Vitalik Keynote: Ethereum and AI: Synergies and Potential Applications l KBW2024

okay vitalic how are you can you hear me yes uh I am a good can you guys hear me I don't know can crowd can you hear vitalic okay vitalic it's working I give the stage to you okay go for it thank you okay amazing um okay I am just uh try to figure out how to share the uh screen here um it's uh H one second um click this is uh so fascinating it is uh somehow not working but that's uh but that's fine um okay then I will uh do what I I will do what I can without the uh uh without without the screen in that case um so great so crypto and AI so I think this is a really hot topic over the past year or so I think uh for and some very obvious reasons right which is basically that blockchains are becoming something that's uh a really big deal and lots of people are so to build applications more in ethereum um we've had the E 4844 this month and so since then we've had like actually affordable transaction fees like under one cent uh and that's making it even easier for people to build more applications there's been all this progress in ler toos and then at the same time in the AI space we've just seen this massive boom in what you can do with AI where I think it's gone from being something that people like barely really PID attention to back in 2021 to this extremely big deal that I know a lot of us use every day in our lives in 2024 and uh so but at the same time I think it's like important to really be rigorous in understanding like what the actual intersection between these two things is right because uh like at the same time it's not true that if you just take two random areas that are really cool and fit them together that there's going to be like much of an intersection to talk about right it's like uh you know like rockets are cool and crypto is cool and you know maybe we can talk about cryptocurrencies and chains and later two is running on Mars but otherwise the two areas feel kind of separate and so to me the big question is basically figuring out is this like that kind of situation or is there really something there in the intersection so I wrote this post on crypto and AI back in January and back in January I uh talked about uh this uh like one of the first things that I pointed to right is that there was this vent diagram in my post right so like basically in the vent diagram you know on the left side you have ai and then on the uh right side you have like block chains and like basically the idea is that there's like a lot of things that are different between the between the two right so AI is like basically high on function high on function functionality but it's like centralized not non transparent and so on and so and then of course the yeah flip side of this right is that if you look at blockchains then blockchains are you know uh decentralized transparent uh uh open but the uh flip side of this is like basically that their functionality is limited and what do I mean by limited like obviously you could do really cool things on blockchains but like you know when we get excited about like nfts like nfts are just numbers in a database we've been able to make numbers in a database go up and down since the 1980s right and so to me the question is basically like what actually is in the middle here right because like there's this pretty good arguments to be made that these two things are opposites of each other right but and like one of them is sort of the ultimate technology of centralization one of them is maybe the ultimate technology of decentralization um but the question is like well can you actually use one to try to compensate for the weaknesses of the other right and so in this post I talked about like four major categories for what the the kinds of intersections that I thought that Krypto and AI could have right so the four that I mentioned were basically one is AI what I call AI as a player in the game then there is AI as the interface to the game there is AI as the rules of the game and what I call AI as the objective of the game right uh basically yeah in this case what I mean by the game is like some kind of system of rules incentives that is implemented by a blockchain right and so if you create a token you are creating a game if you're creating a smart contract system you're creating a game if you're creating a defi protocol you're creating a game uh if you're even doing something like GID quadratic funding that's a game and the arits that I make in this idea of AI being a player in the game right is basically that the like decentralization is supposed to describe the game it's supposed to describe what happens at the center it's not supposed to describe what happens at the edges right and so the way that you combine together the benefits of decentralization with the performance benefits of uh some like more centralized AI approaches is you basically make these centralized AI approaches be players in the game right so what does this mean um so the blockchain sets up some set of economic rules for participants to interact and then AIS just interact in those roles so this is actually a category that's existed for at least a decade right like we've had unchain decentralized exchanges for about a decade and unchain decentralized exchanges like I think they are the first crypto AI intersection use case and we kind of don't think about them but they've just been passively making markets more efficient the whole time uh so this is one example another example is AI participating in prediction markets right uh so this one I think is uh is fascinating uh basically so know we've seen prediction markets becoming more popular recently right so we see poly Market I know we've been continuing to see some of the others in is continuing to make progress augur is still around right and one of the challenges is that if you want to get like some like a decent level of quality on an answer to some question you need to have a pretty high level of volume and the reason is that you need to have a pretty high level of volume in order to get human beings interested in doing the work of figuring out how to actually answer on this question what if we could have Micro prediction markets prediction markets where the amount at scale in each one is maybe $50 or maybe even $1 or potentially 5 cents and the use case of this is uh I think at the medium level you basically have this kind of Oracle where you can go and ask it questions and you could get answers um and then at the small level I think there's a lot of really interesting applications in decentralized social media and the one like example of an application just to give people like something concrete is uh so on Twitter right there is this uh thing called Community notes and Community notes does a very good job of uh know basically putting notes on posts that provide like various forms of context often it's basically saying like you know like the content of this post is like actually totally fake and like here's the real information sometimes it's more nuance than that sometimes it's actually saying well the post is completely correct but here is even more information about it and so like I think a lot of people from across the political Spectrum have had a very high opinion of uh Community notes but when people criticize it the biggest criticism is always that notes don't appear fast enough and one way to think about prediction markets as a building block is that they can make other mechanisms faster they can be like it's basically like the equivalent of distilling an AI right basically yeah you take some existing mechanism and then you make some like much faster thing that gets incentivized to do an as good job as possible of predicting the yes lower thing right and so on Twitter right imagine uh you know you make a post um and let's uh let's say I don't know the post is going to be know metallic Endor um endorses uh unicorn ma uh Magic dog dog uh dog dog dog coin and let's say someone tweets this and people um you know like start uh like retweeting this right and then if you just have Community notes then what's going to happen is like basically yeah I mean you guys have seen this right like basically after some amount of time you know you're going to see a note and uh the note just says like actually italic doesn't uh and like Community notes are required to have links and so they'll have like links to some example.com blah something that I said and uh that like Bas basically says that I uh like Haven to actually endorse anything or that I I don't endorse stupid coins in general um so but that the problem is right that this happens after some delay right this happens maybe let's say yeah I don't know could be 12 hours could be 24 hours whatever the delay is but the delay is too much right and so the question is like well here's what you can do you basically create a prediction market and the goal of that prediction Market would basically be like you could ask a question of like will there be a note um saying that this post is uh misleading just to keep things simple right and that yeah that that that note is uh something that uh you know you and then people could start betting on the prediction market right and the prediction Market here could have like a very small size like it could literally be 0.1 when it's 0.1 there's no point whatsoever in human beings participating in this right this is basically just llms Bots AIS algorithms that are fighting it off against other ones right and then the nice thing about this is that like you actually have a ground Truth at the end right and the ground truth at the end is you can have a more democratic mechanism that's much much slower that involves a lot of people um that making a decision and then the this mechanism will actually might put the note at the end and then it would then give rewards to people who predicted correctly right so over here you um basically what we have here is a mechanism that is low cost uh but fast but it's uh in terms of uh like the like it's missing incentives right basically because if you just like get llms to do things on their own then there's a question of like well are the LMS going to be biased what are their motivations who's actually training them and all these things right and then over here what you have is you have a highly legitimate mechanism um that is like like accurate robust but it's uh expensive and slow and so what you use is you basically say you're using this thing like the prediction Market as a predictor for Comm for for the community notes and that way like you actually get signals uh basically yeah saying know instead of like this being in uh 24 hours you know you might have a note you know look over here saying uh prediction uh markets show 93% chance Miss uh um misting and uh and so you actually get like something that is both democratic and fast at the same time uh so this is uh like just one example of prediction markets as a player in a game so the so this is one section right now let's talk about some other sections um AI as an interface to the game so this is uh also a really important category right here basically yeah the idea is that like this one thing that already exists is that crypto wallets have all kinds of features in them that show them that um know a page that a user is about to visit could be a scam could be misleading could be something else right and like scams and fraud in general like starting to be an increasingly a really big problem and AI is unfortunately going to keep making the problem worse and scammers are really sophisticated so I actually wrote a post about a couple months after this where I talk about the for people to start understanding and using security questions everywhere but what if AI can also help you here right and the way that AI would help is basically you could have models that are running client side so fully on the side of the client not sending any information off to any centralized server purely acting to protect you and uh these models would analyze all of the messages that are coming in and they would give you a warning if they is something that's potentially misting right and this is something that I thinkes sense for like text messaging it's something that makes sense for social media I think it's something that uh also makes sense for crypto wallets right so this is uh this would be wallets incorporating features right now there are there isn't really any use of AI right now it's just like very simple like black lists and white lists but other potentially AI could make this uh much more powerful so AI is the interface to the game now ai as the rules of the game this is like something that's a lot more tricky right so when we say AI as the rules of the game what we mean is like there's this idea that people get excited about sometimes of like let's have ai judges because AI judges are going to be less biased uh for example right and like that's uh and I see the promise but also that's risky because we know that like actually AI models definitely have biases their biases come from their training data and actually if the training data is opaque then the biases of thei are going to Bea as well and that's one of the problems the other problem is that there's this issue called adversarial machine warning and uh the issue with adversarial machine warning is basically that like you have like there are things that people have like I'll just give one example directly right people have literally created masks like masks that you can wear on your face right like combination of like Co mask maybe like something on your eyes and if you wear this mask then at least for like some subset of cameras those cameras will not recognize you as a person or potentially they might even misreport you as being a different person right so like these kinds of attacks are like actually pretty challenging and one of the issues is that well AI if you have ai where the pipeline of like how the AI was trained is insecure then you have centralized attacks but on the other hand if you have AIS where the way that the AI was trained the data the entire process was open then that actually makes it easier to do other kinds of attacks which is basically attacks where you try to like optimize these kinds of masks locally and then you run it against your local copy of the AI a billion times and then you will take that attack and you apply it in the real world right and so the question is like can we create AI models where you get like you get confidence that the trading process was fair but at the same time like you the AI model can sort of contain some kind of private key like some level of data that can be used to like actually fool these kinds break these kinds of attacks and so with cryptography you actually can right and worldcoin has actually been exploring a lot of this stuff really heavily right and and so like basically you could use multi-party computation FH other kinds of things to try to make like AI training and AI inference actually be like both privacy preserving either on the model or on the data or on both and reasonably efficient at the same time and if you add zarch ideas that you can also make it provable now the fourth piece is AI as the objective of the game and in that case what I mean is if we can use some comb of blains and cryptographic tools to make a then we don't just need to use that for projects like worldcoin that are themselves ctor projects we can use that for other kinds of AIS as well right and so potentially there's even applications here and just uh improving the safety and improving The credibility of uh AIS much more generally so and then finally you can have Dows that try to do like Democratic governance of like what goes into the trading data and the trading process of AI right so those are some of the specific intersections that I think are promising I think uh the big takeaways are right that I think you really want to be rigorous in trying to understand you know what actually the intersection of these two things is Right basically if you're building an application that incorporates some level of crypto and some level of AI then you need to make sure that you know it actually makes sense the AI is being used for the things that the AI is being used for and that you're actually getting guarantees from uh the crypto part of your application right there's uh a lot of AI applications where or crypto AI applications that I've seen that basically just look like oh you know you just published some data on chain and you like put hashes of trading data and but the question is like well is that actually yeah going to be that like by itself right like what's the actual security property that do that gives you and so I think if you want to build an application that incorporates sub combinations of both then like it's really important to have like a really clear and a really technically strong story about how those two pieces actually fit together and I think if they do then like I think there's a lot of opportunity to do some uh really powerful and amazing things [Music]