{"id":"1jLH5GLp5fc","title":"NEAR AI Office hours #23 - EigenLayer w founder Sreeram Kannan","description":"EigenLayer Build open innovation ∞ play infinite sum games.\n\nIn this episode of NEAR AI Office Hours, Sreeram Kannan, Founder of EigenLayer, explores the intersection of AI, blockchain, and decentralized coordination. The discussion about the future of AI-driven innovation, the role of blockchain in coordinating research, and how decentralization can help solve the tragedy of the commons in large-scale AI development.\r\n\r\nGuest: Sreeram Kannan\r\nhttps://x.com/sreeramkannan\r\n\r\nProject: https://www.eigenlayer.xyz/\r\nTwitter: https://x.com/eigenlayer\r\n\r\nNEAR AI: https://near.ai/\r\nIllia Polosukhin https://x.com/ilblackdragon\r\nAlex Skidanov https://x.com/AlexSkidanov\n\nTakeaways:\n• Decentralization & AI Coordination: The future of AI lies in coordinating sovereign agents, not just GPUs, to unlock innovation through decentralized systems like blockchain.\n• Tragedy of the Commons in Large Firms: Traditional centralized models face internal inefficiencies as companies scale, creating diminishing returns in large AI labs.\n• Crypto’s Role in AI Innovation: Blockchain provides a commitment layer, allowing for flexible, decentralized coordination that can drive faster innovation and incentivize individual contributions.\n• Free Market AI Development: Kannan argues that a free-market approach, where smaller AI models can be built and valued independently, is more scalable than relying on one monolithic model.\n• AI Safety & Future Challenges: The rise of AI poses serious risks, including security vulnerabilities and the potential misuse of AI for harmful purposes, requiring careful coordination and new safety frameworks.\n\nTimestamps:\n[00:00] Introduction – Kicking off the episode and introducing Sreeram Kannan.\n[01:06] Sreeram’s Background – From computational genomics to AI and EigenLayer.\n[04:00] Discovering the Power of Neural Networks – A pivotal moment that shifted Sreeram's view on AI.\n[06:20] Coordination as Humanity’s Superpower – The evolutionary significance of human coordination and how blockchain can enhance it.\n[09:25] The Role of Blockchain in Coordination – How blockchain combines communication and commitments to facilitate decentralized coordination.\n[11:30] The Challenges of Decentralized AI – Why decentralized research and AI development are harder to achieve than simply coordinating hardware.\n[14:10] Free Market Models for AI Development – The advantages of a free-market approach to AI, where independent models compete and contribute.\n[20:00] The Tragedy of the Commons in Big Labs – Internal inefficiencies in large-scale centralized AI organizations.\n[24:30] AI Safety Concerns – Kannan and the hosts discuss the potential threats AI poses, including security risks and the need for formal verification in code.\n[32:00] EigenLayer and Its Vision – Sreeram explains EigenLayer’s vision for decentralized validation and its role in AI computation.\n[45:00] Closing Remarks – Final thoughts on the future of AI, blockchain, and decentralized innovation.\n\nJoin NEAR's community:\nWebsite: https://near.org/\nNEAR AI: https://near.ai/\nTwitter:   / nearprotocol  \nBlog: https://near.org/blog/\nMedium:   / nearprotocol  \nGitHub: https://github.com/near \nDev Docs: https://docs.near.org/\n#NEAR #nearprotocol #NEARAI","channel_name":"NEAR Protocol","published_at":"2024-10-22T00:00:00.000Z","thumbnail_url":"https://img.youtube.com/vi/1jLH5GLp5fc/hqdefault.jpg","transcript_text":"near to one we're close to won't stop us now together we'll redefine the game with near will find the flame empowered Minds rise and store use your own AI we'll explore hold the future hand in hands in this Tech driven Wonderland okay hello hello episode 23 holy crap this is a lot of episodes but uh uh it's really amazing to have shiram here uh we're gonna have an amazing episode diving in into AI I'm sure we'll touch on Ian later and uh lots of fun stuff I'm I'm sure lots of people already know shiram but do you want to introduce yourself quickly and absolutely thank you so much like really excited to be here uh start as an introduction I'm shiram I started the Ian L project like three and a half years back uh and uh used to be a professor at the University of Washington Seattle before doing uh crypto I used to work on computational genomics and through which I got into to uh AI a little bit we even had a spin-off from our lab uh you know startup working on applications of AI to 5G um but super excited to be here it's one of the my most uh passionate areas so and I don't think there's anybody better than Ian and Alex to discuss this with I was just telling them before the conversation that I am a fan of the white board series that near did like four four years back learn everything about every protocol out there and how how they think about designing it so this is a modern version of that with the with yeahi stuff so super excited sprinkle yeah not sprinkle AI everywhere and like sprinkles yeah well I mean first of we haven't had actually a whiteboard about igen l so I think no that um that's a mistake yeah well we're GNA be in Bangkok but I actually just F to mention because you do have like I think I don't know if it's your one of your lat last p i mean last non-blockchain paper papers but you had the most cited paper is improving Federated learning uh personalization so uh yeah you have quite quite a few of uh kind of machine learning papers back in 1819 which somehow gets more cited than blockchain papers I don't know what it's clear right like there's no everybody in one of the reasons I switched from a you know what happened was I my PhD was in peer-to-peer Wireless Systems and and know peer-to-peer Wireless didn't go anywhere so after I finished my PhD uh even though I we Sol some like theoretical problems in that thing I was looking for what's a high impact area and then got to genomics which I entered into and doing like protoc you know algorithms for an analyzing genomics data mostly statistical learning and as as I was doing it you know the funny episode is in 2017 I used to have like a very what do you call cursory negative opinion that like you know armchair academics used to have about like neural networks and Ai and I'm like my student came in and we were doing this thing where you you have a DNA sequence going through a nanopore and you sequence what the the DNA strand is and uh you know we were doing like all these statistical learning algorithms we were working with actually like uh the the physicist who had the strongest uh empirical measurement of the gravitational constant G so I was very confident about our physics models about our inference models everything and then my student comes in and says hey I'm going to use like you know rnns to like do this thing like what kind of an idiot are you like you know maybe good for things which don't have mathematical models you know here we have like the absolute best model so like you know you're going to go knock yourself out waste your time two weeks later he beats our Benchmark okay this is that so I my estimate of like it just like flipped within that thing and then I said hey can you invent like communication algorithms using the same thing and then he's like oh yeah you can and then like we we did a bunch of stuff there and that that became the spin-off he became founding engineer there and then I decided man this is going to like take over our job soon so I need to find an area which doesn't have this same issue and my model for AI became basically anything that you can actually evaluate the optim the objective function easily I call this a high-speed Oracle where you can evaluate how good your output is quickly you're going to optimize the heck out of it like it's not it's so that was my like operating model and then like okay what is going to give me like a five to 10 year Horizon to do something I'm like okay maybe blockchains and cryptography and stuff you can't like evaluate whether a cryptographic scheme is good just by like empirically simulating it you to have some mathematical proofs which as we we can talk about they're getting really good at so you know we have like a few more years here to do something fory our blockchains what that that's interesting yeah kind of the opposite of ours which was like hey we need we need to coordinate a bunch of people to do the machine machine learning so and like the tool to coordinate is you know like peer-to-peer payments and kind of uh some some form of crowd crowd work that facilitated by blockchain yeah amazing I I you know somewh unrelated but I just want to call out when I remember uh recently the a16z uh report came out and I saw near is you know number two on like number of active addresses Kudos like you guys kicking ass and getting there on the way to coordinating you know agents and people to to do this so super excited for that but you know I I mentioned the kind of technical reason why I got into crypto but also I think that there was a similar kind of philosophical reason which is that you know I used to have this mantle model that like the most important thing people do is you know we're intelligent and be innovate so Innovation being the highest value so you can think of like AI as giving super super powers to our intelligence our Innovation abilities and I kind of flipped from that model to a different model which is that yes you know you have Innovation and intelligence but the the defining evolutionary advantage of humans is our ability to coordinate flexibly and so accelerating coordination is a different value and maybe even more fundamental than accelerating Innovation or intelligence and particularly the thing to think about coordination is you know I think of like the long Arc of evolution you know Innovation is like you know this exponential growth one building on top of the other and then but suddenly coordination is this Quantum Leap so you suddenly have a new coordination mechanism and then everything like flips and as an example like you know in in like a cell you have you know a procario to ukar right this kind of a transition where like you know inside a cell you don't have a nucleus now you have a nucleus specializ ing in like information storage and mitochondria specializing in energy production they have their own dnas but they coordinate because they they live and die together this kind of a thing happens very rarely and then once it happens it just explodes the scope of innovation that can be built on top so maybe maybe a close example is you know pre- internet and then we have internet and like the coordination on internet right just exploded so yeah so great great example internet the the way we think about like I I started thinking about like coordination and the internet is internet so the coordination requires two things communication and commitments communication the internet basically played the major role where like we can send information to everybody else for example you could not train any of these AIS if you don't have the world's information on the internet at your disposal right so you know that was one kind of like base level of coordination the second one is commitment like you know we can come together and coordinate only if not only if we can communicate with each other but also so if we can enter into binding commitments hey we produced this episode we we spoil the shares equally whatever the division rule is and then like that's a binding commitment once we know there's a binding commitment the game theory of cooperate versus defect completely changes so the internet the communication super substrate crypto our commitment super substrate and then together you can actually like now build arbit new interesting coordination structures on top Alex you want to say something before no no are you waiting for me I see yeah we need a better coordination mechanism here um yeah so that was basically like the reason we kind of I got into crypto the philosophical kind of reason being coordination super structures the technical reason being like you know this these things require cryptography and cryptography is not going to be like you know firstly broken by AI or invented by AI at least for the short while so it gives us a little bit of an edge for for us to do something I think now like as I've gone deeper into this Rabbit Hole I have like deeper and crazier theories and like we can talk about some of them we we Aller crazy theories but actually so so staying on the topic of coordination because one of the things with Alex we always discuss is um the kind of for for you know we we have this like user own internet right as like a a main thing that I think like all of the different blockchain work is fitting into and as part of that you need a user own AI which is like optimized for you as individual but and like everybody's kind of okay let's you know have decentralized training right let's you know train models in decentralized way because that gives us kind of lots of guarantees but the challenge that uh is really there is actually it's not the coordination of bunch of gpus to train the model that's hard it's actually getting to really good model that's hard that's like the reality is there is you know there's open Ai and like still pretty big gap you know meta Google uh and there's like hundred other companies that are training models right now and they're not like actually matching right so like they don't even have the problem of coordinating gpus they just like buy a cluster and training it the getting good model is the hard part and um so I think like the coordinating of research is kind of the conceptually missing piece and you know right now the only examples that we've seen that work is you know you staff everyone in this the same office and you know you inspire them that they building AGI and now they're like you know crashing it but like that seems like you forgot the multi-million dollar uh pay yeah and then and then you run a sorry for-profit company um yeah so so so the question is like kind of what what is what what can we bring right from what we've learned I mean also like you've been in Academia coordinating a bunch of people doing research right I mean we actually worked in those organization centralized organizations on our side so like what can we bring kind of Best of Both Worlds leverage blockchain leverage all the tools we're building uh so that's kind of the topic I thought we can like dig in a little bit absolutely la la The Prompt um so you know I think you know the phrasing I I would like to rephrase we I think coordinating a bunch of gpus is not interesting okay so let me just start with some slightly controversial opinion right I think it's it's similar to what you said but just phrasing it like that coordinating a bunch of GP is not interesting why anybody can buy gpos okay you can buy gpos their their commodity and there's an economy of scale you can just put like you know Elon Musk is buying a bunch of them like people just go buy them right it's really easy easy to coordinate because they're not sovereign agents coordinating Sovereign agents is what is really interesting and you know the not yet talking about AIS but humans right so when you think about it like building AI one model of coordinating agents is hey everybody come here you all have equity in this like common Enterprise you know open AI whatever like you know number go up let's make the number go up together by building this common you know objective everybody comes in they contribute and then like they can scale um firstly I think this is reliant on this coordination structure the company as the core unit of scaling human coordination and if you think about it actually like back in 2017 some people had observed it but I think it is kind of like there are many like gems of ideas that have kind of Fallen by the wayside of like crypto and one of them is this theory of Kus law like Kus law is like why are companies sized at the size that they are and you know Ronald Co and Economist right wrote this paper the theory of the firm and it's basically like oh you know company is this big because you know it's easier to coordinate individuals internally than it is to coordinate individuals across organizations and this lays the foundation for how open a Google Deep Mind everybody else is operating today is hey you know we're coordinating by the super structure the equity Corp and we are going to actually okay so and you start thinking about it and you know you've been there I my I've never worked in a company before so I have limited experience but I have friends you know I'm not living in under a rock so saying I can layer Labs is not a company uh the not one I made up the ones that existed before um the the way to um think about the um the the problem is when your company grows big right Beyond 100 you know at Beyond thousand what happens is there's an internal tragedy of the commons right because you have inside the company your contribution is marginal relative to the contributions of everybody but you receive an equal payout irrespective of your contribution and this disincentivizes the most contributing ones unless the company has a very strong internal system of like incentives which is very difficult to build because you know it's a collaborative Enterprise that it's fundamental so what happens is like you know you have this marginal incentive problem every unit of new person contributing to grow the stock or you know output doesn't feel their output impact their like you know the outcome equally so so there's an internal tragedy of the this is why we see like companies do really well at small scales you know at a reasonable small valuation 10 to 200s at like a reasonable big valuation 100 to th and then you go beyond that it's like massive tragedy of the comments like people are doing rant work they're doing like skating on minimum work all kinds of things happen they just fight with each other all these fundamental problems come together and what we do know is the best mechanism for resolving units of value is the free market like this is the biggest like lesson we've learned in coordinating humans over the last 300 years right free market puts a number a value on each unit like oh how much is near valuable how much is I Like You can actually like do this because the free Market's kind of like expending energy in allocating value and betting on oh you know this unit is actually producing more let's allocate value more so the thesis should should not necessarily be to coordinate gpus it should be to coordinate humans in different architectures where you can maximize the net output rate of growth and rate of innovation and so what are things that can actually do this so zooming back to your pro problem can we build uh the universal AGI as a composition of smaller models like that's one like you know angle at it right you say oh you know there is some context free general intelligence and then there is contextual specific intelligen is the best intelligence for evaluating where a tomato from a best intelligence for you know say singing a song in Tamil which is my native language like you know each one could be something very specific we do know this like from General observation that specific intelligen the you take the person who has the most general intelligence like IQ or something and they never played chess and asked them to play chess they're going to lose to somebody who has lower general intelligence but higher contextual intelligence right so they've just play many many years they remember all kinds of like structures and movements so this thing about like are there architectures of universal general intelligence which are compositions of like some some smaller units of intelligence and if there are let them build in the free market and coordinate value across a blockchain is way better then we have to get these 10 million people sitting together in open AI suffering a tragedy of common and so this is like our thesis of crypto and why I think why I think it is you know uh I think crypto is going to impact every endeav in every human and free agent Endeavor and specifically at the Endeavor in AI is that you create units of work and each unit of work is measurable and attributable by the free market for example the general intelligence Tunes to like different composable models and allocates value based on like how much each model is used now if you build a model and it's kind of like sitting on the super structure and you know your model is used more your token your like you know unit of value acral increases so this is I think the kind of my thesis for where like you know how AGI will be built on the free market not on as compositions of individual agents rather than like one you know giant agent I think it's also better for us it's minimizes risks there's lots of of other like second order consequences but the first order thing is I think this is actually maybe the best way to coordinate like you know 10 million people to like come in and contribute to this common end yeah I uh I've been thinking uh for a while along the same lines how do we um so effectively if you think if you think of uh um like a large lab working on AI they have let's say they have 300 tier one researchers working full time but but the number of you know important breakthroughs happening is relatively small like like let's say let's say they have like 10 10 10 impactful breakthroughs per year that move the needle right so that means that first of all that means that bringing in you know something on the order of 10 people who are good to work on an effort is relatively meaningless right because the the the you know the expected time before a breakthrough is relatively high right so you want to but also the problem is um those labs are capable of bringing the absolutely best talent right but in the wild we will have it it will be in order for us to capture people of that quality uh it will need to be an even bigger sample of people right so we need to find a way to orchestrate thousands of people potentially working on the same goal which is a um hard especially given you know you know the the there is a little bit truth to that there's also a little bit of the opposite to that which we're seeing already in open AI right like many of the the the founding team is going out doing their own thing it's a kind of like a sense of a signal of the tragedy of Commons right hey we come together towards this common Enterprise and then like hey actually I'm the one who's doing all the like why should I like you know and I think this is the the the fact that each one of those like 200 whatever people that you're talking about they believe their thing is going to be the thing that moves the ne the most and I as long as there is a kind of coordination super structure that can help like compose the values created by each one of them in some like meaningful way it might so anyway just like a observation there on the inter like an empirical observation on the tragedy of Commons which is that they can bring these people together but to keep them together and you know because of the kind of same type of problem that I I think it's already happening empirically yeah it is definitely it is definitely the case that lot of people who uh like the employees uh how do you distinguish between those like you know people who actually dig the the trench a lot of them start companies right perplexity is a good example right our event was open back in 2022 I think but uh yeah whether whether people on the high level live for that reason that's a that's a different question I think that but maybe like Alex if you want to contrast because I mean we had that conversation like hey is this going to be a lot of kind of specialized AI let's just call them agents models you know Services systems that you know you kind of interacted coordinate with or this is going to a big model that you need a lot of human intellect coordinated together to advance kind of what it does right because I think that's like a a pretty polar uh view on on where this is going and how to coordinate these efforts at the same time it maybe that both are true right they don't mean like that you know either or but I think it would be interesting Alex if you want to highlight like what's the other like hey a big model coordinating lots of researchers to kind of push the envelope kind of look like interesting so so I definitely do believe in one model future I uh I I don't know how how important my opinion is I don't think we have enough information to predict right it's just a good feeling that it will be it will be one large model that is connected to all the apis and whatnot and uh just like if I you know like sort of GPT 5 scenario but hopefully it's not going to be gp5 hopefully it's going to be you know some sort of like like either intellect one or whatever I I don't remember the name of the prim intellect one or something like that right something built by the community um but I don't uh the the part about that model orchestrating researchers is a um I don't remember the reference but I do think that it's a the interesting problem there is that you don't need to just lot of researchers you know the each each unit of research requires resources right you cannot you cannot just like a is alchemy in the blockchain you know I can be sitting in a like you know case in point the consensus that is driving near today I came up with it sitting on a beach in Puerto Rico right I didn't need any resources to to come up with it right but but a is alchemy right so your hypothesis is not worth anything until you actually tried it and and something you know on low scale doesn't necessarily you know transfer to the high scale so it's a it's interesting how do you orchestrate if you have thousands of people you probably don't have resources to actually power them you know continuously pushing so it's it's an interesting problem not not quite a it is a great uh great prompt uh there are I I think the the scaling problem right which is that basically unless you have a lot of resources you can't even experiment because these are emergent at only large scales I think that's a really uh you know pressing problem um the so going to like aia's question of whether there's going to be like a single model to rule them all or like you know is there uh compos composition of models that and and contextually switching across them is going to be better I think it's going to be like the basic line which which shows whether like the approaches in crypto can succeed versus because you know if if it's one model it's much better like you know you have thousand people sitting in an office and trying to kind of hash it out right like there's no no world in which like a a committee based approach for like what that model is is going to work as well as a kind of you know much better organization that so I do think that that will basically draw the boundary it might also be that you can actually build composition of models which are pretty close to the to the or original one like it's maybe the the big Central one is the first one that is produced and in six to nine months or one year whatever like there is an alternative which is built on this composition Theory because at the end of the day you know that is a kind of like a internal composition right a big network is made of many small networks connected in some more complex ways but this um the yeah so so the more uh we the other thing that I think about there is how do we make sure that so there is this open closed asymmetry the open closed asymmetry is closed can copy from open open cannot copy from closed okay like this is a fundamental kind of problem it's a it's old problem in like you know uh open source software and you know copy left was kind of born as a kind of result like you do a GPL license and then GPL you know close Source can't copy from GPL but what we need is even GPL right like didn't give incentives right so I contribute code to GPL I don't have incentives other people can create GPL they don't have incentives so that whole thing kind of like dies out instead it's much better to build on MIT or some other like non- copy left license and then like have versions which are like you know closed and like this is the open core or whatever model that you know software has kind of tended towards I think reexamining and understanding if there are improvements to be made in that level can also play a role here because you know if the open closed asymmetry is not broken you have the problem that you know always the one Central model will incorporate every idea that's there in the the open decentralized world and so the one one thing we've been playing a lot around is like a new licensing model we call PDL permissionless derivators license the idea of permissionless derivators license is you create a software and anybody can create a derivative of that software but both the software any derivative created has to be launched on the same platform so the PDL is Tethered to a platform which means like I take this software and then I that this this platform helps let's say do influence whatever of that software and if I any fork or like derivatives have to be launched back on that platform what this solves is the value uh measurement problem right the value measurement problem is you have a model I created a derivative now I launch it on the same platform now which means this platform is measuring how much value how much utility each of these derivative models have and then can pass like relative value across ancestors of models right so I think a lot about this question which is you know permission Innovation can actually grow things exponentially I can build on top of you it's one of the reasons where like you know in newbs you know one of the big you know AI conferences which is where I heard about Helia paper first I was at newbs 2017 in Los Angeles and the crowded crowded poster stand uh I I don't know if IIA was there or not IIA were you there I was there yeah I mean I was we were Manning it I think in two shifts yeah yeah okay I I badly knew anything about any of this stuff but I did see a very crowded stand in Los Angeles it was also like a Tipping Point for me to know like okay this is going to go crazy and you know I was November 2017 I think November December 2017 and then I shifted to crypto but uh the uh the the reason I mentioned that is when we're thinking about these model derivatives if you can allocate value to like these derivatives and each small unit of improvement can ACR value because they're being built one on top of the other it's actually like way better for for like people to contribute back to the commons the reason I mentioned nebs is I used to observe this trend which is you know I come from a kind of very mathematical oriented subset of Statistics information Theory and where like people take three years to publish one one paper and what new set is a new pattern where like every three to six months people just like batch whatever ideas they have and then publish it and initially I thought oh these are like you know this is not a good idea but then now I see that as actually the accelerant for the rate of innovation because you you you publish like small snapshots other people can build on top of your Snapchat rather than only you can build on top of your own things so you have way more p but each paper is uh you know the net rate of growth that you have as an area actually grows much faster it's the same thing with Innovation right do I need like everybody to come together and build this one big model there are only like 10 companies in the world that could try to do that whereas if each of us can contribute a diff on top of like an existing model oh this is a fineing this is this thing this is that thing and then like each one like passes value across through the ancestors then I think you get like a much more interesting thing where as a creator of like new ideas new models you you're also incentivized heavily and maybe each of these like new models has their own token because you know all value allocation kind of can be passed back to the kind of like the the thing all of them happening on chain programmably like so there is I think uh really interesting things to be done in this intersection at least as an experiment but the thing that I feel like very deeply is we have a very narrow time window to do this so I'm really glad you guys are taking like a leadership position here and pushing this and I think we have a very narrow time window to actually like do do this I think 2027 is going to be crazy so just backing out from there we have like two three years and it's if if we can show that crypto is actually can be very powerful in these this kind of a setting there absolutely no other better use case for all of crypto yeah I mean you kind of spoil some of our talk talks at redacted so uh I don't know how much CH want to spill here but yeah I think the kind of the way we've been thinking about it is is very much aligned with this kind of from from multiple directions one is we don't know which of these two worlds will actually win and so as as kind of uh you know being now in financial industry we hedge and we do both and so we need a way to coordinate AI agents and even Beyond AI agents how to coordinate any Market participant um effectively and create effectively a protocol to for them to exchange information pay Char have the P have the commitments um to do that and then on the other side is well it's still this is still about research and Innovation and how can we accelerate the speed of innovation in this open way right and so as you said nips kind of established a new setup I mean it's actually I thinkr with uh open review. net right actually y I think Yan leun introduced this idea of like even more open even faster like put your preprint publicly let's just go you know let's just on it I mean uh comment on each other papers in public and created a Twitter like environment for academic papers yeah but but it kind of it's all about like accelerating that feedback loop but the thing is like what we see still is like if you work at one of the big Labs you actually see the source code of that paper right even before it's published and you can like build on top of it but in open source right now it's all completely like everywhere and and the the you know the Frameworks are different the data is different you cannot actually like always compare them Apple to Apple even when you do you like you know the data prep was different the prompts was different something always is different right and so how do we actually establish a like a kind of you know coherent enough framework that you know is extensible enough for people to build on uh how do we promote this more faster like hey I want to experiment out here's like here's a finetune or here's one thing I did the funny part is we actually I started thinking about the license as well I think it was like last year because that does sound like something very much like Hey how do we actually like yeah BR that I think I kind of got stuck because I talked with a bunch of people and for license to really truly work you kind of need uh you kind of need to buy in from commercial organizations and Commercial organizations hate like new licenses and like their lawyers are really like you know their brain melts when you like introduce anything uh and uh and like for non-commercial right you you I mean you end like you kind of don't want to like over like if it's a research you don't want to actually like charge them effectively so you want them to use it so you kind of like the same reason why copy left I think doesn't function very well is because the people who actually want to open source stuff already don't care about the financial like directly and the people who care about financial they use open source more to get the attention and then they end up building stuff close source so the question being what can we do actually to have maybe not open source but provable compute or provable you know uh processes provable things and then incentivize that and kind of create an alignment where may I mean some things I end up sharing between people reuse it Etc so you can track it but at the same time the outcomes are actually uh available to everyone permissionless but not open so like open source right and that's an interesting you know on its own kind of challenge uh and then on the other side yeah like how do you coordinate research to be faster to like build on top of each other um so yeah I mean I think there is a there's also another angle I faced when talking about licenses is like the the the crypto Cipher punk people don't like oh you know IP is and this and that but I think it's important to um to create like in so my like one of my highest kind of values is uh is you can you know in a crypto language you'd call incentives in like my traditional language I'll call it kerma which is like if you contribute more to the commons the commons should give back more to you and if you contribute like negative things to the commons the commons should hit you back and this is a very important thing to regularize these common you know systems of Commons and when you're creating knowledge for for example and Publishing it in the comons and right now if you don't have any incentives to do it you know a person sitting and thinking hey you know I I go join open AI I get $2 million a year I just sit and contribute here in open source like you know it's like I have to do this like on the site right like it's not not exciting and so yeah totally so there's a lot of angles in getting this right and you know is is it is it like do we need more powerful onchain mechanisms like private private interest and other things can actually like help a lot um one uh question I had for you guys is uh you know where are you guys on safety right what's AI safety Doris like how how far on the curve are you guys I uh I have my own opinions about safety uh it's it's interesting I was uh I was talking to I was at a dinner and there was a bunch of guys from different safety Labs one one was naturally and the one was a I wouldn't remember the name but it's something like you know nation National AI safety Institute not like that but you know along the lines same embedding uh and both of them were effectively thinking towards how do we make the model we train safe right so in case of entropic they say Hey how do we make CLA not accidentally not say sorry you know at the end of the sentence uh or you know the the National Institute didn't have their own models but they were like working with the labs you know how do we make sure that their model is safe which I don't think is the concern at all I think the concern is how do we make sure that a threat actor who has access to 100,000 gpus doesn't train a model that will annihilate us all and we cannot do anything about it so you know from this perspective uh the way uh I think myself personal about safety is how how can we how can we address a particular problem that will arise right so I will give you an example of of this thinking right one of the big so obviously misinformation is a very big place where models will make it even worse than it is today and it's already quite bad uh but another another place where AI in a short term can do a lot of damage is uh is effectively breaking the internet in the sense that all the code that the internet is built upon uh it's like you know millions of lines of code which is uh written over many years by many people it naturally has a lot of bugs right and the security model of this code is that a very smart person by looking at this code will not find it right because a lot of very smart people looked at it and they didn't find it and and like every now and then you know like hard bleed happens where you know someone does find something right but the thing is that the model doesn't need to be a GI complete uh in order to start finding bugs better than us right so you know like it would like models already clearly showing faster pace of advancements in code in other areas of life right so for one they are better than us of painting bugs the only question is who will Who will invoke it first a good person who will immediately patch all of the issues right or a bad guy who will just uh right and and the bad guy could be different here it could be a bad guy who who seeks profit right like like presumably you know it's it's my assumption I have no idea but some of the two L agencies probably have zero zero days which allow them to SSH to my machine right but they they're not going to do that or maybe they will but they're not going to reveal the existence of those right uh but but you know if some script Kitty or or like someone who is just you know H yeah not a reasonable actor gets access to that right that then we can get be in very big trouble where where you know like you can just bring down the whole internet by finding a zero D in the networking stack and would like like would like recovering from this state when you cannot communicate with people anymore so that's one of the fears they have and we can work directly towards addressing it right like uh the simplest way would be investing right now heavily into formal verification and you know start rewriting or very you know formally verifying bits and pieces the most fundamental ones so that by the time the models are good enough to penetrate our defenses we have formally verified stack you know that powers the core of the internet and then you know we know that it's very unlikely to to be broken so so that's something that I'm uh I'm personally spending a lot of effort know going that way super interesting I I I think of like this fundamental problem in in the following way think about like any kind of complex system it's like there is a layer of free agents and Innovation built on top of like some kind of a coordination substrate usually it's the coordinations off things like the government and the uh the free agent economies like the free market and but I think uh so what's happening is technology like particularly AI will accelerate massively the free agents right so that's that's what we're seeing and the free agents are going to find every kink in the coordination structure and then try to break it right like that's that's the issue that that we are worried about and I think the the solution that uh some of these safety people are gravitating towards is let's go try to get the government to regulate it better or whatever like you know that that thing and I think instead I think what we should do as a kind of crypto space as trying to Pioneer coordination mechanisms is what crypto coordination mechanisms can we kind of bring to bear on this uh on this problem and right now actually my D you know diagnosis is it's actually making it worse okay like imagine if Alex's attack goes live right so an agent knows how to attack the internet it's going to first like try to steal a lot of people's e and BTC I think uh we lost yeah yeah we lost you for a couple minutes when you started talking about the plan of the agents they they shut you down so that you don't reveal it accidentally yeah I think that continues yeah much Silicon Valley internet you know Seattle internet and uh like everywhere else yeah that's fascinating how you know two of the two of the capitals of the of the tech in United States have the worst internet okay I guess we lost we're 30 seconds away from from from ending the episode anyway right what where is it okay you have closing thoughts well I think okay we're back we're back to worried the agents know we're talking about them and EXA I was just going to say that you know right now crypto is actually potentially making it worse not making it better because you know if if you do the Alexis attack and it it's going to steal our e and BTC and then these are sensorship list and immutable systems and you know it's it's chaos so I think we need to think through like for example something that we think about a lot at Aon is Fork cability right what set of fork ability conditions should exist for these systems and for you know one of the ideas we've been toying with is we should write asimov's LW something like that into the FOC ability conditions of the chain so this is I think another like way of thinking about it is like how do we set incentive structures where like these agents want to play our game rather than co-opting us into their game right so we set the incentive landscape and they play according to that is a kind of Meta Meta idea I'm fascinated with uh I think we we did a lot of high level stuff I'll just spend at least 10 seconds explaining what IG layer has to do with all the crypto stuff I think the one simple thing is when you want to run any offchain service and bring the results back to a blockchain like ethereum or near or any other chain you you need like the nodes to be incentivized and like actually be able to run it and if they do do the thing that they don't they claim to do then they should be rewarded if they don't do the thing they claim to do then they should be punished this is really what the aen layer system is designed to do so we have a you know several a projects working with us on you know inference training benchmarking uh you know which we we'll have announcements about a bunch of them coming up but that's that's where like naturally AI computation cannot be done onchain it has to be done offchain and so what are the best mechanisms for doing it what are the mechanisms for adjudicating that they've done it correctly you know is it you know cryptoeconomic onchain punishments is it cryptoeconomic for cability punishments is it unconditional correctness can we actually do things like that so that's what we we think a lot about at a so really excited for seeing some of these things how much for sure yeah I think we out of time but yeah I mean thanks uh thanks for coming I think it was very interesting discussion there's uh there's definitely a lot of topics to dig in I think accelerating coordination I think we should all change you know to a a cak cak let's go cak yeah cak yeah so uh I think that's that's kind of the the thing I took out but yeah I'm sure we'll see you in Bangkok and for everyone come to redact it we'll talk more about kind of what we're doing on near ey side how it all fits into near as well as uh I think you know some of the projects who are working with i l going to be there as well uh like na Labs presenting what they working on so uh thanks and uh yeah see you guys next time thank you so much","transcript_synced_at":"2026-11-03T05:28:46.497Z","segments":[{"start_seconds":3,"text":"near to one we're close to won't stop us","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=3s"},{"start_seconds":10,"text":"now together we'll redefine the game with near will find the flame","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=10s"},{"start_seconds":18,"text":"empowered Minds rise and store use your own AI we'll explore","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=18s"},{"start_seconds":32,"text":"hold the future hand in hands in this Tech driven","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=32s"},{"start_seconds":39,"text":"Wonderland okay hello hello episode 23 holy crap this is a lot of episodes but","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=39s"},{"start_seconds":47,"text":"uh uh it's really amazing to have shiram here uh we're gonna have an amazing","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=47s"},{"start_seconds":54,"text":"episode diving in into AI I'm sure we'll touch on Ian later and uh lots of fun stuff I'm I'm sure lots of people","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=54s"},{"start_seconds":63,"text":"already know shiram but do you want to introduce yourself quickly and absolutely thank you so much like really excited to be here uh start as an","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=63s"},{"start_seconds":72,"text":"introduction I'm shiram I started the Ian L project like three and a half years back uh and uh used to be a professor at the University of","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=72s"},{"start_seconds":80,"text":"Washington Seattle before doing uh crypto I used to work on computational","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=80s"},{"start_seconds":87,"text":"genomics and through which I got into to uh AI a little bit we even had a spin-off from our lab uh you know","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=87s"},{"start_seconds":95,"text":"startup working on applications of AI to 5G um but super excited to be here it's","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=95s"},{"start_seconds":102,"text":"one of the my most uh passionate areas so and I don't think there's anybody better than Ian and Alex to discuss this","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=102s"},{"start_seconds":111,"text":"with I was just telling them before the conversation that I am a fan of the white board series that near did like","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=111s"},{"start_seconds":117,"text":"four four years back learn everything about every protocol out there and how how they think about designing it so this is a modern version of that with","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=117s"},{"start_seconds":127,"text":"the with yeahi stuff so super excited sprinkle yeah not sprinkle AI everywhere and like","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=127s"},{"start_seconds":136,"text":"sprinkles yeah well I mean first of we haven't had actually a whiteboard about igen l so I think no that um that's a","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=136s"},{"start_seconds":146,"text":"mistake yeah well we're GNA be in Bangkok but I actually just F to mention because you do have like I think I don't","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=146s"},{"start_seconds":153,"text":"know if it's your one of your lat last p i mean last non-blockchain paper papers but you had the most cited paper is","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=153s"},{"start_seconds":162,"text":"improving Federated learning uh personalization so uh yeah you have","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=162s"},{"start_seconds":169,"text":"quite quite a few of uh kind of machine learning papers back in 1819 which somehow gets more cited than","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=169s"},{"start_seconds":178,"text":"blockchain papers I don't know what it's clear right like there's no everybody in one of the reasons I switched from a you know what happened","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=178s"},{"start_seconds":187,"text":"was I my PhD was in peer-to-peer Wireless Systems and and know peer-to-peer Wireless didn't go anywhere so after I finished my PhD uh even","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=187s"},{"start_seconds":195,"text":"though I we Sol some like theoretical problems in that thing I was looking for what's a high impact area and then got to genomics which I entered into and doing","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=195s"},{"start_seconds":204,"text":"like protoc you know algorithms for an analyzing genomics data mostly statistical learning and as as I was doing it you know the funny episode is","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=204s"},{"start_seconds":213,"text":"in 2017 I used to have like a very what do you call cursory negative opinion that like you know armchair academics used to","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=213s"},{"start_seconds":222,"text":"have about like neural networks and Ai and I'm like my student came in and we were doing this thing where you you have a DNA sequence going through a nanopore","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=222s"},{"start_seconds":231,"text":"and you sequence what the the DNA strand is and uh you know we were doing like all these statistical learning","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=231s"},{"start_seconds":239,"text":"algorithms we were working with actually like uh the the physicist who had the strongest uh empirical measurement of","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=239s"},{"start_seconds":247,"text":"the gravitational constant G so I was very confident about our physics models about our inference models everything and then my student comes in and says","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=247s"},{"start_seconds":254,"text":"hey I'm going to use like you know rnns to like do this thing like what kind of an idiot are you like you know maybe","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=254s"},{"start_seconds":262,"text":"good for things which don't have mathematical models you know here we have like the absolute best model so like you know you're going to go knock","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=262s"},{"start_seconds":269,"text":"yourself out waste your time two weeks later he beats our Benchmark okay this is that so I my","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=269s"},{"start_seconds":279,"text":"estimate of like it just like flipped within that thing and then I said hey can you invent like communication algorithms using the same thing and then","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=279s"},{"start_seconds":287,"text":"he's like oh yeah you can and then like we we did a bunch of stuff there and that that became the spin-off he became founding engineer there and then I","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=287s"},{"start_seconds":295,"text":"decided man this is going to like take over our job soon so I need to find an area which doesn't have this same issue","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=295s"},{"start_seconds":302,"text":"and my model for AI became basically anything that you can actually evaluate the optim the objective function easily","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=302s"},{"start_seconds":310,"text":"I call this a high-speed Oracle where you can evaluate how good your output is quickly you're going to optimize the heck out of it like it's not it's so","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=310s"},{"start_seconds":318,"text":"that was my like operating model and then like okay what is going to give me like a five to 10 year Horizon to do","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=318s"},{"start_seconds":325,"text":"something I'm like okay maybe blockchains and cryptography and stuff you can't like evaluate whether a cryptographic scheme is good just by","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=325s"},{"start_seconds":333,"text":"like empirically simulating it you to have some mathematical proofs which as we we can talk about they're getting really good at so you know we have like","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=333s"},{"start_seconds":342,"text":"a few more years here to do something fory our blockchains what","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=342s"},{"start_seconds":351,"text":"that that's interesting yeah kind of the opposite of ours which was like hey we need we need to coordinate a bunch of people to do the machine machine","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=351s"},{"start_seconds":359,"text":"learning so and like the tool to coordinate is you know like peer-to-peer payments and kind","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=359s"},{"start_seconds":367,"text":"of uh some some form of crowd crowd work that facilitated by blockchain yeah","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=367s"},{"start_seconds":374,"text":"amazing I I you know somewh unrelated but I just want to call out when I remember uh recently the a16z uh report came out and I saw near is you know","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=374s"},{"start_seconds":383,"text":"number two on like number of active addresses Kudos like you guys kicking ass and getting there on the way to","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=383s"},{"start_seconds":390,"text":"coordinating you know agents and people to to do this so super excited for that but you know I I mentioned the kind of","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=390s"},{"start_seconds":399,"text":"technical reason why I got into crypto but also I think that there was a similar kind of philosophical reason which is that you know I used to have","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=399s"},{"start_seconds":409,"text":"this mantle model that like the most important thing people do is you know we're intelligent and be innovate so Innovation being the highest value so","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=409s"},{"start_seconds":417,"text":"you can think of like AI as giving super super powers to our intelligence our Innovation abilities and I kind of flipped from","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=417s"},{"start_seconds":424,"text":"that model to a different model which is that yes you know you have Innovation and intelligence","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=424s"},{"start_seconds":432,"text":"but the the defining evolutionary advantage of humans is our ability to coordinate flexibly and so accelerating","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=432s"},{"start_seconds":440,"text":"coordination is a different value and maybe even more fundamental than accelerating Innovation or intelligence","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=440s"},{"start_seconds":448,"text":"and particularly the thing to think about coordination is you know I think of like the long Arc of evolution you","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=448s"},{"start_seconds":455,"text":"know Innovation is like you know this exponential growth one building on top of the other and then but suddenly coordination is this Quantum Leap so you","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=455s"},{"start_seconds":464,"text":"suddenly have a new coordination mechanism and then everything like flips and as an example like you know in in like a cell you have you know a procario","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=464s"},{"start_seconds":473,"text":"to ukar right this kind of a transition where like you know inside a cell you don't have a nucleus now you have a nucleus specializ ing in like","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=473s"},{"start_seconds":480,"text":"information storage and mitochondria specializing in energy production they have their own dnas but they coordinate because they they live and die together","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=480s"},{"start_seconds":489,"text":"this kind of a thing happens very rarely and then once it happens it just explodes the scope of innovation that can be built on top so maybe maybe a","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=489s"},{"start_seconds":496,"text":"close example is you know pre- internet and then we have internet and like the coordination on internet right just exploded so yeah so great great example","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=496s"},{"start_seconds":506,"text":"internet the the way we think about like I I started thinking about like coordination and the internet is internet so the coordination requires","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=506s"},{"start_seconds":515,"text":"two things communication and commitments communication the internet basically played the major role where like we can send information to everybody else for","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=515s"},{"start_seconds":523,"text":"example you could not train any of these AIS if you don't have the world's information on the internet at your disposal right so you know that was one","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=523s"},{"start_seconds":531,"text":"kind of like base level of coordination the second one is commitment like you know we can come together and coordinate only if not only if we can communicate","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=531s"},{"start_seconds":538,"text":"with each other but also so if we can enter into binding commitments hey we produced this episode we we spoil the shares equally whatever the division","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=538s"},{"start_seconds":547,"text":"rule is and then like that's a binding commitment once we know there's a binding commitment the game theory of cooperate versus defect completely","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=547s"},{"start_seconds":554,"text":"changes so the internet the communication super substrate crypto our commitment super substrate and then","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=554s"},{"start_seconds":562,"text":"together you can actually like now build arbit new interesting coordination structures on top","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=562s"},{"start_seconds":571,"text":"Alex you want to say something before no no are you waiting for me I","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=571s"},{"start_seconds":578,"text":"see yeah we need a better coordination mechanism here um yeah so that was basically like","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=578s"},{"start_seconds":587,"text":"the reason we kind of I got into crypto the philosophical kind of reason being coordination super structures the","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=587s"},{"start_seconds":596,"text":"technical reason being like you know this these things require cryptography and cryptography is not going to be like you know firstly broken by AI or","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=596s"},{"start_seconds":605,"text":"invented by AI at least for the short while so it gives us a little bit of an edge for for us to do something I think now like as I've gone deeper into this","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=605s"},{"start_seconds":614,"text":"Rabbit Hole I have like deeper and crazier theories and like we can talk about some of them we we Aller crazy","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=614s"},{"start_seconds":622,"text":"theories but actually so so staying on the topic of coordination because one of the things with Alex we always discuss is","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=622s"},{"start_seconds":630,"text":"um the kind of for for you know we we have this like user own internet right as like a a main thing that I think like","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=630s"},{"start_seconds":639,"text":"all of the different blockchain work is fitting into and as part of that you need a user own AI which is like","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=639s"},{"start_seconds":646,"text":"optimized for you as individual but and like everybody's kind of okay let's you know have decentralized training right let's you","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=646s"},{"start_seconds":654,"text":"know train models in decentralized way because that gives us kind of lots of guarantees but the challenge that uh is really","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=654s"},{"start_seconds":662,"text":"there is actually it's not the coordination of bunch of gpus to train the model that's hard it's actually getting to really","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=662s"},{"start_seconds":670,"text":"good model that's hard that's like the reality is there is you know there's open Ai and like still pretty big gap","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=670s"},{"start_seconds":678,"text":"you know meta Google uh and there's like hundred other companies that are training models right now and they're","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=678s"},{"start_seconds":685,"text":"not like actually matching right so like they don't even have the problem of coordinating gpus they just like buy a cluster and training it the getting good","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=685s"},{"start_seconds":693,"text":"model is the hard part and um so I think like the coordinating of research is","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=693s"},{"start_seconds":700,"text":"kind of the conceptually missing piece and you know right now the only examples that we've seen that work is you know","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=700s"},{"start_seconds":708,"text":"you staff everyone in this the same office and you know you inspire them that they building AGI and now they're","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=708s"},{"start_seconds":715,"text":"like you know crashing it but like that seems like you forgot the multi-million","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=715s"},{"start_seconds":722,"text":"dollar uh pay yeah and then and then you run a sorry for-profit company um yeah so so so the question is like","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=722s"},{"start_seconds":731,"text":"kind of what what is what what can we bring right from what we've learned I mean also like you've been in Academia coordinating a bunch of people doing","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=731s"},{"start_seconds":739,"text":"research right I mean we actually worked in those organization centralized organizations on our side so like what","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=739s"},{"start_seconds":746,"text":"can we bring kind of Best of Both Worlds leverage blockchain leverage all the tools we're building uh so that's kind of the topic I thought we can like dig","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=746s"},{"start_seconds":755,"text":"in a little bit absolutely la la The Prompt um so you know I think you know the","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=755s"},{"start_seconds":764,"text":"phrasing I I would like to rephrase we I think coordinating a bunch of gpus is not interesting okay so let me just","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=764s"},{"start_seconds":772,"text":"start with some slightly controversial opinion right I think it's it's similar to what you said but just phrasing it like that coordinating a bunch of GP is","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=772s"},{"start_seconds":779,"text":"not interesting why anybody can buy gpos okay you can buy gpos their their commodity and there's an economy of scale you can just put like you know","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=779s"},{"start_seconds":787,"text":"Elon Musk is buying a bunch of them like people just go buy them right it's really easy easy to coordinate because they're not sovereign agents","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=787s"},{"start_seconds":795,"text":"coordinating Sovereign agents is what is really interesting and you know the not yet talking about AIS but humans right","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=795s"},{"start_seconds":803,"text":"so when you think about it like building AI one model of coordinating agents is hey everybody come here you all have","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=803s"},{"start_seconds":811,"text":"equity in this like common Enterprise you know open AI whatever like you know number go up let's make the number go up together by building this common you","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=811s"},{"start_seconds":819,"text":"know objective everybody comes in they contribute and then like they can scale um firstly I think this is reliant on","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=819s"},{"start_seconds":827,"text":"this coordination structure the company as the core unit of scaling human coordination and if you think about it","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=827s"},{"start_seconds":834,"text":"actually like back in 2017 some people had observed it but I think it is kind of like there are many like gems of ideas that have kind of Fallen by the wayside of","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=834s"},{"start_seconds":843,"text":"like crypto and one of them is this theory of Kus law like Kus law is like why are companies sized at the size that","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=843s"},{"start_seconds":852,"text":"they are and you know Ronald Co and Economist right wrote this paper the theory of the firm and it's basically like oh you know company is this big","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=852s"},{"start_seconds":861,"text":"because you know it's easier to coordinate individuals internally than it is to coordinate individuals across organizations","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=861s"},{"start_seconds":869,"text":"and this lays the foundation for how open a Google Deep Mind everybody else is operating today is hey you know we're","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=869s"},{"start_seconds":876,"text":"coordinating by the super structure the equity Corp and we are going to actually okay so and you start thinking about it","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=876s"},{"start_seconds":884,"text":"and you know you've been there I my I've never worked in a company before so I have limited experience but I have friends you know I'm not living in under","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=884s"},{"start_seconds":892,"text":"a rock so saying I can layer Labs is not a company uh the not one I made up the ones that","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=892s"},{"start_seconds":901,"text":"existed before um the the way to","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=901s"},{"start_seconds":908,"text":"um think about the um the the problem is when your company grows big right Beyond 100 you know at Beyond thousand what","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=908s"},{"start_seconds":917,"text":"happens is there's an internal tragedy of the commons right because you have inside the company your contribution is marginal relative to the contributions","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=917s"},{"start_seconds":925,"text":"of everybody but you receive an equal payout irrespective of your contribution and this disincentivizes the most","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=925s"},{"start_seconds":932,"text":"contributing ones unless the company has a very strong internal system of like incentives which is very difficult to build because you know it's a","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=932s"},{"start_seconds":940,"text":"collaborative Enterprise that it's fundamental so what happens is like you know you have this marginal incentive problem every unit of new person","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=940s"},{"start_seconds":948,"text":"contributing to grow the stock or you know output doesn't feel their output impact their like you know the outcome","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=948s"},{"start_seconds":956,"text":"equally so so there's an internal tragedy of the this is why we see like companies do really well at small scales you know at a reasonable small valuation","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=956s"},{"start_seconds":965,"text":"10 to 200s at like a reasonable big valuation 100 to th and then you go beyond that it's like massive tragedy of the comments like people are doing rant","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=965s"},{"start_seconds":973,"text":"work they're doing like skating on minimum work all kinds of things happen they just fight with each other all these fundamental problems come together","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=973s"},{"start_seconds":980,"text":"and what we do know is the best mechanism for resolving units of value is the free market like this is the","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=980s"},{"start_seconds":988,"text":"biggest like lesson we've learned in coordinating humans over the last 300 years right free market puts a number a","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=988s"},{"start_seconds":996,"text":"value on each unit like oh how much is near valuable how much is I Like You can actually like do this because the free Market's kind of like expending energy","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=996s"},{"start_seconds":1004,"text":"in allocating value and betting on oh you know this unit is actually producing more let's allocate value more so the thesis should should not necessarily be","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1004s"},{"start_seconds":1013,"text":"to coordinate gpus it should be to coordinate humans in different architectures where you can maximize the","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1013s"},{"start_seconds":1021,"text":"net output rate of growth and rate of innovation and so what are things that can actually do this so zooming back to","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1021s"},{"start_seconds":1028,"text":"your pro problem can we build uh the universal AGI as a composition of","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1028s"},{"start_seconds":1035,"text":"smaller models like that's one like you know angle at it right you say oh you know there is some context free general intelligence and then there is","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1035s"},{"start_seconds":1043,"text":"contextual specific intelligen is the best intelligence for evaluating where a tomato from a best intelligence for you","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1043s"},{"start_seconds":1051,"text":"know say singing a song in Tamil which is my native language like you know each one could be something very specific we do know this like from General","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1051s"},{"start_seconds":1060,"text":"observation that specific intelligen the you take the person who has the most general intelligence like IQ or something and they never played chess","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1060s"},{"start_seconds":1068,"text":"and asked them to play chess they're going to lose to somebody who has lower general intelligence but higher contextual intelligence right so they've","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1068s"},{"start_seconds":1075,"text":"just play many many years they remember all kinds of like structures and movements so this thing about like are","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1075s"},{"start_seconds":1083,"text":"there architectures of universal general intelligence which are compositions of like some some smaller units of","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1083s"},{"start_seconds":1092,"text":"intelligence and if there are let them build in the free market and coordinate value across a blockchain is way better","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1092s"},{"start_seconds":1101,"text":"then we have to get these 10 million people sitting together in open AI suffering a tragedy of common and so","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1101s"},{"start_seconds":1109,"text":"this is like our thesis of crypto and why I think why I think it is you know uh I think crypto is going to impact","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1109s"},{"start_seconds":1115,"text":"every endeav in every human and free agent Endeavor and specifically at the Endeavor in AI is that you create units","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1115s"},{"start_seconds":1125,"text":"of work and each unit of work is measurable and attributable by the free market for example the general intelligence Tunes to like different","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1125s"},{"start_seconds":1132,"text":"composable models and allocates value based on like how much each model is used now if you build a model and it's kind of like sitting on the super","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1132s"},{"start_seconds":1139,"text":"structure and you know your model is used more your token your like you know unit of value acral increases so this is","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1139s"},{"start_seconds":1148,"text":"I think the kind of my thesis for where like you know how AGI will be built on","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1148s"},{"start_seconds":1155,"text":"the free market not on as compositions of individual agents rather than like","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1155s"},{"start_seconds":1161,"text":"one you know giant agent I think it's also better for us it's minimizes risks there's lots of of other like second","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1161s"},{"start_seconds":1169,"text":"order consequences but the first order thing is I think this is actually maybe the best way to coordinate like you know 10 million people to like come in and","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1169s"},{"start_seconds":1177,"text":"contribute to this common","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1177s"},{"start_seconds":1186,"text":"end yeah I uh I've been thinking uh for a while along the same lines how do","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1186s"},{"start_seconds":1193,"text":"we um so effectively if you think if you think of uh um like a large lab working on AI they","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1193s"},{"start_seconds":1202,"text":"have let's say they have 300 tier one researchers working full time but but the number of you know important","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1202s"},{"start_seconds":1210,"text":"breakthroughs happening is relatively small like like let's say let's say they have like 10 10 10 impactful breakthroughs per year that move the","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1210s"},{"start_seconds":1218,"text":"needle right so that means that first of all that means that bringing in you know something on the order of 10 people who are good to work on an effort is","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1218s"},{"start_seconds":1227,"text":"relatively meaningless right because the the the you know the expected time before a breakthrough is relatively high right so you want to but also the","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1227s"},{"start_seconds":1233,"text":"problem is um those labs are capable of bringing the absolutely best talent right but in the wild we will have it it","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1233s"},{"start_seconds":1242,"text":"will be in order for us to capture people of that quality uh it will need to be an even bigger sample of people","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1242s"},{"start_seconds":1250,"text":"right so we need to find a way to orchestrate thousands of people potentially working on the same goal which is a um hard especially given you","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1250s"},{"start_seconds":1259,"text":"know you know the the there is a little bit truth to that there's also a little bit of the opposite to that which we're seeing already in open AI right like","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1259s"},{"start_seconds":1267,"text":"many of the the the founding team is going out doing their own thing it's a kind of like a sense of a signal of the","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1267s"},{"start_seconds":1274,"text":"tragedy of Commons right hey we come together towards this common Enterprise and then like hey actually I'm the one who's doing all the like why should","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1274s"},{"start_seconds":1281,"text":"I like you know and I think this is the the the fact that each one of those like","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1281s"},{"start_seconds":1289,"text":"200 whatever people that you're talking about they believe their thing is going to be the thing that moves the ne the most and I as long as there is a kind of","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1289s"},{"start_seconds":1298,"text":"coordination super structure that can help like compose the values created by each one of them in some like meaningful","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1298s"},{"start_seconds":1305,"text":"way it might so anyway just like a observation there on the inter like an empirical observation on the tragedy of","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1305s"},{"start_seconds":1312,"text":"Commons which is that they can bring these people together but to keep them together and you know because of the kind of same type of problem that I I","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1312s"},{"start_seconds":1321,"text":"think it's already happening empirically yeah it is definitely it is definitely the case that lot of people","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1321s"},{"start_seconds":1329,"text":"who uh like the employees uh how do you distinguish between those like you know people who","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1329s"},{"start_seconds":1336,"text":"actually dig the the trench a lot of them start companies right perplexity is a good example right our event was open","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1336s"},{"start_seconds":1346,"text":"back in 2022 I think but uh yeah whether whether people on the high level live for that reason that's a that's a","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1346s"},{"start_seconds":1353,"text":"different question I think that but maybe like Alex if you want to contrast","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1353s"},{"start_seconds":1360,"text":"because I mean we had that conversation like hey is this going to be a lot of kind of specialized AI let's just call","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1360s"},{"start_seconds":1368,"text":"them agents models you know Services systems that you know you kind of","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1368s"},{"start_seconds":1375,"text":"interacted coordinate with or this is going to a big model that you need a lot of human intellect coordinated together","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1375s"},{"start_seconds":1384,"text":"to advance kind of what it does right because I think that's like a a pretty polar uh view on on where this is going","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1384s"},{"start_seconds":1394,"text":"and how to coordinate these efforts at the same time it maybe that both are true right they don't mean like that you","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1394s"},{"start_seconds":1401,"text":"know either or but I think it would be interesting Alex if you want to highlight like what's the other like hey","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1401s"},{"start_seconds":1408,"text":"a big model coordinating lots of researchers to kind of push the envelope kind of look","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1408s"},{"start_seconds":1415,"text":"like interesting so so I definitely do believe in one model future I uh I I","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1415s"},{"start_seconds":1422,"text":"don't know how how important my opinion is I don't think we have enough information to predict right it's just a good feeling that it will be it will be","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1422s"},{"start_seconds":1430,"text":"one large model that is connected to all the apis and whatnot and uh just like if I you know like sort of GPT 5 scenario but","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1430s"},{"start_seconds":1438,"text":"hopefully it's not going to be gp5 hopefully it's going to be you know some sort of like like either intellect one or whatever I I don't remember the name","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1438s"},{"start_seconds":1446,"text":"of the prim intellect one or something like that right something built by the community","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1446s"},{"start_seconds":1452,"text":"um but I don't uh the the part about that model orchestrating researchers is a um I don't remember the reference but","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1452s"},{"start_seconds":1460,"text":"I do think that it's a the interesting problem there is that you don't need to just lot of researchers you know the each each unit","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1460s"},{"start_seconds":1469,"text":"of research requires resources right you cannot you cannot just like a is alchemy in the blockchain you know I can be","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1469s"},{"start_seconds":1476,"text":"sitting in a like you know case in point the consensus that is driving near today I came up with it sitting on a beach in","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1476s"},{"start_seconds":1484,"text":"Puerto Rico right I didn't need any resources to to come up with it right but but a is alchemy right so your","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1484s"},{"start_seconds":1492,"text":"hypothesis is not worth anything until you actually tried it and and something you know on low scale doesn't necessarily you know transfer to the","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1492s"},{"start_seconds":1499,"text":"high scale so it's a it's interesting how do you orchestrate if you have thousands of people you probably don't have resources to actually power them","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1499s"},{"start_seconds":1506,"text":"you know continuously pushing so it's it's an interesting problem not not quite","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1506s"},{"start_seconds":1513,"text":"a it is a great uh great prompt uh there are I I think the the scaling problem","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1513s"},{"start_seconds":1521,"text":"right which is that basically unless you have a lot of resources you can't even experiment because these are emergent at","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1521s"},{"start_seconds":1528,"text":"only large scales I think that's a really uh you know pressing problem um","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1528s"},{"start_seconds":1538,"text":"the so going to like aia's question of whether there's going to be like a single model to rule them all or like you know is there uh compos composition","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1538s"},{"start_seconds":1547,"text":"of models that and and contextually switching across them is going to be better I think it's going to be like the","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1547s"},{"start_seconds":1554,"text":"basic line which which shows whether like the approaches in crypto can succeed versus because you know if if","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1554s"},{"start_seconds":1562,"text":"it's one model it's much better like you know you have thousand people sitting in an office and trying to kind of hash it out right like there's no no world in","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1562s"},{"start_seconds":1570,"text":"which like a a committee based approach for like what that model is is going to work as well as a kind of you know much","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1570s"},{"start_seconds":1578,"text":"better organization that so I do think that that will basically draw the boundary it might also be that you can actually build composition of models","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1578s"},{"start_seconds":1587,"text":"which are pretty close to the to the or original one like it's maybe the the big Central one is the first one that is","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1587s"},{"start_seconds":1595,"text":"produced and in six to nine months or one year whatever like there is an alternative which is built on this composition Theory because at the end of","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1595s"},{"start_seconds":1603,"text":"the day you know that is a kind of like a internal composition right a big network is made of many small networks","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1603s"},{"start_seconds":1609,"text":"connected in some more complex ways but this um the yeah so so the more uh we","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1609s"},{"start_seconds":1620,"text":"the other thing that I think about there is how do we make sure that so there is this open closed asymmetry the open","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1620s"},{"start_seconds":1627,"text":"closed asymmetry is closed can copy from open open cannot copy from closed okay like this is a fundamental kind of","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1627s"},{"start_seconds":1633,"text":"problem it's a it's old problem in like you know uh open source software and you know copy left was kind of born as a","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1633s"},{"start_seconds":1641,"text":"kind of result like you do a GPL license and then GPL you know close Source can't copy from GPL but what we need is even","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1641s"},{"start_seconds":1649,"text":"GPL right like didn't give incentives right so I contribute code to GPL I don't have incentives other people can create GPL they don't have incentives so","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1649s"},{"start_seconds":1657,"text":"that whole thing kind of like dies out instead it's much better to build on MIT or some other like non- copy left license and then like have versions","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1657s"},{"start_seconds":1665,"text":"which are like you know closed and like this is the open core or whatever model that you know software has kind of tended towards I think reexamining and","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1665s"},{"start_seconds":1673,"text":"understanding if there are improvements to be made in that level can also play a role here because you know if the open closed asymmetry is not broken you have","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1673s"},{"start_seconds":1681,"text":"the problem that you know always the one Central model will incorporate every idea that's there in the the open","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1681s"},{"start_seconds":1688,"text":"decentralized world and so the one one thing we've been playing a lot around is like a new licensing model we call PDL","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1688s"},{"start_seconds":1696,"text":"permissionless derivators license the idea of permissionless derivators license is you create a software and anybody can create a derivative of that","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1696s"},{"start_seconds":1705,"text":"software but both the software any derivative created has to be launched on the same platform so the PDL is Tethered","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1705s"},{"start_seconds":1713,"text":"to a platform which means like I take this software and then I that this this platform helps let's say do influence whatever of that software and if","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1713s"},{"start_seconds":1723,"text":"I any fork or like derivatives have to be launched back on that platform what this solves is the value uh measurement","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1723s"},{"start_seconds":1732,"text":"problem right the value measurement problem is you have a model I created a derivative now I launch it on the same platform now which means this platform is measuring how much value how much","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1732s"},{"start_seconds":1740,"text":"utility each of these derivative models have and then can pass like relative value across ancestors of models right","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1740s"},{"start_seconds":1748,"text":"so I think a lot about this question which is you know permission Innovation can actually grow things exponentially I can build on top of you it's one of the","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1748s"},{"start_seconds":1756,"text":"reasons where like you know in newbs you know one of the big you know AI conferences which is where I heard about","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1756s"},{"start_seconds":1763,"text":"Helia paper first I was at newbs 2017 in Los Angeles and the crowded crowded","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1763s"},{"start_seconds":1770,"text":"poster stand uh I I don't know if IIA was there or not IIA were you there I was there yeah I mean I was we were","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1770s"},{"start_seconds":1777,"text":"Manning it I think in two shifts yeah yeah okay I I badly knew anything about any of this stuff but I did see a very","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1777s"},{"start_seconds":1784,"text":"crowded stand in Los Angeles it was also like a Tipping Point for me to know like okay this is going to go crazy and you know I was November 2017 I think","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1784s"},{"start_seconds":1793,"text":"November December 2017 and then I shifted to crypto but uh the uh the the reason I mentioned that","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1793s"},{"start_seconds":1802,"text":"is when we're thinking about these model derivatives if you can allocate value to like these derivatives and each small","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1802s"},{"start_seconds":1809,"text":"unit of improvement can ACR value because they're being built one on top of the other it's actually like way","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1809s"},{"start_seconds":1816,"text":"better for for like people to contribute back to the commons the reason I mentioned nebs is I used to observe this trend which is you know I come from a","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1816s"},{"start_seconds":1825,"text":"kind of very mathematical oriented subset of Statistics information Theory and where like people take three years","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1825s"},{"start_seconds":1832,"text":"to publish one one paper and what new set is a new pattern where like every three to six months people just like","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1832s"},{"start_seconds":1840,"text":"batch whatever ideas they have and then publish it and initially I thought oh these are like you know this is not a good idea but then now I see that as","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1840s"},{"start_seconds":1847,"text":"actually the accelerant for the rate of innovation because you you you publish like small snapshots other people can build on top of your Snapchat rather than only you can build on top of your","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1847s"},{"start_seconds":1856,"text":"own things so you have way more p but each paper is uh you know the net rate of growth that you have as an area actually grows much faster it's the same","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1856s"},{"start_seconds":1864,"text":"thing with Innovation right do I need like everybody to come together and build this one big model there are only like 10 companies in the world that could try to do that whereas if each of","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1864s"},{"start_seconds":1872,"text":"us can contribute a diff on top of like an existing model oh this is a fineing this is this thing this is that thing and then like each one like passes value across through the","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1872s"},{"start_seconds":1881,"text":"ancestors then I think you get like a much more interesting thing where as a creator of like new ideas new models you you're also incentivized heavily and","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1881s"},{"start_seconds":1890,"text":"maybe each of these like new models has their own token because you know all value allocation kind of can be passed back to the kind of like the the thing","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1890s"},{"start_seconds":1898,"text":"all of them happening on chain programmably like so there is I think uh really interesting things to be done in","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1898s"},{"start_seconds":1905,"text":"this intersection at least as an experiment but the thing that I feel like very deeply is we have a very","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1905s"},{"start_seconds":1912,"text":"narrow time window to do this so I'm really glad you guys are taking like a leadership position here and pushing this and I think we have a very narrow","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1912s"},{"start_seconds":1921,"text":"time window to actually like do do this I think 2027 is going to be crazy so just backing out from there we have like","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1921s"},{"start_seconds":1928,"text":"two three years and it's if if we can show that crypto is actually can be very powerful in these this kind of a setting","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1928s"},{"start_seconds":1937,"text":"there absolutely no other better use case for all of crypto yeah I mean you kind of spoil","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1937s"},{"start_seconds":1948,"text":"some of our talk talks at redacted so uh I don't know how much CH want to spill","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1948s"},{"start_seconds":1955,"text":"here but yeah I think the kind of the way we've been thinking about it is is very much aligned with this kind of from","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1955s"},{"start_seconds":1964,"text":"from multiple directions one is we don't know which of these two","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1964s"},{"start_seconds":1970,"text":"worlds will actually win and so as as kind of uh you know being now in","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1970s"},{"start_seconds":1977,"text":"financial industry we hedge and we do both and so we need a way to coordinate AI agents and even Beyond AI agents how","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1977s"},{"start_seconds":1985,"text":"to coordinate any Market participant um effectively and create effectively a protocol to for them to","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1985s"},{"start_seconds":1993,"text":"exchange information pay Char have the P have the commitments um to do that and then on the other side is well","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=1993s"},{"start_seconds":2002,"text":"it's still this is still about research and Innovation and how can we accelerate the speed of innovation in this open way","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2002s"},{"start_seconds":2010,"text":"right and so as you said nips kind of established a new setup I mean it's actually I thinkr with uh open review.","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2010s"},{"start_seconds":2020,"text":"net right actually y I think Yan leun introduced this idea of like even more open even faster like put your preprint","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2020s"},{"start_seconds":2028,"text":"publicly let's just go you know let's just on it I mean uh comment on each other papers in public and created","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2028s"},{"start_seconds":2036,"text":"a Twitter like environment for academic papers yeah but but it kind of it's all about like accelerating that feedback loop but the thing is like what we see","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2036s"},{"start_seconds":2044,"text":"still is like if you work at one of the big Labs you actually see the source code of that paper right even before it's published and you can like build on","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2044s"},{"start_seconds":2053,"text":"top of it but in open source right now it's all completely like everywhere and and the the you know the Frameworks are different the data is different you","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2053s"},{"start_seconds":2061,"text":"cannot actually like always compare them Apple to Apple even when you do you like you know the data prep was different the prompts was different something always","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2061s"},{"start_seconds":2070,"text":"is different right and so how do we actually establish a like a kind of you know coherent enough framework that you know is extensible enough for people to","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2070s"},{"start_seconds":2078,"text":"build on uh how do we promote this more faster like hey I want to experiment out","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2078s"},{"start_seconds":2085,"text":"here's like here's a finetune or here's one thing I did the funny part is we actually I started thinking about the license as well I think it was like last","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2085s"},{"start_seconds":2094,"text":"year because that does sound like something very much like Hey how do we actually like yeah BR that I think I","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2094s"},{"start_seconds":2102,"text":"kind of got stuck because I talked with a bunch of people and for license to really truly work you kind of need uh you kind of need to buy in from","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2102s"},{"start_seconds":2111,"text":"commercial organizations and Commercial organizations hate like new licenses and like their lawyers are really like you know their brain melts when you like","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2111s"},{"start_seconds":2120,"text":"introduce anything uh and uh and like for non-commercial right you you I mean you","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2120s"},{"start_seconds":2127,"text":"end like you kind of don't want to like over like if it's a research you don't want to actually like charge them effectively so you want them to use it","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2127s"},{"start_seconds":2134,"text":"so you kind of like the same reason why copy left I think doesn't function very well is because the people who actually want to open source stuff already don't","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2134s"},{"start_seconds":2144,"text":"care about the financial like directly and the people who care about financial they use open source more to get the","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2144s"},{"start_seconds":2151,"text":"attention and then they end up building stuff close source so the question being what can we do actually to have","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2151s"},{"start_seconds":2158,"text":"maybe not open source but provable compute or provable you know uh","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2158s"},{"start_seconds":2165,"text":"processes provable things and then incentivize that and kind of create an alignment where may I mean some things I","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2165s"},{"start_seconds":2172,"text":"end up sharing between people reuse it Etc so you can track it but at the same time the outcomes are actually uh","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2172s"},{"start_seconds":2180,"text":"available to everyone permissionless but not open so like open source right and","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2180s"},{"start_seconds":2187,"text":"that's an interesting you know on its own kind of challenge uh and then on the other side yeah like how do you coordinate research to be faster to like build on top of","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2187s"},{"start_seconds":2196,"text":"each other um so yeah I mean I think there is a there's also another angle I faced when","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2196s"},{"start_seconds":2205,"text":"talking about licenses is like the the the crypto Cipher punk people don't like","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2205s"},{"start_seconds":2212,"text":"oh you know IP is and this and that but I think it's important to um to","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2212s"},{"start_seconds":2219,"text":"create like in so my like one of my highest kind of values is uh is you can","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2219s"},{"start_seconds":2228,"text":"you know in a crypto language you'd call incentives in like my traditional language I'll call it kerma which is like if you contribute more to the commons the commons should give back","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2228s"},{"start_seconds":2236,"text":"more to you and if you contribute like negative things to the commons the commons should hit you back and this is a very important thing to regularize these common you know systems of Commons","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2236s"},{"start_seconds":2246,"text":"and when you're creating knowledge for for example and Publishing it in the comons and right now if you don't have any incentives to do it you know a","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2246s"},{"start_seconds":2253,"text":"person sitting and thinking hey you know I I go join open AI I get $2 million a year I just sit and contribute here in open source like you know it's like I","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2253s"},{"start_seconds":2261,"text":"have to do this like on the site right like it's not not exciting and so yeah totally so there's a lot of angles in","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2261s"},{"start_seconds":2270,"text":"getting this right and you know is is it is it like do we need more powerful onchain mechanisms like private private","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2270s"},{"start_seconds":2277,"text":"interest and other things can actually like help a lot um one uh question I had for you guys is uh you know where are","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2277s"},{"start_seconds":2285,"text":"you guys on safety right what's AI safety Doris like how how far on the curve are you","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2285s"},{"start_seconds":2295,"text":"guys I uh I have my own opinions about safety uh it's it's interesting I was uh I was talking to I was at a dinner and","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2295s"},{"start_seconds":2303,"text":"there was a bunch of guys from different safety Labs one one was naturally and the one was a I wouldn't remember the name but it's something like you know","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2303s"},{"start_seconds":2312,"text":"nation National AI safety Institute not like that but you know along the lines same embedding uh and both of them were","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2312s"},{"start_seconds":2319,"text":"effectively thinking towards how do we make the model we train safe right so in case of entropic they say Hey how do we","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2319s"},{"start_seconds":2326,"text":"make CLA not accidentally not say sorry you know at the end of the sentence uh or you know the the National Institute","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2326s"},{"start_seconds":2335,"text":"didn't have their own models but they were like working with the labs you know how do we make sure that their model is safe which I don't think is the concern","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2335s"},{"start_seconds":2341,"text":"at all I think the concern is how do we make sure that a threat actor who has access to 100,000 gpus doesn't train a","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2341s"},{"start_seconds":2348,"text":"model that will annihilate us all and we cannot do anything about it so you know from this","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2348s"},{"start_seconds":2355,"text":"perspective uh the way uh I think myself personal about safety is how how can we how can we address a particular problem","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2355s"},{"start_seconds":2364,"text":"that will arise right so I will give you an example of of this thinking right one of the big so obviously misinformation","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2364s"},{"start_seconds":2372,"text":"is a very big place where models will make it even worse than it is today and it's already quite bad uh but","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2372s"},{"start_seconds":2379,"text":"another another place where AI in a short term can do a lot of damage is","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2379s"},{"start_seconds":2386,"text":"uh is effectively breaking the internet in the sense that all the code that the internet is built upon uh it's like you","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2386s"},{"start_seconds":2394,"text":"know millions of lines of code which is uh written over many years by many people it naturally has a lot of bugs right and the security model of this","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2394s"},{"start_seconds":2403,"text":"code is that a very smart person by looking at this code will not find it right because a lot of very smart people looked at it and they didn't find it and and like every now and then you know","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2403s"},{"start_seconds":2411,"text":"like hard bleed happens where you know someone does find something right but the thing is that the model doesn't need to be a GI complete uh in order to start","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2411s"},{"start_seconds":2420,"text":"finding bugs better than us right so you know like it would like models already clearly showing faster pace of","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2420s"},{"start_seconds":2428,"text":"advancements in code in other areas of life right so for one they are better than us of painting bugs the only question is who will Who will invoke it","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2428s"},{"start_seconds":2436,"text":"first a good person who will immediately patch all of the issues right or a bad guy who will just uh right and and the bad guy could be different here it could","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2436s"},{"start_seconds":2444,"text":"be a bad guy who who seeks profit right like like presumably you know it's it's my assumption I have no idea but some of the two L agencies probably have zero","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2444s"},{"start_seconds":2452,"text":"zero days which allow them to SSH to my machine right but they they're not going to do that or maybe they will but they're not going to reveal the","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2452s"},{"start_seconds":2460,"text":"existence of those right uh but but you know if some script Kitty or or like someone who is just you know H yeah not","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2460s"},{"start_seconds":2468,"text":"a reasonable actor gets access to that right that then we can get be in very big trouble where where you know like you can just bring down the whole internet by finding a zero D in the","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2468s"},{"start_seconds":2476,"text":"networking stack and would like like would like recovering from this state when you cannot communicate with people anymore so that's one of the fears they","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2476s"},{"start_seconds":2483,"text":"have and we can work directly towards addressing it right like uh the simplest way would be investing right now heavily into formal verification and you know","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2483s"},{"start_seconds":2491,"text":"start rewriting or very you know formally verifying bits and pieces the most fundamental ones so that by the time the models are good enough to","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2491s"},{"start_seconds":2500,"text":"penetrate our defenses we have formally verified stack you know that powers the core of the internet and then you know","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2500s"},{"start_seconds":2508,"text":"we know that it's very unlikely to to be broken so so that's something that I'm uh I'm personally spending a lot of","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2508s"},{"start_seconds":2516,"text":"effort know going that way super interesting I I I think of like this fundamental problem in in the following way think about like any kind of complex","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2516s"},{"start_seconds":2525,"text":"system it's like there is a layer of free agents and Innovation built on top of like some kind of a coordination","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2525s"},{"start_seconds":2532,"text":"substrate usually it's the coordinations off things like the government and the uh the free agent economies like the","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2532s"},{"start_seconds":2540,"text":"free market and but I think uh so what's happening is technology like particularly AI will accelerate","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2540s"},{"start_seconds":2548,"text":"massively the free agents right so that's that's what we're seeing and the free agents are going to find every kink","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2548s"},{"start_seconds":2555,"text":"in the coordination structure and then try to break it right like that's that's the issue that that we are worried about and I think the the solution that uh","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2555s"},{"start_seconds":2564,"text":"some of these safety people are gravitating towards is let's go try to get the government to regulate it better or whatever like you know that that","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2564s"},{"start_seconds":2573,"text":"thing and I think instead I think what we should do as a kind of crypto space as trying to Pioneer coordination","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2573s"},{"start_seconds":2580,"text":"mechanisms is what crypto coordination mechanisms can we kind of bring to bear on this uh on this problem and right now","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2580s"},{"start_seconds":2588,"text":"actually my D you know diagnosis is it's actually making it worse okay like imagine if Alex's","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2588s"},{"start_seconds":2597,"text":"attack goes live right so an agent knows how to attack the internet it's going to first like try to","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2597s"},{"start_seconds":2603,"text":"steal a lot of people's e and BTC","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2603s"},{"start_seconds":2621,"text":"I think uh we lost yeah yeah we lost you for a couple","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2621s"},{"start_seconds":2631,"text":"minutes when you started talking about the plan of the agents they they shut you down so that you don't reveal it accidentally","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2631s"},{"start_seconds":2643,"text":"yeah I think that continues yeah much Silicon Valley internet you","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2643s"},{"start_seconds":2653,"text":"know Seattle internet and uh like everywhere else yeah that's fascinating how you","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2653s"},{"start_seconds":2660,"text":"know two of the two of the capitals of the of the tech in United States have the worst internet","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2660s"},{"start_seconds":2670,"text":"okay I guess we lost we're 30 seconds away from from from ending the episode anyway","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2670s"},{"start_seconds":2678,"text":"right what where is","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2678s"},{"start_seconds":2687,"text":"it okay you have closing thoughts well I think okay we're back we're back to","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2687s"},{"start_seconds":2696,"text":"worried the agents know we're talking about them and EXA I was just going to say that you","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2696s"},{"start_seconds":2703,"text":"know right now crypto is actually potentially making it worse not making it better because you know if if you do the Alexis attack and it it's going to","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2703s"},{"start_seconds":2711,"text":"steal our e and BTC and then these are sensorship list and immutable systems and you know it's it's chaos so I think we need to think through like for","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2711s"},{"start_seconds":2720,"text":"example something that we think about a lot at Aon is Fork cability right what set of fork ability conditions should exist for these systems and for you know","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2720s"},{"start_seconds":2729,"text":"one of the ideas we've been toying with is we should write asimov's LW something like that into the FOC ability conditions of the chain so this is I","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2729s"},{"start_seconds":2737,"text":"think another like way of thinking about it is like how do we set incentive structures where like these agents want to play our game rather than co-opting","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2737s"},{"start_seconds":2745,"text":"us into their game right so we set the incentive landscape and they play according to that is a kind of Meta Meta","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2745s"},{"start_seconds":2751,"text":"idea I'm fascinated with uh I think we we did a lot of high level stuff I'll just spend at least 10 seconds explaining what IG layer has to do with","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2751s"},{"start_seconds":2760,"text":"all the crypto stuff I think the one simple thing is when you want to run any offchain service and bring the results","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2760s"},{"start_seconds":2768,"text":"back to a blockchain like ethereum or near or any other chain you you need like the nodes to be incentivized and","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2768s"},{"start_seconds":2775,"text":"like actually be able to run it and if they do do the thing that they don't they claim to do then they should be rewarded if they don't do the thing they claim to do then they should be punished","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2775s"},{"start_seconds":2784,"text":"this is really what the aen layer system is designed to do so we have a you know several a projects working with us on","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2784s"},{"start_seconds":2792,"text":"you know inference training benchmarking uh you know which we we'll have announcements about a bunch of them coming up but that's that's where like","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2792s"},{"start_seconds":2800,"text":"naturally AI computation cannot be done onchain it has to be done offchain and so what are the best mechanisms for doing it what are the mechanisms for adjudicating that they've done it","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2800s"},{"start_seconds":2809,"text":"correctly you know is it you know cryptoeconomic onchain punishments is it cryptoeconomic for cability punishments is it unconditional correctness can we","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2809s"},{"start_seconds":2817,"text":"actually do things like that so that's what we we think a lot about at a so really excited for seeing some of these things how","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2817s"},{"start_seconds":2825,"text":"much for sure yeah I think we out of time but yeah I mean thanks uh thanks for coming I think it was very","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2825s"},{"start_seconds":2832,"text":"interesting discussion there's uh there's definitely a lot of topics to dig in I think accelerating coordination I think we should all change you know to","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2832s"},{"start_seconds":2841,"text":"a a cak cak let's go cak yeah cak yeah so uh I think that's that's kind of the the","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2841s"},{"start_seconds":2851,"text":"thing I took out but yeah I'm sure we'll see you in Bangkok and for everyone come to redact it we'll talk more about kind","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2851s"},{"start_seconds":2859,"text":"of what we're doing on near ey side how it all fits into near as well as uh I think you know some of the projects who are working with i l going to be there","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2859s"},{"start_seconds":2867,"text":"as well uh like na Labs presenting what they working on so uh thanks and uh yeah","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2867s"},{"start_seconds":2874,"text":"see you guys next time thank you so much","url":"https://www.youtube.com/watch?v=1jLH5GLp5fc&t=2874s"}],"youtube_url":"https://youtube.com/watch?v=1jLH5GLp5fc","url":"https://near-pedia.xyz/videos/1jLH5GLp5fc"}