Shashank Yadav on Why AI Agents Will Replace Hedge Funds

Transcript

This is AI Supercycle episode 31 presented by Near and today we've got our friend Shashank from Fraction AI XYZ. We're going to be talking about autonomous AI trading agents and everything happening in the space as Fraction AI recently recently announced. And so Shashank, congrats on the announcement and welcome to the show. Yeah, thanks a lot for having me. Really appreciate it. Absolutely. Well, I I'd love a bit of background on yourself, how you ended up as a founder of Fraction AI and then we'll we'll get into Fraction AI and how you're viewing the entire AI and agentic space nowadays.

Yeah definitely. I started in AI back in 2014 in college. It used to be called machine learning back then. We barely had any deep learning models. So we were doing research in computer vision and our goal used to be identifying things from images. So if given an image of cat, the model has to say it's a cat or a dog like super simple models. And so I've seen the growth how industry has progressed over the years. And transformers was barely a thing back then. And especially in natural language processing, we used to believe non-neural models were actually better at prediction because generation

itself was super bad. Wasn't possible at all. And the kind of growth we have seen in the industry over time, it's just fascinating. After college, my first job was at Goldman Sachs as an AI researcher. There I worked with Professor Charles Elkin. He was head of machine learning at Goldman. So our team was called called core machine learning team and we were embedded into several different finance teams. So I got to learn how finance actually how AI actually embeds into finance at the world's best investment bank firm.

So yeah, that that that that was that and after that I worked for an early stage startup a hedge fund and finally decided to start Fraction AI because I saw how how how quickly AI was moving, how quickly it was improving. And it was necessary to ensure that everyone had equal access to AI, equal access to agents so everyone can have their own agents that creates value just for them and not not just in the hands of few companies. Absolutely. And I and I think this is a general trend from you know those that we're seeing even with the larger labs, companies.

You know, the the idea that intelligence should be ubiquitous among everyone on the planet. I think this is something that we all share. Obviously some of them are going about it you know different ways and you know two of them Elon and Sam are in court right now battling it out about how it should be done, right? So obviously you know we here believe in open source and we believe in a lot of these a lot of these pillars for open source intelligence and AI. And so if we're talking in particular about you know the AI trading agents here, I'm curious before we you know really dig into a fact fraction, what was it that you saw at Goldman Sachs you know during your time there? How were they as much

as you can share? I'm sure you know they've got trade secrets and and things NDAs and whatnot. But you know what what was your experience at Goldman? How were you using AI in the market making and trading business at the largest scale? And what were some of your largest your most important takeaways from that experience? Yeah. So trading is a major part of Goldman's business. Market making is mostly done by HFT firms so it's not that relevant. So they create very exotic products and like directly sell it to high net worth individuals. So you have to create engines to manage that risk. So for example, there's a portfolio of watches, paintings and like several other exotic stuff. It is very difficult to identify

a true value and sell it at a price where you can actually like hedge the risk with respect to your entire firm. So So what you learn from there is managing risk itself is way more important than the the returns curve. So yeah, that that kind of got stuck there. And regarding agents for finance, perps are like the biggest growth drivers in crypto right now. In in due time possibly all assets, all RWAs indices equities, commodities would be there in perps. It's just a better way to trade.

Options are super difficult. You have to know a lot of maths to actually do options right. But with perps, it's super simple. Anyone can do it. So I believe for perps is going to be a big growth driver and agents for perps is how it's going to play out because the way agents can trade, it's way more efficient than humans ever can. However, the there has to be intelligence behind how trading happens. So if you read Marcos Lopez de Prado, he has been the the authority on using ML in finance. He talks about how you shouldn't directly try to put fit models onto the data. Because the model will always fit on noise since in

finance there is very high noise to signal ratio. Most of it is noise, very little signal. So he talks about how you should always start with an idea and then try to validate that idea through data. And we used to do that at Goldman and that's what we have done with our new product product index. So there you just get started with with an idea. You just say what you believe about the market, what what is going to happen, why is it going to happen? And the AI just talks with you. It it discusses why it thinks you're wrong or why it thinks you're right. And then it converts your idea into a strategy a coded strategy. And then you can back test that strategy across different time

periods, see how well it has performed. And the then AI even recommends you based on the results how you can actually improve that strategy. And then you don't have to improve it. AI does it for you. So at the end it's you who is working with AI to create a good strategy, but just through your ideas. You don't need to code, you don't need to do anything. You just supply with your ideas and it gets validated through AI. And finally when you are done, you can just make it live and it trades directly on Hyperliquid. And yeah, that's it. It's it's just that simple.

So very I mean it's fascinating. The AI is effectively farming us for our ideas. We're we're willingly obliging. And so we we come up with this idea. Now it it you know will talk back and forth with us. Ultimately it will it will codify the strategy and then it will go and it will execute the strategy. And so I've got I've got a couple questions that immediately come to mind. Um you know let's say I have a terrible idea, right? Like I don't know what I'm talking. Let's say you know uh NFTs are going to make a massive comeback.

Bored Apes, right? And and I I I say this because a lot of AIs right now they feel like they're going to agree with whatever we say. It's like it I don't well if it hallucinates, you know it'll say it it it it'll say all these kinds of things to just affirm our beliefs. If I have some radical idea that I think Bored Apes are going to you know or I think If I think what what's one that might you know if I think Solana is going to flip Bitcoin. Ridiculous things, right? Will it push back? Will it tell me like look man, you probably got you probably got the wrong idea here. Maybe you want to you want to you want to check yourself. Or is it just going to affirm and it's like oh I

can go lever up on Solana for you if that's what you want? Yeah, that that's a real problem and I guess that's why harnesses are becoming more and more important because most of the models are converging. If you look at like top three models, they have similar issues, almost similar performance in all things of value. So what matters is how you design workflows around those models so they get validated at at each step and are sure about making a choice whenever they are making a choice. So probably in our case if you say okay, I am really long on Bored Apes. It'll probably say really bro? I mean >> [laughter] >> Something like that.

If you ask it whether Solana is going to outpace Bitcoin, then it'll look at the past data, look at the current news and then it'll come up with why it won't happen probably. And if it does happen, what has to happen for it? And then it'll say yeah, it's yeah, really unlikely to happen. Don't I won't bet on it. Yeah. Yeah, so maybe like a confidence interval or something like some statistical probability that it it it could happen. And and so let's let's talk a little bit more about why perps are such an effective modality at expressing some of these ideas.

Yeah, I mean Hyperliquid obviously has started to incorporate more assets, more traditional equities and these sorts of things, commodities, oil, gold all available on Hyperliquid in addition to traditional digital assets. What is it about perps and Hyperliquid that are such an effective venue at expressing some of these ideas and allowing AI models to carry out and execute these trades? Yeah, I would say leverage. And if I have to compare perps, I would definitely compare them with futures and options.

So you can trade with leverage on futures and options. In futures, you have to carry over because they expire at the end of month. So it's a hassle. You there's an extra step involved there. In case of options, they're just super difficult. You need like super cool quants to work all time to make sure you actually make money in options. There's those Greeks. There's so many equations, maths. You need actual physics PhDs to trade options. Whereas in case of perps, it's just up or down, long or short, and and a leverage, and you are in.

Everything is taken care of by the system. It's just a funding rate which is pretty well defined. You can't really do there much. And yeah, so you are just taking up your directional position. It's like the purest form of taking a position. So as things get simplified, they are they're just better for everyday people. Man, how how long until we see we see Goldman, you know, you go back to your old boss and you're like, "Hey, I built this thing. It's ready." Oh, actually my old boss, Professor Charles, he's one of our advisors. And we do connect like probably every week to talk about what what what's happening in crypto. So he's he's an AI expert. He is a professor at UCSD.

So he's like super deep into AI. He has been in AI since decades. So I am the one who connects him with crypto. And he's he's super excited in the direction we are going. Very cool. And and the index product is on mainnet, right? It you know, it's ready for people to say their ideas and and execute strategies. And so what is do you have any insight into what people are currently using it for? Can do you know like what some of the more popular strategies that that people are saying into into index are?

There are some cool power users in our community who have been kind enough to share some ideas in our Discord. And by the way, the the product was in beta for a few weeks and it was open to like 50 people. And they tested like thousands of strategies. I myself since I had access to API, I just paralyzed things and ran like 25,000 strategies of my own and I figured out what actually worked. And then those are kind of live right now on index. So I I I super believe in dog fooding your own product because if you don't use it yourself, how can you expect them others to use? So let me I just want to clarify. So you were you you ran 25,000 strategies in parallel.

I I mean, Shashank, you're a smart guy, but there's no way you had 25,000 ideas off the top of your head. Did you have another model generate the ideas and then you put them into index to to to run these strategies? Oh, yeah. As you said, AI is farming us for ideas. So I would just start with a very simple basic thread. Yeah, I want to trade something related to momentum with some probably volume or RSI based filters. And then it would generate like 20 different ideas. Then we would backtest those with like different set of parameters. See if if those are working. If they are working, are they stable? Because if something works for one parameter and doesn't work for other, it's it's not really stable.

So it would run and then it would learn where where it failed. Then it would say, "Okay, this didn't work, but this part of that strategy works, so I'll I'll combine these two." So yeah, it it kind of got out of hand over time. And then I had like 25,000 strategies running in parallel. But yeah, it was a great experiment. Okay. And so back to the original question. What's working? What you know, you you backtested all these strategies, you ran all these ideas, your community got in during the beta program. Now you're seeing these things in production. What are some of the most successful strategies on index right now?

Oh, yeah. So if this whole space is the space of all strategies, the working strategies are like few dots here and there. So when you search for those strategies, you like kind of go like this and then you hit some good strategy here. So that doesn't mean all strategies here are good. There's something here, but you have to start from a different point to get there. So the point where I got the the best strategy is something like trade on momentum when the market actually mean reversion. So if the market has gone down for several days and also in several hours and the volume right now is pretty high, then if you buy, you make money most of the times.

So going against the trend. If you have momentum Yeah, so the market has fallen for for quite a while and now the volume is high. Relative volume is high. So it's like it's it's building up. And now if you buy, it's it's more likely to go up. Oh, interesting. And this is true across a wide spectrum of assets, crypto, equities commodities. Most of crypto, most of equity equities, not on BNB. Nothing works on BNB somehow.

>> [laughter] >> Yeah, but yeah, it it it works for most assets. And another great insight is don't trade too much. The the transaction fee eats a lot of your profits. So a strategy that that only trades when it's confident is way better than something that trades a lot. Wow, very cool. And so you know, obviously you're you've got the infrastructure now where anyone can trade anything just their idea. The index AI will effectively, you know, coordinate the execution of the strategy.

Do you plan on, you know, kind of open sourcing some of your dog food ideas? Like, you know, some of the ones that you were able to do, give people some suggestions early on, kind of help get them started, and then they'll start to, you know, use that as a starting point and build from there? Yeah. So we do have documentation and we are working on that. But my principle is your product should be such that it shouldn't require any documentation to use. So when you go to index, you see several options there. Like, do you want to analyze a token? Do you want to start with some pre-built ideas? So you are not presented with a blank screen. You have some prompt to get started with. So

initially if someone comes there, he doesn't know what to do. They can just click, get started with, get a feel of the system, practice a few things, see what's working, what's not. Then they can even ask questions, "Why is it not working?" So the goal is to be make it back and forth rather than asking user to go to our documentation, go through it. I just just interact with the AI. It will help you create a strategy agent. It will also tell you how likely that agent is to do well. And so yeah, just talk to the system rather than reading documentation. Yeah, makes makes complete sense. Now we can talk directly to systems, you know, given these LLMs.

I want to I want to switch gears a little bit because it it it feels like there's an untapped opportunity here. Now that you have the the the portal, the front end, the user interface, if you will, to interact with these venues. I don't I'd imagine that you can now expand the breadth of venues. So you have hyperliquid perps because you have leverage at your disposal is a great modality to execute some of these strategies. You know, prediction markets also have a very broad set of markets at their disposal.

Do you have plans to integrate prediction markets into the back end so that you can carry out the execution of some of these strategies there as well? Yeah, definitely over time because our goal is to bring this intelligence to all products on chain. As we have seen like the RWAs and perps and all these financial products are coming on chain. It's going to be worth in trillions, but it doesn't really matter if if no one can actually trade those products in the right manner. If you just have to deposit your capital to a fund and then then that fund decides where your capital goes, it's not really decentralized at all. So I believe it's going to be agents

who are doing who are going to do the allocation. And we should enable users to create good agents. So they can use it. They can make their friends use use their agents. And yeah, that's how we we enable intelligence for everyone. And so what do you think the market structure looks like as we build out the rest of the agentic trading economy? And and you know, I I speculate because the the ideas like, "Hey Claude, build me a trading bot that is going to make me a millionaire in a month. Make no mistakes, right?" Everyone types this into to Claude or index. Now we all have intelligent AI trading systems.

>> [clears throat] >> Maybe the market gets a lot smarter, but then you have smarter market participants all competing against one another, and it becomes difficult once again. How do you think the market structure will shake out once everyone has agents at their fingertips? It's funny because that was part of our introduction video where it starts with index making me a millionaire. It's like, "Yeah, that's not how it works." But we have something that that works. So >> [laughter] >> yeah. Um So the the way we look at it, it's it's still going to be like few people who are really good at creating these agents.

They'll have those agents, and everyone would be using those agents. But those agents change over time. So it's like how YouTube democratized video consumption. Earlier you had to be dependent on Hollywood for getting movies. And there's only few studios who were in charge of making movies. And only everyone who was good at creating content worked there. If you worked there, you could work in movies, otherwise not. But YouTube ensured there was a platform where you could create content, and that content could be consumed by everyone. So, I believe it's going to be same for agents rather than just 10 firms which have been same for like several decades who are managing

trillions in capital. People would be creating agents on Index and people would be putting their money into the best performing agents. The best agents are going to change over time, but everyone has a fair shot at creating a good agent. So, if you have a good idea, just go experiment with your agent, see how well it is doing, and then you can make it open for everyone to manage their capital through your agent. So, just make it open for anyone to to be able to manage capital through agents. Yeah. I I really like this, you know because things will continue to come out.

People, you know, uh they will will generate a a meritocracy of of sorts, you know, just through their ideas. Some people are going to have good ideas, doesn't matter if they're, you know, they're they're young, old, you know, where they come from. Anyone can have a good idea and they can embed those good ideas into these agents. I I'd also imagine some people are too lazy to come up with their own ideas. And so, you know, let's say, you know, Shashank, you you've already kind of done a lot of the backtesting, obviously have, you know, strong AI background.

I can come up with an idea, but the likelihood of it being successful, I don't know. Maybe, you know, I have some money and I want to use your agent, right? Can I I want to like I want to rent or license your agent on the Index protocol. Do you think this is something that will will see a lot more of? Oh, yeah. This is like the one of our major features that's not not live yet, but we are working on it. But yeah, that's how we look at it evolving over time because not everyone can create good agents, but everyone should have access to good agents.

So, it's like everyone specializes in different things and the job of a platform is to connect capital to someone who manages good capital in the right manner. So, even you could be someone good at creating agents, then create good agents. You might just want to manage your capital better, so just select based on past performance, actually look at the strategy how it has worked and and deposit your capital there. Maybe even diversify it. So, yeah, both part of the market are important and yeah, that's that's why we we are making in this manner. Yeah. Yeah, it feels like there's a marketplace developing here for agents as well.

Once they're created, then they can do multiple things for multiple different entities. So, very, very cool to see. I also want to just ask, you know, we've obviously we've seen a lot of exploits as of late. It kind of feels like AI has gotten into the hands of the bad actors and and a lot of the bad actors are using AI for evil things more than the good actors are using AI for protecting and defending against those evil things. Um you know, AI because it's still somewhat immature or naive or nascent, um presents a lot of opportunity, but it also, you know, there's been somewhat of some some apprehension when it comes to

giving AI private keys, you know, putting a lot of money into these agents or into these models. What is the what is the security mechanism or the funding mechanism for how money's deposited into Index? Is it just a matter of, you know, I have my Hyperliquid account, I fund my Hyperliquid account, and then the AI has permissions to trade on my Hyperliquid collateral. Can I set up limits or risk control so that it doesn't spend all of my balance in one trade? Like what is what are some of the parameters here that that can protect against something going wrong?

Yeah. So, this is super important to us and we didn't want to leave this to the user because a platform shouldn't be like, okay, yeah, yeah, I can make mistakes. You you just take care of your money, keep limits, and so on. So, the way we made it is agent never has access to your wallet. It never it can never directly call your wallet or use your wallet. It can only interact within a small set of parameters. It has it's like its own sandbox. It has its own language, it our own proprietary DSL. So, what it it just writes code in that particular language and it calls the APIs and it can only do certain actions and it can never send

your money from your wallet to some other wallet. It can only place trades, take actions on the platform, do whatever a normal user would do on Hyperliquid, and nothing else because it doesn't have access to any API outside those. So, it's it's yeah, so since it doesn't have access to that API, it cannot call that API and that's why it's safe. Yeah. Yeah, makes makes sense. Shashank, really really appreciate you coming on today. I'm just curious, what's the best way for our audience to get involved? Where should they go? How do they get started?

Yeah, just join our Discord. There's a lot of power users there who are sharing their experience, sharing some of the strategies that worked. A lot of users ask why their strategy is not working and these kind people are super helpful. They always say, okay, try this and not this, this doesn't really work. So, yeah, just just go there, talk to people, you'll you'll you'll learn a lot. You'll learn a lot about finance, you'll learn a lot about AI, and you'll create super cool strategy agents that grow your capital. Very cool.

That's fractionai.xyz. Try out the product, it's live. You can test out these ideas and join the discount join the Discord, learn a lot more about what's going on here. Very, very cool. Shashank, thank you for joining this episode of AI Supercycle. We really appreciate you coming on and and congrats on both the raise and the launch. Very cool to see and wish you the the best of luck and success here in the future. Thanks a lot. It's been a pleasure. See you next time. Thanks, Shashank.