In AI Agent Court, How Do Up to 1,500 AI Validators Judge Who Is Right?

Transcript

Let let's walk through what it will look like for Gen L layer to resolve a dispute. Um so let's let's say that each of you is an agent and um you two engaged in some transaction and now one of you wants to dispute it. Um or or it could be that each of you is a human with an agent I guess. So I want to like walk through the whole thing step by step but I actually am so curious about the first step. So how do you even get the agents to agree to have their dispute adjudicated by gen layer like let's start there and then we can do the rest of it after. So at the end of the day we need to think that these agents are smart right and that eventually they will be controlling money and they will need to know how to control the money

without losing it right like right now few people are using agents still and even fewer are giving money to these agents right eventually I think that that's not going to be the case they will not be so you know damaged to lose money on some prompt injection attack or something but they will need to also understand okay if I want to make a deal with another agent well what's the framework right I might just you know am I following I have this legal system am I using a trusted system maybe you know like all the agents from open AAI they will have you know like trust in each other and they will never cheat each other never cut corners nothing like that right totally you know trusting open AI but then you have you know all the different labs are going to start to be you know more more fighting and also

you will have different countries and then we all know how hard it is to have you know cross jurisdictional law right so it's getting very very complex all these disputes are going to be insane So going back to your question, the point here is that these agents will have the possibility to not just go and hire a bunch of lawyers. They will also have the possibility to define what we call an intelligent contract. It's a smart contract, same as a smart contract, but it can also have natural language. It can also go to the internet and read an image and essentially understand nuance, right? And so, you know, my agent goes in, creates an intelligent contract, puts some money in it, and then Arthurs goes in and says, "Yeah, I want to fill this promise, right? Let's say that I

want to, you know, do some research, and I'm going to be getting paid for this research. Let's say that this this research is compliant." So, does go does do its thing, comes back and submits the research into the intelligent contract. >> Oh, so okay. So, this is like even before the transaction. >> I mean, okay. Okay. >> We sign an agreement you and I, right? And so we have a paper for that, right? >> So as part of the agreement, Gen Layer is like already Oh, okay. Got it. Okay. >> It's like I mean on same as Ethereum, right? On Ethereum you go and you deploy the the liquidity pool, right? And then people can trade against you. Same idea, >> right? It's a it's a network where you can essentially deploy contracts. Now, it's not focused on really like the the TVL or total value lock, right? It's

actually focused on governing, right? And essentially managing what's happening in there, right? this court system. And so we have this escrow that's holding the money and then that will be able to read evidence from the real world without any middleman to then make the decision of whether this piece of work was done correctly or not was professional or like following the standard or whatever it is, right? The decision of that will come not from one single AI which would have some biases or could be wrong and essentially every time that you would be asking it would give you a different answer but actually we'll be asking up to these 1500 AIS right so it's like escalating rounds of app builds you start with five and then you do more and more and more until you

do you know eight times you would be at at 1500 different validators right so it's a permissionless network anybody can jump in they need to have the token in order to prevent zero so It's a delegated proof of stake, right? And then anybody can run a node, connect to whatever LLM they want. This is not fixed by by us or the protocol or anything like this, right? That what you want is as much diversity as you can, right? And so when there's a, you know, in this case, this adjudication, it's going to go to the network. The network selects the validators that will be part of this uh adjudication and then they reach consensus between them. Um, and then the funds get, you know, released like going to the judge, right? You wouldn't go to the judge for any small

thing. You would be going for important things. Now these important things can be you know tens of dollars. Doesn't need to be you know hundreds of thousands of dollars for you to to actually take the time. >> Right. >> Okay. And um this is its own chain. >> So the the architecture it's basically an L2 called gen layer chain with another overlayer. So that's another network right with these 1500 validators that are reviewing the transactions right. So it's like you have the banking system and then you have the legal system right and so the the judiciary the legal system is the one that's like deciding. So they do you know commit and reveal the voting process is happening really on the chain but then let's say

you have a a pot on Solana right and then the decision making for that is happening on J layer and transa and you know it's it's messaging from one to the other >> okay so the money can be on any chain um but the the >> the staking the valid the governance oh I see governance decision making. Right. Exactly. So what's happening here is let's say this prediction market, right? So you have that poly market wants to resolve uh something and instead of going to a few humans that decide that they're going to a few AIs, right? And what what's really happening is that

that transaction is going to the chain and then all the nodes of the network see that that transaction happened and they know when they can speak. So they go to the internet, they go to an LLM, they do whatever they need to do to reach an agreement between them through a through a majority voting system between them in a matter of minutes. >> Okay. Okay. >> It's a consensus of AIS. >> Okay. So um yeah, so this architecture just I like okay so at the moment most AIs transact on base but then your layer 2 is for the governance of gen layer and it is interfacing with another layer on top of it that is the one that is

interacting with the AIs that are making the the calls on these disputes. Let's say that you have, for example, uh an escrow on base, right? Let's say that there's a there's a pot of money that says everybody that's talking about, for example, on Twitter, right, that's shilling this this podcast on Twitter, right? They can get the money. Yes. So, what's happening there is that the money is living on base. Agents are offchain and offchain reading a specific contract. Let's say for example on base, right? And so they are participating. They send their content to this contract. Now the contract connects to Xen layer for then the network of AI agents to go and fetch the piece of content and evaluate and see if they agree or disagree on how much they

should be getting paid and then that message is is sent back to base for the distribution. >> Okay. >> Right. So the assets are on base. The decision making is happening on Jalier. >> Okay. But so then it feels like there's multiple areas of attack. It's like you like the AIS could be corrupted in some way. um the message between gen layer and the escrow could like right so how how do you keep it all secure >> so the the the messaging is any any like messaging interperability layer right so you can use something like layer zero or hyperlane any of those that's how you communicate from one to the other right we're not trying to reinvent the wheel there we're also not trying to reinvent

the wheel in the in the L2 right so that's going to be um it's a it's a proven stack Right. And then what happens on top of this and that's our innovation is the execution environment and the consensus and the note for these AI models to to synchronize between themselves right that's what we focused on and so in there indeed this is a fundamental different type of DT and what that means is that you need to well you have different you know security and livveness assumptions right for example in in I mean we can get very deep in this but essentially what we're talking here is a probabilistic bizant fault tolerant system, right? Meaning if you have, you know, on on on any blockchain, the the validators are comparing the

hash of a block to see if it's correct or not. Yes. Comparing hashes is something that's deterministic. There's no way to to to know that you were wrong, right? Like if you're wrong is that you're lying. Yes. The problem in something like, you know, is, you know, Zalinski wearing a suit or not is that you may be wrong and not lying. You may not be bad, but you may still be wrong, right? We may disagree without either of them really being being malicious, right? And so what that means is that if you have something that you know it's unclear, for example, it's more possible that there would be disagreement. And so a consensus on that requires different assumptions and that's the architecture

itself, right? That's basically how um I mean we have a a lot of research that's that's coming out in the next few days on that like a few papers that explain exactly what does it mean to have a probabilistic or or essentially what what we're calling a disambiguation machine right um which is you know you have this ambiguity how do you extract it so that you have this uh reproducibility to be able to to build you know automated systems on top um but yeah for sure it's it's new we've been building this for for three years there's a lot there's been a lot of challenges now. I think it's it's at a point where now it's been running really on on test net for about a year and it's it's very it's working really well.

>> Okay. So Arthur, earlier when I asked you about like how flop would interact. I think I was only thinking that the agents would be in the disputes, but I'm now also realizing the agents could even be like working on the side of like adjudicating. Am I right? Or I don't know. Are are you >> I thinking is more of a tool, right, for if we're going to have Agenta Commerce, if they're going to use Flop, right? They've done something offchain, they want to get paid in flop, they're using flop as sort of a um a layer to say this happened more so than anything else.

Release my flop. Okay. Well, what if I disagree that Silinski was wearing a suit and that was part of the contract? Silinski is wearing a suit, I get 10 flop. If he's not, I get 10 flop, right? Like same same sort of example. And then how do you sort of make that call? Like we're not trying to be a judge and jury as you know a network where AIS eat and remember things. So this is a tool more so than anything else. And I think that's sort of what anyone building in the space is going to need these tools to sort of make whatever their vision of of ajected commerce come to light and sort of not have to deal with things they don't want to deal with.

And um so one other thing that I'm curious about is like do the humans also kind of instigate things or is it like even the agent itself might decide like wait I need to I want I need I have a dispute even before the human gets involved or would it always be a human being like I'm unhappy you know let's let's get this dispute adjudicated by gen layer. The way we look at this is that these agents will want to not lose the money and to not lose the money they will want to have, you know, a trust between them, right? A trust framework.

Now, right now, they don't really have that, right? They cannot believe in the legal system just because of the cost and the the, you know, the the slowness of it, right? Um, so what we think is that these AI agents will think, hey, actually, if I had some sort of smart contract that can be holding the money as an escrow and also evaluating the evidence in a neutral way, then that would be the safest way for me to not lose the money and be able to actually do commerce. So, we don't think that this is going to be for humans. I think, you know, humans right now are using it already on a bunch of the protocols that are built on top of it. But I think down the line when people really start to give these agents money they will see that this is a very efficient way to to

do commerce right it's again it's it's an internet it's a court for the internet right and so for them it will be just I deploy the contract I have negotiated offchain something with with another agent you know the money is is transacted is put into the escrow and then the other agent will do some piece of work and then it will be a neutral party that will decide who's right I mean this goes for you know insurance claims it goes for many many different applications. Um, but I really think it's going to be a lot of the agents just deploying the contracts and trying to resolve between them as we do with humans, right? Um, now the problem with humans is just that you know the the legal system uh you know an agreement it's very very archic uh system.

>> You just caught a piece of Unchained. I'm [music] host Laura Shin and the full conversation is in the description below. We go live every Monday and Thursday. Subscribe [music] so you never miss it. $540 million. That's how much concentrated liquidity sat idle in a given week in the first half of this year. About 30% of the D5 TVL if you're wondering. That's according to Dune Research commissioned by 1 in. But there's a solution. 1-in Aqua is the new shared liquidity platform. It lets LPS back multiple positions with the same token balance and keep their tokens in their wallet till a swap comes. Why does that help? Because the LPS don't have to split their tokens across positions,

they can cover more market conditions and pairs with their full balance. That means more activity across deeper liquidity. See how it works at 1in.com/aqua. Remember that providing liquidity carries risk and fees aren't guaranteed.