Ex-Citadel Quant on Trading the Most Asymmetric Market - Neel Somani

It's table stakes to put down seven figures of collateral in order to seriously trade power. Joining me today is Neil Somani, a next Citadel quant who traded the power markets. One of the most opaque markets in the world. There's different types of hedge funds, and the role of a QR depends on the type of hedge fund that you're at. Citadel in the commodities group, it's not fully automated. The role of the QR is to build models that are ultimately used by traders. In the power markets, one cold week can move the price 100 fold, and being on the other side of that trade can wipe you out. If you're always long, you're just blindly long power, on average you lose money. It's no coincidence that people refer to power options, startups, asymmetry, all in the same sentence. We'll do my best to give you guys the best content we can with the best guests and best interviews we can do. Thanks and back to the episode. Neil, thank you so much for coming on the pod. Thanks for having me. You worked at Citadel as a quant researcher. I'm just curious, what does the day-to-day look like actually in that role? So I'll try to speak in generic terms because I don't want to answer what it's like to work at Citadel specifically, but there's different types of hedge funds. The role of a QR depends on the type of hedge fund that you're at. So if you're at a hedge fund like 2 Sigma, DE Shaw, then generally you're building out a model that is fully autonomous and it outputs relative scores for different assets that you might be interested in trading. Those scores ultimately get converted to a position. And then there's some execution model that tries to make that position into reality. So you'll place the actual trades in order to have that position on your book. So that's how 2 Sigma, DE Shaw would work. There's other shops that are more discretionary where you build a model and maybe that model is directly outputting a price. And then the question is how much is your estimated price already priced in? Are there factors that you're not accounting for, how accurate is it, and then the discretionary traders will take your model and decide what to do with it. Maybe they have other models that they're comparing it with. Maybe they have some intuition on why your model might be missing something and they'll make a decision from there. So sit it all in the commodities group. It's not fully automated. That's public information that it's a discretionary fund. And the role the QR is to build models that are ultimately used by traders. So I was on central research. There's also analysts that sit directly on desks like PowerDess, desks. And their work is more like day to day following the news. If you're a QR on central research, then you're mostly building out your model and you're not necessarily following the news as closely as the guys that actually sit on desks. And since you're working with the discretionary traders, I guess how attached to the actual P&L are you? You know, you're building the model. They're using it to trade. In general, I guess what's the separation to look like? So you're definitely aware of what everyone's P&L looks like. And you sit in on the desk meetings. So they have PM meetings that you would be listening in on. You chime in if it's appropriate. You'd understand what are their sources of uncertainty. You try to provide your own estimates and you might build out models to help them understand where they're lacking information. So to that degree, you're plugging into P&L. On the other hand, the amount of slope you get, meaning that the actual percent of P&L then actually ends up in your pocket largely depends on your seniority. So if you're super junior, you likely don't have slope at all. And that's true for funds across the board. It's not like a set it all thing. In general, if you're junior, you have some amount of bonus and it's sort of up to the PM's discretion or it's up to your manager, your manager manager. So that's how your comp ends up working. As you get more senior and as you're actually placing trades yourself, then you might negotiate for some percentage of slope. So that's kind of how P&L works from the perspective of a quant. And I guess from junior trader in junior quant researcher, are those two paths that both directly that you know sit in that seat for a while, keep getting promoted? Do you eventually end up in a PM seat or from the researcher role is it a lot harder? I'd say that from almost any role in a hedge fund, the ambition of most people is to move closer to the trader side of the desk. So everyone wants to be a PM eventually. If you're a QR, probably the next logical step is to be a head analyst at a desk. If you're focusing on gas, for example, you'd probably want to be head analyst on the gas desk. And then from there, you have the option to place your own trades. You might have like a miniature book yourself. You might become like an APM or you might become a full PM from there. So that's the direction you move in. Even like an engineer, for example, could transition into a junior analyst role and then move vertically from there. The most straight shot to becoming a PM is by joining as an analyst. And then you would just start placing trades from there. Before joining Citadel and working as a quant, what was the difference between the way you thought the role would be and how it actually was. And I know you can't talk in specifics, but just in generalities, what is your take on the perception versus reality of it? Well, I thought that a lot of it was based on market hours before I actually entered finance. But what I realized is that a lot of the time that you spend is more like process improvements. And that can be done at any time. You can even work weekends when the markets are closed or not actively traded. And you can improve your process, improve your model. And a lot of it is like looking at retroactive data or data in hindsight. So you don't really need the market to be open in order to do that. So that was probably the biggest misunderstanding that I had. I pretty much thought that the day was over when the market closes or when peak trading hours are over. If you just listen to Neil break down the quant path and thought, that's the job I want. Listen up. Also an open is partnered with Onix Capital Group, the world's largest market maker in oil derivatives. This July in London, they're hosting an in-person event for university students attend and you'll have a shot at an internship. They're also running a trading competition, a thousand pound prize, winner announced on the night. I personally know people who progress extremely quickly to senior roles while working at Onix. You do not want to miss this. Apply the link below. Also, I'll be at the event. See you there. As a commodities quatt, you mentioned something they're along previously, something along the lines of you're not looking at the news all the time. You're not actually the one sending trades. But I imagine especially for something like commodities, especially for that asset class, you need to have an understanding of the fundamentals of the drivers, even of the story. And what the different participants are doing. I guess what's the process for staying in touch with that side of the coin of like understanding that deeply when you're doing the deep research as well. So the fundamentals in the market structure don't really change. So that's kind of like table stakes. Everyone who's working in power or gas needs to understand where the price is coming from. When I say following the market, I mean like if there's a new headline that the start date for this specific generator has been pushed back by a month, that's the type of granularity that the top analysts or the top quants will keep up with. If you're just like starting junior quant, you likely aren't following it that closely unless you're specifically working on that desk and maybe that power plant is one that you've been paying a lot of attention to. But yeah, if you talk to a head analyst or something, they could probably name every major plant on the grid and they can tell you all the latest news about every single one. And that's certainly something that they follow pretty closely. And as far as getting the fundamentals down and power price saying or natural gas, a lot of the basics are just micro-econ. So understanding that just to begin with. And then the power market in particular varies on the region. So for example, I'm in California, this is California. I so I so was short for independent system operator. So you can literally look up the California ISO or the ISO rules. They'll tell you this is the auction that we conduct every day. And this is exactly how we compute the power price at each location. There are some reliability constraints that are not publicly released, but you can infer a lot of it, basically. When trading power, if you were to boil down what the foundation of edge is, what is it? Where does competitive advantage come from? So by and far, the foundation of edge in power is almost always congestion. And congestion means when you're not able to produce all the power you need in a specific region and you have to import it. And then that power line eventually gets congestion. You can only put so much power through a power line. And the reason why that matters is because so first of all, physically why do you why is there a limit to how much power you can put through a power line? Heat causes things to expand. So when you push more power through the power line gets gets a little longer and eventually might group down and might hit a tree or let it on fire or something. So that's the reason why power lines have their limits and they're called ratings. So once it's hit that limit, then you're not supposed to send any more power through it. And that causes the difference in prices in two regions. So there's two reasons why there's a difference between two regions. One is that power is literally lost. on the power line from point A to point B. But the second more important reason is that sometimes you literally can't move the power there. So people will trade the difference between those two prices and they'll say, I think that this specific power line is gonna be traded. So I'm gonna buy at one location and sell at the other because I basically think of that spread is gonna increase or decrease. - And how is that modeled on a granular basis? - So you definitely need a sense of how much generation is gonna be demanded at each of those two locations between the transmission line. But you also kinda need a global view of how the power price is being computed nearby. So for example, if a bunch of powers demanded at, so I'm trying not to refer to some topology in my head that's too complicated to communicate in words. But let's say some region is already exporting as much power as they can. So their local generators don't have any capacity. Then that means that they have to start importing power to meet any local demand. So that's gonna impact the power price because it means that importing can cause a line to be congested. Whereas conversely, if there are really no imports or exports that are relevant, then that means that you can meet all the demand locally. So you need a global view of how powers being demanded are supplied at all the nearby regions and generally you solve the entire grid at once. So you don't just solve for the price of one location. - What are the steps in the call it trade generation or process? You know, actually from ideation, you know, you identify that you think there's gonna be a dislocation to actually sending an order, right? What does that pipeline look like? So it depends on what you're trading. So if you're trading what's called the hub price? So the average price for a region, that's different than the sort of congestion trading that I was referring to before. So let's talk about hub trading because it's probably the easier version, which is obviously the power grid has many different prices. There's a bunch of different points. And like I mentioned, they all have different prices given that there's congestion. So the hub price is sort of some even average across all of those. And if you're modeling the hub price, then a simple approach that many people take is they start by first modeling demand. And they try to figure out how many megawatts or gigawatts of demand are gonna be in this region. And they usually do that by looking at the weather. The weather matters because when it's really hot or it's really cold, people turn on their AC so they start demanding more power in order to support that. So you get some estimate for what demand is gonna look like. And based on that, they'll chip if you're like a gas trader and this is how you'll think about power, you'll think about which power units need to turn on in order to meet that demand. And if, for example, there isn't that much demand at all, then it's all gonna be met by renewables. If there's a lot of demand, then you're gonna have to start turning on less and less efficient units and that's gonna drive the price up. So using all that information, you'll ultimately come up with some estimate for what you think the price is gonna be for some specific period in time. And then the final step is you compare that to the market prices. And if the market price is way different than your estimate, then you should probably take a look in the mirror and figure out whether your model's totally off or whether you have ads and whether you should actually place that as a trade. - Our orders actually set for a power trading business or power trading fund. - So you would onboard directly to the independent system operator called the ISO. So California has like a way that you can onboard. You'd put up collateral given that if you're going short, these power prices can spike really high, especially if you're in a region like Texas or Alberta, Canada, where they sort of build that into the energy pricing. So for that reason, you put up a bunch of collateral. So people sometimes say that it's table stakes to put down seven figures of collateral in order to seriously trade power. - To seriously trade power, it's a very capital intensive business. And it's not something that I guess retail can access or even a smaller fund can access, would you say? - I mean, you could go long. The issue is that being always long for power is a structurally losing trade. And the reason for that is if you have any skew asset, so an asset that has a 1% probability of the price being really high, then in general, the price of that asset is gonna be a little bit higher than expected value. And the reason for that is twofold. One is that people want access to that asymmetry. So think about startups, for example, startups always raise at these ridiculous cops, but when you think about it, it's kind of logical because nowhere else can you get access to some asset that goes from zero to billions of dollars. So people want assets like that. And the second reason why it's systematically a little bit higher is because no one wants to go short. Because if you're short, then there's a chance you can lose your shirt. So for those two reasons, if you're always long, you just blindly long power, on average you lose money unless you have some specific reason while you're picking one region over another. So that's the only viable way that you could theoretically trade power without putting up a bunch of collateral. But yeah, I'd argue that you're missing out on most of the opportunities in power, which are like I mentioned congestion trades. - How do you think working as a quant researcher within power trading shaped how you think about edge and competitive advantage in, I guess, every domain? - Yeah, so I guess after I left power, I've been an entrepreneur and it definitely influenced the way that I think what startups in general, given that startups are also skew assets. So the way that they're priced is sort of similar to options is sometimes how it's described. It's priced similarly to power. And I use all these terms interchangeably, but it's not as loose as I'm describing it because sometimes you literally do price power as an option. There's this concept called a heat rate call option, which is effectively the P&L derived from a power plant. And sometimes you'll sell one of those if you're financing a power plant. So my point is that it's no coincidence that people refer to power options, startups, asymmetry all in the same sentence. So that definitely influenced my thinking, just the deep understanding of asymmetry and how that impacts the prices of various things. The second thing that it influences that whenever I see headlines related to supply chain constraints. So this is a popular thing that's happening in AI right now. People say, oh, we don't have enough samis or GPUs or power or whatever else. It seems like there's a million things that people need. And the question is, are those all binding constraints? I think in terms of binding constraints a lot. And what I mean by binding constraint is when you look at the price for a given item, the cost to increase or basically the price of that item in theory should be the cost of producing an additional unit of that item. And the cost of an additional unit of that item is basically the cost of the item that you're whatever input you're missing. So for example, if a cost, let's say I have a ton of wood and I want to sell tables, then wood is not going to be the binding constraints. I'm not going to price the table based on the price of wood. Instead, it's going to be labor. Because to produce another table, I need to pay for some person to actually construct it. And at some point, I'm going to run out of people. So I've got to raise the wage. And that's what causes the price for tables to increase. So that's the way that I think about pricing for a lot of supply chain related problems. And that was influenced because of you analyzing constraints within the grid. Was that-- was that what caused that? Yeah, that's basically how the power price is set. And that's how a lot of commodities prices are set as well. They're based on these sorts of binding constraints. And you might not hear that phrase as much if you talk to traditional analysts, given that they might not have a full optimization model that they're building for every asset that they're pricing. But a lot of the quants in this space will basically think in these terms. So let's sit it out. You work on power trading. What else were you doing? So every power person also does gas. It's pretty much impossible to understand power without understanding gas. And the reason for those who don't even know the basics is that you have to burn natural gas in some cases in order to produce power. And in some cases, you burn coal, like what we're seeing in Miso, which is in the Midwest. And those are two of the thermal sources of fuel. So if you need-- or need to have the thermal sources for power. So if you need steady power, then oftentimes solar, wind, or not going to be suitable. So you have to turn on one of these types of units. Of course, there's also a nuclear, but nuke isn't available everywhere and it's super expensive to build. So that's the reason why natural gas and coal are so popular. And those are often the last unit needed to turn on to set-- and therefore, they're the ones that set the price by that reasoning that I was giving earlier. So because natural gas and coal are often the Marshall unit, you have to understand those deeply in order to have a good estimate on what the power price is going to be. Why is it that hedge funds often trade power and natural gas and not really something like oil? And at least in oil, from what I know, they tend to struggle. So there's a few reasons. One is that if you're just looking at power in general, oil is not generally used to produce power. That would be like in extreme scenarios like in New England, for example. In the winter, all the natural gas becomes very constrained because people also use natural gas for heating their homes. So that uses up some natural gas. And then there isn't enough natural gas to meet the power demand. So once you're out in natural gas, then you have to start turning to-- dirty or units like diesel. And actually a lot of people are starting to use diesel power for all these temporary power plants that they're using for AI data centers like colossus, just XAIs power plant my understanding I don't want to over speak, but I think that they're using a lot of temporary sources like diesel. So that's basically one reason why people don't necessarily do oil just because they do powers because oil is not really used to produce power under normal circumstances. So that's the first reason. The second reason is that petroleum is much more globally influenced. So as we saw with the straight of her moves, a closure on the other side of the world can dramatically impact petroleum and petroleum product prices here in the US. So as a result, people like to stay away from geopolitical risk. It's really easy to lose a lot of money on stuff like that because it's hard to predict the behavior of governments. So see, those are some of the reasons why I think trading natural gas here domestically and trading powers a lot easier. And there's not really like a global element to natural gas in the United States because we already output we ship as much natural gas as we can in the form of liquid natural gas. So it doesn't matter if the price abroad went up by $10 or $100 or whatever because we were already shipping as much as we could to them. It doesn't impact the price domestically. What are the biggest risk factors in natural gas and in power trading? Well, the biggest thing is that it's so asymmetric. So the worst thing that can possibly happen in power for a power grid is not meeting demand because people can't really curtail their demand that easily. Like, of course, there's things like data centers which in theory could be shut off. We have, you know, my little brother playing his PlayStation. I'm sure he can stop doing that. But there's also power demand like hospitals. There's things like keeping the the AC running when it's intensely hot and people will die of heat stroke if they can't have their AC running. It's things like keeping the heating running when it's super cold like what happened in Texas on if you don't turn on your heating then your house could freeze over and you could die. So for those reasons and those are those tend to be the times when powers race scarce. So power demand is in elastic and not being able to meet that demand means that you have to start clinging to extreme sources. If diesel is around then fair enough you're going to go to that. But what if there's not even diesel. What if there are no units that can turn on then we have no choice but to start having blackouts or brownouts you have to shut down parts of the grid. The power price is going to hit pretty much the highest price possible. Really extreme things can happen. So if you're short during periods like that. Then of course it's going to be very dangerous. You're going to lose your shirt. It's also really attractive to be short during periods like that from a sense of the price is already going to be really inflated. People are really afraid about these freak events happening. So going short means you can collect that risk premium. Of course sometimes the risk goes against you and then you lose all your money. Can you walk me through an example I guess story of a typical trade a power slash natural gas PM would put on and if things went bad the things went wrong how they would navigate. Taking that position off. So probably the easiest example is like March like February 2021. So this is in Texas where the natural gas units froze over some of the wild heads even froze so they didn't have enough natural gas. It's really cold so people need their heating on and if you looked just a few weeks earlier there was basically no indication that it was going to get that cold to the point where there was going to be extreme temperatures that led to a total shortage of power. So if you were a little bit overconfident you might see the prices spiked a little bit but maybe my weather model says we're going to be okay. So you decide let's go short and then like a week or two later then you'd update your weather estimates and you might see that actually now it's a little bit more likely that we're going to have this freak power event. And the prices and go a little bit higher because the market's pricing that in as well. But you might still think it's your right. So you might actually increase the size of your position might say let's just double down and we're going to we're going to make up for the loss that we've already incurred. And then you're like a week before the actual free is actually happens and the problem with weather is that even a week before we still don't really know what the weather is no look like. So now you think there's like a 30% chance of this event occurring and at that point if you're a good trader you might say you know what it's gone against me here it's not worth the risk of blowing up because I know that in Texas if the power price really is at the maximum is going to go to $9,000 per megawatt hour. As opposed to a normal price of like 50 to $100. So it's going to go like you know 100 times or you know 10 to 20 times higher than what it's supposed to be. So so I'm going to I'm going to just close my position. I'm going to take the loss. And then a week later happens the day of happens and you actually get to see the weather and that's when the free is happen and and the price actually settles. So so that's the way that a power price or a bad power trade would plan. What are the most important things to model I remember you mentioned demand but you also mentioned whether they're I'm just curious I want to get a picture of what it is you should be looking at if you're sitting in this seat. So the most obvious factors what you just mentioned which is weather almost all power traders are watching the weather very carefully because that influences demand. The second thing that you're watching is things like outages if a unit is not online then you need to subtract that from your estimate or from your model you that supplies not going to be there. You're looking at so so let me just break it down so you have demand which is basically whether you have supply which is influenced by outages you have fuel costs. So gas price for example is also directly influenced by the temperature given that some people have natural gas heating. So you need to account for that if there's some local coal price like a coal shortage or something you had to account for that as well. So fuel prices. The last factor in supply would probably be like renewable availability so if you have a bunch of wind generation but it's not a really windy day then that wind capacity is essentially going on used or if it's solar but it's not a very sunny day then you don't really have access to that solar power. So you'd want to model things like renewables fuel prices outages new plants coming online you have to have the good picture of what supply looks like in order to see how you're going to meet that demand. And in general once hedge funds model these things do they tend to be short or do they tend to be put on long positions I guess in moments when they think the price will shoot up. So it totally depends those are two different types of trades the one trade is if you're long one area and short another that's called a basis trade and basis trade to some people they might think oh that's like buying spot and selling the one year forward that's how what a basis trade means and bonds but a basis trade in gas or in power is generally referring to two different regions. So that's one type of trade the others directional meaning you live in the prices and you go up or down and it depends on the desk depends who you're working with I do think of like a lot of the multi asset managers these days tend to be delta neutral so so they basically are always both long and short. Is there not more money to be made in the directional positions taking a big one and if if the if you have edge and if your PM has done this for a long time I guess you know what are your thoughts on that. Let me think about that so I think betting on congestion is an easier thing to model so so a basis trade is inherently a congestion back right where's a directional trade maybe you do that on like the hub price if you do it for a specific node I just think there's too much variance it's it's not a super common thing to do in my experience that said trying to think about cases where you would just straight up go long maybe in like a specific zone if you know a lot about what's see the thing is that if you're modeling one zone you're almost always modeling other zones as well so you might as well put on a position on one of the ones where you think that it's over price or something and go short that region it's safer and it can also be more profitable so so I think it's part of the reason why people tend to do basis tritz. What leads to disaster in power trading in natural gas trading what is what are the mistakes people make in this space. So the big thing is not accounting for for skew events so a low probability event that sends the price shooting that's the number one way that almost everyone loses their money in power or gas trading and beyond that. If you're like working in a normal regime meaning that there are no binding constraints there's no crazy supply shock you're probably not going to lose a ton of money but you also don't make a lot of money during periods like that either so it's like those skew events are the most interesting times to trade that's when you're supposed to be trading aggressive way it's just also the time when you're most likely to lose a lot of money. You know for setting up for those situations what are the ways top hedge funds build competitive advantage you know what are the specific things they build out what town do they recruit I guess how does a well oil machine look like for power trading. So I think a lot of it is learning from history. So if a hedge one missed a really major event, then they might maybe even over index to events like that looking forward. So maybe there was some massive geopolitical event that really influenced or impacted someone's book, then they might think, "I know we need to hire an expert in that field so that we don't make the same mistake next time." Weather is one of those common places where people make acquisitions that will aggressively hire weather startups, the Ohio weather experts or professors at different universities. You can model the amount of water that comes from glaciers or from all sorts of different sources for hydropower. So that people who are specialists in specific details that drive a lot of the variance, whether it's on the supplier or the demand side, those tend to be the people who are hot commodities in the field. Why do you think Citadel has been so successful in not just power, not just commodities, inequities, and almost everything? So I think a lot of that is about attracting great talent. I will say that the approach between different groups is probably not the same. I can't speak to any of the groups that they said it all. I can't speak to commodities because of NDA. And then I don't know anything about the other groups. So I can't say what their process looks like, but I do think that Citadel tends to attract a lot of top talent, just partly because of its reputation, partly because they do pay well. They have good mentors. It's generally like a good, reputable fun. I think there's sometimes due to this controversial, if you look on like the GameStop community or something, but among people in finance, I think it's a pretty uncontroversial hedge fund to work on. What are you doing now? What research are you working on? I guess. And how did your experience as working as a quantum researcher, our quantum researcher at arguably the greatest hedge fund of all time? How did that shape? But you want to do today. So when I first left, I started a blockchain company. I was doing that for a couple of years. And these days, I find myself thinking mostly about AI like a lot of other people. AI is impacting commodities a lot, just given the massive power demand here in the US. There's different inputs to AI. There's data, compute, which is related to power and commodities. There's fundamental research. So I find myself thinking a lot about those three things. I think the data play these days is basically like you define some benchmark. So you define some task that AI models are sort of okay. Some of them get 60% right. Some of them get 30% whatever. You get a lot of attention on your benchmark. And then you announce it publicly and then different labs want to saturate it. They want to perform really well on your benchmark. Because you're the one who defined it, you have the best data for it. And you sell that data. That's almost always the shape of these RL environment or data type companies that are popping up in AI. And they're super popular. And the question is, what's the longevity to ideas like that? Because eventually labs do have enough data. And their model will start out performing humans. That being said, I guess that same critique could apply to commodities plays. So I think a lot about the power demand and the stated needs by these AI labs. If you ask people at opening IR, if you see online, they basically say that their power demand is like unbounded. It's just going to keep growing and growing. And that might motivate you to go and build a power plant right now. And I'm definitely thinking about a lot of natural gas plays. To lab had this tweet, those recently, I thought it was pretty insightful, which is that almost every glut that he's observed, or no, it's that he's seen many gluts that are not followed by shortages. But every shortage he's observed is almost always followed by a glut. So whenever there's a shortage, people overdo it because it's hard to predict the behavior of other market participants. So everyone rushes to supply power. And a lot of it is mis timed. So maybe a lot of it will come too late. And as a result, it will be more than enough power is according to to lab stereo applied in this case. That said, I think that there's, indisputably, a lot of power demand right now. So question is like, how do you meet the six to 12 months demand for power without necessarily locking yourself in? Or maybe you should lock yourself in. Maybe that is the play that you should do. If you think that the lab's power demand will keep increasing. So I spent a lot of time thinking about that. And I think that temporary generation, even dirty or use sources like diesel, start to become more interesting in cases like that. And it kind of depends on what the labs propensity to pay is. Like people can say they really need power. But there is a price where it becomes unaccommodal. If you quote someone $150 per megawatt hour, maybe that's too high. Whereas something like 50 bucks per megawatt hour, there's no way you could finance a power plant on the order of six months in a way that's actually economical. Because let's say you're building like a 50 megawatt plant or something, I think conservatively that's in a cost, let's say 80 million, let's say 100 million for simplicity. So maybe you're paying like 20% interest year over year on that. So you're spending 100 million, you have to take out debt. Then it depends on your local fuel costs. Maybe it costs you 50 bucks per megawatt hour to produce each megawatt. If they're paying only like $80 per megawatt hour, then you're really not making that much per year. It's probably not enough to really justify this huge, this huge liability if you're taking on. So my point is that the actual $1 amount the labs are willing to play to pay heavily influences, which projects make sense and where they should be built. So I think a lot about site selection and things like that. And the last thing I spent my time on is basically machine learning research. So I do some stuff in reasoning related work. So it's AI and math that's some interpretability work. I don't know if you're familiar with this field called mechanistic interpretability, but yeah, it's some subfueled within AI interpretability to make sense of these models. I've become less interested in that that's specifically area for some more theoretically grounded reasons, but yeah, that's kind of how I split my time between commodities related work, thinking about machine learning research and some data related stuff. I would imagine you have differentiated takes on the future of labor as well. Is junior talent fucked? It's a question. So I think it depends on the field. And what's good about junior talent is that they have an invested decades of their life into some skill set and they're flexible. So junior talent can always pivot into different things. I think that there's two different types of labor concerns in the US. And some of it is driven by the coastal elite. So these are people who have held jobs like maybe government jobs that are being like, oh, or maybe there are software engineers. They've gotten used to a cushy lifestyle or certain type of lifestyle. And they don't know if they really want to learn a new skill set. And then you have the second set of people that are concerned about jobs, which is blue collar people. And I think the latter bucket is a lot easier to appeal to. In the sense that there's been some mismanaged messaging, I think related to data centers. If you look at the bills that are being passed in congress related to data centers, very few are pro data centers. There are things like in PJM, which is one of the major tradable regions. They're accelerating the interconnection queue. The interconnection queue is one of the bottlenecks to building a power plant that's actually connected to the grid. We can go into all the power plant and behind the meter compute stuff if you're interested in that. But there are a couple initiatives on accelerating federal pathways to getting right of ways and things like that. But by and far, the vast majority of data center related regulation has to do with not impacting retail or residential consumers of power, whatever that means because commodities are all interconnected. So it's really hard to account for every externality. They're related to transparency. So in some cases, I don't remember what I think it's like Senator McGiver or something. There's some bill related to releasing the water demands for any power plant you build or whatever people care about. And that's actually not that big of a deal to me because whenever you create a power plant, you have to do all that deconflictation anyway. So it's just making that information public. So yeah, basically, I think that the emphasis from politicians for data centers should really be on the jobs that are being created. Blue collar drops in particular. We need people to man the power plants. We need people to construct them. There's a huge immediate power demand right now so we can pay pretty well for blue collar labor robotics isn't there yet. So I think that's that fear is probably a little bit overstated from blue collar workers that they're getting replaced. The more legitimate fear is from people like software engineers. And I think a lot of that seems to be moving toward like agent management in any of these AI labs. Almost everyone says that they barely right code themselves. It's all agent management. So I think being really proficient in that getting more proficient in distribution and go to market seems more relevant now. But I think if people basically need to broaden their skill set beyond just plain software engineering or at a minimum, this is kind of a cliche thing to say, but they have to redefine what software engineering means to them because it isn't what it was five years ago. For an engineer today, how would you think about building a knowledge mode? Because I see a lot of advice online about getting up to date on the tools or like you mentioned learning how to work with agents. But the bar keeps rising. And I do think it's honestly a crime that the advice given to young people, to junior people is to focus on using the tools better because I think that's a game where only the top can win. How would you think about building up personal differentiation today? So I think a lot of it is about being more ambitious and maybe going beyond what colleges are emphasizing right now. Berkeley has a big emphasis on software engineering as the way that they teach computer science. But if I were restarting today, I'd almost certainly be looking more seriously at hardware. So things like an minimum GPU kernel optimization. So the way that you write code for GPUs is called a kernel and optimizing that so that they're extremely performant, understanding Nvidia GPUs or alternative GPUs very deeply. I think that's going to be a valuable asset at least in the short term. And having a valuable asset in the short term might be the best you can ask for. Given that life moves, things change. You'll have to adapt. But at least for the next couple of years, I think being a GPU expert is a really big deal. Robotics is still useful. All this LLM stuff is not super useful for robotics. People are trying to do sort of like end to end different differentiable models for robotics and there's VLMs and things like that. But for the most part, robotics literature hasn't had the same step function increases in the same way that language models have. And a robotics seems a little bit more robust at least in the near term. And it's hard to do robotics without doing actual robotics literature. You can't just come from an LLM background and suddenly start doing robotics. I would be looking at as an entrepreneur, much more ambitious ideas, things like nano-tech, if it's possible to innovate in that. I think biological applications of LLM. So there's a guy who supposedly cured cancer in his dog or something like that looking at genomics. I don't know if that dog cancer thing was legitimate by the way. I didn't like to be deeply into it. But my point is that if that is legitimate, then that's a very interesting direction. And I would seriously pursue that if I were a student and if I were building a new startup or something. I think setting the bar much higher than I'm building a SaaS startup and I'm just going to write some code and expect it to start printing money. I think if people have to start raising the bar for what they're building. That's that there's a lot of really near-term opportunities like the ones I was matching earlier. Like selling data, creating a benchmark, going viral and just selling data for it. You don't necessarily need a company for that. You can sell eight figures of data. Literally, that's the price of the people I'm quoting. I've talked to some of these data vendors. And I think that's like a great way to make a bag if you're like a young person who just wants to tell at least be financially comfortable. Meeting short term power demand. I think that's a little bit more operationally intensive. Maybe that's not the best area for someone to venture into. But picking up skills that are in the next six to 12 months demanded by the labs or going to be useful applications of what the labs are building. I think that's probably the skill set that people should be optimizing for. Do you think a quant job today is something that people should shoot for? I just think that the opportunity to cost from AI is way too high. And I think it'd be a shame for someone to go become a QR and miss out on what seems to be like the greatest technological development, possibly ever. It's like an extremely important technological development that I personally wouldn't want to miss out on. And even if you're really interested in commodities, go do some AI related thing for commodities. I think there's plenty of place to be done. And that's that. It kind of reminds me of how software was, I don't know, like 10, 15 years ago. If you're interested in like medicine, one of the best ways to contribute to that rather than going and getting a medical degree is to work on like a software startup in the medical space. Now I think it's like all AI adjacent instead, which rings like a similar battle. It's still software. But I think if there's just a bunch of immediate applications and opportunities that would be a shame to miss out on by having your heads down, just looking at trading charts. How do you think about risk taking in your personal life today when things are so uncertain? What is the optimal way to ride this trend, I guess, in a general sense? So what's funny is that you'll find that guys who take risk for a living, especially those that had funds, tend to take the least risk in their personal life. They buy index stocks and they're in index funds. They don't do anything that creative. I don't want to advise on risk taking because I also don't know if I take risk in an optimal way. So I don't want to advise people wrongly. But I just think that as a student right now, because it's so difficult to get a job anyway, I would really be thinking about ambitious ideas that you could build using LLMs. And even the, even things that the LLMs are not capable of now, think like one or two years out and start building for that future. And eventually the LLMs will catch up. That's kind of how I think about the types of risks that are worth taking right now as a young person. What's an example of something once, two years out, the LLMs can't do right now. The LLMs don't have the capability for right now. But in one, two years, this could be the future. This could be something viable. Any ideas, any examples? So I think biology is the most obvious thing that comes to mind. And I think there's a lot of people talking about that. I think thinking about the widespread proliferation of robotics, lands a few more ideas like machine repair, for example, building a business, maybe it starts as a services business. There aren't a lot of automated machines that are deployed in the wild right now. But started thinking about how you could build your own robotics to repair all their robots that will be deployed. And having a really serious company that's well positioned for when that future arrives and partnering with the right companies right now, that seems like an idea that would make a lot of sense. Things like warehouse automation, maybe the robotics aren't quite ready to automate all the warehouses, but picking like a small subset that works now and just trusting that robotics development will get better over the next one or two years to eventually automate all of them and doing like third party logistics or last mile delivery, things like that. And there's the types of ideas that I think are forward looking and are also pretty viable. You mentioned before we started the podcast that when you left Citadel, there were a bunch of other people who were there who also left to build out their own companies. How did it pan out in general? Just curious. So it's always true that most people who leave and start a company, the company doesn't work. It's the first time and I just met with someone who's at OpenAI. I live in Mission Bay, so I'm close to their office. And he was sort of in that same cohort as me and started a blockchain company, blockchain company didn't work out. Then he built some really cool audio model, that model ended up getting acquired by OpenAI, so he works there now. I have friends where they started something, maybe it didn't work. Now they're working at some other company, like a hot startup. Those companies have really ripped folks working at AI coding companies like cursor cognition. A lot of folks that ended up at XAI getting acquired by them. I think a lot of these startups tend to have soft landings, meaning that in the worst case, people will get acquired. There are a few cases where it didn't work out and the person is sort of still iterating and they're trying different things. Maybe it shows early signs of traction. So I'd basically say like maybe 5% of cases, fewer than 5% of cases really ended in a home run outcome where now that person that I know is a billionaire. And in the remainder of cases, maybe 20% of cases, the person ended up financially steady, but they didn't necessarily have a home run outcome. And then the remaining 75-ish percent of cases, there was some sort of aquahire or maybe they started another thing that ended fine, but that's usually the distribution, I think. What do you think is the stupidest way to play one's career in this regime today? I think building something that's unambitious, like an unambitious software company seems to make little sense. I think it'd be really difficult unless the idea is you're building many of them, like you have a portfolio of like 100 vibe-coded apps and you're just kind of throwing darts and seeing where they land and you're doing aggressive go-to-market. Like if you're someone like the Calais, I guys and you have some massive go-to-market muscle, then maybe that's fine. But I will say that like for most people, I think that's sort of not the best use of their talent. If they're a really strong research thinker or they're a really strong engineer, then I'd still lean into that ability, but I wouldn't try to apply it to software because you're forcing a square peg in a round hole. Software is not designed for that anymore. Software is designed for like quick wins that you can turn out quickly, make a small bag and then move on. But it's not super, I don't want to dismiss all software altogether. That's a little bit too extreme and Paul Graham had some tweet earlier, like a hardware startups are getting funded more aggressively because investors are scared about the longevity of software startups. He views that as a mistake. I don't know what his special insight is on when software is still defensible, but for that reason I don't want to write it off because he's better at the stuff than I am. But I will say like my instinct as a young person who's more like engineering research leaning is to not build like it quick software startup. I don't think that that's interesting enough. I don't think that it has enough of a mode. I think it's worth building something that's like something that's at the limit of what you're able to do and like really stretching your abilities because that means that it's at the limit of what anyone can do. What do you think software looks like in 2030? What is the what is a profitable software company look like? I just think of like the concept of apps and like individual websites and like these isolated silos that exist today might not exist along the future. 2030 is a while from now so you can know what happens the next few years and maybe 2030 might be even too soon for this kind of transition but I think that I like that idea of software being generated on the fly as we need it and maybe having some shared infrastructure layers like the way that Twitter has tweets and we're all kind of surface different views based on our personal algorithm. I think that direction but for all content seems to make sense to me. TikTok is sort of influenced by thinking on this. I feel like the TikTok recommendation engine is so good and there's also a lot of trash posted to TikTok that basically no one sees. Probably 99.99% of the videos posted to TikTok are just things that it's just garbage but TikTok is able to intelligently surface what's actually useful. So just like shared data wells and I don't think that it would be on a blockchain or anything. I don't think that it would necessarily be like shared infrastructure but some way of having like base APIs that agents can quickly build software on seems like a logical end state for a lot of the stuff. What skills win in this current regime for JunuTon, for JunuTon, for everyone in general what is the most important thing that you see studying these things and researching these things deeply. I mean people who are interested in and building things that are within the lab's opportunity sets on a six to 12 month horizon. Those are by and far the juniors who have leapfrogged the fastest. So someone who's building video generation models and open AI really wants to improve their video generation model and they just hired this guy with a huge package or maybe they the kids started a small start up and it gets acquired. Things like that I think are those types of approaches tend to be very profitable right now. So that's what I think a lot of young people should be indexing on if they have the research chops for it. If they don't and they're more of like a go-to-market person then leaning hard into that and developing a huge platform of hundreds of thousands or millions of followers. But whatever you're doing I do think that there's like massive power laws right now. So you've got to be the top 0.1%. So I relate to the sentiment that you mentioned earlier like you don't want to blindly tell everyone she just become really good at the tools because only the talk people really profit. That said if you think that you could be one of the top people at just using agent to coding tools then you should by all means do that. If you don't then you should go pick some other skill where you are going to be the top 0.1% and there's enough different types of skills especially useful skills out there that there should be enough room for everyone as long as someone has a high enough agency. But yeah I wouldn't settle for being like I don't know median the median software engineer. I don't think that the median software engineer is in a great position. What do you do if you're the median software engineer? There's median talents in general because my thinking on this is your median you're screwed. You know that is like my those are my honest thoughts. I just think it's that's just the the world we live in. I think it's a shame but it's it's true. What are your thoughts? So I think like the way that people should think about their careers is similar to how startups kind of think about their own positioning because the top 0.1% of talent is kind of like the open AIs of the talent world. The guys who are the winners right now they have some opportunities that they have some things that they want to do and it applies to every field. I think in academia you have the Terrence Tows who produce so many more proofs and so many more interesting results and everyone else. Yet like a lot of mathematicians still produce meaningful results despite the existence of Terrence Tows. And it's because the only Terrence only has so much bandwidth and he only has some fields that he's more interested in than others. So there are other areas for people to focus on but if you're the median software engineer I would not just like hold still and hope that things get better. I'd seriously start looking at other skills that you can develop that aren't in the direct opportunity set of some of these labs like for example I was talking with one of my friends who has a bio-related startup and there's a lot of people who are building web labs and things like that for LOMs to reason through experiments and I think that category or that shape of company is so popular because open AI is unlikely to build a web lab. They might start even designing their own hardware. They might build their own power plants. They might build their own benchmarks and collect their own data. But bio like a wet lab is just like it's one step too far like developing a physics lab building like a mineral mining facility. Like there's so many like random things further and further down the list that it seems unlikely that a major lab is going to seriously consider this as one of their top priorities to allocate resources to and opening I had this initiative called AI for science. I think that's been largely deprioritized. I do think that there's like some emphasis on bio because I think there's some PR benefit to it too. If you can cure cancer or do something like that that's like good for PR in general and it's also a useful thing to do. But I think there's a lot of other things which are worth the time for them. They should just take that energy, put it back toward recursive superintelligence and try to make that happen one month faster. That's I think a lot of the labs are thinking. Final question. Do you have any contrarian ticks on where the world is headed or the state of quantitative finance or anything at all that we've talked about. What is one thing you think you're right about that everyone you know would disagree with you off? Well that's a good question. Let me think for a moment on that. Okay so I have a couple of takes. One is that I think that coding tools or coding as a use case for models is not going to end up being a very profitable use case. So all the coding agents start ups things like cursor which is likely going to get acquired by XAI. They're probably going to exercise that option. Quad having a lot of demand from software engineers. That use case I think is basically going to sharply disappear. And my reasoning for that is one the switching cost is super low. I switch between these models like literally day to day. If a model fails to do something I switch immediately. And then I just start using the new model until that solves the problem or until that fails. And then two I think at open source models will end up being good enough at coding where people will just pick it because it's free or it's exceptionally cheap and it's good enough. So I think if there's been a lot of investment into coding related tools like cognition cognition growth has been incredible and it's like really cool to see that. I wonder if like they need to expand into something else before the rug gets pulled beneath coding related tooling or maybe I'm just wrong. So that's one take that I have. And then another take that I have is I think that we need a new paradigm for valuing entire sets or categories of companies. So for example a lot of these RLN or the data type companies that I was referring to. I think that we do see it a little bit like the multiple they get on their revenue compared to like their actual startups valuation when they raise a series A or B. It's a little compressed compared to what software valuations used to be or compared to some of their peers and other categories. And I think that that's justified because the data sales like I mentioned are sort of one off. Eventually these labs do have enough data in whatever category they're selling in. And it becomes kind of a cat in mouse game where they have to keep getting more and more rare and marginal sets of data which are also less and less valuable. So I basically think that that's one example. But there's other types of companies where they're selling right now to humans soon they're going to be selling to agents. And when you sell to humans you can do things like incur a cat. You're going to incur customer acquisition costs. You invest in growth and then you have some retention. And that's how you justify the initial investment you made. But when you're working with agents if you incur a cat so you give them some promotion like I'll give you a hundred bucks for free. And then I hope to retain you the agent might just switch immediately the switching cost goes way down because the switching cost is now reduced to what's the token spend required for this agent to switch from product A to product B which is likely a lot lower than the human labor required to switch from product A to product B. So my point is that a lot of companies rely on growth economics will also need to be revalued. So I think there's going to be some sort of repricing event. I don't know what that's going to look like and whether it's favorable to some companies or unfavorable to others. But generally I feel like the like lined approach of applying software pricing from five years ago to the companies of today is probably going to end with some people losing their shirts. Exciting times, fascinating. Thank you so much for coming on the pod. Neil, this was awesome. Cool. Thank you.