Fintwit读书会 2025年2月：高级投资组合管理：基本面投资者的量化指南

Byrne, how’s it going?

Hey, great to be here. I always wonder what the right reader-frequency statistic to track is, incidentally, because I feel like what you want is an audience that will not act—an audience that is too busy doing cool things to actually read 100% of what any one person writes. Within that audience, you want to get the highest reader percentage you possibly can. If somebody is reading 100%, either they’re not busy enough, or you’re not publishing enough, and possibly it’s both.

That’s a great point. If somebody’s reading 100%, either they’re not busy enough, or you’re not publishing enough, and possibly it’s both. I also listen to The Diff pretty religiously—not to sponsor a competitive podcast, but I like hearing your deeper thoughts on it. Any man who can bring down a multibillion-dollar domestic bank purely through the power of his newsletter has to be worth paying attention to.

The book we’re going to talk about is Advanced Portfolio Management. How do you say the author’s name?

I think he has alluded to being called Gappy a lot, so that tells me I should just call him Gappy.

Great, we’ll call him Gappy. Let’s start with your overall impressions. What did you think of the book?

I thought it was great. It was a really clean explanation for why the factor-neutral model works as well as it does in the contexts where it does. There’s a lot to unpack, but I thought it was a really good explanation of how stock picking can work in a context where we recognize that if you pick one stock—if you buy one stock—you do get some kind of excess return from the fact that you picked NVIDIA and someone else picked Apple, or vice versa. But a lot of what drives those stock prices is not specific to that company.

Sometimes people just get lucky. I forget who originally made the joke that the best financial decision you could ever make in your life was just to get a job in finance, probably in bonds, in the mid-1970s. Anyone who got their first job out of college working in something bond-facing or stock-facing between 1975 and 1985 was just set for life. Once that’s true, investors start to ask themselves, “Why are we paying someone for mostly being lucky?”

If someone has an exceptional career over multiple decades and earns 133% a year owning stocks, while the market did 10% a year over the same period, they’re getting paid on that 13.3% in many cases, but they actually deserve credit for about 3 percentage points of it. This helps explain some of the rise of indexing: If we’re paying fees to people who don’t necessarily beat the market, and often don’t beat the market, why don’t we just buy the market as a whole?

The other way to take it seriously is to say, “If you are skilled, you want to minimize luck because you want to get paid on skill.” If you can measure your skill well and deliver a return stream that is as pure skill as possible, then you actually deserve a much larger cut than before.

These trends are things you can’t avoid if you regularly read The Wall Street Journal, the Financial Times, and Bloomberg. You’ll see stories about larger funds, read about them being factor-neutral, and have some awareness of what that means. You’ll know that indexing is getting more popular. The book talks about a lot of trends that exist, but what it actually does is walk through the math of why you would do things that way and explain a lot of the idiosyncrasies of how these firms are structured.

There’s also a vibe that a lot of people have when they’ve been in a field for a while, done well in that field, and seen enough things to develop recurring jokes, observations, and other indications that they have a lot of experience and know what they’re doing. I liked that aspect of the book, too. When you’re reading it, you’re reading the output of thousands of conversations with risk-takers, risk managers, LPs, and people who allocate capital. You’re getting a lot of condensed wisdom from the book.

I was skeptical when you chose this. I thought, “What does that have to do with me?” But when you read it, a lot of it is applicable—at least, that was my understanding of the book. I think he knew that, because he talks about many different ways of saying, “Even if you’re not on the risk-management side at a pod shop, here’s why you need to be thinking about it, and here’s how you can improve it.”

I knew this book was going to hit with me because there was one specific line in the first chapter where he said, “If you collect all the trophies when you’re playing a video game, here are exactly the parts of the book you’re going to read.” I thought, “Oh, man, this guy knew me. He knew I was coming to read this book.”

My listenership and your readers are mostly individual investors in some way, shape, or form, or maybe people running concentrated fundamental funds. Which pieces did you think about implementing in your own investing after reading this book?

One thing this book really clarified for me is that, even if I’m not trying to neutralize my exposure to every factor, I should be thinking about what factors I have exposure to and what factors I will tend to have exposure to if I have a particular process.

The process isn’t, “I’m going to sort everything by momentum.” But if part of the process is, “I know that I have a tendency to double down on things when I’m obviously wrong,” then I need to account for that. Most of my big losses started out as small losses. I felt like I would be totally vindicated if I just doubled the size of the position after it was down 10%, and then I had a larger position that went down another 15%.

For me, I’ve dealt with that by trying to be more aggressive with stop-losses. But if you do that, what do you end up with? You end up with a portfolio that is always, by default, long momentum. It always has momentum exposure because you’re stopping out of everything else.

Can I pause you there? The stop-losses piece was one of the most interesting parts of the book. You said it sounds like you’re implementing stop-losses in some form. How are you implementing them?

I do the lazy thing where I just have a mental stop-loss. The rigorous way to do it would be to write out exactly what my thesis is, what would prove it, what would disprove it, and in what cases the absence of evidence would disprove it.

This gets to why stop-losses work and where they work. The more you are making a bet on very near-term changes in sentiment, the more what the stock does immediately afterward tells you whether or not you were right. If you are a deep-value investor who is somehow digging up the last annual report that a company issued 15 years ago, figuring out what real estate it owns, looking at the real-estate records, and finding out whether it sold the building, then you’re buying random pink-sheet companies at one-tenth of what they’re worth. There’s also no momentum.

There’s no real information from the stock price moving, other than that respect for MNPI is probably a lot looser in that corner of the market. If the stock doubles, it will probably keep going up, but other than that, there’s very little new information.

But if you have some view on NVIDIA and think this is the cadence of new model releases, this is how fast people will be building out data centers, and this is what investors are getting wrong about it, then you get information every day about whether investors are converging on your viewpoint or developing stronger conviction that something else is true.

If you mapped out all the right variables, did the pod-shop approach of figuring out everything that moves the stock, and then got a variant view on a handful of those things, you get information from the stock price. If the stock isn’t moving in the direction you expect, then either you don’t have a good model for how changes in fundamentals affect sentiment, or you don’t have a good read on those fundamentals.

I use a hybrid of your two approaches on the ones I struggle with the most. Something like Dish Network, which is now EchoStar—SATS is the ticker, if anybody wants it—I do not have a position in it, but that’s a play on Charlie Ergen figuring out a way to monetize the spectrum.

Whether he monetizes it by selling it for $100 billion and everybody’s happy, or sells it for $20 billion but divvies it out through the bondholders in a situation where the market is soft, it’s not like NVIDIA. With NVIDIA, a lot of it is people playing games around whether a new $500 billion data center is coming online. There are very short-term checks, whereas with something like this, I don’t know if it could go from $25 to $20 in a heartbeat.

Then I’m asking myself, “Am I stopping out, or am I stopped out if I’m applying about a 20% stop?” Should I be doubling down because there’s been no new news? But if you tell people that a 20% move wasn’t a big deal, I promise you that 20% is about where it really starts to hurt.

Those are the types of situations where I wonder how a fundamental investor can implement stop-losses when they might just be selling out of one of their top ideas because of random noise. I’m not going to say an earnings puke down 20% is always random noise—most earnings pukes down 20% are because there’s something going on—but as a fundamental investor, it feels like you’re always getting punched out. The counterargument is that I don’t know many stocks I’ve written down 50% that I’ve ever ended up making money on. It feels like there’s somewhere in between, but I struggle with that.

I think it gets down to the question of how much you are trying to understand a business and buy a piece of a business that you’d like to own for a very long or indefinite period. In that case, you probably want to bet more on mean reversion, especially if part of why you own the stock is that you think management will be diligent about buying back shares when the stock is cheap, and that when the stock is not cheap, there are opportunities to reinvest in the business. In a case like that, your expected return goes up when the stock price goes down.

But this is also a classic way for a lot of value investors to blow up. They keep telling themselves, “I thought this was a dollar bill trading for 7 cents, and now it’s a dollar bill trading for $0.50, so I need to take it from 10% of my portfolio to a quarter of my portfolio.” If you’re right, that’s great, but a lot of the time, when you start coming up with that kind of justification, you’re not thinking about what you could have done to avoid the loss or what the actual distribution of outcomes is.

I have a watch list that is very annoying because I put it together a couple of months ago. I had looked at a lot of high-quality companies where my thought was, “This is really interesting, and I should just buy it later. I should buy it when they have a bad quarter because I think they’ll probably recover from that.” Late last year, I started putting that list together, then I forgot about it for a couple of months. When I looked back at it, one company had gone up 50% and the other had gone up 20%. That was immensely frustrating.

My last big loss was a company where an event just resolved in an unpleasant way, and the stock dropped by half in a day. You may know this one.

Target Hospitality.

Oh, yeah. No position anymore, but I know that one extremely well.

The loss before that was a short position in a particularly scammy category of companies. I just wasn’t covering fast enough, and the stock went vertical. It went from $7 to $50 in about a week and a half.

That’s what short positions do, especially. That’s what really breaks me. If you hesitate a little bit when a long position moves 20% against you, at least the position has gotten smaller if you’re not adding on the way down. You have to be so glued to it with a short. That’s one of the tough things about shorts.

This is actually a case where I feel the pod shops have the right framework for handling a portfolio that has, in terms of number of positions, mostly short positions. A lot of what they’re doing day to day is figuring out the next thing to short. If they’re constantly adjusting their portfolio and trying to get closer to their target exposure to different factors, then they are pretty much automatically covering the short positions that are moving against them and pressing the short positions that are working. That is what you’re supposed to do.

It also depends on whether it’s a short position in the sense of, “The stock is at $20, I think demand is getting softer in this industry, and the stock will probably be at $15 in 6 months,” versus, “The stock is at $20, the CEO is committing fraud and should be in prison, and the only question is how much I pay in borrow costs before it goes to zero.”

There are companies like that where I look at them as entertainment because the borrow is too high. In those cases, it’s really annoying to recalibrate because you feel like you’re giving somebody else money when you cover part of the position because it has gone up and your conviction should be higher. But nobody gets credit for finding a stock at $10 and realizing it’s going to zero, but first it’s going to $200. You don’t get credit for being early to recognizing it as a short if it blows up beforehand.

My favorite example of all time is GameStop. Somebody was saying, “GameStop is at $50. Sell the $100 calls next week for $2,” or it might have been 6 weeks. Six weeks later, you’re closing the loop on this profitable trade, and GameStop has gone from $50 to $600. At peak, you were down, but you died on the way there.

That’s another part of the pod-shop model. They’re looking at what path gets you to that long-term return, and they care about that a lot. First, it makes more sense to compare strategies by looking at their volatility and at how correlated those volatilities are. But there is also a meaningful difference, especially when you lever up, between 15% plus or minus 5% and 15% plus or minus 40%.

If you look back at financial-media coverage of good investors in the 1990s or early 2000s, in many cases these people just really liked beta. They owned the more volatile slice of the market, did really well when the market did well, and did badly when the market did badly.

They had a nice narrative, which was that they generated good returns over time, and the investors who kept the faith with them and didn’t redeem when the market was down by a third and they were down by two-thirds did make money. I would suspect that if you ran a regression on what you got from investing with them versus using S&P futures to get 200% long exposure to the market, you’d find that they were about as good as being 2 times long the market, but a little more volatile.

The thing I thought was most interesting was this line, which he used in a few different ways: It’s not enough to have great ideas. He says, almost directly, “Having good ideas is useless without the knowledge of how to turn them into money.” He’s basically saying that it’s not enough to have ideas; you have to have portfolio and risk management on your side.

I thought that was interesting because Charlie Munger is famous for saying that it only takes one. You have one great idea, plow it all into that one great idea, and that’s all it takes. Both ideas have a lot of elements of correctness, but when you mash them together, they couldn’t be more divergent. This comes back a little bit to the difference between a pod shop and running a concentrated, 6-stock book.

I think it’s a realistic look at how the world works in general. You can look back at someone’s career and say, “If I had made the one big decision they made, I would be very successful.” If you had realized that electric cars were feasible in 2008, or that PCs were going to be a really big deal in 1975, or that Bitcoin was digital gold, you could say, “I knew.”

But you can go back and find that a lot of people had the same key insight. The people who executed well on those insights were often just making a lot of iterated, good decisions along the way.

There was a thought experiment going around Rationalist Twitter many years ago about having a time machine that could send a message to yourself 10 years in the past, but only a certain number of bits of information. What would you send to make as much money as possible?

I thought about that in light of the Winklevoss twins. You could imagine them getting a message from their future selves saying, “The 2 best conceivable investments you can make in the early 2000s are, first, buy Bitcoin,” and they got that one right. “Second, buy as much equity as you can in whatever social network Mark Zuckerberg is working on.” They nailed it: They were trying to maximize their equity position in the Mark Zuckerberg social network before anyone else realized that it was a way to make a fortune.

Yet they did not execute perfectly and ended up not making as much money as they probably could have. It’s obviously a very carefully chosen example, but I think it is revealing. There’s a narrative fallacy around investing where you look at someone’s one best idea, but you also have to ask how they were in a position to monetize that idea.

If you had the same idea at the same time, but your net worth was $1,000 in your checking account and you were otherwise broke, you could have multiplied that money by putting some of it into Netflix. But being in a position where you can actually have a material stake in a company means having made a series of pretty good, though not necessarily narrative-generating, decisions beforehand.

There’s natural resistance to the idea that coming up with good trading ideas is nice and important, but not the key thing. The key thing is coming up with a portfolio construction that gets you the most value out of those ideas.

People who get into finance, particularly the capital-markets side and certainly stock picking, don’t dream of coming up with the right way to diversify a portfolio. They don’t dream of coming up with the right cutoff where they figure out whether their 20th long idea will be a net contributor to expected return or to their Sharpe ratio. What people dream of is being the person who figured out that Netflix was a good buy in 2011, or who bought Google right at the IPO, or Visa or Mastercard at the IPO, or who figured out that GameStop was going to rocket higher.

Those individual, discrete decisions are what get people excited. But if you’re running a fund and selling a stream of returns, what you’re actually selling is the output of a process where there shouldn’t really be one big winning decision. Those don’t come along every year, so you can’t deliver consistent year-to-year returns with one big win.

Ackman has been in the market with his Howard Hughes bid recently. I’ve owned Howard Hughes in the past, but I don’t currently own it. Bill Ackman offered to take the company private. The stock was at $70 unaffected, and he offered almost $1 billion at $90, but then you give him a management contract and he charges 1.5% of the equity market cap.

As I was saying this, I was worried Byrne hadn’t noticed it, but I remembered that Byrne had one of the funniest lines about it. I’ll let you tell the line if you want. I’m shocked by it because, if you think about it, Ackman has made some killer decisions over the past 10 years. He held Chipotle all the way up, among several others, but the real reason his returns are even passable over the past 10 years is that he nailed an inflation trade in 2022 and made perhaps the best investment of all time with his COVID puts and COVID CDSs.

It’s interesting because that’s what he wants to monetize when he goes to Howard Hughes and says, “I’ll do the management contract.” He’ll have some really interesting trades, put the company into them, and make multiples off those trades to pay for everything. I’m just shocked by that because that’s what he’s trying to manage. A comet’s going to hit the Earth, and I’m going to be able to monetize that in some way, shape, or form.

In one sense, it makes sense to backtrack slightly and say that in some domains, you do want to make a handful of really good calls over your career. If you’re doing early-stage investing, it is pretty much, “Did you invest in Stripe? Did you invest in Databricks?” If you got one of those deals, you were pretty much set.

On the other hand, there’s a big difference between being able to put $5,000 into the seed round for a company and being able to put $500,000 into a very early round in the same company at a similar valuation. You get very different outcomes from that. I forget which Huffington Post cofounder I was reading about, but they had also made a very early investment in Uber and actually made more money from their $5,000 check into Uber than from cofounding and playing an important role in running The Huffington Post.

In cases like that, you have these nonlinear returns. In macro, there’s a combination of one-off opportunities. We didn’t have much data on what to do during a pandemic. You could buy anything with China exposure because it was all going to bounce back. That was not the COVID trade. It was a COVID trade at various times, but it was not the main one.

In cases like that, it does come down to making the right call in one special circumstance. But thinking about being in a position to make that call matters. If your destiny is to predict some out-of-left-field recession and make a ridiculous amount of money on CDSs and deeply out-of-the-money index puts, the biggest impact that decision will have on your net worth comes down to how much capital you have to deploy in that trade and how much capital you have access to.

Even if you are the kind of person who can swing for the fences, the way to maximize the value of that is to have a good track record with smaller-scale bets that may use a similar thought process. There’s also this paradox of alpha, which the book talks about a little bit. Some strategies used to be alpha—you used to have to pay someone to implement them—and now they’re wrapped in an ETF that costs tens of basis points.

Things like simply buying value stocks used to be fairly difficult. You had to get the Moody’s manual and page through it. Today, somebody can come to me and say, “This stock is really cheap; it trades for 8 times pre-tax earnings,” and I can say, “In the 1970s, that was a great analysis. Now there’s a computer that has done that.” Unless you have more than that, you’ve provided no edge. I’ve almost felt like the right way to do stock screens at this point is to screen for everything you don’t like and then look at a random selection of companies to figure out which of them could actually inflect from low margin to high margin or from no growth to growth.

I had a question and insight on that. He mentions the GICS standard—there are 4 levels and a bunch of different subsections beyond that. If you buy Apple, you might need to short some Microsoft to get some of that factor exposure out.

I’ve had this debate with people who look at a company and say, “60% of its revenue is from the lowest-multiple thing you can do, producing coal, and 40% is from AI generation.” One issue is that the company doesn’t get labeled with the AI GICS classification, and then all the pod-shop money will be able to rush in.

Do you think there’s anything to buying things that are classified incorrectly in order to generate alpha? Or are people deluding themselves into thinking they have some clever way to do this? The GICS classification isn’t unknown.

The synthesis is probably that you should analyze a lot of companies and have some process where either you have some source of randomness in what you look at, or you are constantly traversing the graph of companies.

Let’s say you spend a lot of time on AI and your big uncertain question is how fast it gets implemented on the enterprise side. You start looking at a bunch of enterprise AI users, and maybe you find that one of them is not just using ChatGPT in smarter ways than everybody else, but has also been doing a bunch of other smart things. Your AI thesis turns into, “I’m going to buy this well-managed industrial conglomerate over the other ones.”

You do want some of that randomness. I was being a little cute when I said you should screen for everything you hate and then find companies, but it is the case that everything you can screen for is something that someone is trying to price in. They might be mispricing it, but there can be cases where you screen for high margins and what you look for within that is a qualitative view that margins are not just high but will keep going up.

That gives you a very pod-shoppy kind of thesis. You have a fundamental view, it’s a variant view that you can underwrite through whatever data you’re able to gather, and you have a view for how that flows through into changes in the price.

Instead of saying, “I like the stock, and here’s the DCF,” you say, “People are valuing this on EBITDA. Here’s what they think EBITDA will be at the end of this year. Here’s why I think it will be higher, and here’s my event path for getting to a higher number. Once they hit that higher number, they will probably be at a higher multiple.” You win on both sides of that.

That kind of thought process—taking the things you screen for as a given and asking how they’re changing and whether you can have an edge in predicting those changes—is pretty valuable.

Going back to the point about one-off trades and the nature of alpha, there are a lot of trades that were repeatable but required a lot of manual work. They have since become more commoditized. Those used to be alpha; they’re now a different flavor of beta. As we go through the strategies you could turn into purely systematic strategies, what we’re left with is the stuff you can’t really describe as repeatable. You can’t really describe the big mortgage short or the Magar trade in mortgages as a repeatable strategy. All you can describe is that you want to be smart, pay attention to what’s going on in the world, find things that don’t make sense, find the best leveraged way to express that they don’t make sense, and either get the timing right or find a way to make the trade with little negative carry, ideally positive carry. Magar did that incredibly well: they figured out the stuff was unsustainable, but also figured out how to earn positive carry while betting on it. That meant they didn’t have to call the top in housing or figure out exactly how the bubble fell apart.

Over the past 8 weeks, I think you were one of the first people to really note this in the market. There was the DeepSeek news, and a lot of the AI stocks fell, but what fell more than Nvidia or the companies directly exposed to it were the power plays—the utilities that had nuclear plants and were going to have unlimited demand from data centers.

You noted that if you were a pod shop running a fund, there was a limit on how much Nvidia and chip exposure you could buy. One way to increase your AI exposure even more was to buy utilities. You weren’t getting penalized as much on the risk factor. Nvidia might be 5 units of risk, while a utility might be 1 unit of risk, so you might have been able to lever it up 5 times more.

In your experience, how gameable are these factor and risk models? When I talk to my pod-shop friends, they’ll say, “I like this company over that company,” and one of the reasons is that it gets them a backdoor play with more beta than they’re being charged for.

It’s good to be cautious about that because you read a book like this, everything makes perfect sense, and you think you have a model in your head for how these companies operate. But by the time the book was written, the specific way it gets made obsolete is that people are constantly trying to figure out how to game it.

It is fundamentally true that if you give someone a cut of P&L, and it’s not their money that’s invested, you have given them a call option on the performance of their portfolio. You can construct whatever constraints, incentives, and other things you want, but you’ve still given them a call option. They have an incentive to seek volatility.

You’re also hiring people who are really high performers. They basically haven’t had any big career mistakes, or if they have, they’re impressive and original mistakes. You’re selecting for people with very high egos, so you’re hiring people who will think, “These risk rules are meant for people who are dumber than me. If I find some edge case where I have not only a good trade but a trade that outsmarts the risk system, of course I’m going to take it.”

Then you hire all your risk people to stop them from doing that. From the perspective of the actual manager of that fund, someone doing a trade like that is basically stealing. They’re stealing office supplies. In this case, the office supplies are market beta, beta exposure, factor exposure, or whatever.

I don’t know for sure who would have been positioned that way, but if you’re looking at these stocks as pretty low-beta, low-volatility companies whose exposures you understand, they’re a quiet corner of the market. If you have a variant view on how much their profits will grow or how much demand will grow, you have a very clean trade.

But if you and your peers push these utilities up 20% or 30% in a couple of months, there isn’t a utilities-focused buyer waiting for a 5% or 10% pullback. There’s just someone mystified that you’re so wildly bullish on this. They may not pay attention until it’s back down to where it was 6 months earlier.

There was an air gap where there weren’t many people who were somewhat optimistic about the AI-utilities trade. To bet on the utilities trade, you were betting on the Situational Awareness paper’s model of the world: We’re going to have this massive deployment, the scaling laws will hold, and so much value will be created that U.S. electricity consumption will increase by a third over a fairly short period.

I completely agree with you. I wonder how much being a successful pod-shop manager over the past two years involved having a bullish view on AI and expressing it in every way possible. The easiest way is to buy NVIDIA, Microsoft, or whatever, but the system limits you. In the rest of your book, you find ways to get exposure to the AI play, even if it’s clunky. Then, when you’re right, you get rewarded. There’s hidden AI beta because AI is probably a factor now, but it wasn’t necessarily recognized as one when the exposure was put on.

The steelman argument against my view is that AI is a factor. People clearly weren’t hedging it, which I think is broadly true. But the population of people who weren’t hedging it might not be the pod shops. If there were investors who were the first to put a label on new factors and not just call them momentum, it was probably one or all of the top pod shops. Part of the job of a pod-shop manager is to identify emerging factors, bet on them when they’re underexploited, and capture some of the upside from people realizing that AI is a theme to allocate to, just like oil or midstream energy. If you’re early to that, you’re also in a good position in terms of information gathering. You’ve been following NVIDIA slightly longer than many peers and thinking of it as an AI play slightly longer than many others. When you go to a conference and hear really good questions you hadn’t thought of, that means other people are thinking past where you are. At that point, your alpha has completed most of its evolution to beta.

That is a really fascinating way to think about it. I’m not saying I’m the smartest person in the world, but there have been one or two companies where I’ve really known the thesis. Someone would say, “I just made this big position,” talking about things I’m an expert in. When all your conversations with peers are asking the same questions, that’s when it’s played out. When people are asking what to you are basic facts, that’s when you have the most potential edge. That’s a fascinating way of framing it. Let me ask one last question here. There’s one line that jumped out to me: “The ability to combine these alpha forecasts in nontrivial ways from a variety of sources and to process a large number of unstructured data is a competitive advantage of fundamental investing, and one that will not soon go away.”

That’s a really interesting line because, if I’m remembering correctly, this was written before the current AI period. We have AI today, and it is getting markedly better. You put it best: A good macro person reads an article today about the Japanese yen, matches it to an article they read a year ago about the Bank of Japan’s interest-rate policy, and realizes, “Oh, my God, the yen’s about to break out one way or the other.”

That is an advantage for fundamental human investors. How much do you think matching uncorrelated data will be taken over by AI versus remaining an advantage for fundamental investors?

I think we’ll develop a more elaborate taxonomy for what that looks like. There are already things I used to do manually that I can now automate. For example, parts of my idea-sourcing process used to involve getting a long list of things to read and skimming through them. Now I can feed a long list of things to read into one of the OpenAI APIs and get a summary that’s much quicker to go through.

For a fixed process where you know roughly the frequency of needles and the size of your haystack, you can automate it pretty straightforwardly. I don’t know exactly where the human-in-the-loop component will remain, but part of what will happen is that it will be as if you had an extra 10 hours a day or an extra 50 hours a day to read. You’ll have the knowledge base you would have if you had almost infinite time to read and were looking for particular things.

What gets difficult is getting the right level of serendipity. Whenever you automate a process, you’re implicitly saying that you know roughly what that process is. In some cases, you don’t know what the process is until you’ve done it manually.

This will be an interesting barrier, or an interesting trait to look for, among new investors who started doing fundamental analysis after LLMs became available. They’re used to the idea that some things are read by them, some things are read by Claude, and there’s a mix of different kinds of content.

It will be harder for them to do the boring work of reading a bunch of 10-Ks from different companies in an industry and trying to understand them. They know they can get a summary, but if you ask an LLM to summarize a 10-K, it will probably give you a high-level, bullet-point description of what the company does and may tell you a little bit about how it’s growing. You won’t know the weird distinctions between companies until you’ve read a 10-K and realized, “This is the first time I’ve ever heard a company mention this thing. I wonder if it’s unique to this company, or if nobody else talked about it.”

You still want a lot of tokens in your own personal context window, even though you have these other context windows. What it means is that a lot more people will be promoted to management. My first piece on how AI is going to change a lot of the way that we work was called “Working with a Co-pilot,” in April 20123. Your direct reports are all electronic, but they’re much less wise than you. They need a lot of specific instructions, but they’re also a lot smarter than you and have a higher energy level.

The company’s standards have kept going up since you joined, so everyone who reports to you is objectively better than you in every sense except that they have less experience and therefore worse judgment. Your job is to impart a lot of judgment but outsource as much of the actual cognition to them as possible because they’ll be better at it. You still need to know enough about how they’re thinking to spot flaws in it.

I would go back to the point about low-multiple stocks and how it used to be alpha to calculate the earnings and the price and realize that the P/E was only 8. That’s no longer a source of alpha in the way it once was. It certainly did not reduce the total amount of time people spend analyzing stocks; it just changed what they spend that time on.

The world is also going to get more complicated as AI tools make it faster to analyze the world as it is, but also make the world more complicated. There are a lot of things that will be harder to measure. Think of the network effects in a social network. That used to be a fairly straightforward concept, but it’s fuzzier now if people increasingly interact with LLMs and use LLMs to produce comments.

At one level, there’s more user-to-user interaction because the LLM will suggest something for you to say in response to your friend’s status update on Facebook. It will offer prefilled options, which I always wanted. If someone posted about a tragic life event, I’d want to say something comforting, then I’d see a comment saying, “So sorry for your loss,” and think, “That’s a really good one. I wish I had come up with it first.” Then I’d see, “You’re in our thoughts and prayers,” and think, “What a good thing to say. I wish I had said that.” Now an LLM has solved that for you.

But when you receive that, you recognize that Llama 3.7 feels really bad that your dog died, and maybe your friends do, too. You’re more conscious that you’re experiencing this through AI, with your friends giving their stamp of approval to comments written by something else. That’s a fuzzier concept of the network effect and the connectivity we get from social networks, which means we have to think about what the social-network model actually is in an AI context.

A lot of other industries will also have to rethink what they do, what the economic drivers are, and what they charge for. Even though AI speeds up that process, its existence also means there is more to learn and figure out. It’s trickier because the barrier between a deterministic computer process and a somewhat random, unpredictable human process is now a continuum.

As you say that, it brings me back to our conversation about GICS. As more investors are trained this way—and I think you’re right that investors just a few years younger than me put everything into ChatGPT first—I wonder if they demand a higher risk premium and the companies trade cheaper. Is there more alpha in finding something that doesn’t code properly?

I’m probably underestimating AI’s ability to rationalize. I know companies that have been accused of not screening on Bloomberg because their financials are formatted differently from every other company. When AI glances through the financials, it might not pick them up in the same way. I wonder if there will be a way to beat the bots.

The tough thing is that I don’t have a good model for what I’d be good at that an LLM is not better at. One way LLMs help me is with things like, “This company was cheap because it only produces financials and annual reports in Japanese, and it uses Japanese accounting.” Now you can save the PDF, put it into ChatGPT, and get a nice summary.

With sin stocks, the AIs may be less willing to talk about certain things, although ChatGPT will probably analyze a liquor company just fine. It might say, “Alcohol is dangerous, so I’m not going to give you a cocktail recipe.” For a porn company, maybe it wouldn’t do that, but I don’t think there are any pure-play porn companies left. Reddit is probably the closest thing to that, but that’s not the main driver of the business. Cannabis is another example where ChatGPT is probably fine talking about the business.

For everything else, you almost want to invert your instincts. Things that used to be inconvenient because they weren’t digitized or easily searchable can now be processed by AI faster than you can process them yourself.

Let’s say there’s an interesting conglomerate—one of those mini-Berkshires with a great capital allocator at the helm—but the CEO only does podcast interviews, and the annual letter doesn’t tell you much. The podcast interviews have a lot of depth and information about the thought process. You might think, “I’m not going to listen to 20 hours of podcast interviews to get 5 nuggets on capital allocation.” But if you can convert them to text, put the text into your LLM of choice, and say, “I’m an investor looking for information on this company’s capital-allocation approach. Please go through this transcript and tell me what they said about it,” you can probably get your answer.

Those things become efficiently priced very quickly. The closest you get is to ask yourself what a big, sophisticated fund with an enormous technology budget and many employees would build to make investment researchers more productive and allow them to ingest more data—increasing the number of tickers one analyst can cover.

Then you build a janky, hacky version of that and use it only for companies with a market cap of $300 million or less. You’re pretty confident that Point72 is probably not going to have one of its very expensive analysts looking at a nearly bankrupt clothing retailer, a Bulgarian energy company, or something similar.

I’m laughing because I know a few funds that have argued, “What’s our edge?” They’ll say, “We apply credit-card data to companies from $500 million to $2 billion in market cap, and that’s our edge. We’re more sophisticated, and the big guys won’t play in these small ponds.”

But then you ask, “Why won’t the big guys play there? Why is there enough liquidity for you and not for them?” That’s the pitch.

Byrne, let’s wrap it up there. Our internet connection keeps getting a little sticky, so I don’t want us to miss any big points, and we’re almost at an hour anyway. This was awesome, and you’ve given me a lot to think about with this book.

I’ll say again that there were so many pieces for somebody who runs an ideas podcast. Hearing the line that it’s not sufficient to have great ideas hit me a little bit in the chest. At the same time, it was a really thoughtful point.

I was skeptical as a fundamental investor coming into it, but I was really pleased with the book. Byrne Hobart, the person whose work I read 85% of the time—he tells me maybe I should get a little busier and read a few less things. This has been awesome.

If you’ve got suggestions for Book 3, we’re going to do another one in March. Byrne and I will have to read it and discuss it next week.

All righty. Talk soon.