{"id":"7kc4Mv75u5Y","title":"NEAR AI Office Hours #4 Whiteboard Session with Illia Polosukhin & Alex Skidanov","description":"Illia and 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\nAI Data Optimization to User-Owned Internet\n\nIn this episode of AI Office Hours, Illia and Alex talk advanced AI research topics, including optimizing training data through curriculum and active learning. They explore the future of user-owned internet and the importance of open-source models, discussing potential vulnerabilities and the need for formal verification in AI development.\n\nTakeaways:\n• Optimizing AI Training: Leveraging curriculum and active learning can significantly enhance the efficiency and effectiveness of training large language models.\n• Open Source AI Models: Open source models, while limited by current resources, are essential for fostering innovation and maintaining transparency in AI development.\n• Symbolic AI for Precision: Integrating symbolic AI with current models can improve reasoning accuracy, providing a hybrid approach to problem-solving.\n• User-Owned Internet: Building a decentralized, user-owned internet infrastructure using AI and blockchain technology can ensure security and data ownership.\n• Importance of Formal Verification: As AI systems become more complex, formal verification of code and smart contracts is crucial to prevent vulnerabilities and ensure robustness.\n\nChapters\nIntroduction and Overview (0:00 - 1:18)\nAI Training Optimization: Curriculum and Active Learning (1:18 - 18:36)\nDiscussing inefficiencies in current AI training methods (1:23)\nVisualization of data smartness and learning (3:11)\nActive learning and data scoring techniques (12:06)\nOpen Source AI Models (18:36 - 33:46)\nComparison of proprietary vs. open-source models (21:01)\nFuture-proofing AI development with open-source (27:06)\nBuilding an open-source research lab for AI (36:02)\nSymbolic AI and Formal Verification (33:46 - 57:03)\nImportance of symbolic AI for precision (39:00)\nAI developer and product manager workflow (42:09)\nEnsuring correctness with formal verification (48:03)\nBuilding a User-Owned Internet (57:03 - End)\nRewriting internet infrastructure with formal proofs and cryptography (58:02)\nThe need for a decentralized, secure internet (1:00:00)\nConcluding thoughts and future directions (1:01:08)","channel_name":"NEAR Protocol","published_at":"2024-06-06T00:00:00.000Z","thumbnail_url":"https://img.youtube.com/vi/7kc4Mv75u5Y/hqdefault.jpg","transcript_text":"I feel the custom feel hack the world and make it tight blockchain flows near the one we're close to won't stop us now together we'll redefine the when near we'll find the flame empowered Minds rise and St use your own AI we'll explore hold the future hand in hand in this Tech driven Wonder okay well welcome to a special office hours we actually are in person here in San Francisco is Alex and so we're going to do a wideboard deep dive uh we're going to improvise so uh feel free to kind of enjoy the show uh hopefully it's educational we'll try to go a little bit deeper than usual into some of the topics but uh yeah welcome and let's start uh so I tweeted this few different hypotheses that I think uh will be pushing the envelope on Research in Ai and so we wanted to kind of launch off that to also discuss some of the other topics we're working on so maybe to start um so one of the kind of core challenges with Z car models is that because they're trained on you know massive amounts of tokens the um it takes you know extremely a kind of lots of compute to actually produce that produce a model and train it and obviously the model itself is not training the most optimal way right it's you know getting random pie of content like it's usually chunked kind of in the middle of a document sometimes like in the second EP it sees well not e but yeah early on it is like yeah a proof of ran hypothesis yeah like something about medical stuff you know some article about royalty and then closer to the end you know like it's it's being fed something from forchan and it's not not contributing as much into yeah so so kind of there's like few of the research topics called curriculum learning and active learning around how to prioritize which data you feed to the model so wanted to discuss that and kind of Alex been thinking about it for a while yeah yeah I I like this um uh visualization of that which is by no means scientific uh but uh imagine that you plot your um you have some sort of uh omnicient Oracle who looks at your entire data set and for every um for every document uh it assigns a value of how of how smart um that document is right so so like uh and how much intelligence it requires right and then we will assume that smartness um because the data set is sorted right if you look at and and this is this is your data set as you're feeding it into the model right so this is time uh and assuming the data set is somewhat uniformly shuffled if you look at like uh distribution of smartness uh during different times it will be more or less the same right so like some very smart data uh on the very top slightly less smart data and this is like ready the rest of Internet yes yeah um now um the way I like to think about it especially if I find the Eraser is um um let's say that exactly half of the data uh is only useful um when it comes to learning uh grammar of English and and has no other usefulness to it uh and then half of that half of the rest so this is 50% not very useful 25% slightly more useful than 12% even more useful uh and now I will exaggerate it significantly but it will it will sort of deliver the point imagine that in order to even start grasping usefulness of this half uh you need to to go through half of this data set or like in general well I think in maybe two way to specify right so when you're training these models like generative pre trining right what it does it's trying to predict the next token and so in the beginning like the basic grammar rules is enough to like reduce the loss right yeah so like if you understand at least how English senten are constructed and like basic vocabulary you know that like an article goes before word Etc right that already reduces loss right but obviously as you train more and on more complicated things you need to actually start reasoning and making inferences and remembering facts right so in this case let's say let's say that this is like a uh that this 25 % start involving basic reasoning and let's say that you need um so that will be an exageration of the actual percentage but let's say you need to uh this 50% of the uh lowest level data set um consumed before you even start comprehending reasoning right so in this case this data this part of the data is only useful here uh and up to this point you were only using it to learn grammar which is you know not the most efficient way of using this particular data like it would have been much better uh to swap them right so like to move uh this block of um data which is only useful From perspective of learning grammar here and this that gives you reasoning here right in which case you will double your effective you know effective learning from text and then here you you have something even more uh smart right but then it takes you half of this time to start comprehending it and so here you only use 1/8 of this data set and then you only use 116th but but the actual the actual coefficient is 150% right it's it's less than that but the but the point here is that somewhere on top you have like let's say you want to you want the model to be good at uh reasoning about um you know general relativity differential geometry tensor calculus all those things right and let's say that this is is like level seven of this diagram and so you do have I mean people do talk about it talk about it on the internet write the textbooks people talk about it on Reddit in specialized subreddits people talk about it on like physics stack exchange and this is all like in here but you only use like one over you know 256 of this data set uh and with each consec with each consecutive level you use exponentially less and less of the data usefully right and if you were able uh if you know if you had those labels from the Oracle and you were able to reshuffle your data set uh you would be able to be you know like you wouldn't even need new data which doesn't exist yet um you would just use whatever you have there efficiently yeah and so as a humans right we're also not getting thrown you know an article of theorical physics an article about you know uh linear equations and an article about uh how I don't know how different uh species are related to each other before we learned you know how to read to read yeah and kind of maybe basic math and so so the idea is like how do we design a system and the thing about this is actually this doesn't need to modify any of the architecture of the model this is just the training procedure that needs to be modified and so the question is how to get that Oracle right that given a document returns well you called it smartness but um nonscientific smartness score right and the thing is like you probably don't need this to be you know 100% like obviously I mean it's really hard but if you get get a you know reasonable correlation um you can use this and so one interesting point is that if you take you know any of the document anywhere and you take existing model right let's say we take Lama 3 and we run that model on this document if the document doesn't require kind of the lower it is on this level the more that uh that model already is good at predicting that's kind of what we train them to do right and and so you actually can use the loss of uh let's say llama which it give it gives you a predictive loss uh if you score the document as a way to approximate this now it's indeed not always going to be like you know completely correct but yeah because complete garbage will have a pretty pretty high loss yeah and so so there are a few things that you need to um kind of correct for right so there's um if you look at any document and this is kind of for those who are actually training this or or looking at this models um you can imagine let's say it's it's a news article right and there is like you know a title and then there's an author and then there's like something and then there's an ad probably and there's something else right and so the reality is even in in a general article there's like already a bunch of stuff that actually almost impossible to predict like if I give you this title and ask you to predict an author uh like even with those intelligence uh it will be really hard I mean maybe if you learn who works where in which Media Company and which topics they cover but like generally there's already parts of the document that you ideally want to kind of ignore so so there's another part part and again similarly if there's a document that has complete garbage right which sadly internet is full off literally just Spam of like you know comma comma comma like some words repeated somebody like optimizing for SEO there's like a bunch of just you know random words like you want to figure out how to filter those documents anyway right this is definitely not helpful for anything as well as on a document level there's parts that are completely unpredictable and again the model will output so there's a kind of if you look at the Sloss and the correlation with smartness um they're going to be kind of a a distribution as well right so loss smartness um so really high loss probably low smartness and so I mean I mean I haven't plotted this but yeah you know the meme yeah except this is Spam but yes yeah so anyway so so there's a kind of some procedure we can use here now the other way which is more active right so this is kind of you know you score this data ahead of time now the problem is this is also computationally really expensive because you literally need to run your model over the whole data set exactly what the training yeah is but you need to do it once presumably to sort this now you can train every model using that function right now there's an interesting option which is can you do it actively right so as you're training a model right so right now let's say you have a model you are you know feeding some data it you know uh you know you compute the loss and you kind of back Pro like update the model from this um and so right now kind of generally I mean this is this is another um hypothesis so the idea of active learning is you instead of um uh kind of feeding just random pieces of data you actually try to find like you kind of ask the model what would be what would be the most ambiguous for it right so in a way you can say like oh well if losses let's say three you know what would be the inputs and you sample the inputs from this right and then you kind of uh now generally Active Learning is used with human feedback right and so for pre-training it may be not very effective because like ideally you have you know you have a question for example uh you know question answering system or something similar and you are uh uh trying to find you know what questions to ask to like what pairs of question answers to add to find to improve the model uh but indeed it's a bit hard now can you do the if you have like this big data set right so lots of documents uh pretty much can you cheaply find something in it that is you know the most promising for you to update um and so you can have like a smaller model that's one option uh you can have if you if you're training so right now you know this is let's say 15 trillion tokens which is a lot if you're training on a smaller subset let's say you just take Wikipedia for example only right which is uh 2.5 billion tokens so yeah right now it's trained on 7,500 vikia size of content uh so you can train multiple times on it right so you can actually sort data um you know from a previous ug based on how you model scored to feed it in a in a in an order that's more useful for it to update because the other thing is this models right now need to like back propop thousand times something to actually memorize it effectively so like a simple idea would be you know instead of grabbing a random document you grab random 10 compute the loss and train on the five that have the the low Med the median yeah so that presumably drops the stuff that you understand very well which might potentially bring you back and removes the stuff which is either spam or something beyond your comprehension yet so the only problem now is that uh so early on that would make sense right like if you hear like early on you grab some documents uh the spam which we don't even like let's say there's also spam here will have high loss something you don't understand will have high loss you will be picking this ones what you want to happen is that later down the road um you kind of want to be able to find out because you were running into them right like as you were moving along you were running into smart data so it would be nice to somehow in yeah fetch it later yeah that like this was something I didn't know how to do before minus spam yeah but hope hopefully you remove the the spam ahead of time right spam can be potentially detected like by like you know like to high or something like this yeah yeah I mean and yeah I mean one option is you use like a down sample model like you know 1B model for example to score data as prescore data find things that are like complet you know not even gr like grammar predictable and then filter out that okay sounds like something to be done yeah and then the other interesting thing I mean this is just like a thought that when you're training right right now you're applying laws kind of evenly across the whole document right and so there are places which are like either unpredictable or really hard to predict like a first token in the document right like I you know tell me how's this document going to start from you know a start of a document token it's really hard so in a way you want to kind of also update a little bit how the loss is propagated based on in general how predictable can this be in a training procedure yeah I think I think another interesting completely tangential to that idea here is that um like like obviously for the lower levels of smartness uh you have to mine internet MH uh but like you know if you go to llama today and you ask her something about Cuda it will give you something pretty meaningful right so like it's grasping at least some concepts of Cuda but naturally the amount of data in the internet for Cuda was much smaller right so so the higher the level the more tractable it becomes to just uh you know grab humans and annotate data from scratch right so instead of uh trying to find a way to autom atically identify smart data you can create smart data and put it at the end or I guess I don't know F tune on it yeah I guess that I mean relevant to this question is we want the model to understand how to you know how to read this how to reason about it they don't always needed to remember all of this right because the other part right now with this models we're trying to actually get them to remember every single fact and you know thing that happening on internet because that would be the low like then at the you know training time it will get the low loss uh on predicting next token and so the interesting question there is like how can we modify the training procedure in such a way that um it's not it doesn't need to remember it can actually use something in a context to uh for for example better predicted documents and so one idea that I had actually is as you you know starting to train on documents in in kind of these places is to find relevant information that may be at uh either previous levels of smartness but that contains kind of some information and putting that in a context before predicting right so let's say I mean you're reading a Wikipedia article about you know linear equations maybe you want to put in well maybe linear equations is uh I don't know some fact like you you're generating like a news article about Nvidia and so you want to put in maybe like Nvidia Wikipedia article and something else so you want some retrieval at training time at training time yes because the big benefit of this models is that there you know can use the context to retrieve information right and you just only you know back prop loss here and use this more just for context y well shall we talk about how easy it is to train model from scratch today in 2024 no you can okay so so so let so there's another um somewhat fundamental uh uh area of thinking which is which has nothing to do uh with the science of training models and more about the current state of the world uh and um let let me let me use my favorite analogy which is uh let's say that today in 202 four I want to create a document yeah like IL and I are preparing for this office hours which we absolutely did not as you can see but let's say we were you know put more more effort into it and actually try to prepare then we would create a document right and we would have multiple ways of putting it together right one of them would be and that's likely what we would have actually done is to use Google dos however I might have ahead of time say well that document I would prefer to preserve it for posterity like you know when I'm 90 years old I would like to look at that document preparing for the office hours and Google Docs is not the way I mean I can probably save it but right and so I would open up my copy of Microsoft Word uh and prepare it in the Microsoft Word um or I can say well I can prepare it my in Microsoft Word uh but then by the time I'm 90 Microsoft could be gone Microsoft Word is you know not supported anymore uh and nobody will be there to pick up the torch because Microsoft Word is not open source right so instead I would open Liber office that escalated quickly right and so now let's let's see how much of Open Source each of them is right so Google Docs is not open source uh and also I have no ownership of data of anything that is stored on Google dos right so no ownership data belongs to and no ability to Tinker with the with the Google Doc binary like you know if Google tomorrow says now you know you want these buttons to be swapped because someone needed to get a promotion at Google I can do nothing about it right Microsoft Word not open source for sure but they do get to have binary right I have I have a binary of which I can make no sense at all uh but I have it locally right so if tomorrow Microsoft Word says hey instead of toolbar we will have a ribbon which they did to us I can say well I'm not going to use your Ribbon I will use the old Microsoft Word right so um well I guess it's it's separate from ownership of data right like I can Tinker with it no tinkering there's no tinkering and there's like there's like limited tinkering right well it's more like you're not for wor thing to to upgrade to yeah yeah but I guess also importantly I have ownership of data like you know nobody nobody's looking into my documents I'm editing locally actually don't know if Microsoft Word is that's a the time have changed yeah tinkering okay and Libra office it's open source I mean it's as open source as it gets I can download the code I can you know if I don't like something about it I can change it I own my data own my data I can Tinker as much as I want so uh even though in Microsoft Word I can tinger a little bit with the binary I can I don't know I can write a plugin for Microsoft Word I don't know if I can I own my data nobody would ever say Microsoft Word is open source right uh so now if we I don't know how far this goes but let's say this is more space if we compare it to the state of the artificial intelligence today Google Docs would be open AI or like entropic any anything that goes through the API right if I if I use open AI like a good example with open AI uh is when they introduced this button that says you know we're not going to save your history uh and then you know uh in my particular instantiation one day it just fli it back because they changed the control and the new control by default was again saving my history and saving history means staining on my data right so like open AI I have no no control over how my data is being used at best I can you know they can promise not to use my data and I can trust them uh and then llama and mistr would not be in the last row for sure right because in reality with Lama in what we get is we get this binary that we cannot make much sense of right so llama would be um more like Microsoft Word right so I own my data everything I do locally more like stable diffusion uh if I locally do inference I know for sure that data doesn't go anywhere it stays with me I can do some limited tinkering I can f tune do laora um but it's not open source right even though they called open source models they are not open source because I cannot like if I want to change something in them fundamentally that is not something I can do training Loop is not open source well you don't actually know what went into it yeah yeah we don't know the data that went in we don't know the training Loop right so like if I say well I like clama but I want I wanted to understand images now so how about I change the training uh distribution a little bit right or like apply more filtering none of that is possible right and like and also the flip of side of this is this paper on sleeper agents where you pretty much train the model to have a specific behavior at a specific instance and that doesn't show up at any valve that or uh you know persists over fine tuning yeah and so like there can be like things that are you know when the condition triggers it installs for example npm package in in your python when it's get cold that has a malicious uh for example code yeah okay and so now well I haven't been following the space very closely recently but the only model I knew as of like a month ago wasmo which is Allen's Institute of Technology right yeah so that one is you know you can reproduce it right you can download everything uh and um the problem is that here's the problem that as the as the um industry goes forward we need to make sure that we have open source models that are competitive with the stateof the art right so so clearly you can be in in three different regimes one is that where even open weight let's call them open weight even open weight models are not competitive with state of the earthart right this is this is complete dystopia uh that effectively means that uh and here I make an assumption that AI will actually drastically change the way we experience the internet and will become you know the fundamental part of our interactions with it then we at a mercy of two companies we always go through their Gateway yeah and we always go through their Gateway which which has been the case in multiple other areas of Our Lives social networks you know search Etc and that is not great um second one is open weight models are competitive this is strictly better right so then um um at least at any at the point of time when open weights models are better uh there's a possibility for competing products the problem with open weight models is that there's no uh it is not it is not future proof right so so so these guys they they will be um innovating right uh and in the moment we have open weight models and in the moment we have those entities who publish them aligned in terms of incentives with the with the big guys uh sorry they aligned in terms of incentives with the world to publish those models uh for various reasons like for example for some of them the reason could be that they get attention today because attention is important for some others uh I can only guess monitors for Facebook but I think Facebook is generally has been known for open sourcing stuff because then people Tinker with it and Facebook gets it like like you know if there's llama and and GPT and Lama is open source then Facebook gets a whole world working on their stuff for right they get v l and whatnot for free effectively and that's that's like a very nice Synergy right people get open weight models Facebook gets a development on it but it's not future proof because if at some point incentives change right so suddenly you know like hypothetical I don't want to name any names but like any company that was interested because of the attention the moment attention is not as valuable anymore and they want to start generating profits the license change um don't know if it ever happened yet um right but if you have an open source model it is future proof right so let's say for the company let's say let's say alen's Institute of Technology says tomorrow well we don't want to work on Alo anymore it's not it's not anymore you know meaningful for us anyone can pick up from where they've left and train a new model right the Next Generation so it's it's very important that there is a open source model um well I competitive it's it's it's not on itself future proof right because the future proofness here comes from the flywheel of you know like they have a model they put it into a product that generates Revenue yeah that they put back into Data compute people people as product which then improves the model right so so I think an important yeah there there was a couple of steps in my head that I forgot to to put on the Whiteboard but in this case for the product to exist right so we have the model right let's say it's Lama uh and then we want a product and let's say the product would be even just like you know a Char GPT alternative right so for a company to truly commit into making it to be a very you know appealing competitor they need to be that company needs to be certain that they will not die out if the model they depend on gets discontinued uh and suddenly uh you know they once CH GPT switches you know to GPT 5 uh they will be left behind because there's no GPT 5 for them to switch onto and there as a company right so so like if you want to build a future proof company it's much better to know that the infrastructure on top of which you're building will be evolving yeah but ideally what you want is you want kind of this system to exist in open source where more people data and compute are being kind of gathered to build a better model that and not just one but many products are built around which then fund development of this y hence the economy of blockchain like you need some kind of platform that is encouraged by the fact that more people are building on top of it versus competing kind of product that they pretty much trying to eat everybody else which is generally how any single profit company needs to exist Y at which point of course it's very desirable for normal people to be able to interact with the blockchain for yeah you were going somewhere no more or less train on the training Loop no not no I think I think that's more or less the the stage I wanted to build up well I think the I mean the question we kind of mentioned that I think last of is ours around like training Loop um because right now indeed like I mean I think this is there's this and then there is a megat Megatron which is from Nvidia and uh I don't know if there anything else and uh I mean all of them obviously require like you know data center level of compute I mean you you can start with like obviously smaller uh setup but if you want to train like I mean 70 70b model which seems like you know reasonably powerful um this is what 3 months of so the original Lama paper reported I think something like 29 days on 248 A1 100s mhm however uh my limited knowledge of those training Loops suggest that that was an extremely optimized code which not everybody has access to yeah so so I think for more realistic uh well I mean the question that would be like around couple month yeah yeah the question is can we in open source right the benefit we have is that we can have more people actually looking at it and optimizing it yeah um and so again I mean we briefly discussed it but it was you know you can you can imagine like a training Loop that is open source you have a kind of few in a way plugins right on like on the model on data sampling procedure maybe you have the trainer SL optimizer that are kind of easy to plug in and so the kind of the biggest benefit is right now as this you know the researchers who are like hey I have this idea right we just discuss something I could code it up you know and ship it tomorrow I just don't have the this level of compute yeah and so ideally we would have a place where they can come in and say like hey here is a pretty much full request with some modifications train it on maybe like 1B first um if it shows improvements or shows interesting results then you know scale it to let's say 7B to 70b right so um and then you kind of have you know behind this you have few clusters of compute that are Al so there is like a um it's like an open you know AI research lab where actually people can come in and suggest no it's like open end of token AI not open AI as a token yeah yeah you know token you need a better tokenizer yeah what is so this is what is it it's like 2K per hour right which is for one month where is my calculator I think it's uh six million you can talk about some other thought in the me time yeah so I mean this is kind of something we've been uh thinking about like to to build an open source um all right yeah so you can imagine one B you can probably train for way cheaper well let's just say it scales linearly for Simplicity right that that should be easy to divide by 70 it's like2 yeah yeah 0 Z to 20K yeah that that becomes tractable yeah I mean you probably start with like even smaller but yeah I think it should be cheaper right now I think 7B you can train with like 128 for a week or something so anyway um yeah so there's like a I mean in the way this is like a you know a research Grant to someone and then you kind of experiment with a few of those and then exp explore probably combination of this for a bigger model yeah so as as a community we should build this to truly be the home for use your own in open source AI yeah sounds great all right so that's solved clearly all um do we want to talk about symbolic stuff sure I mean we don't have much probably to say about it but uh so so I guess like a very short bit of trivia is Ilia and I when we met uh one of the things that we shared in common was that both of us during our master's days were trying to use symbolic AI for whatever and so for program synthesis program synthesis yeah so so so both of us kind of dream about still about using it one day are we any closer to knowing how to do that no but but like do do you want to draw your diagram with the uh user and the system you mean for for the developer oh yeah yeah for a developer but I think it applies to anything yeah so well maybe yeah there's like few few things but specifically I I actually talked about this at Town Hall earlier we working on this thing we call the developer right so you can assume there's a person and they want to build some application right so there's an application here and you know without loss of generality we assume it's a vi3 application yeah right which has you know some kind of smart contract some kind of uh you know we call middleware which and then some front end and so we want an AI who will you know go from build uh I don't know freelance freelancer Marketplace that will actually give you a full application now if you imagine on this side you actually have a human that is like get this I mean they'll assume something right they'll they can build a freelance application but it's probably not the one that user wanted right and so you first need kind of a process which you know um usually somebody like a product manager or uh or the person sagee which is actually write out exactly what they want to be built right so we call this a specification and so the reality is though there's a lot of detail that this specification should go into to actually fully describe this and so uh you actually have a kind of interative process where as the user you know hey build a fencer Marketplace let's imagine there's we you know we call it aipm product manager who writes a specification for you you're like well actually that's not exactly what I wanted I actually want maybe to you know I want people to have nft badges for every time they complete their task right and so you kind of update and add features and maybe modify the user flows and so at some point when you're you know content with the specification you then say okay let's build it and so that's get shipped to proverbial AI developer who is then write the code now potentially still coming back to you to update the spec as it goes yeah it potentially either comes back to you and you can update this back right so um in a way you can almost model it as if you had like a Dev team behind the chat and you in a group chat with them and uh you had two outputs for them one is documentation of the project and the other one is the app itself and you can kind of iterate with them on this uh that's why there's potentially also like AI designer who you know helps with the front end Etc so that's kind of you know we we building this system out we're going to be opening up the um opening up this part to play around for the community in a bit um so you can actually like write out your ideas see the specification iterate on it and then we're going to start leveraging already what we have but Al people to develop this and kind of iterate on and to see how best train the developer to do this but one of the biggest issues with all this and and kind of generally with the current state of llm writing code is that they're not very good at it at like at a you know sufficient detail of reasoning and and logic not very rigorous very rigorous yeah and so they and with smart contracts it's especially important that you are very rigorous and you don't lose Lo people's you know money assets and uh so you have set of you know conditions that you want right now the way it's done is it's sent to an auditor which is another you know person who like reviews the code finds one you know issues and then kind of developers iterate on it but that is I mean like usually a very cumbersome process and Auditors also need to know the specification and they also make mistakes and they also make mistakes like if you have an AI auditor here yeah so but even human Auditors like they make mistakes yeah they miss things right I mean how many smart contracts got hacked which had an audit so the question is can we have a some form of guarantee that this smart contract is indeed correct up to a specification plus you know some set of conditions and variants around not losing money not changing ownership of the uh of the assets without the originating user uh confirming it Etc yeah so what would be interesting here is if uh imagine that there is a uh highlevel um language on which or rather I actually don't know if there's any formal proof systems for stability but they do exist for us for example right so people people for many years now people have had uh ways of formally proving that the code does something you expect from it but you can probably go even one level higher and say so so now uh there's a specification language which is uh not uh which is a uh an actual symbolic language which can be you know reasoned about um using some sort of proof system right and so when the user comes into the system and says I want to build a freelance Marketplace that is natural language so there's an llm but the the llm the output of the llm is not in the language space as it is today but instead it uses this symbolic um like sort of let's called domain specific language in which it builds some sort of understanding of what the user wants uh and for for this system for this DSL we want two properties property number one is that it can be uh converted to a Smart contract SL middleware um in in uh like without using any leaky llms right so like there's there's an actual routine that converts the DSL into the smart contract mware and similarly there's a way to convert it to language uh which conveys exactly what the DSL is right so like what you want to happen is that you go to the eipm eipm builds this DSL and then it communicates back to you saying uh on the high level this is what I think you want and then the user can say well either yes that's exactly what I want or no that's not quite what I want uh but once they done uh what a IPM has uh is something that user signed off as what they want which can be directly translated into into code um the the problem here is that you want it to be sufficiently low level uh to be able to convert to to the smart contract middleware but sufficiently high level to communicate back to the user so so potentially uh this like single step is not quite possible maybe it requires well I think the the more problematic statement is that the there's a specification stage right where I say like um again I I'll just use some conditions right you don't like for any token uh for every token right the you know the own the ownership of a token kind of uh I don't know oh I'll mess up right now the the formal specification but the owner of a token you know uh doesn't change after transaction unless there is a like clear you know spe like clear intent by the owner right so that that would be for example a very specific condition that you want to insert that actually user may not even like explicitly State yeah but the solution to this right is that in every transa in every function call right in you know in Rust like there is some something that needs to be H happening to satisfy this so so there is this kind of separation between um um I mean in the way specification right and the solution so there's an implicit set of assumptions yeah which the user doesn't even think about yeah which which ideally we have just because like I mean they're in the system overall right um or I mean maybe user does want this to be different right maybe like they run an you know a whatever something that randomly changes owners every tick of a time but like they need to specify that explicitly so and so there's the step of going from you know first of all solving the specification with the solution which ideally indeed the solution is also formal so you can and but you can go to a actual code and you can have a proof right that this formal solution is indeed satisfies all of the specifications condition and variance I thought you were saying the problem is that the user will not even notice that well yeah but I mean those can be added like you know almost implicitly to the to the specification right I mean presumably our IPM is very much like crypto whereare Crypton native D gen and so they know this right but but we need this kind of U System and so there is actually a really interesting paper Alpha geometry which is kind of in in this direction I don't know if you want to describe it right okay so so so I think Alpha geometry is the most interesting paper today uh which so so if we if we move like one level higher the general um property we desire from the system is that the system when it does reasoning it does reasoning in the symbolic space where it cannot mess up because in the language space models mess up and people mess up and as of today we have actually no evidence that uh it is possible to reason in a language space without messing up but we do know that it's possible to do it in symbolic space but of course symbolic space uh you know it blows up um because uh the search space is huge so ideally what you want is you want symbolic space for reasoning and large language model as an intuition and Alpha geometry is very close to that they they say Well it happens to be the case that for geometry the symbolic the DSL is relatively limited I think it's still massive I think the amount of effort that they put into creating it was uh it's hard to estimate but if you watch the video they scroll through the definition of it and uh it doesn't end scrolling while they show it uh but it's uh but it is tractable right they manag to uh to describe it uh and they also limit significantly what intuition means right it doesn't even happen in the language space it still happens in the symbolic space but they effectively uh I I don't even think it's still called can be called intuition instead they they just say that every now and then when you solve a problem you need to what they call pull the rabbit out of a hat and say and now I'm going to introduce this concept which comes out of nowhere but then once it is present the you know the solution becomes tracable well I mean same with competitive programming right like you have a problem yeah right and like the solution is you use black HS or what something yeah like a I think in programming quite frequently the rabbit is the whole solution well it's it's like it's it's some algorithm that if you apply it then everything else kind of fits into place right right and so if we go back either to competitive programming or even to aipm right so so effectively let's use a am as an example because we already introduced it right so you have the user I will do the same as you did and then you have code right and so as I said I want and then we want the user to communicate to the system uh to the IPM and a IPM well I cannot okay the user sad sadly after after Co we are not used to the whiteboards meaning we are used to mirror where you can just zoom in yeah and like Drop arrows precisely right yeah yeah and so you you communicate to the cipm uh which here's language and back also goes language but internally IPM like if you zoom in into their uh sort of hat they have this DSL right so uh so naturally uh there is ADSL uh such that any code can be expressed through it like there's a there's a um I don't remember the name but there's a system built on top of rust that allows you to provide you know formal guarantees for every function right so we know that there R core itself is actually formally verified but I'm saying like you can annotate function saying you know the um you know like this function will have this invariant and then there's a compiler that will um verify verify it for you right however it's so low level that any meaningful like if we convert it back to natural language it will not convey anything to the user right at the same time for any particular idea right so let's say I want to build an order book there's a D cell which has exactly one operation which says build an order book right which can be converted back to language and to code right but but it's not very generalizable right and so you can uh you can like imagine how generalizable it is and how understood it can be by user right and so here's our example of the order book right and here's the lowle that low level language and so the question is how this looks like and whether there is U and so like let's let's say there's like a space of sufficiently generalizable and sufficiently understood by the user like it goes some somewhere here right and like like the question is does this curve touch it or does it go entirely below uh and if it does go entirely below can we split it into multiple steps or introduce some other um this is noted to here or can we introduce you know some other components to the system that will make it look more like this yeah but hopefully I mean this is we saying if if the spec already is convertible to code one to1 yeah yeah yeah so exactly so like so this is for the we're saying yes what we get the dsls and for every level level of generalizability we have an oracle which says what is the best DSL in terms of being understood by the human but I think I mean I think the problem here is that you have you like the the the What's Understood by human is a problem specification what's what code needs to contain is a solution right so there is a step of uh like solving yeah which I think that's what I mean you still you still have the same yeah like for the for the description of the problem statement uh like how how generalizable is that problem statement how easy it is to modify right because like it's not build an order book it's build a specific order book that H is in your mind right there a very specific handling of all of the edge cases versus you know if you modify even one thing now you need like a completely different language versus like the I mean this point with Co with like rust annotations it's already a solution right it doesn't actually have like a you know right in my hand they not quite separated I think ideally you want to be in a situation where for problem statement you can go in the symbolic space to the I mean I think they should they should be in the same language space but but I do think there's a which Alpha geometry was the case the problem definition was in the same space as the but I think there like there's an expected at least in my for me there's an expected step of that the solution that is convertible to like this is one to one very explicitly yeah this actually requires like yeah and this is where you need symbolic reasoning with the intuition yeah which like this is where you need to come up with this and then prove that it's actually satisfy as oldin variance y yeah so this is where like actually like solver coming and uh yeah I think like why is this is important is because if you think of any code that is written all of it is so far been written by people and people have I think like there's like two three bugs per 10,000 lines of code something like that that sounds extremely optimistic okay maybe play a thousand lines of code so so the the CH I mean well written code with lots of tests Etc yeah and so the problem is like you know there's millions and millions of lines of code right now in your you know like you can you can have a net you have networking stack you have operating system you know and you have your browser and then you have your applications right assuming everything is just web apps uh like in all of this right starting with networking there's vulnerabilities and as the progress of AI actually goes forward AI actually will be way better at finding bugs already in the code kind of just because it's like a statistical problem and like the introdution problem versus doing this versus like writing a correct code on itself again with LMS right now and so so we're going to see more and more vulnerabilities being found and potentially exploited by um by the using llm that's one of the kind of the AI risk uh directions right that uh people exploring and so this is extremely important because in a way we need to rebuild the internet with this and cryptography and other pieces kind of from scratch to ensure we have kind of not breakable internet with AI so so to give a little bit of a different perspective right like Linux networking stack certainly has vulnerabilities there's absolutely no way that that many lines of code were written without vulnerabilities uh and the way I see the security model of the of the networking stack is that a very smart person cannot find that vulnerability right because a lot of very smart people looked at it and they didn't find them um however um the model can become much better than people at finding vulnerabilities way before we it becomes like AGI or it because that's a very narrow uh yeah it can run the code it can write tests it can simulate it like somebody is really targeting that specific problem yeah like like as an example of that the models they very good at writing code right not they still not applicable in many cases but but clearly if you show the current model to a person like in 2022 in the early 2022 they would be mindblown right and so correspondingly but but in some other areas of um of something that we consider to be trivial the models still mess up right so so the same the same argument applies like they will become maybe they already have became become uh much better at U breaking code in people right and uh well finding a zero day in the networking stack of Linux could be relatively um disastrous colossal right for the for the world right and so we either uh need to ensure that uh the good guys find them first and and Patch it um or I I I don't think we can in any way prevent the models from developing in into that in that direction or yeah or or like alternative would be to just uh build something yeah build something that is you know fully verifiable and just rewrite the whole networking stack and you know preemptive preemptively deploy something that is provably correct yeah I think in general like from the kind of yeah general safety approach right the whole point of like ma math is that yeah like the you know if you can have a proof and it's you know logic sound then there's no alternative way to kind of work around that with doesn't matter how much more intelligence or compute you have yeah versus if you have anything outside of that right outside of like formal logic and kind of mathematical proof potentially you can kind of brute force a way around it hence why again cryptography being one of the fundamental blocks as well um of creating this system I mean and right now it's used everywhere here as well and it doesn't help that the networking is written in C no which is which is generally the disaster language yeah well and so is a lot of Linux in general yes exactly yeah so yeah so we have a lot of work um we'll start here because uh there's a clear value there and and and the I mean smart contracts themselves are smaller and so at this point of the you know solving this problem space uh there's also just like a lot more people and and kind of um coordination contribution we can get through building applications but the idea of this model and of the system is to go down the stack and I mean that's why I kind of started talking about you know user owned internet because that's kind of what we need to do right like across every stack leveraging cryptography leveraging developer and uh rebuilding it in a new way I think we're approach ing the time or we already over time so uh thanks for tuning in hope this was educational as Alex said we have not prepared but uh we will promise you we will continue not being prepared uh but uh you know you will actually glean into what we're thinking discussing kind of day today uh and what we're working on so tune in for next time thank you","transcript_synced_at":"2026-11-03T05:51:08.788Z","segments":[{"start_seconds":0,"text":"I feel the custom feel hack the world and make it","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=0s"},{"start_seconds":11,"text":"tight blockchain","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=11s"},{"start_seconds":18,"text":"flows near the one we're close to won't stop us","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=18s"},{"start_seconds":26,"text":"now together we'll redefine the when near we'll find the flame empowered","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=26s"},{"start_seconds":35,"text":"Minds rise and St use your own AI we'll","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=35s"},{"start_seconds":47,"text":"explore hold the future hand in hand in this Tech driven Wonder","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=47s"},{"start_seconds":78,"text":"okay well welcome to a special office hours we actually are in person here in San Francisco is Alex and so we're going","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=78s"},{"start_seconds":85,"text":"to do a wideboard deep dive uh we're going to improvise so uh feel free to kind of enjoy the show uh","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=85s"},{"start_seconds":93,"text":"hopefully it's educational we'll try to go a little bit deeper than usual into some of the topics but uh yeah welcome and let's","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=93s"},{"start_seconds":101,"text":"start uh so I tweeted this few different hypotheses that I think uh will be","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=101s"},{"start_seconds":108,"text":"pushing the envelope on Research in Ai and so we wanted to kind of launch off that to also discuss some of the other","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=108s"},{"start_seconds":115,"text":"topics we're working on so maybe to start um so one of the kind of core","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=115s"},{"start_seconds":123,"text":"challenges with Z car models is that because they're trained on you know","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=123s"},{"start_seconds":129,"text":"massive amounts of tokens the um it takes you know extremely a kind of lots of compute to","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=129s"},{"start_seconds":138,"text":"actually produce that produce a model and train it and obviously the model itself is not training the most optimal","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=138s"},{"start_seconds":147,"text":"way right it's you know getting random pie of content like it's usually chunked kind of in the middle of a document","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=147s"},{"start_seconds":154,"text":"sometimes like in the second EP it sees well not e but yeah early on it is like yeah a proof of ran hypothesis yeah like","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=154s"},{"start_seconds":162,"text":"something about medical stuff you know some article about royalty and then closer to the end you know like it's it's being fed something from forchan","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=162s"},{"start_seconds":170,"text":"and it's not not contributing as much into yeah so so kind of there's like few of the research topics called curriculum","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=170s"},{"start_seconds":179,"text":"learning and active learning around how to prioritize which data you feed to the model so wanted to discuss that and kind of Alex been thinking about it for a","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=179s"},{"start_seconds":187,"text":"while yeah yeah I I like this um uh visualization of that which is by no means","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=187s"},{"start_seconds":195,"text":"scientific uh but uh imagine that you plot your um you have some sort of uh","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=195s"},{"start_seconds":204,"text":"omnicient Oracle who looks at your entire data set and for every um","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=204s"},{"start_seconds":210,"text":"for every document uh it assigns a value of how of how","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=210s"},{"start_seconds":216,"text":"smart um that document is right so so like uh and how much intelligence it","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=216s"},{"start_seconds":223,"text":"requires right and then we will assume that smartness um because the data set is sorted right if you look at and and","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=223s"},{"start_seconds":232,"text":"this is this is your data set as you're feeding it into the model right so this is time uh and","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=232s"},{"start_seconds":240,"text":"assuming the data set is somewhat uniformly shuffled if you look at like uh distribution of smartness uh during","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=240s"},{"start_seconds":247,"text":"different times it will be more or less the same right so like some very smart data uh on the very top slightly less","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=247s"},{"start_seconds":254,"text":"smart data and this is like ready the rest of Internet yes yeah um now um the","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=254s"},{"start_seconds":263,"text":"way I like to think about it especially if I find the Eraser is um","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=263s"},{"start_seconds":272,"text":"um let's say that exactly half of the data uh is only","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=272s"},{"start_seconds":279,"text":"useful um when it comes to learning uh grammar of English and and has no other usefulness to","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=279s"},{"start_seconds":287,"text":"it uh and then half of that half of the rest so this is 50% not","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=287s"},{"start_seconds":295,"text":"very useful 25% slightly more useful than 12% even more useful uh and now I will exaggerate it","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=295s"},{"start_seconds":303,"text":"significantly but it will it will sort of deliver the point imagine that in order to even start grasping usefulness of this half uh you need to to go","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=303s"},{"start_seconds":313,"text":"through half of this data set or like in general well I think in maybe two way to specify right so when","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=313s"},{"start_seconds":322,"text":"you're training these models like generative pre trining right what it does it's trying to predict the next token and so in the beginning like the","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=322s"},{"start_seconds":330,"text":"basic grammar rules is enough to like reduce the loss right yeah so like if you understand at least how English senten are constructed and like basic","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=330s"},{"start_seconds":339,"text":"vocabulary you know that like an article goes before word Etc right that already reduces loss right but obviously as you","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=339s"},{"start_seconds":346,"text":"train more and on more complicated things you need to actually start reasoning and making inferences and remembering facts right so in this case","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=346s"},{"start_seconds":355,"text":"let's say let's say that this is like a uh that this 25 % start involving basic reasoning and let's say that you need um","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=355s"},{"start_seconds":364,"text":"so that will be an exageration of the actual percentage but let's say you need to uh this 50% of the uh lowest level data set um consumed","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=364s"},{"start_seconds":373,"text":"before you even start comprehending reasoning right so in this case this data this part of the data is only useful here uh and up to this point you","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=373s"},{"start_seconds":383,"text":"were only using it to learn grammar which is you know not the most efficient way of using this particular data like it would have been much better uh to","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=383s"},{"start_seconds":391,"text":"swap them right so like to move uh this block of um data which is only useful From perspective of learning grammar","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=391s"},{"start_seconds":399,"text":"here and this that gives you reasoning here right in which case you will","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=399s"},{"start_seconds":407,"text":"double your effective you know effective learning from text and then here you you have","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=407s"},{"start_seconds":414,"text":"something even more uh smart right but then it takes you half of this time to start comprehending it and so here you","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=414s"},{"start_seconds":422,"text":"only use 1/8 of this data set and then you only use 116th but but the actual the actual","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=422s"},{"start_seconds":430,"text":"coefficient is 150% right it's it's less than that but the but the point here is that somewhere on top you have like let's say you want to you want the model","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=430s"},{"start_seconds":438,"text":"to be good at uh reasoning about um you know general relativity differential geometry tensor calculus all those","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=438s"},{"start_seconds":447,"text":"things right and let's say that this is is like level seven of this diagram and so you do have I mean people do talk","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=447s"},{"start_seconds":456,"text":"about it talk about it on the internet write the textbooks people talk about it on Reddit in specialized subreddits people talk about it on like physics","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=456s"},{"start_seconds":464,"text":"stack exchange and this is all like in here but you only use like one over you know 256 of","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=464s"},{"start_seconds":472,"text":"this data set uh and with each consec with each consecutive level you use exponentially less and less of the data usefully","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=472s"},{"start_seconds":480,"text":"right and if you were able uh if you know if you had those labels from the Oracle and you were able to reshuffle your data set uh you would be able to be","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=480s"},{"start_seconds":488,"text":"you know like you wouldn't even need new data which doesn't exist yet um you would just use whatever you have there","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=488s"},{"start_seconds":497,"text":"efficiently yeah and so as a humans right we're also not getting thrown you know an article of theorical physics an","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=497s"},{"start_seconds":504,"text":"article about you know uh linear equations and an article about uh how I don't know how different uh","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=504s"},{"start_seconds":513,"text":"species are related to each other before we learned you know how to read to read yeah and kind of maybe basic math and so","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=513s"},{"start_seconds":522,"text":"so the idea is like how do we design a system and the thing about this is actually this doesn't need to modify any of the architecture of the model this is","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=522s"},{"start_seconds":530,"text":"just the training procedure that needs to be modified and so the question is how to get that Oracle right that given","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=530s"},{"start_seconds":538,"text":"a document returns well you called it smartness but um nonscientific smartness","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=538s"},{"start_seconds":547,"text":"score right and the thing is like you probably don't need this to be you know 100% like obviously I mean it's really hard but if you get get a you know","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=547s"},{"start_seconds":556,"text":"reasonable correlation um you can use this and so one interesting point is that if you take you know any of the document","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=556s"},{"start_seconds":564,"text":"anywhere and you take existing model right let's say we take Lama 3 and we run that model on this document","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=564s"},{"start_seconds":572,"text":"if the document doesn't require kind of the lower it is on this level the more","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=572s"},{"start_seconds":579,"text":"that uh that model already is good at predicting that's kind of what we train them to do right and and so you actually","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=579s"},{"start_seconds":587,"text":"can use the loss of uh let's say llama which it give it gives you a","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=587s"},{"start_seconds":594,"text":"predictive loss uh if you score the document as a way to approximate","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=594s"},{"start_seconds":601,"text":"this now it's indeed not always going to be like you know completely correct but yeah because complete garbage will have","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=601s"},{"start_seconds":608,"text":"a pretty pretty high loss yeah and so so there are a few things that you need to um kind of correct for right so there's um if you look at any document and this","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=608s"},{"start_seconds":617,"text":"is kind of for those who are actually training this or or looking at this models um you can imagine let's say it's","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=617s"},{"start_seconds":624,"text":"it's a news article right and there is like you know a title and then there's an","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=624s"},{"start_seconds":631,"text":"author and then there's like something and then there's an ad probably and there's something else right and so the","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=631s"},{"start_seconds":638,"text":"reality is even in in a general article there's like already a bunch of stuff that actually almost impossible to","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=638s"},{"start_seconds":645,"text":"predict like if I give you this title and ask you to predict an author uh like even with those","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=645s"},{"start_seconds":652,"text":"intelligence uh it will be really hard I mean maybe if you learn who works where in which Media Company and which topics","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=652s"},{"start_seconds":660,"text":"they cover but like generally there's already parts of the document that you ideally want to kind of ignore so so there's another part part and again","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=660s"},{"start_seconds":669,"text":"similarly if there's a document that has complete garbage right which sadly internet is full off literally just Spam of like you know comma comma comma like","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=669s"},{"start_seconds":678,"text":"some words repeated somebody like optimizing for SEO there's like a bunch of just you know random words like you want to figure out how to filter those","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=678s"},{"start_seconds":687,"text":"documents anyway right this is definitely not helpful for anything as well as on a document level there's parts that are completely","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=687s"},{"start_seconds":694,"text":"unpredictable and again the model will output so there's a kind of if you look at the Sloss and the correlation with","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=694s"},{"start_seconds":701,"text":"smartness um they're going to be kind of a a distribution as well right so","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=701s"},{"start_seconds":710,"text":"loss smartness um so really high loss probably low smartness and so I mean I mean I haven't","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=710s"},{"start_seconds":719,"text":"plotted this but","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=719s"},{"start_seconds":726,"text":"yeah you know the meme yeah except this is Spam but yes yeah so","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=726s"},{"start_seconds":735,"text":"anyway so so there's a kind of some procedure we can use here now the other way which is more active right so this","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=735s"},{"start_seconds":742,"text":"is kind of you know you score this data ahead of time now the problem is this is also computationally really expensive because you literally need to run your","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=742s"},{"start_seconds":750,"text":"model over the whole data set exactly what the training yeah is but you need to do it once presumably to sort this","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=750s"},{"start_seconds":757,"text":"now you can train every model using that function right now there's an interesting option which is can you do","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=757s"},{"start_seconds":764,"text":"it actively right so as you're training a model right so right now let's say you have a model you are you know feeding some","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=764s"},{"start_seconds":773,"text":"data it you know uh you know you compute the loss and you kind of back Pro like update the model from this um and so","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=773s"},{"start_seconds":783,"text":"right now kind of generally I mean this is this is another um hypothesis so the idea of active learning is you instead","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=783s"},{"start_seconds":792,"text":"of um uh kind of feeding just random pieces of data you actually try to find like","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=792s"},{"start_seconds":801,"text":"you kind of ask the model what would be what would be the most ambiguous for it right so in a way you can say like oh","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=801s"},{"start_seconds":808,"text":"well if losses let's say three you know what would be the inputs and you sample the inputs from this right and then you","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=808s"},{"start_seconds":816,"text":"kind of uh now generally Active Learning is used with human feedback right and so for pre-training it may be not very","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=816s"},{"start_seconds":824,"text":"effective because like ideally you have you know you have a question for example uh you know question answering","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=824s"},{"start_seconds":830,"text":"system or something similar and you are uh uh trying to find you know what","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=830s"},{"start_seconds":838,"text":"questions to ask to like what pairs of question answers to add to find to improve the model uh but indeed it's a","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=838s"},{"start_seconds":845,"text":"bit hard now can you do the if you have like this big data set right so lots of documents","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=845s"},{"start_seconds":853,"text":"uh pretty much can you cheaply find something in it that is you know the most promising for you to update um and","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=853s"},{"start_seconds":863,"text":"so you can have like a smaller model that's one option uh you can have if you if you're training so right now you know this is let's say","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=863s"},{"start_seconds":871,"text":"15 trillion tokens which is a lot if you're training on a smaller subset let's say you just take Wikipedia for","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=871s"},{"start_seconds":878,"text":"example only right which is uh 2.5 billion tokens so yeah right now it's trained on","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=878s"},{"start_seconds":888,"text":"7,500 vikia size of content uh so you can train multiple times on it right so you can actually sort data um you know","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=888s"},{"start_seconds":896,"text":"from a previous ug based on how you model scored to feed it in a in a in an order that's more useful for it to","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=896s"},{"start_seconds":903,"text":"update because the other thing is this models right now need to like back propop thousand times something to","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=903s"},{"start_seconds":912,"text":"actually memorize it effectively so like a simple idea would","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=912s"},{"start_seconds":921,"text":"be you know instead of grabbing a random document you grab random 10 compute the loss and train on the five that have","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=921s"},{"start_seconds":930,"text":"the the low Med the median yeah so that presumably drops the stuff that you","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=930s"},{"start_seconds":937,"text":"understand very well which might potentially bring you back and removes the stuff which is either spam or something beyond your comprehension yet","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=937s"},{"start_seconds":946,"text":"so the only problem now is that uh so early on that would make sense right like if you hear like early on you grab some documents uh the spam which we","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=946s"},{"start_seconds":955,"text":"don't even like let's say there's also spam here will have high loss something you don't understand will have high loss you will be picking this ones what you","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=955s"},{"start_seconds":963,"text":"want to happen is that later down the road um you kind of want to be able to find out because you were running into them right like as you were moving along","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=963s"},{"start_seconds":972,"text":"you were running into smart data so it would be nice to somehow in yeah fetch it later yeah that","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=972s"},{"start_seconds":980,"text":"like this was something I didn't know how to do before minus spam yeah but hope hopefully you remove the","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=980s"},{"start_seconds":988,"text":"the spam ahead of time right spam can be potentially detected like by like you know like to high or something like this","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=988s"},{"start_seconds":996,"text":"yeah yeah I mean and yeah I mean one option is you use like a down sample model like you know 1B model for example","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=996s"},{"start_seconds":1003,"text":"to score data as prescore data find things that are like complet you know not even gr like grammar predictable and","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1003s"},{"start_seconds":1011,"text":"then filter out that okay sounds like something to be","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1011s"},{"start_seconds":1026,"text":"done yeah and then the other interesting thing I mean this is just like a thought that when you're training right right now you're applying laws kind of evenly","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1026s"},{"start_seconds":1034,"text":"across the whole document right and so there are places which are like either unpredictable or really hard to predict like a first token in the document right","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1034s"},{"start_seconds":1042,"text":"like I you know tell me how's this document going to start from you know a start of a document token","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1042s"},{"start_seconds":1049,"text":"it's really hard so in a way you want to kind of also update a little bit how the loss is propagated based on in general how","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1049s"},{"start_seconds":1057,"text":"predictable can this be in a training procedure yeah I think I think another","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1057s"},{"start_seconds":1065,"text":"interesting completely tangential to that idea here is that um like like","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1065s"},{"start_seconds":1072,"text":"obviously for the lower levels of smartness uh you have to mine internet MH uh but like you know if you go to","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1072s"},{"start_seconds":1082,"text":"llama today and you ask her something about Cuda it will give you something pretty meaningful right so like it's grasping at least some concepts of Cuda","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1082s"},{"start_seconds":1090,"text":"but naturally the amount of data in the internet for Cuda was much smaller right so so the higher the level the more","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1090s"},{"start_seconds":1097,"text":"tractable it becomes to just uh you know grab humans and annotate data from","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1097s"},{"start_seconds":1104,"text":"scratch right so instead of uh trying to find a way to autom atically identify","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1104s"},{"start_seconds":1110,"text":"smart data you can create smart data and put it at the end or I guess I don't know F tune on","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1110s"},{"start_seconds":1120,"text":"it yeah I guess that I mean relevant to this question is we want the model to","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1120s"},{"start_seconds":1128,"text":"understand how to you know how to read this how to reason about it they don't always needed to remember all of this right because the other part right now","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1128s"},{"start_seconds":1136,"text":"with this models we're trying to actually get them to remember every single fact and you know thing that happening on internet because that would","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1136s"},{"start_seconds":1145,"text":"be the low like then at the you know training time it will get the low loss uh on predicting next token and so the","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1145s"},{"start_seconds":1154,"text":"interesting question there is like how can we modify the training procedure in such a way that","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1154s"},{"start_seconds":1161,"text":"um it's not it doesn't need to remember it can actually use something in a context to uh for for example better","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1161s"},{"start_seconds":1170,"text":"predicted documents and so one idea that I had actually is as you you know starting to train on documents in in","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1170s"},{"start_seconds":1178,"text":"kind of these places is to find relevant information that may be at uh either","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1178s"},{"start_seconds":1185,"text":"previous levels of smartness but that contains kind of some information and putting that in a context before","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1185s"},{"start_seconds":1193,"text":"predicting right so let's say I mean you're reading a Wikipedia article about you know linear equations maybe you want to put","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1193s"},{"start_seconds":1201,"text":"in well maybe linear equations is uh I don't know some fact like you you're generating like a news article about","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1201s"},{"start_seconds":1209,"text":"Nvidia and so you want to put in maybe like Nvidia Wikipedia article and something else so you want some","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1209s"},{"start_seconds":1217,"text":"retrieval at training time at training time yes because the big benefit of this models is that there you know can use the context to retrieve information","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1217s"},{"start_seconds":1225,"text":"right and you just only you know back prop loss here and use this more just for context","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1225s"},{"start_seconds":1234,"text":"y well shall we talk about how easy it is to train model from scratch today in","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1234s"},{"start_seconds":1255,"text":"2024 no you can okay so so so let","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1255s"},{"start_seconds":1263,"text":"so there's another um somewhat fundamental","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1263s"},{"start_seconds":1270,"text":"uh uh area of thinking which is which has nothing to do uh with the science of training","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1270s"},{"start_seconds":1279,"text":"models and more about the current state of the world uh and um let let me let me use my favorite analogy which is uh","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1279s"},{"start_seconds":1287,"text":"let's say that today in 202 four I want to create a document yeah like IL and I are preparing for this office hours which we absolutely did not as you can","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1287s"},{"start_seconds":1296,"text":"see but let's say we were you know put more more effort into it and actually try to prepare then we","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1296s"},{"start_seconds":1304,"text":"would create a document right and we would have multiple ways of putting it together right one of them would be and that's likely what we would have","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1304s"},{"start_seconds":1312,"text":"actually done is to use Google dos however I might have ahead of time","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1312s"},{"start_seconds":1321,"text":"say well that document I would prefer to preserve it for posterity like you know when I'm 90 years old I would like to look at that document preparing for the","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1321s"},{"start_seconds":1328,"text":"office hours and Google Docs is not the way I mean I can probably save it but right and so I would open up my copy","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1328s"},{"start_seconds":1337,"text":"of Microsoft Word uh and prepare it in the Microsoft Word um or I can say well I can prepare","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1337s"},{"start_seconds":1345,"text":"it my in Microsoft Word uh but then by the time I'm 90 Microsoft could be gone Microsoft Word is you know not supported","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1345s"},{"start_seconds":1355,"text":"anymore uh and nobody will be there to pick up the torch because Microsoft Word is not open source right so instead I would open Liber","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1355s"},{"start_seconds":1364,"text":"office that escalated quickly right and so now let's let's see how much of Open","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1364s"},{"start_seconds":1371,"text":"Source each of them is right so Google Docs is not open source","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1371s"},{"start_seconds":1378,"text":"uh and also I have no ownership of data of anything that is stored on Google dos right so no","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1378s"},{"start_seconds":1387,"text":"ownership data belongs to and no ability to Tinker with the with the Google Doc binary like you know if Google tomorrow","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1387s"},{"start_seconds":1395,"text":"says now you know you want these buttons to be swapped because someone needed to get a promotion at","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1395s"},{"start_seconds":1402,"text":"Google I can do nothing about it right Microsoft Word not open source for sure but they do get to have binary right I","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1402s"},{"start_seconds":1409,"text":"have I have a binary of which I can make no sense at all uh but I have it locally right so if tomorrow Microsoft Word says","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1409s"},{"start_seconds":1417,"text":"hey instead of toolbar we will have a ribbon which they did to us I can say well I'm not going to use your Ribbon I will use the old Microsoft Word right so","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1417s"},{"start_seconds":1426,"text":"um well I guess it's it's separate from ownership of data right like I can Tinker with it no tinkering there's no tinkering and","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1426s"},{"start_seconds":1433,"text":"there's like there's like limited tinkering right well it's more like you're not for wor thing to to upgrade to yeah yeah but I guess also","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1433s"},{"start_seconds":1442,"text":"importantly I have ownership of data like you know nobody nobody's looking into my documents I'm editing locally actually don't know if Microsoft Word","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1442s"},{"start_seconds":1449,"text":"is that's a the time have changed yeah tinkering okay and Libra office it's open source I mean it's as open source","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1449s"},{"start_seconds":1457,"text":"as it gets I can download the code I can you know if I don't like something about it I can change it I own my data own my","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1457s"},{"start_seconds":1465,"text":"data I can Tinker as much as I want so uh even though in Microsoft Word I can","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1465s"},{"start_seconds":1472,"text":"tinger a little bit with the binary I can I don't know I can write a plugin for Microsoft Word I don't know if I can I own my data nobody would ever say","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1472s"},{"start_seconds":1481,"text":"Microsoft Word is open source right uh so now if we I don't know how far this goes but let's say this is more space if","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1481s"},{"start_seconds":1489,"text":"we compare it to the state of the artificial intelligence today Google Docs would be open AI or like entropic any anything that goes through the API","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1489s"},{"start_seconds":1498,"text":"right if I if I use open AI like a good example with open AI uh is when they introduced this button that says you","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1498s"},{"start_seconds":1507,"text":"know we're not going to save your history uh and then you know uh in my particular instantiation one day it just","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1507s"},{"start_seconds":1514,"text":"fli it back because they changed the control and the new control by default was again saving my history and saving history means staining on my data right","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1514s"},{"start_seconds":1521,"text":"so like open AI I have no no control over how my data is being used at best I can you know they can promise not to use","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1521s"},{"start_seconds":1528,"text":"my data and I can trust them uh and then llama and mistr would not be in the last","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1528s"},{"start_seconds":1536,"text":"row for sure right because in reality with Lama in what we get is we get this binary that we cannot make much sense of right so llama would","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1536s"},{"start_seconds":1544,"text":"be um more like Microsoft Word right so I own my data everything I do locally more like stable","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1544s"},{"start_seconds":1554,"text":"diffusion uh if I locally do inference I know for sure that data doesn't go anywhere it stays with me I can do some","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1554s"},{"start_seconds":1561,"text":"limited tinkering I can f tune do laora um but it's not open source right even though they called open source models they are not open source because I","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1561s"},{"start_seconds":1569,"text":"cannot like if I want to change something in them fundamentally that is not something I can do training Loop is not open source well you don't actually","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1569s"},{"start_seconds":1577,"text":"know what went into it yeah yeah we don't know the data that went in we don't know the training Loop right so like if I say well I like clama but I want I wanted to understand images now","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1577s"},{"start_seconds":1585,"text":"so how about I change the training uh distribution a little bit right or like apply more filtering none of that is possible right and like and also the","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1585s"},{"start_seconds":1593,"text":"flip of side of this is this paper on sleeper agents where you pretty much train the model to have a specific behavior at a specific","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1593s"},{"start_seconds":1601,"text":"instance and that doesn't show up at any valve that or uh you know persists over fine tuning yeah and so like there can","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1601s"},{"start_seconds":1609,"text":"be like things that are you know when the condition triggers it installs for example npm package in in your python","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1609s"},{"start_seconds":1617,"text":"when it's get cold that has a malicious uh for example code yeah okay and so","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1617s"},{"start_seconds":1624,"text":"now well I haven't been following the space very closely recently but the only model I knew as of like a month ago","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1624s"},{"start_seconds":1632,"text":"wasmo which is Allen's Institute of Technology right yeah so that one is you know you can reproduce it right you can","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1632s"},{"start_seconds":1639,"text":"download everything uh and um the problem is that","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1639s"},{"start_seconds":1650,"text":"here's the problem that as the as the um industry goes forward we need to make","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1650s"},{"start_seconds":1658,"text":"sure that we have open source models that are competitive with the stateof the art right so so clearly you can be in in","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1658s"},{"start_seconds":1667,"text":"three different regimes one is that where even open weight let's call them open weight even open weight models are not competitive with state of the earthart right this is this is complete","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1667s"},{"start_seconds":1675,"text":"dystopia uh that effectively means that uh and here I make an assumption that AI will actually drastically change the way we experience the internet and will","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1675s"},{"start_seconds":1684,"text":"become you know the fundamental part of our interactions with it then we at a mercy of two companies we always go through their Gateway yeah and we always","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1684s"},{"start_seconds":1692,"text":"go through their Gateway which which has been the case in multiple other areas of Our Lives social networks you know search Etc and that is not great um","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1692s"},{"start_seconds":1702,"text":"second one is open weight models are competitive this is strictly better right so then um um at least at any at","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1702s"},{"start_seconds":1709,"text":"the point of time when open weights models are better uh there's a possibility for competing products the problem with open weight models is that","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1709s"},{"start_seconds":1718,"text":"there's no uh it is not it is not future proof right so so so these guys they they will","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1718s"},{"start_seconds":1727,"text":"be um innovating right uh and in the moment we have open weight models and in the moment we have those entities who","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1727s"},{"start_seconds":1733,"text":"publish them aligned in terms of incentives with the with the big guys uh sorry they aligned in terms of incentives with the world to publish","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1733s"},{"start_seconds":1742,"text":"those models uh for various reasons like for example for some of them the reason could be that they get attention today because attention is important for some","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1742s"},{"start_seconds":1749,"text":"others uh I can only guess monitors for Facebook but I think Facebook is generally has been known for open sourcing stuff because then people","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1749s"},{"start_seconds":1757,"text":"Tinker with it and Facebook gets it like like you know if there's llama and and GPT and Lama is open source then","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1757s"},{"start_seconds":1764,"text":"Facebook gets a whole world working on their stuff for right they get v l and whatnot for free effectively and that's that's like a","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1764s"},{"start_seconds":1773,"text":"very nice Synergy right people get open weight models Facebook gets a development on it but it's not future proof because if at some point","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1773s"},{"start_seconds":1781,"text":"incentives change right so suddenly you know like hypothetical I don't want to name any names but like any company that was interested because of the attention the moment attention is not as valuable","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1781s"},{"start_seconds":1790,"text":"anymore and they want to start generating profits the license change um don't know if it ever happened yet","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1790s"},{"start_seconds":1798,"text":"um right but if you have an open source model it is future proof right so let's say for the company let's say let's say alen's Institute of Technology says","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1798s"},{"start_seconds":1807,"text":"tomorrow well we don't want to work on Alo anymore it's not it's not anymore you know meaningful for us anyone can pick","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1807s"},{"start_seconds":1816,"text":"up from where they've left and train a new model right the Next Generation so it's it's very important that there is a open source","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1816s"},{"start_seconds":1823,"text":"model um well I competitive it's it's it's not on itself future proof right because the future proofness here comes","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1823s"},{"start_seconds":1831,"text":"from the flywheel of you know like they have a model they put it into a product that generates Revenue yeah that they","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1831s"},{"start_seconds":1840,"text":"put back into Data compute people people as product which then improves","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1840s"},{"start_seconds":1849,"text":"the model right so so I think an important yeah there there was a couple of steps in my head that I forgot to to put on the Whiteboard but in this case","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1849s"},{"start_seconds":1857,"text":"for the product to exist right so we have the model right let's say it's Lama uh and then we want a product and let's say the product would be even just like","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1857s"},{"start_seconds":1865,"text":"you know a Char GPT alternative right so for a company to truly commit into making it to be a very you know","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1865s"},{"start_seconds":1872,"text":"appealing competitor they need to be that company needs to be certain that they will not die out if the model they depend on gets discontinued uh and","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1872s"},{"start_seconds":1881,"text":"suddenly uh you know they once CH GPT switches you know to GPT 5 uh they will be left behind","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1881s"},{"start_seconds":1890,"text":"because there's no GPT 5 for them to switch onto and there as a company right so so like if you want to build a future proof","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1890s"},{"start_seconds":1898,"text":"company it's much better to know that the infrastructure on top of which you're building will be evolving yeah but ideally what you want is you want","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1898s"},{"start_seconds":1905,"text":"kind of this system to exist in open source where more people data and","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1905s"},{"start_seconds":1914,"text":"compute are being kind of gathered to build a better model that and not just one but many products are built around","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1914s"},{"start_seconds":1921,"text":"which then fund development of this y hence the economy of","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1921s"},{"start_seconds":1930,"text":"blockchain like you need some kind of platform that is encouraged by the fact that more people are building on top of","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1930s"},{"start_seconds":1937,"text":"it versus competing kind of product that they pretty much trying to eat everybody","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1937s"},{"start_seconds":1944,"text":"else which is generally how any single profit company needs to exist","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1944s"},{"start_seconds":1952,"text":"Y at which point of course it's very desirable for normal people to be able to interact with the blockchain for","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1952s"},{"start_seconds":1964,"text":"yeah you were going somewhere no more or less train on the training Loop no not no I think I think","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1964s"},{"start_seconds":1971,"text":"that's more or less the the stage I wanted to build up well I think the I mean the question we kind of mentioned that I think last of is ours around like","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1971s"},{"start_seconds":1979,"text":"training Loop um because right now indeed like I mean I think this is there's this and then there is a megat","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1979s"},{"start_seconds":1987,"text":"Megatron which is from Nvidia and uh I don't know if there anything else and uh I mean all of them","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1987s"},{"start_seconds":1996,"text":"obviously require like you know data center level of compute I mean you you can start with","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=1996s"},{"start_seconds":2005,"text":"like obviously smaller uh setup but if you want to train like I mean 70 70b","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2005s"},{"start_seconds":2013,"text":"model which seems like you know reasonably powerful um this is what 3 months","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2013s"},{"start_seconds":2024,"text":"of so the original Lama paper reported I think something like 29","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2024s"},{"start_seconds":2032,"text":"days on 248 A1 100s mhm","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2032s"},{"start_seconds":2040,"text":"however uh my limited knowledge of those training Loops suggest that that was an extremely optimized code which not everybody has access to","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2040s"},{"start_seconds":2050,"text":"yeah so so I think for more realistic uh well I mean the question that would be like around couple month","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2050s"},{"start_seconds":2056,"text":"yeah yeah the question is can we in open source right the benefit we have is that we can have more people actually looking","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2056s"},{"start_seconds":2063,"text":"at it and optimizing it yeah um and so again I mean we briefly discussed it but","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2063s"},{"start_seconds":2070,"text":"it was you know you can you can imagine like a training","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2070s"},{"start_seconds":2078,"text":"Loop that is open source you have a kind of few in a way plugins right on like on","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2078s"},{"start_seconds":2087,"text":"the model on data sampling procedure maybe you have the","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2087s"},{"start_seconds":2094,"text":"trainer SL optimizer that are kind of easy to plug","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2094s"},{"start_seconds":2101,"text":"in and so the kind of the biggest benefit is right now as this you know the researchers who are like hey I have this idea right we just discuss","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2101s"},{"start_seconds":2109,"text":"something I could code it up you know and ship it tomorrow I just don't have the this level of compute yeah and so","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2109s"},{"start_seconds":2118,"text":"ideally we would have a place where they can come in and say like hey here is a pretty much full request with some","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2118s"},{"start_seconds":2126,"text":"modifications train it on maybe like 1B first um if it shows improvements or","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2126s"},{"start_seconds":2136,"text":"shows interesting results then you know scale it to let's say 7B to 70b right so","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2136s"},{"start_seconds":2143,"text":"um and then you kind of have you know behind this you have few clusters of","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2143s"},{"start_seconds":2150,"text":"compute that are Al so there is like a um it's like an open you know AI research lab","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2150s"},{"start_seconds":2162,"text":"where actually people can come in and suggest no it's like open end of token AI not open AI as a token yeah yeah you","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2162s"},{"start_seconds":2171,"text":"know token you need a better tokenizer yeah what is so this is what is it it's like 2K per hour","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2171s"},{"start_seconds":2178,"text":"right which is for one month where is my calculator I think it's uh six million","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2178s"},{"start_seconds":2189,"text":"you can talk about some other thought in the me time yeah so I mean this is kind of something we've been uh thinking about","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2189s"},{"start_seconds":2196,"text":"like to to build an open source um all","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2196s"},{"start_seconds":2203,"text":"right yeah so you can imagine one B you can probably train for way","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2203s"},{"start_seconds":2212,"text":"cheaper well let's just say it scales linearly for Simplicity right that that should be easy to divide","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2212s"},{"start_seconds":2220,"text":"by 70 it's like2 yeah yeah 0 Z to","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2220s"},{"start_seconds":2230,"text":"20K yeah that that becomes tractable yeah I mean you probably start with like even smaller but","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2230s"},{"start_seconds":2237,"text":"yeah I think it should be cheaper right now I think 7B you can train with like","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2237s"},{"start_seconds":2243,"text":"128 for a week or something so","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2243s"},{"start_seconds":2250,"text":"anyway um yeah so there's like a I mean in the way this is like a you know a research Grant to someone and then you","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2250s"},{"start_seconds":2258,"text":"kind of experiment with a few of those and then exp explore probably combination of this for","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2258s"},{"start_seconds":2265,"text":"a bigger model yeah so as as a community we should build","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2265s"},{"start_seconds":2274,"text":"this to truly be the home for use your own in open source","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2274s"},{"start_seconds":2281,"text":"AI yeah sounds great all right so that's solved","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2281s"},{"start_seconds":2292,"text":"clearly all um do we want to talk about symbolic stuff sure I mean we don't have much probably to say about","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2292s"},{"start_seconds":2301,"text":"it but uh so so I guess like a very short bit of trivia is Ilia and I when","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2301s"},{"start_seconds":2308,"text":"we met uh one of the things that we shared in common was that both of us during our master's days were trying to","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2308s"},{"start_seconds":2316,"text":"use symbolic AI for whatever and so for program synthesis program synthesis yeah so so so both of us kind of dream about","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2316s"},{"start_seconds":2324,"text":"still about using it one day are we any closer to knowing how to do that no but but like do do you want to draw","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2324s"},{"start_seconds":2333,"text":"your diagram with the uh user and the system you mean for for the developer oh yeah","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2333s"},{"start_seconds":2340,"text":"yeah for a developer but I think it applies to anything yeah so well maybe yeah there's like few few things but specifically I I","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2340s"},{"start_seconds":2349,"text":"actually talked about this at Town Hall earlier we working on this thing we call the developer right so you can assume there's a person and they want to build","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2349s"},{"start_seconds":2357,"text":"some application right so there's an application here and you know without loss of generality we assume it's a vi3","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2357s"},{"start_seconds":2364,"text":"application yeah right which has you know some kind of smart contract some kind of uh you know we","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2364s"},{"start_seconds":2373,"text":"call middleware which and then some front","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2373s"},{"start_seconds":2379,"text":"end and so we want an AI who will you know go from","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2379s"},{"start_seconds":2388,"text":"build uh I don't know freelance freelancer","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2388s"},{"start_seconds":2396,"text":"Marketplace that will actually give you a full application now if you imagine on this side you actually have a human that","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2396s"},{"start_seconds":2403,"text":"is like get this I mean they'll assume something right they'll they can build a freelance application but it's probably","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2403s"},{"start_seconds":2411,"text":"not the one that user wanted right and so you first need kind of a process","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2411s"},{"start_seconds":2418,"text":"which you know um usually somebody like a product manager or uh or the person","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2418s"},{"start_seconds":2427,"text":"sagee which is actually write out exactly what they want to be built right so we call this a","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2427s"},{"start_seconds":2435,"text":"specification and so the reality is though there's a lot of detail that this specification should go into to actually","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2435s"},{"start_seconds":2442,"text":"fully describe this and so uh you actually have a kind of interative process where as the user you know hey","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2442s"},{"start_seconds":2451,"text":"build a fencer Marketplace let's imagine there's we you know we call it aipm product manager who writes a specification for you you're like well","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2451s"},{"start_seconds":2460,"text":"actually that's not exactly what I wanted I actually want maybe to you know I want people to have nft badges for","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2460s"},{"start_seconds":2468,"text":"every time they complete their task right and so you kind of update and add features and maybe modify the user flows and so at some point when you're you","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2468s"},{"start_seconds":2477,"text":"know content with the specification you then say okay let's build it and so that's get shipped to proverbial AI","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2477s"},{"start_seconds":2485,"text":"developer who is then write the code now potentially still coming back to you to","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2485s"},{"start_seconds":2493,"text":"update the spec as it goes yeah it potentially either comes back to you and you can update this back right so um in a way you can almost model it as if you","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2493s"},{"start_seconds":2502,"text":"had like a Dev team behind the chat and you in a group chat with them and uh you","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2502s"},{"start_seconds":2509,"text":"had two outputs for them one is documentation of the project and the other one is the app itself and you can kind of iterate with them on this","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2509s"},{"start_seconds":2518,"text":"uh that's why there's potentially also like AI designer who you know helps with the front end Etc so that's kind of you know we we building this system out","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2518s"},{"start_seconds":2526,"text":"we're going to be opening up the um opening up this part to play around","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2526s"},{"start_seconds":2533,"text":"for the community in a bit um so you can actually like write out your ideas see the specification iterate on it and then","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2533s"},{"start_seconds":2542,"text":"we're going to start leveraging already what we have but Al people to develop this and kind of iterate on and to see","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2542s"},{"start_seconds":2551,"text":"how best train the developer to do this but one of the biggest issues with all this and and kind of generally with the","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2551s"},{"start_seconds":2559,"text":"current state of llm writing code is that they're not very good at it at like at a you know sufficient detail of","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2559s"},{"start_seconds":2568,"text":"reasoning and and logic not very rigorous very rigorous yeah and so they and with smart contracts it's especially important that you are very rigorous and","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2568s"},{"start_seconds":2576,"text":"you don't lose Lo people's you know money assets and uh so you have set of you know conditions that you want right","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2576s"},{"start_seconds":2584,"text":"now the way it's done is it's sent to an auditor which is another you know person who like reviews the code finds one you","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2584s"},{"start_seconds":2591,"text":"know issues and then kind of developers iterate on it but that is I mean like","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2591s"},{"start_seconds":2599,"text":"usually a very cumbersome process and Auditors also need to know the specification and they also make mistakes and they also make mistakes","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2599s"},{"start_seconds":2607,"text":"like if you have an AI auditor here yeah so but even human Auditors like they","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2607s"},{"start_seconds":2615,"text":"make mistakes yeah they miss things right I mean how many smart contracts got hacked which had an audit so the question is can we have a some form of","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2615s"},{"start_seconds":2623,"text":"guarantee that this smart contract is indeed correct up to a specification plus you know some set of","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2623s"},{"start_seconds":2631,"text":"conditions and variants around not losing money not changing ownership of the uh of the assets without the","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2631s"},{"start_seconds":2640,"text":"originating user uh confirming it Etc","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2640s"},{"start_seconds":2647,"text":"yeah so what would be interesting here is if uh imagine that there is","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2647s"},{"start_seconds":2658,"text":"a uh highlevel um language on which or rather","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2658s"},{"start_seconds":2666,"text":"I actually don't know if there's any formal proof systems for stability but they do exist for us for example right so people people for many years now","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2666s"},{"start_seconds":2673,"text":"people have had uh ways of formally proving that the code does something you expect from it but you can probably go even one","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2673s"},{"start_seconds":2682,"text":"level higher and say so so now uh there's a specification language which is uh not uh which is a uh an actual","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2682s"},{"start_seconds":2690,"text":"symbolic language which can be you know reasoned about um using some sort of proof system right and so when the user","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2690s"},{"start_seconds":2699,"text":"comes into the system and says I want to build a freelance Marketplace that is natural language so there's an llm but the the llm the output of the llm is not","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2699s"},{"start_seconds":2708,"text":"in the language space as it is today but instead it uses this symbolic","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2708s"},{"start_seconds":2716,"text":"um like sort of let's called domain specific language in which it builds some sort of understanding of what the","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2716s"},{"start_seconds":2723,"text":"user wants uh and for for this system for this DSL we want two properties property number one is that it can be uh","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2723s"},{"start_seconds":2731,"text":"converted to a Smart contract SL middleware um in in","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2731s"},{"start_seconds":2740,"text":"uh like without using any leaky llms right so like there's there's an actual routine that converts the DSL into the","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2740s"},{"start_seconds":2748,"text":"smart contract mware and similarly there's a way to convert it to language uh which conveys exactly what the DSL is","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2748s"},{"start_seconds":2756,"text":"right so like what you want to happen is that you go to the eipm eipm builds this DSL and then it communicates back to you saying uh on the high level this is what","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2756s"},{"start_seconds":2765,"text":"I think you want and then the user can say well either yes that's exactly what I want or no that's not quite what I","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2765s"},{"start_seconds":2771,"text":"want uh but once they done uh what a IPM has uh is something that user signed off as what they want which can be directly","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2771s"},{"start_seconds":2780,"text":"translated into into code um the the problem here is that you want","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2780s"},{"start_seconds":2790,"text":"it to be sufficiently low level uh to be able to convert to to the smart contract middleware but sufficiently high level","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2790s"},{"start_seconds":2798,"text":"to communicate back to the user so so potentially uh this like single step is not quite possible maybe it requires","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2798s"},{"start_seconds":2806,"text":"well I think the the more problematic statement is that the there's a specification stage right where I say","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2806s"},{"start_seconds":2813,"text":"like um again I I'll just use some conditions right you don't like for any token uh for every","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2813s"},{"start_seconds":2822,"text":"token right the you know the own the ownership of a token kind of","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2822s"},{"start_seconds":2830,"text":"uh I don't know oh I'll mess up right now the the formal specification but the owner of a token you know uh doesn't","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2830s"},{"start_seconds":2840,"text":"change after transaction unless there is a like clear you know spe like clear","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2840s"},{"start_seconds":2848,"text":"intent by the owner right so that that would be for example a very specific condition that you want to insert that","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2848s"},{"start_seconds":2855,"text":"actually user may not even like explicitly State yeah but the solution to this right is that in every transa in","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2855s"},{"start_seconds":2863,"text":"every function call right in you know in Rust like there is some something that needs to be H happening to satisfy this","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2863s"},{"start_seconds":2871,"text":"so so there is this kind of separation between um um I mean in the way specification right and the solution so","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2871s"},{"start_seconds":2881,"text":"there's an implicit set of assumptions yeah which the user doesn't even think about yeah which which ideally we have just because like I mean they're in the","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2881s"},{"start_seconds":2890,"text":"system overall right um or I mean maybe user does want this to be different right maybe like they run an you know a","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2890s"},{"start_seconds":2899,"text":"whatever something that randomly changes owners every tick of a time but like they need to specify that explicitly so","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2899s"},{"start_seconds":2907,"text":"and so there's the step of going from you know first of all solving the specification with the solution which","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2907s"},{"start_seconds":2914,"text":"ideally indeed the solution is also formal so you can and but you can go to","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2914s"},{"start_seconds":2921,"text":"a actual code and you can have a proof right that this formal solution is indeed satisfies all of the","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2921s"},{"start_seconds":2930,"text":"specifications condition and variance I thought you were saying the problem is that the user will not even notice that well yeah but I mean those can be added","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2930s"},{"start_seconds":2939,"text":"like you know almost implicitly to the to the specification right I mean presumably our IPM is very much like","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2939s"},{"start_seconds":2946,"text":"crypto whereare Crypton native D gen and so they know this right but but we need this kind of","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2946s"},{"start_seconds":2953,"text":"U System and so there is actually a really interesting paper Alpha geometry which is kind of in in this direction I","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2953s"},{"start_seconds":2961,"text":"don't know if you want to describe it right okay so so so I think Alpha geometry is the most interesting paper","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2961s"},{"start_seconds":2968,"text":"today uh which so so if we if we move like one level higher the general um property we","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2968s"},{"start_seconds":2977,"text":"desire from the system is that the system when it does reasoning it does reasoning in the symbolic space where it cannot mess up because in the language space models mess up and people mess up","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2977s"},{"start_seconds":2986,"text":"and as of today we have actually no evidence that uh it is possible to reason in a language space without messing up but we do know that it's","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2986s"},{"start_seconds":2994,"text":"possible to do it in symbolic space but of course symbolic space uh you know it blows up um because uh the search space is huge so ideally what","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=2994s"},{"start_seconds":3002,"text":"you want is you want symbolic space for reasoning and large language model as an intuition and Alpha geometry is very close to that they they say Well it","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3002s"},{"start_seconds":3011,"text":"happens to be the case that for geometry the symbolic the DSL is relatively limited I think it's still massive I think the amount of effort that they put","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3011s"},{"start_seconds":3019,"text":"into creating it was uh it's hard to estimate but if you watch the video they scroll through the definition of it and uh","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3019s"},{"start_seconds":3027,"text":"it doesn't end scrolling while they show it uh but it's uh but it is tractable right they manag to uh to describe it uh","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3027s"},{"start_seconds":3035,"text":"and they also limit significantly what intuition means right it doesn't even happen in the language space it still happens in the symbolic space but they effectively uh I I don't even think it's","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3035s"},{"start_seconds":3044,"text":"still called can be called intuition instead they they just say that every now and then when you solve a problem you need to what they call pull the rabbit out of a hat and say and now I'm","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3044s"},{"start_seconds":3052,"text":"going to introduce this concept which comes out of nowhere but then once it is present the you know the solution becomes tracable well I mean same with competitive","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3052s"},{"start_seconds":3060,"text":"programming right like you have a problem yeah right and like the solution is you use black HS or what something","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3060s"},{"start_seconds":3067,"text":"yeah like a I think in programming quite frequently the rabbit is the whole solution well it's it's like it's it's","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3067s"},{"start_seconds":3075,"text":"some algorithm that if you apply it then everything else kind of fits into place right right and so if we go back either to competitive programming or even to","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3075s"},{"start_seconds":3083,"text":"aipm right so so effectively let's use a am as an example because we already introduced it right so you have the user I will do the same as you did and then","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3083s"},{"start_seconds":3092,"text":"you have code right and so as I said I want and then we want the user to communicate to","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3092s"},{"start_seconds":3099,"text":"the system uh to the IPM and a IPM well I","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3099s"},{"start_seconds":3108,"text":"cannot okay the user sad sadly after after Co we are not used to the whiteboards meaning we are used to","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3108s"},{"start_seconds":3115,"text":"mirror where you can just zoom in yeah and like Drop arrows precisely right yeah yeah and so you you communicate to","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3115s"},{"start_seconds":3124,"text":"the cipm uh which here's language and back also goes language but internally IPM","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3124s"},{"start_seconds":3132,"text":"like if you zoom in into their uh sort of hat they have this DSL right so uh so","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3132s"},{"start_seconds":3139,"text":"naturally uh there is ADSL uh such that any code can be expressed through it like there's a","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3139s"},{"start_seconds":3147,"text":"there's a um I don't remember the name but there's a system built on top of rust that allows you to provide you know","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3147s"},{"start_seconds":3155,"text":"formal guarantees for every function right so we know that there R core itself is actually formally verified but I'm saying like you can","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3155s"},{"start_seconds":3161,"text":"annotate function saying you know the um you know like this function will have this invariant and then there's a compiler that will um verify verify it","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3161s"},{"start_seconds":3170,"text":"for you right however it's so low level that any meaningful like if we convert it back to natural language it will not convey anything to the user right at the","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3170s"},{"start_seconds":3179,"text":"same time for any particular idea right so let's say I want to build an order book there's a D cell which has exactly one operation which says build an order","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3179s"},{"start_seconds":3188,"text":"book right which can be converted back to language and to code right but but it's not very generalizable right and so you can uh you can like","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3188s"},{"start_seconds":3197,"text":"imagine how generalizable it is and how understood it can be by user","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3197s"},{"start_seconds":3207,"text":"right and so here's our example of the order book right and here's the lowle that low level language and so","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3207s"},{"start_seconds":3216,"text":"the question is how this looks like and whether there is U and so like let's let's say there's like a space of sufficiently","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3216s"},{"start_seconds":3224,"text":"generalizable and sufficiently understood by the user like it goes some somewhere here right and like like the question is does this curve touch it or","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3224s"},{"start_seconds":3234,"text":"does it go entirely below uh and if it does go entirely below can we split it into multiple steps or introduce some","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3234s"},{"start_seconds":3240,"text":"other um this is noted to here or can we introduce you know some other components","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3240s"},{"start_seconds":3248,"text":"to the system that will make it look more like this yeah but","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3248s"},{"start_seconds":3256,"text":"hopefully I mean this is we saying if if the spec already is convertible to code one to1 yeah yeah yeah so exactly so","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3256s"},{"start_seconds":3264,"text":"like so this is for the we're saying yes what we get the dsls and for every level level of","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3264s"},{"start_seconds":3273,"text":"generalizability we have an oracle which says what is the best DSL in terms of being understood by the human but I think I mean I think the problem here is","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3273s"},{"start_seconds":3282,"text":"that you have you like the the the What's Understood by human is a problem specification what's what code","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3282s"},{"start_seconds":3290,"text":"needs to contain is a solution right so there is a step of uh like solving yeah which I think that's what I mean you","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3290s"},{"start_seconds":3298,"text":"still you still have the same yeah like for the for the description of the problem statement","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3298s"},{"start_seconds":3305,"text":"uh like how how generalizable is that problem statement how easy it is to modify right because like it's not build an order book it's build a specific","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3305s"},{"start_seconds":3314,"text":"order book that H is in your mind right there a very specific handling of all of the edge cases versus you know if you","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3314s"},{"start_seconds":3322,"text":"modify even one thing now you need like a completely different language versus like the I mean this point with Co with","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3322s"},{"start_seconds":3330,"text":"like rust annotations it's already a solution right it doesn't actually have like a you know right in my hand they not","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3330s"},{"start_seconds":3340,"text":"quite separated I think ideally you want to be in a situation where for problem statement you can go in the symbolic space to the I mean I think they should","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3340s"},{"start_seconds":3347,"text":"they should be in the same language space but but I do think there's a which Alpha geometry was the case the problem definition was in the same space as the","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3347s"},{"start_seconds":3356,"text":"but I think there like there's an expected at least in my for me there's an expected step of that the solution that is convertible to like this is one","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3356s"},{"start_seconds":3365,"text":"to one very explicitly yeah this actually requires like yeah and this is where you need symbolic reasoning with","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3365s"},{"start_seconds":3372,"text":"the intuition yeah which like this is where you need to come up with this and then prove that it's actually satisfy as","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3372s"},{"start_seconds":3380,"text":"oldin variance y yeah so this is where like actually like","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3380s"},{"start_seconds":3389,"text":"solver coming and uh yeah I think like why is this is important is because if you think of any","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3389s"},{"start_seconds":3397,"text":"code that is written all of it is so far been written by people and people have I think like there's like two three bugs","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3397s"},{"start_seconds":3404,"text":"per 10,000 lines of code something like that that sounds extremely optimistic okay maybe play a thousand lines of code","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3404s"},{"start_seconds":3413,"text":"so so the the CH I mean well written code with lots of tests Etc yeah and so the problem is like you know there's","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3413s"},{"start_seconds":3421,"text":"millions and millions of lines of code right now in your you know like you can you can have a net you have networking stack you have operating","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3421s"},{"start_seconds":3430,"text":"system you know and you have your browser and then you have your applications right assuming everything","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3430s"},{"start_seconds":3439,"text":"is just web apps uh like in all of this right starting with networking there's vulnerabilities and as the progress of","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3439s"},{"start_seconds":3446,"text":"AI actually goes forward AI actually will be way better at finding bugs already in the code kind of just because","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3446s"},{"start_seconds":3455,"text":"it's like a statistical problem and like the introdution problem versus doing this versus like writing a correct code","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3455s"},{"start_seconds":3462,"text":"on itself again with LMS right now and so so we're going to see more and more vulnerabilities being found and","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3462s"},{"start_seconds":3469,"text":"potentially exploited by um by the using llm that's one of the","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3469s"},{"start_seconds":3476,"text":"kind of the AI risk uh directions right that uh people exploring and so this is extremely important because in a way we","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3476s"},{"start_seconds":3484,"text":"need to rebuild the internet with this and cryptography and other pieces kind of from scratch to ensure we have kind","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3484s"},{"start_seconds":3492,"text":"of not breakable internet with AI so so to give a little bit of a different perspective right like Linux networking","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3492s"},{"start_seconds":3501,"text":"stack certainly has vulnerabilities there's absolutely no way that that many lines of code were written without vulnerabilities uh and the way I see the","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3501s"},{"start_seconds":3509,"text":"security model of the of the networking stack is that a very smart person cannot find that vulnerability right because a","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3509s"},{"start_seconds":3516,"text":"lot of very smart people looked at it and they didn't find them um however um the model can become much better than","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3516s"},{"start_seconds":3525,"text":"people at finding vulnerabilities way before we it becomes like AGI or it because that's a very narrow uh yeah it","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3525s"},{"start_seconds":3533,"text":"can run the code it can write tests it can simulate it like somebody is really targeting that specific problem yeah like like as an example of that the","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3533s"},{"start_seconds":3540,"text":"models they very good at writing code right not they still not applicable in many cases but but clearly if you show the","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3540s"},{"start_seconds":3548,"text":"current model to a person like in 2022 in the early 2022 they would be mindblown right and so correspondingly","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3548s"},{"start_seconds":3555,"text":"but but in some other areas of um of something that we consider to be trivial the models still mess up right so so the same the same argument applies like they","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3555s"},{"start_seconds":3563,"text":"will become maybe they already have became become uh much better at U breaking code in people right and uh","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3563s"},{"start_seconds":3572,"text":"well finding a zero day in the networking stack of Linux could be relatively um disastrous colossal right","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3572s"},{"start_seconds":3579,"text":"for the for the world right and so we either uh need to ensure that uh the good guys find them first","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3579s"},{"start_seconds":3588,"text":"and and Patch it um or I I I don't think we can in any way prevent the models from developing in","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3588s"},{"start_seconds":3596,"text":"into that in that direction or yeah or or like alternative would be to just uh build something yeah build something","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3596s"},{"start_seconds":3604,"text":"that is you know fully verifiable and just rewrite the whole networking stack and you know preemptive preemptively deploy something that is","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3604s"},{"start_seconds":3612,"text":"provably correct yeah I think in general like from the kind of yeah general safety approach right the whole point of","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3612s"},{"start_seconds":3621,"text":"like ma math is that yeah like the you know if you can have a proof and it's you know logic sound then there's no","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3621s"},{"start_seconds":3628,"text":"alternative way to kind of work around that with doesn't matter how much more intelligence or compute you have yeah versus if you have anything outside of","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3628s"},{"start_seconds":3637,"text":"that right outside of like formal logic and kind of mathematical proof potentially you can kind of brute","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3637s"},{"start_seconds":3645,"text":"force a way around it hence why again cryptography being one of the fundamental blocks as well um of","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3645s"},{"start_seconds":3654,"text":"creating this system I mean and right now it's used everywhere here as well and it doesn't help that the networking is written in C no which is which is","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3654s"},{"start_seconds":3663,"text":"generally the disaster language yeah well and so is a lot of Linux in general yes exactly yeah so yeah so we have a","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3663s"},{"start_seconds":3671,"text":"lot of work um we'll start here because uh there's a clear value there and and and the I mean smart contracts","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3671s"},{"start_seconds":3678,"text":"themselves are smaller and so at this point of the you know solving this problem space uh there's also just like","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3678s"},{"start_seconds":3686,"text":"a lot more people and and kind of um coordination contribution we can get through building applications but the","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3686s"},{"start_seconds":3695,"text":"idea of this model and of the system is to go down the stack and I mean that's why I kind of started talking about you","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3695s"},{"start_seconds":3702,"text":"know user owned internet because that's kind of what we need to do right like across every stack leveraging cryptography leveraging developer and uh","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3702s"},{"start_seconds":3711,"text":"rebuilding it in a new way I think we're approach ing the time or we already over time so uh thanks for","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3711s"},{"start_seconds":3719,"text":"tuning in hope this was educational as Alex said we have not prepared but uh we will promise you we","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3719s"},{"start_seconds":3727,"text":"will continue not being prepared uh but uh you know you will actually glean into what we're thinking discussing kind of day today uh and what we're working on","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3727s"},{"start_seconds":3736,"text":"so tune in for next time thank you","url":"https://www.youtube.com/watch?v=7kc4Mv75u5Y&t=3736s"}],"youtube_url":"https://youtube.com/watch?v=7kc4Mv75u5Y","url":"https://near-pedia.xyz/videos/7kc4Mv75u5Y"}