Vitalik on Ethereum's Role in the AI Agent Era

So, I see Ethereum providing like two big functionalities for applications. Right, one of them is the public billboard. It's a place where anyone can post data on-chain and then applications can interpret that data in many different ways. And then the other is the on-chain computation. And on-chain computation includes a lot of financial applications, includes other applications, includes DeFi, includes many other things. I think with AI it's fundamentally the same use case as well. It can continue to be important. I think AI will definitely make the way that we interact with the blockchain and with other tools that we have very different. Like there's many differences, but one of them is that in a pre-blockchain world as a user you're interacting through one interface and then that one interface is like an interface to one thing. And with AI, as it exists right now and probably the next few years, what we're seeing is that we're seeing tools and then you have skills and then you can have like a user-side AI that can use these different skills and it can just do things that combine like interacting with many different things at the same time. Intuitively like that should increase the amount of workflows that we see that interact with Ethereum and also interact with other kinds of things. And so I think to me, you know, like the operating system is like almost the wrong metaphor in some ways, right? Like I think the operating system will continue to exist, but like it will be smaller, it will be more simple. And then we will just have a collection of like different tools, different skills that like AI will help people like use and do things with. And blockchain to me is like the most natural way to allow like applications to happen that allow cooperation between many different people and that kind of work in the long term and work without needing to like agree on like who to trust. And then the other thing is also the economical layer, right? This is the layer where like blockchains can support AIs, right? Because like there's this difference of like can you have more decentralized AI and more centralized AI? And if you have more decentralized AI, that means you have the like different AI things like agents, programs, whatever you call it that are built by different people that are controlled by different people that needs to interact with each other. And for that interaction to be possible you need to have an economical layer, right? Because either cooperation is based on like economic incentives, economic rules, rewards, and like penalties, or it's based on central control. And so usually one of the two, right? And if we can set up the economic system, then like that makes more decentralized interaction between AI possible.

I mean, I personally like I still think of people as being the users. And I think of AI as like being the replacement for UI. Like it's the new ways through which like people can interact with the chain. And so like if you think about like for example, if you use the internet to like learn about something, right? And then if before AI, then you might search for it on Google, you would access the page on Wikipedia. If it's technical, like maybe search Stack Overflow. But then with AI, it's like a very different workflow, right? Like you ask a question to you know, the bot, and then the bot does things, and it answers. And like I think, you know, we're going to see similar kinds of changes, right? And this does like mean big changes in terms of what properties we want the blockchain to have. So, like for example, one of them is latency, right? Like when a human is interacting directly, like often low latency is very important, right? And I think with agents like there are some situations where low latency is good and like actually you want even lower latency. Like if an AI can think a thousand times faster than a human, then like there is benefits in pushing it down to 1 millisecond. But also other situations where latency actually does not matter. And like because if you think about like for example, when you ask Claude like something or GPT or DeepSeek or anything, like a complicated question. What it does is it thinks, it makes internet searches, it gets responses. Like basically you ask the question, then you go get a cup of coffee and then you come back. And so it's like actually okay if the latency is a 1 minute. And so like we're going to see both, right? So like this is a big change in you know, like how we need to think about like even the chain. How we think about the interface layer will change a lot. And like if we look at the wallet for example, right? Like the piece that I think will stay the same is the SDK, like the core software that is just it's like an API layer like in as a function. Send coins, you know, tell me my balance, like do something privately. And but then everything around it like maybe you know, like we don't need a wallet, right? Like you just need you know, you talk to the SDK, right? What we're going to have is we're going to have these like different packages of software that's like people work hard on them, make sure they're secure, make sure they're formally verified and you have like a skill file and then they're designed so that no matter how you use them, like you have a strong assurance that it's secure. And then like AI works around that, right? So, yeah, I think at the layer kind of in between the chain and you know, like the user, we're going to see very large changes.

Yeah, I think it's always good to try to like decompose identity as much as possible and understand like what are the claims about yourself that you actually need to prove to me to do one particular interaction. Like I think for most interactions, it actually does not make sense to fully reveal who you are. I think in most cases it makes sense to reveal like some things about who you are. And like you can make zero knowledge proofs about your reputation, you can make zero knowledge proofs about like the source of your funds, you can make zero knowledge proofs about all kinds of things. And I think it makes sense to like figure out some framework that makes it easy to do these kinds of things. And like making it easy to zero knowledge proof things about Ethereum, about like things that people have, making it easier for wallets to help people store private information. And then we also want to figure out like for different applications, like what is a version of this that like actually satisfies the needs of the application and then at the same time protects the user's information as much as possible.

No, I think it's fine to continue to build on top of the same framework. Let's say like agents themselves can like figure out you know like what things they need to do on L1, what things they need to do in different kinds of L2s. And at least right now, I know the reasoning that they're using is not that different from like the reasoning that people are using. And I expect that you know the market will figure out you know the right distribution of tasks to these different platforms.

Yeah, this is tricky because figuring out how to fund public goods is a governance problem and to have governance you need to have a theory of who the stakeholders are. And you know I think there's still stakeholders, right? Like for every agent there's people who are running those agents who have who had some reason to run them. I think like AI and ZK both open up many more possibilities for how that kind of governance can run. And then at the same time I think you know AI does break a lot of the previous mechanisms because basically a lot of things that used to require a lot of high amounts of effort to break, you know we can break them automatically. And that's something that we will just have to like keep figuring out, keep iterating over time.

Yeah, I think I mean L2s are important, but I think we need to be more imaginative in how we build L2s. Like I think the thing that we have been doing for too long is basically saying, let's take the EVM, let's make a copy of the EVM, make it 1,000 times bigger, make it more centralized, and then publish like hashes onto Ethereum, and then we're done, right? And like I think one big thing that the Ethereum community has been realizing is that this actually does not make much sense. Like the thing that makes sense for L2s, I think, is start from the point of view of an application, and then ask basically what are the pieces that the application needs that the L1 does not provide. And then try to provide those pieces, right? Like to me, in an ideal world, applications would have some pieces on L1, some pieces on L2, right? Like ideally, even for example, I think like user accounts should all live on L1. And the reason is that user accounts are something that gets updated like very rarely, right? And it's like okay for that to be more expensive. And then individual transactions, like trading, order matching, like that can happen on L2. And or maybe you know, like L2s can provide like functionality for privacy. So, like for example, like I already like actually like I think of things like Tornado Cash or Railgun or privacy pools as already being L2s. Like they're not L2s for scalability, they're L2s for privacy, but they are L2s. And so, yeah, I think we're going to see like more different kinds of L2s going in different directions.

Yeah. I think that has many dimensions, right? I mean that has to be yeah, something that is intuitive like for people to use. It has to be something that I think like works well as part of a bigger ecosystem. Like it doesn't you know like try to take over a user's entire life. I think like taking privacy and security very seriously, right? Like by default I think even the open source AI world is very bad at this, right? Like we see you know like open clients, like very cool that like people do a lot of things, but then also like extremely easy for people to just lose like all of their data and get hacked in all kinds of ways, right? And so, we have to think about all of that in different aspects of this. I think you know the game definitely changes. Like one thing that I'm excited about with AI is that I think like we actually can have AI that's like intentionally aligned with the interests of the user. Like I think we need companies and like teams to build AIs that are not attached to you know like any company or like any application. Just people that try to make things that like serve the user as well as possible. And you know the more intelligence there is sort of representing and helping the user I think is like that can really help make sure that you know like people are treated fairly and like we don't see as many attacks, you know, like we don't want to see like applications that are exploiting users. And then at the same time, you know, like we'll be yeah, you would want applications that are able to interact with the setup that a user already has, which might be customized in different ways. And to be able to interact with like other things that a user wants to use and to be secure. And like if you have all of those things, then I think that's like ideal.

Yeah, I think I mean AI, we've been interacting with it in two ways, right? One is also using AI to start like building Ethereum, right? So, the Ethereum team is starting to use AI to do formal verification more and more. And like we think it'll be both possible to achieve like much higher levels of security than before, but also it will be basically mandatory to do that. Basically, if you have any bugs, then super powerful AI will discover bugs and break the network. And that's something we have to be really on top of. Yeah, and also we've been starting to you know like also do agentic AI wallets inside of the Kohaku team and trying to figure out how to combine like that with making sure it's privacy preserving, making sure it's secure, making sure that supports all of the different kind of like security properties that we care about. Making sure it's friendly to local AI. Yeah, because you know, like if your agentic experience is always dependent on like calling into some third-party server, then like you don't have decentralization, you don't have privacy. And so but then also at the same time, you know, like AI is kind of vulnerable to attacks and you know, you want to put limits on like how much AI can do and like doing that well is definitely the wallet's responsibility. One other thing I would think about is that I think it's probably not healthy to think about like interacting with Ethereum as being like a separate universe from interacting with everything else on the internet. I think what we want to move toward is like an operating system where local AI helps a user do anything. Ethereum is one of those pieces, but then you also have ways to, you know, securely and privately like search other things on the internet, interact with other kinds of protocol, like have local files, like learn information. And yeah, think about this in a very full stack way.

Yeah, I think for payments, probably the next step is ZK payments. So, we've been working with something called the ZK API, published a few different things. And the goal is basically that's like if you make any kind of API request, which can include an LLM, it can include internet search, it can include RPC, anything, then like basically like that request, like every individual request is private and it's not linked to any other request. And like that's actually important because like with especially with AIs, like having pseudonyms, like having anonymous identities, you know, like if they're persistent identities, then like still like you lose all your privacy. Like you lose enough information, right? And so, what we need is like to separate each request from each other request. And you know, that's harder to do. And we figured out like how to do it in a ZK without requiring an on-chain transaction for each request cuz obviously that would be too expensive. And time-consuming. Exactly. Well, I think the time-consuming is fine, right? Because like you just put it all into a program and then the program does it, right? But yeah, definitely too expensive and highly insecure. So, that's something we're working on, I think. And then also there's ways of like combining that with security deposits and then you can basically protect against kind of abuse both on the user side and on the application side without actually breaking anyone's privacy. Yeah, so we're doing a lot of interesting work in that direction.

Yes, I'm so too. Thanks for the interview and thanks for your time.