Intersection of AI and Blockchain

Transcript

invite Miss queenn and vitalik Burtin founder of ethereum to the stage together we will open a special gift box as a symbol of our appreciation for their remarkable contributions to the blockchain community now let's give them a round of [Music] applause one more big Applause [Music] [Applause]

[Music] hope we going to be have a really exessive year ahead thanks Miss Gwen and vitalik Burtin and right now I would like to invite to the stage co-founder of ethereum a man who intellect has reshaped the digital landscape Mr vitalic [Music] Burtin gu let's give him a big Applause [Music]

[Music] okay hello everyone and good morning so no glad to see so so many people coming out here and uh definitely a uh great job on the organizers for choosing such actually interesting and fascinating venue um so looking forward to spending the the the rest of our two days here to here together I hope we're going to have

lots of fun um so today um I wanted to um talk about um the a topic that I think a lot of people have been thinking about for a long time which is the uh intersection of blockchains and AI right now this is uh one of those areas where a lot of people have uh been trying to look for intersections for probably almost 10 years now I yeah some almost every time I go to a conference and I and and when people ask me many questions often at least one of the questions is about you is there some

kind of AI application that fits in here and at a very high level like you can kind of see why people think that that there might be good synergies right Ai and blockchains are both very important fields in technology um and um they have some kind of interesting properties that seem like they complement each other right so AI is centralized blockchains are decentralized um AI is not very transparent blockchains are all about providing uh transparency um so there's a lot of people that are the there's a lot of people that have this kind to very intuitive idea

that there's like some kind of natural connection between the two that make that just makes a lot of sense right but I think uh the Big Challenge here is um basically asking the question we have some of these very high level ideas about how blockchains and AI might connect with each other but how do these ideas um how can they be realized in actual applications right like what actually is a specific example of an actual application that has some kind of blockchain component um maybe some kind of ZK component and also has some kind of AI and where it actually makes sense

to put those two things together right like you're not just combining buzzwords you're creating a solution that like actually makes sense and actually works for people so my view about this is that there's four main categories that we can think about and the way that I think about these four categories is uh the first is AI as a player in the game the second is AI as the interface to the game um the third is AI as the rules of the game and the fourth is um AI as the objectives of the game now when we build blockchain applications we often think about blockchains as being platforms that let us create games now sometimes when we

say game we mean I know the kinds of games that people play to have fun um but more generally yeah we mean some kind of environment where different people can come in and interact with each other that environment has some kind of rules to it and the rules are designed so that when people come in and when people interact there's some some kind of outcome gets accomplished and so the question is if you have some kind of blockchain based system could be a smart contract could be a dow could be something else then like in what piece of that is the piece where an AI can actually fit in um so the first um

answer here that I have is AI as a player in the game right so one thing that we already have and actually that we've had for about 10 years is AI is doing Financial Market Arbitrage right any kind of platform that has any kind of Finance you're going to have market makers you're going to have Traders you're going to have Bots um and this is true in traditional Finance this is true with financial obligations on top of blockchains it's true anywhere right so this is just like one kind of intersection and we don't really think about it as an intersection um but it's something that people have been doing have been doing for 10 years and like it is something that makes sense but in the

blockchain space I think uh we often like to say that like Finance is only the first app right the most interesting thing about the space is what other kinds of applications you can build on top of it and I think here too right the you can ask the same question we already have a playing in financial markets what are some other examples of AI playing in games that might make sense to come up with and uh try try to implement one thing that I think is going to happen soon is um AIS in prediction markets right so prediction markets are becoming an increasingly Hot Topic over the last few years basically the idea is that if

you want to try to get like more accurate right information about things that are going to happen in the future you create an open market people can make bets on it um and you have a market where people who have different opinions um about what's going to happen can trade against each other one of the challenges with prediction markets is that they're very um some they they have a hard time um succeeding unless there's a lot of people who are participating right if you have hundreds of participants if you have thousands of participants they can work well but if you only have a very small number of participants often they

don't they don't work very well and so for questions that are not extremely popular often prediction markets so far have been weaker in practice than they are in theory so this is one of the places where AI can come in basically instead of just having people uh participate and make Shades you can have AIS come in and make Shades and you can have AIS that can read the internet and that can make more complicated judgments about all kinds of questions and like this becomes a higher level game right instead of having a game that is about you coming in and making your own answers to each question individually the game is about who can make the best

AI that can answer these questions the most accurately and if you can answer the questions the most accurately then you you can get the uh you can get the most Revenue um so this is something that um um Omen which is a prediction Market from the gnosis team has been experimenting with this is a screenshot of uh AI Omen it's a a live demo it's uh actually um it's actually live and already working but this is also just only kind of one one other um example going further Beyond this right another example is you can make markets for arbitrary questions that AIS might want to ask or or that people might want to

ask right if you have some kind of a question that you wants to ask and you wants to get some kind of highquality answer to it then you can a you can create an open market and you can ask the AIS now the purpose of the blockchain and the purpose of uh people use the participating in the blockchain mechanism will be to basically provide the incentives right so people can do spot checking people can spot check a small number of of different questions and you can have AIS participating and you can have AIS potentially answering a much larger number of questions very high volume very low cost this is May is

can be useful in decentralized social media um it um Can could be useful in dows it could be useful in all kinds of different applications so AI as uh as players players in the game this is one option um AIS as an interface uh to the game is um one of another one of these uh really important things right so what do I mean here so today we have uh we have BL chains and when users are interacting with blockchain users are generally interacting with a wallet right so who here has used an ethereum wallet in the

last week okay in the last day good so the wallet I mean the word wallet I think is a little bit incorrect because uh you know the word wallet like means money but it's really not just about money right a wallet is like your entire interface to this big decentralized ecosystem that has all kinds of applications on it right and at the beginning maybe 5 years ago or 10 years ago wallets were like fairly simple fairly automated pieces of software right a wallet I you know you would just uh store your key if you want to do something you would U you know you would click and then the wallet would uh have you type in a password it would have you

approve and it would be pretty basic today wallets are becoming more and more and more sophisticated right um so what's happening with wallets today is uh you're starting to have things like for example metamask has an anti-cam feature right they have a list of uh different uh of addresses that metamask knows are connected to scams and if you try to access them then it gives you a big warning um there's another wallet called Rabbi that has a built-in transaction simulation feature so when you are signing a transaction it actually like shows you these are the effects that the transaction is going to have and if the transaction looks

dangerous then it shows you more warnings right so wallets are becoming more and more complex pieces of software that are really uh not just being these kind of completely neutral gateways to just let the user do what they want but to like try to actively inform the user actively protect the user um like actively kind of be this layer in between user and the applications that's there to be this more sophisticated actor that actually looks out for the user and protects their interests today this is all done with simple algorithms I expect over the next 5 years this will be done with AI as well so that so

that's the value of AIS as the uh as an interface to the game so third category is AI as the rules of the game right so this is where where I think uh we start getting into some of these kind of deeper um and more challenging applications of AI basically you know imagine um using an AI as an oracle to give final judgments about some kind of complex questionability outside world right so for example you could have a price Oracle right and um you could imagine an AI that searches the internet and actually tries to determine what some particular price is um another good example of this is um in

the worldcoin project right so the world coin project is trying to give people um like one world IDE to uh to every unique uh person and if you have a world IDE then you can get some quantity of world coins every two weeks um and the way that you sign up for World coin is I you know by scanning the orb right so who here how who here has a world ID okay a few people that's good now in order to um tell you know who actually is a person and who is fake right in order to kind of tell the difference between a person scanning and you know making like a deep fake of a person and like printing out a picture and then trying to like scan with the

picture they have pretty sophisticated a AI algorithms inside right right now this is all kind of fully uh trusted like you have to trust them to have ai algorithms where those AI algorith are cannot be exploited by random attackers and cannot be exploited by the team right so right now there's a lot of trust in the system and I think there's an interesting question of uh can we use cryptography to try to reduce the amount of trust right can we try to use cryptography to try to actually make AIS that are more secure and that are closer uh to being trust free in the same way that other kinds of programs are um so

the big problem here is aders attacks right two kinds of adversarial attacks um they can come from the inside and they can come from the outside so in the traditional World generally you worry about outside attacks and you solve outside attacks by just like hiding everything right the less you know the the the harder it is to attack but in the crypto space you know I think when we believe in decentralization we believe in making systems secure against the inside as well right and making system secure against inside attacks is uh I think really worth doing but it is also more challenging um so adversarial attacks um here are some two interesting

examples of this in of this out in the real world right um so on the left here you have a picture of a cat on the right you have a picture of a cat but if you ask an uh if you ask a particular AI the AI will tell you it is very convinced that the picture on the right is a do why right basically what you can do is you can train an attack where you can identify very tiny changes that you can make to the cat on the left that will trick the AI and make it think that the cat with a few very small changes is actually a dog basically it's like optical illusions but optical illusions for robots the problem is if you can see

the inside of an AI model and if you basically the more you know about an AI model the easier it is to train extremely effective attacks now what we have on the right here is um basically adversarial attacks that take the form of a face mask that you can wear right basically what they did here is they took an AI um Ani for facial recognition and they optimized an attack where the attack is that you wear this ma this mask on your forehead and if you have a mask on your forehead then the AI yeah the facial recognition AI does either does not recognize you as a person or recognizes you as a completely different

person so this is challenging right because if you imagine a system like worldcoin you could imagine training an attack against it where something that's not a person can like basically wear a mask and pretend to be a person so this is challenging right so basic and I think the big reason why these attack are challenging is that they can be trained with only a limited amount of access to the model like you don't even need to see the full model to trade an attack right these attacks are like they're actually pretty scary and you know they can be like you don't have to fully control an environments to to provide an attack you can just wear a mask

so the challenge here basically is um you know can we use cryptography to try to protect against these attacks um so actually we can um so what we have over here is this is a diagram that shows the math of what goes on inside of an AI model right and what I think is interesting about um about the math is that it's a very specific and like very structured kind of computation right so who here has worked with Z Nars before so if you work with Zas Nars you know that zekas Nars have very high overhead right if a computation takes 1 millisecond to do maybe it takes 1

second or maybe 10 seconds to prove now ai computation is basically a combination of a large amount of matrix multiplication and a smaller amount of like these very simple nonlinear calculations matrix multiplication is actually very easy to prove if you have um a very large M Matrix Del m multiplications you can optimize proof systems that do a very good job of proving them very quickly so making um AI computations be provable is actually much easier than making regular comp regular a hard computation provable

so you can do this with zi snarks you can also do this with multi-party computation right you can make a system where you have three computers and each of those three computers has one part of a model um and you in order to compute you need you need all three of them to you need any two of the three to participate um but this the system continues to work even if any one of them is dishonest right even if any one of them is dishonest you continue to get correct answers and no single one of them actually has can extract any significant amount of information about the mod

right so you can do an NPC and you can actually preserve privacy you can basically models where no one can see the inside not even the Creator and if you make a model like that then training attacks against it actually be actually becomes much more difficult um so one interesting long-term goal is basically ask the question can we make a system where we can use some kind of Dow on the blockchain plus some kind of of cryptography offchain maybe MPC um to both to train a model and to run a model and if you do this then because the process the for for who sub submits the training data is an onchain mechanism

it's a dow and and there's fair and transparent rules that show who can submit people are going to know that the AI model was created in a way that is open and that is transparent and they're going to have that guarantee without even being able to see the model right no one can see the model not even the Creator can see the model but because you have this process for creating the model where the the Integrity of the process is guaranteed with a blockchain based Dow and using cryptography you can still have guarantees that this is something that's trustworthy so you can imagine applications that wants to use AI as the rules for the game to try to

do something like this um finally AI as the objective of the game right basically what if we want to use blockchainbased mechanisms and cryptography to train better AIS that we can then use for other purposes so one example of this is what I talked about right now using cryptography for privacy another example of this is using blockchainbased markets and blockchain based incentives to create open markets that allow anyone to come in and participate in traing a model so four types of um four categories of application um I think uh so some just

general conclusions about this right uh so first of all like I think intersections of uh blockchains and AI are real they are important but at the same time like I think it's also important to like remember that there's a lot of challenges and probably should not try to consider this as kind of the narrative that will carry the entire space forward right there's lots of other things there's lots of other amazing and important types of blockchain applications decentralized social like in my opinion is super important um continuing to um to um you know improve on I you know decentralized identity reputation proof of personhood

also important um making better versions of all of the existing applications also important you know games also important def also important like I don't think that there is like one single category that's going to you know kind of Dominate and I carry the entire Space by itself right and I think if we're building a blockchain AI intersection then the easiest kind of thing to build is um AI as a player in a game right the rules of the blockchain Define a game and the rules of the rules and the people participating like the humans participating create the incentives and

then the AIS participate as the players I think this is a design that is fairly simple it's uh it doesn't depend on any kind of hard technology that's still under development and you can just build it and there's a lot of different applications for which I think this kind of form factor makes a lot of sense so if you want something simple if you want something that's doable now I think uh AI as a player in a game is is a great place to start now going forward um AIS as an interface are I think important um we are I think in the next 10 years of ethereum ethereum is going needs to really reach users on a much larger

scale than before and I think to reach users well we are going to need much more sophisticated forms of wallets and interfaces to try to make the uh experience of interacting with ethereum both more understandable and easier and at the same time more safe um but the harder ones are AI as the rules of the game right basically because if you make the AI actually be the rules then like you're putting a lot of trust into it and you're putting a lot of trust into something where we sometimes it still breaks right and uh we you know we know how to train attacks against it and so you can try to use cryptography to

protect against these things but the challenges in trying to do this are greater right so I think uh benefits are really great but challenges are also really great as well and uh there's a lot of uh you know really fascinating math and really fascinating cryptography that goes into doing these things well um so if anyone here is um interested in building a I uh blockchain applications um you know first you know make sure you have an idea that actually makes sense an application is not automatically good just because it has an AI and it has a blockchain um you know and first it has to be good as an application it has to

provide something that's useful then the AIS have to provide something that's needed and the blockchain part also has to provide something that's needed and if you can figure that out then great right um and then the next question is like how do these two different pieces fit together and I think you know I talked over here about a couple of different uh strategies that you can use um and uh you know there's a lot of uh room for create for creativity here but I'm and I'm looking forward to seeing what kinds of things people here come up with thank you [Music]

thank you Mr vitalik bin for sharing useful information with your future vision and myself and all the guests here are looking forward for the coming upcoming explosive development of AI and blockchain and hopefully those of us who can be said be the uh to be the Pioneers will gain the great achievements and hope vital will have a lot of health and bring ethereum more and more successfully [Music]