Vitalik Buterin I The Promise & Challenge of Crypto + AI Applications I Agentic Ethereum 2025 Summit

[Music] and for this next talk I want to bring on italic to uh do a one-year later review of a blog post he originally wrote talking about promise and the challenges of crypto and AI so January 30th to February 7th 2025 a lot has changed and uh welcome thank you uh so great so a year ago right I wrote this blog post that was on uh the intersection between crypto and Ai and at the time of course that area was still very small and uh I think you know there have been a lot of people that had like tried to force the two things to come together and a lot of the time there's a challenge in like actually making the combination work right but of course since then AI has become powerful enough that AI has basically been well on its way to eating everything and so you know we've been seeing crypto and AI intersect quite a lot in the last year um and so I thought it would be valuable to look at some of the uh predictions that I uh made about cryp applications in a year ago and then basically see well where are we today and can we say some something more definite about some of these categories right so in the block post last year I had described four categories of crypto AI interceptions um and I categorized them as AI as a player in the game AI as the interface to the game AI as the rules of the game and AI as the objective of the game right and basically the first two are like you know AI participating or the first three are AI participating in uh blockchain uh constructions and the last the last one is of course blockchain constructions helping to build or improve or maintain AIS um and sometimes you actually need to do both at the same time right and uh the thing that I said last year is basically that the first one AI as a player in the game is like actually the category that is the closest uh to uh being able to like really seriously launch and get adopted and I even made the arguments that actually a we've had an AI crypto intersection use case for almost 10 years which is arbitrage Bots trading on dexes and uh like if you think of that kind of category as a starting point then you can ask the question well what other kinds of games um onchain games potentially offchain games can AI participate in and then the other categories are a little more uh like still blue sky and still dependent on further technology but of course AI is improving quickly and we've also seen cryptography improving quickly as well right so let's talk about AI as a player in the game right so the the OG AI crypto intersection use case is um um Arbitrage and Market making on decentralized exchanges right and since then of course we've seen AI evolve from being like relatively simple algorithms to being primarily llm driven and so in 2024 we've seen a lot of Twitter Bots also of course foraster Bots and other kinds of bots as well and uh we also saw last year a couple of experiments in some interesting areas right so AIS uh participating in prediction markets so Martin kman and like Omen together with aot noas and that whole cluster basically tried to do AI participating in prediction markets and that was pretty interesting and so the question is well what are the kinds of games that we can make this year that AI could participate in as well right so I'll say something more definite this year than I said last year um and the definite thing that I want to say is basically that I think one very particular unique opportunity is AI Dows right so Dows I know they've been part of the dream uh since basically before the beginning for ethereum uh right I think I literally wrote The Three posts on decentralized autonomous organizations and Bitcoin magazine like literally right before I wrote the original ethereum white paper but then of course we've seen Dallas have a lot of problems and I think a lot of the problems come down to the fact that uh there's a lot of things that need to be decided and and basically you end up very easily having quality decisions right so the question is well can we have higher quality decisions and uh actually on top of the low quality decisions problem one of the other problems that's starting to pop up recently is this bribery problem right like there's a like an actual thing I think it's like literally called bribe da or something like very similar and it's like becoming one of the largest delegates in arbitrum D like basically people are like literally doing like bribery and that turning turning coin voting into auctions right so can we make AI mechanisms that are both more high quality in decision- making in peace time and more resistance to capture in war time and so the core generator of ideas that I'm using here in this Paradigm is this concept of uh AI as the engine humans as the steering wheel right so humans are there to provide some kind of like alignment provide some kind of value signal and AI is basically just there to optimize and the optimization is Guided by the signal that the humans provide so this can be done in a few different ways right and I think a lot of the time like people just tend to instinctively gravitate toward this approach where you just have a single AI model and you have humans somehow feed their preferences into the single AI model and like you can do that right it's like an intuitive and easy construction to understand but it's also a construction that I think has a lot of pitfalls and we'll talk a bit about exploitability um a bit later and and the other Pitfall is that realistically AIS are going to improve and like the optimal stru uh design of an AI is going to be completely overturned pretty much every three months from here all the way up to the singularity and so like you want some kind of function that accommodates new AIS being able to come in and like basically being able to upgrade the engine without that upgrade process itself being a strain on the decision function right and so I've suggested these two other ways of achieving the same goal so one of them is of course old idea from Robin Hansen and many others over the last decades a f tarkey basically humans vote on an objective function and the prediction Market which I think realistically can be played by AIS can identify which decisions could optimize the objective function right so the standard example that gets used to illustrate this is like the fire the CEO Market basically you create a you know three markets one of them is uh um a prediction market for will the company fire the CEO one of them is um a token that gives you one share um if the CEO gets fired and one of them gives you one share and the third one is a token that gives you one share if the CEO doesn't get fired and then you can basically divide one price by the other and you can get the markets implied expected share price if you fire the CEO and the expected share price if you don't fire the CEO and then like basically based on whichever price is higher you uh make that decision right now of course like a thing that's really important to caveat here is that like share price is not the only objective function right it's like a very natural what in a for-profit context but we create organizations for many different purposes and uh we uh want um like some kind of like whatever your objective function is which for a country it could even be something more complicated like average self-reported happiness plus the exponential of life expectancy over 10 uh plus you know cube root of GDP like you could you know basically go crazy right you can Pro you should probably combine different metrics like basically you have humans that decide what that objective is and then you have a market mechanism that tries to optimize for that objective and realistically AI end up playing right and the reason why this is sort of aligned is that if you have an AI that is wrong right so that identifies decisions that don't optim for that objective function then it's going to lose money in the market and actually a an incorrect um participants coming in creates the incentive for correct participants to come in and even put more funds in in order to take their money right and so in a sense like manipula any attempts at manipulation actually pays for countering the manipulation distilled human judgment is this uh other idea that I've been suggesting recently where basically yeah the idea is that if you have a large number of decisions that you need to make a human jury provides answers to a random subset of those decisions and then you have ai models that provide answers to all the decisions and you choose the linear combination of AI models that is the best fit to the jury answers right so basically people submit you know AI submit all their answers and then humans they're only providing potentially a few hundred bits and so humans could potenti we spend like an entire five minutes potentially even 30 minutes or longer on one single bit you make sure each individual bit is high quality right so you think of it as being like a court system court systems output relatively few bits but they put a lot of effort into each bit and then AI becomes this like low effort hypers scalable thing that tries to like do a best fit to approximate what the human jury does and so deep funding is uh an effort to try to apply this to a graph where the the the question you're asking is what percent of credit for thing X goes to thing y so what percent of a credit for ethereum goes to gu what percent of credit for GI goes to like some particular Library dependency and so on and so on and the idea is that you could use this for things like retr funding right you basically just like have AIS create this graph of uh who deserves what percent of credit for what and then any funer can just come in and say click I want to support this node because I believe you know ethereum is good or I believe gu is good or I believe know cryptocurrency is good and then the credit automatically gets broadcasted down recursively to all of the nodes that were Upstream of it and the goal of this is to basically have something that's able to surface potentially multi-layer dependencies potentially dependencies that are difficult to uh detect without needing uh projects to engage in that popularity contests deep human distilled human judgment for Community notes is like one of those ideas that have suggested right or you can think of it as like dist distill human judgment for like social media posts in general right so with Community notes it's you know a lot of people love it when people criticize it usually the number one criticism is um that it's too slow right by the time a community know it appears everyone's already seen the misleading content and so the question is well can you create an incentivized Market of AIS that try to predict which posts will get community noted in real time right and so in all of these cases these are all examples of this AI as the engine humans as the steering wheel Paradigm and what they do is they allow for a highly distributed Dow to instead of asking each person to have an input on thousands of decisions which leads to either low quality or on decisions or um like 1% participation rate or everyone Outsourcing the delegacy to get centralization back instead you have each person providing a much smaller number of inputs either through a random sample mechanism or through like a futarchy type objective function and then you have AIS that use that as a signal and try to create optimized decisions for everything else so the like basically I think this is like one approach that's really worth trying and seeing you maybe this can be the like public governance mechanism of the future and like maybe this is one really valuable way to try to make Dows great again to try to make public goods funding great again and a lot of other things AI as the interface to the game right this is um and one of those interesting ones so I mean people uh you know use the word agent um a lot and uh I I don't know I personally always feel like une with the word agent uh because it implies sort of one particular form factor and it like implies that uh you like at least to be this that this kind of long running autonomy is the thing that we're striving for when I actually don't believe that right but of course in reality you know the word agent gets used to apply to like basically a whole spectrum of AI related applications where basically the AI can make kind of one or multiple calls out to other services after you give it an instruction right this cute little cartoon it's actually from Salesforce which recently released this like very lovely marketing branded thing called agent force and so everyone's getting in on the game I'm an agent you're an agent Everyone's an agent um but I think if we want to like think about it more concretely right the basically yeah the the challenge here is that like people need help interacting with the blockchain world and uh there's lots of different contracts that you can interact with and potentially AIS can help users discover smart contract systems uh that they might want to interact with for a particular function help uh users evaluate their security and um participate in them uh potentially you could even have something that like activates once in a while and so they serve this kind of Watchtower function as so they do something if something goes wrong and basically make it easier to interact with the whole crypto world and I think it also has this other interesting consequence which is that like the the interface like the AI logic is uh not something that is being provided by whoever wrote the smart contract system um but rather it could be a provided by a third party and or potentially provided by something that is more aligned with the user right and so I think it has these interesting political implications in terms of you know who has certain forms of power in in interaction with with adap right so I think this is something that's really worth worth expl exploring as well AI has the rules to the game right so this is uh you know when I go back over here right this is the concept of like let's have a dow where we just make a single AI mod model humans feed it their preferences maybe just feed it in a text file and then the AI model just like goes and like actually runs things right so the concern I have with a lot of this is that it's very easy to underestimate the exploitability of these things especially if they are open source right so there's this entire science called adversarial machine learning where basically you can take a model and then you can gradient descend attacks against that model like basically tiny things like random bits that you can insert that that optimize for tricking the model into believing something false about about what it's seeing right and so you know this this is like an attack that was trained against one particular classifier and so you know you start with a cat you add a tiny bit of random noise and it's a cat that it thinks is a dog right and it's actually fascinating how sophisticated these attacks can get like there was one that was used against facial recognition models where the attack was actually made in the form of a mask like a mask that you can wear like a face mask uh that doesn't actually cover that much and there is a version that makes you think that makes the security camera think you're not a person and there's a version that even makes it think that you're a different person right and if a model is open source that actually makes it easier to train attacks against these kinds of models so how to fix this there's many ideas that depend on cryptography uh currently this stuff is pretty far away from being sufficiently efficient F and all that stuff has overhead that's very high but we will see right um and so one approach is to like try to work on cryptography potentially work with te potentially other things and like try to fix this kind of issue but the challenge is like you're solving for two problems at the same time right you want the model to sort of be closed enough that it's not exploitable and you also want the model to be open enough that what the participants have confidence that whoever created the model did not themselves inser a bug that lets them eventually take out all the money right and so there are cryptographic ways so I get the benefits of both but it's like a challenging game right and this is one of the reasons why like I'm personally really interested in trying to encourage like more re more development in these approaches where instead of enshrining a single AI engine you're creating an open market of AI agents where it's a it's a free entry ecosystem where um every like any can join in they're being judged against this value signal that's being provided by humans um AI is the objective of the game so using crypto to create AIS so we're seeing some distributed trading runs happen I'm not seeing much progress yet in this idea of like using crypto technology to get stronger guarantees about what's happening in a trading run I think one of the challenges is that a trading run is just really expensive and uh any cryptography adds an extra cost on top of that um but I think it's interesting to start thinking about these things too right so conclusions right I think in the short term the biggest focus is on this spectrum between AI as the player and AI as the interface and I think one place where I was wrong last year is I thought about player and interface as being two separate categories right and I think the reality is that the it's actually a spectrum right like you could have a system where AIS are relative highly autonomous and it's like a script that just goes and does its thing even know like Market maker Bots they've been mostly autonomous for the last decade or so uh but then on the other hand you could also have AIS that are like very active that are very engaged um where uh and that are constantly getting inputs from their user in terms of what they want and what they don't want and the reality is like there's a sliding scale of how much input the GI participant is providing right so AI as a player in the game AI as the interface to the game I think is really the place where we want to look and I think there's a a unique big opportunity in governance so AI Dows AI based deep funding uh so I recently made a demo um if people want to just like play around with deep funding ideas if you go to github.com De fundings scoring I have a couple of fun demos there that going show what deep funing might look like in action and just like experiment with some of these ideas of course experiment with futarchy as well um and I think there's some really valuable stuff there and then of course uh you know in the future continue to work on the Egyptian god cryptography they try to make uh you know the other AI crypto intersections more efficient and you know work on the other things as well so yeah wish everyone trying to build a a applications on ethereum the best of luck oops uh I think we lost your video we'll add you back on there we go uh it was an interesting kind of uh w a lot of what the thing I'm kind of curious about is knowing has that your thinking on what the evm needs to do to adapt a lot of these applications or do you kind of still that a lot of this is around the fact that things have we should happen offchain and we should think about blockchain for uh preserving the data integrity versus uh enabling the uh the live less or the execution I think realistically not too much right like if you like analyze some of these existing like these ideas for application right you know you have whether it's def or whether it's prediction markets or or whatever it's like ultimately still you have transactions transactions are going on chain I mean maybe there is an argument that this means that secur scalability needs to go even higher right uh because uh you know with if AIS are going to send uh participate in prediction markets then like they'll be willing to do that even if the expected return from a transaction is like one cent and so there's value in the transaction fee being like 0.01 cents instead of 0.1 cents right no that makes sense um I asked this question to um an earlier panel with Ilia and Ed from arbitrum and kind of the question was around understanding what are some interesting ways to think about leveraging zero knowledge proofs here um world of centic um execution and I'm curious if you have any thoughts around how do you uh see the overlap of AIS and ZK because we've seen a lot of tees being the the way they kind of think about message passing but uh how do you gonna introduce determinism or other various kind of proof systems for compression yeah I mean I guess I it depends on what you're trying to do right um like if you're trying to to the cont of like interacting with llms and and specifically kind of uh token generation right right yeah because uh if you're trying to like the thing with ziki proofs I mean obviously again you know they have a significant overhead but then also the question is like what is the actual assurance that you're getting right and uh then because if a model was created by someone else then like you like you also wants to have assurances over know what the actual model is that's doing the the computation um you know you might wants to have potentially other kinds of assurances as well and then one of the challenges of course is that all of this is inherently competing against the like the very natural alternative which is to just like do whatever computation you want it to do locally right and there's two valid reasons to not do that one of them is like hyperscale but then with hyperscale the problem is like you can't do snarks I mean you could do tees because like the overhead is fairly low right but sarks you can't um and then the other reason is Computing over private data and like I think a lot of people are definitely interested in this use case of like trying to see if you can combine you know private data from different participants um potentially like a private model from one participant with private data from another participant I think uh there's definitely a bunch of specific applications where that Mak sense I mean potentially like you could even imagine it being used in something like deep funding right like you could you know imagine like something where like there might be models would be able to get input from um individual participants and uh like that uh feedback would be provided privately in some way where I there's some guarantee that the model can only use that that information in some like very limited way and like basically extract like at most one bit information uh so I think like that's definitely interesting intersection like the intersection of governance and privacy is definitely important um the uh I mean also even the intersection of governance and unprovability right because I guess if see and right governance gets taken over by bribes and uh I mean recently the other worst thing that's happening is like defition what's definition of unprovability like what's the Nuance here there so this is the idea that like you need to provide an input to a governance mechanism in such a way that you can't prove which input you provided even if he wants to make the proof so it's like basically the same as the M use case right but like to me I think about this as like basically any mechanism that is not a financial mechanism is a mechanism where it's valuable to have this kind of like like limit to proof because you want like a limit to trade right and uh that's uh I mean even public discourse is a good example right like I saw a couple of days ago like hu was complaining on Twitter that kaido which is this know Twitter leaderboard thing that like might have an airdrop is like basically ruining crypto Twitter because people are optimizing for the algorithm right and like that's the same thing right because like culture ultimately is a type of governance right and so it's the same sort of thing where like if you have a side input where the side input is offering you an incentive then if you take the incentive you're claiming 100% of the benefit but then you're only suffering like 0.01% of the cost of like you know like basically degradation of quality of discourse right and so the things is unbalanced so I think like yeah like I I definitely think we need to explore unprovability more and then I mean unprovability is also like a te application because I me as we've saw in the complete knowledge paper one of the use cases of a te is basically using a te to prove that you're not inside of another te that's selling your account so and like if you're you're going to be participating in something that involves like talking directly to ANM like the exact same sort of thing applies no that makes sense uh I I think it's a good transition to uh sort of get ready for the next panel because ultimately governance and understanding how do we do things more interestingly on chain uh with AI is pretty hot topic and actually relevant to you because uh apply named uh yeah Dows are missing the a [Music]