投资中的 AI：与 Daloopa 创始人 Thomas Li 对谈

All right. Hello. Uh, today I'm posting a video of a webinar that I did with the loopa uh, talking about AI in investing and its implication for fundamental investors. Uh, we posted this webinar last week. It got a lot of it got a lot of good feedback. So, we figured, hey, we've already got the video file. Might as well post it on the podcast channel so anyone who's interested in, you know, my work or just thinking about AI and investing can go ahead and listen to it and get some ideas. I thought it was just a really informative conversation and Thomas Lupio sees how companies and investors across the board from very smallcale hedge funds to you know super large investment banks prop trading desks hedge funds that are managing billions are using AI. So I thought it was a really informative conversation and I'm happy to share it on the channel.

Hello, everyone. Welcome to the webinar. I'm Andrew Walker, the host of the Yet Another Value podcast. With me today, I'm excited to be talking with Thomas Li, the CEO and co-founder of Daloopa.

Thomas, I'll give an overview, and then I'd love to dive into it with you. One of the most frequent questions I have when I'm talking to my friends and thinking about investing is: AI is everywhere now. It's getting used for everything. I think ChatGPT recently surpassed Google as the most-used search function.

In terms of finance, I'm always wondering: How are my peers and competitors using AI to improve their work methods? If you're not using AI as a financial analyst, I personally think you're going to get left behind. But there are so many different ways to use it, and I'm always worried, as a one-man shop or a small shop, that somebody has figured out a way to use it 10× better, 100× better, or 1,000× better.

I really wanted to talk to you because I remember when we met back in 2020 over Zoom, in the dark days of August, you were talking about using AI and LLMs inside the Daloopa products. You have a very wide view and range of a lot of different institutions through your perch, so I'm really curious about how you're seeing AI get used by the best.

That's a broad overview. That's why I'm excited to talk to you. I'll flip it over to you for high-level thoughts on what you're seeing in terms of AI's use in finance, and we can dive in from there.

Hey, Andrew. Thanks for having me on. There are a couple of questions there, and I think the first level set I'll provide is: How do you think about what AI is built for and what it's capable of doing?

The foundational models keep getting better, and they keep doing something different. But the reality is, at the end of the day, what AI really is doing is a prediction model of what the next item is. Over time, we've gone from the next item being the next word, to the next sentence, to the next pixel, to the next set of pixels, to the next paragraph. It's able to generate, check its own work, regenerate, check its own work, and generate again, so it gets better and better.

At the very core of it, what hasn't changed in the AI world is: Based on everything I know, what is the next object? How do I generate the next object? The key to all of this, where it becomes really smart and why it's so much better than search, is the concept of generation.

As humans, what we're typically doing when we have conversations or thought processes is generating stuff. I'm creating new sentences as we speak, and it's very humanlike. The problem is that not everything we do is generation.

For instance, if you're a financial analyst and you're sitting there trying to model, a lot of what you're doing isn't generating stuff. It's actual processing. It's not processing how to write an email or how to pitch a story to your PM. It's actually processing to understand something better.

It's going through numbers, extracting data, thinking through inconsistencies between how costs are growing relative to how revenue is going, thinking through how CAC is changing for a software business, and thinking through how occupancy rates are changing for a hotel company. You might also be thinking through how all of those things come together in one big picture. That's really what an analyst's work is about.

That's not generation of concepts or generation of objects, and that's fine. Our brains have different compartments for different things. When I think about what AI is really good at, it's really good at the generation piece. There are a lot of applications where you want to use that, but trying to use it for everything is probably not a good idea.

I think every time—and this is a common human fallacy—when you're a hammer, everything looks like a nail. When you're a really, really shiny hammer, everything definitely looks like a nail. We are in a little bit of that, but very quickly you start to realize that if you try to use ChatGPT to do a lot of tasks, it just fails spectacularly in a way that it succeeds spectacularly at other tasks.

One common example I always tell people is something I absolutely hate doing: trying to plan a trip. Because of the volume of literal words that I have to go through—travel reports and reviews, Reddit, Google reviews, and whatever else—that takes a lot of time to process and plan. ChatGPT can solve that for me by giving me a one-pager: Here's what you should do if you want to go to Croatia. That's incredible.

But if I'm saying, "Hey, I have a financial model I want to build, and I want to capture all the intricacies in a financial model," all of a sudden that becomes really, really hard for ChatGPT to do. It fails. It starts generating things where they seem correct, but I don't need them to seem correct. This is a public company with public disclosures. I just want correct.

When I think about how to use AI, it really comes down to what AI is built for, what the philosophies of AI are, and applying those philosophies. One really smart use case I've seen a lot is when you have a bunch of internal notes that you've written about a business—your internal understanding of a business—and you say, "I want to correlate my internal understanding of the business with every single earnings call that has taken place in the last 4 quarters since I started covering this company. Are there any discrepancies?" AI is phenomenal at that.

You gave a really interesting use case there. I'm an analyst, and I've got a company I've been following. I might have pages and pages of notes on it. You're saying one of the really effective use cases is to upload your notes into ChatGPT and tell it, "These are my notes on the business. Tell me if there's anything inconsistent with what you're seeing in the way the company has been talking about itself on the earnings calls." Is that one of the use cases you're seeing that's interesting?

That's a pretty common use case. Most buy-side firms don't do that because they're not allowed to upload anything to ChatGPT.

I haven't thought about just writing my own thesis—which I do all the time—writing my own thesis on a company, uploading it to ChatGPT, and then saying, "Tell me if this thesis is disproven by something that's happened in the past 6 months." Is that what you're thinking?

Generally, you want to be more specific than that. You would say, "Here is the latest earnings transcript. Here is the latest conference transcript. They were presenting at the Morgan Stanley TMT Conference, for instance. Here's the transcript for that. Here are my notes prior to these 2 events happening. Where are the inconsistencies?"

It's a really smart blacklining function, if you think about it. That works.

Let me ask one more question, because I want people to have a takeaway about how they can start using AI and improve their use of AI through this process. If you're uploading your notes, as you said, a lot of what you're going to get out is what you put in. I've joked that with AI, it's garbage in, garbage out. How would people structure their notes in a way that would make this use case—which seems really interesting to me—as efficient as possible?

The good thing about how these language models are set up is that you don't actually have to structure your notes in any way, shape, or form. When it was ChatGPT-3, there was some structuring that you should do, but with the latest iterations of the models, as we've moved away from just regular transformer architecture, the structuring becomes less important.

I won't bore you with the technical details, but as we've moved away from transformers, what that means is that we're able to generate multiple items instead of just one. We can check the output ourselves and then regenerate the output. That looping process allows you to loop over 2 words, 50 words, 15 words, or whatever. How you structure your notes doesn't matter to the AI anymore, so long as they're words.

Where it doesn't work, unfortunately, is with numbers. It just doesn't work because the logic for numbers isn't quite there yet—or isn't there at all. But for notes, it works incredibly well.

Let me stick with this theme. One of the cool things about Daloopa is that you guys work with a bunch of different clients, right? You have big investment banks, big hedge funds, and small hedge funds—you really run the gamut.

You mentioned the reason I asked this: when we were talking, you said, “Hey, a lot of buy-side firms don’t do this practice that I think is pretty useful” because of internal restrictions on uploading notes. I wouldn’t be surprised if analysts are writing their own notes and uploading them on the side, but one thing I’d love to start with is this: how do you see—let’s keep it to investors—big pod shops that probably have a big research budget and are trying to be at the bleeding edge versus, let’s say, more traditional, sleepy long-onlys? How do you see differing use cases with AI, and how they’re interacting with AI, between those 2 broad buckets of customers?

Yeah. Actually, I wouldn’t categorize it that way, because what we’ve seen is that the big pod shops and the big long-onlys have both really just put their foot down and invested in AI. I think they’ve seen the benefits. To your point, it’s like Google all over again. If you don’t use Google as a research analyst, you’re just not doing your job, right? It’s like that all over again.

So I think there is no hesitancy between the pod shops and the long-onlys in saying, “We’ve got to do something here. We’ve got to figure out how to create applications and how it works.”

But you’re absolutely right on the research-budget front. There are firms with huge research budgets, and there are firms with much smaller research budgets. Generally, big-AUM firms—it doesn’t matter if they’re long-only or long-short—have much bigger research budgets. The big discrepancy that we’re seeing is whether or not they build internal tools.

The thing about AI that is very different from a search engine is you will never build your own search engine, because the foundational algorithm of a search engine was always locked within the confines of Google and Microsoft. To this day, we don’t have access to Google’s search algorithm, but we have access to a ton of foundational models.

We can literally pick foundational models today and 3 weeks later find something that’s 4 times as good and half as expensive, and you can switch on a dime, right? It’s almost like if Google just released its algorithm and 20 other companies were releasing their search algorithms, too.

What that implies is we can all build search in the way we want to build search. We can build extremely fine-tuned search, internal search, and external search. We can build all sorts of search, right? So that’s what’s going on with AI.

The cost of building an internal AI capability now is super low because of what people like to call ChatGPT wrappers, right? A ChatGPT wrapper is really just software that you build on top of a foundational model that OpenAI provides, or Anthropic provides, or one of these companies provides. So what’s very obvious now is that there are a whole suite of buy-side firms that are willing and able to build these internal tools.

So the question is, why? Why would you build these tools? Why would you not use ChatGPT? I think a year ago, the answer to the question was compliance. When people are hesitant about a lot of software, the first answer to why you don’t do anything is always compliance.

But over time, as people understand this better, real economic decisions start getting made. What we are seeing today—the real economic decision—is what we call the context problem.

If you wanted to do that note-comparison example that we talked about, where you compare what you’ve written with what the company has disclosed, you need access to 2 different sources of data. You need access to the company’s transcripts, which are not actually available on the public internet.

So if you try to do this with ChatGPT, it actually doesn’t work, right? And you need access to your own notes, which also do not exist on the internet. So if you try to do this with ChatGPT, it doesn’t work.

Obviously, if you go into ChatGPT today, even if you upload your own notes, it doesn’t work because it doesn’t have access to Morgan Stanley TMT, and it doesn’t have access to the company’s earnings-call transcript. Amazingly, it doesn’t even have access to the 10-Qs and 10-Ks, right? It has access to blogs and posts written about public companies, but not the actual filings of public companies. It doesn’t pull them off EDGAR or anything.

No, it does not. Hmm, interesting. I didn’t realize that, because I’ve definitely searched, and I feel like I’ve gotten links from EDGAR before, but I didn’t realize it wouldn’t pull from there.

You would have gotten links from a blog that links to EDGAR, but it doesn’t actually go into EDGAR.

Yeah. So why is that?

AI is very good at processing information, but the act of obtaining a document is still an old-school technique, right? AI doesn’t solve the problem of how you go get a document. How do you go get a document from the SEC? How do you go download an investor presentation from a company’s website? All of these are difficult problems, or just very manual, difficult problems.

There are systematic ways to get it, but it’s not a large-language-model problem to solve, right? If you went to ChatGPT and said, “Hey, can you pull me NVIDIA’s latest investor-relations deck?” and you went on Google and asked a question, you would get it immediately because it gets you to NVIDIA’s IR website, and then there is a PDF button, and you hit that and download it. If you try it on ChatGPT, you just won’t go anywhere.

So that’s another problem with AI: the document-obtaining piece is very difficult, but Google has effectively solved the problem super, super well, right?

The real problem is that you don’t have access to all these different buckets of data. You want your foundational model to sit on top of all of these different buckets of data so that if your analyst asks that question—“Here are my notes. What are the inconsistencies?”—not only is it able to compare them with Morgan Stanley TMT and the latest earnings release, it should also theoretically compare the data between your model and what the company has reported.

It theoretically should be able to do it for consensus. It should do what an analyst is really trying to accomplish, which is: How has my company changed? That’s the real question.

What we are seeing the really big shops doing is building this.

You’re saying the most difficult part of building this historically has always been the foundational model, the logic. But now someone has invented a logic, so all I need to do is feed it the information.

Feeding it the information is a harder problem than you think, but it’s solvable because you can leverage the fact that your analysts generate the notes. You can leverage the fact that you can purchase transcripts from data vendors. You can leverage the fact that there are guys like Daloopa that have done the work of extracting all the data with AI into a database.

So you can say, “Hey, I would like to compare how my estimates are relative to company historicals. Am I getting more accurate over time, or am I getting less accurate over time?” Right? Now this exercise is completely doable, assuming that you have access to a foundational model, your analysts’ actual models, and the fundamental data from Daloopa.

In 1 sentence, you can answer the question where historically that work is a research associate’s 2 weeks’ worth of just manual grinding through in Excel.

How are you seeing this? It strikes me that if I went into a fund right now—let’s use age gaps, right?—there might be a 50-year-old portfolio manager who’s been doing this for 25 years, a 35-year-old portfolio manager or very senior analyst who’s been doing this for 10 years, and then a 25-year-old analyst or post-MBA professional who’s very early in their career.

Each of them—I could imagine the 35-year-old kind of falls into that bucket: we came up and AI came in once we’d already gotten our process going. The 50-year-old saw the internet come when they’d already started and gotten their process going, so maybe they’re less adaptable, or maybe they’re more adaptable because they’re just willing to throw it out there.

And then the 25-year-old is just like, AI came while they were in school, completely native. How are you seeing those 3 buckets? Obviously, all of them have different job responsibilities. How are you seeing those 3 buckets differ in their use of AI?

Yeah, it’s a great question. I think it comes down to how strategic the person is. Over time, you notice that the more senior a person gets, because of job requirements and demands and experience, you tend to get more strategic in how you think about your business. Versus at a very junior, just-graduated-from-college level, you’re just trying to get the work done, right? You’re just working on activities versus working on long-term strategic visions.

What we’ve noticed is the people with the long-term strategic visions are basically saying, “This is a monumental shift in how things work. We’ve got to adapt to the shift, because you don’t want to be the 1 bank who didn’t move to Microsoft Excel when Excel happened. You don’t want to be the 1 research associate who said no to Google when Google was happening. You certainly don’t want to be the person stuck on BlackBerry when everybody was moving to iPhones, right?”

There are these monumental shifts, and I don’t think the age gap matters. I think, actually, the senior people we talk to fully recognize the power.

In fact, the more senior the person, we've noticed, the more likely they are to want to be AI adopters. The more junior the person, the more likely they are to be skeptical. One of my friends pointed this out. He was like, “Look, if you're a senior person—if you're the PM of the PMs—and you're just having everyone email you the research, and you give them 2 questions back and then say yes or no, you might be more likely to be an AI adopter because you're just like, ‘Oh, I'm just replacing calling the 25-year-old on the phone with putting it into ChatGPT.’ It's the exact same thing: I ask a question, then give it a yes or no.”

Whereas if you're that 25-year-old, you're used to, “Hey, I read every 10-K all the way through. That's what my boss taught me: read it all the way through, because I need to have all these special insights.” You might be a little more skeptical or a little more scared to push off those insights.

That brings me to another one. With investment banks especially, I remember when AI first started coming out, there were a lot of people saying, “Oh, AI—investment bankers, famously, when you're a junior, you're basically just making pitch decks nonstop, right? We don't need juniors anymore.” The counter to that was, “Hey, maybe you need 1,000 more juniors. Maybe you need them doing a lot more.”

Which direction are you seeing, especially at the junior level, that AI is pushing headcount? Is it pushing toward less because AI can take over so many of the jobs, or is it pushing toward more? If I switch from investment banking to research, I'd say, “Hey, maybe you need more people calling up franchises to say, ‘How are your sales this quarter?’” and putting that in to give AI more bespoke information. Which direction do you think it's going?

I think that's a great question. The reality is, that's not how we see our customers think. I would draw a parallel to investment banking before Microsoft Excel. People forget that Microsoft Excel was created in the generation of what we would know today as the MD generation. So when the MDs of today, or the SMDs of today, were analysts, they were in the migration phase from Lotus Notes to Microsoft Excel.

If you've ever used Lotus Notes, it is really not meant to do what spreadsheets are supposed to do today, what we think of as financial models. To your question, what we've noticed is the amount of hours worked doesn't really change. When Lotus 1-2-3 became Microsoft Excel, the argument was, “Now we can build models way more efficiently. It's so much easier. We can build much more detailed models. We don't have to work 100 hours a week anymore.” Guess what? People still work 100 hours a week.

When PowerPoint became way better, you had hotkeys to align logos—all these things became easier. Should you be working less in theory and be more efficient? Yes. But the reality is, in an apprenticeship model, which Wall Street is, that's just not how it works.

When firms think about their revenue structure, the question is never, “Hey, if I could replace this piece of cost, can I save cost?” Firms fundamentally are growth engines. We're always thinking about how do we expand business, how do we become more competitive, how do we become first to market, and how do we scale our firm.

So when you find a new capability that has tremendous cost savings, as it does today, the question is, “Okay, what do I do with the cost that's saved? What do I do with the associate's time that isn't spent updating models, summarizing things, or transcribing earnings calls? What do I do with the time? Can they be more productive?” It's rarely, “Okay, now I need 4 fewer associates. Let's shrink the firm down, let's keep our revenue steady-state, and let every other bank outgrow us.”

One of my biggest worries with AI, or any product that automates the process, is this: I think a lot of people have taught me, “Hey, if you're covering 35 companies and you need earnings reports and you just need a quick 3-statement model, great—go get somebody else to build that model, and then you can study it and think about it. But if you're saying, ‘Hey, this is the company I want to bet on,’ you need to go build that model by hand, because building it by hand, and the act of actually thinking it through, really helps you think through your assumptions, understand it, and get a better feel for it.”

One of the worries I have with AI is—you mentioned it earlier, where you had an analyst who was just tossing everything into it—one of the worries I have is that you do that and AI... I guess there's 2 problems. There's the echo problem, which I'll come back to in a second. But the second is, you're just outsourcing all your thinking, and then when you come to a real edge case, you kind of haven't thought it all through. You think you understand it, but you understand it only at a very high level.

There's the famous example where the guy goes and gives 100 speeches, and one day he has his bodyguard go give the speech for him because his bodyguard's just got it memorized but can't understand questions. How are you seeing people—especially more junior people—avoid that problem, where they've outsourced so much of their thinking? Yes, they're getting a lot done, their job's going, but they're not actually learning and maybe developing?

I think we don't actually see a lot of people outsourcing their thinking process. We do see a lot of people outsourcing to AI the mundane parts of the job, and finance is a really mundane job. Anyone who listens to this and works in finance probably won't disagree with that. There are just a lot of hours upon hours where you spend basically spinning your wheels. It's necessary, but it is very boring, and I think a lot of that work is getting farmed out to agents today, or should be farmed out to agents for those who aren't doing it.

Do I see people trying to farm out the intellectual piece? Yes, but it's not very common, because finance is also the industry of corner cases. Rarely do you look at something and say, “Oh, yeah, the same thing has really happened before.” History kind of rhymes, but not all that much. There's a lot of human assumptions and human judgment that need to happen.

Frankly, that's the fun part of the job, right? That's why people want to be in finance. That's the piece of the job that rarely do we see people trying to automate away. Where we see people really interested in automating away are the things that get in the way of that.

The loop exists because we update people's models, because we help generate automated comp sheets and industry models and whatnot. But there are also cases where you can say, “Your average earnings call is about 15 pages long. It takes a long time to read that.” So if you're trying to say, “I don't cover this industry, but I want to understand how tariffs are affecting this entire industry because this industry feeds into what I cover,” do I really want to spend 16 hours reading through every transcript of every company? Probably not, because I don't get the key points anyway.

But I can very quickly run a summary system where I can get the highlights, Q&A through it for 30 minutes, and just get all the information that I need. So that's saving on process, but not saving on the intellectual horsepower.

That brings me nicely to my next question. My biggest worry is that you use AI to get to a yes on something, and then you've got the echo problem. I've had this before, where I got into it with ChatGPT and asked it a bunch of questions. Then I said, “Hey, what was the earnings number in this quarter?” and it gave me a number. I was using that to build a lot of stuff.

Then I went back and checked, and it said, “Hey, earnings for 2023 were $500 million.” You go back to 2023 and you're like, “Hey, the revenue was $200 million in 2023. I'm pretty sure they couldn't have earned $500 million.” It was like $25 million or something.

So when you talk to clients and they're using AI, how worried are they about the echo problem, and what are they doing to ensure that? You could imagine a deep-research project where you and a team are spending a week on it, and then you come out at the end and there was an echo problem at the beginning and all of your premises were completely mistaken. How are you seeing them think about the echo problem, worry about the echo problem, and avoid the echo problem?

I think the reality is, we don't actually see a lot of people worry about the echo problem.

All right, Thomas, let me ask you. I asked you how your pod shops versus long-onlys differed, and you actually said they're kind of using it the same. Let me ask you this in a different version: your best clients—the ones you think are incorporating AI the best—what are 1 or 2 things they're doing that your clients who are spending a lot of money, but who you don't think are incorporating it as well, aren't doing? What are they using AI for that your average clients, in terms of AI, aren't?

Yeah, great question. I think it's what we call a context problem. There are people who realize that the difference between a great algorithm and the second-best algorithm is really not that big. The nuances are really, really small. But it's how you drive context into the algorithm that makes all the difference.

Once you realize that what makes an output really good is giving the foundational model access to a ton of data and creating guardrails around how to display the output, that's really what makes the output powerful, versus picking the best algorithm, the cheapest algorithm, or spending time on the foundational-model piece.

This is not just for hedge funds. You see this across the entire stack. Hedge funds and VC funds invest in AI and whatnot. The foundational models are important, really, but the differences are small. Where the money is made, where the rubber meets the road, where the applications are created, is in how we assemble the most amount of context around the model I pick.

Because if you think about the Google analogy I used before, if this were closed-source—if all the foundational models were closed-source—then yes, we should end up with a Google where the guy with the best algorithm just takes the entire market. But we don't have that world. We have an open-source world where the foundational models effectively are—I wouldn't say free to use, but cheap to use—and anybody can use them to build any product. Then it comes down to the product.

This is like the internet or mobile data: mobile data is cheap for anybody to use to build a business on top of. Then you get evaluated on the merits of your business and the creativity with which you are leveraging the power of mobile internet. That's where the world changes.

The context problem is: how do you build a product that solves the problem for your customers using the most amount of data that you can assemble, as opposed to thinking about the actual AI problem?

You should be thinking about a product problem, right? It’s so easy to get stuck in, “Oh my God, AI is new; let’s use AI,” and forget the fact that, at the end of the day, people build businesses because they want to solve problems. If you can solve a problem, it doesn’t matter if you’re using a hammer, a nail, or a saw, so long as you’re solving the problem.

What is one thing that someone on the outside of one of these big shops with a big research budget—I mean, me—that I would be surprised my competitors at big pod shops are using AI for nonstop? What is a use case that would be surprising in your mind?

That’s a good question. I don’t know what you see, so it’s hard for me to figure out what is surprising.

Just an average person: If somebody’s listening to this and they’re a part-time investor—they’ve got a day job and they like investing—I do hour-long deep dives on podcasts, and they’re really into that. Their mind would be blown if they heard, “Oh, professional investors just do all of that with AI now.”

I think the most surprising thing is how little AI adoption there is. That’s probably the biggest surprise.

Great. Let me ask my next question, then. What is one thing where you think there has been hesitancy or slow uptake on AI among investment firms—again, there’s a huge amount of research—that you think would materially improve people’s jobs if they spent a little more time shifting that work to AI?

I think if we can get around the hurdle that, if you kept querying an AI agent, the AI agent can figure out what you’re trying to do, right? The compliance hurdle is a very real problem, because if you had an enterprise account with ChatGPT and someone on the backend asked ChatGPT, based on all the queries that you’ve seen, what do you think the next big investment is, ChatGPT probably has a really, really good guess.

So you have that problem, right? It’s a very real problem, and if you’re the chief compliance officer, you should be freaked out, because that means someone could theoretically be front-running all your trades. That’s a big part of why adoption is low—not because there is no desire to adopt, but because there’s hesitancy to adopt a model that’s totally running wild. People want their own internal model.

That being said, if you look at how we use Google today, we don’t seem to really care, right? If Google really wanted to figure it out, they could. But the way most firms have gotten around that is essentially by partnering up with a vendor and having the vendor be so compliant and so trustworthy that, even though they could figure it out, you trust that they wouldn’t.

In my mind, I’m thinking of a company like Bloomberg. If Bloomberg really wanted to figure out what trades we’re going to put on, they would know, because half these trades happen on Bloomberg anyway, but you trust that they won’t.

Let me ask you this: If you and I were talking 60 years ago—maybe this is more like 100 years ago—security analysis was not very sophisticated. You could outperform in the Ben Graham era by going and saying, “Isn’t that interesting? That company has $100 in cash on its balance sheet, trades at $450, generates good earnings, and pays out a dividend yield of 12%.” You could outperform by doing that.

60 years ago, there was no Reg FD; insider trading was allowed a lot. You look at the history of how Buffett took over Berkshire. The CEO basically promised him that he was going to tender at one price, then did it a quarter cheaper. Buffett was so incensed that the CEO would lie to him about a tender of which he had advance knowledge that he took over Berkshire.

50 years ago, you could still do very well by looking at a company and saying, “Oh, this trades at 8 times price-to-earnings. Its peers trade at 15 times price-to-earnings. I buy; I short.” A lot of that has been done away with—the insider trading by Reg FD and stuff—but a lot of the other stuff has been done away with by quantitative models.

One thing I tell investors, especially college students who come to me and pitch a stock, is that they’ll say, “It trades at 10 times price-to-earnings,” and I’ll say, “That’s great, but that’s not a thesis. You gave me something that a computer can spit out. The quant models are all over that.”

I worry that, when it comes to a lot of the edge cases and thinking things through, AI is increasingly better at that. We talked earlier about AI displacing jobs. I’m not talking about that anymore. I’m more worried about how you generate an edge as these AIs get better and better, can replicate things, can think faster, and can pull in information from different sources.

A biotech company announces results. It takes the best scientists I know several hours to look through the results and really analyze them. AI could do it in a quarter of a second and be trading that stock. Is there room for humans in 3, 5, or 7 years? In public markets, yes, there’s room; in private markets and structuring—everything there, handshake deals and all that type of stuff—but in public markets, is there going to be room for humans?

Yeah, absolutely. There will be a ton. When you think about the hardest funds to get into, they are very often the fundamental funds. Think about the funds that, as an LP, you absolutely want to get into because the Sharpe ratios are just off the charts. They typically are the super-big long funds and the super-big multi-managers.

Why are the multi-managers able to consistently outperform? If you look at the returns of some of these multi-managers, they’re so uncorrelated: Markets are good, they’re up 15%; markets are bad, they’re up 20%. It’s like clockwork, and it’s every LP’s dream because it’s the true promise of uncorrelated, hedged returns. Everybody calls themselves a hedge fund; not everybody actually provides hedged, uncorrelated returns.

Why are they able to do that? I think, at the end of the day, it’s really sitting down and asking yourself, “What are the true sources of alpha out there?” A true source of alpha needs to be something that exists systematically, is true, and is something that you can protect for a reasonable period of time—not forever, but a reasonable period of time.

Buffett’s famous source of alpha is basically holding horizon. He’s willing to buy a company and let the company compound, and because he has a horizon so much longer than everybody else, he’s able to just hold. I would argue that, if you’re a retail investor, you have that same source of alpha, too—one that most professional investors don’t ever have, because your marks happen daily, monthly, weekly, or whatever.

But a professional investor has sources of alpha that retail investors can never get. One of the single largest sources of alpha is risk modeling. You’re able to factor out all sorts of risks through risk models. Let’s say you run a PCA and can figure out the few main factors of risk, hedge them out, and what you’re left with is alpha.

If you’re able to say, “Okay, if what I’m left with is alpha, then how do I increase the number of trades? How do I increase trade velocity?” and just play a numbers game, which for the most part is what the multi-managers are doing, they’re saying, “We will do a lot of trades. We’re not here for home runs. We’re not here for 10-, 20-, or 30-year compounding returns. What we are here for is uncorrelated, hedged residual alpha that I can leverage on a cheap basis and provide uncorrelated, consistent returns for LPs.”

Can an AI actually create those little sources of hedged alpha?

Not really, because that’s just not what the foundational models are trained to be able to do. A lot of those sources of alpha are not obvious. It’s not as if you can just crack into someone’s warehouse and see inventory flowing out the windows.

Historically, a lot of those sources of alpha in hedge fund shops have been channel checks, lots of expert calls—which are proprietary—sourcing expertise, figuring out whether you trust the expert, asking the expert the right questions, doing expert checks, and going into the field.

There’s satellite imagery of a parking garage or parking lot, but you can also go into the store and see, “Hey, are these stores clean or dirty?” Whatever it is.

How much do you think those types of proprietary data—people getting out into the field, gathering things, seeing things that maybe AI can’t—will increasingly become the job, especially of the analyst, versus, if you watch Billions, most analysts are at their computers from 8:00 a.m. to 7:00 p.m., clicking away, building models, or something like that?

Do you think it shifts more toward qualitative people skills as analysts try to generate proprietary data to put into the AI, or am I missing something? Am I imagining something or creating a story?

I think alternative data is not a source of alpha. I think there was a point in time when it was—2017, 2018, 2019—it was absolutely a source of alpha. Today, I don’t think it’s really a source of alpha.

Are you saying alternative data that you pay for? Because I think I was more talking about alternative data that you generate on your own, right? That’s more proprietary than alternative data.

Yeah, even the proprietary stuff, like channel checks—all of those. I think because it’s done in such high volumes...

So yes, absolutely. When it was first being done, satellite data and credit card data were an absolutely huge source of alpha. Today, it's just what everybody does. It's like reading a 10-K; no longer a source of alpha. Warren Buffett would famously say, back in the day, that if you read a 10-K, that is a source of alpha, right? Just knowing what the company does is better than the average.

Sometimes I wonder if it's a source of anti-alpha, where you should just close your eyes and not look at the 10-K. There's a lot of information in there that you don't want to know, and it would be better to just not know it.

But what I think is true today, and what the pod manager is so good at adapting to, is this fundamental idea of what makes a stock move in the short term. I think what makes a stock move in the long term is very well studied. Most people appreciate the fundamentals, multiples, interest-rate environment, and whatnot. But what makes a stock move in the short term is actually much simpler than that: it's buyers and sellers. If there are a lot of buyers and not a lot of sellers, the price of a stock goes up in the short term.

So if you're an analyst, really what you're trying to figure out is what the next guy is trying to do, right? An AI can help you get all the knowledge about, hey, do you think this stock is long fundamentally? What are they saying? Is there a discrepancy? But ultimately, what you need to do as an analyst is basically play the game.

You're saying, based on the information I have, who do I think has access to the same amount of information that I do? Who do I think is on a 2-day delay from the information that I have? Who do I think is on a 5-minute delay? If I were to buy the stock now, who am I selling it to 5 minutes from now, 30 minutes from now, 4 days from now, 2 weeks from now, a month from now? Where does the demand-and-supply game shake out?

The best funds you're thinking about—are they spending more time on that game theory? Here's a hypothetical example: this company announced awful earnings. The stock is down 30%, but it's generally a good company. If I buy now, this afternoon the quality long-onlys are going to have reviewed the thing and talked to management. Are they spending more time on that game-theory aspect—who's my buyer in 5 minutes, who's my seller in 5 minutes, who's my seller 2 days from now—than anything else?

I don't think they're spending more time, but I think all the best funds do this religiously.

Is it implicit or explicit in the process?

It should be explicit.

So if I was at a pod shop and I went and pitched my portfolio manager, I would have at the top, “Here's who I plan to sell the stock to.” Is that kind of how explicit it is?

Yeah. I mean, because when you sell, you don't know who you're selling to, but it would be like, “Here's who I think the buyer is going to be. Here's what I think makes the stock go up.”

And what makes a stock go up could be fundamental, but the right answer is rarely fundamental. Based on everything I see, I'll give you a good, high-level example. Let's say, based on your channel checks, you're like, “Wait, there is a structural shift because people are afraid of, let's say, potential retaliatory tariffs on this sub-piece of my business.”

The company is about to go to a conference in a week. I'm sure they will either talk about it or be asked about it. So my catalytic event is 7 days from now, when management goes to this conference. That's where my channel check today comes to fruition. We have a 7-day window, right?

I definitely get that. But then when I hear that, just to echo what we said 5 minutes ago, channel checks are no longer a source of alpha. So my worry would be: our firm believes that, but Thomas's firm believes that, and XYZ firm and everyone else believes it.

Maybe the answer is we buy because all of them are shorting because they think they're going to come out negative on tariffs, but I don't think they're going to be as negative on tariffs. So when they come out, all of them are looking to cover. So we're selling into their cover. That type of pod-monkey knife fight just gets really complex.

How deep into the “he said he thinks that I'm doing this, so I'm going to do this” are people getting?

Pretty deep. I would say pretty deep. I would say that's a big part of the job, trying to figure out what the market is actually going to do. How much of the market are other pod shops playing with each other, right? Who's sitting at a poker table is the most important thing to figure out. And once you figure out who's sitting at the poker table, you need to figure out who's in the hand.

Let me ask you another question on that. There have been a lot of people recently talking about the pod shops, and a big piece of the pod shop right now—you mentioned, hey, they're going to go to this investor day, analyst call, whatever it is, and I think they're going to say this.

I mean, a huge piece of the pod shop right now is: I think this company—the channel checks are great, all the credit card data is through the roof, there is no way they're going to miss the quarter. I'm going to buy because there's no way they're going to miss the quarter. They're going to have a huge beat. Outlook's going to be great.

Then you'll see these companies, and the way I've heard it described—everything makes sense—is that the week before earnings, the stock almost can't do anything but go higher because every pod shop is buying. And then even when they have great earnings, sometimes the stock will be down because all the pod shops are so full on the beat and then they have to sell.

So I guess, from a game-theory perspective, are you seeing people start to reverse that, or how are people playing that? It does seem—I've seen some companies report great quarters and then everyone's just so full of them, the stock's down 10% the next day.

You are absolutely right that that happens. But what I think a lot of people don't see is that what you just described is an actual risk factor. It's called the crowding factor, right? And you can model out crowding factor.

So if you are really a well-established pod shop, what you are almost always doing is actually hedging out your crowding factor because it's such a big factor of risk. So is it fine to be in crowded names? Absolutely. But then what you also want to do is short a bunch of crowded names too. So from a crowding-factor perspective, you're neutralized.

Now, how good people are at executing neutrality around crowding factor is something else entirely because it's really hard to short highly crowded stocks or go long highly crowded shorts, because everything fundamentally is telling you not to do it, right? But that's why the long-short market-neutral guys are so good at what they do: they recognize that this is a big risk moment, and they take the other bet as much as they take the crowding bet.

Yeah. A lot of people don't realize that crowding is actually a risk vector that is modeled across all the pod shops, and it's pretty well neutralized.

No, I just keep thinking back to the DeepSeek day, when DeepSeek came out and they announced, hey, we did this model. I don't know if people fully appreciated it—there are debates about how much went into it and everything—but it was crazy that Nvidia and the typical AI players were not down the most. It was the power plays—Talen, Vistra, all these big power plays—that were down the most.

Yes, they were exposed to AI, but what it really was was this hidden crowding factor, where if you had any exposure, you could just say, “Hey, I'm going to go long these IPPs because I get extra AI exposure for that.” So I just always think about that day and then all the beats that we were talking about.

I've got a few more questions, but you talk to way more clients than I do, way more people you interact with. So I just want to ask you: as you see where I've taken this conversation, is there anything when it comes to AI and investing that we haven't even touched on or have glanced over that you think listeners should be thinking about or learning about as they think about AI and investing?

Yeah. Something I get asked a lot, especially by senior management at banks and very big hedge funds, is what my view is of AI going forward. And I don't think they care about my view. I think what they care about is, let's figure out what all the AI founders are thinking, because the world is probably somewhere in the ballpark of where they all think the world's going to go, right?

So I always find it interesting to not only provide my view, but also ask, what have you been hearing from other people in my seat around where they think the world is going to go? And the short version of it is fascinating, because there isn't close to a consensus on where the world's going to go.

There is a group of people who think that the most important piece in AI is really just spending money on the foundational models, and the biggest foundational model wins. It's a winner-take-all model, and one guy is just going to win. I think you see that with the funding rounds, right? You see extreme funding rounds happening. Some of them are very, very justified. Some of them are not quite as justified, but you see a lot of those.

But you also see a second group of founders, and I think I fall into this camp, where we basically say, look, there are enough foundational models out there that it will probably be competitive for an extended period of time, if not always be competitive.

So for these foundational models to collapse into a monopoly seems unlikely. I'm thinking about cloud when I say that. There was a period of time when people believed that there could only be 1 cloud provider, and today we know that's not true. You cannot have too many because the cost of building cloud infrastructure is insane, but you could use Google, Amazon, Azure, or Oracle almost interchangeably today.

I think AI goes into that world, and where I think the rubber meets the road—where I think the money gets made—is on the application layer. How do you leverage the fact that these foundational models exist to start building an application?

Said differently, the question really is: Is the future of AI the creation of these tools, the guys who actually create AI, or is the future of AI the software and social media apps that get created on the back of the fact that these tools exist and supply you with the infrastructure to do what you need to do?

No question Instagram cannot exist without AT&T and Verizon. It just wouldn't work. But there's also no question that Facebook is a bigger company than the infrastructure providers that have enabled Facebook as we know it today. So it's 2 different bets. Some people think it becomes the Google of the world, and some people like the way it shakes out. Some people think it goes into the applications of the world, the software companies of the world.

Look, a natural question from that: As you mentioned, there are some people who think it's win-or-take-all, and ChatGPT gets so good and so far ahead that eventually it's, “Hey, ChatGPT is better than your 5 competitors.” Because of that, it gets more data and more usage, and then 2 months later it's significantly ahead and just scales, and everyone uses ChatGPT.

There's another view: “Hey, yes, ChatGPT, but all these little add-ons and apps that you're building on top of it are where the real thing is.” I could imagine you saying, “Hey, I've built this model. I've built this integration onto ChatGPT, and it's so much better at picking stocks than ChatGPT,” or something like that.

When you see these pod shops, long-only funds, or whatever it is, spending big money building AI integrations and AI apps, do you think 2 years from now it's all kind of wasted because general-purpose ChatGPT will just be so much better that it incorporates so much more? Or do you think the way of the future is that every pod shop has this huge research budget and its own integrated little apps built on top, running on its proprietary data, that's actually really bespoke and useful? Does that make sense?

Yes, that makes total sense. I'm very confident in this answer because I think the people with the best ability to build AI are going to be the people with the most access to data.

The very big pod shops have access to a lot of data sources. One, they just buy a lot of data, and 2, they generate a lot of data in their day-to-day workflow. That will allow them to build internal AI tools that are superior to the smaller shops that don't have those resources.

Do I think ChatGPT is going to get 10 times better? Absolutely. The trajectory has been insane. There's no reason to believe anything else. But do I think ChatGPT, by having a better model, can all of a sudden get access to proprietary data? No. I don't think that world exists.

If you're trying to solve internal problems for yourself, you will always need internal data. You will always need to feed your AI with internal data. I'm even willing to bet that a second-tier foundational model with the right context fed in will solve more problems than the best foundational model with no access to data.

Last question. There's always a debate in AI: Have we hit the scaling laws? Have we hit the peak? Are we slowing down? Sometimes it will be yes, and then they'll hit a new vector, and it'll be, “Oh, it's accelerated again.”

Ignoring improvements in AI and finance, when you're working with the big banks and the big pod shops, obviously 24 months from now there's going to be more AI than there is now. But do you think usage is accelerating from where we are today? I could also imagine a world where you say, “Hey, they're already doing a bunch of AI, and 24 months from now, will they be doing more AI? Yes. But it'll be from 50% to 52%, not from 50% to 250%.” Does that make sense?

It does. Interestingly, I'm referencing a lot of banks when I say this, thinking about a lot of conversations with management at different banks and the big investment banks.

I think what they have all collectively acknowledged is that the real problems they want to solve might not be AI-solvable problems, but they want these problems solved nonetheless. To what degree can some of my problems be solved using AI?

I think there is a very serious willingness to adopt AI right now. We're seeing that. We're getting bought by all these banks. The seriousness and the willingness to put money down are absolutely there. The question is: Are the problems being solved?

I think where we are today with foundational models is that they solve some problems, but there are a whole litany of hyper-nuanced finance problems that you simply cannot solve. I cannot speak for other industries, but finance is nuanced. A model is nuanced. An investment memo is nuanced. Why you buy a stock is nuanced.

It's not, “I think the company's going to grow. I think the company's going to be a bigger company tomorrow. Therefore, I buy the stock.” It's not that simple. Especially in the public markets, there are expectations, demand and supply of shares, liquidity, and risk factors. What happens if a news event comes out? Do you double down or do you sell? How do you know you're right or wrong?

There are all these nuances in finance that need to be taken into account, and a lot of the models today, at least not yet, have been trained on these nuances. If you're sitting there at the top of a bank and looking at all the problems that surface to you every day, rarely are the problems big, simple blocks of problems. They're always nuanced.

You mentioned banks there, and I might be showing my bias here, but I realize investment banks are not the same as the big commercial banks. I would probably guess that investment banks are a little slower and a little behind in incorporating AI versus pod shops, long-only funds, and investment firms, just because pod shops, long-only funds, and investment firms are smaller. They're flatter, more flexible, and probably adapt a little bit quicker. They're also certainly less regulated.

I want to frame the question slightly differently. You mentioned banks and basically said, “Look, banks are running into AI, and even when problems might not be natural fits for AI, they're figuring out ways to attach them to AI.”

I just want to ask the question again for pod shops and investors writ large. If we fast-forward 24 months, the way I said it, AI usage is going up—no doubt about it—but is it incremental from this point? Or do you think 24 months from now we'll be saying, “Oh my God, these AI guys don't even touch the trading screens anymore. They just put all their notes into the AI model, and then it says, ‘I think this company is a buy,’ and the AI model says, ‘Andrew is a dummy. He thinks it's a buy; it's a short,’ and we've got all this data.”

Or, “Wow, this is really great. We combine Andrew's work with Thomas's work. He thinks it's a buy, he thinks it's a buy—boom, it's a double buy.” Is it accelerating, or is it incremental progress?

12 months ago, I had the same thought as you just mentioned, which is that banks are slower than hedge funds. Why would anything else be true?

We've noticed that that's not true, but it's not false either. We've noticed that the people who move fastest are the people with the most agency—where there is a single person who's invested in building a process and building a product for the company and saying, “Let's go.” Those are the firms that are doing so.

We've seen that at big investment banks, we've seen that at big pod shops, and we've also seen that at very, very small funds. It's really agency. If you buy AI expecting that tomorrow it's going to monumentally change your business model, it's not going to. It could, but it's probably not going to.

If you buy AI thinking, “I'm just going to get rid of 10 of my associates tomorrow because it's going to do the job,” that's just not really how it works. That's a dream, but it doesn't come to fruition all that much.

If you say, “I'm going to believe in the future, and I'm going to build today and iterate with my people on how to enhance their experience, how to help them research companies better, cover more companies, and react to situations faster,” and you iterate through that, chances are you're probably going to be right to some degree.

But that is a process, and you have to be committed to the process. In order to be committed to a process in an industry where generally we are relatively short-term in terms of how we think about everything, it requires agency. It requires someone at the top to say, “We're going to do this. It might take 12 months; it might take 36 months, but we're going to do this.”

This is like when the internet happened. I'm not going to miss out on this wave and play catch-up later because that's going to be painful.

This is the actual last question, and then I'll turn it over to you for closing thoughts. It strikes me that a lot of what we talked about through this conversation was how the people who are doing AI best are putting their information, proprietary notes, everything, into the AI and letting AI run with that.

Obviously, if our firm was you and me and we were putting in notes, that's 2 people. If it was you and me and 5 other people, that's 7 versus 70. Are you finding that the big firms—let's assume everybody's very competent at using AI—are using AI better than small firms, just because they have more people putting in data? Is there that famous scale data advantage they're getting? Is that a thing? You can tell me 1 versus 5, yes, there's a difference, but maybe 5 versus 50 is vanishingly small—or is it even better because more data really serves the purpose here?

Yes. Loosely speaking, there are 2 types of hedge funds that we serve. One type of hedge fund has a central risk model or a central book, and the other type does not.

For the hedge funds that have a central risk model or a central book, they have very sophisticated people sitting on these central models or central portfolios. What the central portfolio is often trying to figure out is which analyst is good under what circumstances. Maybe Andrew is really good with mid-cap internet companies going into earnings but does not have a good track record with large-cap internet companies going into conference season. Can we both backtest and forward-test these hypotheses? Can we avoid data mining?

If you are in a central book and you know the quirks of all of your analysts and all of their coverage, then you can say, “When Andrew takes a mid-cap internet company into earnings, I'm just going to double the stake. When Andrew takes a large-cap internet company into conference season, I'm just going to halve it, or I'm just going to hedge out all the factors on the back end.”

There are a lot of these biases that you might not even know. Maybe you just like stocks with high dividends in low-interest-rate environments, and it's a bias that you don't even realize. A really good central risk model is going to be able to figure that out.

The second thing the central risk model is trying to figure out is performance evaluation. Are my guys getting stocks right because they're lucky, or are they getting it right because they're right? Are they getting stocks wrong because they're unlucky, or are they getting it wrong because they're wrong?

Is there a way to systematically prove that he or she is a really good analyst and that this just wasn't their day? They took 4 companies to earnings, and all 4 were wrong, but they were actually precisely right in their assessment. There was just some big macro event that rocked the book. A good risk model tries to figure all of that out.

This might be a conversation for a different day, a different time, but I am curious about that. I gave the example of the increasing crowding factor, where you have a company and you know they're going to beat, so you're long it. It beats, and then sometimes you're seeing the stock come down on pristine earnings because it's so crowded.

How does a risk book, in your experience, judge this? Andrew thought consensus was 5 cents. Andrew thought it was going to be 20 cents. It was 20 cents, spot-on. The outlook was incredible, and the stock was down 10%. Let's ignore that Trump announced 250% tariffs, so the market was down. But he got the positioning wrong. How does it judge positioning versus forecasting?

Positioning is hard to judge, but you can basically judge everything else, and what is left is positioning. It's sort of like PCA, or principal component analysis, back in the day.

How do you know if an analyst is precisely right or wrong? The short answer is, you know what the analyst forecast is, you know what the historicals are—or the newly updated historicals are—and you can do it side by side. You can do, basically—and every analyst does this—forecast versus actuals, or budget versus actuals.

The question is, we know that each analyst does that, but that analysis is not consumed on a global level, at a fund level. So what risk books want to do is say, “For every single analyst, for every single PM, can we consolidate their forecast versus actuals so that we know who has the widest variance and who has the narrowest?” We can adjust it by sector.

There are some sectors where revenue variance is basically nothing because it's all contracted revenue, and there are some sectors where revenue variance could be relatively huge because they're consumer companies and they're chasing trends and fads. How do we adjust for those spreads and figure out if Andrew is a really good analyst who gets the numbers right but gets the positioning wrong, or if Andrew's numbers are just totally off? Even though he got the stock right, either he got lucky, he got the positioning right, or it was a combination of both. You can actually strip those factors out.

Thomas, we've gone over time. I was really enjoying it, and we went to a lot of different areas that I didn't think we were going to. We got schooled on risk. But Thomas Li, founder of Daloopa, this has been really insightful. Again, you've just got a really wide range of how a bunch of different firms are using AI, so I really appreciate you taking the time. I'm looking forward to catching up soon.

Of course, Andrew. Thanks for doing this. This is fun.

All right, this is Andrew coming in to wrap it up. Uh thank you for joining and listening to the full podcast. Uh I really enjoyed the conversation about how AI is transforming finance. Hopefully it gave you some fresh ideas for improving your workflow. Uh look, I'll just wrap it up with the Loopa. If you're interested in learning more about them, seeing how they can help you focus on analysis, save you time, enable smarter, faster decisions, incorporate AI, uh you can sign up for a free account by visiting dupa.com. That's dupa da.com. Uh and you can check that set up and get a uh free trial there. Thanks so much. A quick disclaimer, nothing on this podcast should be considered investment advice. Guests or the hosts may have positions in any of the stocks mentioned during this podcast. Please do your own work and consult a financial adviser. Thanks.