Epicenter - Learn about Crypto, Blockchain, Ethereum, Bitcoin and Distributed Technologies - Numerai & Predictoor: How AI Changed Crypto Prediction Markets - Trent McConaghy & Richard Craib

Episode Date: September 20, 2024

Prediction markets were one of the first use cases of smart contracts, yet their popularity only recently spiked, following Polymarket’s social media spread. However, while prediction markets are generally zero-sum games, the rise of AI models trained on large datasets led to AI-powered prediction feeds and hedge funds. Crypto offers a unique array of use cases as it allows data scientists to not only share their data sets and models with complete privacy, but also access decentralised computing and model training. While Predictoor employs Ocean’s data infrastructure to run AI-powered prediction bots on lower timeframes, Numerai developed its own AI hedge fund for stocks, that recently also expanded to crypto (Numerai does not trade cryptocurrencies, and Numerai’s Hedge Fund(s) have no relation to Numerai Crypto).Topics covered in this episode:Predictoor & Numerai overviewPrediction marketsPrediction feedsPrediction markets in TradFi and other use casesImplementation of Ocean’s tech in PredictoorPredictions vs. FuturesMarket participantsNumerai hedge fundNumerai’s trust assumptionsThe role of AIThe evolution of AI and how it might solve market inefficienciesNumerai crypto and how it differs from PredictoorPredictoor x Numerai collaborationsThe future of prediction marketsEpisode links:Trent McConaghy on TwitterRichard Craib on TwitterPredictoor on TwitterNumerai on TwitterOcean Protocol on TwitterOasis Protocol on TwitterSponsors:Gnosis: Gnosis builds decentralized infrastructure for the Ethereum ecosystem, since 2015. This year marks the launch of Gnosis Pay— the world's first Decentralized Payment Network. Get started today at - gnosis.ioChorus One: Chorus One is one of the largest node operators worldwide, supporting more than 100,000 delegators, across 45 networks. The recently launched OPUS allows staking up to 8,000 ETH in a single transaction. Enjoy the highest yields and institutional grade security at - chorus.oneThis episode is hosted by Friederike Ernst.

And so everyone was very excited about it because it was going to be a prediction market. But the special thing about a prediction market is it can actually make an oracle. The money going in equals the money going out among the different people making predictions. So it's zero sum among them. In prediction of feeds, it can be positive some because there's new money coming into the system from the people making money trading. 500,000 transactions per month, about $123 million worth of value per month. We've actually spiked as high as, I think, like, 500 million per month. Data scientists sign up, download the data.

You could say they're actually the consumers of the data, but we're the consumer of their predictions, which is based on our data and the models that they built upon the data. And that ends up being very highly predictive and also lower volatility than individual models. If you're looking to stake your crypto with confidence, look no further than course one. More than 150,000 delegators, including institutions like BitGo, Pintera Capital and Ledger trust Corace won with their assets.

They support over 50 blockchains and are leaders in governance or networks like Cosmos ensuring your stake is responsibly managed. Thanks to their advanced MEV research, you can also enjoy the highest staking rewards. You can stake directly from your preferred wallet, set up a white label note, restake your assets on eigenayer or symbiotic or use their SDK for multi-chain staking in your app. Learn more at chorus.1 and start staking today. This episode is proudly brought to by NOSIS, a collective dedicated to advancing a decentralized future. NOSIS leads innovation with circles, NOSIS pay and Metri, reshaping, open banking, and money. With Hashi and NOSIS VPN, they're building a more resilient privacy-focused internet. If you're looking for an L1 to launch your project, NOSIS chain offers the same development environment as a

Ethereum with lower transaction fees. It's supported by over 200,000 validators, making nosis chain a reliable and credibly neutral foundation for your applications. NOSIS Dow drives NOSIS governance where every voice matters. Join the NOSIS community in the NOSISDAO forum today. Deploy on the EVM-compatible NOSIS chain or secure the network with just one GNO and affordable hardware. Start your decentralization journey today at NOSIS. I.O.

Welcome to Epicenter, the show which talks about the technologies, projects, and people driving decentralization in the blockchain revolution. I'm Frederica Ernst, and today we have a very special episode because I have not one, but I have two fantastic guests, namely Trent McConaghy, the founder of Ocean and Richard Crape, the founder of Numeri. And we're here to kind of talk about prediction markers and prediction feeds. Hey Trent and Richard it's good to have you both on It's really good to come back It's been a long time And yeah We just checked this out

And Richard was on seven years ago Which is like 70 years and like regular years So yeah and Trent I mean you've been on regularly since Yeah I guess pretty common But it's great to be on again And you know always speaking about something different and crazy And it's great to join Richard today too

So looking forward to it. Super nice. So Trent, I think you probably don't need an introduction on this podcast, but let's do one anyways. So you are the founder of Ocean Protocol, which is a decentralized data marketplace enabling data sharing while preserving privacy.

And you recently, a year ago or so, launched a Prick Tour. So tell us about that. Yeah, Predictor is crowdsourced time series prediction. So there is basically two types of users. There's data scientists that run scripts or call them simple bots that basically predict, will Bitcoin go up or down, yes or no, five minutes from now? And specifically this would be, say, Bitcoin, USGTP, and finance.

And when they predict, they actually stake against it. They make money if they're correct and they lose money if they're wrong, basically. and so it's sort of this game among these different data scientists. And on the flip side, traders can come along and buy these signals and use them for alpha in their trading, for different trading bots, etc. And this is really meant for people, you know, running scripts and bots and that sort of thing on both sides, on both the prediction side and the trading side because it is every five minutes or every hour.

So that's what predictor is about. Ultimately, it's about time series prediction. We have focused initially on defy because we see, you know, it's got, it's high volume, and it is low latency. But we have interest in exploring time series prediction in other domains to weather, energy, and more. So that's ocean predictor. Yeah, super nice.

And, Richard, you founded Numerai while back. So basically, it was the first time I kind of heard about homomorphic encryption. I remember reading up on this before the podcast that we did seven years ago, apparently. And I was like, oh, this is fantastic. This is so cool. So basically the idea is that kind of you have, that kind of, it kind of provides like a hedge fund like thing powered by crowdsourced predictions from data scientists.

And kind of they, they kind of, the way that they kind of predict things is on encrypted data. So kind of, so as not to kind of leak any alpha. And so kind of you've kind of come to this prediction space from the Tradify arena. more recently you launched Numerai crypto but maybe let's kind of stick with the traditional Numeri for now. So tell us about that. Sure, yeah, Numerai,

that was the key innovation of Numerai. We found a way to share this data set that we had that was very expensive and we wouldn't want other people to take our data and just go and start their own hedge funds. And it's actually right at the moment the version of the data set, it's really more obfuscated than

encrypted. In the early days of Numeri, we were trying to use homomorphic encryption. It was about a billion times too slow, and then it got to a million times too slow, and so it's still too slow. But I like seeing this many entrepreneurs out there still trying to apply homomorphic encryption and crypto. But yeah, the important thing is that no one can see or know what the data is on Numerai, even though they can still model it. Like all the structure is there, but they can't know what the features mean or what the stocks mean. And Numeri has grown a lot since then. We've become basically the biggest online community of people approaching this problem. And there's millions of dollars staked on models. He uses stake their models. And we've also traded billions of

dollars of volume in our global equity fund. So it is quite remarkable how much it has kind of worked since, you know, there's definitely been a lot of bumps along the way, but it has pretty much we're doing the exact same thing we were doing probably in the last call seven years ago, just at a much bigger scale. Yeah, super cool. I assume most listeners will be familiar with prediction markets. maybe let's give like a super short summary.

Who of you wants to do that? I'll give a bad one, trying to give a better one, but I'll scratch some of the surface area. You know, one of the things that's interesting with Numeri was I had just invested in and met with Joey Krug, who started Auger. So you have to be maybe quite OG in crypto to remember Auger. but Auger was something that came out shortly before Ethereum was launched. And so it was like the type of thing that was a project that was going to happen on Ethereum

once Ethereum had launched. And so everyone was very excited about it because it was going to be a prediction market. But the special thing about a prediction market is it can actually make an Oracle. And so it's a way of getting information that's off-chain. on chain and that is a really nice property of it because it's not just about hey people are staking and betting and trying to win money that's not the cool part the cool part is that that market surfaces new information auger was my first experience with that and you can see some indication of that in sort of being something a project that did kind of inspire numerai to have

users stake on their predictions so that the quality of those predictions were go up. So brief, not very brief, but yeah, intro to prediction markets. I can perhaps compliment that. And of course there were two OG projects on prediction markets, one, you know, on American soil and the one on German soil or Berlin soil, to be precise. And that's, of course, with, and was, yeah, Nosis with Frederica and Markin and the rest, of course. And being in Berlin, we got to know the NOS folks quite well and still do, of course. And, you know, similar to Augre, they were going for prediction markets. And kind of the way that I see prediction markets are, there's a broad variety,

but the way, you know, generally I'll describe by example, let's say you want to know, what's the chance of Biden winning as president? That's zero these days probably. With the chance of Harris winning as president versus Trump, right? And then what if you can put your money where your mouth is? So, and so you put, you know, you can bet on this. You can put down $100 and maybe a bunch of other people have put down, you know, $100 here and there for up and saying, you know, Harris and other people have put down their money, $100,000, whatever dollars for Trump. And of course, if here's a key thing in prediction markets, if more and more people are betting on Trump, then your potential amount that you can win if Trump actually does win is lower.

whereas let's say 10% of people are going for Trump and you bet on it on Trump, then you could win a lot more, right? So you can view it sort of like there's two tokens, one token for betting on Trump and one token for betting on Harris as an implementation. And the value of those tokens fluctuate with time. And that also, the market cap of the tokens total tells you the relative chance of them winning, right? And it's sort of put your money where your mouth is.

But the signal itself of whether Trump or Harris will win is public. and people don't have to buy it. And so it's basically this sort of zero-sum game among people predict Trump versus Harris, and they can speculate and buy and sell along the way. So it is an article, like you say, and all the things you say, and it happens to have some of these characteristics. It tends to be for one-time predictions, who will win as president, and then with betting along the way all the time,

with the price of the bets adjusting based on the supply and demand. and eventually later it resolves. So that's kind of how I see it overall. Probably we can elaborate more later. Probably also long-winded, but hopefully it helps to set the stage as we start talking about prediction feeds. Fantastic.

So I will not make a third, maybe I'll make a third one, a third explanation. You guys can synthesize. So in a way, it's kind of a thin. So a prediction market kind of is a trading venue for synthetic assets, that kind of refer to usually future events

where the payout is one if it's true and zero if it's not true and then if you have market dynamics along the way that kind of reflect the probability of that event becoming true or not becoming true. So that was kind of like the compliment, I hope. So let's talk about prediction feeds. How do they, because kind of in a way

they are similar to prediction markets in a way that they help you anticipate future outcomes. How are they different? Yeah, they are very different but there's definitely a lot of similarities. Numeri receives

equity predictions. So it's not like anybody is saying I think Apple is going to be worth more than $110 on this date and so doing it post. stock on numerali. It's it's 5,000 predictions on every equity in the world that is tradable with enough volume. And so it's just sort of set up not on a per asset basis on numerai. And it's also happening every day. So it's it's really like we're getting a feed of predictions,

a stream of predictions each day on these different assets. And so it makes it a bit different than the event-based type prediction markets. And I can perhaps elaborate. So definitely the feed is probably the key to finding characteristics. It's not these one offsets over and over every five minutes, every hour, every day, whatever. Besides that, at least in the implementations that Numerai and Ocean have of prediction feeds compared to prediction markets, prediction markets, thus the feed itself is public, and it's a zero-sum game. among the participants that are investing.

In prediction of feeds, the feed is private unless you pay for it. In the case of Numeri right now, there's exactly one customer for that feed, and that's Numeri, the hedge fund. In the case of Ocean, people have to come along and pay money to get access to a feed, typically a 24-hour subscription to access a feed. And that tends to be traders. So basically, that part of the feed, prediction feeds, the feed is public, prediction markets, that signal of, you know, wind versus losing.

etc. is public and no one has to pay for it. And then this also means that if it's private, you know, part of the reason it's private is so that people can make money on the other end or you get them to pay for it, of course, right? So in the case of Numeri hedge fund, it's implicitly paying for it as part of the overall flow in the case of, oh, should I mention the pay for. So, but yeah, the second part here is zero sum. So imagine with a, yeah, a prediction market, It's the money going in equals the money going out among the different people making predictions. So it's zero sum among some people who predict well make money, other people lose money. In prediction of feeds, it can be positive some because there's new money coming into the system from the people making money trading.

So if it was just betting among the predictions every five minutes, there's sort of people getting money, people getting slashed, etc., back and forth among the predictors making predictions. but when you have this revenue coming in from people paying for the feeds or paying you paying for your feeds based on the profits then that flows back to the predictors and then they make money right so it's a positive sum game with for the the data side is the field making predictions within prediction feeds so yeah to summarize then it's prediction feeds are feeds not one-time events they are typically paid for and private not public and it's a positive sum game among the people making predictions versus a zero-sum game in the case of prediction markets.

But you can actually turn prediction markets into positive sum games the same way, right? You could pay wall them and kind of have a bunch of experts kind of place money on things and kind of then incentivize the experts afterwards depending on kind of like how they did, right? So I mean, so basically I think that's kind of like maybe an arbitrary distinction in that kind of that's how they're used, but it's not imminent to kind of what they are. It is a little bit maybe in that to know the price that the current odds are, to know if you're going to participate in this prediction market against Harris, you would really need to see the current price.

And so you maybe wouldn't want to like bet blind without or have to pay to see the current price before you actually participate. And to be honest, I would see like you can, you know, view prediction feeds as iterated prediction markets, but you have to change three characteristics, right? So generally when you change one engineering artifact, one technology artifact by several factors, if it's sufficiently different, then it probably is helpful to have a new name so as not to confuse it with other things. So we thought long and hard when we were launching predictor and asked, you know, is this a prediction market?

And, you know, it chatted with Robin Hansen and with nosis folks, et cetera, and realized it probably does make sense to call this something unique, right? and talked it over with Richard and stuff too. And that's why we've just, you know, really focused on using this label prediction feeds. But we're not saying, you know, we're not going to be really just about this saying, it's definitely not a prediction market. It's just much easier to think of them as two different things

with those characteristics as the defaults. And then, you know, there can be a blurring as you toggle characteristics one way or the other. And probably the main distinction is the one-offs versus the feeds. Yeah. kind of one of the things that we haven't talked about yet is that for prediction streams, the data is often actually generated by algorithms or AI models. And I mean, this can also be the case for prediction markets, as we see with kind of like

NOSIS AI stuff, but it's the norm for prediction fields, right? I would say that's right, yeah. It's definitely the norm on Numeri. You can't provide any predictions that's like based on your impressions of Apple or something. You really have to build a model on all of the equities in our data universe. So definitely very much leaning towards AI models. And there's no value to human insights onto stocks on Numerai. So prekshire markets, the idea of them has been around for a while.

and they never really took off until very recently. And we can talk about why that is so maybe a little bit later. Tell us about prediction feeds in traditional finance. Are they things that kind of existed pre-numeri and kind of Web 3? Yeah, I would say they have. I mean, in some sense, even signing up to Bloomberg, it's a subscription to a feed of data. it happens to be the live prices, but it's kind of an estimate. No one has all the details of the

live prices of stocks because there's dark pools and there's all these small details. Many data vendors would have different price numbers and things like that. So that's in some sense the same type of model. And then many data vendors would sell signals. And they'd say, we've made this thing. We've used machine learning and we've used a data set. We use the datasets. We use the data sets. we don't have a hedge fund, but we are going to sell this for $50,000 a year. And so that was kind of common. But in some ways, those often were very low quality products. And so why would you sell something?

If it was so good, why wouldn't you trade it yourself? And what's so nice about the staking mechanisms in Ocean and Numerai is that it's saying, this actually is high quality because I've staked a lot on it. And you can see its high quality because look at this track record I've developed. And so that's where it starts to make a particular sense for a blockchain use case. So if you look at systems that kind of consume prediction feeds, can you kind of give us an idea which domains they span? Is it just financial or is it also?

So I could imagine things like weather forecasting prediction, is like can I travel a specific passage with my shipping vessel? Or I could imagine kind of like insurance-related things that kind of could be predicted and tracked and so on. So is it mostly financial or do we also see them in other venues? I could talk to that. Sure. Yeah, so, you know, when we were exploring building predictor,

we actually looked at a whole bunch of different domains. and I yeah we found that there were prediction feeds in many domains and typically quite often they had different labels in different domains and different ways of consuming so for example if you've got an iPhone you open up your Apple weather app and it tells you you know what the forecast is for you know sun versus rain etc one hour from now two hours from now and so on so you can do that as a prediction feed if you want right and it's rolling so one hour from now it's going to tell you the feed one hour in the future and two hours in the future from that, right? And under the hood, there's a lot of data and modeling going on, right? Weather prediction is a huge sub-industry,

but we actually see that it's mostly via governments, right? So governments, they pay national forecast services, things like NOAA as part of the USA and more. And then that outputs predictions as APIs that get consumed by things like Apple weather and stuff. And Apple might have its own AI too. So that's one example within the weather domain. In other domains, it's much more private. So for example, in energy forecasting, it turns out there's an $80 billion industry for management of energy. So this is basically the software and systems around it to manage energy, you know, balancing the grid, all of this sort of thing. And a subdomain within that is a forecasting of energy demand and energy prices.

But it's highly regulated. This is where energy trading is, you know, of Enron fame, for example. But it's highly regulated. So you don't hear much about it, and it's very opaque, very hard to see. You know, if you talk to people in the energy space, they can tell you a bit about it. But it exists. And so if you look around, you know, this exists in various places. You know, very, very different shapes and sizes.

You know, Bloomberg kernels versus your Apple iPhone versus some rest API, right? And we realize that given that where, you know, we realize that where, you know, we're in the crypto domain, whether the application is crypto or tradfey or otherwise, it made sense. There's a lot of prediction markets are well-known term there, so it makes sense to have a term that compared and contrasted what a prediction market is versus this stream, this feed, right? And that's why we chose that label that way, as opposed to just calling it forecasting, because there's different stuff going on too, right? You know, as you call it forecasting, it's like, okay, well, how does that differ from traditional

Web 2 approaches, et cetera, et cetera? So it's useful to sort of put a stake in the ground for this qualitatively new thing in many ways. You can view it as even a, you know, a new crypto building block, a new atomic primitive, if you will, that can be used in crypto, whether it's, you know, shaped the way that Numerai has this technology or ocean technology or otherwise. Can you give us an idea of kind of how the ocean technology is used? So give us an idea of kind of what the things are that are being forecasted, who are they being forecasted by, and who are the consumers of these feeds? Yeah, so our focus for starting is DFI.

Like mentioned, we did look at a lot of different industries and, you know, I talked with a lot of different people. And we saw in terms of the different industries, our criteria were what is the amount of volume that can go through and what's the latency, you know, if you predict well, how quickly can you turn around and make money from those predictions? You know, so for example, if you're predicting medical interventions, you might, it will take you five years, seven years for those predictions to get SD approval, et cetera. So you're talking like a, you know, a clock rate of one clock cycle of five years, say, versus if you're doing, say, gaming,

you know, marketing for gaming, you know, prediction of marketing metrics is actually super useful. Maybe you can get that down into a month or maybe even a week. But then if you start thinking about trading, tradfire, D. Fire. otherwise, you're looking at, you know, minutes if you want, right? And you can even go to two sub minutes if you want to get in the realm of high speed trading or MEP extraction, right? The MAB searchers. So it's, but the point is it's very high value flow and it's very low latency. So from that, we given that, Ocean had a pretty thorough Web3 stack. We had a lot of stuff around, you know, decentralized data sharing, all that, you know, tokenized data access control,

including selling data feeds. So we had a lot of off-the-shelf tech. then we, you know, started with, we used that as the starting point and then basically built an app slash mini stack on top called predictor. So at the heart of it, you know, whereas, yeah, basically the technology is at the heart of it, it's a smart contract running on a confidential EVM chain, so a privacy preserving EVM chain, Oasis Sapphire. When we went live, it was the only one in production and it's, you know, it's been in production of a year and a half and quite stable. and the OAS system has been great to work with. So that is privacy preserving. It happens for using TEE right now, Intel SGX, but that keeps expanding into or is on path to expand to other technologies.

So myself, as a predictor, a data scientist, I submit a prediction, say that Bitcoin, UST, Pair, and finance will go up five minutes from now. I stake, say, 100 ocean. And, you know, maybe Richard, he predicts that it will go down and he stakes, say, 200 ocean. Okay. And maybe you, Frederic, you stake 100 ocean going down. So now you've got 300 ocean going down, 100 ocean going up for five minutes for now, Bitcoin, US, TEPR and finance. Anyone who has bought that feed can see that. They'll see that it's a 25% chance of going up, 75% chance of going down, and they'll see the total stake up and total stake down, right? 300 down, 100 up. So that's

what happens. And then when the actual true signal emerges five minutes later, then everyone gets paid out or slashed accordingly. So let's say it goes down. Then I get slashed. I lose my 100 ocean. And you guys get my 100 ocean. Richard would get two thirds of that. Frederick would get 100. So Richard would get 66, Frederick would get a hundred. So let's say that there's sales from traders buying this and maybe it's another 100 ocean worth of sales. Then Richard would get another 66 ocean from that and Frederic another 33. And you would get your stake back from before to you're outslashed. So to summarize, I predicted wrong. I lost money. I was slashed. You guys predicted right. You made money. So you got my slash stake and you got your cut of the revenue. And how much you got from me, from my stake and from the revenue, that's pro rata to your estate. So that's how ocean works at the very heart. That's all in a smart contract, relatively small. I think a few other lines of code.

All running on Sapphire as a ERC 20 is solidity smart contract. And yeah, it's ERC22. So it behaves like an ERC20 token too in some ways, which is pretty neat. And that's actually for the subscriptions, right? So you buy a subscription. When you buy that, you get 1.0 data tokens, basically, that gives you access to this feed, right? So that's what Ocean Predictor is at the very heart. And then you're the only, because of the OASIS, privacy tech, you're the only one, once you buy this, you're the only one who can see it. Correct. Yes. At first, yes. And then that's exactly it. So there's hand-shaking with Oasis to make this visible to the person who's bought it. And there can be more than one buyer right now, right? you know, we'll see in the future, because if too many people come along and buy and take advantage,

it might diminish the value of the signal, of course, but we haven't seen that yet. We expect to, we hope to. It should happen game theory-wise. Explain to me why I can't resell. So kind of like, say, I buy this token that kind of lets me see the feed. Yeah, that's a great question. So, you know, you could, but this would add a lot of complexity to the system. And if you get closer and closer, you have to submit your prediction, five minutes, are greater than when the prediction is due, right? So if you start wanting to speculate and buy and sell and resell in the four minutes, three minutes before, then the prices would change, etc. And that gets you closer to prediction markets. And that's fine. But we decided to keep it very simple and straightforward,

where it's just, you know, every five minutes boom, every five minutes boom. And you can have a rolling too. It could be every minute you have five minutes ahead. Every minute you have five minutes ahead, rolling like that, right? But I can only break five minutes ahead. I can, but, but I can only break five minutes ahead. say 24 hours ahead? We actually have it right now where it's five minutes ahead and it's 16 minutes ahead. And we didn't, you know, right now, you know, it's the systems a year old. We've been focusing our energies towards making sure that everyone in the ecosystem can make money. So we haven't bothered expanding beyond that.

We will see. Once you get beyond 24 hours, then it makes much more sense where humans might want to intervene. If it's every five minutes, it would be very tedious for humans to interview. And like Richard was hinting at, right? So robots getting involved. But beyond 24 hours, I'm not sure how much value there is in a prediction feed versus a prediction market. So prediction feeds are, in my view, tuned better for less than 24 hours. Basically, you know, they can be much shorter timescales.

And by not having a speculative aspect in between, it makes it really clear how all the incentives work, etc. But who knows, maybe in the future it will be better to have that, right? Just the system that design they have right now doesn't have that. how does this materially differ from a futures exchange, right? So kind of say, if I think about say Derbit, and kind of there's kind of like the prediction for, I can see that as a prediction of the USC Bitcoin price 24 hours from now, whatever the intervals are that they offer.

Couldn't I just kind of use that as a basis and then hedge? Yeah, so absolutely. So they're very complimentary. If you squint, you could even call it a future, you know, a type of derivative that happens to be futures-ish. But I would say it's different. And probably the best example of how to illustrate it simply you can use predictor prediction in order to buy, to make a futures bet.

And in fact, we actually have a lot of successful experiments internally where it is betting on Binance, futures exchange, et cetera, right? So if you squint at when I'm making a prediction, I am making a bet on something happening five minutes from now, say. But it's a slight, you know, the shape of that bet is different than the shape of a traditional future. Of course, if you say you're just doing derivatives of stuff in the future, then it's completely wide open, right?

So you can view that this, that the prediction feeds we talk about are especially shaped type of derivative, right? But it makes it different than the futures as well. Yeah, that's how you see it. And I guess the other thing, and I know Richard, you can probably call it. comment on this. In the way that it's shaped, it makes it very easy for non-traders to engage because they're not having to put traditional trading money and to trade against the price itself or the future price. They're in sort of a more direct way. It's sort of indirect,

allowing data scientists to engage on, you know, was my prediction accurate yes or no, without having to engage in the trading more directly. And in some countries, some jurisdictions you're limited from doing that too, right? So, and for many people, it's just much less scary to operate on predictions. You know, people that like to participate in the data science competitions like Kegel and so on, you can view this like a data science competition with teeth, right?

Just before we pass back to Richard, tell us about the people or the AI's kind of issuing the predictions and the traders consuming it. Do you have kind of like a user profile? Yeah, partly, I mean, we have the user profile we targeted, but we don't.

exactly he was doing it because it's fully decentralized and all anonymous, etc. So this is all on chain. We get stats and if you go to DAPRater, we have a Stapsin Ocean predictor. So we see that there's 76 weekly active users. So you can view it as 76 wallets that are making bets any given week. And maybe some users are running 10 wallets. Maybe it's all a completely individual. We don't know.

And then the volume, it's about 500,000 transactions per month, about $123 million worth of value per month. We've actually spiked as high as, I think, like, $500 million per month. So the market is getting a lot of press around this. We're not really pushing ourselves to get the press. We just know that it's growing. And part of the reason for that is goes back to the target profile, which is we really are targeting the defy traders who know their way around. on Python and data scientists, data science. So maybe you're, you know, an AI pro or a data scientist by trade and you know a bit of crypto, or maybe you're a crypto pro by trade and really know

Python and are dabbling in data science. So you basically have to have some Python skill and some crypto skill to get into this. But we make it really easy from there, you know, going from, you go to predictor.com. On the top right, there's a button run bots. It takes you straight to a read me. And, you know, 20 minutes, you've got a Python script, running that's running your bot on test net, right? And then another 10 minutes, you can have it running on main net, making predictions. So it's really meant to be low friction, but targeting people that demographic, right, people who aren't scared of, you know, some degree of professionalism in crypto and in Python, but we also want these people not just to run the bot straight out by

default. You know, then you can make some money. But if you want to make really good money, you have to have more accurate predictions than the next person, right? Because then you can be basically grabbing their slash stick more often than they grab your slash stake and make money over time from that, right? So that's the demographic. In terms of the traders, it's probably some of the same folks, but even more so, it's, you know, there's a lot of professional hedge funds out there that quietly operate that you never hear about. Sometimes one mad shows, often teams of five, 10, 15 people and, you know, being in crypto for so long, we know quite a few of these different teams, right? And we talk to them and get feedback back and forth. And how we're making sure we do a good job

all this is, or dog's footing it internally. So, you know, internally in Ocean, I'm running a team where it's a bunch of engineers, several and no data science very well, the own Python well, and we've got, you know, some web app capabilities and whatnot too. And our number one KPI inside that team is make money trading, right? And then we pass that off to our users. And by doing that, you know, we can have very good empathy for the users. But it's also, yeah, and a strong incentive for us to just keep getting better, better, better, right? What about you, Richard? So I know who the consumer of your data is, namely a Numrii hedge fund.

I remember when we talked about this seven years ago, kind of you were targeting data scientists. How has that panned out? Yeah, so it's still data scientists. Data scientists sign up, download the data. You could say they're actually the consumers of the data, but we're the consumer of their predictions, which is based on our data and the models that they built up. upon the data. So everybody downloads the data, they train the models, they submit them, and then they stake them. And what we get is thousands of models that are all ensembleed together,

and that ends up being very highly predictive and also low volatility than individual models. And so we're really relying on the ensembleing of the different models to have it, make it, make a good hedge fund at the end of it. And how successful has this hedge fund traded over the last seven years? Yeah, it's been really good. I mean, we're not supposed to really talk about performance, but, you know, you can say some general things. We have very, very good start to the track record.

Even though the company started a while ago, the actual investing side of the business took a while. because we need to build the community, we need to get a competitive amount of data, and we need to kind of pioneer staking on our NMR token. And so once we got all those things right in about 2019, the fund did very well and grew investor interest. A lot of the investors in our fund are, you know, endowments or pension funds or people like that.

The fund is really an institutional product. And so we had a, yeah, we've had a very, very good period of the last five years where I think all but all but one year has been very strong performance. It's been one down year. There's all sorts of qualifiers there. Talk about the volatility and the sharp ratio and all that type of thing, but not going to get into that during this podcast. Yeah, absolutely. So it was very clear when Trent just talked about kind of like how pretty. is set up that everything is super decentralized, right? And I think this is less of a focus

point for Numeri. So tell us about kind of what are the trust assumptions using Numeri and why do you think it's fine as is? Yeah, well, it's, there's some definitely good things and this important things that are required. I mean, we were not hoping for thousands of individual people to start their own institutional hedge funds. To do so would cost like $30 million each. And so in fact, I think a lot of society's capital is wasted on new marginal hedge funds sucking up talent when really you just need one hedge fund that makes it possible for any talents to contribute. And that's really always been numerized thesis. So we like the fact that if you have NMR, we cannot take it back from you, right?

Because it's on a blockchain. That's a nice tool for us. It's much better than having points in the database, having the actual cryptocurrency. We like the fact that for the majority of users, when they're uploading their predictions, they do not have to divulge the IP used to create their model. So their IP really is theirs. We can't take it from them. we like the crypto asymmetry of we can give you this incredible data set that you can train your model on

but you can't steal it from us and use it for your own trading so we're kind of using all these different elements of crypto but ultimately it's much more efficient to have the execution and operation structure of a hedge fund be kind of centralized and and that doesn't mean you know, all these other pieces can't be decentralized. I mean, many users, for example, our users put together a website called Numer Bay. It's like eBay for Numeri, and you can buy and sell Numeri models.

There's nothing we can do to shut that down, and we don't want to shut it down. Like, we don't think it's cool, but we didn't think about it, and that just kind of popped out of the community. And there are many other examples of that, too, where the community's made open source things. There's something called Numer API

that a lot of users use to connect to our API. We didn't even write Numer API. It was written by the community. So all of those aspects definitely makes for quite a healthy decentralized community. But the core investment product, you can't trade equities on a blockchain

legally. It has to be a centralized hedge fund. Let's talk about the elephant in the room, and that's AIs, and AI. models and agents, what roles do they play in your systems? Well, that's, yeah, Trent and I had a lot to talk about when we first met. I think I was maybe Fred Ersum from Coinbase who introduced us.

And we just had a lot of similar background. We were both machine learning people. We happened to get into blockchain, but we both knew AI. And I think we both saw a lot of what was coming in AI. And to me, the magic piece is somehow, is actually the intelligence being built in these systems. Not the user interface or the, all that stuff's good. But it's really like, did it actually work to predict this very hard to predict thing? And with more investment in AI, there's more data scientists.

in the world than they were when Numerai started in 2015. And so there's just so much more going on. There's so many new architectures. Some Numerai users, for example, are using transformer architectures in the development of their models. And so it's just like somehow it's kind of all been a long time coming, but it's somehow all being proved out. And I think a lot of the things Trent had done early on and Numeri had done set aside.

up for this kind of time. And yeah, maybe I can add to that too. Yeah, it's, you know, Richard and I, it was great to be introduced, right, when both Ocean and Numeri were getting going, and we were swapping white papers and giving feedback and so on. We've been in touch ever since. So, you know, I think there's a lot of really cool, quiet sort of collaborations that people never hear about, but that actually do affect, you know, what gets built, right?

And that's really nice. While your question, you know, the elephant of the room AI. Yeah. So in predictor, I didn't mention this in the onboarding when you go through this readme, under the hood, the script that you're going to be launching and running on is building an AI model of fly, right? You can have super simple models that are, you know, not even AI, just like a linear regression model or a linear pacifier. But then you can get fancier and fancier with things like boosted trees, XG boost, this sort of thing, or Gaussian process models. And we have all of that out of the box for you to just put in and set your own parameters.

and then you can have to fancier yet with, you know, large neural networks, transformer type stuff and so on all you want. And that's pretty easy to add in with your own without changing the architecture much. And that's nice. Part of the reason, like, we went for a prediction was, you know, in Ocean, we thought a lot about data marketplaces, of course. We built them over the years. And often the things that we built, we saw that people, you know, it was chicken and egg.

Do people want to, you know, do you try to get lots of supply of data first and then get people to can buy it? Or do you want to get, you know, create people to buy it first and then supply? Or do you want to have sort of like one ecosystem where you've got buyers and sellers somehow together? And we realize, yeah, you do want an ecosystem of buyers and sellers together. And then you need to make it really tuned towards that ecosystem. And then we realized like it was a key constraint. And maybe this is obvious to any entrepreneur. But it kind of hit us in the head like a bigger rocks.

You know, you have to be able to make money, right? So we, we, we, so you put this in a loop. it's, you know, debt data, you know, you spend money to buy data or create data. From that, you build a model, you know, and then from that you, from that model, you run inference and you get a prediction. And the prediction is the final piece of data. From that prediction, you take action to make more money. And then you loop around. You make more money, create more data, build another model, make prediction, take action from the prediction, make money, loop, loop. Taking that money made to buy more data, that closes the loop. And overall, then, we call that

the data valuation loop and it's actually been sort of a core thesis for us overall saying, okay, you know, how much money is potentially flowing through this loop depending on the domain? How quickly can you go through the loop? And that's what led us to, well, predictions as a first cut. Why? Because it's much more valuable to be at the tail end of a supply chain versus at the beginning. It's better to be a Starbucks than a coffee farmer in Costa Rica, right? And I grew up as a wheat farmer in Canada, right?

where you're getting a dollar for five bush for a bushel of wheat. You know, that's the equivalent of say, yeah, five gallon pail, so 25 liters worth, right? So for a dollar, right? And then you go to buy a loaf of bread and it's $3 and you know, you can really make, I don't know, 100 loaves of bread from that wheat. So overall the point is it's much more profitable to be at the tail end of the supply chain. And that in the data land, that's predictions, right? And that supply chain is both AI and data.

It's not just a data supply chain. It's AI and data. And so basically numerine and Ocean, I think part of the reason we're having success with these is that we focused on the high value stuff first and then the rest of the supply chain implicitly gets filled in. People are willing to spend the money to do centralized things or locally centralized to themselves. But if you have 50 or 100 of them, then it becomes decentralized to get the data to build the models, et cetera. And you provide, you know, give them some guidance to make that low friction. We do see that as time goes on, the rest of the. this supply chain will thicken out, right? Just like initially, you know, you had basic coffee supply

chain back in the day and you needed coffee buyers, you know, people who want to drink coffee, but then over time, the whole supply chain thickens out. And I see this now as this is how the data supply chain thickens out over time. So working backwards, you know, you can have decentralized model inference, you can have decentralized model training and so on. But it's not a prerequisite, right? There's a lot of great work happening in that. I'm paying close attention. And, you know, we've even thought a lot about doing that ourselves in ocean, but our focus right now is still focusing on the predictions, simply because that's the tail end.

And if we do a great job in that, everything else kind of half takes care of itself, right? So that's AI and the deep supply chain, and of course this will extend. Where I'm most excited up, but where this can extend is weather prediction. Imagine predicting weather for every one of the 500 million square kilometers on the planet. You've got 500 million square kilometers times about seven key weather metrics,

temperature, precipitation, humidity, etc. That's 3.5 billion feeds. So if you start training a model for that, you're going to have a huge-ass model. It's going to be super intelligent. Just not super-intelligent in human-shaped intelligence, but super-intelligent as in sort of all knowing a lot of the dynamics of the world

at sort of a weather time scale, right? So to me, that's super exciting. And, you know, fours of artificial intelligence that the world hasn't seen before, but can be extremely beneficial to society. And that's what I really hope where we can take this technology and along the way, you know,

help to grow a decentralized economy. Do you see the quality of AI predictions kind of leveling off? If it were possible to kind of make perfect predictions that there wouldn't be a market left, right? Kind of the market kind of thrives

on the inefficiencies to a certain extent. So basically reframing this, how good do you think these models can get where kind of the minimum is kind of the noise that's kind of inherently there that kind of you can't get to vanish. Yeah, well, that's a very good question.

It's very hard to know. It will be hard to even write a proof about, you know, this is the ceiling of what you could do. And I think obviously it's dual with data, though. So that's been the story of numeri. If you kind of look at numerize back test results, you might say, okay, at this date five years ago, we're seeing sharp of 1, and then it went to 1.2 after we added some more features.

And now you just keep adding features, and suddenly you add a sharp of a back-tested sharp of 3 or something. And so that's the type of growth you can see even by holding the model constant. So if you hold the model parameters constant and just increase the data, that scales the performance. But you could also scale them both. And so you can have the model get 10 times bigger and the data get 10 times bigger. And then that you get. So basically, I would say like it's impossible to say there's no better way to make a model. But it is remarkable to me how much we've been able to improve on, you know, we would have very, say, tough periods in the market, say, March 2020, where there's so much hedge fund crowding.

and maybe some of the numerine models all went down at the same time. Well, it doesn't seem like if you had those models now upgraded to the latest data, that that would happen again. And so there's always that type of thing happening. But of course, the same story might be happening inside some other hedge fund. So they would have told you that five years ago, the Sharpest won, and then they improved it. And so there is always the specter of competition. but one of the special things that we like about Numerai is that the returns are quite uncorrelated.

I mean, the times that we have found to be difficult for us often being the times the market's doing well and factors are doing well and all the other hedge funds are doing well. And somehow for us, that's a bad. Those are the tougher periods. But that's a very good thing because it's almost the goal of an investor is to combine a set of uncorrelated return streams. and so it's very good to us that we're different and different in a way that's kind of like increasingly better. Do you think transformer models are the best we'll see for this,

or do you think there's going to be another step change? One of the great things is, you know, I'm used to be younger. Really? Did you and me both? I was 29 when I started a numeraline. And I felt like I kind of knew a lot of the frontier things about machine learning and AI. And then I had to become a, you know, as somewhat of a jack of all trades being the CEO of a fund that had a crypto and a community. And I actually right now couldn't easily whip up a transformer model myself.

But luckily, my past self plan for this moment by creating an open hedge fund where anybody who suddenly is an next. expert in Transformers can add that intelligence to Numerai to the extent that it can be added. And so that's kind of how I think about it. And I would say that, yeah, the one best user on Numerai right now, his username is the human peep. And he claims to be using neural networks and Transformers. Do you think this is kind of the end or do you think there's going to be something better than transformers? I think there will be something better

in a way that's probably a lot better. And I do you know, in some ways Numerai, we are betting on model innovation, right? If we just had, there was one best model, we would just run it

ourselves. But we have constantly seen model innovation be possible not just compute and data but model parameters. and the type of model can have a very large effect, especially in this high, low signal to noise environment.

What do you think, Trent? Yeah, so overall, I think it's actually worth giving a shout to Richard and the Numeri team for increasing their sharp ratio on all these other metrics over the years relatively consistently, right? So that's a really nice demonstrator of the thesis. So really, like, they're not as well known in the crypto space

because they really have focused on their customers the LPs into the hedge fund, but it's really pretty impressive. You know, predictor has been live for a year, and we have seen the accuracy steadily creeping upwards. And that's a sign, you know, there's this argument of people submitting predictions, and they want to make more money.

So that's happening in practice. And, you know, the incentive structure allows for it, right, such that there's, you know, more and more people continuing to try different things and whatever works better they use. And, you know, at the same time with our internal team, And we've also been making advances. You know, we were, we didn't have great solutions to handle non-stitionary data at the beginning, where the underlying dynamical system itself changes.

And, you know, we've cracked that in a pretty nice way now, for example, internally. And that allows us to train on much more large datasets, et cetera. It's a big unlock, actually. And I'm sure that some of our community members have done that, too. None of the traditional time series methods work. So, but I've, you know, I've been in the AI professionally since the late 90s, and I've seen a huge evolution in AI. So there's no reason that, you know, the current state of the art of AI will end.

As a baseline, Moore's Law is going to keep rocking along because there's so much money to be made. And, you know, we're going to get improvements in compute power. We're going to get improvements in data storage. We're going to get improvements in the amount of data. And we're going to get improvements in the AI algorithms themselves. And each of them has their own sort of sub-Mores law and they all combine together. And if you look at the numbers, it leads to about a 10x improvement in AI capabilities year to year to year, right, from these different components. So, you know, Transformers were a nice innovation, and there's some other, you know, things coming down the pipe, you know, things around, um, uh, this is logic around planning, um, you know, making plans of, um, how to, how to decompose a task and then work it out. And there's, you know, some really great planning research going back decades. But, you know, now people are putting this into an LM context, um, with some pretty cool results. And, you know, chains of logic and trees of logic are reasoning. And, and, and, and, and, and,

and reasoning as synthetic data, rationales synthetic data, sorry. So there's a bunch of stuff coming in the pipe right now, you know, that sort of pressures in the last half a year and finding its way into the latest, greatest LMs. But that's still only a piece of it, right? We're going to see more and more and more in the next few years, and it's just not going to stop, right, until we hit, you know, AGI, which is basically par with human across the board and beyond that, too. And because in finance,

Tradfi and Defi, there's lots of money to be made. For sure, you know, a lot of these models are going to find their ways directly into here more and more of two. So I'm pretty bullish for AI. I think it's, you know, there's lots of money to be made in many many domains. And we see that it definitely will happen here. And there's a huge design space to be explored for shapes of AI systems. And we're still only at the tip of the iceberg.

We've got, you know, we've only covered a fraction of what we're going to see. The world's got to change in huge ways. Yeah, I tend to agree. Let's hope that most of them are positive. Richard, you recently launched Numerai Crypto, which kind of veers into the direction of pre-cdo. So tell us about that and kind of how is similar and different to what trend is built. Yeah, so when Trent was building predictories,

like, by the way, we're going to be doing something like this. And then when I was building Numerai crypto, I was telling Trent, you know, we were thinking about doing this. So we definitely, like Trent said, we do feel like this is a collaborative space, crypto. It's amazing how much it works that way. But we just had, with Numeri, a lot of data scientists who signed up over the years. And simply because we've paid them in cryptocurrency, they've learned about cryptocurrency. many of them hadn't when we first paid them.

And so we kind of onboarded a few thousand very bright people to Ethereum by giving them NMR. But the question was, okay, well, numerize getting very hard. Why is it getting hard? Well, you have to add intelligence to a system that's already very intelligent. And so it becomes harder and harder the more models are being added to the ensemble. and we thought, well, let's create Numerite crypto, which is exactly like Numeri, except the universe of stock, the universe of stocks of 5,000 stocks, which is what Numeric normal has, is now just 500 cryptos that are the highest market cap cryptos that are deduplicated and a few other things. And we made that universe and we said, you know what, we don't have. any crypto data. So we're just going to say, bring your own data. Go and find data. You can use

news data. You can use any data you want. But you just have to submit predictions on this universe of 500 stocks. Now, where it's different from predictor is predictor is these five-minute predictions. And many, many firms operate on that horizon, including, you know, like a Renaissance, there would be many models on this short time horizon. But Numerai had basically specialized in making one-month predictions. Now, it's not because we like hate short predictions or something. It's simply because one-month predictions have higher capacity for hedge funds. So most people don't care.

If they can turn a million dollars into two million dollars, they're happy. But a hedge fund that can only turn a billion dollars into one billion and one million dollars, that doesn't actually move the needle very much. So if something's got a stressed capacity, then usually small cap cryptocurrencies have stressed capacity, then it's not so valuable. But if you can trade over a few days, and it doesn't matter how long it takes to get into the cryptocurrency,

well, then you can make quite a lot more money until you have higher capacity. So we decided to create this thing, Numerary crypto, you've got these 500 stocks, everybody started joining and submitting and staking. They're still staking to say that they believe their predictions will work. And we launched it in July. The prediction horizon is 20 days out. And we did another weird thing, which we've never done before, which is we give away all the predictions. So we don't use the predictions because Numerai doesn't trade cryptocurrency.

we just said we'll just give them back to the users that generated them but we're going to give them the ensemble of all the predictions added together and the incredible thing is just how well it's worked there's been almost no sort of one or two week periods of the correlation of the models being negative with subsequent returns and so that doesn't mean it's going to be positive forever. It's very likely to be negative at some point. And this is really an educational, interesting thing. We wouldn't recommend betting all your money on these predictions. But you can go to our website and download the live predictions. And you can start playing with them yourself and see how they work. And see, oh, well, yes, the meta model did predict this stock right,

but it's a very low volume stock. So maybe that's not that valuable. But it also got Bitcoin right. And so that's a very high volume stock. So I could have made a lot of money on that. And so ironically, by the way, there are any Bitcoin maxis out there, but right now the best coin across all 500 is Bitcoin. The second best is ether, I think.

So it's skewed towards high market cap cryptos right now, whereas sometimes it's very different. So that's what it is. And it's just been really fun to watch because it's been a kind of, it's just being kind of full fun. We wanted to have another use case for NMR. And so now $600,000 per day is staked on these predictions. And it's just been a really cool thing to watch. And it's nice to see people using them and thinking about these predictions and trying to maybe turn them into a trading strategy. But for now, Nibirai doesn't have plans to do so ourselves. But if I get this correct, then currently you're kind of paying for these predictions that you're not actually using. So how is that sustainable? Yes, so we pay them from our treasury and it's really to cede the whole thing and see what happens. We don't pay that much. The vast majority of our payouts are going to the equities predictions.

but we are paying some to the crypto predictions to cede it and see what happens. And I think it benefits Numeri if we have more stuff for users to do, more reasons to stake NMR instead of sell it. And we can always do anything we want with these in the future. There are many crypto funds that have, say, recently maybe bought NMR or something.

Now, they might trade. they might trade cryptos and they could ask us, you know, hey, can we get this exclusively or something like that. Now, I hope it doesn't go that way. I wouldn't want to just be a kind of exclusive thing. But right now, it's just very interesting to watch. And it's amazing how good it is. I mean, for a perspective, it's very good to get sort of 2% correlation with equities.

but this has 9% correlation with the subsequent returns of crypto. So it's just very, very much, much, sort of worked way better than we thought it would. And it almost kind of gives me an indication why Trent is excited about what he's doing because he's probably seeing similar things on the shorter horizon. Yeah, I can jump in. Certainly, and I'm very excited for the results.

you guys have had with Numeri on these longer timescales. And yeah, we've seen, you know, in the five-minute horizon, actually it's hard to get a lot of signal. But once you get to one hour, four hours, eight hours, it gets really nice. You know, probably the, you know, it starts to be really nice at about four hours, right? So, but even on, you know, five minutes, you can get on the order of, you know, baseline 52%, but as high as 65% depending on the token and stuff too. accuracy, that predicting up to talent, right? And so, yeah, I mean, overall, I think we both agree that there's a lot of

maybe not low-hanging fruit, but medium-hanging fruit for predictions in Defi. And what we're doing, you know, it's complementary on purpose, you know, in at least a couple dimensions. One is timescales. Numeri is focused on longer timescales. And the second dimension is just the structure and incentives of Numeri versus Ocean predictor. So, Numire is a hedge fund that is that single customer for those feeds.

And so you get paid in that way, that way. In Ocean's case, it's a feed with many, many buyers. And so it's, you know, just different payoffs. You could have, I'm pretty sure there's people that are participating in both the Numurai competitions and the Ocean predictions, you know, for, you know, the 20-day stuff as well as the five-minute stuff and everything in between. That's great, right? And I think those are probably the two main dimensions of difference. And then overall, though, and I guess you'll get to this, Frederick, but, you know, what are the, there's many potential points of collaboration. And we've started going down this path more so, right?

So around the same time Numerai crypto was launched, I think a week later, we launched with Numeri a data challenge called CryptoFactor Modeling Data Challenge. So this is basically some of the stuff Richard was describing, but where people kind of submit what they did into this ocean data challenge describe what they did you know because with if you're just pure numerai you get to stay opaque but if you want you can put this into a PDF describe what you did and then um how well you do as well as your description etc can lead to prize money and it was on the order of 20 000 um or dollars worth the prize money and came a pretty interesting right um and that's yes it just uh it just ended so yeah it was a really nice uh kind of collaboration.

Everybody knows the factors in equities. Like, you know, it's sort of like in textbooks. It's like size, momentum, value, beta, all these factors determine some of the stock returns. But in crypto, it's not actually really known what the factors are. And so by kind of doing this collaboration, this sort of competition to find factors in partnership with Ocean, we got a good. a great deal of new interest in that field.

And if we get good at that, it helps us do all kinds of things, including have better scoring on numerai, where we neutralize factor risks from the, from predictions and so on. So, yeah, it's been a good start to doing something together because we always talk about doing things together,

but we haven't really done it properly yet. Yeah, and one of the things like, you know, Frick, you had mentioned, you know, and pointed that a lot of numerized infrastructure is centralized. So Richard and I talked about this and realized for their business model, there's less incentive to have it fully decentralized. It just doesn't matter as much, right? And that's fine.

And if there was specific aspects where decentralization is really useful, then cool. Now you've got, you know, the ocean predictor contracts that you can basically, you know, numerai can deploy and we can work with. And that's an obvious point of collaboration. But interestingly, because of the slightly different business models and incentive structures, it's not necessarily needed as much, which is kind of interesting, right?

But that also points to then, you know, complementary user basis, at least on the, on the consuming side, right? And on the supplying side, I think it's actually, you know,

people who become experts at one can cross over their expertise to the other and vice versa, right? So, and that's, that's I see what, what, you know, is happening for sure,

and at least one or two people and probably more, and will happen more of the future too. So, yeah, and but because, that, you know, Ocean is now doing stuff in the last year with Criticter are much closer to what Numeri is doing. And Numerai is also

dabbling in crypto. It's the perfect time to, you know, find excuses to work together and then, you know, make extra high values. Yeah, well, tell us more about the potential future collaborations that

kind of you're envisaging. There's so, yeah, I would say that Trent has has stayed ahead of multiple things. So for example, there's no one at Numeri who's an expert at OASIS, which

predictor are users. And you could see versions of Numerai Crypto or something like that where we might want to use that. And therefore, working with Trent's team on a project to do with that could be very, very useful.

And we would be sharing the right skill sets. In some sense, Numeri's team used to be, you know, maybe half blockchain engineers and half website engineers. And now it's like 90% trading and hedge fund execution engineers and researchers. And so it's sort of like the makeup of our company is designed around our customers, which are our investors, which we care the most about and have a fiduciary duty to put them first. Whereas some of the crypto stuff, especially because we've been, yeah, like sometimes we get excited about doing something in crypto and then the gas prices would be 10 times higher than we could do it and manage it and so on.

But that's the way I think it could work. Basically, his team knows more about blockchain. We might know a little bit more about quant finance. Yeah, you guys certainly know a lot more about quant finance. But we're learning and we're bringing in, we find ourselves bringing in more trading and try to fight people too. just to buff up our own expertise. I think a good example of collaborations and where this can go,

I had mentioned earlier in this podcast about prediction of feeds as a new type of crypto building block, right? So we've got oracles, chain link style, where, you know, the traditional definition of Oracle is actually a future prediction, right? But in Crypto land,

it happens to be that a chain link style Oracle is simply a feed, typically of crypto prices or of existing data, right? So if you want to have forecasts or predictions on top of that, perhaps you call it a future Oracle, if you want to think about it as a crypto building block, or if not future Oracle, then simply prediction feed, right?

But once you have that, then there's lots of crypto products that can be built on top. For example, there are quite a few decentralized derivatives projects, right, where people are building different sort of features, trading strategies, options, and so on. all of that could leverage decentralized predictions

quite well, right? Especially once predictor extends to have continuous valued predictions and volume volatility, which is one of the directions for us,

of course. Beyond that, too, if you're doing a loan protocol, you might be able to reduce the collateralization ratio from instead of 180%, maybe down to 120% if you can predict

volatility better, because then you need less time to react. And similarly for stable coins that are based on collateralization. And then also for Dex's, where the LPs are adjusting their weights in things like Balancer or adjusting their threshold in things like the pseudo-orderbook he set up that Uniswap has and so on. All of these could leverage predictions or do a better job. And of course, if you can run this stuff, all as its own new protocol on top that combines these existing protocols, plus this new building block of prediction feeds,

then you can get basically a much more interesting behavior, you know, optimize all of this existing defied infrastructure, like I mentioned, whether it's derivatives or loans or stable coins or dexes and also point the opportunity to new primitives at a higher levels. So that's an example. Yeah, and Ocean can do that on its own, but then we with numerai, we can pull in stuff from, you know, imagine the numerai trad-fi stuff and the numerary expertise.

So there's a lot of back and forth there. Or, you know, Numeri itself building some of these things on top, right? So there's just basically because of where predictor is the Numeri is, we've increased the surface area of possible technology artifacts, right? Increase the size of the adjacent possible, if you will. And from that, you know, Ocean can reap rewards, New York can reap rewards, and any other adventurous builder. So, final question. Prediction on prediction feeds in the next two years.

What's going to happen? What are we going to see that may be surprising? Yeah, I mean, I must say I haven't followed. Is it Polymarket? Yeah. Yeah, I was sad that Ogre didn't end up working, but I'm very, very happy. that someone's taken the taken and they seem to be doing unbelievable job but I wish I knew more of the technical details but I'm very happy to see that because it's always been for me it's like

stable coins and prediction markets and prediction feeds you know oh that's very very 2017 of you I know but it's like it's like what's better than those things um and and so you know because those things actually catch AI in their web and that's what's so cool about them. So yeah, that would be my prediction is that there's more of, there's a lot more of this. And it is the vector for AI to kind of express blockchain. Yeah, no, I can explain to you how a polymarket works because it actually uses NOSUS's conditional token framework and we built it years ago. No way. But that, yeah, but that's a story for another time. Okay. There's probably another podcast about that one.

Trent, what about you? Prediction on prediction feeds. Yeah, I mean, I think overall prediction feeds will find themselves playing a role in the broader world of AI. You know, mainstream has, you know, learned what AI within the last couple of years in spades, right? You know, I mean, my friends would hang out all the time going back to the 90s talking about it, right? So to me, it's wonderful that a lot more people talk about it and what the future holds with AI and the philosophical implications and societal risks and societal gains, potential, right? Right. And as these AIs, you know, more and more powerful AIs get built, some will be centralized, some will be decentralized.

And both the centralized and decentralized ones are going to want to be using the prediction feeds to make better decisions, right? So it's not just for pure trading, you know, that's the near term stuff. But for any sorts of decisions, right? Already we're seeing, you know, like chat to BT and stuff, hooking into different APIs to do what it does. well, there's nothing stopping it from looking into this API, this stream of ocean predictions and numerary predictions. Potentially, you never know. So there's that.

But this will actually unlock a lot more power, right? Because you can have AI systems if you have like super awesome weather prediction, which I hope, you know, maybe not, you know, global on a two-year time frame, maybe. But three, four, five, who knows? And incentives are a powerful thing. Then with super awesome weather prediction, it will, you know, have big impacts on improving safety, whether you're flying or driving otherwise. And a whole bunch of other stuff we probably can't even predict, but just sort of macro level,

I see that prediction feeds will catalyze the AI space. That's probably the biggest thing. And I also see that they can help to catalyze the defy space by increasing the capital efficiency. And also the triad-fi space, as Richard and the numeric here are showing. And, you know, so that's some of the basic. So improving AI generally, improving the phi spaces, and then beyond, like I mentioned earlier, you know, wet weather, prediction, et cetera. And one of my pet dreams is to have superintelligent AI emerge by that is actually, you know,

a single large model trying to predict the weather dynamics of every square kilometer on the planet, right? I think that would be really cool to see. And by the way, there's going to be another big side effect, which is part of the point of predictor, which is this will be one of the leading horses pulling all. along the whole decentralized data economy, right? You know, now there's already, there's more good money in making predictions, right? Doing it at Numeri or Ocean.

But then working backwards, just like Richard had talked about, there's this market, Numer Bay for the models, we can see more and more of that stuff and thinking out the rest of this decentralized data supply chain. And so overall, all of this stuff is going to catalyze decentralized AI for really the large-scale models and that's pretty exciting. So yeah, that's where we're headed.

Fantastic. So where do we send people to find out more about predictor and numerai? So numerai, the one I was talking about crypto, I would go visit that one first. It's crypto.nomeru.a.i slash meta-hyphen model. That will give you what the model's predicting right now

and you can download the predictions, and you'll also find a way to sign up there. But there's also Numeru.aI for main Numeri, if you're a machine learning person or a data scientist, without any data, then that's great because you can download our data from there. What about Predictor? Yeah, so Predictor.aI, so Predictor has two O's.

And from there, you'll see a table that looks a bit like Coin Gecko or coin market cap and, you know, scanning the table and looking at it, you can sort of quickly grok how predictors behaving and so on. And you actually get the Bitcoin feed five minutes are free. So play with that too. And then like mentioned earlier, you can go to the top right, click on run bots.

That will take you to a readme for running prediction bots. And actually trading bots too. So you can have your own trading bot running live on finance, etc. You know, so that's the main way to get going on predictor. Going to predictor. dot AI. Your Bitcoin wins free. Yeah.

So yeah, maybe that's interesting. I mean, maybe someone's already using our API to predict on yours or vice versa. And that's very interesting. Yeah, I could see that. Cool. Thank you both for coming on. That's been super interesting. And, yeah, have a lovely evening.

Thank you very much. Thanks for having us. Thank you for joining us on this week's episode. We release new episodes every week. You can find and subscribe to the show on iTunes, Spotify, YouTube, SoundCloud, or wherever you listen to podcasts. And if you have a Google Home or Alexa device, you can tell it to listen to the latest episode of the Epicenter podcast.

Go to epicenter.tv slash subscribe for a full list of places where you can watch and listen. And while you're there, be sure to sign up for the newsletter, so you get new episodes in your inbox as they're released. If you want to interact with us, guests or other podcast listeners, you can follow us on Twitter. And please leave us a review on iTunes. It helps people find the show, and we're always happy to read them. So thanks so much, and we look forward to being back next week.