Fireside Chat: Verifiable AI Agents with Sreeram Kannan and Nikita Sachdev

Transcript

[applause] Welcome everyone. Um I I was already introduced. I'm the CEO of Luna PR. My name is Nikita Sachev and I am joined here by Sriram, an icon in the Ethereum ecosystem and the crypto space. Um uh Shirram, maybe we can start off with a quick introduction. Uh hey everybody thanks Nikita really excited to be here I'm Shirram uh I started IGN labs uh four and a half years back before that I used to be a professor at the University of Washington Seattle um we we built at Igen labs IGEN layer which is a mechanism to extend Ethereum to run all kinds of other networks on on top on top

of IGN layer we built a bunch of other things IDA is a data network if you heard of a blockchain called meth rise these blockchains run on top of DA. Uh we also built IGEN compute, IGEN AI, many other services. I'd say like for our part, we've built a lot of infrastructure. So the question is what the hell do we do with all this? >> Well, today we're talking about the verifiability verifiability layer that um the a AI agent economy needs. It's a little bit of a mouthful. And um what that means really is something that I want to ask and put towards you guys to the audience today. So just with a show of hands, I would like to see how many

of you would trust your AI agent to book a flight for you. Just show of hands. Raise your hands if you trust your AI agent to book your flight. Right. Most hands. If you're not raising your hands, then get with it. [laughter] It's just a flight. Keep your hands up if you would trust your AI agent to make a $100 transaction for you. Okay, a lot more. Now, keep your hands up if you would trust your AI agent with your passport, your identities, full bank account, all your documents, everything else that you own.

Very few hands. I think I see one. What needs to change? >> Yeah. Um the when you think about an agent and uh its action space, how do you verify that the agent's doing actions within a constraint space? I think this is a very important problem. I would say there are two layers of verifiability to it. The first one is when somebody says they're running an agent, is the agent running exactly the code that they're saying that they're running? Is it running the model that you think that that is being run and so on. So that's one level. The second level is is the does the agent do what it said it'll do right how do you constrain and enforce these constraints

on top of these agents that's the core trust problem and today when you think about the agents the second problem which is how do you enforce constraints I think is the dominant problem the the reason is when you have s you can stratify agents into agents that are used by a single player a person using an agent for themselves or you can have agents which are shared across multiple different people and mutually untrusting parties and these two require very different agent deployments. When it's a single person running an agent, you can run the agent yourself. Set the constraints whether it is through your uh you know some kind of a shell you

know uh Nvidia came up with a concept called Nemo shell which is basically a constraint system. You can run it yourself or you can run it remotely by a cloud or a provider that you trust. But when there are multiple people interacting with the same agent, for example, I want to build an agent that runs um that adjudicates a prediction market. Okay. So when you think of an agent like this, it's a multiplayer agent. It sits in between multiple players. Now the trust becomes even more paramount because who decides where is it run? Who decides how the prompts are changed? All these things have material impacts on how the prediction market will resolve. So agents need [snorts] higher grade of trust when they're

multiplayer than when they're single player. And this is a we at IGEN we built a lot of solutions to actually uh solve this problem. If you run like an agent on top of IGEN compute you can actually guarantee that the agent runs exactly the same set of prompts uses the specific model uses the specific data in order to actually make decisions. But um so Shirram like you know we're I was walking outside the hallway and I every time I walked by I heard someone talk about AI. A lot of the booths here a lot of crypto companies have pivoted into AI. But what makes you guys the right company to actually do what you guys are doing? There's a lot of

competition now. Everyone is trying to incorporate AI. >> Yeah. Um, I think it's the the AI is like the tsunami that is hitting everything and I'm glad the crypto space is responding to it by incorporating it and adapting to this thing. Uh, my own relationship with AI even AGI I would say goes way farther uh like you know people talk about their AGI moment when they uh feel their job can be done by an AI. I had this in 2017. I was actually working with this uh physicist um doing DNA sequencing. So as a professor we were working on this project where uh you know we were building mathematical

algorithms to take a DNA transported through a nanopore and then infer what the DNA sequence was and my specialty was doing mathematical algorithms for this. One of my students came up to me and said you know Shri Ram you should really run AI for this. At that time it wasn't even called AI. This was called deep neural networks. [snorts] And he said, "Oh, you should use a deep neural network." And I said, "This is a stupid idea. Maybe you should use deep neural networks for problems where you don't have any any idea what to do." But here we have this physicist. The physicist I was working with actually had invented that nanopore. He also has the sharpest measurement of the gravitational constant G. His name is Yanskll. I said, "We know all the

physics. We know the math. This is, you know, to try to actually do um AI. Student went home, came back 2 weeks later and told me he beat my benchmark. I said, "This is not possible. You can't beat my benchmark by just running blackbox AI." Turns out he was right. Then I started looking into it. And whenever a problem is objectively verifiable, when you can actually measure if you're doing well, an AI can optimize the heck out of it. And that's exactly what was happening. You we we had a neural network. We had this DNA problem. There was a lot of data. You just optimized the hell out of it. And it was better than what the physicist and I could put together myself ourselves. And so this formed this

informed why I actually got into crypto and blockchains because firstly I felt like I kind of lost my job to an AI. So this is 2017 November and um it led us into many different directions. you know, we ended up doing way more in AI and biology. I ended up writing a paper with uh um David Baker who ended up winning a Nobel Prize with Demis Hassabis. But this is all like in the past. But what I did is I when I switched into blockchains, it was very clear that if you if you build a blockchain and you can't run a machine learning system, a neural network system on top, you're going to be left out. So in the very first white paper of IGEN layer we said what's layer IG layer is

an extension protocol for Ethereum. So you can run any network you want on top of it. What network should you run? You should run machine learning AI agents these kinds of things. And so it's just been a journey to actually kind of one of the things we realized is we built this like layer on top of which other people can build infrastructure and it has been a progression for us to go further and further up the product stack. We built layer anybody can build networks on top. We built DAG compute all these things. So that those were services we built ourselves. Now we're going one layer further to say no actually we should do the most important

use case that can be built on top and like and and that's our big new thesis we call it agentic companies. >> So you've always had AI in mind like you've it built um IEN with with the um I guess the thesis that you guys would eventually move into AI. It just wasn't called AI. I was what you say deep neurological networks. >> Deep neural networks is what people used to call >> neural networks. I guess um I haven't so see I haven't heard anyone talk about deep neural networks and I guess that just proves that you're um very wellversed in what you do. I mean I'm talking to a guy who has a PhD from an IV league school. So of course but I'm still going to hit you with hard

questions otherwise this panel is not going to be interesting enough. Um so would I en layer move as quickly as it is today without AI? Yeah. So I I'll talk about IGN Labs, the company behind IGN layer, right? We are basically seeing the pace at which AI can enable product development and growth. Like one really fun story I'll [snorts] share is two months back Google came up with a major quantum breakthrough and uh so they had a quantum algorithm which is much better than anything that exists and they said oh it's so sensitive we are not going to release it. So they didn't release this algorithm. They said it's 10x better

than anything that exists out there. And what one of my, you know, uh, colleagues, he's actually not yet finished his college. He decides to do is to say, oh, if Google has this result, it's, we know that this result exists. Can we have a bunch of agents to go and figure out what the hell they did? Okay, so he takes the problem, sets up an objective benchmark and unleashes a bunch of agents to go improve this solutions to this problem. And his agents are working day and night and they they make some progress. And then he comes and tells me, hey, you know, I improved the state-of-the-art without Google's result from 25 billion to 10 billion just by using a bunch of agents and but I'm stuck there. So and then I

suggested why not we run create an open network where anybody can come and use their agents their knowledge to actually improve this benchmark. So we launched this challenge called ECDSA.fail ECDSA. So the importance of quantum is if quantum computing exists you can break you know bitcoin signatures you can break all kinds of internet encryption. So we we started this challenge and then we launched it and within 6 days right now we've not only beaten Google's benchmark, we're actually 40% better than anything that exists. So it's been unbelievable to see that you can take

one of the hardest scientific problems, building a better quantum computer, and then have a bunch of anons. Basically, none of the people that were involved in this are actually experts in in this stuff. Even myself, I wouldn't call anywhere near an expert. I'm on the quantum advisory board for Coinbase, but really like I don't really know quantum computing. Each of the areas are very specific and deep. And what I found is like a network of anons collaborating, coordinating on a problem can actually solve something so important because agents have democratized the ability to build. So why am I giving this example?

If a network of anons using agents can actually solve, you know, a big problem in quantum computing, we should all ask ourselves how much are we taking advantage of this intelligence bonanza that exists and accelerating our own product roadmap and and and building. And I think we are trying our best, but I think there is a 10x to 100x improvement that is possible. we already are like you know 10 to 20x better than where we were 6 months back in the in terms of how we build how we ship but I think there's another 10 to 100x to go but um Shirram can we move quick enough because you're talking about AI being able to solve um you know quantum

security issues but AI is cracking into crypto and is already causing crypto issues like with what happened with Zcash right so um are are moving fast enough. Yeah, this is going to be one of the hardest questions for crypto is I think what happens is when you know I think eventually we will be actually much stronger for it because it's kind of like this thing where we say let in some sense Bitcoin was a public bounty for many years before people started trusting Bitcoin right because it had to run in a public permissionless setting for long enough before people know okay you can't attack it there is a similar version that's going to happen

with AI is People are going to try to attack every protocol that exists and uh protocols that can actually withstand this will actually stand the test of time. But how do we actually do this? I think the biggest tension is between open-source. It's not between open-source and security. It is actually between instant settlement and security. When you have a protocol that has to settle instantly, then if there's a hack and you know somebody stole the money, they just ran away with it and then you're just like you know forever lost.

But if protocols can control the settlement lag by saying okay it's going to take one day to settle in this day some agents might review it some other like network will review it whatever the protocol is. I think those kinds of solutions will actually stand this period much better than protocols that optimize for instant settlement. This is one of the big I think uh the next 6 months to one year is going to be very scary for DeFi because protocols that have instant settlement will have a pretty strong attack surface >> and you believe that you know we're moving into an era where AI agents will become their own companies. Um you said earlier that you got into crypto because

AI was already doing your job. What about what about other people's jobs? [laughter] >> Yeah. So this is like you know we were talking about D5 security but of course agents are going to have a massive so societal scale impact. Uh before we deal with this maybe like a quick highlight of the thesis. What is this thesis that agents will become companies? The idea is when you look at a company, what is going on? Somebody puts in money. There's a capital layer. And then there's a governance layer. How do the people govern the company? That's another like governance layer. Then you have the execution layer. Usually comprised of teams that work together.

You have the CEO, you have the you have the uh you know execs and then you have the execution people and all of us together go and execute on something. You have the execution which is usually human and then you have like the company can own various properties you know Facebook owns Facebook.com Mr. beast owns his YouTube channel, you know, uh uh a game might own their website, whatever, right? And when you think about the company, like right now, companies have really nothing much to do with crypto. So, this also relates to one of my core thesis for crypto. Like the place where crypto's biggest value proposition shines is capital, right?

[snorts] the movement of global capital and enabling new kinds of instruments for capital. So when you think about it like this uh right now crypto is only the capital layer for onchain protocols but onchain protocols are very very limited. there is like five D5 protocols worth investing. That's why everybody keeps talking about you know one or two it's the hyperlquid and friends that everybody is talking about because those are the only onchain things which are investable but to actually have a much larger surface area of investable things by crypto capital we should make every digital company investable how do you do this if I can take the digital company's credentials and put them all onchain the

website the YouTube account the stripe accounts, the Roblox games, any one of these can be encoded and owned by smart contracts. Right now, there is no framework for doing this. That's what we're building at IGEN, which enables any one of these companies to be represented and owned onchain. And then the execution layer, which is usually a big thing for crypto, is you have tokens on chain, but people build companies and sometimes companies get acquired and the tokens just free running on the other side. And instead of this, if you had AI agents which [snorts] are fully onchain, AI agents that are fully onchain, which are doing a part of the execution, which

are orchestrating both humans and agents to do execution, then suddenly there is no issue that somebody just ran away or like somebody went somewhere else. The agent is always on chain, always owned by the protocol. They can appoint a new human as a steward. They can appoint other people. So the new thesis answering your question on the role of humans is agents at the center and humans at the edge. It's already you can see this in Jack Darcy's thesis for how he wants to run block his company is agents maintain context and information and and and things at the at the center and people plug in at the edge and you can now remove the entire management layer. But imagine this agents not just

offchain but these agents are living onchain. Now suddenly these companies become fully investable by crypto capital. >> So you feel that AI, blockchain and crypto need to function altogether. Is that a necessity? >> Absolutely. >> Awesome. I think we've ran out of time. I could go on and on and this panel was so interesting. Thank you so much. Hope you guys Thank you. [applause]