Building Synthetic Trust for the Onchain AI Era | Albert Castellana - GenLayer

All right, thanks everyone for being here. Um, my name is Abra Castayana. I'm the co-founder and CEO at at Zen layer. I've been in this industry for about uh 13 years more or less on different L1s, different D5 protocols. Um, and about three years ago, I started looking into LLMs.

I started to build a bunch of agentic frameworks and also got into some pretty heavy litigation. Right? So what I'm going to be talking about today is a bit of a different thing that what you guys are used to. Right? We've been in this industry for for a while.

Right? Uh we have a lot of solutions on how to make agents be transacting with other agents, right? We've been working for over a decade to make sure that everything is cheap, everything is safe. We have a lot of different solutions for that, right? And I want to propose a question that I think that many of you haven't heard before and like why are we not hearing that before and actually that's basically our solution to this question right the thing here is we live in an amazing moment there's basically you know we've been able to turn sand into thought that's Mark Andre saying this right I think it's absolutely un like unbelievable what's happening right now what's going to happen as well is that all these agents as you know my colleagues just explained before are going to be taking over going to start to be transacting with each other and that's Well, very nice, right?

It's going to be accelerating a lot of things, but I think we need to remember this. These guys are pretty smart, right? So, they're going to be starting to save on cost, cutting corners, lying, cheating. And so, the question is, if two agents disagree, how do they trust each other? How do they have a judge?

Do they have a decision maker? How do we make it so that we can have actual real contracts that, you know, can be agreed among two different parties without having to rely on, you know, the typical mechanisms that we've been using so far, right? Because the problem right now is that the legal system is already broken. There's like five billion people in the world that don't have access to that. It's very expensive, very slow, it's quite biased.

It's very much uncomfortable if you have ever been in that situation, right? So, do we think that they're going to be using lawyers? How will that work? That's a problem, right? Maybe smart contracts and fix this, right?

That's why we guys are here, right? We all think that blockchain, you know, have these smart contracts. They're 247 self- enforcing, very cheap. That's very nice. The problem of that is that smart contracts are quite limited in what they can do, right?

What you see is that smart contracts if you try to create one smart contract that is trying to understand you know is this thing that has been de delivered professional quality probably will not be able to do that right if you say has this thing you know that has been delivered for example is it legal or not this ambiguity that normally you have in any type of real world contract it's something that smart contracts cannot do at all so the problem here is that Ethereum and all other blockchains require a deterministic setup. Every validator needs to see exactly the same thing. That's why you have that, you know, all verification and crypto pretty much goes from comparing hashes, right? And that basically means that anything that is nondeterministic, well, that kind of like falls through the cracks, right? Like we're trying to see that, you know, everything is deterministic, meaning that given an input and an output, the same input will give you always the same output.

But the problem is that these AIs are not like that. These AIs are probabilistic. They're stoastic. They give you different responses every time that you ask them, right? And also, if you start to think about the type of contracts that they will need to be engaging on, many of those contracts are actually quite, as I said before, ambiguous, right?

Um, say, for example, something like high quality. Can it be that the one that is servicing something thinks that what he delivered it delivered is high quality and that the customer thinks is not high quality? happens all all the time in the real world, right? How can we imagine a world in the future with all these billions of agents if they cannot have a disputed resolution system that essentially works at this type of machine speed? Because in reality, most of the contracts in the planet work like that are ambiguous or not clear, right?

And so the thing here is again if they disagree who decides, right? Are we going to go to the legal system is expensive? Are we going to go to smart contracts? They have some limitations. Are we going to just, you know, trust them?

Are we going to have just reputation system? That's something that's actually happening quite a lot, right? Are we going to go back to just trusting in, you know, counterparties, platforms, right? I mean, I think that if you guys are here for the most part, you would like to see that, you know, that we have decentralization to to some degree, right? And the problem here is that this is not getting fixed.

It's actually getting worse because we're just trying to, you know, like pretty much what you see is like most of the things are working right now are kind of like leaning towards centralization. Why? because the primitives that you can do in crypto if you're just playing with deterministic tools are quite limited. So the question here is we live in this AIH right execution is cheap writing code is cheap everything is cheap what is expensive what's the bottleneck the bottleneck is actually making decisions is a judge is actually how do you have a judgment right how do you have common sense how do you trust without having to trust a single entity and to my knowledge basically nobody's working at this except for us so what we're proposing Today is this new type of consensus. Essentially a consensus over interpretation, right?

This what Zelier is is a trustless decision making protocol. You can think about it like a digital judge. Right? You have Bitcoin giving you trustless money. Ethereum giving you trustless applications.

Generally gives you trustless decision making. again a digital judge in a sort of synthetic jurisdiction that works 247 self- enforcing like a smart contract but also can understand natural language also can go to the internet fed some information and make a nuanced decision right and so let me tell you a bit how it works it's not too I'm not going to go super deep um but basically the stack works like this there's this generator chain that's like the banking system and on top of this generator chain there's this consensus algorithm I will talk about it in the next slide right? Where basically, you know, all the different validators in the network, that's up to 10,000 validators are going to be selected at random to do something. I'm going to talk about this in just a second, right? Each one of these validators, this is a permissionless network.

You all can join whenever you want. I mean, it's not live yet, but very soon, right? You can join the test net if you want. Um, all of these validators are going to be running this thing called the JamVM, which is an execution environment that can do these nondeterministic operations, right? So what I'm proposing here is essentially a new type of DT that is capable of non-deterministic operations.

Something that basically everybody else can't do that because they're trying to do something in a deterministic fashion. Right? We'll see some use cases in just a second. You will I think understand it more. But essentially what's happening is that the GMV this execution environment made in WAM can perform nondeterministic operations such as go to the internet and fetch some you know data an image or even like go to an LLM and essentially make a decision and that's when the consensus kicks in right because again the problem is that you need all this network to have a mechanism to reach an agreement a consensus right and if everything is kind of like probabilistic if everything is different if you don't have like onetoone comparison of How do you make them agree on something?

That's the tricky bit, right? Because we're talking about ambiguous clauses in a smart contract. How do you do that? Right? That is subjective.

And so the consensus was like this. There's a thousand validators. Transaction comes in from those validators. Five of them are selected at random. One is the leader goes do something.

This something is essentially an arbitrary function. Can be in Python, can be in natural language, can go to the internet, anything. and it needs to make a proposal to the other guys. The other guys vote on that proposal in a majority vote. It's a simple majority vote.

And if they disagree, they rotate the leader. If they agree, they basically the transaction is accepted with a 30 minute finality window in which anyone can come in, appeal, and double the value that is set. That's basically it, right? What I'll kind of like repeat that what's happening is that you have a thousand validators out of which only five are selected at random at the beginning. It's a DPOS network.

they one of them makes a proposal the other vote on that and if they still disagree you can just escalate escalate seven times until you basically have the entire validator set to be able to make a decision and why do we do this right because like if you're thinking about how to make decisions on chain I'm absolutely sure that you have heard about different approaches on how to have a single AI model making the decision let's say proof of inference right there's a bunch of different approaches to this trying to make a process that's naturally nondeterministic deterministic Right? How do we make that? Well, the problem is that if you have a single model that by definition is a centralized model. It doesn't matter how you train the model. It doesn't matter how you run the model.

It doesn't matter if you know exactly how it was run because it's a single entity, a single quot, right? And so what we're doing is we're aggregating the minds of up to a thousand different AIs to reach a consensus together. Different points of view basically mean that you have higher correctness, you have less bias, you have higher security. It essentially you can trust it. Right?

If you guys want to read about this is called the condored theorem, right? And it basically means that you know depending on the type of question if you get different judges that actually come together to a consensus on a majority vote. Well, the the correctness of the system is higher. And like this we can actually build defi on top and we can build new types of primitives because you cannot have defi on something that works 99.9% of the time.

You need something that works 100% of the time. And the way we're doing it is again if you have this thing that cost like you know 95% of the time if you ask 50 different validators it will be essentially 100. Doesn't make sense to you to basically go beyond that because it's essentially 100. Right? So it's a probabilistic consensus.

You guys are used to having deterministic consensus. This is a probabilistic consensus. It's a dynamic consensus. What does that mean? Okay, the problem is that because all of this is probabilistic.

You need them to have a way to agree. And the way that they agree is essentially because the developer can define the threshold of acceptance of a specific thing. For example, you know, is this tweet funny? I see I don't have a lot of time. So, is this tweet funny?

If they agree, that's the developer said, you need to check a tweet. Is it funny? If you guys agree in a majority vote, that's good to go. This is called the currency principle and this enables basically different type of primitives trustless decision making. It enables different types of primitives.

I will show you a bunch of them but it's essentially a judge as I said before right you can have a marketing protocol where basically anybody can permissionlessly on any blockchain come in put in a pool some funds and tell the pool that if somebody creates content on your product it can take the money. Now the judge is essentially looking at the tweet and seeing that the content is correct. And what you see here is you know it's basically doing that right about 20 different things. This is live today rally that fun you can go and visit. There's basically 125,000 people right now.

Um like in the weight list there's like you know thousands of of decisions being made every single day. Um just know for the sake of time there's also argued fun. It's a debate prediction market. Basically you're betting on what the AI will decide after the whole consensus goes through. It's a completely new type of prediction market, right?

You just have two ends like two parties can be debating something, betting on that something and at the period at the end of the period the AI decides who won. You can have a dispute resolution system. So this is an escrow. You can deploy this on base on whatever you want really because gener is not trying to compete with other blockchains. It's trying to ex expand what they can do right through layer zero any type of inter interperability protocol.

You're going to be able to call gener to make a trustless decision for a very very low cost like let's say rally cost like about a dollar or something like this per decision. This is holding the escrow deciding who to release it to and it automatically releases the the escrow. And then also we have another one called merge proof. This is coming soon which is essentially u you know a way to manage um you know like PRs and issues in an automated fashion in a way that you have you know economics behind. So you you have an issue, you put some bounty, anybody in the planet can come in and try to take that bounty by doing a PR and if it gets um you know like debug it then can lose this money.

Anyway, I'm out of time. Uh thank you all so much for the attention. If you want to reach out um I'm you can find me here in the socials. Uh thank you so much. Yes, thank you.