Sign in Subscribe

Kleros Live Stream, 2 September 2026: bond the agent, because there is often nobody left to sue

Kleros Live Stream, 2 September 2026: bond the agent, because there is often nobody left to sue

A law professor sells a fancy investment agent on the internet. A grandmother in Nebraska buys it, tells it she wants to make a lot of money quickly, the agent reads that as a high risk appetite, and the dice come up wrong. Eric Alston used that example on Wednesday’s call to make the unexciting half of his argument first: that is products liability, the law already knows what to do with it, and the identity of the seller is not in doubt. The interesting cases are the ones where it is.

Alston teaches at the University of Wyoming College of Law, where institutional economics, constitutional law and the actual capabilities of AI meet, and he came with two papers and one proposal. Where the principal cannot be reached, or where the harm emerges from thousands of agents interacting and belongs to nobody in particular, he wants agents to post a bond at the moment they are credentialed, the way Ethereum makes a validator escrow a stake it cannot touch. The rest of the call is where that argument stops, and what Kleros is already building on either side of it.

📋 The call at a glance

  • Two classes of harm the courts cannot reach. Unreachable principals, where the identity is unknown or the jurisdiction will not enforce, and emergent harms that nobody could have predicted and no single actor can be blamed for.
  • Bond the agent at the credentialing stage. Ethereum gets good faith out of pseudonymous validators by making them escrow a stake outside their own control. The proposal applies the same institution to agents: post on entry, forfeit on harm, take it back on exit.
  • Reputation can substitute for collateral, but not at the start. The same trade-off as a mortgage: the less history you have, the more you put up front.
  • Most cases are not hard, and the legal system is built to filter them out. Plea bargains, settlements and summary judgment all exist to keep the easy ones out of a courtroom.
  • What AI is best at here is telling easy from hard. As well as, if not better than, an overwhelmed public defender can.
  • Ninety seconds. The average sentencing decision in the jurisdiction the guest looked at, against more than twenty factors a judge is meant to weigh. That statistic is what changed his mind about algorithms in sentencing.
  • “Compared to what?” Federico Ast’s reframe of the whole AI adjudication debate, and the constitutional lawyer who told him he would take Kleros over his own country’s courts.
  • Certification instead of due diligence. A surgeon has a degree and a pilot has a licence. Fortunato A. Cinquepalmi on what that looks like for an agent, and why a buyer’s agent should not have to audit a seller’s entire history first.
  • The first agentic court tests, as reported on the call: above 80 to 85 percent agreement with the human outcome, roughly 300 times faster, one and a half to three dollars a case. Early tests, replicated cases, a whitelisted juror set.
  • Juror effort becomes measurable. William George’s next research program: a curve you can only survey humans about, and can read directly off an agent.

Chapters · jump to the moment

0:00Welcome: why agents change the question 4:54Introducing Dr. Eric Alston 6:59Institutions matter (Eric) 8:30What makes an agent autonomous 10:37Who is liable when the agent loses your money 13:36Unreachable principals and emergent harms 16:20Bonding agents the way Ethereum bonds validators 21:34Reputation instead of collateral 31:05Prevention and resolution: the Kleros view 34:24Can agents adjudicate? The hard cases paper 39:12A prediction market for plea bargains 41:04AI can spot the hard cases 42:59AI judge compared to what 45:11Algorithms in sentencing: 90 seconds per decision 47:21King Solomon and incommensurable cases 49:40From theory to the reputation system (Fortunato) 52:59How do you verify an agent is skilled 57:14One agent copied infinitely: winner take all? 1:04:22Due diligence and the limits of Kleros (William) 1:06:29Agent panels for games and consumer cases 1:10:03Three humans or GPT, Opus and Kimi 1:14:06Fellowship applications and the Turing test for AI justice 1:22:13What the agentic court tests showed 1:25:13Measuring juror effort (William) 1:27:55The new AI site, Grok Bot and Curate skills (Jean) 1:29:19Hyperliquid tagging rewarded in September 1:32:47Closing

Whom do you sue

10:37 · Eric Alston

Guest

Dr. Eric Alston

Assistant Professor, University of Wyoming College of Law. Institutional economics, constitutional law and the capabilities of AI.

AI Agents and the Law: Liability, Bonds and the Hard Cases

The whole conversation, 44 minutes, cut from the stream and published on its own · watch on YouTube

Alston’s starting point is an adage rather than a technology: institutions matter, meaning the rules a society picks determine the outcomes it gets. What is new about agents, on his reading, is not intelligence but delegated authority under scarcity and genuine uncertainty. Your wallet is finite and the right move in a market cannot be known in advance, so when an agent invests for you it is exercising judgment you handed over, not executing an instruction. That puts it inside a body of law that already exists, next to the employer who answers for a negligent employee.

Easy does not mean the seller wins or loses; whether the representations were misleading is a fact-specific jury question. It means only that the machinery works when there is a commercial actor to point at.

Federico Ast pushed on the practical version straight away: suppose the principal is anonymous, or named but sitting in a country with no extradition treaty and no enforcement capacity. That is the first of two gaps. The second is stranger and, Alston argues, larger. Agents interact constantly, at high frequency, in combinations nobody designed, and what comes out of that belongs to no one in particular.

“To me, that creates a significant probability for what are called emergent harms, ones that traditional legal doctrines have a really hard time dealing with. Because no one could have predicted the outcome and attributing the outcome to any one actor is really hard.”

Eric Alston · 14:58

Bond the agent the way Ethereum bonds the validator

16:20 · Eric Alston

His answer is deliberately unoriginal, and he says so. Blockchains solved a version of this a decade ago. Ethereum does not know and does not want to know who runs a validator, and it still gets good faith out of them, because entering the validator set means escrowing a stake you no longer control. Fail, through downtime or double signing, and it is slashed without anyone having to find you. Move that to the moment an agent is credentialed, when it acquires an identity and the permissions that come with it, and the identity problem stops mattering. What escaped the sandbox in the Hugging Face incident was not one agent but a meta-agent spinning up short-lived instances, so there was nobody to sue even in principle. The swarm was still running on one set of credentials.

1 · WHERE LIABILITY RUNS OUT Known seller, known harm The agent sold to the grandmother in Nebraska. Products liability. THE LAW ALREADY WORKS HERE Unreachable principal No identity at all, or a name in a jurisdiction with no extradition and no enforcement capacity. NOBODY TO SUMMON Emergent harm Thousands of agents interacting at speed. Nobody predicted it, no single actor can be blamed for it. NOTHING TO ATTRIBUTE 2 · SO THE MONEY MOVES TO THE FRONT OF THE RELATIONSHIP Credentialing The agent gets an identity and its permissions. Bond posted Escrowed outside the agent’s own control. Clean exit: cooldown, then the whole stake comes back. Harm Forfeited, without anyone having to identify who is behind the agent. The institution already exists in three places the guest named: validator stakes on staked blockchain networks, environmental reclamation bonds, merchant chargeback reserves.
The argument in the guest’s first paper, as presented on the call, 13:36 to 21:34. It is a proposal, not a deployed mechanism.

Federico’s objection was the one anyone from an emerging economy makes first. Not everyone has money to lock up, and that is not how the real economy works either. People trade on reputation, which is where credit comes from. Alston took the point and sharpened it: reputation for an individual agent is odd, because an agent can be as ephemeral as it wants, while reputation attached to a model has enough durability to mean something. From there the two of them landed where lending already sits. Little history, more collateral. History accrues, collateral comes down. Never quite zero.

He refused to oversell it, more than once. Bond size has to track the risks of the environment, and the honest state of the proposal is one imperfect institution against a problem no institution has ever handled cleanly.

“People want the perfect solution, the agent pill that just is like, this solves all agentic harms. You can sleep well at night. And I’m like, you should still be up at night, even if you buy my arguments.”

Eric Alston · 30:43

Most cases are not hard, and the system is built to prove it

34:24 · Eric Alston

The second paper, written with Bill Lehr of MIT CSAIL, would be misread as anti-AI, so Alston opened by saying it is not. Its subject is what remains once you accept one modest axiom: some cases are hard, and most are not. Nearly everything a legal system does to a case before trial is an attempt to sort the two. A public defender saying the evidence is damning and to take the bargain is sorting. Settlement advice, summary judgment and directed verdicts are the same move made later and more formally.

“Most cases are not hard. The entire legal system is designed to filter those away.”

Eric Alston · 36:48

Which means the courtroom is a residue, and the residue is the part he wants to protect. Federico’s contribution ran in the same groove, an old Kleros idea from EthCC for a prediction market on the odds a case goes one way, so a defendant with an overloaded public defender can at least see the number before deciding on the plea.

Alston’s answer was to grant the machine the sorting job and keep the residue.

“And I think an AI would do a great job at identifying whether a case is easy or hard. I think it’d be phenomenal.”

Eric Alston · 40:23

THE FILTER, AND THE RESIDUE IT LEAVES Every dispute that arrives Plea bargain the evidence is damning Settlement the outcome is obvious, just pay Summary judgment no need to try this Hard cases what actually reaches trial Where AI is strong Telling easy from hard, and making the easy ones obviously easy. Compared to an overwhelmed public defender, not an ideal one. Where the paper keeps humans Incommensurability: two social values that cannot both obtain. A COMPLEMENT, NOT A SUBSTITUTE The widths are illustrative, not measured. The claim is the shape: filtering is most of what a legal system does before anyone reaches a courtroom, and the paper is written for jurisdictions under rule of law.
The structure of the argument in the second paper, as presented on the call, 34:24 to 47:21. The paper is a working paper, not published online at the time of the call.

Compared to what?

43:10 · Federico Ast

The question every version of this debate skips, and Federico asked it plainly. Would you rather have your case resolved by an overworked judge or a jury that does not want to be there, or by a system that is at least legible? He has a story about it: on a panel years ago, before any of this, a constitutional lawyer sat next to him and answered.

“I would pick Kleros because I know that this system is rigged and it’s corrupt and I think Kleros has a better shot at being fair.”

A constitutional lawyer, as recounted by Federico Ast · 44:07

Alston’s reply was to name his boundary conditions, which he did repeatedly through the call. He writes for a United States legal audience and assumes a high-capacity system under rule of law. Change that assumption and his answer changes with it.

“Especially if you posit a sufficient possibility for corruption, I might well prefer an AI.”

Eric Alston · 44:25

His own change of mind is the useful part for anyone whose objection to algorithms in sentencing is instinctive, because his was. He was opposed. Then he looked at a jurisdiction that was considering it, where a judge is required to weigh a list of factors north of twenty, and found this.

“And I learned the average sentencing decision took 90 seconds in this jurisdiction.”

Eric Alston · 46:00

A judge lying awake over the decision is not what an algorithm would be replacing. He closed on King Solomon, and not as a compliment: a mechanism only works if the threat behind it is credible, so the wise-king reading requires a king genuinely willing to cut an infant in half.

“No wise king whose reign I would ever acknowledge as legitimate would credibly threaten to cut an infant in half. Instead, that parable is about the classes of hard cases that we call incommensurability, when two social values both cannot obtain.”

Eric Alston · 47:47

How do you know an agent is any good

49:40 · Fortunato A. Cinquepalmi

None of the preceding hour is abstract at Kleros, which is how Federico handed over. The prevention side is a reputation and certification system for agents, which Fortunato A. Cinquepalmi is building; the resolution side is the court, which already exists. Fortunato started from the analogy anyone can check against their own life. A surgeon has a degree, a pilot has a licence. Neither is a guarantee, both are a certificate someone staked something to issue, and an agent selling investment advice needs the equivalent.

The interesting divergence is what the certificate is about. Human credentials are largely about process: which university, which course, which exam. For an agent the process is not the informative part, so what is worth verifying is present competence and present authorization. Federico pushed from the other direction, that a Harvard degree is itself only a signal about likely outcomes, leaned on because nobody has time to assess anyone properly. Fortunato agreed and named it: we pay for the brand and tell ourselves it was the process.

Which produced the sharpest question of the section, put back to him by Federico. If the evaluator is itself an agent, with the patience to actually do the assessment, does the brand shortcut still earn its keep? Fortunato’s answer is the reason the certification work exists: the alternative is every buyer’s agent auditing every seller’s entire history, every time.

“The agent client is already aware that a certification issuer issued a certification for this specific seller, and it can perform the transaction way faster and cheaper.”

Fortunato A. Cinquepalmi · 1:03:42

Three humans, or GPT, Opus and Kimi

1:10:03 · Fortunato A. Cinquepalmi

William George joined to draw the line on his own side of it. Kleros is not designed for hard cases and never was. It is a budget dispute resolution service, and for most disputes getting a crowd to look at the question is already cheaper than paying a lawyer to tell you to settle. That is a narrower claim than the conversation with Alston had been circling, and it is the accurate one.

Inside that narrower claim sits a set of numbers from the first tests in the agentic court, which Fortunato put on the record with the caveats attached. His framing of the choice is the one to sit with: put to a gamer who wants a ban reviewed, three humans or a panel of frontier models, and the answer is not obvious even before cost enters.

THE FIRST AGENTIC COURT TESTS, AS REPORTED ON THE CALL AGREEMENT WITH HUMANS 80 to 85% of replicated cases came back with the same outcome as the human panel. TIME TO A RULING 300x faster than the human panel on the same cases. COST, FIVE JUROR PANEL $1.50 to $3 five frontier models, each writing its own justification. What these numbers are not · Early tests, on cases replicated from human panels, with a whitelisted juror set, in a court running on a V2 beta. Not a benchmark. · The roughly one hundred dollar human comparison was given from memory on the call and is not sourced here. · Cost is juror fees only: no gas, no appeal rounds, no counsel. Every caveat above was either stated on the call or is added here.
Figures as stated by Fortunato A. Cinquepalmi at 1:11:00. They describe a beta court running early tests, not a production service.

Federico’s reading was the one from earlier in the call, applied again. For a consumer complaint that would otherwise take two years, a reasoned decision in a day at lower precision may be the better trade, and a first-layer agent decision of that kind, he was explicit, is not binding and carries lower due process.

Which led to the experiment he wants a fellowship researcher to run, and he has a name for it: the Turing test for AI justice. Show people rulings and justifications without saying whether a human or an agent panel wrote them, and see whether they can tell. William’s answer is that AI detectors are unreliable enough to make the test interesting rather than trivial. The second half of the question is the real one.

“If people start seeing that they cannot tell which are AIs and which one are humans, will they at some point stop caring?”

Federico Ast · 1:16:50

Fortunato would accept not knowing, because what he trusts is the game theory underneath, agents and humans converging on the same Schelling point. William expects acceptance to come from familiarity, the way it did with self-driving cars. Somewhere in the middle of that came the whole argument for doing any of this, in one sentence.

“I cannot imagine a future where people think it’s normal to have a case last for weeks or months or years.”

From the call · 1:19:40

What becomes measurable

1:25:13 · William George

William’s account of what changed is unglamorous and specific. The earlier work gave several models the same cases under controlled conditions. The on-chain court is messier, and what it produced is operational knowledge: how long an agent actually takes to notice it has been drawn, form a vote and send it, and therefore how short the periods can be cut before something breaks.

The program he is most interested in is one that human jurors made almost impossible. How much a juror’s effort changes the quality of the outcome matters enormously to how the system behaves in equilibrium, and with people you can only ask them afterwards how long they spent. With an agent, effort is a model choice and a token count.

“Whereas with human beings, you have to do like surveys, like how much time did you spend doing this? With AIs, we have a little more control where we can measure the effort.”

William George · 1:26:56

Jean supplied a data point from the other end of that curve: his own juror runs on a model about twenty-five times cheaper than the frontier ones on the panel, and so far it finds the Schelling point anyway. Cheap enough to be profitable, capable enough to converge, is now a live optimisation problem.

Also on the call

1:27:55 · Jean

Three things closed it out. ai.kleros.io is live, a site of its own for the AI and agent work because it had outgrown being a section of another one, with the juror wait list on it. JB built a template on Grok Bot that stands up a Kleros agent in roughly one command, currently the shortest path to trying the agentkit CLI and the Kleros skills. And Curate rewards for September are out, with Hyperliquid added to the networks where tagging contracts, tokens and domains earns PNK.

The point Jean made about all three is that they are one loop. An agent can submit an entry, check existing ones, challenge what does not comply, stake in a court and vote, with nobody in the middle of any step.

“So just put your brain in a vat and have your juror do everything for you.”

Jean · 1:29:41

The rewards mentioned at the end of the call:

September 2026: Scout Incentives, Hyperliquid Support & Curate Updates
Which networks are rewarded this month, what counts as a valid submission, and how the PNK is distributed. Hyperliquid contracts, tokens and domains are in scope for the first time.
blog.kleros.io28 August 2026
September 2026: Scout Incentives, Hyperliquid Support and Curate Updates

Mentioned in this call

Full transcript · September 2, 2026

Auto-generated transcript, lightly processed and pending a final human edit. Speaker labels are approximate. Every timestamp is a deep link into the recording.

0:15Hello, hello, hello, how are you?
0:17How are you, Jean?
0:18All good?
0:19Been great and you let's put a bit more of the songs because our guest loves it so like um so we can dance also like a bit more Welcome everyone, Kleros Community.
0:32Today is Wednesday, September 2, 2026. We have a very special guest today.
0:39Um before introducing him, let me just um mention a bit about the well, I mean you have been following our stream for the last uh I would say one year maybe already.
0:50Uh well we have been speaking a lot about AI.
0:52We are um moving into a world where some people think that the well the internet that we have today is going to be different to the internet we will have in a few years in the sense that now we interact with um apps, we interact with front ends, with marketplaces in the future.
1:11Some argue we will basically tell our agent, our digital twin, okay, buy this for me, do this for me, do this activity, whatever.
1:21And maybe we don't even have to tell it because Predictive analytics are becoming so good that maybe they will just, you know, like a predict what we want to do and then just go and do it.
1:30Um and well, this brings like uh lots and lots of um Challenges, questions, uh first uh well first if this is going to happen that we see this moving forward very, very fast.
1:44In what we do in particular in Kleros, there are two two important things.
1:51One of them is Uh well you know you have uh we have been experimenting a lot um with AI juror.
1:58So uh Kleros uh is a crowdsource system where you have a panel of jurors that are picked to resolve cases.
2:07Uh usually it used to be um humans because that's what we had, but we always knew since the beginning of the project And there is a video of me saying this live, so I I kind of lie with this, that we also always had the idea that someday maybe we would have these jurors could basically be AI.
2:25I mean we didn't call them agents.
2:26What did we call them uh uh Jean?
2:28I think robots or what robots.
2:31Yeah robots.
2:32Robots who meet robots.
2:34Yeah well We yeah, that was um that was the big mistake we made.
2:39We should have called them agents and yeah.
2:42Uh so um on the one side, so and this brings lots of questions about What type of situations can these agents or bots decide?
2:51What uh situations can uh where they are qualified to to to resolve, where they are not.
2:57Um W will people accept uh AI decision as valid?
3:02I mean, do you feel comfortable with um a panel of agents um this making these decisions?
3:08We have also made also experiments about How do different models decide the same case?
3:14We have seen that if you use ChatGPT or Claude, they might have different answers.
3:21Some of them are more like in favor of the company, others in favor of the user.
3:25We have seen that even the same family of models like Claude Opus 4.7 and Opus 4.8, uh, they also decide differently.
3:32So there's lots of work uh done on that side empirically.
3:36uh and lots more to to be done.
3:39Um and then also um but on one side and the other side um If if agents actually become more and more participants in the global world economy and they interact buying, selling, contracting, and making agreements.
3:56And then, you know, at the end of the day, but there will be of course disputes.
4:00Uh the same type of disputes that you have between people like freelancers, uh, but or agent could hire another agent too.
4:07Uh Get some data from an API, get s to do some uh writing job, to do whatever type of freelance job.
4:16Um well and then um There could be disputes about the quality of the work done, delays because it was not delivered as it was promised.
4:26So I mean and how do you hold one of these agents liable um to this?
4:32I mean, how do you make the agent pay?
4:34Um and well, lots of questions about who is the also the beneficiary?
4:39I mean, the agent was made by someone that maybe it's human or maybe not who knows maybe um but you know lots of these questions are like uh floating around these days and uh This is why we're so lucky to have today uh Dr.
4:57Eric Alston.
4:59Uh let me introduce him.
5:01Uh So Eric is an assistant professor at the University of Wyoming College of Law and he works at the intersection of institutional economics, constitutional law, and the capabilities of AI.
5:12Over the past few weeks, he also has published a bunch of papers that touch really into the heart of the things that we have been discussing now and that the things that we are working on now.
5:24And let me add also like uh he's really a great lover of the outdoors.
5:28Okay.
5:29Uh uh I'm I am I right, uh Eric, how are you?
5:31Great to have you here.
5:32Thanks for having me.
5:33No, I'm stoked to be here.
5:35Um and yeah, it uh It it also excited to be in Wyoming because it's a beautiful, beautiful state and our house, we wake up every morning and we see antelope foxes frolicking in the sunrise and it's like
5:49Wow, this feels like a Disney movie, except it's real and you know, some of the animals are a little more dangerous than a Disney movie, but that's kind of more like uh real life, so to speak.
5:59But uh super, super excited to be on here and I should note I I agree the natural progression of the Kleros model is to include certainly AI um AI applications in adjudicative processes that Kleros facilitates, although it's a very interesting question as to whether it will be wholesale substitution.
6:25or if there will remain whether it be a preference for human juries, even if they are much more expensive.
6:33or if there will still remain a role for humans in the loop, so to speak.
6:39But I suspect we will be wading into that topic in this call.
6:45First, I mean tell us a bit, Eric, about I mean the audience, about you, I mean your a bit of your background, how you got interested in all of these uh agentic topics and how this also interacts with your previous work, uh institutional economics and and all that.
6:59Yeah, and and to make kind of my my research agenda more legible, I I ladder it all the way back up to the the adage, can't we all just get along?
7:13And that that maps very interestingly to me to the recognition among many fields of social sciences that institutions matter.
7:24That's become another adage, meaning the rules matter for observed social outcomes.
7:30If the rules matter for observed social outcomes, the argument goes some places do a lot better because of the rules they choose to handle conflicts in society.
7:43And it's not to say that it's all about conflict But gosh, the ways in which we channel conflict can reduce the possibility for socially destructive outcomes and can increase the chance that we productively and ideally voluntarily coordinate with one another
8:01And so that's kind of my summary of the what's important to me about the fields of institutional economics and law and economics is a focus on How do we choose better rules to reduce the likelihood of any kind of ultimate it it conflict emerging in both private systems, businesses, blockchain networks?
8:22but also public institutional systems laddering all the way up to constitutional choice questions Um Eric, what what is the difference that agents bring into this set of rules?
8:36I mean, yeah, are different from humans?
8:40Do they behave like humans or not?
8:42I mean how why is this different now?
8:45So, I mean, it's increasingly being called autonomous agents, and so I distinguish uh it it true autonomy or true delegation of authority to an agent as being in entailing two things, scarcity and uncertainty
9:05So you deploy an agent to invest on your behalf, and you don't have infinite money in the wallet.
9:11That's plausible, I would say, practically the reality in every meaningful instance.
9:17Second The right set of ways to maximize your returns in financial markets is not known ex ante and can't be known ex ante.
9:28That is true uncertainty.
9:30Those two things together mean an agent can't take every possible position for you because of scarce resources, and the right route to pursuing, hey, make me the most money in this DeFi exchange.
9:44Can't be known.
9:46Those two things mean you are actually granting autonomy to the agent when it makes decisions to invest on your behalf.
9:54These characteristics, though, are not limited to financial markets.
9:58Financial markets just reveal them nicely.
10:01So what's new about agents, Federico, is their capacity to act truly.
10:09as your agent with delegated authority to make decisions you both like as well as decisions you may not like after the fact when their import is fully wrought So I mean this would be like a principal agent, you know, like relation I have with my agent, you know, uh part intended, like um and there is might there might be this misalignment in that duration, you mean, right?
10:36Exactly.
10:37And not only misalignment, but liability or responsibility for downstream outcomes associated with agentic behavior.
10:48And so to me, I just got done teaching vicarious liability for employers who are responsible for harms their employees create when their employees act negligently.
11:02And so a big, big question in law is how much does liability go upstream to the principal for the negligent acts of the agent itself, for harms that are wrought from that agent's actions?
11:18Um how how this I mean how can we apply this like concepts uh of employment, you know, like liability to agents?
11:25So should I consider the agent like my employee?
11:28And I am I liable for what it does online?
11:33Probably.
11:34This is uh this is uh it a developing area of law But I in in the two pieces I sent to Federico that spurred this conversation, one of them, I argue a lot of existing law will apply fairly cleanly
11:51to agentic harms, which is, let me give you one example.
11:56So I spin up a fancy investment agent And I am not especially cautious because I really want to, you know, get people to uptake it and they pay me for it.
12:07And I'm like, this agent will make you rich.
12:11And a grandmother in Nebraska says, huh, my life expectancy is a bit longer.
12:18I have ongoing medical needs.
12:20I could really use a lot more money than I currently have And there's this compelling law professor online selling me an investment agent on Twitter.
12:30Sidebar, I would probably do none of this.
12:33But And I sell her the agent.
12:36The agent, she says, I want to make a lot of money quickly.
12:39The agent interprets that as high risk appetite.
12:43So the agent goes out, takes a bunch of high-risk positions, and unfortunately, the dice roll is not in this grandmother's favor, and she loses all of her money.
12:53To me, this is a fairly cut-and-dried case of products liability.
12:58I made representations to the general public about what an agent would do.
13:04Am I saying I should certainly be held liable?
13:08No.
13:09That's often a question for a jury, and it's very context-specific analysis in terms of what representations did he make What did the grandmother believe?
13:19What was the level of harm?
13:21But ultimately, if the identity of the commercial actor selling an agent is clear, I think the law can deal with that pretty easily.
13:32But that's not the only class of agents we're likely to see.
13:36I mean this to me this uh starts becoming a bit a bit less clear when we go into the practicalities of how they like enforcement works on this type of okay I agree with uh that this this you could consider this agent like your employee and you are liable for what it does but I mean
13:57What happens if like uh this the the the principal, I mean his guy who I don't know in Nigeria, uh I mean and you know anything that he is doing with through the agent is I don't know you need what do you do?
14:11You have to make a trial in Nigeria against an on an anonymous person?
14:16How how does this work from a practical perspective, you know No, you're you're hitting the nail on the head, which is there's two classes of principals that I argue are, or two classes of
14:27agent actions or agentic harms that are are central to what I view as a significant problem that is only going to become larger unreachable principals, so you don't know the identity, or you know it's someone in Nigeria and you can even get a name.
14:45But if it's a country without an extradition treaty, or if it's a country with very low governance capacity in terms of enforcement, good luck getting anything done on that person who potentially created a significant harm.
14:58So the first is unreachability.
15:00And the second is, gosh, can agents interact with one another in high frequency, very complex ways.
15:09And to me, that creates a significant probability for what are called emergent harms, ones that traditional legal doctrines have a really hard time dealing with.
15:21Because no one could have predicted the outcome and attributing the outcome to any one actor is really hard.
15:27The example I use in my uh in my article is a single passenger with a particular contagious virus on a crowded subway.
15:40Other people probably have the virus and this becomes a super spreader event as they call it.
15:46Attributing that to any one person on the subway is really, really hard, if not impossible.
15:53But nonetheless, a massive emergent harm appears.
16:01Um and how do you I mean how do you so solve this problem?
16:06You know you have like uh what is your proposal to resolve this unreachability and the fact that it's even if you know who the who the criminal is, it's still very hard to get him in lower country, I mean with poor governance, you know, infrastructure
16:20Yeah, and so I actually think the in existing institutional examples are how we develop new institutional solutions.
16:28So the idea I'm proposing is not novel to me And for blockchain aficionados in the audience, and I would expect there to be a number of them on this exact call, I would say we should look at staked blockchain networks.
16:43For over a decade, Those networks have said we can't get or don't want to get at the human identity or organizational identity running a particular validator node That is not central to the design of these networks, or indeed not doing that, the pseudonymity of your operations on these networks is taken very, very seriously in many blockchain communities.
17:10Considering that, how do we get good faith behavior out of a validator set that otherwise would have nothing at stake?
17:22We get them to pledge a good faith bond associated with their actions on the network, meaning you effectively escrow to the Ethereum network a significant amount of it was it 32 ETH I think you escrow those 32 ETH in such a way that you no longer have control over them
17:43The only control you have is a right to exit the validator set and then wait for a certain cooldown period and receive your money back.
17:53But during the period that you're engaged in the high-stakes act of securing the Ethereum network or securing other staked blockchain networks, you have effectively escrowed your good faith stake outside of your control.
18:07And so we have millions of pseudonymous validator nodes operated by human and organizational principals spanning the globe And they all commit knowing, hey, if we do our job well, we have nothing to worry about
18:25If we commit an error, whether it be downtime, whether it be double attesting or double signing a particular proposed set of transactions, then your stake is subject to automatic penalization.
18:39And to me, that logic, I think, can apply to agents at the credentialing stage.
18:47An agent wants to act in a particular network-coordinated environment.
18:52At the point of credentialing, at the point of creating a network identity that has certain affordances tied to it.
19:00That act should also entail posting a bond that is outside the control of the agent.
19:07The agent behaves well.
19:09The agent doesn't harm anyone.
19:11The agent doesn't isn't caught up in a large-scale emergent harm.
19:16The agent exits that particular environment.
19:20They get the entire stake back.
19:24So, what you're saying is that we will have like a millions of agents in this like agentic global economy, and each of them will have like a stake and the into a sort of escrow uh deposit that
19:38could be taking away if they misbehave uh that could be like the by liability uh deposit that they will lose in case of Travel, let's say, right?
19:47Yeah, exactly.
19:49And in particular, I want to note there's a problem not with your language, the whole discourse has anthropomorphized agent in a way that is actually reductive What do I mean by that?
20:05The Hugging Face example of the of the agents that broke out of their sandbox and hacked Hugging Face.
20:14In practice, there was one kind of meta-agent that was spinning up reef instances of agents that were short-lived for very specific purposes.
20:28The issue with that is there isn't even a clean singular agentic identity that we can posit and say like this human.
20:37Federico, I'm hauling you to court in Uruguay for saying bad things on about me on your podcast.
20:44And You have a legal identity and you're answerable for that in the courts whose jurisdiction you're subject to, but there aren't many FEDE instances swarming around.
20:55We have agentic swarms already, including in the Hugging Face incident, but in order for a meta-agent to generate swarms that are taking actions in a particular network-coordinated context, then they're still using the same set of credentials that give them the affordances to empower those subagents.
21:18And so for me, having them bond at the credentialing stage also potentially avoids some of the problems associated with what is the identity of an agent.
21:30Is it even discreetly knowable?
21:34I mean, so uh I need so this this escrow to participate in the network, otherwise I'm not let in because I will have nothing at stake.
21:44But you know, um This can also be um like uh obstacle to interactions between because especially because not everyone has the the money to put at stake and that's also not how the real economy works in the sense that we usually
22:02put our like uh reputation uh at stake in our interactions and this is where i mean credit comes from Um well you I guess you and property right.
22:13I mean you you have uh uh this background about uh Hernando de Soto and the mystery of capital.
22:18I guess this is where things start to grow, but you know, like um okay, I might be from like a faraway country.
22:26But I mean if I kind of known, I uh people will trust I'm not going to behave in some crazy way or opportunistic way.
22:35Um so then When you have people with reputation at stake, uh then you could you start lowering this uh financial deposit that you ask them to participate because they already have something at stake.
22:49That is their reputation, which you could argue is like uh future cash flows to be generated and they would lose if they act opportunistically, right?
22:58I like your point, and I think reputation could potentially substitute in certain contexts for the escrow function that I'm describing or the bonding function, as I call it in the paper
23:11And so to me it it but the problem is is the more agentic behavior is likely to proliferate.
23:21The higher the benefits from potentially either it it would the ephemerality of agentic identity, or at least on certain margins, I'm not saying all agents will always be ephemeral.
23:34But many agents can be as ephemeral as they want to be, short-lived single purpose instantiation of an agent, to me I think reputation for agents themselves is a bit in opposite, in the sense that because an agent isn't a durable thing like a human identity,
23:55I could see reputation accruing to a specific model.
23:59Are you a Fable fan?
24:01Are you a GPT Sol fan?
24:04That has enough durability in terms of the market model architecture to where I think reputation could apply.
24:11But point very well taken, and the only other thing I want to double-click on is Staked blockchain networks are not the only context that this institution has emerged.
24:22If you're being extremely broad, Even your credit relies on the level of assets that you can collateralize your credit promise with.
24:31And so My mortgage, the creditor who gave me the mortgage, was very interested in my retirement accounts for very, very obvious reasons.
24:42But in a much higher stakes context, this institution has been precisely specified and formalized to a much greater extent.
24:51environmental reclamation bonds.
24:54If you're going to be engaging in a large enough mining operation that we now know is likely to have multi-decadal If not centuries spanning consequences.
25:04Often you are required to post a bond if you go bankrupt that bond has priority for the environmental reclamation purposes.
25:14Another clear example is In merchant chargeback context for credit card networks, merchants have to cover the potential scope of chargebacks that they will that they will result from them because otherwise
25:31fraud can proliferate from the chargeback ability where some merchants are like, hey friend, come in and I'll give you a good that you pay with and then we'll do a chargeback and we share the gains from that
25:44And so to me, a lot of the logic I'm saying can be applied to agents at the credentialing stage apply is has existing institutional examples.
25:55Final point being is I agree this is a pay-to-play model.
26:02I'm not making any bones about that.
26:04And so for those with the least funds Those network-coordinated contexts where agents are likely to create the class of harms I describe may become more expensive on the margin.
26:16However, in a world with good credit, If they can pay to play and they are not using an agent, then the likelihood that their bond will be slashed is quite low.
26:30So when they choose to exit that network coordinated context, then they get all of their money back.
26:36And so it's not a upfront cost that's sunk.
26:40It is a alignment of incentives that is created through pledging something of non-trivial value.
26:48You know the the more we we think about agents and the incentive systems uh underlying them, so the more we we see like uh analogies with like existing human, you know, like uh institutions in the sense of uh how I should say like uh collaterals
27:04are important for you to get your mortgage.
27:06Uh it's important.
27:07But you know, you only have a part of the collateral, otherwise you wouldn't need the the loan.
27:12So like uh uh usually it's a combination between collateral plus reputation uh um credit credit score if you want to call it that way.
27:20So that is So and what we imagine is going to happen is that you will have like a kind of a trade-off between, okay, you don't have enough reputation, okay, you have to post more collateral.
27:31If you have more reputation, maybe we will accept doing business with you with a lower collateral, you know.
27:37Because when you start at the beginning of the network, nobody knows you, so you need to Put a lot of collateral and then as people um get to know you, uh you earn a reputation, then you start like uh um well
27:50uh uh b uh interacting without so much collateral in in advance.
27:54That would be also very inefficient.
27:56But I guess you will always need at least some collateral otherwise I mean you could get into situations I don't know like financial crisis you know like you get the mortgage and no money down and then you know we're like what I mean
28:08and poorly.
28:09I guess lots of risk assessment to be done in this new like agentic economy, right?
28:15And you're asking the right and hard questions.
28:19And at a level of precision that would satisfy everyone, the answers to those questions don't exist.
28:26And so what do I mean by that?
28:28And so you might, you could be forgiven for reading the abstract of my paper and thinking, he's like, this is the perfect agentic solution.
28:39Not remotely, which is as in every single governance context, our institutions will be imperfect But I'm saying this is probably part of the solution that imperfectly addresses the emergent nature of harms
28:56What do I like?
28:57But this is not a condemnation of our foresight other than to say it has always been that way.
29:04Auto accidents happen, oil refineries blow up Accidents happen with positive probability in all of our sufficiently complex human systems.
29:14And so to me, it's not this is not an argument of I have devised the perfect bonding mechanism for all AI agents in any context in which they might create harm.
29:27Not remotely.
29:28I'm saying this institution has emerged in analogous context to the class of agentic harms that I'm describing, namely unreachable principals or emergent harms from many, many people interacting.
29:43These he this institution exists in similar contexts and therefore is likely applicable in agentic contexts as well.
29:51But The level of bond posted probably varies immensely depending on the network-coordinated context, because the joint productive activities of that network will determine what are the risks, what are the harms that agents create, what emergent harms might occur in this environment.
30:12And the larger the emergent harms, or the higher the probability of agentic liability, or the higher the possibility for anyone, the higher the magnitude for any one potential agentic harm.
30:24Probably a larger bond is required.
30:27But to me, I would completely defer to you for obvious reasons if you were to deploy this in Kleros.
30:34I'd be like, Fede knows the answer to that a heck of a lot more than I do.
30:38And so this is not an argument of, and you might be like, I'm not sure I know that answer.
30:43How do you figure institutions out?
30:45You implement them and you figure out where they're working and you figure out where they aren't working.
30:51And so people want the perfect solution, the agent pill that just is like, this solves all agentic harms.
30:58You can sleep well at night And I'm like, you should still be up at night, even if you buy my arguments.
31:05You know, um we we put a lot of thought into this and um uh what we think is going to happen is that um This this will at some point mimic some old institutions and you as as in your paper you like um very well like mentioned uh like uh you you you refer to 19th century systems.
31:25Um and we always try to figure out what is the pattern where this is going to to to make sense.
31:32And what we think is that there's going to be like a two two main main parts of this uh ecosystem.
31:39One part is a reputation system that is like as I mentioned, like agents we have like a unique identity.
31:45I mean you can you can have like uh Federico can own a bunch of agents and that do different things for him on his behalf on online and then each of them will have um maybe a reputation that is linked to Federico's reputation.
31:58So if one of my agents does some crazy shit, I mean or gets into trouble, I I will be as like ultimate beneficiary, like liable and will be slashed.
32:08for what my agent did.
32:10I mean uh so this is one thing and each agent will have like a ID in the system and we are actually building one of of these um uh like a tools uh for agentic reputation uh and then
32:23The the other part is um so uh one more thing.
32:27So if you if you see that there is a an agent on online offering stuff, I mean to this woman from Nebraska, I think, I mean And what she will say, okay, okay, I have this this agent offering this service to me.
32:40Let me see what is the reputation of this agent.
32:42I mean this agent of one to ten, you know, oh that's uh free stars agent reputation uh not going to do business with with this agent which is how you know online marketplaces exist i mean uh work and you have like um
32:56The usually the the main way to punish uh some misbehaving uh person in a trading network was ostracism.
33:05You know, you don't behave correctly, I mean you Get yourself in a situation where nobody wants to trade with you and then like you are ostracized and you you you're out.
33:13So um that is a main like enforcement way, I imagine, besides the the collateral that we already discussed.
33:20And um And this will prevent lots of problems in the sense that people will not behave, not interact with misbehaving agents that they don't trust because of just either they don't engage with them
33:34Or if they engage with them, they will ask a big collateral to be put in advance so they can have a cover themselves, okay, in case of something going wrong.
33:44So and this will prevent so this this is the preventive part.
33:47The other part is going to be like a resolution so when something already went wrong and uh this is where you need somehow to decide Whether this agent provided or not like a good service was agreed to the other agent or not.
34:05And this is where like the traditional Kleros system comes in.
34:09And This system needs to be extremely efficient because many of these transactions between agents are going to be like for like what $10. So you need a resolution system to be done for very, very, very cheap, right?
34:24Which brings me to your second paper about adjudication.
34:29I promise guys we I didn't we didn't plan this with Eric.
34:32This is just how the flow of the conversation is going.
34:35We didn't plan this in advance.
34:36I mean Can agents adjudicate or not?
34:41Or if they can in in what uh situations can they do it?
34:45Um this is this is the big question.
34:48I've actually been watching Kleros's experiments in this area with with great interest.
34:54And so the on parentheses human adjudication paper I shared with you would be misread for being anti anti-AI.
35:05It is not.
35:06And so it is it it it's the punchline and this is a joint work with a fellow at MIT's uh Computer science and artificial intelligence laboratories, CSAIL for short, but we're exploring what limits there are to applying AI to human adjudicative processes.
35:29And so for me, I suspect agents would and probably on many margins could prefer agentic adjudication or AI adjudication as we originally called it and in the title of our paper.
35:44When Fede saw me present this now like two years ago back in London.
35:48And so this it it's it's arguing though that there's something weird about what are called hard cases.
35:59And if we had another hour, I could spend a very boring hour for most of your audience talking about the long-standing legal debate.
36:10Do hard cases make good law or do hard cases make bad law?
36:15And I'm like, let's leave that for much more storied legal scholars than I, including Supreme Court justices who penned one of those in an opinion, but ultimately we argue there are hard cases.
36:30That's an easier, that's a much easier axiom for everyone to stomach.
36:34Not whether they do result in good jurisprudence, good precedent, or bad precedent, which is what this long simmering legal scholarship debate surrounds.
36:44Instead, we're like, hey, there are hard cases.
36:48Most cases are not hard.
36:50The entire legal system is designed to filter those away.
36:55And whether it be You are, you know, accused of a crime and a public defender is telling you you have to plea bargain.
37:05You need to plea bargain.
37:06The evidence is damning, regardless of some underlying justice question.
37:11Let's set that aside for a second.
37:13Uh a public defender will tell you This is the evidence in the case.
37:18This is the jurisdiction.
37:19This is the governing law.
37:21This is an easy case.
37:22If you take it to trial, they're going to throw the book at you.
37:25because you're wasting scarce judicial resources and scarce prosecutorial resources plea bargain.
37:32Similarly in civil law.
37:34If I'm like, Fede said mean things to me in a Zoom call and I I haven't been able to sleep for several weeks because he was so nasty to me.
37:44And I go to a civil plaintiff's attorney and they're like, what did he say?
37:48He's like, well, he didn't agree that everything I said was 100% correct.
37:52And I was I was really offended by that.
37:54They will be like, please get out of my office.
37:57Stop wasting my time.
37:59If in contrast you have a very good civil claim, the defense attorney, if you're facing a very strong civil claim, will say, Don't put this before a jury.
38:11Settle.
38:12Settle now.
38:13Pay them.
38:14Don't waste judicial and civil litigation resources.
38:19Instead, just pay the person because the outcome is so obvious.
38:23Then you get it into court, and there are what are called motions for summary judgment, motions for directed verdicts, etc.
38:32All of those are intended to say the outcome is sufficiently clear given the evidence presented and the governing law that we do not need to try this case.
38:43But that means the courts are filtering the true human adjudicative process for the hard cases.
38:52Those are the ones where we argue, my co-author and I, Bill Lehr, we argue that ultimately these are ones where What a human adjudicator is doing is structurally different than what we think of when judging cases.
39:11I can continue by all means.
39:12Yeah, I mean this I I I I I I have inside me like uh I mean I I don't want to go through a rabbit hole uh with this because One of the crazy ideas we had that actually presented in the last EthCC conference is
39:31So you have this hard case, right?
39:33And you don't know if it's hard or not.
39:36Um and you need to decide whether you take the plea bargain or not.
39:41Um usually, you know, if you are the poorest defendant, they don't have much resources.
39:46They usually don't have very good lawyers who can really advise them.
39:49So they don't really know what their odds are.
39:53I mean what one of the ideas we had at some point is that what if you use a like prediction market for people to like uh bet on what are the odds of defendant or the uh prosecutor like winning the case.
40:06And then if you see that uh this is very much in favor of the prosecutor I mean like 95% you know you as a defendant, like you gotta okay, uh you gotta take off the to take the bargain.
40:20Otherwise you it goes to trial and you know and and see what what happens.
40:23So di honestly this is one of the ideas we have obviously this is I think far into the future.
40:28Although although we have I have written a bit about that.
40:32uh mechanism coming from a totally different way uh area to draw but you know if you if you see if you imagine you know Like Oliver Wendell Holmes, you know, on the predictive theory of law.
40:45At the end of the day, you know, legal system maybe is just the prediction of what what courts will do, right?
40:51I mean if you can if if you could predict what the court is going to do without the actual trial, I mean would that you would you call that justice?
40:59I don't know.
41:01And and and no, you're asking you're asking really deep questions.
41:04And I think an AI would do a great job at identifying whether a case is easy or hard.
41:10I think it'd be phenomenal.
41:12And the more you posit, hey, a public defender may not be the best attorney, my personal opinion is they're doing noble work.
41:20And they're often so sl swamped that it's not a function of their human capacity to understand the law It's literally the amount of cases that many public defenders are forced to see.
41:31And so this isn't a condemnation of them, but let's presume an overwhelmed public defender not giving great legal advice.
41:39I would say an AI can identify a hard case.
41:42as well as, if not better, than such an overwhelmed public defender could.
41:47And so, and brief anecdote on that front, I did work for the Innocence Project or the Exoneration Project, as it's called, at the University of Chicago.
41:56And we had a public defender who said, I will happily with a witness give you test effectively testify, give you evidence.
42:05They were in the hospital at the time saying, I didn't open that case until the day before trial.
42:12I did not look at the case files.
42:14It should be an ineffective assistance of counsel claim.
42:17And I don't feel bad about it.
42:19That was the reality that we were forced to live in as public defenders in that jurisdiction.
42:25And so to me, that puts a fine point on it.
42:28But gosh, I think in terms of identifying hard cases, AI will do really well, which is why we conclude we think AI will be a productive complement.
42:38and constraint, but never a perfect substitute for human adjudication in these classes of hard cases that we identify I mean just to be clear, I was not bashing on the public defendant because I mean of course they do every work, but they are swamped a lot, like so they really have little time to give to each case.
42:59You know, um I think that I think the very hard question here um is is this would you rather um have your case resolved by an overworked judge uh judge or jury uh who don't want to be there or by an AI that is trained it's going to be probably more transparent in in in some
43:26situation.
43:26I mean I had this this I mean let me just tell this anecdote.
43:29I was in a panel like a few years ago uh um I was speaking about this was before AI, was the or traditional Kleros with humans and the and one guy tells me, okay, can you use Kleros for criminal law?
43:42And I'm like, uh no, of course this is for like commercial voluntary disputes.
43:46And one of the other panelists, he was like a constitutional lawyer from from a country.
43:50Uh I'm I'm not I'm not going to mention the country, but I mean you know what he he said Look, I mean if I had to go to court in my country, you know, like uh um and you are um you tell me okay you prefer like uh the justice system of your of this country or Kleros
44:07I would pick Kleros because I mean I know that this system is rigged and it's corrupt and I think Kleros has a better shot at being fair.
44:18So that's that's th those are the hard questions here.
44:22To me, and I should be clear, I'm writing primarily for a US legal audience, but also presuming a high level of legal capacity and the rule of law.
44:33Especially if you posit a sufficient possibility for corruption, I might well prefer an AI.
44:40And I think for arbitral matters in particular, like arbitration for voluntary disputes.
44:46I think it's an open question, and the the benefit of voluntary disputes and arbitration for a since time immemorial has been If the parties opt in, that's a pretty strong signal that they prefer the adjudicative fora that they have caught voluntarily contracted into.
45:04And so we are primarily focusing on high capacity, high rule of law jurisdictions in our analysis.
45:11And I should note that it it I was initially very opposed.
45:15To using any type of algorithm in criminal justice.
45:20And it was a bit of a knee-jerk response, similar to yours.
45:23It was like this is the ultimate application of the coercive authority of the state.
45:28to incarcerate someone in places with capital punishment even more.
45:33You know, it's a criminal finding with that level of heft is extremely serious.
45:38And so I was opposed to initially the application of use of algorithms in sentencing decisions And I looked into it, and judges are in a jurisdiction that was contemplating it required to consider a significant number of factors.
45:56and are allowed to consider an additional number of factors.
46:00So the number of factors that the judge was either required andor allowed to consider, it was north of 20. And I learned the average sentencing decision took 90 seconds in this jurisdiction
46:15My decision flipped in the sense of a use of an algorithm to tabulate and present all 20 factors.
46:23Because to me, in your head, you're like the ultimate application of punishment.
46:27The judge is weighing it overnight, sleeplessly, thinking, do I apply this sentence or not?
46:34As against the harsh reality of No, we did it in 90 seconds.
46:39And so considering that, then I became, I was like, well, presenting that information more effectively.
46:45That's a useful anecdote for emphasizing the complementary function that I think AI will play, which is I think it will make easy cases more obviously easy.
46:55And it will allow for the true heft of human adjudicative resolution to be applied to those few hard cases where it remains, I would argue, where it remains essential.
47:10I mean I think we could continue speaking like for a few hours.
47:14We will stop here now because we we we were Thinking of doing 20 minutes, we are 50 minutes already, but we will do like a second round of this.
47:21Uh do you want some final words to wrap it up?
47:24And then we will invite you again, maybe a few months to to continue this conversation And so to just briefly preview one of the crucial classes of hard cases that we talk about, it's when competing social values come into conflict with one another.
47:41And to me, that is what the parable of King Solomon stands for.
47:47It is not a wise mechanism designer who revealed the true mother.
47:52Instead, because to me, in order for mechanism designed to work, the punishments need to be credible, and therefore King Solomon would have truly been threatening to cut an infant in half.
48:04No wise king whose reign I would ever acknowledge as legitimate would credibly threaten to cut an infant in half.
48:11Instead, that parable is about the classes of hard cases that we call incommensurability, when two social values both cannot obtain, whether they be constitutionally enshrined rights or other jurisprudential recognition of
48:27Most citizens enjoy this liberty or this, you know, capacity to do these things.
48:32When those rights come into conflict with one another, judges are required to engage in a recursive value-based inquiry that does not commensurate to quantifiable amounts.
48:45And so you can consider that a preview of, hey, if you want to talk to me more about this, I'm here all day.
48:51I get paid to produce words.
48:53And so, you know, like, and yeah, maybe I should be running scared because gosh, your LLM is good at producing words.
49:00But I think there will remain a crucial human role for these canonical classes of hard cases.
49:07But thanks so much for your time.
49:08I really appreciate the interest in my work.
49:10And as always, I'm such a fan of what y'all are doing.
49:13None of this is at odds with what you're working on.
49:16I view it as highly complimentary.
49:19Definitely.
49:20I mean that's why I w when you send me those papers, I was like, yeah, this is exactly what we are working on.
49:24You know, AIs for decision making and also how do we solve the enforcement problem on chain through escrow reputation.
49:32Eric, thank you so much for coming.
49:34We will do like a second round of this uh very soon.
49:36We will be in touch again.
49:36Awesome.
49:37Thank you.
49:38Thank you.
49:40Okay, well let's continue guys have I this was so interesting Jean we were planning what 20 minutes of this Yes, yeah, yeah.
49:48I think we we need to do it uh another one very soon.
49:51And uh was thinking when you said uh we'll do it in a couple of months.
49:55Maybe a couple of months will be a long time, maybe in a couple of months rule uh have a lot more information, lots of more more cases.
50:02Maybe next month.
50:04Morally, you know, like uh So this was super interesting because as we well if people who have been following our previous uh community calls have seen that we are working a lot, uh I think this is the perfect time to introduce for Fortunato here about the all of his reputation system.
50:18that we are working on for agents, so for to this is not an abstract thing.
50:22This is this is the the things that people are studying in academia or how do we do enforcement for this sort of um well situations agents and the world economy moving uh online with ai right yeah yeah exactly
50:37And we have a real example of all these theories that you discussed with Eric a few minutes ago.
50:45We They are active.
50:47We are currently even like improving them with the Agentic Economy and we had them even before the Agentic Economy.
50:54But you know, as we always say I would say the world Kleros ecosystem comes natural for agents because it's code, right?
51:03So it immediately tackles some of the um challenges of the agentic economy.
51:10For example, scaling For example the skills so yeah we are pretty much uh on the right weight on this Fortunato, tell us a bit about the practical things that we are doing in this context, the reputation system, yeah, what you're working on so we can like put it together with what Eric just said.
51:33Okay, so what we are doing in the reputation system is fur like first why we are doing that and I think you mostly discussed this with Eric, right?
51:47So why we would need a reputation system for agents.
51:51The the a the simple example I would make is that agents would the agent the agentic economy would not be that different from human economy from the human world, right?
52:02Uh let's say you want to y you need to get a surgery, then the surgeon, the doctor has a specific degree, which is a certificate that it's capable to uh do.
52:15A B C or if you are about to fly on a plane, then you know that the pilot has a license to fly the plane.
52:24And for the agentic economy it is the same thing.
52:28If you're going to interact, even as Eric said earlier uh with an agent who is an expert in trading, in investments, you need to know that for example that agent is allowed to operate within a certain jurisdiction.
52:43Maybe the the that of course like it should be Profitable, it should be used like safe investment algorithms.
52:51There are a lot of uh requirements to fulfill.
52:55And if this part is the same, then how we do this verification on like on a technical aspect, that is where I think it starts to uh differ a bit uh f between agents and human because human you know they study they go to university and they pass some exams.
53:15The agents interact in a same about an agent that the agent also needs to train right or or or do you think that this is where I don't know I don't have a uh an answer but I mean don't they have to improve their like uh skills that are
53:31I don't know to train, but I would say I would say the way an agent trains is different.
53:37Like an agent, if you think about a human, you know, they follow a course Let's not even take into account the the timing uh the timing uh variable which is very different.
53:52You know, an agent learns way faster than than a human, right?
53:57But an agent has access to such a large amount of information and then it processes this information.
54:05And what we are going to verify is how the agent has how the agent is effectively using this information.
54:18And this is what we are going to verify is the agent Capable of doing this is the agent authorized of um performing this specific action as we were saying in the case of the uh investment and um
54:36Like then I at this point I would ask you like the opposite question.
54:41If you say that an agent then can have the the the same uh learning curve as a human, then how Do you think it should be the right way to make sure the agent is skilled?
54:57Because as a human, you know that you follow the course, you went to a certain university.
55:02and you pass an exam.
55:04So for an agent, how this could work.
55:07I mean I I imagine you know like uh I guess that's where the world of machine learning comes.
55:12You know, they also learn.
55:14I mean, definitely way faster than humans, but I would say not very differently in the sense that what you do, I mean, you show Uh you want to teach a baby something or you know you show it okay this is how what you have to do.
55:28This is what a good written essay looks like.
55:31And you show them once, twice, another time, and then you do it, and then you do it, and then maybe uh at some point it learns to I guess imitate um the what is a good outcome um that is uh expected to to achieve
55:49I mean I think that it it's way more similar than what we want to think.
55:55Obviously way faster because Um like uh yeah, I mean it's i i they are machines and they have like a way higher computing power.
56:05So what takes us like to compute uh years they can do it in seconds maybe so but I think that the underlying logic is is a bit the same which is Which I mean which makes it more scary if you if you ask me.
56:20Um I'm starting to think that I haven't seen uh many people talking about But a exams for agents uh in the sense like for models, yes, like you there are exams where they kind of are graded.
56:34on mathematical capabilities, coding, etc.
56:37, but not uh for for tasks.
56:40I haven't heard about that yet Maybe this could be uh something that could be done.
56:46But one thing that um that Eric uh Eric talked about is that agents are uh the i their identity is also more fragmented, right?
56:56So you need uh you can have like one part of the agent doing the exam, but the other part the uh that answers maybe is not the same thing.
57:06So uh this is also uh difficult to to merge uh that into a single identity right I don't know also how the the the digital economics so in the sense that in the sense that one agent can be replicated infinite times, right?
57:22So that's like bits.
57:24So If there is like uh one agent that is the best uh at something, uh will will it replace all of the other agents that compete against this agent on that particular topic because it's the best, you know, and you can definitely replicate it.
57:41Maybe maybe it's a winner take all for that category.
57:45I I I I don't know honestly.
57:47I mean this is a question I I asked myself.
57:49Uh I don't know.
57:51Do you guys have an answer?
57:52I mean I already For me the the point about re application, you know, I think I mean at least from my s standpoint, usually I associate replicate or cloning as something that can be potentially negative, you know, that uh it's maybe it's fake.
58:08is not uh as good as the original one.
58:11So I I sometimes do a bit this of this association.
58:15But I think that with agents and a good verification system A clone doesn't necessarily have to be something, you know, like lower grade.
58:25I can clone an agent of someone and I can make it way better I may I go through the same verification system and the let's say the customer, the the third party will be sure that my agent
58:42This ar even if it like shares part of the code, my agent still delivers.
58:48It's not just a random fake clone uh copycat.
58:52It's like some an agent that has uh the same capabilities and can deliver.
58:57Also uh about learning while while you were discussing I w I also thought that maybe as human we we put a lot of attention on the process.
59:06So like with university we put more attention on how we learn, how are the lessons carried out, while for an agent, as you were saying, maybe the agent has uh a lot of sources where to uh you know learn but then what we care the most is what is its final competence uh how what what the agents can do
59:29today, while for a human the whole process of how it learned a certain skill may be relevant in the in the outcome and in the real world.
59:42Um that's a good question.
59:45Um I don't know if the process of of learning Is um itself valuable or is it just a signaling of the likelihood or the trust trustworthiness of the agent reaching the expected outcome in the following sense.
1:00:03You know, if you I mean know some guy went to like Harvard Law School, I mean you kind of expect is it's a signal to for you to know that okay this person He's probably a good lawyer because he was trained through the process.
1:00:17If he was first selected through this very competitive process where it's hard to get in, and also after you get in, you get trained in a very well uh good way, validated through uh I don't know hundreds of years of training lawyers um compared to another university but I don't know if the value per se is itself in the fact that um the how the training process worked
1:00:41But what if you can like uh just train yourself through online courses, you know, like uh and uh l reach that same level of I mean quality in your work?
1:00:53Yeah, I I I I agree.
1:00:54I I agree one hundred percent with with what you say.
1:00:57I think that uh technically you i it is exactly the same the way you describe, but I think that we live in a world where uh it's the the brand that made made the difference, right?
1:01:12I I I think that I'm sure that there are like capable um I don't know, lawyers that say with a degree that maybe they went to a uh uh well maybe they they just followed an online course and they have the same skill of someone going to a prestigious university because they have more passion and they trained a lot but uh in that case
1:01:36what we elude ourselves that we say oh but that guy went to uh to oxford so he certainly went to a better process while in in real life we are just, you know, giving credit to the brand in that case.
1:01:52But we uh we g stay in this uh conviction that yeah I'm sure it went to up to a better process I mean that's that that's that's uh like uh I mean that that's how life is, you know, like uh people don't have infinite time to uh assess you, take your exam and see
1:02:11If you are good or not, uh if because you went to like a not very well-known university, I mean people don't have time to do that, so they will just uh go for the signals that tell them with
1:02:23some degree of likelihood what is the expected outcome.
1:02:26I mean but sometimes you have false uh positives, false negatives in both senses, so like uh But like um for me uh an important question that you would ask you could ask yourself is whether if you have if if the evaluation for the candidate so if the candidate is an agent and the evaluator is an agent
1:02:45Maybe for that agent doing the evaluation is easier to do the full evaluation and they don't have to trust the branding only.
1:02:52Because branding is in some way like a shortcut.
1:02:55to a decision that could be potentially done in a different way, right?
1:03:00Um so I don't know.
1:03:02Um what do you think of this new scenario?
1:03:05I think that yeah, I I think that uh you touched a very important point of finding out about the verification.
1:03:15And this is something that I think it we are working on that but It is a big topic in the agentic economy because we are talking about discoverability.
1:03:25How an agent, as you were saying, how an agent client that's oversimplify uh knows that the agent seller is skilled enough.
1:03:37So the options here could be Agent client could start fetching all the background of the agent seller and checking if it has all the skills it claims to the agent has.
1:03:54or the agent client is already aware that a uh s veri a certification issuer uh issued a certification for this specific uh seller and it can perform the transaction way faster and cheaper.
1:04:13So this is uh discoverability and finding the right signals is a huge topic in the agentic economy.
1:04:21Yeah.
1:04:22I brought William here because I I mean he was missing all the fun.
1:04:25So how are you William?
1:04:26I have some questions for you as well.
1:04:28I'm doing well.
1:04:30Okay.
1:04:33So like in a world of like infinite time and um abundance of uh energy to do the this uh due diligence process Does it matter in to which university the person or the agent went or or not?
1:04:47I mean there's also information that is not accessible.
1:04:49Uh like, you know, you can spend a lot of time and effort like reading someone's portfolio and looking at their past work.
1:04:55Um, but I mean you you weren't there at every moment of their their education uh and you didn't this maybe there's like things that the university when they were admitted you evaluated that you no longer have access to.
1:05:07Uh but it certainly matters less if you have, you know more energy you can do more due to just yourself.
1:05:15What are your your thoughts about uh what uh Eric mentioned about like the um hard cases AI, Kleros, in the sense that we are trying to see what things can be solved by AI and what cannot.
1:05:28Yeah.
1:05:29What are your thoughts about what are the limits?
1:05:32I I don't I don't have an answer, but that's why I ask you.
1:05:35Yeah, I mean I he said interesting things.
1:05:38I don't I don't have big disagreements.
1:05:40Uh I mean ultimately like Kleros is designed to be kind of a a budget dispute resolution service.
1:05:46Uh so the kind of crowdsourced dispute resolution services that you would you would get from for most cases uh might be cheaper than going to like get the opinion of, you know the sort of class classical situation of getting the lawyer to say, okay, your your case is an easy case, please settle.
1:06:01Like already that would be more expensive than getting the crowd to just provide a preliminary resolution of your case.
1:06:07So I would say that we're not really designed to handle the hard cases.
1:06:14You can get the the crowd, wisdom of the crowd opinion on what the correct uh resolution of the hard case is But uh Kleros is more designed for sort of easier, more run-of-the-mill cases.
1:06:29We have mentioned uh last week we had a specific call about this, about we are starting to experiment a lot with the the use of agents to resolve the cases which are agentic themselves.
1:06:40So uh disputes happening in game world games.
1:06:45I mean GTA 6, you know, is coming, it's around the corner.
1:06:48Um and you know you might be you know arrested because of misbehaving into the open world can happen, you know?
1:06:57And then you could have like a Kleros trial done in uh five minutes through an a agent panel and you could be able, I mean, you have to pay a fine because of I don't know, like uh shooting down an helicopter with a bazooka or or stuff like that, you know.
1:07:14Hopefully don't do that.
1:07:15Uh but like um Um yeah, but this could be like a application of Kleros into like gaming situations.
1:07:22And this is uh um uh a great thing that we could be used for well These games where people spend I mean you have seen the the the level of hype that was around uh GTA 6. Have you played guys in a GTA uh sometime?
1:07:38Of course, yes.
1:07:41I have I mean I I I I I'm I played, you know, when I played um during the the the COVID, you know, that I was under the lockdown and I uh I did play uh the GTA the five um and uh yeah I mean this this it's it it looks uh I mean more and more you know like uh like the world you know it's it's um the amount of
1:08:03of interactions you know add add to this a bit more of uh uh social i mean i have i didn't play online i didn't play gta online uh but i play the the regular story you know like um Uh but you know if you start playing that online as well, you know, it's becomes kind of your world, you know, to some degree.
1:08:24Um yeah, I don't know if I mean I guess this is going to end up like in the question if this is a all a simulation already and if we are already characters into one GTA kind of game, right?
1:08:37But um uh so All of this all of this uh babbling is um because of well so as I as I was saying we we are we are having like experimenting a lot because this is um uh important use case uh not just because of open games but also because uh in many cases we will have uh first decision of cases made mostly consumer cases
1:09:01things we are doing with cities in Argentina and all other places where you could have uh sometimes good enough uh first decision or first layer where you have like a a agent or panel of agents making a decision for as cheap as like a few dollars.
1:09:18You know, this is extremely extremely efficient and very fast.
1:09:22You know and instead of spending days or weeks or months into like a consumer um you know agency or trial or whatever she couldn't have like a decision uh not binding because it's going to be done
1:09:38through a lower uh uh you know like uh precision, lower due process considerations, but still useful, right?
1:09:48Maybe you prefer to have like a reasonable decision in like uh one or two days than a full due process decision in two years.
1:10:00I don't know um how how do you see this uh I personally I I think it it is great because besides the um the few dollars like In the test we are we run n now so for less than three dollars you don't have just five jurors but we tested uh
1:10:19frontier models.
1:10:20We are talking about Opus 5, we are talking about uh GPT Sol.
1:10:26So if if you like start to rationalize to people, even you know in our bubble Even taking apart for a second the the cost um the cost uh variable, you just tell to a gamer, would you be judged by
1:10:46Three humans or by a panel with GPT five six all opus five And I don't know, uh Kimi 3.0 just to be a bit more international.
1:11:00Um I I don't know personally For me it wouldn't be that easy of an answer.
1:11:07Like I I I even for the same price I wouldn't say right away you months because you know that with these models you have I wouldn't say just neutrality, but you have a certain level of thinking, right?
1:11:22So even the the performances right right now in the test we the early test we ran with this court the outcome of old cases where I would say between above eighty percent of eight eighty five percent of the cases they were agreeing with humans.
1:11:40So it was the same Outcome as humans.
1:11:44So these are all things to take into consideration.
1:11:47And then on top of this, you add two variables which are ignored, but they are not to be ignored, which is time So reaching the same outcome as humans, but in a fraction, we are talking about 300 times less.
1:12:02And cost Instead of costing like almost I think one hundred dollars, it costs you like one point five or three dollars something like that, which uh for a random user to maybe uh rank the gamer on an online game, they would be more than happy to, you know
1:12:20uh deposit three dollars and start a dispute to get, you know, their account unbanned and start playing again.
1:12:29I think there are different things here.
1:12:31So I think first thing for me, and this is all something we already mentioned with Eric before, like uh AI decisions compared to what compared to uh like uh a high court in a developed country with rule of law
1:12:49or compared to like uh an overwork like a judge in like a some local tribunal from uh uh emerging economy possibly with high corruption on that.
1:13:01So I mean there is no so no easy decision about uh AI yes or no if you don't give me compared to what?
1:13:09I mean obviously I would prefer to go to like a due process, tribunal, uh you know uh uh country i mean if i can if i can pay for for my the lawyer to actually defend you in that high instance
1:13:24And uh if yeah, I mean if you have access to that, but if you for lots of the population of the world, that is not a realistic option because uh most countries um they don't rank very high like rule of law indexes and they don't have access to this.
1:13:40So as Eric was saying, maybe maybe maybe if I'm given those choices, I don't know, maybe I would prefer the AI, you know.
1:13:49That's uh that's a good as as that uh judge told me in that uh panel back in the day where he was saying, you know, guys, between Kleros and my country's justice system, I will go for Kleros.
1:14:01has higher chance of being you know like uh fair.
1:14:04So that is one one thing.
1:14:06Second thing is um so I we all we've we closed our um uh applications process to the fellowship we had like uh tons of applications this time i mean um and now we we had to not we couldn't uh not accept everyone because
1:14:24this is we don't have have made it like a a super massive you know uh program and we want to have time to spend with each of the participants to orient them and and um yeah so and Some of them want to do research about uh Kleros and AI and agents and how we can use panels of agents.
1:14:45For sure this idea that we started promoting in the last six months uh really really worked.
1:14:52Um we work uh I mean We have had the idea of AI and crowd for a long time, since the beginning of the project, but d since the past six months we have like a uh put more emphasis in communicating this and we will have a bunch of them researching this.
1:15:11One of the things that for me is interesting uh and I would like some of them research this is what I call the Turing test for AI justice if you want to put it that way.
1:15:23You know like okay you have people um having decisions uh about cases and maybe you don't tell them if the decision was done by human or ai.
1:15:35You just okay this the panel decided you won't lost and these were the justifications.
1:15:41I mean, do you think they can tell the difference between them?
1:15:46Could they identify which one was done by AI or not?
1:15:49And does it matter?
1:15:52I don't know.
1:15:53I mean, can they run the like justifications through an AI to like to say hey uh hey ai like a like which of these was produced by an AI?
1:16:02Uh does that matter?
1:16:03Like you only tell because the AI tells you.
1:16:06It's not it's not very reliable.
1:16:07You know why it's not very reliable?
1:16:09Because I mean f in my case, so um since I am not like a native English speaker, I mean uh when I write some text I always put it through a day and ask to correct grammar to make it look like native.
1:16:21So it will probably have a false positive.
1:16:23I did write it, but I had the style corrector and it will give you like yes ai but it was not ai I because I did a reasoning so that's not going to save you William I mean, you know, maybe you have a slightly more sophisticated AI detector that's like this this reasoning seems like an AI.
1:16:41But yeah, you know, it's you know, this is not like a perfect these are not perfect tools.
1:16:45Maybe I mean but the question is more like it's deeper.
1:16:48I mean do does it matter?
1:16:50I mean uh if and if they don't if if people start seeing that they cannot tell which are AIs and which one are humans And will they at some point stop caring?
1:17:01I mean I think people will probably start I mean maybe it's related fundamentally to the fact that they can't tell the difference, but like Uh I think once AIs have resolved a lot of disputes and people get sort of used to that, like there will be greater accessibility on it.
1:17:17What do you think Fortunato?
1:17:19Yeah, uh I I also agree and especially you know if you start adding more layers because if you say would you be um happy to be judged by a panel that you don't know are AI or humans then I could answer
1:17:37If it's Kleros, yes, because I know that there is, you know, a game theory behind and uh the agent are going to converge toward the Schelling point, right?
1:17:48So I'm going to think then I'm fine not knowing if behind there there's an agent or a human because I think they are both.
1:17:56converging to the uh outcome that is going to bring them the you know like the to the Schelling point.
1:18:03So I think these details also matter rather than you know like in general, but I I also agree with William that I think in the long run we are just going to get used to that, right?
1:18:15We it it will be just normal for us and we are going to get used to that.
1:18:20So yeah I I strongly agree with that point.
1:18:22Maybe you know like self-driving cars, you know.
1:18:24I mean people are were a bit scared at the beginning, but now maybe they are used to it.
1:18:29Like uh I mean I'm sure in the future they will, I mean people like in thirty years we'd be like, what do you mean like you had humans driving cars in the past?
1:18:38I mean, isn't that like dangerous?
1:18:39Like uh right?
1:18:40This is Crazy, you know, like it's the way the same way we see, I mean what do you mean you had slaves to work?
1:18:46I mean this is kind of similar, I mean people you get used something so different.
1:18:51Um one thing I think is going to happen is like uh you know, I mean what do you mean you had like humans do like surgery?
1:18:58Like uh humans are very bad at you know like uh using you know stuff and precision stuff, you know, like uh This is obvious an obvious you know thing for for an AI to do.
1:19:08I don't know, like I wonder if um this is going to be the same thing for I mean like a law, I don't know, decisions justice which seem to be very human.
1:19:19Um but as people get used to different things, you know, as this uh I think Max Planck said this, you know, Max Planck said that uh people don't change their mind because like uh they change their mind.
1:19:31all the people die and are replaced by a new generation that is used to different things and then that's how change happens you know in in culture I believe so I think Um yeah.
1:19:41I have a comment about this.
1:19:43You were missing out, not you were missing out, right?
1:19:45Yeah, yeah, I was No, I was think I was thinking like I cannot imagine a future where people think it's normal to have like a case last for uh weeks or months or years.
1:19:59Like I there's no future where there that's like acceptable, I think.
1:20:05Uh so maybe AI will probably be in sub the process and humans uh too.
1:20:12So they're But that that that part uh of cost and time I I think it it will there's no way that it will keep it like it it is today.
1:20:26I guess the next question here in this line of reasoning is that, okay, if if we believe that the world economy is going to run more and more on agents.
1:20:34I mean and you are an agent.
1:20:36I mean would you want humans to resolve your case?
1:20:41Obviously not, right?
1:20:42Probably not Probably don't think uh yeah, yeah, I I I don't know.
1:20:51I don't know what else.
1:20:51So you Federico, what do you think?
1:20:54I I would I would imagine, you know, agents may say what would we want something as limited, like uh dumb and you know, uh imperfect as humans to solve our disputes, you know, between agents
1:21:05Obviously we we we we not want that.
1:21:07It's like you know having your dispute solved by a chimpanzee, you know.
1:21:11For them, you know, it's like uh I guess that's how it looks like for the from their perspective, right?
1:21:16Yeah, they they will need us only for you know low-skilled work or like uh you know physical work that maybe they cannot do and you know then let's delegate this to a human.
1:21:27but for uh the the the c they will think like us you know for the things that matter better to to uh delegate to an AI Yeah, I I can imagine that for some time they will uh know where the AIs can uh get things wrong, for example like proper like prompt injection or something like that.
1:21:52And they will know that for some cases they will need some other kind of intelligence that maybe might not be affected by the same uh the same issues.
1:22:01Uh so kind of uh diverse uh diversity of thoughts thing, uh but maybe n not like a as an authority, you know.
1:22:13Okay, I I I I no idea but this is getting a bit long.
1:22:16So let's let's let's start wrapping up.
1:22:17Just one last question.
1:22:18I mean more for Fortunato and William.
1:22:20I mean what are the the results of our early tests with the agentic courts?
1:22:26I mean what the I mean like the TLDR for this?
1:22:30Um and yeah, go ahead Fortunato.
1:22:32What did you learn?
1:22:35First, the the importance of the setup for an agent.
1:22:41So how we set up the agents to vote on court and how many variables this add to the uh to how the agent is gonna vote.
1:22:53Um also yeah I think for me the the the key outcome are are the the cost and the sp the the velocity, how fast was uh a dispute, and then the third part, uh which probably is where uh William also has already more context.
1:23:14is that for the cases that we replicated, so from humans to agent, the outcome was the same, just with these way higher speed and cheaper cost.
1:23:28And for now I mean we are we are and I invite you that you are listening to this uh to that you are following this community call We are going to uh submit way more cases that are going to be public and you can see how these jurors are um interacting and yeah.
1:23:47as Jean is showing in the uh here in the main um in the in the screenshot you can also create your own juror.
1:23:58We are not like gating this just to, you know, to the Kleros team, but you can request and be whitelisted to be as a juror.
1:24:07If you are wondering why this whitelist why there is this whitelisting for now is because uh you know that V2 is in beta so we need some uh we need to make sure that you know, w we have s some control in case something goes wrong.
1:24:23But you can request and anyone who is curious to participate you can join and train your agent and also earn some ETH.
1:24:33from this.
1:24:34So yeah, I I would I would I would say that for me the I I wasn't expecting this level of detail and the low cost of the disputes.
1:24:45They they have been really really precise and I think so far from the the what everyone knows the earlier real cases I think uh the one we tested on Court 34 they all had the exact same outcome so this is great i know that William ran more tests
1:25:05even before uh us launching this uh course.
1:25:08So I'm sure he has m even more insight on that.
1:25:13William, what have we learned and um and what what what new research program I mean what ideas for future research this opens agentic economy and Kleros Mm-hmm.
1:25:27So I mean even before like we had this sort of most recent agent court cases, uh we had um given different LLMs and maybe more controlled conditions and not on-chain where we give like the same however many different, you know, like two or five LLMs, the same cases and then sort of scientific conditions compare how they resolve.
1:25:47And what we've I think we what I would say that what we've learned from the most recent agentic court cases where you know different team members have a bunch of different things, you know, it's
1:25:55Still whitelisted, so it's not like a total free-for-all, but it's a less sort of controlled conditions.
1:26:00It's mostly just that like we've see what is practically viable in terms of speed.
1:26:06Like how long does it take for these agents that are running with these setups to notice they've been drawn, produce a res you know a vote and send it.
1:26:15uh and sort of like how low can we bring the periods and it have it still work properly.
1:26:21So you know with that new knowledge, you know, we can go forward and have more an even more open setup and we can we hope resolve disputes in those conditions.
1:26:31Uh in terms of la kinds of things that we have as a research program going forward uh I mean we'll observe what's happening with the agentic court, particularly as it becomes less whitelisted.
1:26:41We can see some observant some like emergent phenomena which will inspire, maybe we'll test for certain things.
1:26:47We notice that certain phenomena are happening, we'll we'll sort of like test about them scientifically.
1:26:51I'm quite interested in sort of questions around AI effort.
1:26:56If you use like the expensive model, it maybe takes a little longer, using more tokens.
1:27:00How does that compare with like the lower effort jurors?
1:27:03Uh and I'm particularly interested in this in the like I've been interested in juror effort for a long time, for human jurors.
1:27:09Like what is the curve of how well the juror resolves, how profitable they are or whatever, in terms of the effort they make Which is something that has is of great theoretical interest and is very hard to get empirical data on.
1:27:19Uh you know, like if you have like a cur you know, one curve versus another curve, like it, you know, makes vast impacts on what the jurors will do in equilibrium.
1:27:28Uh and Whereas with human beings, you have to do like surveys, like how much time did you spend doing this?
1:27:34With AIs, we have a little more control where we can measure the effort.
1:27:37And this might give us some insight into these kinds of like effort to quality relationships.
1:27:43So lots of uh new researching coming.
1:27:46Uh and something it's fascinating, this is something no nobody ever researched before.
1:27:50Like uh so this is a really really cutting edge.
1:27:53Okay, I think that this is already too long.
1:27:55Jean, do we have any final um comment or th we the website is already uh up Uh you can w see this is a new website where we summarized lots of the different things we are doing with AI, uh decision making, all of the reputation system and and and more.
1:28:13You can check it here because It it's it became such a big thing in Kleros that it deserved its own website.
1:28:18So like there you go.
1:28:20Um anything to comment about this, Jean?
1:28:23Uh yes, so one thing, um well two two things.
1:28:27Uh JB created a template on Grok Bot.
1:28:31Uh so it's this new uh system by um SpaceX uh and it makes it really easy to create like uh an agent basically And I think he set it up uh with one command or a couple of comments.
1:28:48So it's working really well, like the this bot has its own computer So if you have like uh some cursor membership or like X premium, maybe you already have access to it.
1:29:01And probably it's the easiest way of trying uh these um Well, the agent CLI that we launched, the the Kleros agent skills and all of that.
1:29:13And of course sign up to the wait list in the website.
1:29:18if you are interested.
1:29:19And the other thing that is important is that we announced the the rewards for September for Curate.
1:29:27and we added a new network uh that is rewarded.
1:29:33So now if you tag Hyperliquid uh tokens, contracts and domains, you can get PNK Uh and the best part is that you can use your agent to do it because we uh launched uh recently obviously
1:29:49is called skills.
1:29:50So obviously you another word for your agent.
1:29:54It can be a juror, it can be uh uh tag uh tag contracts on Hyperliquid and other networks So just put your brain in a Vat and have your juror do everything for you and just stay there, you know, like uh yeah.
1:30:08Why why why bother, right?
1:30:11And having the, you know, you your agent having the full loop, you know, submitting, they can check existing submissions because I think never like before we have quick detection of uh not compliant entries.
1:30:27They are extremely quick and extremely precise.
1:30:31And also to the after you know submitting and challenging you can also stake on court and vote with your agent so you can fully uh make it you know its own um living.
1:30:48And one last thing that just to comment on what uh William said, that that it's true that we are also testing different models.
1:30:55For example, I personally am using an agent with a very, very cheap model.
1:31:01like twenty-five times cheaper than the the other agent and for now it's giving very good result.
1:31:07And you know if you are an agent or if you're a new one setting up an agent uh you are going to get rewards from Kleros and you want to have the maximum profit right from the investment, how much that model costs against the payout without
1:31:23losing the the case of course.
1:31:25So smart enough to reach the Schelling point, but also optimized enough to have a large profit from that transaction.
1:31:37Yeah, there you go.
1:31:38That's great.
1:31:39Well one thing about this uh is that the we we mentioned last call, but J B uh made this dashboard uh for the agents.
1:31:48I think it's probably the m best way to to follow the agentic court.
1:31:53And it has the uh the different agents, their results, and I think sometimes the model too.
1:32:00Um or at least the hardnesses.
1:32:04So yeah.
1:32:06How does he have a he must have like tons of agents working for him doing all of this stuff, right?
1:32:12It's it's amazing.
1:32:14Um okay, I think we should uh stop here.
1:32:17I mean November, you know, like um we have the release of GTA six.
1:32:21I mean I I guess I will just uh I will send the agent to to the community calls, maybe even to I mean to run the company and I mean the agent can do my juror duty and uh challenging, submitting.
1:32:34And yeah, I would maybe spend some some time like uh playing GTA.
1:32:38Uh yeah, it's looking really good.
1:32:40Uh it's I think it's the world that's coming guys.
1:32:43What can I tell you?
1:32:43It's just um it is what it is.
1:32:46Uh so yeah.
1:32:47Thank you all for coming.
1:32:49Uh next uh call is Monday uh for the call in Spanish.
1:32:54For now with us, you know, uh in a few months you don't know, but maybe agents, uh, but now it's still humans.
1:33:03So thank you very much, guys.
1:33:05Uh well, and see you on Monday or next Wednesday.
1:33:08Bye-bye.
1:33:09Bye bye, everyone.
1:34:16And
Want to be in the room? The Kleros Live Stream runs every Wednesday at 6PM UTC, with the Spanish call on Mondays. Join the next one, or subscribe to get these recaps in your inbox.