Featuring Jan Czarnocki.
I didn’t think this would end up being a debate, but I think it was quite interesting and worth a listen!
My talk at Jan’s event in Switzerland:
Transcript:
[00:00:10–00:00:34] All right. Welcome back to The Supercycle. Today’s guest is Jan from Elastics AI, which is a startup working on AI agent tooling for prediction market trading.
[00:00:34–00:00:48] And I know Jan because his company had an event near Zurich in May, and they flew me out there. We stayed at a hotel in Zurich.
[00:00:48–00:00:58] And I gave a little presentation at the event about insider trading in prediction markets, which you can find on our site, supercycle.blog.
[00:00:58–00:01:04] And that was a great event. So hopefully we’ll be able to do that again sometime.
[00:01:04–00:01:13] Hello, Eli. Pleasure to meet you. Yeah, it was fun. It was fun in Switzerland. We got great photos and your presentation was great.
[00:01:13–00:01:18] I’m happy to see you here and hear more about your progress. Where are you? Where are your thoughts?
[00:01:18–00:01:29] And I’ll be also happy to tell a bit more about Elastics, where we are, what’s our approach to AI agents, and where we see everything going from here now.
[00:01:29–00:01:39] Yeah, great. Yeah, I have some questions about that. I think should be pretty interesting for our listeners here. So if I can start off, that works for you.
[00:01:40–00:01:42] Sure. Just shoot at me.
[00:01:42–00:01:47] All right. So you trade geopolitical markets, right?
[00:01:49–00:01:50] Well...
[00:01:50–00:01:52] Oh, yeah. About trading.
[00:01:53–00:01:57] I’m doing my best trading, let’s say. But like, sorry, I interrupted you.
[00:01:57–00:01:59] All right. Well, yeah. Okay.
[00:01:59–00:02:10] What’s an example as a trader of a trade where you might trust your own judgment over your agent’s judgment, and maybe a trade where you just trust the agent over yourself?
[00:02:11–00:02:18] Well, what’s kind of the line there between what the agent is good at and what still needs a human in the loop kind of managing it?
[00:02:18–00:02:31] Yeah. So just for the record, I’m an aspiring geopolitical trader. I’m doing my best to learn. It’s very, very hard, actually. And it’s very, very hard to reconcile full-time startup job plus serious trading.
[00:02:31–00:02:44] But it helps to trade even a bit to empathize with the problems that traders are having, basically. And there’s still a lot of issues and things to be solved on prediction markets and also from the tooling perspective.
[00:02:45–00:02:54] And in general, as you hinted already upon, our core thesis is that AI agents are amazing. AI is amazing.
[00:02:54–00:03:16] But we really believe that AI is as smart as you are, not more, no less. So the usefulness of AI is the property of the skill, IQ, wisdom, knowledge, everything of people who are using AI, basically.
[00:03:16–00:03:45] And my idea is that judgment. Judgment when you’re trading, when you’re not a quantitative trader. So basically, if you’re not relying on models and fast trading and, let’s say, formalizable, let’s call it like this, rules of, okay, I buy this and just operate on very fast intervals in time.
[00:03:45–00:03:46] And space.
[00:03:46–00:03:57] And space. Then, unless you’re a quant trader, basically. It’s not yet automatable, basically.
[00:03:58–00:04:09] But, so, if you’re a quant trader, AI is here and now very helpful because it can analyze data faster and it can, it makes stuff faster, basically.
[00:04:09–00:04:21] But there’s a whole world of macro traders and discretionary traders who are like, okay, I think inflation will be like this. Okay, I think there won’t be a peace deal with Iran.
[00:04:21–00:04:29] Because they read some article, they read a lot of books, that’s what they feel and that’s where their edge is.
[00:04:29–00:04:36] And they, let’s say, time horizon is broader than next few minutes or next few seconds.
[00:04:36–00:04:39] And they’re just not operating on these repetitive models.
[00:04:39–00:04:45] And here the question is, where AI can be useful for this kind of people?
[00:04:45–00:04:51] And our answer is basically that AI can extend your cognition.
[00:04:52–00:04:53] It can help you with your judgment.
[00:04:53–00:04:55] It can pile up all the data.
[00:04:55–00:04:56] It can pile up all the news.
[00:04:56–00:04:58] It can help you to browse stuff.
[00:04:59–00:05:08] It can help you with simple workflows with, let’s say, defensive positions or defensive quitting from contracts.
[00:05:08–00:05:09] It can alert you.
[00:05:09–00:05:12] It can be an extension of you.
[00:05:12–00:05:16] So we are building a tool in which you land.
[00:05:16–00:05:18] You don’t need to move anywhere else.
[00:05:18–00:05:25] And you have everything you need as a macro trader, discretionary trader, geopolitical events trader.
[00:05:25–00:05:27] You have everything you need to trade.
[00:05:27–00:05:29] You have news.
[00:05:29–00:05:31] You have AI with the best models.
[00:05:31–00:05:34] Let’s say we have plugged in CloudFable.
[00:05:34–00:05:38] Right now it’s quite expensive, but that’s what we are plugging in.
[00:05:38–00:05:40] We are using Cloud, but you’ll be able to switch models.
[00:05:41–00:05:43] Of course, it will be consuming a bit more tokens.
[00:05:44–00:05:49] But still, we are giving you a specialized tool for trading.
[00:05:49–00:05:52] So imagine Cloud, but for prediction markets.
[00:05:52–00:05:59] And you may ask, okay, I might as well just connect some MCP servers or whatever new sources to Cloud,
[00:05:59–00:06:00] and I’ll be the same.
[00:06:00–00:06:05] No, because you need stuff that is, let’s say, on-chain, aware of what’s happening on-chain.
[00:06:05–00:06:13] It needs to be secure in a particular way, and it needs to have tools very specifically made for trading.
[00:06:13–00:06:18] And we believe the first step in finance is helping you with the judgment,
[00:06:18–00:06:25] and then you can build agentic workflows on top of it when you build enough trust and enough context.
[00:06:25–00:06:29] And also a person who’s wielding this AI is now what it’s doing.
[00:06:29–00:06:37] So we are very much not as this crazy AI bots on Twitter claiming that, oh, we made so much money just on arbitrage.
[00:06:37–00:06:44] I mean, it’s tempting to go there, but we think that lifecycle of these tools is quite short, basically,
[00:06:44–00:06:52] because whatever quant strategy you come up with, you will exhaust this alpha structure later, of course.
[00:06:52–00:07:04] I think part of what Elastics has been framing their product as part of marketing and stuff is that every prediction market trader
[00:07:04–00:07:12] who’s not automating their workflows and using AI for data collection and to kind of automate their process,
[00:07:13–00:07:19] all of those traders are at disadvantages because they’re wasting their time on stuff that could easily be automated.
[00:07:19–00:07:28] So let’s say Elastics was wildly successful and every prediction market trader was using it.
[00:07:28–00:07:31] Doesn’t that disadvantage kind of neutralize?
[00:07:31–00:07:37] And then if everyone has these institutional grade AIs that can do all of this fantastic research,
[00:07:38–00:07:40] where does alpha come from after that?
[00:07:41–00:07:42] From your brains.
[00:07:43–00:07:44] Literally.
[00:07:44–00:07:47] Well, where are you ingesting the data from?
[00:07:47–00:07:58] I mean, the thing is about this event on prediction markets is that the reason why quant models are so good at predicting,
[00:07:59–00:08:07] let’s say, stock prices is that stock prices, commodities and everything on financial markets is fairly formalizable
[00:08:07–00:08:12] because these are, let’s say, discrete entities with certain mathematical properties.
[00:08:12–00:08:14] So let’s say price, yes?
[00:08:14–00:08:23] And you can fairly directly take historical data and can learn something about this price given this or that event.
[00:08:23–00:08:30] But then in a bigger scheme of things, you also have this whole financial global order,
[00:08:30–00:08:36] which assumes that you have respect for contracts, you have an open trade system, and there’s no wild stuff.
[00:08:36–00:08:39] So with every quant model, you have so many assumptions.
[00:08:40–00:08:44] But to have this quant model, you have to have this formalizable assumptions.
[00:08:45–00:08:50] And the thing with prediction markets is that with events, historical data are close.
[00:08:51–00:08:56] They’re not, let’s say, patterns formalizable into quantitative model.
[00:08:56–00:09:07] So the thing is that it’s about sharpening your judgment about what will happen.
[00:09:07–00:09:10] Because with events, it’s about human psychology and stuff.
[00:09:11–00:09:17] And with a lot of events on prediction markets, it’s a judgment of one guy, let’s say Donald Trump,
[00:09:17–00:09:21] under certain systematic, historical, however you call it, forces.
[00:09:21–00:09:25] So there’s less math in this, basically.
[00:09:26–00:09:30] So that’s why AI, and AI is running on maths.
[00:09:30–00:09:35] It’s probabilistic engine, or however you will call it.
[00:09:36–00:09:40] And we humans, we are biological entities, so we feel realities.
[00:09:41–00:09:46] And maths and AI, it filters a lot of qualitative stuff out of reality.
[00:09:46–00:09:54] So my point, basically, is that, however, in the current AI paradigm, when you have large language models,
[00:09:55–00:10:00] there is a limit to which they can be good at forecasting events.
[00:10:00–00:10:07] I think they can just, you know, maybe run you, propose you alternative scenarios as to what may happen under,
[00:10:07–00:10:10] and let’s say some assumed probabilities.
[00:10:10–00:10:15] But then, at the end of the day, you need to make, you need to make, you need to judge.
[00:10:15–00:10:20] And they are only as good as the data based on which they are trained.
[00:10:20–00:10:23] And it’s also a matter of judgment, what data do you choose?
[00:10:24–00:10:27] What kind of, let’s say, international relations theory do you follow?
[00:10:27–00:10:31] Or what kind of mix of sources or commentators do you follow?
[00:10:31–00:10:38] So there’s a lot of fine-grained ideas which sets prediction markets apart from financial markets,
[00:10:38–00:10:43] which make them less quantifiable and less quant-like.
[00:10:43–00:10:49] Unless you’re doing short-term contracts on Bitcoin, short-term on prediction markets equities.
[00:10:50–00:10:51] I’m talking about events right now.
[00:10:51–00:10:58] And I think I’m focusing on events because I truly believe that the biggest social and economic value of prediction markets
[00:10:58–00:11:00] is in events forecasting.
[00:11:01–00:11:06] Yeah, I’d actually like to counter that a little bit.
[00:11:07–00:11:12] Because I think that we’re already seeing, especially with the new Frontier models coming out,
[00:11:13–00:11:17] we know Fable came out, was taken away, now it’s back.
[00:11:17–00:11:21] And GPT 5.6 Soul is coming out tomorrow.
[00:11:21–00:11:30] And if we look at here, I can pull up the latest Metaculous Cup.
[00:11:31–00:11:40] And what’s been happening is that with very little human input, just the forecasting models that are ingesting data
[00:11:40–00:11:46] and they’re being trained by whatever company is making them,
[00:11:46–00:11:51] they’re actually becoming more accurate than a lot of the humans.
[00:11:52–00:11:55] Okay, so I can see.
[00:11:55–00:12:00] I mean, they are better at what is measurable, at what you can measure.
[00:12:01–00:12:04] But at what you can’t measure, they are not.
[00:12:04–00:12:05] And you don’t even know.
[00:12:05–00:12:09] So it’s a matter of what you measure.
[00:12:10–00:12:11] And a matter of what kind of question you ask.
[00:12:11–00:12:13] You can measure the calibration.
[00:12:15–00:12:16] Against what?
[00:12:16–00:12:16] Against what?
[00:12:17–00:12:18] Against real world events?
[00:12:19–00:12:20] Yeah.
[00:12:20–00:12:26] My point is that whatever is quantifiable, these models might be good at.
[00:12:26–00:12:26] That’s fine.
[00:12:27–00:12:28] I’m fine with this.
[00:12:28–00:12:30] Because these things might be quantifiable.
[00:12:31–00:12:35] But for the things which are not easily quantifiable, which are qualitative,
[00:12:35–00:12:36] they’re not.
[00:12:37–00:12:37] And they won’t be.
[00:12:37–00:12:41] And I think that events, a lot of events happening in the world,
[00:12:42–00:12:45] they’re just not that much quantifiable, basically.
[00:12:45–00:12:51] You know, it’s like this with Laplace’s demon and Goodall’s theorem.
[00:12:52–00:12:56] Just some things, you just cannot, you just, it’s because the assumption would be
[00:12:56–00:12:59] that you can just pile up the whole universe into the model,
[00:13:00–00:13:02] which is self-contradictory, just impossible.
[00:13:02–00:13:05] Okay, well, I was just pointing out here.
[00:13:05–00:13:08] So if we see forecast by a forecasting research institute here,
[00:13:09–00:13:15] the median forecast of the super forecasters, they ranked the accuracy.
[00:13:16–00:13:21] So this score is the Breyer score index across all of the questions,
[00:13:21–00:13:25] where 100 is perfect, 50 is predicting 50% of the time,
[00:13:25–00:13:28] and zero is maximally wrong.
[00:13:28–00:13:29] So a higher score is better.
[00:13:29–00:13:36] You see, super forecasters got an average score of 69.2 with the best forecasting model,
[00:13:36–00:13:39] only 0.1 point behind that.
[00:13:39–00:13:44] Wouldn’t you expect with the advancement of AI progress that they’ll get better scores
[00:13:44–00:13:47] and begin to beat human forecasters with this?
[00:13:47–00:13:49] They’re basically there already.
[00:13:49–00:13:55] I think that they’ll be good at enhancing your judgment and sharpening your judgment, basically.
[00:13:56–00:14:01] And that’s pretty much about this, basically.
[00:14:01–00:14:03] Because they’re operating on historical data,
[00:14:04–00:14:08] and they just are not aware of stuff that they haven’t saw.
[00:14:09–00:14:12] Well, they can just ingest as much as they want.
[00:14:12–00:14:12] Of what?
[00:14:12–00:14:13] They kind of have free range.
[00:14:14–00:14:16] They can ingest as much data as they choose to.
[00:14:16–00:14:17] But they don’t have eyes.
[00:14:17–00:14:18] They don’t have eyes.
[00:14:18–00:14:19] They don’t have ears.
[00:14:19–00:14:20] They don’t have skin.
[00:14:20–00:14:21] They don’t feel the reality.
[00:14:21–00:14:22] They do have internet.
[00:14:23–00:14:25] They have just digits.
[00:14:26–00:14:27] Whenever there’s a geopolitical event,
[00:14:28–00:14:30] you’re not flying to the Strait of Hormuz.
[00:14:30–00:14:31] You’re reading the news about it.
[00:14:31–00:14:34] You’re reading the same stuff that the AI agent is also reading.
[00:14:35–00:14:37] Yeah, but I can talk with people.
[00:14:37–00:14:39] I can hear you feel the mood.
[00:14:39–00:14:40] Yeah, the people are also reading the same news.
[00:14:40–00:14:43] Why is the agent not getting it?
[00:14:43–00:14:45] But he doesn’t understand the news.
[00:14:45–00:14:46] They don’t understand the news.
[00:14:46–00:14:47] Why not?
[00:14:48–00:14:51] They’re calculating the probability of the good answer.
[00:14:52–00:14:53] It’s something different than understanding.
[00:14:54–00:14:57] They’re not understanding the practical impact
[00:14:57–00:14:59] of whatever they’re doing on the real world.
[00:14:59–00:15:01] They’re just computing probability.
[00:15:01–00:15:05] And we are not computing probability when we think.
[00:15:06–00:15:07] Yeah, okay.
[00:15:07–00:15:11] But they’re also trained on basically all human writing of all time.
[00:15:11–00:15:12] And so they don’t.
[00:15:12–00:15:15] Yeah, so if you stop human judgment, they will be useless.
[00:15:17–00:15:20] Because you won’t have any more qualitative data.
[00:15:20–00:15:22] They will just hallucinate themselves to death.
[00:15:23–00:15:24] Okay, well, that’s true.
[00:15:24–00:15:26] But we’re not at the end of data collection.
[00:15:27–00:15:29] But it will be false data.
[00:15:30–00:15:30] Meaningless data.
[00:15:31–00:15:32] The thing about this data as well.
[00:15:32–00:15:35] If you think hard enough about AI,
[00:15:37–00:15:41] whatever AI outputs is only meaningful because we say so.
[00:15:42–00:15:44] It’s not inherently impactful.
[00:15:45–00:15:47] It’s just kind of a mirror of human thinking.
[00:15:48–00:15:52] It’s like we are projecting our thoughts against AI.
[00:15:53–00:15:55] And then something is outputted.
[00:15:55–00:16:01] And like whether AI makes sense, whatever AI outputs or not make sense or not,
[00:16:01–00:16:03] is only the property of human judgment.
[00:16:04–00:16:09] So we are the evaluators of whether AI works or not, basically.
[00:16:09–00:16:12] Well, we do have benchmarks.
[00:16:12–00:16:21] But your point is that the way that we decide the benchmark is through a human judgment.
[00:16:21–00:16:27] But on the other hand, for forecasting in particular,
[00:16:27–00:16:33] we do have actual ways to measure the agent’s forecasting accuracy over time.
[00:16:34–00:16:34] Right?
[00:16:34–00:16:37] With the prior score and calibration.
[00:16:38–00:16:41] And we can see when the agent is predicting this.
[00:16:41–00:16:45] Well, if the agent predicted something 50%,
[00:16:45–00:16:49] all of those events should happen about 50% of the time.
[00:16:49–00:16:50] That means the agent is calibrated.
[00:16:51–00:16:59] It’s very, very quantifiable whether a human or an agent or any sort of entity is good or bad at forecasting.
[00:16:59–00:17:02] And right now, it looks like the agents are good at forecasting.
[00:17:03–00:17:06] Based on the data humans provided.
[00:17:06–00:17:16] And they were calibrated according to human judgment and human taught them to be good forecasters, basically.
[00:17:17–00:17:18] So they’re just piling up the human judgment.
[00:17:18–00:17:23] You’re saying that they’re good because humans trained the model in the AI.
[00:17:23–00:17:27] If we stop feeding them human interpreted data, they’re useless.
[00:17:28–00:17:28] Okay.
[00:17:28–00:17:29] Well, that’s true.
[00:17:29–00:17:32] But there’s always going to be more money in AI.
[00:17:33–00:17:33] Yeah.
[00:17:34–00:17:34] Yeah.
[00:17:34–00:17:38] But they will need human data, human made data, human interpreted data.
[00:17:38–00:17:42] So they will need still like human written books, interpretation of news.
[00:17:42–00:17:42] Oh, yeah.
[00:17:42–00:17:42] Yeah.
[00:17:43–00:17:43] Yeah.
[00:17:43–00:17:43] For sure.
[00:17:43–00:17:51] At the end of the day, you will always need some kind of human just being there, seeing, interpreting stuff, talking.
[00:17:51–00:17:52] I completely agree.
[00:17:53–00:17:58] I doubt that the human is in your website.
[00:17:58–00:18:04] I doubt that the human chatting with the agent is actually going to help it very much.
[00:18:04–00:18:06] I think it’s more about its training data than.
[00:18:06–00:18:14] No, it’s just my point is basically that they are very formidable and impressive tools, which will be wielded by humans anyway.
[00:18:15–00:18:18] And they’ll be developed by humans and they’ll be enhancing humans.
[00:18:18–00:18:30] I just really, I’m very skeptical about all these narratives trying to anthropomorphize them basically and give them autonomy in a human sense.
[00:18:30–00:18:31] It doesn’t make sense.
[00:18:31–00:18:36] I mean, perhaps there’s a future in which they’ll be trading and forecasting events on polymarket better.
[00:18:37–00:18:37] Okay.
[00:18:37–00:18:38] Good enough.
[00:18:38–00:18:38] Yeah.
[00:18:38–00:18:58] But I think that underneath, and that’s also what we’re thinking about in Elastics, you need to provide, you need some human who will provide proper data, harness, proper harness, context, direction, taste, judgment, however you call it.
[00:18:58–00:19:03] So there will be actually extensions of these humans ultimately.
[00:19:04–00:19:15] Because at the end of the day, you just need this, let’s say, for the lack of the better words, let’s say human discernment, human biological discernment of things.
[00:19:15–00:19:20] That I am here, I’m embodied, and I’m not just, you know, the logical gate.
[00:19:21–00:19:21] Okay.
[00:19:21–00:19:23] Without consciousness.
[00:19:23–00:19:34] It seems like the recapping the disagreement here is whether you just need training data or you also need a human steering the AI at all times.
[00:19:34–00:19:35] But we can...
[00:19:35–00:19:36] Maybe not at all times.
[00:19:37–00:19:42] But, and it’s up to, it will be also the human judgment when do you steer.
[00:19:42–00:19:46] And there’ll be more successful approaches and less successful approaches.
[00:19:46–00:19:54] It just, I just think that humans with AI will be beating just AI.
[00:19:55–00:19:56] Okay.
[00:19:56–00:19:59] Well, we’ll see in the near future.
[00:19:59–00:20:00] Yeah, yeah, yeah.
[00:20:01–00:20:01] Yeah.
[00:20:01–00:20:06] And we’ll need to, let’s say, curate the battlefield carefully.
[00:20:06–00:20:07] But it’s an open question.
[00:20:08–00:20:08] Okay.
[00:20:08–00:20:13] So, let’s talk about the data a little bit that the AI is going to be ingesting.
[00:20:14–00:20:17] How do you vet the quality of the data that’s coming in?
[00:20:18–00:20:33] And how can you prevent, how can you defend against information attacks against the AI if someone is writing, if someone’s like planting headlines or faking posts to try to target AI models to trade on bad information?
[00:20:33–00:20:45] So, what we’re doing in Elastics is that we have curated news sources, curated news sources providers from only reputable sources.
[00:20:45–00:20:50] And we, from, I mean, your question could be a broader question.
[00:20:51–00:20:54] It’s a societal question, but I don’t have a good answer.
[00:20:54–00:20:55] I just can’t tell you what we are doing at Elastics.
[00:20:56–00:20:57] So, we have reputable news providers.
[00:20:57–00:21:09] And we’ll also let you, and that’s what we are working right now, that you’ll be able to connect, choose the Twitter accounts you want to follow and from which you want to have news sources.
[00:21:09–00:21:13] So, it will be also up to you to select the news sources.
[00:21:14–00:21:21] And I think this is the important part that, and it’s also a part of this process where you’re building up the context for your own AI.
[00:21:22–00:21:24] So, you choose, I want to hear Reuters.
[00:21:25–00:21:27] I don’t want to hear this news.
[00:21:27–00:21:28] I want to follow this guy on Twitter.
[00:21:29–00:21:38] And then your AI, you can chat with your AI and perhaps build automations and agents built on your chosen context and data.
[00:21:39–00:21:45] So, I think here it’s very important for a lot of traders to, let’s say, prepare the proper environment.
[00:21:45–00:21:54] And in Elastics, we are creating this environment in which you can, let’s say, imagine that you’re interested in Iran world-related contracts.
[00:21:54–00:21:56] You choose, let’s say, five, six contracts.
[00:21:56–00:21:58] You plug in news sources.
[00:21:58–00:22:02] AI is correlating these news sources and giving, let’s say, impact score.
[00:22:02–00:22:07] And you can chat and then make trades, make automations, make indices out of it.
[00:22:07–00:22:12] Make some agents working based on these news triggers, others for you.
[00:22:12–00:22:14] So, that’s how we approach.
[00:22:14–00:22:21] We don’t want to just give you an LLM wrapper and connect it to Polymarket and good luck, son.
[00:22:21–00:22:35] We just want to provide you a proper environment in which you can create enough confidence and capacity in whatever is happening there that your AI can gradually make bolder moves.
[00:22:36–00:22:38] And you’ll see whether you can trust it or not.
[00:22:38–00:22:39] You can switch models.
[00:22:40–00:22:42] The better model, the more expensive this will be.
[00:22:42–00:22:44] And then you can make judgment.
[00:22:44–00:23:01] Okay, maybe I’ll set up some stop loss or some defensive, let’s say, cut losses automation or maybe indice between this and that venue that will be realizing my portfolio based on if this kind of information hits.
[00:23:01–00:23:05] So, there’s a lot of details there.
[00:23:05–00:23:14] And we are often discussing AI agents in a very, very broad way and understandably because it’s a very hot and interesting topic.
[00:23:15–00:23:18] But then you start to build these things because you’re excited.
[00:23:19–00:23:23] You want to serve this AI wave and you want to fix problems that people have.
[00:23:23–00:23:36] And then you realize that on the ground, it just takes a lot of, lot of, lot of small step and details to make people trust this and make it actually workable.
[00:23:36–00:23:42] So, someone is actually, you know, putting 500 grand on the agent.
[00:23:43–00:23:49] That’s not something, unless you have this quantitative model in big hedge funds.
[00:23:49–00:23:53] That’s not something that’s frequent right now.
[00:23:54–00:23:59] People are still much more into this clicking, checking, et cetera.
[00:23:59–00:24:00] So, it’s a process.
[00:24:01–00:24:03] It’s perhaps the future in which we are going.
[00:24:04–00:24:15] But I really think that the best agents will be, let’s say, the mirror versions of their owners.
[00:24:15–00:24:18] So, you will have Eli, Eli Army of Agents.
[00:24:18–00:24:21] I’ll have my young army of agents.
[00:24:21–00:24:32] Maybe based on our podcast, they can meet in the background and based on what we are talking about, they can be enriched with this context and maybe say,
[00:24:32–00:24:33] okay, Eli, let’s go trade.
[00:24:34–00:24:36] We have this idea and these agents are already trading for us.
[00:24:37–00:24:37] You know what I mean?
[00:24:38–00:24:43] So, I believe in these AIs as an extension of us, basically.
[00:24:43–00:24:50] And this actually is interesting because these are AI agents which really demand human agency.
[00:24:51–00:24:56] So, for agents to work, you need agency because you’re an initiative.
[00:24:57–00:24:58] So, yeah.
[00:24:58–00:24:59] Maybe that’s the future.
[00:25:00–00:25:00] That would be interesting.
[00:25:00–00:25:14] Do you think that – so, when we think about prediction markets, they’re accurate because they harness the wisdom of the crowds to provide forecasts.
[00:25:14–00:25:21] And those forecasts are aggregating lots of independent errors from lots of different traders, right?
[00:25:22–00:25:27] But a lot of the trading agents are going to be prompted in sort of similar ways.
[00:25:27–00:25:30] People have similar ideas.
[00:25:30–00:25:39] But more importantly, they’re all trained on probably the same set of five or six foundation models that you’re using.
[00:25:39–00:25:53] And if agents dominate volume, do you think these humans are going to have different enough ideas to counter the effect that essentially all of the traders have the same brain, essentially?
[00:25:53–00:26:04] So, do you think that could actually make prediction markets less accurate overall as more agents scale to the volume?
[00:26:04–00:26:11] Do you think that there just might not be enough traders because all of the agents are going to have the same ideas?
[00:26:11–00:26:12] There’s not enough counterparties?
[00:26:13–00:26:18] But don’t you think that this is a matter of whether these agents are losing or winning?
[00:26:18–00:26:25] If they’re winning, maybe they will have good streak and then they’ll die or they’ll just die immediately or they’ll be amazing.
[00:26:26–00:26:36] So, I still think that the success of agents will be the property of the humans behind them, how they’re programmed, how they’re trained, what ideas are there basically.
[00:26:36–00:26:44] And because just you will be prompting them, either through code, natural language, however you want.
[00:26:44–00:26:54] So, just imagine this successful hedge fund guy, which is just giving this amazing prompt, basically saying, avoid this, you should do that, ABC.
[00:26:54–00:26:58] Just thinks, I cannot even fathom because I never worked at hedge fund.
[00:26:58–00:27:05] And just, let’s say that he has the best model, just a lot of compute.
[00:27:05–00:27:14] And then he’s just competing against some random dudes who haven’t even read the finance textbook.
[00:27:14–00:27:18] And his prompt is, make me a million dollars, don’t make any mistake.
[00:27:19–00:27:22] So, I guess we know who will win.
[00:27:23–00:27:26] And I think this is, so, AIs are shells.
[00:27:27–00:27:29] They are just like tools, weapons.
[00:27:29–00:27:39] It’s just kind of a bullet that you will shoot and maybe a rifle and just who will have bigger gun and who will have the better gun, who is shooting better, yeah?
[00:27:39–00:27:41] Who have bigger missile.
[00:27:42–00:27:42] It’s a weapon.
[00:27:43–00:27:43] It’s a tool.
[00:27:44–00:27:47] How you calibrate, they won’t be equal at all.
[00:27:47–00:27:50] They’ll have access maybe to this server, to that server.
[00:27:50–00:27:55] Maybe there’ll be some rules discriminating between, let’s say, spamming agents.
[00:27:55–00:27:58] So, it’s an open field.
[00:27:58–00:28:04] But maybe about the efficiency of prediction markets.
[00:28:07–00:28:18] If efficient prediction markets, I guess, are those where you have very high probability of something happening, I guess.
[00:28:19–00:28:19] I don’t know.
[00:28:20–00:28:21] Correct me if I’m wrong.
[00:28:21–00:28:29] Like, efficiency here is a, because at the end of the day, something happened or not.
[00:28:30–00:28:30] Yes?
[00:28:30–00:28:31] Yeah.
[00:28:31–00:28:33] And how do you measure efficiency?
[00:28:34–00:28:37] So, I can better conceptualize what you’re doing here.
[00:28:37–00:28:38] So, you’re right.
[00:28:38–00:28:39] How efficient is the market?
[00:28:39–00:28:41] How accurate was it, essentially?
[00:28:42–00:28:48] Like, how efficient are prediction markets in, how well they are predicting something, basically?
[00:28:48–00:28:48] Yeah.
[00:28:48–00:28:51] So, there’s a couple ways of measuring this.
[00:28:51–00:29:00] The most common way is a Breyer score, as I mentioned a bit earlier, where it’s basically an aggregate of the calibration that shows.
[00:29:01–00:29:05] So, let’s say we take all prediction markets on Polymarket and Kelsey, right?
[00:29:05–00:29:09] Let’s say it’s like 100,000 prediction markets, let’s say.
[00:29:10–00:29:20] And let’s say 1,000 of those at a certain time frame out from the resolution were at exactly 50%.
[00:29:20–00:29:32] So, if the market was efficient, then we’ll see that those markets that fell in that 50% bucket would actually happen 50% of the time.
[00:29:32–00:29:39] And we kind of aggregate all of those across each of the different probability buckets into a single score that scores how efficient the market was.
[00:29:39–00:29:49] So, maybe if agents will be so good at predicting future, they’ll be sitting in some contracts where there is some money to make, basically.
[00:29:51–00:30:02] And the better traders will be either looking for marginal gains or this or that market or will be looking for markets where there is not that much information.
[00:30:02–00:30:09] So, I guess it will be more a property of yet undiscovered information.
[00:30:09–00:30:25] And this is maybe the take that our common friend Adi Rajaprabraharan made, that there might be more and more different and exotic markets about which there’s not that much information ready at hand.
[00:30:25–00:30:34] And maybe there will be people just going there digging this information, feeding it to agents and just having more and more exotic hobbies.
[00:30:34–00:30:39] Or there will be like Citrine researchers seeing whether there’s a traffic in Ormond Strait.
[00:30:39–00:30:40] I don’t know whether you heard about this.
[00:30:40–00:30:46] There was a guy just literally went to see whether the tankers are passing straight or not.
[00:30:47–00:30:52] And he made trade based on this information, I guess, on Polymarket or maybe something like this.
[00:30:52–00:30:55] So, there will be an incentive.
[00:30:56–00:31:09] I know that I cannot directly answer the question about efficiency, but the trend would be then that for the markets about which there’s an abundant data, there will be just priced well.
[00:31:09–00:31:20] And then there will be a strong incentive to just go deeper into more exotic markets to just discover more information about stuff we are just not knowing this much about.
[00:31:20–00:31:24] And just one more thing and interesting dynamics.
[00:31:25–00:31:33] Like if you look at Polymarket right now, the liquidity is where the global attention is because people are talking about it, people are writing about it.
[00:31:33–00:31:37] So, you see Iran, maybe Russia, Ukraine.
[00:31:37–00:31:42] These are fairly, let’s say, efficient liquid markets.
[00:31:42–00:31:50] But then there are other big stuff, big things where, let’s say, the public attention is not yet there.
[00:31:50–00:31:52] Like right now, the issue of Greenland.
[00:31:52–00:31:55] Greenland, it’s still the thing.
[00:31:55–00:31:56] It’s still the thing.
[00:31:56–00:31:57] Greenland is still the thing.
[00:31:58–00:31:59] But there’s not that much liquidity.
[00:32:00–00:32:05] But there will be now liquidity because suddenly Trump’s at Greenland and there will be journalists writing about it.
[00:32:05–00:32:06] And you’ll have reports and stuff.
[00:32:06–00:32:09] So, suddenly you’ll have this, let’s say, liquidity.
[00:32:09–00:32:17] So, the thing I saw with prediction markets is that the liquidity information there is also property of, let’s say, global attention.
[00:32:18–00:32:20] So, where the information is produced and collected.
[00:32:21–00:32:26] So, the people who will be perhaps making a lot of money will be there who will be, let’s say, ahead of the curve.
[00:32:26–00:32:32] That they’ll be not even able to predict what are odds on this particular contract.
[00:32:32–00:32:37] But broadly speaking, where the attention will go given, let’s say, structural forces.
[00:32:38–00:32:40] And there won’t be that much data about this.
[00:32:40–00:32:43] Only, let’s say, broad patterns perhaps.
[00:32:43–00:32:47] Or maybe there will be, like in this previous data, there will be, maybe it will be discoverable.
[00:32:48–00:32:48] I don’t know.
[00:32:49–00:32:51] But that’s what I would say.
[00:32:51–00:33:06] So, do you think that could make it so smaller exchanges that are doing more specific markets, a specific category of markets, could end up having more liquidity and more volume because of agents?
[00:33:06–00:33:12] And do you think that could make it so Calci and Polymarket aren’t really the only large exchanges anymore?
[00:33:12–00:33:15] I think, I think so.
[00:33:15–00:33:31] Also, because Calci and Polymarket seem to be reluctant right now to list contracts which are relevant, very relevant for the general public in order for the US government not to lose edge in its wars or endeavors.
[00:33:32–00:33:39] So, my opinion is that there is less and less interesting contracts from Polymarket related directly to war in Iran.
[00:33:39–00:33:44] So, there are proxy markets like, oh, will be there meeting or not.
[00:33:44–00:33:51] But there are not markets like whether they will invade or much more granular war markets which are very directly informative.
[00:33:51–00:33:54] It’s just we need to make proxies for stuff.
[00:33:54–00:34:05] And proxy markets are, let’s say, telling, there’s some underlying thesis and you need to have bucket of proxy markets.
[00:34:05–00:34:06] So, that’s fine.
[00:34:06–00:34:24] But I would like to have more prediction markets in different jurisdictions that are floating prediction markets which necessarily might not be favorable for this or other government just for the sake of the global public to have an information about this.
[00:34:24–00:34:34] But the thing is, structurally speaking, is that it’s only in the US that you can just, you know, fuel infinite amount of money into unprofitable companies like Calci and Polymarket for them to grow in the future.
[00:34:34–00:34:38] Because you just issue reserve currency so you can just print money and, you know, invest in them.
[00:34:39–00:34:41] So, I hope there will be more prediction markets.
[00:34:41–00:34:51] But I don’t, in short term and midterm, I don’t anticipate anyone as successful as Calci and Polymarket outside of the US.
[00:34:51–00:34:52] Because where else?
[00:34:52–00:34:53] Where else?
[00:34:53–00:34:53] In China?
[00:34:54–00:34:59] They are just, they are jailing professors for telling that the economic data are cooked.
[00:34:59–00:35:00] In Europe?
[00:35:00–00:35:01] Oh, no, it’s a gambling.
[00:35:01–00:35:06] It’s a gambling and we need to regulate everything before we’ll let you do everything, basically.
[00:35:06–00:35:13] So, I think Calci and Polymarket are fine and perhaps there will be some competitors from there.
[00:35:14–00:35:14] Yeah?
[00:35:14–00:35:15] Yeah.
[00:35:15–00:35:20] Well, Laura, I think we can wrap up here.
[00:35:20–00:35:21] But this was a great podcast.
[00:35:22–00:35:27] I’m sure it will be really interesting for our listeners, especially our debate about agents.
[00:35:28–00:35:31] But, yeah, really excited to see where Elastics ends up.
[00:35:32–00:35:34] Hopefully, we could do another one of those events.
[00:35:34–00:35:35] That would be great.
[00:35:35–00:35:36] Let’s do it.
[00:35:37–00:35:38] Let’s do it.
[00:35:38–00:35:40] And always a great pleasure to talk with you.
[00:35:40–00:35:42] And, yeah, it was very, very fun, actually.
[00:35:42–00:35:42] Yeah.
[00:35:43–00:35:49] These talks are inspiring because these are things that are, let’s say, somewhere in my background.
[00:35:49–00:35:54] But you, like, you know, it’s like you prompted me to tell them, like, hey, I...
[00:35:54–00:35:54] Yeah.
[00:35:54–00:35:54] All right.
[00:35:55–00:35:56] All right.
[00:35:56–00:35:57] Thanks, Sean.
[00:35:57–00:35:58] Thanks, Eli.
[00:35:58–00:35:59] Bye.
[00:35:59–00:35:59] Bye.
