Blockcast

GenLayer proposes AI-adjudicated 'intelligent contracts' to resolve subjective disputes for autonomous agents

32 min Listen to the original
AI AgentsSmart ContractsBlockchain GovernanceDecentralized JusticeGenLayer

Summary

Albert Castellana, CEO of GenLayer Labs, argues that traditional smart contracts cannot handle subjective outcomes, necessitating a new system for AI agent disputes. GenLayer's 'intelligent contracts' use 1,500 validators tied to different AI models to adjudicate ambiguous terms like content quality or legal compliance. The protocol employs an appeals process that doubles the validator set to reach consensus based on the Condorcet jury theorem, aiming for consistency rather than objective truth. Castellana positions this as a 'small claims court' for the agentic economy, handling low-value disputes currently ignored by traditional legal systems. He predicts AI agents will eventually operate seamlessly in the background, with liability remaining a uniquely human domain.

Key takeaways

  1. 1
    AI agents are gaining financial agency, creating an immediate need for dispute resolution mechanisms beyond simple payments.

    Unidentified speaker (speaker_1) · 1:09

  2. 2
    Liability and legal accountability remain uniquely human differentiators as AI cannot experience imprisonment or death.

    Unidentified speaker (speaker_1) · 30:42

Episode sections

  1. 0:06 – 1:47

    Introduction and the Problem of Agent Disputes

    Takatoshi Shibayama introduces guest Albert Castellana, CEO of GenLayer Labs, which builds the application layer for the agentic economy. Castellana argues that while AI agents are gaining payment capabilities, there is no efficient way to settle disputes involving subjective outcomes, unlike deterministic smart contracts.

  2. 1:47 – 7:43

    GenLayer's Intelligent Contracts vs. Traditional Smart Contracts

    Castellana explains that traditional smart contracts on networks like Ethereum or Solana cannot handle ambiguity or subjective criteria (e.g., whether content is 'funny' or 'legal'). GenLayer proposes 'intelligent contracts' that use AI to adjudicate these ambiguous terms on-chain.

  3. 7:43 – 11:47

    Use Cases: Insurance, Marketing, and Agent-to-Agent Commerce

    Potential applications include insurance for any verifiable event, performance-based contracts (e.g., paying for content that meets specific criteria), and automated marketing protocols where agents pay other agents for tasks like tweeting, bypassing the need for human review.

  4. 11:47 – 12:46

    Network Architecture and Validator Selection

    The GenLayer network consists of 1,500 validator seats, each connected to a different AI model to ensure neutrality. For each dispute, five validators are randomly selected based on stake to vote on the outcome. If they disagree, the leader is changed.

  5. 12:46 – 22:23

    Appeals Process and Consistency vs. Truth

    An appeal window allows users to escalate disputes by doubling the validator set (5, 11, 23... up to 1,500). Castellana states the system aims for consistency and reproducibility rather than objective truth, relying on the Condorcet jury theorem to reach a majority consensus.

  6. 22:23 – 28:52

    The 'Petri Dish' Analogy and Future Vision

    Castellana compares the current state of crypto innovation to a 'Petri dish' where new primitives are tested. He predicts that in five years, AI agents will operate seamlessly in the background, handling daily tasks and negotiations, with liability remaining largely with humans.

  7. 28:52 – 30:02

    Future of AI Agent Interfaces and Human Agency

    Takatoshi Shibayama and Albert Castellana discuss the transition to agent-driven interfaces where traditional software front ends disappear. Castellana predicts a seamless, one-tap interaction model where agents handle authentication and data aggregation, contrasting it with the current fragmented app ecosystem.

  8. 30:02 – 31:30

    Liability and Decision-Making as Human Differentiators

    Castellana argues that while AI may replicate creativity, liability and taste will remain uniquely human domains due to mortality and legal consequences. He suggests that retaining decision-making power over AI model usage is a critical frontier for human agency in the coming decades.

Assets discussed

AssetKindStanceContext
EthereumETHcryptoneutralEthereum is also a digital court system, with the difference then it just can check, you know, like hashes, right? Like whether the math is the same.
SolanaSOLcryptoneutralit's impossible basically for a smart contract uh on Ethereum or Solana or or Bay. base to actually say whether something is funny because it can only deal with code, right?
BitcoinBTCcryptoneutraltrustless money like on Bitcoin or trustless applications like on Ethereum
UniswapUNIcryptoneutralAnd so it's tricky because you know what is Uniswap? Well, Uniswap is kind of bypassing the system as well, right?

“Stance” is how the speakers talked about the asset, not a recommendation.

Claims worth verifying

GenLayer has 1,500 validator seats, each connected to a different AI model.

Unidentified speaker (speaker_1) ·2:25

How to verify: Check GenLayer Labs documentation or whitepaper for network architecture details; verify validator count in technical specs. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.

Smart contracts on Ethereum, Solana, or Base cannot determine subjective qualities like whether something is 'funny' or 'legal'.

Unidentified speaker (speaker_1) ·3:30

How to verify: Review smart contract code examples on these chains; verify that standard EVM/SVM logic cannot parse natural language semantics. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.

The appeals process doubles the validator set size with each appeal (5, 11, 23... up to 1,500).

Unidentified speaker (UNKNOWN) ·13:01

How to verify: Check GenLayer technical documentation for the dispute resolution algorithm and escalation logic. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.

The system is based on the Condorcet jury theorem to ensure reproducible results.

Unidentified speaker (speaker_1) ·14:28

How to verify: Review GenLayer whitepaper or technical blog posts citing the Condorcet jury theorem as the mathematical basis for consensus. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.

Uniswap bypasses traditional centralized financial systems by allowing anyone to deposit funds without KYC/AML.

Unidentified speaker (speaker_1) ·20:52

How to verify: Review Uniswap documentation and terms of service; verify permissionless access to liquidity pools. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.

Primary source

Crypto currently represents approximately 0.5% of total monetary volume.

Unidentified speaker (speaker_1) ·26:35

How to verify: Compare total crypto market capitalization against global M2 money supply data from the Federal Reserve or IMF for the relevant period. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.

GPT-6 is currently cheaper to use than GPT-3.

Unidentified speaker (speaker_1) ·27:57

How to verify: Check OpenAI's pricing page for GPT-6 and GPT-3 models to compare cost per token. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.

Five billion people do not have access to the legal system.

Unidentified speaker (speaker_1) ·18:14

How to verify: Search World Bank or UN reports on 'access to justice' and global legal system coverage statistics. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.

Human judges exhibit biases, such as being more lenient after a meal.

Unidentified speaker (speaker_1) ·15:48

How to verify: Search academic literature for studies on judicial decision-making biases, specifically referencing the 'hungry judge' study by Danziger et al. (2026). For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.

GenLayer is building a 'marketing protocol' where agents pay for tweets based on guidelines.

Unidentified speaker (speaker_1) ·19:21

How to verify: Check GenLayer Labs website or GitHub for active marketing protocol implementations or case studies. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.

AI agents will replace traditional software front ends by plugging directly into data sources.

Takatoshi Shibayama ·28:52

How to verify: Monitor industry trends in AI agent frameworks (e.g., LangChain, AutoGen) for shifts from UI-based to API/agent-based interactions. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.

Current digital experiences require users to manage multiple passwords and apps for basic tasks.

Unidentified speaker (speaker_1) ·29:14

How to verify: Survey users on the number of apps and passwords required for common tasks like school communications or banking. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.

AI agents will autonomously log into services to aggregate information, eliminating manual password entry.

Unidentified speaker (speaker_1) ·29:22

How to verify: Review technical documentation for AI agent frameworks that support OAuth or credential management for automated service access. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.

AI agents will become as ubiquitous and invisible as electricity.

Unidentified speaker (speaker_1) ·29:42

How to verify: Track adoption metrics of AI agents in consumer applications over the next 5-10 years. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.

Centralized labs may control which applications can use specific AI models.

Unidentified speaker (speaker_1) ·30:02

How to verify: Review terms of service and API usage policies of major AI model providers (e.g., OpenAI, Anthropic) for restrictions on downstream applications. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.

Liability and taste will remain uniquely human traits because AI cannot die or go to jail.

Unidentified speaker (speaker_1) ·30:42

How to verify: Analyze legal frameworks and philosophical debates on AI personhood and liability. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.

Do your own research

  • Verify GenLayer Labs' current validator count and AI model partnerships via their official documentation or GitHub.
  • Review the technical whitepaper for the specific implementation of the Condorcet jury theorem and validator selection algorithm.
  • Check for any existing pilot programs or live deployments of 'intelligent contracts' on mainnet.
  • Research the legal enforceability of AI-adjudicated contracts in key jurisdictions like the US or EU.
  • Investigate the tokenomics and staking requirements for validators to assess centralization risks.
  • Look for independent security audits of the GenLayer protocol, particularly regarding AI model manipulation or oracle attacks.
  • Compare GenLayer's approach to other decentralized oracle networks like Chainlink or API3.
  • Assess the competitive landscape for AI-agent infrastructure, including competitors like Fetch.ai or SingularityNET.
Full transcriptexpand

0:06Unidentified speaker (UNKNOWN): Welcome

0:07Takatoshi Shibayama: to License the Shield, the podcast by Blockhead, where we discuss what's happening in the ever-changing landscape of crypto, the markets, and anything else that comes to our minds. Before we start, everything we say here is not investment advice, business advice, relationship or friendship advice. I'm Takatoshi Shibayama, the host of this podcast, and also the head of APAC APAT for Ledger. And today I am doing this solo, no co-host today. I'm joined by Mr. Albert Castellano, co-founder and CEO of Gen Layer Labs. Welcome to the show.

0:36Unidentified speaker (UNKNOWN): Hey

0:36Unidentified speaker (speaker_1): Daka, how's it going? Thanks for having me. Exciting.

0:39Takatoshi Shibayama: Yeah, absolutely. And Gen Layer is building the application layer for the agent economy. So can you elaborate that a little bit and then I will kind of discuss uh what you guys are building? Yeah,

0:49Unidentified speaker (speaker_1): definitely. So everybody's talking about how do agents do payments and commerce, right? Agentic commerce is one of the biggest topics I think this year. And I think it's gonna be actually not just this year, it's gonna be going forward, to be honest, right? Little by little we're giving these AIs more agenc capabilities, right? We're giving them money, we will be giving them like passwords, and we're gonna be like access to basically everything. And the problem is that while everybody's focused on you know how do they pay each other, the problem is that things in the real world don't always go well, right? Like sometimes things go south. And then the question that we're proposing is what happens when agents disagree? Right? What happens when you have two agents that are trying to enter an agreement, but they disagree, right? Do they go to the legal system? Very expensive, very slow. Do they use a smart contract? It's great, very fast, It's very cheap, enforceable, and so on. The problem with smart contracts is that they are quite limited in what they can do. So we need an needed a new type of contract, we call them intelligent contracts that will allow them to be able to have adjudication, even if the contract has some ambiguity in it.

1:47Takatoshi Shibayama: So it's just kind of like setting the scene here. So if I picture like an AAI agent that books and pays for my for services on my behalf and the work comes back, you know, disappointing or like just utterly, completely wrong. And what happens today and where does gen layer step in?

2:02Unidentified speaker (speaker_1): Right. So you can think about Gen Layer as a digital court system,

2:05Unidentified speaker (UNKNOWN): right?

2:05Unidentified speaker (speaker_1): Ethereum is also a digital court system, with the difference then it just can check, you know, like hashes, right? Like whether the math is the same. Like two plus two is two. So it can just deterministic code that can check deterministic results to see that it's the same, right? The problem is that AI is a stochastic system, right? It's a probabilistic system. The world is also a probabilistic system. And so how do you build a system for smart contracts that can deal with ambiguity, for example, right? So again, think about it as an AI court system. So you have 1500 validators, each one of them is connected to a different AI model, and they reach consensus, not just comparing the hashes, but actually voting on whether the contract was resolved or not, right? So let's say we have a you know a performance visa vis contract. They say, hey, if you read an article and it's great. and it gets this much traction and it's correct and legal, for example, then you get the money, right? I deploy the pool, let's say a smart contract on on Ethereum, right? And then I set the rules on the contract itself. And then if you go and then I put the money in, you go ahead, you write some content, you send the URL of the content to the contract, and then this network of AIs are gonna going to be downloading the content and deciding if you actually met the criteria to get paid.

3:16Unidentified speaker (UNKNOWN): Right.

3:17Unidentified speaker (speaker_1): And

3:17Takatoshi Shibayama: let me just put a pin on what you just said. You said most like real world contracts are ambiguous. So why can't like regular smart contracts handle these things? And then what makes an intelligent contract that you mentioned different?

3:30Unidentified speaker (speaker_1): Right. Let's say for example the word legal or funny. Let's say funny. Yes? It's very it's impossible basically for a smart contract uh on Ethereum or Solana or or Bay. base to actually say whether something is funny because it can only deal with code, right? So it can run their rules, but anything that is ambiguous, that is subjective, it they cannot do. Right? So smart contracts cannot deal with any type of ambiguity because they need to agree. You have, you know, thousands and thousands of nodes in a network that they're comparing their outputs to see if they agree or not on what the other one said, right? It's not the same if you have a deterministic system, right? That if anybody is disagreeing with you is because they're cheating, they're malicious. Then if you have a system that is ambiguous, that is probabilistic, right? So a probabilistic BFD, right? If you have a situation where you don't know if somebody is lying or is wrong. If I tell you, let's say, you know, is Zelensky wearing a suit, right? You look at the image. and you say, okay, you think yes, maybe I think no. I don't know. What's if you ask everybody, all the population in the planet, what would they say? The reality is that they would say, at each one of them, they would maybe even change their opinion w between one day and the other, right? So how can we have a reproducible and neutral adjudication layer? What we're doing right now is we're going, when we have any ambiguity, we're going to the legal system. We ask

4:49a judge and we say, hey. judge, can you please take a look if this, you know, is this a suit or not? Right? And that's just very, very, very expensive. So what we did is basically we created this network of validators, right? It's a permissionless network. But instead of just comparing the highest of blocks, they're voting after connecting to an AI to see if they can agree. So that means, for example, is Donald Trump and Donald J Trump the same thing or Trump? It's not the same thing or is it? Right? Depends. It's the same person. It's not the same word, obviously, right? Trump and Donald J. Trump is not the same word. Now, how do they agree? Well, you need to be able to understand this, right? You need to have common sense. And you cannot have common sense like programmed into software if you're not using an LLM. This is the superpower that they bring, right? They bring common sense and then you can make this type of judgment, right? So now you have that literally on chain without relying on a single decision maker. Like we're going to this world of AI making decisions for us. AI is making decisions of what you see on YouTube, what insurance you get paid or not, on whether you get hired or not, right? We're just trying to take back some of that, right? Decision making should not just be constant. concentrated power on the hands of just, you know, the big labs, for example. Right? How can we make it so that you don't have just

6:09trustless money like on Bitcoin or trustless applications like on Ethereum, what generally is it trustless decision making, right? Having to not have any single entity make decisions on your behalf.

6:19Takatoshi Shibayama: Yeah. And you mentioned about like probabilistic and deterministic. So like for me, like if I think about um immediate use cases where AI agents can be potentially used today, I'll think of like trading, payments, very deterministic stuff, right? So if I go and give them criteria like what to invest in, what when to sell, when to buy, or if I want them to to shop for things or whatever it is, I mean it's pretty deterministic stuff. But I'm trying to think of like where we're AI agents and maybe I'm lacking in creativity in my mind here, but if I want them to do something that's a little bit more nuanced, what kind of like examples can we use? I mean, you you mentioned about suits, yes. Okay, so well if I give them like clear things like I want to buy a blue suit by so and so brand, or or I want to buy a gray suit that's this measurement and this and that, these are very deterministic in a way, right? But if I tell it I want a cool suit or I want wanna I wanna look like Brad Pitt and buy me something that makes me look like Brad Pitt, you know, those those are very probabilistic. So are you talking about those those kind of cases?

7:24Unidentified speaker (speaker_1): Yeah, maybe uh that's uh you know, like uh at the end of the day the use case needs to make like economic sense and like do you really need a decentralized judge to be able to make that decision for you or you would just send, Hey, is this, you know, okay or not to your Chat GD, GPT, right? But if it's an agent to an agent in commerce or even like with humans, it's the same. You can think about, for example, insurance, right? So here you can basically have insurance on anything that the network has access to. Right? So that's the type of example. Or what I said before about the performance-based contract, right? You can have the contract, hold the money, and then the money just be released when something happens. In this case, what you're saying is that the the cool the suit is is cool. So okay, you get the suit. then you take a photo of the suit, you send it to the contract, the contract goes to these AIs and the AIs decide, right? Now, it's extremely subjective. Would you really do that? Like, do you get the other party, right? My take is this. We're now in the moment where these AIs are still not really controlling that many assets. We're gonna be giving them more and more and more agency, right? So one thing is what happens now, this is useful for prediction. markets, options, insurance. You can have DAOs, you can have many different types of applications. You can have an AMN that changes the price curve based on what's happening on centralized exchanges, right? You can have, you know, reality mapping that's paying people to actually do that. We have this actually going right now. This

8:43is today. In the future, what I think will happen is that you will have your agent, you'll give your agent $10,000 ten thousand dollars to do something. You will just say, hey, agent, here's ten thousand dollars, make them a hundred thousand dollars. And it will need to decide to do something with them. And if it is trading is one thing, but if he's actually going to find, for example, some other reporter that wants to spread your word, for example, then this performance based contract comes in, right? What will happen is my agent will deploy the contract, will put the money there and say, Hey Taca, if you write this following these guidelines, you can take this money out. That's agent to agent commerce. You will not even see it. Like your agent will be trained with your data, right? And your network. And he will just say, Hey, I'm paying for your network, I'm paying for your writing style, and like whatever you have, your expertise already embedded in your agent, probably. And that's how you I start to have agent to agent. My agent with the money just decides, okay, I'm gonna be fetching what are the best reporters in the planet, right? And like what is the price. their pricing? How do I do that? Well, I need to get that list. How do I do that? Maybe I can ask somebody to fetch for me that, do that research, and I can pay for that research, agent to agent commerce. Yes? So it starts to open up the moment that you start to give them money, but we're just not there yet. We need a year, two years, I don't know. Fifty years down the line, I think that But that's gonna be like it will be inthinkable unthinkable that we're still going with like legal contracts as so slowly. It just doesn't make any sense.

10:03It's just too too cumbersome, right?

10:05Takatoshi Shibayama: Yeah. Well I remember in the times of like the when in Amazon and all these uh platforms start coming out and you know, people had to pay with their credit cards and they were very worried about, yeah, giving the credit card details to like an unknown, unreputable platform like Amazon that you never heard of before. But now twenty six years later, I mean everybody's credit card details is all over the place. So I think it's just a matter of time when people get a bit more comfortable with these agents and they know that actually they could do the stuff that you're commanding it to, then I think we're gonna all start trusting these agents to get a whole lot of money. But

10:38Unidentified speaker (speaker_1): the user experience is just great. Right. You're just chatting with natural language, it already knows the password, doesn't need to be clicking on cookies and stuff like that, right? So it's just like it's insane how easy it is to just, hey, here's my wallet, just do whatever you want. Right now they're kinda dumb, so they still they will easily m lose money, right? They will get tricked by others, right? And and that's the problem. How do you prevent them from getting tricked if they don't really know how to identify itself and so on and so forth, right? Each other? Well, here's a way a smart contract, right? I put the money and then the smart contract will be judged if you actually did the thing or not. And anything that can be proved on chain, or everything that can be proved off chain that through a website, for example, then that's gonna be from digital identity to like it just opens up many, many, many new applications that we're not used to. So like you you've only seen blockchain with deterministic contracts.

11:25Unidentified speaker (UNKNOWN): You

11:25Unidentified speaker (speaker_1): have never seen what blockchain is with non deterministic contracts. And there's good things and bad things about this, right? But yeah, we can explore.

11:33Takatoshi Shibayama: Yeah, absolutely. So walk us through like what happens inside the network when the dispute is raised. You mentioned about validators, but what are these validators? How do they work? The proposal, the votes, the appeal, how do the all these things kind of fit together?

11:47Unidentified speaker (speaker_1): Yeah. So again, it's it's like a an AI court system, right? So there's a network, there's fifteen hundred validator like seats, right? And then anybody can run a validator. Right? You can run it, it's just some can you can tell your agent to the deploy price it and so on. But there's a condition is that you need to be connected to an AI. So it doesn't tell you which is the model that you need to connect to. If we were going to a single model, that would be a centralized model, right? I don't want to have just the same model repeat everything, because then it will have only the same kind of biases, basically, right? It will only be like that point of view, right? We're trying to make a neutral system. Yes? And so you have these Of fifteen 1500 hundred validators, each connect to a different AI model, and then when a transaction comes in, from those 1500, based on the stake, right, five of them are selected at random. One is the leader, goes to the internet, goes to an LLM, whatever, makes a proposal, right? Hey, I think that this is a suit. And then the other ones look at it and they say, is it a suit or not? They go to the internet, they go to an LLM, and then they vote. And they say, yes, we think in a majority is a suit. shoot, Then that would be accepted. If They don't, they change the leader. If they disagree with the leader, they change the leader. But if it's accepted, then you have an appeal window. You have 30 minutes to appeal. And so every time you appeal, you're doubling the validator set.

13:01Unidentified speaker (UNKNOWN): And

13:01Unidentified speaker (speaker_1): every time that you do that, then well basically you're getting like 2M plus one. So that's like five, eleven, twenty-three, and then you can appeal a bunch of times until you get to the fifteen 1500, hundred, right? And that's like okay, this cost you a $100, hundred dollars, costs cost you like three hours. In three hours you have finality after asking 1,500 different AIs by different people and so on, the result on whether they think that this is a suit or not, for example. Right? Let's say maybe something more kind of, let's say that you're getting a box, right? A shipment, and you say it needs to be reaching me in good state.

13:30Unidentified speaker (UNKNOWN): So

13:30Unidentified speaker (speaker_1): what is good state? Is one scratch good state? Is ten 10 scratches good state, right? How do you make that decision? Well, you send the photo and then the network. decides.

13:40Takatoshi Shibayama: So the five AI models have to agree. And does that make the decision like correct or is it just consistent? Because you know Yeah.

13:51Unidentified speaker (speaker_1): Is it a suit or not, right? It depends on how well defined the question is. If I tell you is it a suit or if I tell you is it a suit where a suit has a tie? The second one is very clear. Then it's definitely not, right? They will agree like 95% of the time because they see that there's no tie. But the first one is more ambiguous, and then I don't know, maybe it's 60%. The question is, you need something that's reproducible and neutral, right? So I need to be able to ask the network and try to get the same result as much as I can. How do I do that? Well, getting more samples. This is based on this thing called the Condorcetry theory. theorem. Basically, like anything that's more than fifty percent, if you get sufficient samples, well, will turn into a a hundred percent. You don't need, let's say, if you ask if you have something that they agree in seven in 75% in honest uh validators, right? If you ask forty-seven of them, that's it doesn't make sense to go to a million. You will get the same answer with forty-seven, 47, then a million, basically, based, right? So that that's the idea here. Escalating rounds runs of appeals to get something that's reproducible and clear, even if it's not the truth.

14:54Takatoshi Shibayama: But these well, I mean, maybe I'm looking at a snapshot of today of these LLMs. They are kind of like some are like more biased than others, right? I mean, it's sometimes it could be using flawed data because it's uh basically, you know, AI is basically just scavenging the internet and there's like always garbage out kind of issues as well. So maybe today is not the right. snapshot to be using because maybe in the future it's a bit more corrected. But because they're scavenging information where a lot of it could be garbage or mixed with a a lot of things, can you actually rely on it completely?

15:28Unidentified speaker (speaker_1): So let's see that, right? I think that humans are not perfect, really, right? And so like at the end of the day, you go to a judge and judge also has biases, right? Has towards watched the same movie. movies or has specific or got a divorce like last week. Yeah? So like there's a lot of things and it's proven that human judges also have biases, right? You get more lenient settlements, for example, if it's after a glance, for example, right? Like the decisions. And so the point here is not really to get the best answer, it's to get something that is fair, neutral, as good as you can, basically, right? And so when you don't ask one AI, but you ask many different AIs. right? Or even the same, one person, no, many people, then what happens is that these errors get reduced. Maybe one is wrong, the other one is not wrong. Well, are they in the majority wrong? If the majority of these 1500 AIs are wrong, then you will get a wrong answer. That's the point. If they're malicious, that can also be. That's the quirk here, right? This is an adversarial system, right? That's the thing here. You used You're to have having deterministic Byzantine fault tolerance systems. This is a probabilistic one. So it's a fundamentally different type of thing, right? Uh you need to have, for example, simple majority instead of supermajority if you want to have liveness, right? Which is different than what you have on a on a normal uh BFT. Anyway, um, the point here is you're just trying to create a system that you

16:48can have neutral adjudication, and then you use this thing. on subjective things, right? And then you use that to create new type of primitives and contracts, and then just put that on chain so that it's like self-enforcing, right?

17:00Takatoshi Shibayama: Makes sense. I mean, I guess the point here is that you know, all this agentic stuff that we do, this gonna create more and more disputes, which we probably don't see as much in today's world, where not obviously a yeah, you're not there yet, right?

17:14Unidentified speaker (speaker_1): Have you ever been in an illegal situation? I mean, most people have never been in a legal situation, right? Why? Well, because you didn't enter in a con today you went to grab a coffee. You didn't enter a contract with the waiter that he didn't put poison in your coffee, right? You're trusting the waiter. You see the face, you see where they're working. Yes, there's trust. And human trust is one thing. It's this technology we've invented to be able to navigate this very, very, very complex. world we're in. You're gonna think about all the things that can happen. And so we have the legal system to try to allow us to, even when that fails, to actually have, okay, we have now the trust system that's more grounded on things that we can control, right? Yes. So, like, okay, he will go to jail. Yes, that's what's making you comfortable going and actually buying a cup of coffee. The problem is what happens when these agents that you don't see, they're all over the place. They're working 24-7, working like a million times faster than any of us. What happens when that's happening? Yes? Like right now, there's five billion people that don't have access to the legal system. That if they were in any type of political situation, they would be losing simply because they don't have the resources to do

18:21Unidentified speaker (UNKNOWN): it.

18:21Unidentified speaker (speaker_1): If I scam come from you $1,000, a thousand dollars, I mean you have the police, this is one thing, right? Let's say any type of contract, if it's whatever, a few thousand of dollars, are you really gonna go to the lawyer and then chase me for years potentially and spend hundred times more to try to get those $2,000 back? Probably not. There's many, many claims, trillions of dollars that are uneconomical to pursue. And so you can think about Gen Layer not as a substitution of the legal system, what it is is a essentially, yeah, a small court that works globally, where you can just have, okay, you know, let's just put this on an ESCO, it will be able to resolve itself, whatever it is, right? So and then you can still go to the legal system if you want. But imagine the long tail, the long tail of transactions that cannot happen because there's no trust between these AIs. I mean, in the future, right now, even I would say in humans, right? Like a lot of these things doesn't exist. right now. For example, we have a protocol which is a marketing protocol. So you have a pot, you have a smart contract Tesco, you put the money there and you say, anybody talking about my product following these guidelines, sending a tweet to it can take the money out. You can have tens of thousands of people participating, and no one human needs to read every one of those tweets and assess how much you know, they're worth and who to pay. Yes? It's totally automated. Yes?

19:41You wouldn't be able to do this without it.

19:43Takatoshi Shibayama: Yeah. So it's basically it's just the kind of like bypass well I wouldn't say bypass, but you don't everything doesn't need to go to court. You know, you don't need humans always in the middle of everything t for all these small little things. And and I do understand like, okay, let's let's make it much easier and have ha have a system in place where it seems consistent, it's fair and uh don't really need to bring it into court. But if there were situations where these things were handling large sums of money or very important decisions and you do have to go to court, like would the gen layer decision hold up in real court or is it more for like a kind of smaller scale things?

20:20Unidentified speaker (speaker_1): Imagine that I tell you, hey, here's a smart contract on Ethereum, okay? Just put your token here and then I will give you something else. let's say, an OTC deal, for example, right? And you do, you sign the contract on chain,

20:32Unidentified speaker (UNKNOWN): but

20:32Unidentified speaker (speaker_1): then I take the money from you because you didn't read the the code or whatever it is, right? You can sue me. That's the reality right now is that code is not law. This is a fact right now, right? And so it's tricky because you know what is Uniswap? Well, Uniswap is kind of bypassing the system as well, right? You can't cannot really have something like Uniswap in a normal centralized system, basically, right? You don't have like JP Morgan allowing anybody on the planet to go and deposit the funds without any type of KWS and AML and then other people to go and trade against it and make money from that. Who knows what that means, right? Permissionless systems enable new applications, new primitives, right? And so what we're doing is that we're enabling new primitives, but in this case through the use of AI natively on chain, right? That doesn't mean that this is as I said, it's not like competing with the legal system, right? It's just trying to see is there a more efficient way? Is there a more efficient way that we don't need to be fighting? Yes? And here's why, you know, we want to do an OTC deal. So here's this contract, like the contract can take care of this, right? Now, if it's a trick, it's another thing, right? If I'm um scamming you, for example, but here you have a way that we don't need to be having a third party, having, you know, like some some legal agreement or yeah. This is just an automated system. It's just more efficient.

21:44Takatoshi Shibayama: Yeah, no, I do agree. I mean, you know, I I kind of see this as a first instance of like small claims layer, I would say. And then, you know, with the bigger things, you still have the courts as a backstop.

21:54Unidentified speaker (speaker_1): I mean, let's say that Uniswap didn't start with a hundred billion or like a trillion dollars in the first day.

22:01Takatoshi Shibayama: Yeah, absolutely.

22:02Unidentified speaker (speaker_1): Right?

22:03Unidentified speaker (UNKNOWN): So

22:03Unidentified speaker (speaker_1): that's the thing. You start like, you know, whatever. Algorithmic trading at the beginning was, you know, hundreds of dollars, then thousands of dollars, then you know millions, and then hundreds of billions. Right. So that's the thing. Things need to prove themselves. To me, like if I look at at crypto like the last 10 years, it's like you remember probably the DeFi summer, right? Like every single day there was something new. And I'll always say, like, I can imagine like a battery days, right? So you have all these different seeds. Are some of them are flourishing, some of them are not, right? Lots of experimentation, lots of things that are happening, right? A camera and explosion, right? And now for the last, I don't know, four or five years, something like this, it's been quite kind of boring, right? I mean, there's been a couple of new things that happen, right? Let's say, oh, there's like strong oracles. Okay, now you can have perpetuals, right? Or you have something like UMA. Okay, now you have prediction markets. Great, amazing. You see, it flourishes, it enables new things, right? So, what does it enable when you have ambiguous contracts as part of the equation, right? Common sense. Right? Being able to go to any website in the internet and fetch whatever image it is and make a decision on that. So, what I want to build is essentially a bigger Petri dish, like or newer newware Petri dish, right? That will allow for any other blockchain. And what I was saying about this is like. now institutions are looking at the Petri Dish of Ethereum, for example, to pick which are the big the winners and replicating that for themselves.

23:23That's a Robin Hood, right? So what's happening, what I think will happen is we start bottom up, we have this Petri Dish, things start to flourish, and then institutions will start to pick the winners and actually use them as well, right? And this is like the the normal life cycle. And in the middle, you will have all these agents that will, I think, be leaning towards finding the most efficient paths. They will not want to be wasting their resources. If, okay, it cost me 30 cents to deploy this contract that is intelligent, that can just be like between you and us, the judge, instead of going and hiring a uh lawyer or like going to the judge and so on, which will cost me who knows how much. What do I do? I trust it will. just trust blindly some other agent on the internet, it will deploy a smart contract or it will go to the legal system. Our bet is that it will just deploy an intelligent contract.

24:11Takatoshi Shibayama: Hm. Yeah. No, I get it. I mean we're still kinda in that Petri dish moment with these agentic AIs and you know, obviously we'll start to continue to have like these kind of disputes and I think a lot of these disputes are based on public information. But extending into the future where corporations start using this and those information is gonna be more private, confidential data. I think that's kind of the next layer for you guys, right?

24:35Unidentified speaker (speaker_1): TEs or ZKML, there's just so many different approaches to doing privacy, but I think like we just decided okay, let's start totally permissionless, totally transparent. Like let's see what this battery disenables and then we can always improve it. Right? There's there's lots of things to do on top of it. We'll be building for for three years with exactly the same goal, which is is this, right? And then after a time I think we'll say, okay, this makes more sense and where the demand is and then you improve the protocol in in those directions.

25:00Unidentified speaker (speaker_2): Yeah.

25:00Takatoshi Shibayama: So I guess like for your vision in five years, and I'm not really the huge immunitim person on these AI stuff, but five years from now, like what is a normal day look like for someone who's like using AI agents? Like what are the stuff that people can kinda imagine from this?

25:15Unidentified speaker (speaker_1): I think it's gonna be just you will have one or two AI agents that are pretty close to you, probably just on your phone and just literally one button, hey, this is da da da. It will be kind of proactive. And in the background what you will not even notice is that there's gonna be like many, many, many layers of AI, meaning AI agents as well. Some of them doing something every hour, some of them doing it, you know, once a month, like whatever yeah like different system, interacting with other websites, negotiating deals, like moving and trying to basically help you on the day to day. I think it will be totally seamless. I think you will eventually you will not even be saying I mean, I don't know if you will ever be saying like stopping to say AI, but uh maybe SI, right? But the point is I think it will just become seamless. Like nobody's talking, oh all like the internet right now. Like how often do you hear the word, you know, internet right now? So I think it's just like it will trend down. I think five years is a long time. So our objective function really was okay, how do we build something that will be useful after this sim singularity that we're now in, basically? Right. So now in the next two years, a lot of things will change. I have no doubt about that. Like to me, like I'm 100% sure that things will change. And so if that's the case, like how can I see I can't how can I foresee? How can I build something that I think will be useful even then? I just can't imagine, let's say, OpenAI coming and building their own gel layer because if they do

26:35that, well it will be their their jelly, right? It's their court system then, right? It's not a global one. We need a global one. We need a neutral one. Right. And yeah, I think eventually people will see it. There's just a lot of work to do, right? Like the monetary volume today that's moving through crypto is something like point five percent It's very, very, very little, right? So let's get to the same in that education and you'll see.

26:57Takatoshi Shibayama: Yeah. I mean I I was planning a trip for the year end and I was using my AI and trying to see okay what kind of uh itinerary that I can create and and asking AI to like give me various options for hotels and you know flights and all that. It's still not that great. You know, like I have to double check these things and and uh sometimes it hallucinates or it's giving me the wrong information. And it's definitely not something that I could just trust agents to do right away.

27:22Unidentified speaker (speaker_1): I mean it's changing very quickly. So like three years ago, you didn't even use at all AI. Two years ago, really like they could only barely speak pretty much, right? But they were hallucinating everything. And then we discovered that they can actually go to the internet to fact check things.

27:36Unidentified speaker (UNKNOWN): Now

27:37Unidentified speaker (speaker_1): they don't hallucinate so much because they actually go and check the things before telling you, right? And so how is it gonna be one, two, three years down the road? We literally can't imagine. Even on Gen Layer, there's been some basically all of ZenLayer will improve with time as well, right? All the network, well, it right now it's cheaper for you to use GPT six than GPT three. Literally. Try to you can't, literally. Right now use GPT GPT-3, three, right? And so how do you like the network that will be improving? over time as well and like become smarter, right? It enables new applications. For example, about a year ago, we started to see how contracts can resolve themselves by writing code. So you can have a smart contract, intelligent contract, that goes to an AI to see what is the snippet of code that's necessary to be able to run the snippet of code and give you the best answer. Because, for example, back then it was not not good at let's say counting Rs, right? It's very simple to write a snippet of code that counts how many R's in Strawberry. It's very difficult for them, because of the architecture, to respond at it. It was was was like you know a few months ago. Now it's no longer the case, you see? So it just is moving really, really quickly and like the more it moves, the cheaper it is, the more use cases it enables, the smarter it is, right? And the more things you you will be able to do with it.

28:52Takatoshi Shibayama: Mm-hmm. Yeah. And a lot of people say that kind of this software front end is gonna be gone and because of these agents it would be be plugging into different data sources and trying to get information out of it. So the front end is gone. So for us, like then what are we looking at? Like you think we're gonna have our own AI front end that just mobilizes everything and would be creating these things by ourselves, right?

29:14Unidentified speaker (speaker_1): Yeah. Yeah. You will have your phone, one tap literally, and it will just go and do whatever, like, you know, oh, it's like the school app. You'll not need to be all tapping the password and so on. And it's like it will be it's so clumsy right now. It's so bad. Right? Like my kids go to school. I have like five different apps for the for the kids. Like, what's going on here? I don't know the password of any of them. It's just like bandas, blanders, right? Like why do I need to do this, right? So I already have my agent that's actually like logged into all of them. And like every day is okay. Just tell me what's gonna happen. Yes, I wanna understand it because it's just too much. So yeah, I think it will just become totally seamless. It will be like transparent, like electricity, it will be everywhere. You know, my wife hates it. It's like I call her like Sarah Connor, right? Like she's just like absolutely hates it, right? Because but in reality, it's gonna be like this. Like we can I don't think we can stop it. So it's like, how do we keep some of the power for us? How do we keep some of this power for us? Right? Because everybody will try to take it away from us. Yes, let's say the labs are deciding whether you can use or not their model for a specific application or not, right? Why do they do that, right? Like what can they allow or not? Who gets a say on that, right? Can you try to at least like

30:22take some of that decision decision-making making power back? Because really that's all the thing we have left. Like, what will be left, let's say, 50 years from now for humans? What will be left? I mean, maybe creativity. I don't know. Do they really cannot create? Let's say I would disagree. I think that they can absolutely create, right? What will be left? Well, one thing I think will be left is liability, right? Taste as well, right? Like maybe, yes, like liability for sure. They die, they mean humans, right? Like we die, we go. to jail, right? So like that's gonna be the the final frontier. I think that that's gonna be very difficult for them to to replicate, right? So that's basically decision making power again, right? Yeah.

31:00Unidentified speaker (speaker_2): Yeah, absolutely.

31:01Takatoshi Shibayama: Great. This was a really fun conversation and makes me look farther into the future uh with these AI and AI agents. So thank you very much for being on the show. Thanks,

31:10Unidentified speaker (speaker_1): Saka. Have a good one.

31:11Unidentified speaker (speaker_3): Thanks for listening. If you like what you hear, give blogcasts Blockcast a like and subscribe. on Spotify and Apple Podcasts. And for all your juicy web 3 news, keep updated on Blockhead.co. Catch you in the next episode.

This page was produced automatically: the audio was transcribed with nvidia/parakeet-tdt-0.6b-v3+sortformer-v2.1 and interpreted by Qwen3.6-35B-A3B-UD-IQ3_S.gguf. Transcription and interpretation can contain errors; always confirm against the original episode. Nothing here is investment advice.
All insights