บทถอดความฉบับเต็มของ How Hudson River Trading Actually Uses AI

5: We essentially pick up a penny in front of a steamroller, like we are making sort of money from that spread, and we can pick up the pennies in front of a steamroller if we have a really magical device which tells us how what everything should be. 00:05:16 Speaker 2: When the steamroller is coming. 00:05:17 Speaker 4: Yeah, it tells us. 00:05:17 Speaker 5: When the steam roll is coming. And so I think that's kind of the very very sophisticated sort of middleman in some sense. And the same we've had Amazon is Amazon doesn't make stuff, but it's a very valuable, profitable company provides a service. People get value of same thing. We're moving stock spons through time and space between different counterpartties. 00:05:34 Speaker 2: And yeah, we will. 00:05:36 Speaker 3: Ask you about the steamroller in a few minutes. But before we do that, how does AI or the way you're using AI actually differ from the algorithmic or quant trading of olds, Because I guess that one of the questions is, is this, you know, a sort of evolutionary change, you know, maybe a marginal improvement on what already exists, or is this something seismic and a step change the big ship in the way trading actually works. 00:06:01 Speaker 5: Yeah, I mean I don't want to overstate ourselves in some sense, because into space, as you mentioned before, it's very like opaque what sort of different firms of this class are doing. I can siddenly speak to our own experience, which is we've been doing this type of trading for twenty plus years, and much like everyone who was doing this, the way it kind of worked was you handcraft features. It's sort of based on human intuition. 00:06:24 Speaker 4: Oh, I don't know. 00:06:25 Speaker 5: If the order book looks imbalanced, there is more people wanting to buy than sell, the price is going to go up soon, or something like that. And maybe you get a bunch of very smart people and they think very hard, like it's almost like making a very fancy watch kind of artistanally craft all these pieces, and then maybe you use relatively simple mathematical techniques like linear regression to combine those predictors. And I've been going to conferences and things and recruiting for a long time, and even if today's going the internet, you'll people say things like, oh, that's all you can do in finance. For some reason, they'll say this. They'll say something like, oh it's too noisy, or markets are too nonstationary or things like this, and so that's all you can do. And I guess that belief isn't really backed up by anything in my opinion, and like lived experience, I guess, And so we sort of viewed it more for a long time as well, everything that's happening in the world, and ideally you would put this into kind of like a machine that does not have any human biases. 00:07:19 Speaker 4: I don't know how to. 00:07:19 Speaker 5: Trade stocks myself, like I buy broad market ets, what do I know? And so we but if you could put all the data into a box and it kind of could jurn all about data, it would find things that you would never be able to do, this handcrafted thing. And we started doing that very early, relatively in se twenty fourteen twenty thirteen period, and over time, over less a decade or so, much like in other contexts that are not finance, there has been sort of a hockey stick and you can measure by the sizes of the models the compute deployed and over time that way of modeling the markets initially was not like a hybrid with the traditional way. 00:07:57 Speaker 4: Bially kind of just like overtook it. 00:07:59 Speaker 5: Entirely and so now our trading is entirely driven by this magical machine consumes of a data. I kind of keep saying this magical machine. It consumes of a data for a reason, which is that this is how chat GPT is trained, it consumes all the data all the Internet. It's kind of scraped and connected into one place. When you train a model, that kind of takes it all and something emergent comes from it. And that's why I'm kind of a bit of leading. But that's why I'm talking about in the sense, and I think that is materially different from the like I'm using my intuition of the markets to kind of construct a predictive model. 00:08:31 Speaker 3: So just to be clear, how much of usefulness of AI here is about execution and the fact that you can crunch a lot of data really quickly with hundreds or thousands of GPUs versus spotting sophisticated patterns or discrepancies that you can exploit. 00:08:48 Speaker 4: I think it's both. 00:08:49 Speaker 5: I think one of the things that people sort of missed with a whole like do a linear regression type thing is when you really think about how much data there is in financial markets generated. And when I say data, I think it's important to think of it as every event that happens in market. It's not the sort of time serious of prices, but like the actual low level substrate people are quoting trading retracting quotes that like low level stuff is internet scale data set sizes, and one of us sort of bitter lesson ye type things of AI was like, you know, you shouldn't think too hard about how to feature engineer this in pre process, that you should kind of throttle in to something a form of computation that can kind of make use of internet scale data. In the twenty tens, it was like computer vision people used to make detectors for edges of images and things and they would combine them and same thing. It's like that was a good approach, but you know, it's completely dominated by the idea of getting very large umber of GPUs and a kind of a pretty generic neural network form empowering through it. As for like the how is it finding things that other methods could not. It's very hard to say our models are not very interpretable. And I think that's fine because, as Joe mentioned, our sort of trading style and holding times, a bit of it is like minutes, hours, maybe like a low single digit days for the most part, And I guess in my mind it's

unreasonable to expect them to be interpretable because I don't know if I looked at the autobook data for Tesla or something. Am I really going to be able to tell you better than random with the price of Tesla will be in a minute's time? And so I kind of think it like that, if you have something that's clearly superhuman already, what level of interpretability because you expect like this is very different right to normal AI. 00:10:28 Speaker 2: Right, this is gets into some areas that I'm very interested. But just to like establish what we're talking about, you're trading a stock like a Tesla in video, et cetera with your magic machine machine. We had another episode where that was the money box. 00:10:44 Speaker 3: That's a magic bo different, that's a different one. 00:10:46 Speaker 2: With this AI machine, it is sort of arguably grown, right, and it's sort of grown in a lab more than it is programmed, much like a chatbot. I know, it's very different technology, like what is the price of in video going to be tomorrow? Or what is the price of ennvideo going to be this afternoon? What you're saying is with your technology, you have a better chance of getting that right, that you actually might be able to make an informed prediction about the future in a way that you couldn't have done, say ten years ago. Yes, and that people who talked about this they would come up with reasons. Oh, the stock market. It's not like chess or go, and therefore you can't really do predictions the same way. But what you're saying is that with these models, which are different than lllums, there is some, at least on a short time scale, predictive capacity. 00:11:33 Speaker 5: Yes, I think I find this still to mistake a little bit hard to believe. I think you get this kind of efficient market hypothesis stuff jumped into your head. It seems if someone is saying they can predict like the price of a stock in an hour, your instinctual reaction is incredulity, Like just sounds like you're kind of bluffing or making it up. But no, these models can predict this, And I think it's the way to kind of reconcile the like really man like kind of distinction is that the predictions are very bad in some sense. We don't no way to talk about like accuracy. But I think the way to think about it is like the accuracy is like fifty point one percent type thing, like they're only a little bit better than random. 00:12:10 Speaker 3: But I suppose an extra one percent like blows up your profits if you're doing it. 00:12:14 Speaker 5: Doing it scale doing it enough times and over time you kind of realize the biased coin flip. And as for why it might be possible to do this without kind of invoking magic, It's like markets are very beautiful interaction of like many different potties, all the different kind of utilities and risk preferences and things, and the only way you really see what people are doing is by like the actions they take in markets, and you kind of it's sucking up all that like

signal, that micro signal and extrapolating. 00:12:41 Speaker 2: The synicism or the skepticism about the possibility of machines that could predict the price of stocks is a little strange, right, because machines ingest data then whatever, maybe they see a pattern more likely than not, this consolation of data means tomorrow will be green. Humans do this all the time. What else do we have besides data? 00:12:59 Speaker 4: Right? 00:13:00 Speaker 2: You have an analyst and they put out of Tesla or whatever in video is going to go to five hundred dollars a show? 00:13:05 Speaker 3: How dare you insinuate I'm not smarter than a computer show? 00:13:08 Speaker 2: We were like all humans have this data and much less data, and yet humans are making predictions all the time. They have a whole industry of it, so the idea that therefore was for some reason a computer couldn't do this with much more data analysts ever, have I understand why the cynicism comes off as a little strength. 00:13:28 Speaker 3: I think some of the doubt stems from this idea that a lot of these models tend to be backward looking, right, and some of them occasionally are pretty bad at spotting or reacting to big regime breaks. And I guess the thinking again sometimes is that maybe humans are more flexible, maybe more adaptive in their thinking, and they can kind of spot these big cultural shifts. How do you actually, I guess, prepare for those big pattern changes. 00:13:55 Speaker 5: Yeah, I was at HIT for COVID, and I thought that was kind of like the most that was a pattern, that was a big pattern break, and things went totally fine. Actually, it was more of an engineering crisis in some ways. Stock market volumes exploded and every system was just like screaming trying to keep up with the volume of activity. But in terms of the predictions they stayed quite good. And I had like Rickoncilus in my head as well. I guess it is a matter of like horizon and like how far in the future are we talking intra day. I think a lot of the price movement is driven by just observing, like the flows. It's hard for us as humans to observe, but it's like the relative patterns of buyers and cells in the markets. And it's like, yes, during COVID, volatility was massive and prices were moving up and down a lot, but they were going up and down during say March twenty twenty, and so these models it was sort of out of domain for a human. 00:14:45 Speaker 4: But I don't think out of domain in some sense for the models. 00:14:49 Speaker 5: But I guess I also don't know how you would apply this thinking if you were trying to make it a. 00:14:53 Speaker 4: Month ahead predictions. 00:14:55 Speaker 5: I often get like people being like, oh, everyone knows hedge funds, which were not a hedge fund is like like flipping coins, and it's some survivor bias thing. And you know, I genuinely don't know about months out prediction stuff that is not a data rich environment. 00:15:10 Speaker 2: Just by definition, there have been more days than months, so therefore prediction on a day basis you're offered a lot more data that. 00:15:17 Speaker 5: The rule of thumb is basically very useful and it extends all the way down to seconds, and we see that empirically all the time. And so yeah, I guess all the things I'm saying, do you have this heavy out that it does rely on sudden of it being a certain level of signals and noise. I definitely cannot make reasonable claims about the price of things in like a month using the same kind of like AI hammer. I guess also to be specific, I'm talking a lot about using market data to make these predictions. And that's because on the sort of intra day timescale that is the most important thing. It's all about flows and things have been back and forth. If you're thinking about things in a month's timescale, I think that's fundamentals. 00:15:55 Speaker 4: And can AI be used for that? 00:15:58 Speaker 5: I don't know, to be honest, and it's definitely outside my wheelhouse. And I guess people have various opinions about that, and maybe some people very much would like to claim that they can, and you know, others maybe don't. But it's definitely outside of my Arab expertise. And I don't know, Wait, talk. 00:16:13 Speaker 3: To us about the data that you're using, or talk more, because this is another area where people tend to talk in PR speak, sometimes we have access to all this data, unusual data, alternative data, and that's going to enable us to use AI better. What are you actually looking at and what have you found? I guess most useful? 00:16:32 Speaker 5: Well, I think the thing that I found most counterintuitive when I started was that when you're thinking about predicting the prices of anything a minute, an hour out, I far the most useful thing is just market data. This is the market data feeds you can buy from the exchanges for a pretty reasonable price. People often think of some sort of like competitive moat. The data feeds for these exchanges are not particularly high. I mean crypto, you know where it's just like a wild West. But everyone can collect these feeds, and so that is the most useful raw ingredient. That is the most true expression of everyone's intense right they're going to the market, that quoting, the buying, selling, That is the primary ingredient. People get kind of caught up on the whole, like do you have a Twitter feed type of thing? And Bloomberg sells a Twitter feed through a state of products, and it's every now and then obviously something happens, news happens during market hours that moves, the price justicates the price. But if you really coldly rationalize that that is a relatively infrequent thing compared to the overall massive markets. So I thinking entry day, I think these market data feeds, it's literally like a little events someone quoted that this price and this size. 00:17:35 Speaker 4: It's all anonymous. 00:17:36 Speaker 5: Market data feeds are anonymous, and so that is the raw

stuff, and it is vast. There are just millions and millions of events per day, per stock, per future. When you get to the day days timescale, that's where the alternative data quote unquote kind of really comes in as an alternative to market data, the SEC filings, the news feeds, balance sheets, brokers reports. 00:17:57 Speaker 4: Things like this. 00:17:58 Speaker 5: That's where that comes in, a vast sea of data offerings that people try and sell that I think in that kind of situation, you know, it's a very low shop environment you start getting into, and it can be hard to attribute the extra shop needs of these things. But in some sense it's also very democratized where maybe people collecting very secret data sets, but my inbox, and I'm not even the person in charge of buying these alternative data sets, is often full of people trying to sell me the latest alternative data set, and I think a lot of them don't necessarily have much predictive value, but clearly as a market for. 00:18:31 Speaker 3: What's the craziest one you've seen? 00:18:33 Speaker 4: Huh can you remember? 00:18:34 Speaker 5: I mean, people have definitely reacted very strongly to the Wall Street bets Era tried to kind of create a bunch of reddity extracted thing and go beyond just like raw captures of Reddit and trying to distill. 00:18:46 Speaker 4: It into something. It's just even just thinking about it. 00:18:50 Speaker 5: The meme stock thing is talked about more after it happens, and it happens before, and so like. 00:18:54 Speaker 2: I don't know, it's sort of a sideways question. You mentioned interpretability, and just give me something I've been wondering about AI for a while, not even in the finance realm specifically, you're a deep mind, which of course produced a great GOT player better than the greatest grandmaster in the world. I play chess. We know that chess engines are much better than any human. On the other hand, as far as I can tell, there is no good AI chess tutor. So in other words, the chess crush a doo. But like I've never been able to get a thing where it's okay, you did this move, but you know what, you're closing this rook file and down the line because like it doesn't do that. The chess dot com human talk is very rudimentary, et cetera. Can you talk a little bit about why there are these problems where some version of AI or machine learning or whatever can do fantastically well, but then the actual explanation of what it's doing, which I think is kind of what interpretability is, can't articulate in a plane English why it's able to do what it does. 00:20:16 Speaker 5: I think it's just because these neural networks are it's just like a big old blob of numbers, and what we're aiming to do in maturnit these models is to almost like free ourselves from almost all structure, and they might learn things in a way that is nothing at all like how we learn things. And so my best guess for like why it's hot is because they might be reasoning in some sense internally, and people use these words like reasoning. 00:20:43 Speaker 4: It kind of makes me win. 00:20:43 Speaker 5: So I've seen imagination and things used about neural networks. I don't know if it's like kind of anthropomorphization of them as kind of dangerous because they are essentially processing things internally in this way that I think is inherently not like how we do. And that is my best sort of Yes, there are some interesting counter examples. One of my favorite set of things in It Possible years was Golden Gate Claude, which was anthropic made that the model basically get very interested in the Golden gate Bridge. Every question they asked would come back to the Golden gate Bridge, and so they're not completely impenetrable, but it's clear that like, I guess how I'd be on a point to kind of map this back to how anyway, Like we think it's very attempting to and exciting too, and as especially for like AI safety applications, which is not really relevant to me so much, but I think it's for attempting to try. 00:21:29 Speaker 1: Yeah. 00:21:30 Speaker 2: No, it strikes me is that if you could solve that many jobs, would you could actually make a lot of productivity gains. But I do think that's an important hurdle when you're training your models. So your models are different than large language models, et cetera, but what they have in common is there's a credible amount of data, incredible amount of compute demand, how applicable if someone had worked on lllms, would your training process be to them? How interpret how could they move from that environment to yours? Are there enough similarities in the base notions and compute and requirements to train a model such as yours versus people are doing it the major labs. 00:22:06 Speaker 5: I would say now in twenty twenty five, absolutely, But I would not have said that in twenty twenty. And this is something that kind of caught me by surprise having done this for a while now, is that our problems are kind of defined by long sequential strings of information in some sense, and extrapolating from that. If I think back to the pastive AI, it was like is this is it a hot dog or not? It's kind of like the like image classifier, you know, test. Then there was some stuff with audio and things. I was a little bit more familiar robotics, eh. But when we got to the sort of LM error, it got very interesting because suddenly the problems were very similar and that you have you want to think back over like long histories, long contexts. 00:22:49 Speaker 4: Okay, that sounds good. 00:22:51 Speaker 5: You've got a lot of data and you want to turn through it in as efficient way as possible. 00:22:54 Speaker 4: You also have to serve this model. 00:22:56 Speaker 5: One has to run in like relatively reasonable speed, especially for the LM. Is there a million people typing into chat GBT dot com and they want to hear a response in a relatively prompt manner. Of course for us also, the models have to

make their predictions in a prompt manner of voice of predictions that I'm useful. So all these things mean that our sort of way of thinking about it has become very similar to the frontier LM things. It's just a very different modality. We're operating on I guess primarily text, and we're operating on this fileless interpretable but still sequential stream of tokens, except our tokens are market events. And so it's a lot of fun because you know, in terms of like the research that is still published, you can kind of look at it for inspiration and draw our comparisons. But it's also very much its own problem, which is kind of keeps me interested every day because it's like its own unique thing. 00:23:45 Speaker 4: It's different. 00:23:46 Speaker 3: I want to go back to the point you made about data and I guess democratizing finance in many ways, and maybe this is a weird question. But I'm thinking back to the twenty tens and we used to talk about the big investment banks as flow monsters. They see all these orders, they get all these orders, they see all the flow, and that allows them to optimize on funding costs and other expenses. Is the idea that data and AI can kind of replicate that advantage so that everyone, or not everyone but Hudson at least becomes its own little flow monster. 00:24:22 Speaker 5: Yeah, I think there's still some trends and markets that worry me a little bit in terms of I guess our platonic ideal market structure is probably like everyone trades on exchange in a centralized place, but that is not really how things seem to be going. And there's a huge amount of like off exchange dark quasi dark volume, and I think there's still a lot of qunits of the trading world where like being in the room is kind of like this big advantage. And this is a very much anti AI play in some senses. Data is hidden the data, the flow data is hidden it and it's not something that you can feed into a machine. 00:25:00 Speaker 4: This very spas amounts of. 00:25:01 Speaker 5: It, So that's kind of an interesting trend a lot of us did get sales get reported in a centralized place later, but it's not prompt enough to be useful, and so to the AI thrives on data, this is in some sense like an issue for the long run, you need to kind of be in the rooms where the sort of trading is happening. 00:25:20 Speaker 2: I'm glad you brought that up, because that's specifically what I'm curious about from the sort of physical infrastructure side, Like if I have a queria to chat GPT, I don't care if the model is like trained in like Eblene, Texas or wherever it gets back to me and whatever. But I know that for high frequency trading, at least on the execution side, there are certain parts that you want to be literally co located, and you want to have the shortest possible wire, and however short it is, ideally you would like it to be shorter. Can you talk about the differences and similarities between essentially your physical hardware stack verse is what would

be required at a large language model frontier lab. 00:26:04 Speaker 4: Yeah. 00:26:05 Speaker 5: I think at a bulk level there was actually some pretty similar things. So I can think about it as like latency and throughput latency being the time to react and then throughput kind of like how much thinking you can do in a certain period of time. And so you're right that like this space demands like low latency. Early in the twenty ten, so it was a sort of Flashboys book and perception where it was like really kind of about arbitraging latency. I'm happy to report that in some sense all the latency has been arbitraged. 00:26:32 Speaker 2: For the most part, there's no more engine shortening. 00:26:34 Speaker 5: The wire is probably like a little bit, but it's relatively small, and like I think if you look at the big quant trading firms that need to like really make the wires as short as they possibly can is done or are no longer relevant, which is great because I find out stuff pretty boring. Personally, I think about it more as like, for a given kind of like speed of response, you should be the smartest person. So it feels like this curve, if you're going to take a second to come up with your trading decision, it'd be a really really good decision. And then it doesn't kind of matter that it took a second. And if you're going to take a microsecond, well a you probably can't do too much in a microsecond, but you know it'll still be the best response in a microsecond, and so you. 00:27:14 Speaker 2: Could be a little worse. You can be a little worse than the second. 00:27:17 Speaker 4: Yeah, for sure. 00:27:18 Speaker 5: And so essentially for our training, we use the cloud. We have our own training data centers that we've built ourselves. That is basically the same, although much much smaller scale the scale of Googles and sayings. I don't know, it blows my mind the spending on stuff like this. We are, I think big if you're not comparing us to Google a meta, but that's not like bajillions of dollars. So training is kind of the same inference. We need to put a devices close to the exchanges, and we need to think very hot about the power usage and the latency. But we have hardware teams, We make our own FPGAs, we make our own chips, and we use off the shelf GPUs, And what we try and do is we try and make sure of it for any given set of speed or response, making the smartest possible decision you can. 00:28:02 Speaker 2: So you can kind of field programmabile gate rate. Oh yeah, sorry, Yeah. 00:28:07 Speaker 5: Basically, all these different devices have different latencies and through puts. GPUs have very high through puts. They are that's what they're useful for, right and so, but the problem with markets is they're kind of like narrow. The amount of traffic flowing into these like lms from everyone typing into their redbos it's massive, and they do also some clever things kind of batch up requests and process and things. We don't really have that luxury really, like the markets are going to happen at the speed they happen. We can't kind of like duck out for a while and catch up. We kind of need to stay in the game. So we have always sort of interesting design challenges around how do we use GPUs which are relatively high latency. 00:28:39 Speaker 4: They take a while to give. 00:28:40 Speaker 5: Back a result, but they can process the whole stock market on one GPU type of thing versus the fast response, And so we have whole teams dedicated to thinking about, Okay, I've got this like intelligent blob, how do I get ounces out of it in different ways at different speeds? And that I think is where a lot of smarts are going in this world these days, rather than like how do I make sure my microwave towers are like slightly better aligned somewhere in rural Pennsylvania, which is a cool challenge. 00:29:09 Speaker 4: It's own right, but it's done. 00:29:10 Speaker 3: I think. 00:29:11 Speaker 5: I think people have found the straightest line from New Jersey to Chicago. 00:29:30 Speaker 3: Joe brought up some of the cynicism around CME's cloud deal with Google, and this came up speaking of a specific cynic who went on the record in one of our episodes. Don Wilson basically made the argument that matching on a cloud doesn't necessarily make sense because you might put into orders and you're not really sure which order gets filled first. I guess you're kind of back in that black box environment, or maybe it's a latency issue. I don't know, is that a problem that you're seeing. 00:29:57 Speaker 5: That's something that I worry about. A general philosophy is MA should be very like transparent and as fair as possible, So like equalizing access is a good thing in terms of participants. Shouldn't be at to like basically pull weird tricks to be faster. On the other hand, I think you want reliability, so like this concept of like orders arriving at different times and being filled in different orders just doesn't seem like a very sensible way to run a market. It's something that requires a lot of effort to engineer around, and it's just a good market design to have. It is a very widespread though, in existing exchanges across the world. We've traded in like a vast number of countries, and some of the exchanges have such amazing hardware that like, if two orders are sent within like a nanosecond of each other, this exchange will never process them in the wrong order, even if it's one hundred different network ports and they're all connected. They have this amazing time stamping stuff. On the other hand, you might have like a crypto exchange where it kind of feels like a kid learned JavaScript and ran set up a website and you're kind of like you send an order and you may may not be confirmed that they even received it, and then you kind of have to refresh your like account balance page like five. 00:31:02 Speaker 4: Minutes later to see if there's many in it or not. 00:31:04 Speaker 5: And we kind of we'll

take we'll deal with it as it is, but certainly we have a preference for kind of equalized access but sort of predictable outcomes, and I think that kind of leads to like people spending efforts I think it's not astly very great thing for society for people to be like stressing very hot about why it lam. 00:31:21 Speaker 2: Yeah, no, probably, I'm glad. I'm glad that you report that we've moved on a little bit since then. Where are your constraints? You know, when you talk to LLLM people, there's debates about right now, is it electricity? Is that the big constraint? Is it there just aret enough GPUs? Is it talent? Is it whatever? When you think about where you are now versus the optimal version of where or is it I mean, data is the other big one because there's all this concern that lllms are going to run out of training data, et cetera. Where is the big constraint for you that you feel like you're solving for right now? 00:31:52 Speaker 5: I think in terms of like really long term strategic planning, electricity is like quite clearly a very binding consideration. When we think about spitting up new GPU based training data centers, it really feels like, is there electricity? Like finding a piece of land to put a building in. There's a lot of land. 00:32:11 Speaker 4: Yeah, the electricity negotiation. 00:32:13 Speaker 2: That's an issue at agr T even for us. 00:32:17 Speaker 5: You know, because we have a sort of hybrid mix of using cloud providers and building our own data centers, and yeah, the negotiations and thinking about power constraints. We have an existing data center in a very cold place and we want to make it bigger. And the data center people are fantastic to work with, but they're saying like, well, we need to go talk to like the power grid and negotiate this next trunch and so on, and it's just it often feels like that is the bottleneck. And on the terms of a GPU availability, it definitely was a crunch at some point in the past, but I don't feel like that. 00:32:48 Speaker 2: Is a little more the entire stock market. Say a little bit more about how you perceive the GPU market. 00:32:54 Speaker 4: Right, I think. 00:32:55 Speaker 5: I think if we ask for GPUs, we will get them to live in a prompt manner, not necessar early like next day. But I don't feel like that is the thing that we have a long pull and spinning up more. 00:33:05 Speaker 2: When was the when was the worst of the crunch? 00:33:08 Speaker 5: I guess twenty twenty three, late twenty twenty three felt pretty bad. 00:33:13 Speaker 4: I was. 00:33:13 Speaker 5: I guess that was like the Nvidia Hopper generation, and I saw also a number in Ploomberg yesterday that I think it was like Nvidia conference yesterday and I said something like it was like one million Hopper class GPUs have been made, but already like four million Blackwell class gp has been made. So I think there's been a ramp up of supply. But I don't think they're also sitting on unsold inventory either. I

think it is being consumed. But yeah, in terms of like what is the hod thing, I think electricity, And I am it's insane. As a very millennial person, I guess climate change was a big thing growing up in college, but a lot of discussion about climate change, and to see people spinning up data centers very fast by basically buying as many gas turbines as they can and putting them outside, I'm like, WHOA, Like, yeah, what are we doing? It's wild, but this the only way to get electricity promptly. You just have to throw guests turbines outside the building and turn them on. It's pretty radical stuff. 00:34:06 Speaker 4: And I don't know how all the numbers. 00:34:09 Speaker 5: Of people talking about for future data center expansion kind of math out because you just back of the envelope the power usage and things. 00:34:16 Speaker 4: And I know that the sam Oltman's of the world. I've thought about this, We've talked about this. 00:34:19 Speaker 5: So oh, we need to be generating this much new power generation per you of time, but there's such daunting numbers. I just don't know how that is all going to work out. But yeah, even for us in the grand scheme of things, like a much smaller player in terms of power consumption, we think in terms of like tens of megawatts, not gigawatts, which is more than most towns and cities and things. 00:34:41 Speaker 4: But still, but we. 00:34:42 Speaker 5: Find it like a challenge to find electricity at a reasonable price. 00:34:46 Speaker 3: On this note, can you talk to us a little bit more about where competitive advantage actually comes from in this space, because if the GPU crunch is somewhat solved, and if latency isn't as big an issue as it used to be, where are people actually getting their edge from? 00:35:01 Speaker 4: Right? I mean people? Talent is one of your other things. 00:35:04 Speaker 5: You asked that a constraint it is It is a very competitive people market. We're essentially asking for people to know a lot of things, be both good researchers and good engineers. Because I don't know in this AI era of a distinction is pretty blurry. It's not something you can just wipeboard and then the coding is a little bit afterwards. Any kind of research idea you have is intimately connected with how you implement it. So that's already like a tough ask. So people are constrained. People that we like I want to find and we pay well for those people as a result, and it is competitive. But I think the more subtle edge is almost like putting it all together. Do you have people who can, like an engineering team that can collect double data recorded, make it available to the GPU training data center. This is like many I guess it's petabyte scale data sets and just stroying too much data, streaming it from wherever a stored to wherever in the world the training data center is reliably these training runs are very expensive and then once you've got that model serving it, so it kind of sounds to everything and maybe that's kind of like a lame concert, but it really is. I think you need to be just optimizing the whole stack. And so like my team is like the AI team, so net what that really means in practice is we're focused on training the models, which is an important but not sufficient part of a whole stack, because we would be kind of dead in the water without the teams at HIT who think about how to actually kind of get the data and things TV systems and then the decisions out to the markets and keep up when things get busy, all these things. So when I think about our competitors, I think there is a benefit to scale. I can't imagine how you would start a new company like HIT in the year twenty twenty five because of the huge initial lift to kind of build enough engineering scale to achieve this sort of thing. And so I think are sort of peer companies also have invested very heavily in engineering, and we'll continue to do so. And there was an article in the FT like a little like a week or two ago about how firms like HIT are kind of extending themselves more into slower trading and there are firms that are kind of you know, those slower firms that's trying to kind of go faster. 00:37:17 Speaker 2: And yeah, I was just gonna ask about, just like on the prediction standpoint, Okay, maybe you could predict what's with some reasonable confidence what's gonna happen in the next hour. Sometimes, if you're lucky, maybe a day like maybe a month. It's just ridiculous. But do you in your work is that horizon? Has it broadened? 00:37:34 Speaker 4: It is? Yeah. 00:37:35 Speaker 5: I think one of the things for people who are aware of HIT even at all, I think is still a perception is sort of a pre twenty twenty perception of like we are purely high frequency trading firm, but we would say we are both high frequency and medium frequency trading firm, and it's like a big part of our business. One way to think about it, I think is that by if I really have a view on what a stock should be in like five days time, Let's say I want to buy that stock, I'm going to acquire that stock over time, and maybe it's what's the best time to buy that stuck over the five day period, Well, I have a model that tells me that the best pricing an hour. So maybe the shorter term model should inform the longer term trade and cascading all the way down. 00:38:12 Speaker 2: When you're doing this sort of slightly longer term or slightly slower frequency trading, is the fundamental job still the same, which is you're in the liquidity provision service business, just of or longer you want to hold that warehousing or does it some because when I think of a fund, when I think of a hedge fund, I certainly don't think of maybe to some extent, some of their strategies might be sort of liquidity provision. It's more directional. Is it still that or is the fundamental reason why you make money the service you provide? Does it change by definition? Change over that horizon? 00:38:44 Speaker 5: I think the

market making service provision does break down. I think that stretches analogy too far. I think you have to think of it as like liquidity taking, which somehow seems more like aggressive or something. But the we're trading against orders resting on the book. Someone was like, I want to sell this dock, and we're like, we will buy it from you because we think that in the long run will be worth doing it, and so we do cross to spread and we do pay this transaction costs. Sometimes, you know, you can also kind of acquire position by market making, but with a tilt. So really, at the longer horizons, I think the sort of market making service analogy does break down. But in some sense there's always a counterpartty and they wanted to trade for a reason, and I think a mental model that I don't know. You tell me if this sounds like too touishing rushy. 00:39:26 Speaker 2: But love a mental model. 00:39:28 Speaker 5: Yeah, you mentioned go chess, right, So the thing about those is that there there's zero sum games is only one winner. It's truly like a no, like someone someone's unhappy. Someone was in there maybe equally unhappy plus one minus one. I think the reason that trading works is because it is in some sense positive sum. You know, money is conserved, and I guess the little fee goes to exchange, So in some sense money is at that moment of a trade is actually negative a little. But utility people's general happiness I don't know, might paycheck go into my fur one k provider and it bias some ETFs. I'm relatively like insensitive to how exactly that happens. I just I'm not gonna look at it for never forty years, right, No. 00:40:10 Speaker 4: I try not to look at it, especially lately. 00:40:12 Speaker 5: But uh yeah, like the utility, My utility is a very long horizon, and so someone sells it to me like at one cent different, I don't really care. So but like the person who made the sense happy and I'm happy because I got good liquidity didn't cross a huge spread. So that is kind of why I think it all kind of makes sense and white people are trading together. But it's also why like thinking about markets like an alpha go sense doesn't make sense because it's kind of doesn't really apply. If you thought of markets as hit and all our competitives all kind of in some sort of like deathmatch, who's the smartest, who's trying to pick each other off, then markets would be kind of like this giant standof where no one would be trading. Everyone be kind to be like waiting. But obviously markets are very vibrant. I think it's because even when we were crossing the spread, because a crossing is prettingaintone who wanted to sell for whatever reason. If we're right, I guess in five days time they might be like less happy, but. 00:41:00 Speaker 4: Maybe they weren't. Actually, maybe they were just like hedging a position. They don't care what. 00:41:03 Speaker 5: The stocks prices in five days. They just wanted to like hedge their position, and we traded with them. So that's the way I tell rick and styles in my head. But it can still be like a sort of service provision we make mindly only because someone else wants to trade. 00:41:16 Speaker 4: If no one was trading, we wouldn't exist, right. 00:41:18 Speaker 3: And different market participants with different motivations and goals and aims. I want to go back to the talent question. Yeah for a second, and I get the sense that engineers like open source and they like contributing to the research ecosystem on AI. And then I get the sense that trading firms probably do not like open source, and they're much more into protecting their proprietary models or data or whatever. How does a company like HRT, how do you actually balance that tension? 00:41:47 Speaker 4: Yeah? 00:41:47 Speaker 5: I mean this is also like a sort of really honest answer in that many years ago. This is a relative comparative disadvantage for us for recruiting some We often have conversations with, maybe especially PhDs who are graduating, and they would like, well, I can go to Google and I can still publish my research, and that kind of gives me optionality. 00:42:04 Speaker 4: People will know who I am. 00:42:06 Speaker 5: If I go into an HRT or like firm, I essentially go behind this veil and I never emerge and people just had to kind of take it on faith. I did smart things for many years, and I would have basically no strong counter argument apart from the fact that actually writing papers is kind of overrated. I've been there, done that, as when you get older you will not care. Now though, there's this interesting situation where this golden era may be of like being able to be work at a big tech company and be paid for public research is very much over The papers that do come out of the big AI labs are essentially kind of either a very stale or not important, and if you're working on the most important cutting edge things, you can't share what you're doing and it's very secretive. So some since the problem solved itself a little bit for me, and people now recognize that IP should be protected. I've even seen some of us sort of AI lab people think out a lot about non competes in public thinking tweeting about non competes. 00:43:00 Speaker 4: And things, which is an amazing ton of events, because I feel like. 00:43:03 Speaker 2: That was very anesthetical. 00:43:05 Speaker 5: I mean, they're like literally effectively banned in the state of California, and I think people were almost like proud of this fact, and which also kind of hold it against the New York sort of trading world, being like, oh, look at these people, are there non competes and things, And then someone comes along and pays one hundred million dollars or whatever for like your researchers, and a lot of that money is being paid for talent, but it's also in some sense paying for intellectual property. 00:43:30 Speaker 4: And like, those people. 00:43:31 Speaker 5: Know how the soup is made and they are not writing

it down and not committing any explicit sort of IP theft. But if you hire five people who've been making the soup, they know they know a lot of process knowledge, and you might suddenly feel a little differently about protecting that. We spend a lot of time training our employees. Takes a long time for them to be productive. In some sense, it would be a shame if people could just take that knowledge and immediately leave. 00:43:58 Speaker 4: And so, yeah, just. 00:44:02 Speaker 3: Going back to the steamroller. I promised, I promised we would. When I hear AI in trading or I know people are very excited about agent based AI nowadays, part of me thinks back to one of the more amusing events in financial history, which is Joe. I'm sure you remember at the time that one of night Capital's algos. 00:44:23 Speaker 2: Would not find that to be an amusing Yeah, did all the worst Nightmare possible, but using for them the peanut gallery, right. 00:44:29 Speaker 3: Right, schadenfreud. So this algo went rogue and bought like seven billion dollars worth of stuff? Yeah, exactly, what are the guardrails that you put in place to avoid the destiny of night capital. 00:44:42 Speaker 5: So every training cycle we have a talk about the nightmare with a K and we have multiple x the night employees at HIT, as you might expect just from the lineage of a successful trading firm that ended in a kind of unhappy way, and we have many people who were at night. 00:44:57 Speaker 2: This story is crazy and successful trading firm that ended about fifteen yeah. 00:45:00 Speaker 5: Yeah, So it's fair to say that that stuff haunts us, and we try and take as many lessons away from that as possible. Defense and layers. So I think one of the things that I'd like to emphasize with the AI stuff in particular is that it is not like there's some neural network directly sending orders to NIZ. It is in some sense providing a plan and then traditional human, heavily audited, risk checked layers take the actions and that's just kind of how it has to be. And so for us we are kind of on an operational day to day basis. It's just many, many layers of sanity checking throughout the day, and then at a sort of high level it's very careful process including processes to specifically avoid the KCG type scenario of how are you even releasing new versions and what pre released checks do you run? And audits and we even during the day we have some I don't know, I guess you'd call them like sanity checks of the neural networks to make sure that they are producing the values that we expected they would be reducing. And those sort of checking processes are kind of a little bit behind because they can't keep up with the like flow, but like for enough to kind of just again like every tex of a numeric stability of the model saying and things. It's not it's not about losing money or making money in today, because it's not like, oh, like risk in the kind of financial sense. It's like operational risk. But paranoia is deep and that's probably something that's still very different I think, from this market from the sort of other AI world, which I guess anything goes and like failure rates to kind of just priced in. Yeah, but yeah, you could you could imagine just ruining everything, and I guess we worry about losing money, but I think we worry more about taking an action that a regulator would not want us to do, because if you lose that trust of regulators, you lose it for a very long time. And we trade in a lot of markets and we pay very close attention, and I have deep respect for the regulators and their decisions and all those markets and the rules are sometimes very complex, and man, do we watch that stuff like a hawk, because you know, you don't be kicked out of a country for making an operational error. And this is a very low tolerance culture from regulators in terms of making mistakes. So we stress it a lot, and I think we should because it's it's like the profit you make in ten years by still being in the game versus move fast and break things. It's not move fast and break things, PA still want to move fast. 00:47:16 Speaker 2: I have like a million more questions, but for the sake of time, I'll just ask one more. And I don't know even know whether it's something you're in position great position to answer about. It's something I actually want to do an entire episode about at some point. But as you would characterize it, what happens in the second after a jobs report is released. And what I'm talking about specifically is numbers either flash on the screen or a piece of a text appears on a website, and markets move around a lot all that and there's people. Then suddenly it's actually the jobs report was good, and if you actually look at the wage number and then the six But in that instant, in that first micro second after the release, markets are already moving, certainly before any human has had a chance to read the thing or for view. So what I assume is that there's training on here's the text and here are the things and whatever. But to as you would put it, or from the perspective of hr T, what happens in the millisecond after an event? 00:48:13 Speaker 5: Yeah, so yeah, I mean, so we have like a Bloomberg headlines feed that it's like pretty low latency, and if it's like an important articleize like a star and a feed things like this, right, But you can do everything from having kind of a hand crafted logic to look for keywords through to putting it through like an AI model. One of the things that I like still kind of kind of wrap my head around is, I guess, without saying specific company names, there are options trading firms that have thousands of people that are essentially cyborg trading options. They have maybe ten people trading like options for a single big stock like in VIDEOSA, and they are humans staring at the feeds for these things and clicking buttons, and they have user interfaces that will sit up for them to hit the

green button. 00:49:00 Speaker 4: Of the red button. Essentially very fast. It's weird. We actually want for a hackathon. 00:49:06 Speaker 5: We got a PlayStation controller and kind of gave people a chance to try and practice reacting to events. It's really tough, but it's a learnable skill. I think in an efficient market sense, this should be ai able. It is challenging though, because if you imagine to kind of plumbing it into chet GBT, it would be too slow, Like the latency would probably be sufficiently high. 00:49:27 Speaker 4: I mean it's not that fast, right. 00:49:29 Speaker 5: It's fast for any normal day to day thing, but for markets it's kind of slow also. And this is like a very interesting research challenge. Is like you can't literally use chet GBT to back test anything. It knows every jerme, heal speech, and knows what happened afterwards because it's trained on the whole internet. So how do you really get confidence that for the next federals of speech it's going to do the right thing. Traditionally in finance you back to us things to see how you're done in the past. But if in this case it's all could of ensemble, like it's seen it all before. And I've seen academic finance papers if they try and like grapple of this and they say it's still works. They try and account for this, but I know, just this stuff is really that smart. Yeah, The whole kind of thesis is that it's memorized, everything has been trained on, so why would it be reliable? 00:50:15 Speaker 4: And so when if you see someone being like. 00:50:16 Speaker 5: Oh I ran every federal reserves speech through giant GBT and it got it right like nine out of ten times, it's like only nine out of ten times, Like why not one hundred percent? So I do find that I do think that it is interesting there are how many humans that's still involved on relatively high speed trading. There are a lot of people still doing this and instead of niche products. And it's presumably because it's very hard to integrate all the information. It's AGI twenty I don't know twenty twenty eight twenty thirty. I don't know, there's still a lot of humans trading stock and options and so like, I don't know how to reconcile that, but I think about that. 00:50:50 Speaker 4: When I read fun being. 00:50:53 Speaker 2: Dunning, it was a fantastic There really are like hours more of conversation, so we can back next week next week's episode. But no, that was great, thank you for having really appreciate it. 00:51:03 Speaker 6: Yeah, pleasure, Thank you, Tracy. 00:51:17 Speaker 2: I thought that was really great. I like this idea, this sort of anti symicism, because you do hear a lot of people say, oh no, like AI could solve things like chess or whatever, but the stock market is fundamentally different, and I've never been totally satisfied with some of the theories for why. And like I get stocks are not like necessarily like a solvable problem in quite the same way. But humans make money on the market by matching patterns. Why can't smart silicon brains do the same thing. 00:51:45 Speaker 3: Well, there's also history. Now we have many years of HFT trading and yeah, gruthically driven trading where people have made a lot of money, So it seems to be working. The light bulb moment for me was where Ian talked about the timeframe and the importance of the time frame, and I think that's really the key in many ways. It's adapting what you're doing with AI to the data that's available and the data on markets. Most of it is going to be very short term and more seconds and minutes, more minutes than days, et cetera, et cetera. And a lot of the data is also biased to immediacy versus past analysis, which he spoke about as well. 00:52:25 Speaker 2: It is always funny in finance people. It's like, oh, seventeen out of nineteen times there's been this death cross of the S and P five hundred stocks went down. It's like any serious data scientists will spit at that sampoint. It's like beyond a joke level to talk about a sample size of nineteen. 00:52:43 Speaker 3: Yeah, but death cross in a headline. It's so tempted. 00:52:48 Speaker 2: That's true. You all, you cannot advice to journalists. Never pass up a chance to put death cross. I was glad to hear thought a few things interesting. One is I was glad to hear that the wire length problem is no longer. Yeah, it's not just as racing it closer to the extreme. 00:53:02 Speaker 3: I was kind of boring when people were talking about the Cold War and HFT and all of that. 00:53:06 Speaker 2: It's interesting that the GPU market is eased versus where it may have been a couple of years ago. And it's interesting they even at a scale a good trading shop, that electricity is proving to be a main constraint, which does raise questions about are we just going to hit up against a wall given some of the AI plans that so many people are banking on for the chatbots. 00:53:28 Speaker 3: Yeah, I thought also, I guess the cultural shift in some of the last Yeah, it was really interesting this idea that they've become more proprietary and perhaps more mysterious in some ways, rather than the trading firms becoming more open. Yeah. 00:53:43 Speaker 2: Lots of great conversation, answer some questions. Yea plenty more. 00:53:48 Speaker 3: That was helpful, and I'm sure we'll talk to him again, maybe not next week, but soon next year. All right, shall we leave it there, Let's leave it there. This has been another episode of the All Thoughts podcast. I'm Tracy Alloway. You can follow me at Tracy Alloway. 00:54:01 Speaker 2: And I'm joll Wisenthal. You can follow me at The Stalwart. Follow our guest Ian Dunning. He's at Ian Dunning. Follow our producers Carmen Rodriguez at Carmen Arman, dash Ol Bennett at dashbod and kill Brooks at Kilbrooks. More odd Laws content, Go to Bloomberg dot com slash odd Lots with the daily newsletter and all of our episodes, and you can chat about all of these topics twenty four seven in our discord Discord dot gg slash odlines. 00:54:24 Speaker 3: And if you enjoy odd Lots, if you like it when we dive into how companies are actually using AI, then please leave us a positive review on your favorite podcast platform. And remember, if you are a Bloomberg subscriber, you can listen to all of our episodes absolutely ad free. All you need to do is find the Bloomberg channel on Apple Podcasts and follow the instructions there. Thanks for listening 00:55:02 Speaker 4: In