with cabbage go with cabbage but once it","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1363s"},{"start_seconds":1372,"text":"was like oh wait a second go with cabbage and you're done mhm right and that is the correct one as long as you can reinforce it as long as","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1372s"},{"start_seconds":1380,"text":"it has any chance of getting there Tre search will help you but this this is you're saying in at this is at training time","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1380s"},{"start_seconds":1389,"text":"I'm talking about inference time yeah at inference time you like you just wasted 100 Cycles on doing the same thing so the question is like at inference time","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1389s"},{"start_seconds":1397,"text":"when you like go with cabbage go with cabbage like how do you actually force it out of the because most probably it will you","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1397s"},{"start_seconds":1406,"text":"know go with a like in general this models are like you know here's here's an average of Reddit right an average Reddit goes in this direction so like","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1406s"},{"start_seconds":1414,"text":"how do you actually sample like some more uh you know different op like","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1414s"},{"start_seconds":1421,"text":"semantically different options so and I mean there's obviously methods to do that but it's specifically","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1421s"},{"start_seconds":1430,"text":"important in this kind of approach is where if you want to go get some some different ideas","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1430s"},{"start_seconds":1438,"text":"like you're not going to you know keep hammering which again this is what we see the behavior it does it's like it keeps hammering on the same","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1438s"},{"start_seconds":1446,"text":"idea thinking that you know okay well if I do it just you know a little bit differently it's going to","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1446s"},{"start_seconds":1453,"text":"work but yeah I mean the the other approach is indeed like training it to give sufficiently","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1453s"},{"start_seconds":1461,"text":"different approaches in right so in general","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1461s"},{"start_seconds":1468,"text":"uh the experience of of using a model to critique something without specifically training it for critiquing that","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1468s"},{"start_seconds":1475,"text":"particular aspect are like at best okay right so so it will it will tell you something meaningful and it will spot","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1475s"},{"start_seconds":1482,"text":"the most obvious things uh but it's uh that's the same model yeah yeah it's","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1482s"},{"start_seconds":1491,"text":"uh so like I would say I mean these two are specifically benefiting from critic","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1491s"},{"start_seconds":1499,"text":"yeah from a separate critic it would be much better if the critic is trained","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1499s"},{"start_seconds":1508,"text":"yeah true I mean the you know you can also then train CH a Chain of Thought","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1508s"},{"start_seconds":1515,"text":"you can train it in this way right that that's kind of the difference this is like is this the inference time or this","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1515s"},{"start_seconds":1521,"text":"is a training time where you use it to improve your direction of thought because obviously as a person we kind of","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1521s"},{"start_seconds":1529,"text":"we're doing some kind of search but it's all in linear fashion it's like well let me think about that actually that's","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1529s"},{"start_seconds":1537,"text":"wrong let me go back well yeah like you can think of like if this is your tree yeah that","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1537s"},{"start_seconds":1545,"text":"that is continuation obviously tree of thought is convertible to Chain of Thought with like this right yeah so","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1545s"},{"start_seconds":1553,"text":"like so you have all of the all of these things like this and then like this and then","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1553s"},{"start_seconds":1562,"text":"maybe this this and then yeah hey here's a problem I will use dynamic programming to solve it and you like think about it","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1562s"},{"start_seconds":1569,"text":"for a while it doesn't work then you try something else and like maybe dynamic programming was not a bad idea yeah but yeah obviously I'm not literally backtracking my brain to the previous","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1569s"},{"start_seconds":1577,"text":"checkpoint that would be cool but it's but it's kind of what we I mean what generally is happening right now when you're like redoing the prompts","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1577s"},{"start_seconds":1585,"text":"right imagine you can unroll your brain to before you saw the debates that's a separate","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1585s"},{"start_seconds":1594,"text":"line um yeah","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1594s"},{"start_seconds":1599,"text":"okay um what else we can dive in here yeah I think I mean a lot a lot of","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1599s"},{"start_seconds":1607,"text":"it is then end up being in kind of just having very convenient infrastructure to run fine tuning being able to run these","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1607s"},{"start_seconds":1615,"text":"experiments so so part of it so so one interesting thing here is that um having a way to","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1615s"},{"start_seconds":1623,"text":"measure um progress is helpful for both training and at the end of day at the inference time right so if we have a","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1623s"},{"start_seconds":1632,"text":"data set uh where we have a you know definitive way of saying that is the correct result right like let's say we're trying to solve competitive","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1632s"},{"start_seconds":1640,"text":"programming right for every problem we have tests we we we know exactly when it's solved it is much easier to experiment to experiment right first of","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1640s"},{"start_seconds":1648,"text":"all you can measure progress you just run you know you have a hold out set of 1,000 problems um which hopefully is not","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1648s"},{"start_seconds":1655,"text":"in the training Set uh and you can test right and second of all things like three of thought suddenly are way more","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1655s"},{"start_seconds":1665,"text":"easily trainable because now you have feedback you can say oh I solved the problem I can backtrack uh and give reward so sort of like you know when you","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1665s"},{"start_seconds":1674,"text":"do game of Go you know that you won you can unroll positive work but like if you have if you want to","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1674s"},{"start_seconds":1681,"text":"build um when you work on general purpose development right you say well I want a website I want a website that is","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1681s"},{"start_seconds":1690,"text":"like notion but different right but Buton chain yeah you can yeah button chain you can have","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1690s"},{"start_seconds":1698,"text":"some tests uh to to test sanity right but it's uh uh it's much harder to have a comprehensive Suite of tests and it","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1698s"},{"start_seconds":1707,"text":"will be at the end the user will be interacting with it explaining what exactly they actually mean and making some decisions so I think here it sort","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1707s"},{"start_seconds":1716,"text":"of goes uh here I think one one thing is that it helps to to have a more generic agent solving both things that like","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1716s"},{"start_seconds":1725,"text":"adjacent but like let's say our ultimate goal uh is to build uh applications","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1725s"},{"start_seconds":1733,"text":"right not just let's say it is actually you want to be building like web and I think smart contracts in this","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1733s"},{"start_seconds":1740,"text":"sense are somewhat easier with smart contracts you can um you can like generally you have a very comprehensive","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1740s"},{"start_seconds":1747,"text":"set of tests for smart contracts that's that's the whole goal right but then you have a web UI for it uh and you mess","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1747s"},{"start_seconds":1755,"text":"something up in the UI and people say hey I want to swap I don't know uh 10 near for a uh for whatever for 10 W near","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1755s"},{"start_seconds":1764,"text":"uh but in reality conversion goes wrong in you swap you know a th000 and","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1764s"},{"start_seconds":1769,"text":"yeah so um so the idea what would help is if you have a combined data set where a big part of it has is testable and has","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1769s"},{"start_seconds":1779,"text":"good tests right so the the model can continuously improve and another one is sort of adjacent um and improves together right","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1779s"},{"start_seconds":1786,"text":"uh it also would be uh there's like a meta level you can teach the um ideally you would have some sort of system that","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1786s"},{"start_seconds":1795,"text":"interacts with the um with a result uh and provides feedback so that's something that um especially as you move","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1795s"},{"start_seconds":1804,"text":"to the websites uh gp4 might be decently well good at it right now but if if you think of Open Source models there isn't","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1804s"},{"start_seconds":1810,"text":"that many good if any um visual open source models right and and when they will come up I'm pretty sure they will","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1810s"},{"start_seconds":1818,"text":"be trained like on Layon or Json data sets right so it will not have that much of interacting with uis so that's","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1818s"},{"start_seconds":1826,"text":"something that I think would be uh pretty cool if we if we can have again that's this is something we're","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1826s"},{"start_seconds":1833,"text":"probably going to do in the foreseeable future but like uh a good data set and a Model F tun on it of people interacting with the applications providing feedback","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1833s"},{"start_seconds":1842,"text":"on their experience that that can then be like you know like you have this you have like a good uh critique model that is actually properly trained to","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1842s"},{"start_seconds":1851,"text":"criticize right if you have an app you like plug it in then and it's like interacting with it saying hey it's not what yeah yeah doesn't work or yeah like","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1851s"},{"start_seconds":1860,"text":"this flow doesn't make sense yeah so pretty much like what I mean what you're saying there's a whole","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1860s"},{"start_seconds":1866,"text":"yeah like training a critics SL summarization MH are the steps where okay and where you can thoroughly test","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1866s"},{"start_seconds":1875,"text":"test you should thoroughly test that's a I mean that generally should be the case","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1875s"},{"start_seconds":1884,"text":"yeah yes um I guess thinking more though how do we how can","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1884s"},{"start_seconds":1891,"text":"we upscale some of the existing open source models to to have kind of a more","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1891s"},{"start_seconds":1899,"text":"robust thinking well better data never hurts better data more yeah more reliable data I mean we discussed last","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1899s"},{"start_seconds":1908,"text":"whiteboard right that like currently the models trained on kind of random set of data at all times so a more curriculum","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1908s"},{"start_seconds":1916,"text":"where you actually find where it's not y getting it and","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1916s"},{"start_seconds":1923,"text":"uh fixing that y but on near people have been experimenting with the annotating data for a while there's multiple good ecosystem projects right now working","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1923s"},{"start_seconds":1931,"text":"towards it uh so so we will have uh a pretty large effort towards annotating a bunch of data to open","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1931s"},{"start_seconds":1940,"text":"source it later um and generally I think data notation","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1940s"},{"start_seconds":1947,"text":"is generally a very interesting uh area because uh it requires a lot of interaction between people like ideally","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1947s"},{"start_seconds":1956,"text":"if you if you want to annotate you know millions of examples of data uh and it requires people doing it you have to you have to like like build a scalable","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1956s"},{"start_seconds":1964,"text":"system where there are going to be a lot of malicious actors uh without an ability it's almost like designing a blockchain with a difference between blockchain at least certain things are","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1964s"},{"start_seconds":1973,"text":"verifiable like you know like I I can say that this message is signed by that person while in annotating data everything is subjective right so you","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1973s"},{"start_seconds":1981,"text":"have thousands of people doing subjective work verifying subjective work and obviously like for a particular you know like there's a um you there","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1981s"},{"start_seconds":1990,"text":"have there have been some large scale efforts and uh like you will never be able to document every single possible interaction right so people need to","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1990s"},{"start_seconds":1998,"text":"convert to some uh to some level of uh uh agreement between themselves right so","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=1998s"},{"start_seconds":2005,"text":"it's a uh it's it's a very problem which is very um similar to what we design in","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2005s"},{"start_seconds":2012,"text":"blockchain so we have like I would like to believe that we can do it better than","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2012s"},{"start_seconds":2020,"text":"centralized um centralized companies better cheaper and more accessible right because like yeah it can it works at like yeah fraction of a cost yeah","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2020s"},{"start_seconds":2029,"text":"because all of the marshalling of kind of money happens on chain yeah and like a lot of orchestration can be","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2029s"},{"start_seconds":2036,"text":"done on chain as well so it's a yeah like I'm a big believer that blockchain based data notation","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2036s"},{"start_seconds":2045,"text":"tools will win eventually against the centralized uh companies with like armies of salese and","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2045s"},{"start_seconds":2053,"text":"uh um yep","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2053s"},{"start_seconds":2064,"text":"um another what else was okay so another interesting idea we were discussing today more in an AI space than blockchain we will be jumping between","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2064s"},{"start_seconds":2072,"text":"the two uh at the end of the day uh there's no blockchain without","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2072s"},{"start_seconds":2080,"text":"a uh is that's something El was mentioning he was like hey how can we escape from the uh from The Prompt hell","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2080s"},{"start_seconds":2087,"text":"uh from The Prompt hell uh and uh uh one of the engineers here with us in Morocco he suggested an interesting idea","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2087s"},{"start_seconds":2096,"text":"where um so so this is your prompt and uh uh so prompt engineering is a big thing right like you you you try","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2096s"},{"start_seconds":2104,"text":"different prompts and uh some of them work better some of them work worse anyway you you come up with a prompt which is somewhat good um but then you","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2104s"},{"start_seconds":2111,"text":"have some other way of measuring uh the sucess of your model right so like let's say you actually have a data set which has a lot of tests and ideally we should","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2111s"},{"start_seconds":2118,"text":"have a bunch of those uh so then you start training and when you fit the pr into the Transformer by Design right it","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2118s"},{"start_seconds":2126,"text":"will uh it will go through the layers uh and uh it will like populate those U keys and values uh that like future","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2126s"},{"start_seconds":2136,"text":"things will attend to would be nice like to select and clim","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2136s"},{"start_seconds":2142,"text":"it think it also made it bigger for oh yeah probably it's probably like a tiny","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2142s"},{"start_seconds":2150,"text":"yeah I guess anyway um and an interesting thing here","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2150s"},{"start_seconds":2157,"text":"is that for a particular prompt which presumably will be fixed throughout your usage of the model those keys and values","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2157s"},{"start_seconds":2164,"text":"are always the same right so one obvious thing that I'm pretty sure almost everybody does today uh is you can just cach them right you compute them once","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2164s"},{"start_seconds":2172,"text":"you drop them into memory uh and then whenever you actually use your model in practice uh you do not actually unroll","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2172s"},{"start_seconds":2181,"text":"that you just start from the next token um however um right and so in inad of having a full Transformer here you can","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2181s"},{"start_seconds":2189,"text":"imagine now that you have just a bunch of keys and values and the actual layer is not executed right so like let's Transformer has four layers total right","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2189s"},{"start_seconds":2197,"text":"so this is uh this is the actual roll right so this is first layer second layer third layer and so they are","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2197s"},{"start_seconds":2205,"text":"outputting qkv qkv Etc but here before the The Prompt","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2205s"},{"start_seconds":2212,"text":"is you have those Q KS and vs uh but this doesn't exist anymore right you don't not actually compute them they're","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2212s"},{"start_seconds":2219,"text":"just constants right KV KV for each","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2219s"},{"start_seconds":2227,"text":"layer so now the cool thing is that um like you know when you dra in the model the gradients flow through","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2227s"},{"start_seconds":2234,"text":"these things right uh but the gradients update uh this block uh like MLP and attention that compute them but here","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2234s"},{"start_seconds":2242,"text":"instead now that this block is going on K and Via constants you can f tune the con con themselves like the actual vectors so for a specific problem for a","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2242s"},{"start_seconds":2251,"text":"specific problem right so with a downstream task right like you unroll it and you're like oh it passed all my tests you send the you know the","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2251s"},{"start_seconds":2259,"text":"gradients back they update uh your weights probably luras um well or you don't that's or or or even","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2259s"},{"start_seconds":2268,"text":"or yeah maybe you you fix you fix the model so the weights are the same you're just saying actually there is no actual","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2268s"},{"start_seconds":2275,"text":"like effective prompt it's just some set of vectors attend to maybe initially you computed them by enrolling transformer","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2275s"},{"start_seconds":2282,"text":"for a prompt or maybe they're completely random oh I mean if if you wanted to find answer to increase the chance of","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2282s"},{"start_seconds":2290,"text":"Ever Getting to successful trajectory right yeah but then you can find tune the things themselves so at that point assuming your prom was at least somewhat","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2290s"},{"start_seconds":2299,"text":"good to the beginning to get to any success the goodness of your prom doesn't matter anymore because the promp just becomes the perfect prompt for this","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2299s"},{"start_seconds":2306,"text":"particular problem right yeah uh and not just the at the input level right at","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2306s"},{"start_seconds":2313,"text":"every yeah it's like what whatever the key value is that whenever you need to Output the next token would be the most useful yeah which is you know contains","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2313s"},{"start_seconds":2322,"text":"instructions and whatever else it can yep and if the model if you don't actually send","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2322s"},{"start_seconds":2330,"text":"gradients right then you just like drop one prefix at another I guess probably as expensive as","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2330s"},{"start_seconds":2337,"text":"plugging in lower on top of it but yeah so that's that's something interesting relatively easy to try right so like as we like I guess the closest example","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2337s"},{"start_seconds":2346,"text":"where we can play with it is uh with the summaries that I was showing at the beginning once we have them we um there's a good chance we will try this","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2346s"},{"start_seconds":2354,"text":"yeah we can one we can low which I mean one of the challenges with in general fine tuning right that it gets better on","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2354s"},{"start_seconds":2361,"text":"the set of data that you're showing to but it gets you know worse in some other places you don't even know","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2361s"},{"start_seconds":2369,"text":"where speaking of if we try this and it doesn't work you can be sure we will say it unlike everybody else who tries fails and does not like that's something that","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2369s"},{"start_seconds":2378,"text":"should become a norm a norm yeah like if you try something it doesn't work yeah like it makes sense to say hey we tried","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2378s"},{"start_seconds":2386,"text":"this it didn't work this is how we tried it like maybe maybe we messed it up yeah yeah I mean one of the going back to the infrastructure one of the things","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2386s"},{"start_seconds":2395,"text":"we're trying to do is like how do we keep track of experiments including failed ones and then you are able indeed to have like you know hey what what what","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2395s"},{"start_seconds":2402,"text":"was this okay well this was actually this code was this data and you can see kind of the outcome and maybe indeed","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2402s"},{"start_seconds":2409,"text":"maybe this outcome is an Insight on this doesn't work maybe it's you know has a bug in the code or some other failure or","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2409s"},{"start_seconds":2418,"text":"maybe data you know is too noisy and so ideally you should be able to have this Logan be able to easily","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2418s"},{"start_seconds":2426,"text":"navigate all of them um and then you know have this tooling open for everyone so everybody can","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2426s"},{"start_seconds":2432,"text":"collaborate yep also would are we talking about the","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2432s"},{"start_seconds":2440,"text":"router I mean not yet but yeah that's going to be part part of the I didn't I didn't mention anything you unroll","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2440s"},{"start_seconds":2450,"text":"back and the context of everybody watching yeah but uh yeah few weeks","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2450s"},{"start_seconds":2458,"text":"um yeah so I think I mean the the interesting question here in general right like because is there is a prom that kind","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2458s"},{"start_seconds":2468,"text":"of like structures thing like the process of the further kind of processing data and","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2468s"},{"start_seconds":2477,"text":"generating and uh","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2477s"},{"start_seconds":2485,"text":"the the the other interesting question I mean the question I've been always trying to think through is how do we","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2485s"},{"start_seconds":2493,"text":"actually like as as a people as people will not just you know unroll unroll unroll and then maybe collect the data","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2493s"},{"start_seconds":2501,"text":"and fine tune but you know we kind of trying to update our weights somewhat on the F and so if you can imagine you know we have","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2501s"},{"start_seconds":2510,"text":"this models running with lots of people actually hammering at them","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2510s"},{"start_seconds":2518,"text":"even if it's just like okay well let's take all of the sun rolls and kind of fine tune the problem is the signal is","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2518s"},{"start_seconds":2525,"text":"not always there but uh how do you how do you actually like extract the signal and update the","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2525s"},{"start_seconds":2535,"text":"model and uh going back to the problem of fine tuning right again if you like you know there's some some distribution","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2535s"},{"start_seconds":2544,"text":"of knowledge it it has right and then like you you wanted to solve some specific problem that maybe it's","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2544s"},{"start_seconds":2552,"text":"not that good at it may like you know lose some of the yeah oh by the way this is something we were discussing today right that when","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2552s"},{"start_seconds":2560,"text":"you do Laura you inevitably forget something right like naturally if you mess up with the weights in which your knowledge is stored eventually some of","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2560s"},{"start_seconds":2568,"text":"them will forget something yeah yeah but this adds new weights but if you add new weights uh it doesn't necessarily mean","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2568s"},{"start_seconds":2575,"text":"that the performance will not be regressed right so like it is very likely and very possible that certain values in these keys and values will","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2575s"},{"start_seconds":2583,"text":"inhibit particular behaviors which you might want want want to preserve they just weren't in the data set but I think","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2583s"},{"start_seconds":2591,"text":"it is fundamentally the case that uh if this is what you find tune and not the weights then fundamentally model is as","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2591s"},{"start_seconds":2599,"text":"capable As It Ever Was because the model hasn't changed right it's uh it's extra you know strictly new","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2599s"},{"start_seconds":2607,"text":"data that you fine tune that naturally fits into the model so it's","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2607s"},{"start_seconds":2613,"text":"a yeah yeah Ilia yesterday Ilia out of nowhere stood up and philosophically","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2613s"},{"start_seconds":2620,"text":"said I want a function G such that if you fed into it a","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2620s"},{"start_seconds":2628,"text":"model and a batch of data it it yields you a new model such that the new model","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2628s"},{"start_seconds":2635,"text":"knows everything that or model knew plus the information from the batch right and feeling very smart I was like well this","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2635s"},{"start_seconds":2642,"text":"is obviously impossible because then by doing it repeatedly you can learn infinite amount of data into the finite","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2642s"},{"start_seconds":2649,"text":"size of the model but I guess obviously yeah there's no nothing in this problem statement was requiring the model to remain of the same size right and this","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2649s"},{"start_seconds":2657,"text":"is sort of that's a version of the function G yeah yeah so feeling smart and being smart is not","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2657s"},{"start_seconds":2666,"text":"quite quite the same thing yeah but the point of this is then you","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2666s"},{"start_seconds":2674,"text":"know you're able indeed to maintain the Knowledge and Skills but also potentially grow them but I guess I mean","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2674s"},{"start_seconds":2681,"text":"the other question is like how to how to you know how to evaluate this that it","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2681s"},{"start_seconds":2689,"text":"preserved well then like assuming you know what you want not to forget just have a benchmark yeah have like a universal benchmark which T the whole","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2689s"},{"start_seconds":2698,"text":"set of things you think are fundamentally important to know then if it regresses you're like well too","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2698s"},{"start_seconds":2708,"text":"far yeah except I mean if if you want to train this function G then that's a expensive process to evaluate every time","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2708s"},{"start_seconds":2716,"text":"at every step so um well you can hypothesize you can hypothesize that uh as you train","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2716s"},{"start_seconds":2725,"text":"especially like let's say we specifically train Laura right so we're not saying it doesn't forget everything M knows it just doesn't catastrophically forget important things then you can","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2725s"},{"start_seconds":2734,"text":"hypothesize that to random perturbations to the to the weights uh the level of forgetting is the","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2734s"},{"start_seconds":2741,"text":"same um for a particular distance of the new model right so like you you use some sort of distance like a k Divergence or what not and you like you know you do","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2741s"},{"start_seconds":2750,"text":"some experiments and observe how if kale Divergence is up to some value the model doesn't catastrophically forget stuff otherwise does then you know then you","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2750s"},{"start_seconds":2758,"text":"then you can cheaply say uh moreover like as you train you can like push it like like they do with poo yeah you add you add a loss pretty","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2758s"},{"start_seconds":2767,"text":"much I guess that's exactly the problem they were fighting right when you DOL yeah then were simp so small you you just yeah","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2767s"},{"start_seconds":2775,"text":"yeah yeah and this goes back I mean this can you know you can express it as a real problem where you know you have a","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2775s"},{"start_seconds":2782,"text":"policy I mean it literally is an real problem yeah and","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2782s"},{"start_seconds":2790,"text":"uh you know it's almost like there's some kind of AAR algorithm some ke","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2790s"},{"start_seconds":2800,"text":"learning um yeah but yeah I think like to summarize","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2800s"},{"start_seconds":2809,"text":"the kind of we we employing well we're building a set of infrastructure that is","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2809s"},{"start_seconds":2817,"text":"going to be open source and available to everyone to use so you can track experiments","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2817s"},{"start_seconds":2824,"text":"understand how experiments went um potentially run some on you know some shared compute","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2824s"},{"start_seconds":2833,"text":"um that's provided by some decentralized providers then you can use that inside","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2833s"},{"start_seconds":2841,"text":"this kind of agent framework to you know either actually try to solve your problem or generate unrolled data and","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2841s"},{"start_seconds":2850,"text":"that goes to crowd sourcing and then from there you can you know hypothesize some","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2850s"},{"start_seconds":2857,"text":"experiment around new data that would solve for specific","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2857s"},{"start_seconds":2865,"text":"capabilities while also benchmarking for everybody everything else and uh and test it","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2865s"},{"start_seconds":2873,"text":"out yep what ises general what does near","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2873s"},{"start_seconds":2881,"text":"do I guess when we didn't quite I I mean we have a blog post but it's I think it's worth mentioning what we do in","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2881s"},{"start_seconds":2888,"text":"shortterm right so um I guess the shortterm","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2888s"},{"start_seconds":2894,"text":"goal uh is we do want to get as far as possible in helping people build U web","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2894s"},{"start_seconds":2901,"text":"three apps right so so there's a certain uh there's kind of a chicken the neck problem where the number of web apps is","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2901s"},{"start_seconds":2908,"text":"not that large and so when the users come into the web ecosystem they're like yeah it's you know fun place but not much for me to do here uh and then","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2908s"},{"start_seconds":2917,"text":"therefore the like hypothesis is if we had a bigger ecosystem more users would come uh but you know developers like hey the number of users is not yet that high","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2917s"},{"start_seconds":2925,"text":"so um right and so web three remains relatively nich right but but if anyone","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2925s"},{"start_seconds":2932,"text":"can come and say Hey I want to build like the point here is you know having a notion which is decentralized would be outstanding right so today uh like","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2932s"},{"start_seconds":2942,"text":"notion uh you know you use notion five years later that will that will be gone right there's there's no chance of uh I","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2942s"},{"start_seconds":2950,"text":"mean there is a chance but there's definitely a chance of notion not existing five years from now or 10 years from now right so like it's like short-term","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2950s"},{"start_seconds":2957,"text":"ephemeral um stuff uh and like it would be so cool if there was a version of notion that snaps us to chain or or like","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2957s"},{"start_seconds":2965,"text":"to some sort of decentral storage solution which I know is going to survive uh but building it is very involved right like building notion in","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2965s"},{"start_seconds":2973,"text":"general is not easy uh that's why U the competition is relatively uh lacking even in the web","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2973s"},{"start_seconds":2980,"text":"two space right especially like I think the first 80% is probably relatively simple but like making it polished and","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2980s"},{"start_seconds":2987,"text":"whatnot right but if anyone can come and say Hey I want to build notion version on chain and you know some sort of AI agent will get you 90% there with","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2987s"},{"start_seconds":2996,"text":"guidance from you uh that will be a very different world right so suddenly uh you can build much better well you can","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=2996s"},{"start_seconds":3004,"text":"General faster like go to market as well right because like oh I have an idea you know tap it in boom you have a first version of the app and now you can test","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3004s"},{"start_seconds":3013,"text":"the idea with an app yeah and and every day it looks less and less like science fiction um yeah we started on this","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3013s"},{"start_seconds":3020,"text":"problem in 17 when it when it was more science fiction yes but but like now even just like observing you know the this example I gave at the beginning","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3020s"},{"start_seconds":3028,"text":"with an agent tracking code like it's uh it's un canany close to being very good right it just like it messes up every","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3028s"},{"start_seconds":3035,"text":"now and then slightly but you know the the Delta to be close is very small at this point and uh it will happen","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3035s"},{"start_seconds":3042,"text":"um I mean even on the on the simpler programming competition problems it's already doing like 30% yeah on and so","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3042s"},{"start_seconds":3052,"text":"again compared to when we started in 17 it was doing zero yeah so so so so like like generally every day uh what was it","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3052s"},{"start_seconds":3060,"text":"the website we were playing with today web Sim yeah right so like there's this website I will write it down I yeah we","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3060s"},{"start_seconds":3069,"text":"we the only reason we know about it is because it's uh it's using what open router yeah yeah so there's a website called open router and uh I guess webs","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3069s"},{"start_seconds":3077,"text":"was one of the bigger users we were like well what is it uh and the first my first several interactions with web. it feels like you you have a portal to","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3077s"},{"start_seconds":3084,"text":"another Universe from which they ship stuff to you because it just I still have no idea how this works right so like I will give you an example","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3084s"},{"start_seconds":3094,"text":"uh we had access to Deon like to the actual one from cognition and I wanted to build a game of rock paper scissors with spaceships and it took me uh it","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3094s"},{"start_seconds":3103,"text":"built something eventually uh like it wasn't a complete flop but uh it took me a good part of a day uh you know following up and then in websam I just","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3103s"},{"start_seconds":3112,"text":"said you know rock paper scissor with spaceships and it just loaded in 15 seconds a fully functional game with much better animations it","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3112s"},{"start_seconds":3121,"text":"just or or like one thing given the example of of U chess and checkers I said build me a game of checkers where","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3121s"},{"start_seconds":3128,"text":"every piece is a chess Knight and again 15 seconds later did exactly that I didn't try with Devon but it's it would","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3128s"},{"start_seconds":3135,"text":"be an interesting uh interesting thing to see so yeah uh but the example is every now and then in the modern world","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3135s"},{"start_seconds":3144,"text":"you wake up to the news of something that yesterday would have felt like science fiction like Sora when it came out came out right it was like I mean it still didn't come out technically","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3144s"},{"start_seconds":3153,"text":"but right like yeah something completely impossible so I think the the world in which you just go and say build me this","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3153s"},{"start_seconds":3160,"text":"smart contract on near with the web UI and middleware is pretty close uh and there's a good chance it will happen","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3160s"},{"start_seconds":3168,"text":"before General P was developer just because the scope is so contained and smart contracts are much smaller way easier to test right to like some sort","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3168s"},{"start_seconds":3175,"text":"of research uh is pretty likely to to to","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3175s"},{"start_seconds":3183,"text":"yield good results and because we have access to such a you know uh like the data wheel where we can you know very","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3183s"},{"start_seconds":3191,"text":"quickly spin up projects where people who know near and like near and want near can very quickly you know put","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3191s"},{"start_seconds":3197,"text":"together a bunch of good examples so yeah I'm quite optimistic that that one will work out relatively short term um","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3197s"},{"start_seconds":3206,"text":"but longer term as the song as this beautiful song Don't suggested longer term that's like you know just a","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3206s"},{"start_seconds":3214,"text":"stepping step towards having a whole ecosystem like ideally yeah everything","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3214s"},{"start_seconds":3221,"text":"data sets uh inference um servers training servers um data I said data right yeah","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3221s"},{"start_seconds":3230,"text":"still data sets uh pretty much yeah like I mean experiments dat yeah it should be all I","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3230s"},{"start_seconds":3238,"text":"mean there's kind of a data layer kind of model experiments infra layer and then there's an application","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3238s"},{"start_seconds":3246,"text":"layer right and so in data layer you have you know crowd sourcing you have you know data","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3246s"},{"start_seconds":3255,"text":"acquisition yeah data contribution Network yeah you have some synthetic data","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3255s"},{"start_seconds":3264,"text":"probably and uh some form of curation as well and there's more stuff here right obviously and then here you have compute","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3264s"},{"start_seconds":3273,"text":"kind of broadly you have inference training the models themselves kind of","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3273s"},{"start_seconds":3280,"text":"customized models and then application layer you kind of have all all of the fun stuff that people are building and so you want","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3280s"},{"start_seconds":3289,"text":"kind of the full stack right to not depend on you know this guy is going to just like whatever their API was down for example and like all the application","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3289s"},{"start_seconds":3298,"text":"stack wasn't working mhm and similarly you want you know be able to access data","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3298s"},{"start_seconds":3304,"text":"um or both cheaper and kind of in a different way where you know creators potentially getting rewarded","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3304s"},{"start_seconds":3313,"text":"versus it's all you know free freefor all um but again this is kind of more of a","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3313s"},{"start_seconds":3322,"text":"you know this should be a better stack to build on not just hey this is like you know","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3322s"},{"start_seconds":3330,"text":"because of values we want this it's more this is a better way to do it that's why we're building it and it's going to be","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3330s"},{"start_seconds":3336,"text":"you know cheaper more accessible there can be multiple versions of different uh pieces because they're actually you know","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3336s"},{"start_seconds":3345,"text":"having a healthy competition in incidentally the the existing stack is not that easy to use right I don't know","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3345s"},{"start_seconds":3352,"text":"what's the reason for that uh but like you know uh setting everything up even even even installing","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3352s"},{"start_seconds":3359,"text":"python on the Mac on the Mac yeah is not is not that trivial right so so it's like a rare situation where you can","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3359s"},{"start_seconds":3367,"text":"build a decentralized stack uh and you will not be fighting an uphill battle against a very you know like mac level","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3367s"},{"start_seconds":3377,"text":"you know everything yeah very streamlined process of the centralized uh Solutions so so I think that's a well because in centralized it's also all","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3377s"},{"start_seconds":3385,"text":"like kind of different right like you're using like computes from AWS you know yeah the influence is","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3385s"},{"start_seconds":3392,"text":"here the like the data the data in general you know you need to talk to salespeople you need to get API for you know $66,000","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3392s"},{"start_seconds":3401,"text":"a month Etc right so you generally don't have like a good setup unless you're in","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3401s"},{"start_seconds":3408,"text":"one of the big companies where this is all kind of given as a as a one thing yeah so yeah I mean one","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3408s"},{"start_seconds":3417,"text":"of the kind of you know this experiment tracking and in general being able to ideally spun experiments in this setup","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3417s"},{"start_seconds":3427,"text":"is something that like you know internally some of the organizations have because it was worse for them building when you have like you know thousands of people together working but","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3427s"},{"start_seconds":3436,"text":"it doesn't really exist in open source because every every team is themselves kind of are having some version of that","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3436s"},{"start_seconds":3446,"text":"and so you can have like a shared version that everybody's integrating with if you work in open way in open","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3446s"},{"start_seconds":3453,"text":"source way","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3453s"},{"start_seconds":3459,"text":"yeah but uh I mean the the in the blog post right um kind of the idea that the","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3459s"},{"start_seconds":3466,"text":"I developer is also leads to yeah researcher yeah researcher and this is kind of I think there's a lot of the","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3466s"},{"start_seconds":3474,"text":"conversation right now about this I I think I just saw another thread uh by Jeremy Howard about this that's like","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3474s"},{"start_seconds":3483,"text":"like the current LMS are just you know just giving you statistically most probable thing and so yeah clearly it's","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3483s"},{"start_seconds":3491,"text":"not very good at like exploring different options um yet and so but you need you","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3491s"},{"start_seconds":3499,"text":"for sure will need some kind of logical reasoning and this is where","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3499s"},{"start_seconds":3508,"text":"the formal systems come in yeah um I mean you will need some way of yeah kind","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3508s"},{"start_seconds":3516,"text":"getting un stock and figuring out Alternatives like sufficiently different Alternatives um so like this kind of","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3516s"},{"start_seconds":3526,"text":"Search exploration which again we already see we kind of need even on the","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3526s"},{"start_seconds":3534,"text":"simpler examples worth mentioning this is brainstorming we have no idea how to build the a yeah but this is like I mean we we see","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3534s"},{"start_seconds":3544,"text":"this things are needed even fori developer right and they're clearly need for researcher um and","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3544s"},{"start_seconds":3556,"text":"so the I think generally like","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3556s"},{"start_seconds":3564,"text":"um making agents creative do agents in general","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3564s"},{"start_seconds":3571,"text":"um some influential figures on Twitter say uh it will not be","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3571s"},{"start_seconds":3577,"text":"LMS uh U whether it will be LMS or not hard to say right but but like we need we need","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3577s"},{"start_seconds":3585,"text":"creativity um as a yeah someone someone said that you","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3585s"},{"start_seconds":3592,"text":"know like FMA spent 20 years of his life learning math and then I don't know how old he was anyway the rest of his life","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3592s"},{"start_seconds":3600,"text":"creating math uh so it's something that uh uh no idea yet how to go in that direction but that's like a North star","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3600s"},{"start_seconds":3608,"text":"uh to make sure that you know the eventually the the systems we create are capable of advancing science advancing","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3608s"},{"start_seconds":3616,"text":"science and Humanity yeah but you know starting to focus on","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3616s"},{"start_seconds":3623,"text":"these two sides to make sure sure we actually making progress how we doing on","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3623s"},{"start_seconds":3631,"text":"time it says 32 seconds oh 1 hour 34 okay yeah","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3631s"},{"start_seconds":3639,"text":"yeah all right well hopefully this was good and educational um as always you know we're just kind of brainstormed and","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3639s"},{"start_seconds":3647,"text":"discussed some of the things we're working on and we're going to continue doing that especially when we are in person um and as well posting more stuff","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3647s"},{"start_seconds":3656,"text":"on the blog and uh releasing some of the code and experiments from this so stay tuned and uh thanks for joining yeah","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3656s"},{"start_seconds":3665,"text":"hope it was fun yeah","url":"https://www.youtube.com/watch?v=BbrAOE8_BR4&t=3665s"}],"youtube_url":"https://youtube.com/watch?v=BbrAOE8_BR4","url":"https://near-pedia.xyz/videos/BbrAOE8_BR4"}