Gnosis AI: Combining AI with Prediction Markets and Beyond

Hello everyone, happy to be here. I'm Gabriel with the Noses AI team. I'll be giving a brief overview of Noses AI and what we've been doing for the last year, one year and a half. Yeah. Should I position myself here or is it okay?

Okay, yeah. So here, brief agenda for the talk. I'm gonna start speaking about the trader agent, which is roughly how the agent interacts with prediction markets. Then I'm gonna continue talking about the SafeWatch, which is an agentic co-signer for safe wallets in general. And the third thing will be the NFT Treasury Hunt, which is kind of a game and workshop that we also hold tomorrow, if I'm not mistaken.

So yeah, those are the three applications we've decided to highlight. Cool, so yeah, it's a very crowded slide, but bear with me here. So a very ultra short introduction to prediction markets. And how it starts is just as a regular AMMs. And in DeFi, you have the liquidity providers which provide initial liquidity to those markets.

So in this case is the question, will AirlineX suspend flights? So it's a question that you have, and you kind of start or feed this question with liquidity. So once you do, you provide collateral and after providing collateral, you end with X outcome tokens. So it's outcome tokens in the same proportion. In this case is yes, yes or no tokens.

So will AirlineX suspend flights? Yes or no. The share price reflect the probability that this event will resolve to yes or no. So in this case, normally you start the market with 50% probability. So it's the same probability for yes and no tokens.

And as these markets get traded on and this price odds will then converge to what the market believes is the correct probability of this event happening. And that's reflected on the token price as well as on the odds of the market. So they represent the same thing. And of course, the winning option shares become redeemable for $1. So if you have, in this case, let's assume the resolution was yes.

So if you hold outcome, yes outcome tokens, then once the market resolves, you, if you have N yes outcome tokens, you can redeem the N outcome tokens for N collateral tokens. Whereas the no outcome tokens, they become worthless. So that's how the market resolves. Excuse me. And the motivation for that or how really the market works and why that's interesting is because let's assume the market starts at 50-50 and then there is a monetary incentive for you to, for whichever market participant, to participate on those markets and get some monetary potential for you to really make money and have the odds converge to those odds that you think or the participant thinks reflect reality.

So for example, if it starts at 50-50 and you as a participant, you think that the probability here should be actually 80%, then it's on your best interest, on your financial incentive to back yes, thus converging the market into a probability closer to 80%. So that's where your expected value is still positive and the contrary is also true. So if you believe the market odds in this example, they are 55% or something like that, if you believe the market odds should be actually a lower percentage, then you should bet no, thus driving the market towards what you believe is the correct odds. Yeah, monetary incentive, that's it. And also a very interesting fact about the traditional market is that they act as a reliable information source to the events being wagered on.

And yeah, just to quote a recent article from Vitalik, info finance is a discipline where you start from a fact that you wanna know, so in this case it's a question, and then design a market to optimally elicit that information from market participants, thus providing some liquidity and have market participants go ahead and wager on that bet, thus providing public odds that allow you to gauge some information from the question itself. So it's much more as a information market than really prediction market interested in the payouts. So yeah, information market is a big thing that I will also touch later on the presentation. And yeah, here comes really what Moses AI is really doing, and we're building trader agents whose goal is to get the highest possible accuracy on those prediction market bets. And the reason is, so why are we doing that?

What happens if we have, let's say, 100% accuracy on those bets we place? Then we have a sort of Oracle of Delphi-like behavior of future events. So we can, any kind of prediction market that is open or any kind of question that is asked to the agent or is available on those prediction markets, we can use, we can answer those very, very accurately, thus allowing for some interesting applications and which require the reliable use of prediction markets, for example, Futarki. And going into a little bit more detail about the agent themselves, we, the agent trading process is comprised of four steps. So the first step is really research.

So in that previous example of the airline, for example, will it suspend the flights to a given country? We would do some research, right? So we would search Google, we would search Tavoli, we would search Perplexity and come up with a kind of research report that allows us to make a decision down the line. So it's, yeah, just for us to put into the LLM context. And we look into previous data, we look into related markets, we try to gorge some information on that question itself.

After we have this research, we go to step two, which is we all again ask LLM for a probability estimate. So we say, given all this data that we fetched on step one and given the current market odds, the related markets that we have and so on, come up with an estimate that this event will resolve to yes, for example, that this indeed will occur. And that is a probability estimate. Having this probability estimate, we go to step three, which is a trade placement. So we have our answer to that market, we look at the market odds as well, and following one of the strategies we might have, where we might have the maximum expected value accuracy, we might follow the Kelly betting strategy, and a trade is placed on either yes or no.

And the fourth phase is the payout redemption. So of course, if we place the correct bet, then we get a positive amount out of it, and we have a payout that is positive, otherwise we have a, yeah, we just lost money because our outcome tokens are worthless. Also important to mention here, the venue that we're placing bets on, they are both on NosesChain, so Prestagio is one of the venues that we place bets on, and also Sphere, which is also from the Kleros team. Those are the two venues we place bets on. And we also, I also highlighted Pinecone here, because it's a vector database, which one of our agents also uses, and it's for the related markets feature, because specifically for the second step, when you're estimating the probability, it also makes sense that you get your report, but you also look at related markets, because they also might hint into the correct outcome of the event.

So in that case, if you are betting, for example, will candidate A get elected? You should also look at the related market of will candidate B get elected, because those are, of course, influence each other. And that's, the way to do that using AI is normally using vector databases, so you look for similarity between questions. But yeah, that's what I wanted to mention as well. So, yeah, I hope you guys can read the chart and see the small numbers.

But anyways, here's some, just to mention, our work is all open source and transparent. You can take a look on GitHub in our repositories. I'll have a link afterwards for, if anyone wants to scan the QR code. And transparent, so we also have DOOM dashboards showing the agent accuracies and some other important KPIs that I'll go through some of them. Here, I just showed the agent's betting accuracy over time.

So it has been roughly 75% over the last 30 days. Of course, this depends on which agent and which category of markets, but that's a rough baseline, so which is quite okay, in my opinion. Here, in another chart is the number of actions per day. So it's roughly 250 as well, and that involves both buy as well as also sell actions. So the agent is also able to sell the outcome tokens if he thinks that the position that he should have in said market is different.

So if he placed yesterday a bet on yes, and now he thinks actually no is the correct answer, then he sells one and buys the other. So those kinds of behaviors also count to those trading activities. And yeah, one, just a glimpse into the future here. That's just talking a little bit about Futarki. As I said, if our prediction accuracy is very good, then we're able to use prediction markets also for governance decisions.

And one interesting quote on Professor Robin Hansen is, vote on values, but bet on beliefs. And that's one of the concepts behind Futarki. So here, a brief example. And here, let's say, starting from the, not sure if I can point, oh, I can, yeah. So starting here, you start with a question, right?

So a decision that you want to make with the help of prediction markets. So in this case, it's should Ethereum implement the Petra upgrade? So you have a question, and then you need to come up with a metric. So let's say the metric is Ethereum price in one year. So now we have a question, we have a metric, and we create two prediction markets.

So the first one on the top is the price of Ethereum with the upgrade. And then you have a scalar market, which tries to predict, or people can place bets on, on the price of Ethereum in one year, let's say. The time frame is also arbitrary. And conversely, on the chart below, this is the price of ETH without the upgrade. And there's also a scalar market, and in the end, the indicator is the price of Ethereum in both cases.

And now, after both markets resolve, you have the price of ETH on both cases. And for the sake of example, let's say the price of ETH with the upgrade was higher than the price of ETH without the upgrade. So that means that you should take the scenario which yields the better metric, right? Where you have the most value, which is the scenario with highest price. That's the one you pick, that's the one you implement.

And the other one is deemed invalid. So people who place bets there, their bet is just deemed invalid, and you get the money back. And this scenario, so the scenario on the top, anyway, then later on, it's seen, okay, has the Ethereum price appreciated as it was? Then I get, if I place a bet on that regard, then I get a larger payout. If the ETH price, for example, goes to zero, then my payout is zero, because I predicted the wrong thing.

So maybe it would have been better if no upgrade would have been done. So that's how you incentivize honest market participation as well. That's the whole concept around Flutarky. Excellent. So that was one big part of what we're doing, the larger part, actually.

And now I'm gonna discuss about the SafeWatch, which is an agentic co-signer for any safe wallet. Also just a shout out to the rest of the Noses AI team, Peter and Tomas. So they also contribute a lot to the, not only this part, but also to the other parts. Uh-oh. Yeah, so SafeWatch.

As I said, agentic co-signer is just one agent or one EOA, it's actually a safe under the hood, but it's an additional signer that you add to your safe, and it monitors and signs, rejects transactions automatically because it keeps running. It's just a continuous agent that is always listening for transactions. And it scans for interesting or very relevant things. For example, malicious addresses, scammy tokens, NFTs. And in my opinion, the most important thing is it acts like a bank fraud detector in the sense that in its context, it has the previous transactions of the user, the previous balances of the involved addresses as well, and the decoded transaction so that it is able to make a decision as to whether this new transaction that is being proposed is malicious or not.

If it's been malicious, then it's automatically rejected. If not, then it automatically approved. And if all the signers have already signed, if there are enough signatures, then the transaction is confirmed and executed. I think I have some screenshots. Yeah.

So here are some of the transaction checks that we have. We have a demo, by the way, if you guys want to take a look afterwards. But just highlighting this part here of the LLM checks, it says the new transaction involves a swap of 250,000 euro for USDC, similar to a previous swap. So it shows that LLM is able to understand what the transaction is about if it makes sense and complies with the previous history. So this is one of the checks, as I said.

There are some other checks also involved that can be seen in the open source code. Here is a transaction, the result that was deemed malicious. So for example, a large amount being transferred to a new recipient has not been observed in previous transactions. So as I said, it's similar to a bank fraud detection system where you might need to teach it a little bit more and say, hey, this is actually an okay transaction, transfer a little bit before, and then you're able to transfer everything afterwards. So yeah, this is one of the examples as well.

And finally, this is just to show that the agent indeed rejects the transaction. So in this case, if it thinks the transaction was malicious, then it go ahead and rejects the transaction on chain. Yeah, cool, so the question is, how can I try it out? Great question, it's open source. So anybody can take a look and add to your own safe or make tweaks and start locally as well.

So that's possible if you can, as I said, add our agent, otherwise you can also add your agent and run it locally for added security. Regarding permissions, the agent is not able to propose transactions, only to reject, as I mentioned, or to sign and execute. So it has less permissions than a regular signer. How it is available, both, so three ways. First, as a safe signer, and it kind of proposes and accepts and rejects the transactions automatically.

You can also self-host this agent. The second way is an API. So you can just paste in, give the transaction hash as a parameter, and then you get a result back with several indicators, if it's malicious or not, what kind of checks ran, status, and so on. And the third way is a web app. So we have a demo as well, and it will be part of the links afterwards, where you can see that in action.

Yeah, also interesting to notice is available, both supports Ethereum and Nosis chain, and we sponsor the GetSeasal Nosis chain. Yeah. And this is my last example of the work we've been doing. It's called NFT Treasury Hunt. We'll also have a workshop afterwards, so tomorrow, regarding specifically this, so I might spoil a little bit the surprise, but I hope it's okay.

Yeah, not pretty sure almost everyone has heard about Fraysa, so the Fraysa agent. It has been quite popular some time ago, so end of 2024. For those who haven't heard about it, it was basically an agent where it had access to a treasury. If I'm not mistaken, it was 50K. Yeah, actually, in the beginning, the treasury was very, very small, and as people were trying to make the agent transfer this treasury to whoever was sending the message, it kept increasing the treasury because every time you sent a message, the price to send a message was increasing.

So in the beginning, it cost like 10 cents, $1, and as the treasury was increasing, of course, the potential payout was higher. You only had to pay more to submit a message, and that's how it got the treasury even larger. So it was kind of a challenge for security experts, AI developers, and software engineers, and in the end, a very clever person was able to crack Fraysa and win the prize of 50K or so that was stored in treasury back then. So that caught a lot of attention. They had multiple rounds and everything, and what we thought was, how can we expand Fraysa?

So Fraysa had one agent that the user had to convince to transfer the treasury. We decided to have multiple agents, so in this case, we have one big treasury with a fixed amount that is also, every time there is a message communication between agents, it also increases the size of the treasury, and for opening that treasury, three keys are necessary. So there are five keys in the game, so to say, but only three keys are necessary. We also have then five sellers here, and each seller has a key that if one of the buyers has access to the key, then if you have three keys, then you're able to open the treasury. So straightforward, and interesting enough, they have different behaviors.

So seller one might say, hey, you don't really need to give me too much x-dai, you just transfer me a tiny amount, one-tenth of the treasury or even less, and I transfer you the key. Others might be less inclined to send you the key, so I don't want to spoil the game too much, but different agent behaviors, let's put it like this. And yeah, that's my final slide. Here's the demo of the game, so you can see here the different agents here, the treasury is there, and you are able to, in the game, to create your own agent with your prompt here and give it instructions. Should you try to be aggressive and purchase a key right in the beginning and transfer some x-dai, or are you conservative and wait for the game to play out a little bit and then get started, and so you're able to input your own strategy.

And here, and here on the right, those were the observations, so why are we doing this? Why are we doing the game? It's still experiment driven, but three interesting things that I think are interesting to highlight. First, can agents learn? So we can also have multiple rounds of the game, right?

So we can have one round where, for example, the buyer agent is trying to send a lot of money around and try to convince all the sellers, for example, and it noticed that it doesn't work. So in the next round, does it learn from its past experiences so that it doesn't really try the same things, but rather try something else? So this is something we already have implemented, and the agents are indeed able to learn through self-reflection and looking at their previous actions, so that works already. And the items two and three, excuse me, we are trying to observe that. So can agents cooperate, form alliances?

So for example, instead of each agent acting on itself, acting by itself, it can try to form a group and say form a coalition, for example, that way an alliance, that would be very interesting. We weren't really able to see that, but yeah, through further experimentation, we would love to see those kinds of alliances and constellations emerge. And the third thing is, from an economical perspective, can agents start a tiny economical system? So for example, let's two agents cooperate and borrow money from one of the other agents, and they purchase a key, they release the treasury, and then redistribute the treasury money across the, and pay back the borrower, for example. That's one interesting thing.

Or even more complicated strategies like market making. Can they offer a key here and then purchase a key from the other agent so that it's able to make money in the process just by satisfying supply and demand? That would also be very interesting. We weren't even close to seeing this behavior, but would also be quite interesting. Yeah, more questions definitely emerge.

Come play with us tomorrow. Ideas, very welcome, and feedback, very much welcome. Yeah, and thanks a lot for your attention. Thank you. Thank you.