Fintool's Nicolas Bustamante on using AI to improve in investing
Andrew WalkerNicolas Bustamante
Bustamante’s highest-leverage recommendation is to break the investment process into explicit tasks, then decide “which one can I delegate to AI?” A prior-quarter memo can become the template for the next earnings update, while screening can combine ordinary valuation filters with transcript-level conditions: founder leadership, the CEO mentioning future buybacks, cash available for repurchases, and a record of buying below intrinsic value.
AI’s current investor edge is retrieval and synthesis at a scale that turns a day of work into seconds or minutes. Walker’s examples include searching roughly 50 Caesars transcripts for acquisition commentary and assembling a decade of restaurant same-store sales; Bustamante extends the ladder from one-company KPI extraction to peer comparisons and whole-market qualitative-plus-quantitative scans. The next breakthroughs are “offline and parallelization”: hundreds of simultaneous questions and research that continues overnight.
The durable human role shifts from producing analysis to architecting, challenging, and judging it. Bustamante says AI went from writing 5–10% of his team’s code to effectively 100%, and expects finance to follow: an analyst covering 50 names might cover 100 while doing more complex work. The scarce skills become “taste,” pattern recognition, workflow design, and responsibility for the final bet—not mechanical summarization.
Concentrated investing remains resistant because markets are fat-tailed and the winning edge case can look indistinguishable from a deserved zero. Carvana carried leverage, deteriorating fundamentals, governance concerns, related-party dealings, and short research; Tesla could be framed as a carmaker worth more than the global auto industry or as an Elon Musk option on entirely different businesses. “That’s why we need the human in the loop”: a base-rate model can rank a basket, but a ten-stock fund must choose the exceptional survivor.
Trust depends on constrained sources, citations, and verification rather than a model’s fluency. Bustamante calls SEC filings the “ultimate source of truth,” cites Fintool at roughly 98% on FinanceBench versus about 40% for ChatGPT and 50-something for Perplexity, and says broader web retrieval must verify each fact before incorporating it. One hallucination or wrong diluted-share count can make a PM “lose trust.”
Delegating extraction need not eliminate learning if the investor deliberately moves manual attention up the stack. Walker’s compromise is to let AI summarize years of executive compensation and peers, then read the latest proxy himself; in one Chipotle analysis, cross-company comparison would surface a buried provision paying compensation in RSUs when performance reached 200% of target. AI can also monitor changing 14A metrics and potentially off-cycle Form 4 grants, though Walker’s experiment produced only four or five valid cases from 20 Fintool results.
Bustamante sees a large current adoption gap, while Walker fears AI will become table stakes. Some large firms still prohibit ChatGPT, while smaller Kennedy Capital in St. Louis is described as deeply AI-enabled. Walker’s darker conclusion is that non-users may “get your face ripped off,” while users merely keep pace as markets become more efficient.
1. Workflow decomposition turns a chatbot into an analyst
Bustamante’s opening prescription is procedural: map the path from idea generation to investment—or rejection—into perhaps 50 specific tasks, then ask, “Which one can I delegate to AI?” Funds often already possess this map because they use it to train junior analysts.
For a quarterly Home Depot memo, the system can ingest the Q3 example, read the Q4 earnings call, 10-Q, press release, and 8-K, then reproduce the established structure and extract the required figures. On the sell side, the analyst reviews it, adds the summary or conclusion, maintains or changes the price target, and publishes quickly.
Hedge-fund memos are less standardized: the relevant template may vary by company, industry, PM, and thesis. A margin-pressure thesis might demand inflation mentions and specific operating indicators, so the valuable input is not a universal memo but the investor’s own prior example.
Screening shows why task definition matters. A conventional filter might specify market cap below $10 billion and P/E below 30; AI can add founder leadership, management’s buyback language, balance-sheet capacity, and whether previous repurchases occurred below intrinsic value, then rank the resulting opportunities.
2. Personal pattern matching requires more than a small portfolio
Walker has struggled to upload a successful thesis and ask for similar ideas, or compare today’s portfolio with his historical winners and losers. Bustamante agrees this is “super complicated”: Fintool reserves it for large enterprise work requiring portfolios, coverage universes, internal memos, stock-price data, and SharePoint information.
The sparse-data objection is sharp. A concentrated manager holding eight stocks for two years makes roughly one new investment per quarter, unlike Renaissance’s enormous stream of trades; there may be too few observations for AI to distinguish genuine skill from an idiosyncratic winner.
Bustamante’s counterclaim is categorical: with enough SEC filings, calls, presentations, expert-network transcripts, internal criteria, and contextual data, language models will eventually search qualitative opportunities just as machines conquered quantitative ones. “I’m 100% sure it will happen,” although the market should become more efficient and alpha harder.
Fintool itself grew from Bustamante hearing Buffett describe scanning small caps and reading a booklet of Japanese companies to find what made business and management sense. His premise was that this previously human, partly qualitative assessment had become machine-readable.
3. The analyst’s job moves from production to orchestration
Bustamante’s software analogy is deliberately provocative: AI progressed from writing 5–10% of his team’s buggy code, to 30%, then 50%, and now effectively 100%. His conclusion—“AI is the best software engineer on the planet”—supports a finance role modeled on the architect or “meta thinker.”
At a large bank, juniors who summarize 10-Ks in standardized language are directly exposed. The constructive version is that an analyst covering 50 names can cover 100, or surround Chipotle with five relevant peers and produce a richer study rather than another basic note on inflation.
Walker wonders whether embodied “gumshoe” work becomes the new edge: attend a franchisee meeting, sense morale at three out of ten, and compare that observation with the seven-out-of-ten expectations embedded in the stock. Bustamante concedes proprietary data always helps, while anticipating systems that eventually analyze management videos and patterns of excitement or concern.
Public qualitative data remains underused. Bustamante says that at FinChat they downloaded long-form podcasts and asked AI to isolate investor-relevant discussion of capex, product launches, competitors, and whether LLMs are becoming commodities—information a tech PM cannot manually extract from every four- or six-hour appearance.
4. Parallel and offline agents could invert the research workflow
Today’s capability ladder begins with a sourced table—such as stores opened in each of the past eight quarters—then expands to five-company KPI comparison, management commentary, and whole-market scans blending adjusted financial metrics with qualitative criteria. Bustamante says the largest queries can run 20–50 minutes.
Walker calls his own usage “super Google”: Fintool searched years of Caesars acquisition commentary in 30 seconds rather than a day, while AI can rapidly assemble same-store-sales histories for Wendy’s, McDonald’s, and Burger King. Useful as that is, he asks how it makes him smarter rather than merely faster.
Bustamante’s first answer is parallelization. Instead of asking five primer questions sequentially, an investor could request 100 or 200 analyses at once: business model, how the company makes money, financials, valuation, CEO compensation, and peer benchmarking.
The larger inversion is offline work. Once an agent understands that an investor values cash-rich companies, proven capital allocators, possible spin-offs, or CEOs arriving from competitors, it could research from 7 p.m. to 9 a.m., score moat and management, push candidates proactively, and learn from the investor’s feedback.
5. Long-horizon AI backtests are contaminated by hindsight
Walker asks whether 20 years of successful investments create a moat over a 21-year-old beginner. Bustamante notes that anyone can study historical winners such as Apple or Coca-Cola, but Walker’s Philip Morris challenge exposes the problem: its brand, addiction, distribution, and low P/E looked attractive, yet government action could plausibly have destroyed the equity.
Walker argues that a useful qualitative system would need contemporaneous context, not just outcomes: the news, prevailing narrative, policy statements, expert views, and what investors could have known then. Only with that surrounding corpus might it judge whether tariff panic reflects genuine impairment or “Mr. Market” becoming irrational.
Backtesting remains structurally weak. An eight-position portfolio held for three years yields only about 51 observations over 20 years, while a modern LLM already knows what happened after any historical cutoff; prompting it to ignore Apple’s future does not remove that knowledge. Bustamante sees short-horizon news trading as more tractable than proving a ten-year concentrated strategy.
6. Delegation should move learning up the stack
Walker’s strongest objection is that building a model manually creates understanding: entering 5% revenue growth and tracing it through the statements teaches something a finished spreadsheet cannot. Bustamante accepts the risk—“the more we do the manual work, then we learn”—but argues extraction is no longer the right learning frontier.
Their compromise is layered work. Let AI summarize five years of compensation and benchmark peers, then read the focal company’s latest proxy manually; the investor arrives with historical context while still wrestling directly with the document that governs the present thesis.
In the Chipotle example, a customer manually unpacked a roughly $2 million cash incentive, with 75% tied to company-performance factors and 40% of that linked to measures including comparable restaurant sales and cash-flow margin. Bustamante says a five-CEO comparison would flag a missed footnote: if achievement reaches 200% of target, the remaining compensation is paid in RSUs rather than cash.
Bustamante describes an analysis that scans successive 14A filings for changed incentive metrics—say, net income falling from 50% to 10% of the formula—and tests whether changes coincide with deteriorating results. Walker adds off-cycle Form 4 grants as a possible signal of good news ahead, while acknowledging current extraction is noisy.
7. Fat tails preserve the need for human judgment
A Fintool presentation on Carvana surfaced the bearish evidence easily: leverage, deteriorating performance, a related company run by the founder’s father lending money to Carvana, and the Hindenburg report alongside other short research. Walker says that fact pattern would make the base case a zero.
Bustamante’s answer is uncertainty, not retrospective certainty: Carvana “might have been a zero,” might still prove one, yet the rising stock could devastate a short. Finance is “fat tail,” and when being wrong has consequences, “in one second you can be wrong and it’s game over.”
A quant can short a diversified ETF of AI-identified red flags; a concentrated long-only investor with ten positions cannot rely on the basket outcome. Bustamante therefore keeps a highly paid human responsible for the insight and the bet, even if AI performs virtually all preparatory analysis.
Tesla makes the narrative collision explicit. Walker contrasts a roughly $300 billion global auto industry with Tesla around $1 trillion, while the bull case is “Elon Musk is taking us to the moon”; Bustamante says valuation also requires judgment about Musk, regulatory credits, Optimus robots, and a proposed NVIDIA competitor. The stories cannot be reconciled by arithmetic alone.
8. Source control is the prerequisite for investor trust
Bustamante says one large customer initially dismissed Fintool in favor of ChatGPT, then realized after about three months that it produced nonstop hallucinations. On FinanceBench, he cites approximate accuracy of 98% for Fintool, 40% for ChatGPT with search, and 50-something for Perplexity; these are his stated figures, not independently examined in the conversation.
Fintool’s sequence was to master filings as the “ultimate source of truth,” then add investor presentations, earnings calls, and YouTube transcripts. Web search was due within two weeks, but each retrieved fact would require verification before entering an answer.
The commercial standard is unforgiving: returning a hallucination or wrong diluted-share count is “the best way to lose a customer.” That explains why finance retrieval products can feel slow—the answer passes through multiple sourcing and validation steps before a PM sees it.
Uploads are most valuable when the data is genuinely private: internal memos, Excel files, sell-side research, expert calls, or recorded management conversations. Public work can still focus the model; Bustamante cites SemiAnalysis on liquid-cooled data centers and an Abilene site expanding from 100,000 toward 200,000 GPUs as context for prospective GPU demand.
9. Better inputs—and sustained curiosity—remain human advantages
Bias does not disappear when private data arrives. Walker found ChatGPT favored the legal side represented by six uploaded filings over the side represented by five; management transcripts similarly skew bullish because CEOs emphasize transformations and cost cuts. Bustamante’s remedy is to privilege numbers, compare several calls, and benchmark peers rather than score executive language naively.
For recorded calls, Bustamante points to Granola as a leading option that can preserve the raw transcript and create notes. He says extracting additional information ultimately comes from asking small questions.
The best questioning example came from a Datadog investor-relations executive: ask a company about its competitor—Dollar Tree about Dollar General, or McDonald’s about Burger King—then infer what the answer reveals about inflation, inventory, and management’s own preoccupations. Bustamante’s conclusion is that asking good questions remains an art.
The closing disagreement is about how long the edge lasts. Bustamante says tools can change completely in three months and urges investors to “stay curious”; some major firms still ban ChatGPT while Kennedy Capital in St. Louis is already highly AI-enabled. Walker fears AI will become “table stakes”: laggards lose badly, while users keep pace as markets become more efficient and returns compress.
Full transcript
You're about to listen to the Yet Another Value podcast with your host, me, Andrew Walker. Today's podcast is a really interesting discussion with the founder of Fintool, Nick. Nick has a lot of insights into AI and investing. And look, I know what you're thinking. You're saying, "Oh, hey, Fintool is a sponsor. This is a plug podcast." I don't think it is. I ask him a lot of different questions about how individual investors and fundamental investors can incorporate AI and improve their AI workflow in a lot of different ways. Right off the top, at the start of the episode, I'm going to ask him, "Hey, what is the one thing that a fundamental investor can do to improve their usage of AI right now?" I ask lots of different questions about use cases, downsides, everything. I think you're really going to enjoy it if you listen all the way through. So, without hesitation, here we go: my interview with Nick from Fintool.
All right. Hello and welcome to the Yet Another Value podcast. I'm your host, Andrew Walker. With me today, I'm happy to have on Fintool founder Nicolas Bustamante. Nick, how's it going?
I'm good. Thank you for having me on.
Thanks for coming on. Just remind everyone that there's a full disclaimer at the end of the episode. Nothing on here is financial advice. You can listen to the full disclaimer at the end. I'm super excited to talk to you. I've been talking to Fintool for about a year now. You guys suggested coming on, and I'm always looking for ways to improve my use of AI as an investor. Obviously, you guys run an AI tool for investors, so I thought this could be a really good conversation.
Let me start with the first question. If our listeners stop listening after 3 minutes, let's give them something to take away. If you had the average investor who comes on and listens to this say, “Hey, Nicolas sees thousands of investors use AI. What is one thing that Nicolas thinks every investor could improve in their AI usage?” what would that be?
We work with a lot of hedge funds, large banks, and big consulting firms like PwC. Our product is sort of like an AI equity research analyst. What I've seen from successful hedge funds and value investors using AI is that they have a clear breakdown of their workflows.
From idea generation to making an investment—or even not making the investment—they break down their workflow into very specific tasks. Usually, they did that to train their analysts, right? But now they do that and say, “Okay, among my 50 tasks, which one can I delegate to AI?” Some of them are pretty small.
I was chatting with a fund this morning. They do quarterly memos. When Home Depot reports new earnings, they ask the analyst, “Okay, read the earnings call, read the 10-Q, the press release, and the 8-K, and draft a note. Extract the key numbers,” et cetera.
It turns out that if you feed Fintool an example of a memo for, let's say, Q3, and ask it to do the same for Q4, Fintool will do it automatically. They will delegate this task to AI. By the way, they will do Home Depot, but they will also do Lowe's and more companies.
I would say it depends on your task and which tasks you choose to dedicate to AI. I can give you more concrete examples. For screening, I think that's a good one. Today, you do quantitative screening: “Hey, I'm looking at companies below $10 billion in market cap, healthcare, P/E ratio under 30,” blah blah blah.
But then you can say to the AI, “Okay, look for this, and also I want companies that are founder-led. I want the CEO to mention buybacks—future buybacks. I want you to check if they have the cash to do buybacks. I want you to check if they did successful buybacks in the past. Did they buy anything below intrinsic value in the past? If yes, combine it, rate the company, and show me the opportunity.”
So, yeah, it's per task and per workflow.
For screening, I found it really interesting, exactly what you're saying, where you say, “Hey, find me companies that are trading below 10 times price-to-earnings, and the CEO has said, ‘I think we're trading below intrinsic value,’ in the past 3 quarters.” That's really interesting because trading at 10 times price-to-earnings is something you could have screened for on Bloomberg or anything for 15 years, but marrying it to something specific in the transcript is really interesting.
Obviously, this is a very generic example, but those are really interesting applications. The memo thing really jumps out to me, so I want to probe a little bit on that.
You used Home Depot as your example and said, “Hey, Andrew is my analyst. We have a position in Home Depot, so every quarter Andrew writes”—most people are familiar with this—“the 1-page overview.” If you're a very institutional investor, it's going to be: “Home Depot reported $1.50 per share in earnings. The consensus was $1.35. We thought it would be $1.40. Here were the same-store sales trends. Here's what we're seeing. Here's what we're not seeing.”
It's going to be a 1-page summary of everything you need to know. I'm interested: when you talk to people who start outsourcing that to AI, how heavily are they outsourcing it?
Are they doing something like, “Look, Home Depot has 100 different retail competitors. I have a position in Home Depot, so I make the Home Depot memo, and then I have AI make the 99 companies that I follow so that I don't have the hallucination memo”?
Or are they going so far as saying, “Hey, everything is getting made by AI. I don't get rewarded for regurgitating what's in the quarter”? I'm really interested in that.
I think the sell-side is a good example. You're an analyst at a bank, you cover 50 names, and you have to send a report to your customers within a few minutes to an hour about a new earnings release. It's very standard. There is no customization. Maybe the customization is, “Oh, you know, the estimate was this,” and so you feed the document—the initial memo—to Fintool.
Fintool will duplicate it. You review the output, saying, “Okay, this is good.” Maybe you're going to add a line. Usually, what they do is add either the summary or the conclusion, saying, “Hey, we're still bullish, and we maintain our price target at blah blah blah,” and then they publish. So that's very simple.
For a hedge fund, what I've seen in the memo is less standardized. Every PM and every analyst has a custom memo. That's it. Also, AI is popular because there is no such thing as 1 memo for every company. You might have 1 memo per company or per industry, and depending on your thesis, you will look for different things, right?
Sometimes they look for margin pressure, mentions of inflation, and so on. That's very useful when they give us an example of the memo, upload it, and the AI can duplicate it.
I like that you said “example” because one thing I've tried historically, and I haven't had success with, is uploading something like, “Hey, here was a successful thesis I had on a company that played out. Find me a similar comp. Find me a company with a similar thesis.”
I haven't had much luck in terms of that pattern recognition and mirroring. Maybe I'm giving 2 poor examples, but have you had people succeed with that? Let's say you're a mutual fund or even a hedge fund. If you're a concentrated hedge fund, you might have made 50 investments over the past 10 years. If you're semiconcentrated, maybe it's 500 investments.
Have you had people say, “Hey, here's my top 50 ideas. Generate me 5 ideas to research that kind of look like this”?
Or another way I've thought about it is, “Hey, here's my track record. Here's my current portfolio. Tell me which things in my current portfolio look like my best investments, and which things share qualities with my worst investments.”
I haven't had much luck, but just because dumb, dumb me, drooling at the mouth, hasn't had success doesn't mean no one has. Have you seen people have any success with that?
Look, it's super complicated. Even we do that only with super-enterprise customers. We ask them, “Hey, give us your portfolio.”
Obviously, if it's some hedge fund—for instance, HMI Capital in San Francisco, or whatever—those guys are super concentrated, and then you can look at the 13F. But for firms like TCW, they have $200 billion under management, so they give us a portfolio and their coverage universe of companies.
What we want from them is also additional information, and that's why every solution is similar, right? We ask them to connect their data. We go to SharePoint and download the data. In this data, they have memos like, “Hey, here's this on this investment. Here's why we invested in that investment.”
Then we do the work of looking at the data, looking at the stock price, and trying to gather as much information as possible to identify what a winning investment for them is, and then trying to duplicate it across the whole stock market.
It's not something that you can do with a consumer large language model. It's very intensive and very error-prone. It's just super hard.
I had a question on this later in my notes, but I'll just ask it now.
I was talking to a friend the other day about using AI, and he hit on something. Both of us realized at the same time: if you’re Renaissance and you’re making 100,000 trades a day, AI can be super useful for course-correcting you and studying your history. You’ve got so many data points.
My friend was like, “My average turnover—let’s make the numbers easy—is that I hold 8 stocks, and I hold them for an average of 2 years. That means, on average, I’m making 1 investment per quarter.” And he’s like, “I can’t.” He’s not saying AI isn’t useful; he’s saying that he doesn’t generate enough data for AI to go back and say, “Everything you’re doing is so esoteric or so niche or so at the edge.”
He was saying AI just can’t help him in that way. That doesn’t mean it can’t help in a lot of other ways, but he wants AI to make him a much better investor. He’s like, “It can be your super-Google, but it can’t really improve you like it could a quant fund.” What would you say to that?
I’m very familiar with the Renaissance use case. You may have noticed, but French guys like us tend to work either in AI or in math-oriented fields.
When Jim Simons showed up and said, “I’m going to apply machine learning and statistics to the stock market. I’m going to predict prices, identify price discrepancies with machine learning, and trade on that,” the traders who were on the bank floors were like, “No way. I rely on my intuition,” and so on. Obviously, it didn’t work for Jim for maybe 5 to 10 years. Now Renaissance is $90 billion, compounding at 30% a year, so ultimately it worked.
I think you might be low on the compounding there, by the way.
I wish I could have invested. But it works for the quantitative-data side. Now we have the qualitative-data side, where our customers are more long-term-oriented. Obviously, those guys say, “I rely on my intuition a lot. It’s an art, not a science.” But large language models now understand text data and nuances, so ultimately it will work if you give the large language model enough data.
You start with the SEC filings, earnings-call presentations, expert-call networks, and so on. You input your criteria, and at some point the AI will run offline and find investment opportunities. Those opportunities might be different from the investment opportunities you found in the past, but they will be the same in terms of your criteria. If you do small caps, they’ll be small-cap-oriented. If you focus on profitable companies, they’ll only be profitable companies.
I’m 100% sure it will happen. I understand that people have a lot of skepticism, just like back in the day when the quant guys wanted to apply machine learning to the stock market. I can tell you, as a software engineer, I lost my job to AI. You have to understand that 2 years ago, AI was writing maybe 5% to 10% of our code, and it was buggy. We were frustrated with the AI. Then it went to 30%, then 50%, and today it’s basically 100%. AI is the best software engineer on the planet, and it’s inevitable for every profession.
Let me ask you the next question, then. I have written before, and I’ve said before, that I worry about this. Fifty years ago, you could have made really good returns if you—I’m rereading The Snowball right now—and a lot of what Buffett did was literally go through Moody’s Manual and say, “This company trades for $10, earns $5 per share every year, and has $20 per share in cash. What a deal.” Then he bought it.
Obviously, the man did a lot more than that, but those are literally the stories in the book. You could have made tons of money doing that 50 years ago. Renaissance came along and obviously did this on steroids, with a lot of other stuff, but that quantitative stuff is dead, right?
For the past 20 years, you could make money doing qualitative things—things that weren’t in the numbers. If your thesis was, “This is trading for 8 times price-to-earnings, and I think I’m going to make alpha,” you were dead. If your thesis was, “This trades for 30 times price-to-earnings, but the earnings number is actually 8 times once I make adjustments,” you could make a lot of money.
I worry that quantitative investing is dead. Renaissance killed it, as did AQR and whoever else. I worry that qualitative AI is going to kill it. Are professional investors going to have a role in the world in 5, 7, or 10 years?
The reason I started Fintool was pretty much because of the Berkshire meeting. I had worked in AI for a decade. I went to the Berkshire meeting one time, right after selling my previous company. My previous company was a legal AI company, a sort of AlphaSense for the legal industry.
Again, I heard Buffett saying, “If I do small caps, I can do 50% a year,” and so on, just by scanning through opportunities. I showed up to the meeting, and Buffett was like, “I had this booklet of Japanese companies. I read it all, and I looked at what made sense from a business perspective and at the management.”
That’s a very quantitative assessment that was impossible to do before the large language model. Now the AI is going to do it, and I think it will still be possible to generate alpha. It will obviously be harder. The market will be more efficient, and a lot of it will be done by AI.
The same way that, as software engineers, what we do today is that we’re the architects and the meta-thinkers, while the AI is writing 100% of the code. It’s almost like when you interview someone: it doesn’t even matter if the person knows how to code, because the AI is doing it all.
A year ago, I was with my friend in San Francisco. The guy went to MIT and then Stanford, and he was like, “No way AI is going to write my code.” Everyone was like, “Yeah, obviously. It’s way too hard”—similar to some of our customers today. Then it happened, and he was like, “Okay, what should we do now?” Basically, you become the orchestrator of the AI systems.
This actually dives nicely into a thought I’ve had in my mind. Do you watch basketball? I think you and I have talked about basketball before. You follow basketball a little bit, correct?
Especially in San Francisco. We have a great team.
I’m glad you mentioned it. One thing I’ve had in my mind is Steph Curry. He’s a top-15 player of all time, a multiple-time MVP, and all that sort of stuff. If you took Steph Curry, rewound time 50 years, and put him in the NBA, I don’t even think he would be an NBA player.
His greatest skill is shooting. There’s no 3-point line, so he can’t stretch defenses like that. Steph Curry also has a lot of problems with his ankles. If you put him 50 years ago, playing in flat-top Converse shoes with the medical science of that time, I don’t know if his ankles could even hold up. One of the 20 best players of all time might not even be playable 50 years ago.
In today’s NBA, he’s probably the most valuable player. It’s him, LeBron, and 2 other players who have been the most valuable players of the past 15 years.
In contrast, I think of the legacy power forward who wasn’t big enough to be a center but was 6'10", couldn’t shoot, and could rebound. Not Tim Duncan—he’s one of the best players of all time and had a midrange shot—but that Twin Towers style. A lot of those plodding centers and power forwards were super valuable 30 years ago. They’d be played out of the NBA today.
Roy Hibbert was super valuable in 2010 because of verticality. He couldn’t even play in today’s NBA. The reason I ask is this: you framed something where—I think we can talk about the lawyers—but 30 years ago, for a portfolio manager, quantitative skills might have been the most important thing.
AI, in the same way that medicine and sports have evolved, is going to make some skills a lot less valuable and some skills a lot more valuable. What type of skills do you think AI is going to be a leverage point for? If you were an analyst or portfolio manager today, where the puck is going in 5 years, what skills would you say you really need to be focusing on?
That’s a very good question. I would say that today, the only 2 jobs where you can really see the impact of AI—and by AI, I mean artificial general intelligence, this super-smart system—are software engineering and law.
If you’re a software engineer, you see it impacting your job. It’s very hard to get a job as a junior engineer. If you’re a lawyer, I saw that with my previous company. For lawyers, it’s just words. You ask the AI to do an NDA with 100% accuracy, and when it happens to you, it happens so fast because every day you have new models and new capabilities.
Even for us, we’re like, “How do we hire a software engineer? What is the job if it isn’t coding?” For portfolio managers and analysts, I see that with customers all the time. If you’re a large bank and you employ tons of juniors whose job is to summarize 10-Ks in a nice, standardized way, everything will be done by AI.
The bank will say, “Okay, Jim, you cover 50 names.” Now you can cover 100 names. But they might also say, “Jim, maybe you can do more complex work.” Let’s say you were looking at Chipotle. Maybe you can also look at the 5 companies around it and produce a more complicated analysis.
It’s not just an analysis of the impact of inflation on Chipotle’s business. You’re going to do that for your own brands, et cetera. The analysis you produce will be much better. That’s one way you can frame it: you’re doing more qualitative work over time.
Let me propose a different hypothesis. This is one I’ve talked to people about, and I’ve gotten pushback in different areas. Twenty or 30 years ago, quantitative skills were important. Your portfolio manager at Long-Term Capital Management blew up, but you wanted a Long-Term Capital Management-type person—somebody who could do math in their head quickly.
That was in part because there weren’t even a lot of computers back then. Excel and modeling have been outsourced.
But as you get into reading 10-Ks, understanding them, and getting into niche cases—I want to come back to niche cases in a second—I have wondered: if I could upload the most successful investments of everyone for the past 30 years into AI and have it start spinning out ideas, is that sort of stuff—anything that's in the filings, all of that edge, all of that alpha—gone?
Where the alpha for the next generation is almost, “Hey, can you go meet with management in a room and read their body language better?” Or are you better at the gumshoe-type stuff, where you go to the franchisee meeting and talk to 100 franchisees, then plug that into an AI model and say, “I felt pretty depressed at that franchisee meeting”? The AI says, “The stock is forecasting franchises at 7 out of 10, and you felt 3 out of 10. It's a short.” Do you think that gumshoe work and that personal work becomes more important with AI, or could you tell me, “Hey, AI can get on the earnings call and read someone's body language better than anyone else in the world”? So it's actually worse. How would you think about that?
I think having a data edge is always a source of generating alpha. I will say, from my experience working with a lot of PMs, you can generate a lot of information with public data.
You can do it in 2 ways, just by looking at it better. Buffett is known to be able to compute the owner earnings pretty fast for a company, right? But you can say to your AI, “Compute the owner earnings, get rid of the stock-based compensation and all that stuff, and rank all the companies.” Instead of having an analyst do 1 or 2 companies, the AI is doing that on the whole stock market. Then you have, in my opinion, data sources that are not sufficiently used.
I was discussing this with one of our clients. It's a big firm, a big PM investing in tech, and he said, “Hey, there are so many podcasts where the CEO of Microsoft is going out there and commenting on capex, the new AI capabilities, whether an LLM is a commodity, and in that case, whether they're losing money.” The guy cannot listen to all these podcasts. The Lex Fridman podcast is like 6 hours; this will be like 4 hours. Typically, for what we did at FinChat, we downloaded all these podcasts and asked the AI, “Identify everything that's relevant for an investor: every mention of capex, every new product launch, every mention of how competitors are doing.” I think you can generate alpha that way.
But, to your point, I think we'll see a sort of Citadel focusing more on the qualitative-data side. As you say, looking at a video, analyzing management's pattern, and determining whether they're excited or not, then trading on that.
Let me—I’m not sure where to go with that. I guess when you find people using Fintool or AI in general as an amplifier, what skill do you find they're trying to amplify the most?
It depends on how AI-enabled they are. We open an account with, let's say, 50 seats, and we have the guys who are always on ChatGPT and Perplexity. They pick it up and ask thousands of questions, right? We also have some people who don't even know how to use AI. They don't even know about ChatGPT; they don't know how to prompt.
I'm just laughing because I was talking to a very successful investor the other day, and he was like, “Hey, Andrew, I listened to one of your podcasts and you said if you're not using AI, you're going to get left behind.” I was like, “Yeah, I really believe that. I can do stuff in 15 seconds that used to take me a day.” And he was like, “I've never opened ChatGPT before.” I was like, “Oh, buddy.” Sorry, continue.
I was just laughing because you hit the nail on the head. 100%. I mean, you go to this meeting, you meet with a very famous investor, an investing legend, and you're excited, and you realize that he has never used AI. Obviously, the guy who's 30 years old in the room is a bit like the guy who set up the meeting.
I would say there are several levels. At a very simple level, they have a question: “Hey, Chipotle, how many stores did they open per quarter over the past 8 quarters?” Because it's a KPI, they won't find it in Bloomberg or FactSet, and they have to deep-dive into the earnings calls and the 8-K release. The AI will create a nice table with a source. It's kind of 101—a simple question about a company.
Then you have a harder question, where you say, “Hey, same-store sales, but you're going to compare that to the 5 companies in the industry. You're going to read every earnings call and analyze what management says about that.” It's multiple companies, merging the numbers with a qualitative assessment.
Then you go on and on, and you have these massive queries where people scan the whole stock market for a bunch of criteria, both qualitative and quantitative: “Get rid of the stock-based compensation and do this and do that.” The workflow takes 20 or 50 minutes to answer their questions.
I'll just make it personal, and I might even clip this out because I suspect that most of my friends who I talk to—and most of my friends are plus or minus 5 years from me—are investors who run similar concentrated-value or event-driven styles. I suspect most of my friends are using FinChat, ChatGPT, and everything the same way, and I think this might be useful for them.
The characteristic you described at the beginning is kind of how I use this. I might just clip this specific piece of the podcast and put it on Twitter because I think this will be very useful for people. The way you described it is kind of how I use it. I probably spend 30 minutes to an hour of my day, every day, in ChatGPT and Fintool, and basically I'm using them like a super-Google.
My famous example with Fintool is that I was trying to pull what Caesars had said about acquisitions and how well they had done over the past 7 years. If I had done that myself, I would have had to go through 50 transcripts on my own. Looking through them would have taken me at least a full day of work. Fintool did it in 30 seconds. It took 5 minutes to go through all that with ChatGPT.
I'm using it like a super-Google. I'm using it to say, “Hey, quickly build me same-store sales for Wendy's, McDonald's, and Burger King over the past 10 years.” It's fantastic for those things. What do you think I could be doing? Again, I feel limited because it's just improving and grabbing data and summarizing it much more quickly, which is great. It's freeing up tons of my time. What else could I be using it for? Instead of just using it as a super-Google to free up time, how could I be using it to make me smarter, or what else could I be using to improve my job?
I think ultimately you feel limited, and you're limited by the technology. Finance is pretty hard to nail because it's a combination of words and numbers. All these finance software tools, like Fintool, look very basic, right? Data extraction at scale, looking at the earnings calls, and so on. But as time goes by and we have better models, we can do more complex analysis.
The 2 breakthroughs will be offline work and parallelization. With parallelization, you ask 1 question and get 1 answer. What if Fintool could go and answer 100 or 200 questions at the same time? You want a company primer on whatever company, and instead of asking 5 questions, Fintool will return the business description, how they make money, the financials, how they value the company, the CEO compensation, and a benchmark against their peer group.
Offline work is something that exists only in software engineering. It's very recent. You work from 9:00 a.m. to 7:00 p.m. What if Fintool could do research from 7:00 p.m. to 9:00 a.m. the next day? Once Fintool knows how you research companies, what's important to you, and how you ultimately make a decision—which requires you to explain a bit to the AI—I'm excited by this opportunity.
The CEO comes from a competitor. He's a great capital allocator. They're talking about a spin-off, and they have plenty of cash on the balance sheet. The agent will work in the background and try to find new opportunities. I think it has to do something very complicated, akin to a Morningstar with a score. It needs to score the opportunity: “Hey, competitive moat is 5 out of 5; management team is 3 out of 5.”
We go from you asking questions and being limited by yourself—what type of question can I ask?—to the AI just pushing information to you. You say, “Okay, this is interesting; this is not interesting,” and the AI learns and pushes you even more relevant information. That's the future of AI systems, where you don't go and ask a question; the AI pushes you relevant information.
That's an interesting future because what it would rely on is your past interactions with AI and your past investing, right? Are you saying that in the future, as AI continues to scale, there will be almost a moat in, “Hey, I can upload the past 20 years of successful investments here. I can upload them to AI. It can learn from that. So I've got a moat versus a 21-year-old who's listening to this and thinking, ‘I want to break into investing.’”?
When they start working with AI, they're not going to have any successful investments to point to.
So, they’re going to be way behind because they don’t have that historical data. Does that question make sense?
Yes, it makes sense. You can also argue that there is an absolute good investment, right? If you look at the track record of Apple and Coca-Cola, you can deduce from the past that those were good investments, and then you can try with AI to ask, okay, why were they good investments, and rank how much of that is—
How much of that is N-of-1, though, right? Like, yes, Coca-Cola in the ’80s, and Buffett makes great investments. Philip Morris was famously the best-performing stock of the past 50 years or whatever.
How much of these are N-of-1? I could make a very simple argument for it: hey, Philip Morris—addictive product, good brand, big moats. You’re selling cigarettes: distribution, addictive products, low P/E. Great. Yes, it’s awesome. Or I can make another argument: hey, Philip Morris, find me something that trades for an 8-times P/E because they’ve got the threat of the government bankrupting them over their head.
That was very much N-of-1. Historically, it looks great, but things could have gone a different way, and the government could have demanded a Fannie-and-Freddie-in-2008-style pound of flesh, where all the equity belongs to us now. So, how much can AI learn from that?
I have a follow-up question to that, actually. When I discuss with my friends who are working at high-frequency trading firms, they have terabytes of data, and obviously terabytes of price, volume, options, and all that stuff. For AI to be, on the qualitative side, a bit omniscient—a bit like Warren Buffett—it needs a bit of everything. It needs the news, the newspaper at the time, the story at the time, and the more data and context, the more the AI will be able to understand what is a good opportunity, because often Mr. Market is just crazy.
He’s crazy because of China tariffs, and maybe the AI can read what the White House says and what some experts are saying and say, okay, I think humans are panicking and I think it’s a buy, right? So, it needs a lot of data, and it’s still the early days.
I’m laughing because you say tariffs. It’s like, man, I try and train an AI on when and what Donald Trump is going to tweet. Good luck with that.
Let’s—so, we’ve mentioned Renaissance several times. Quantitative has been taken over by computers. There’s no doubt about that. Quantitative is dominated by computers, machine-learning models, AI, whatever you want to call it. Qualitative, to my knowledge, has not. Now, there are things—you can get a quantitative fund that runs on value factors, everything—but concentrated qualitative has not. I haven’t even seen anyone try to do an AI. Have you guys thought about back-testing or launching, like, hey, here’s our qualitative AI concentrated portfolio? How do you think a fund like that will work, or do you think there would be any success there?
What is very hard is when you have machine-learning models and you want to test the models: you need to see the output of the test quickly. So, if you train a recommendation system for, I don’t know, Tinder, you want to swipe right, swipe left, and look at the data. But when it comes to a long-term investment portfolio, you’re not going to wait 10 years to see if this would be back-tested, right? You could back-test it—and again, you get a very small sample size—because even if you back-test it to 2005, we’ve got 20 years, an 8-position portfolio, and you’re holding everything for 3 years. That’s roughly 17 times 3. So, we’ve got—what is that, 54? 51—51 data points. That’s not that great.
But it is interesting because if we’re saying, hey, the future AI is going to be increasingly better at qualitative analysis, launching an AI fund right now would kind of be the model, you know.
Yeah, I think it’s a bit hard with large language models, in the sense that if you show them just past data and try to find some correlation, the LLMs baked into their training set already know the future. You can say to the LLM, “Look only from 2008 to 2015 and give me an answer for that parameter,” but the training data—they know the story. They already know 2023, or they have a guess.
You train on Apple and say, “Look only at this,” but the LLM knows that Steve Jobs will come back, and you have your prompt where you say, “Hey, just consider that time frame.” So, I think the long term is very hard. That’s why I think quantitative finance is focused on the short term.
I did see a large trading operation in New York where they were kind of trading the news. It was a bunch of kids, 25 to 30 years old—I mean, Bloomberg terminals, all of them—and they were reading the news and saying, okay, there is a gas leak here. Then they were trying to find companies that were exposed to this gas thing, and they were looking at the balance sheet and stuff, trying to short them. I do think an LLM can do a better job for that. You read all the news.
I have no doubt it could read a news article and do exactly what you’re saying faster than ever. Though I will say, I remain impressed by how often you’ll see something and an hour later it’ll be like, oh, that might not have been that good for this company, and it’ll take kind of an hour for it to lead into the price. So, maybe there’s still room for humans, or maybe AI can do it better. I feel like we should think about launching, like, hey, here’s an AI 8-stock portfolio that finds the best qualitative things. I think that could be interesting.
Let me ask you separately. Speaking of concentrated plays, I was talking to another friend, and again, it was a similar conversation. I was like, man, I’m not saying all your thinking should be outsourced to AI, but I was giving them my pitch: it really frees up a lot of time. A lot of things you do by hand, it will free up.
He had a 2-fold pitch, and I agree with him. I actually kind of cleaned this model up. He was like, look, there’s the old thing in banking: “I build every model myself.” And it’s not because the Bloomberg models aren’t impressive, or whatever you want to download. You build every model yourself because then you’re building your understanding of the company, right? There’s just something unique about plugging in, “Next year’s revenue growth will be 5%,” and kind of flowing it through all the income statements and seeing how it impacts the company versus just having it presented to you, right?
So, his thing was, once you start using AI, you kind of move away from building models yourself. That was number 1. I’ll let you respond to that, and then I want to hit you with his maybe more powerful point on number 2.
I think it’s very true. I think that’s the unknown: for us humans, the more we take notes and the more we do the work—the manual work—the more we learn. I was chatting with a customer this morning, and they were analyzing the compensation of the Chipotle CEO. They looked at the proxy, and they looked at—he has a cash incentive of like $2 million, and they were like, okay, $2 million: 75% of that is a company-performance factor, and 40% of that is comparable sales, comparable restaurant sales, and cash-flow margin and stuff.
So, they spent a lot of time, and they came up with this analysis on the comp of this guy. We were discussing together, and it was like, yeah, but look, AI can do this. It can also do the previous CEO, benchmark this compensation with the new compensation, look at Starbucks, and look at 5 different companies. So, yes, you can do that manually for 2 hours and learn about his compensation, or you can have an output and start learning about whether the compensation is standard, whether everything is weird, and what the compensation was for the CEO before and the CEO now.
Ultimately, I think you will have to delegate that to AI and try to learn on more complex information. At some point, you just have to have the taste, the pattern recognition, the understanding, where you say, hey, AI, gather me all these trends, and then I’m going to think hard about this. I’m going to stop spending my time on just data extraction.
No, my solution—and not that my solution is perfect or anything—but my solution has been: I love using AI for the broad stuff. I love saying, hey, go summarize the past 5 years of executive compensation and trends at this company and maybe some of the company’s peers, right? Because that information takes multiple hours of digging through these proxies. These proxies are massive statements, and I feel like they’re intentionally a little bit obfuscated, and AI can return it like that.
Then what I try to do is look at that, learn from it, and then, for the company I’m really focused on, maybe spend a little bit more time just on its most recent proxy, right? So, I’m getting the nice AI summary, and then I’m trying to understand the most recent one because the most recent one’s the one we’re on, and it’s the one that really matters, and I can use the insights from the previous ones and the peers to do that.
So, that’s just been my solution. I’ll pause there if you have any thoughts on that process or anything. And I do have 1 more question I wanted to ask.
Yeah, I think it’s exactly right. And back to the conversation I had with a customer about Scott, the CEO of Chipotle.
So, he did all his analysis, and I think he did the same with AI, and he missed something. When you ask Fintool to compare his compensation to 5 other CEOs, Fintool will flag, “Hey, if the guy reaches 200% of his target, the rest is not paid in cash; it’s paid in RSUs.” That’s something you want to know, right? It was probably buried in a footnote or whatever, and he missed it. It was key information.
Yeah. This is off the cuff, but one of the things I’m always interested in is executives with nonstandard compensation. Probably the most famous example would be Elon Musk in 2018 or 2019. Tesla gave him tons of stock options that said, “Hey, if this goes to 300 in the next 8 years, we’re going to give you basically $50 billion. But if it doesn’t, it’s zero.” So, you’re taking a huge upside bet.
And he got it. Then the Delaware judge said no, and this is all pre-split, so the stock’s way up. But that was a very nonstandard package. I just want to know: how have you found—you can say Fintool specifically, AI, whatever it is—how have you found it analyzing nonstandard packages? I think this will flow nicely into my next question.
Yeah. What we do—and we try to run this analysis offline—is, for instance, run models and look at the change in compensation metrics. They say, “He has an annual cash incentive, and here are the factors,” and we’re asking Fintool to look at all the 14A filings and say, “Did they change the factor over time?” Then you can do one more step: is that correlated with the company—maybe net income decreasing or the stock price crashing—and try to identify that as an early red flag?
With a new DEF 14A: “Aha, new compensation.” It was 50% of net income, and now it’s only 10%. What does it mean? What do they know? That’s the sort of complex analysis we try to do. I also like the nonstandard compensation. Sometimes they don’t have high compensation, but then they have a private security budget, a private jet, school for the kids, that type of thing.
Yeah, exactly. Who should I speak to for my private jet?
I can point you to some guys with some real fun ones if you want.
No, one thing I should probably work with you guys on is this. One thing I’m trying to build better is one of my favorite signals: most companies will grant RSUs and options once per year. So, every February, all the executives get RSUs and options, and when they do, they file Form 4.
Every now and then, a company will do off-cycle grants. They’ll say, “Hey, we’re giving them in April this year as well,” or something. That’s a really interesting signal, because they normally do that in April because good news is coming in May and they want to get everybody paid. Every investor loves those signals, but it’s very hard to pick up on.
I’ll tell you, I tried it on ChatGPT, and ChatGPT was terrible. It just gave me 0, and Fintool gave me 20 examples. Of the 20, only 4 or 5 were actually good, but ChatGPT managed to give me 0, and Fintool actually got me some examples. It’s something I can probably modify, but I think it’s a really interesting use case where everything you see is just a Form 4. If you can find the right context and so on, AI can solve it. Let me ask my other question.
My friend, who I was talking to the other day, said—and I can see this—he was like, “Look, Andrew, quantitative AI is so good, right? Renaissance-level quantitative analysis is so good. When you’re running a concentrated value fund, the edge is in the nuance; the edge is in the edge cases.”
This is an example I came up with, but I think you’d agree: Cliff Sosin with Carvana. Carvana is a stock that was down 90%; every short seller in the world had published a report on it. It was way overlevered, the profits were going down, and it was shrinking. If I had given you that set of facts, every base case would have said, “That is a zero.”
But it was the edge case where—and you know this is n-of-1—he was the only person who could make that business model profitable. There are lots of other cases, but he was saying, “Look, you want the nuance; you want the edge case, and that’s where the real money is made.” AI is never going to be able to detect that edge case, because what it’s going to do is read Carvana’s 10-K and say, “Hey, I’ve read 20 10-Ks like this, and everyone has failed. Everything that has all these red flags—the base rate is terrible—so it’s zero.”
Maybe that’s saying Carvana was your negative-EV lottery ticket that actually paid off, or maybe it was good, but he was saying you’re never going to find the edge case. What would you think about that?
Every Friday at Fintool, we do a presentation on a company, and I think 2 months ago I chose Carvana. It’s a great example. I asked Fintool to look at the accounting and compare it with other companies in the space. Then I put in the Hindenburg Research report and all this research. I was super bearish and, obviously, there were tons of red flags: the father runs a company, he loans money to Carvana, and so on.
First, that’s why we need the human in the loop—someone who is highly paid for the insights. That’s true for lawyers now, because AI is writing everything, but you need someone with a deep understanding and someone to make the bet. Second, that’s also why finance is really hard: it’s fat-tailed.
Carvana might have been a zero; it might be a zero now; maybe it’s a fraud—we don’t know—but the stock is rising. If you were shorting it, you lose a lot of money. That’s why we need a human in the loop: it’s fat-tailed, and if there are consequences to being wrong, sometimes in 1 second you can be wrong and it’s game over.
No, and look, it’s not Carvana; it’s Tesla, right? The red flags and the short sellers who’ve been burned on Tesla are unbelievable. The base case is, “Hey, we’re starting up an auto manufacturer. Cool. All of those go bankrupt.”
But the red flags—again, this comes back to—I think my friends would have said, “Hey, if you built a bucket of companies that had all the Carvana characteristics, or all the Tesla characteristics, 98% of them would probably massively underperform. But it’s the 2 that work.” If you’re running a quant strategy where you’re spreading the bet over 100, you can do that. But if you’re doing qualitative analysis, where you’re saying, “I’m picking 1,” it’s very difficult to use AI there. So, yeah, I don’t know.
Yeah, because you don’t have a basket, right? You can make the case where you create a sort of ETF and short the ETF of all the companies that AI has identified. But if the premise is, “I have 10 stocks; I’m a super-concentrated, long-only investor,” then it’s hard.
For Tesla, I tried it with Fintool. I said, “Hey, give me the real value of the car business. How many cars are they selling? Get rid of this EV-credit stuff.” It comes up with maybe a valuation number, but then you need to say, “Okay, you need to factor in the fact that Elon Musk is Elon Musk, that the credit thing is good,” and so on. Ultimately, you make a decision. I’ve done that math.
As my friend Dan once said, the issue with going short Tesla is you do the math and you’re like, “Okay, the entire global auto industry is worth $300 billion, and Tesla’s trading for $1 trillion, so it’s worth triple the global auto industry.” That’s your bear case. And the bull case is Elon Musk is taking us to the moon. You’re just talking about different stories at that point.
No, I was thinking about when I read the Fintool analysis. There was a part about the super goodwill that Elon Musk has and also the long shots. I think it was Optimus robots and the sort of NVIDIA competitors they’re trying to build. Even if the AI was like, “You know, there’s tons of goodwill for the company,” maybe not $700 billion, but that’s why you need the human in the loop at the end.
No, it’s not just Tesla; it’s the edge cases. There’s a company that my friend and I debate all the time, and he’s like, “Look, this company has built this killer model. They’ve got this killer mousetrap, razor-razor blade. The returns on invested capital are going to be crazy, crazy.” It’s a medical-device company, and I’ll be like, “That’s cool, but we’re valuing them at $1 billion, and their R&D to develop this device was $4 million.”
It’s really hard for me to understand how you can make a $1 billion company with a $4 million R&D product. I understand more goes into their sales force and everything, but it feels like somebody else could come and copy this product and sell it 50% cheaper. It’s kind of where, look, Tesla—for years, the short thesis, one of the many short theses, has been, “Hey, their R&D makes no sense versus what they’re saying when they’re trading globally.” But I guess somehow they’re making it work, you know.
Just a couple more questions. This might be the last question, but one thing I think Fintool might be about to change is that right now, Fintool is only focused on SEC filings. I go back and forth on whether it’s better, when I’m evaluating a company, to do it in Fintool and not have to worry about hallucinations.
With ChatGPT, I’ve asked, “What were this company’s earnings 3 years ago?” It’ll pull a different company’s earnings, or it’ll pull an earnings number from Seeking Alpha that wasn’t the company’s earnings. What are the advantages and disadvantages—and it doesn’t have to be Fintool-specific—of saying, “Let’s only use company filings”? If the company is lying, they will be held legally liable and go to jail for it. The information there can generally be trusted unless it’s an outright accounting fraud. Or should we go broader and incorporate things on the internet when I’m studying this company and using AI?
Right.
Yeah. No, you’re right. One of our biggest customers initially said, “Hey, don’t bother. I’m using ChatGPT.” I think 3 months later, they realized there were nonstop hallucinations. With the search, we actually have a benchmark.
There’s a benchmark called FinanceBench. It’s the leading benchmark for equity research, where they provide the question and the answer. Then you run an LLM model: you run ChatGPT with search, you run Claude, you run Fintool, and you get an accuracy score. I think Fintool is around 98%, ChatGPT is around 40%, and Perplexity might be in the 50s.
Our approach is that we have to start with the ultimate source of truth and be very good with that, even if it means the product might be slower and limited in its data sources. We added investor presentations and earnings calls, and now we’ve added all the YouTube transcripts. We consider those a source of truth because if Zuckerberg is on a podcast talking about his vision for AI capex in 2030, that’s him. It’s a source of truth.
Now we’re going to release web search in the next 2 weeks. This is a tricky one because if it searches sources like Seeking Alpha and Motley Fool and returns a bunch of data, it’s game over because it could be wrong. The way we have to build it is for it to search the web for very specific information, and then we have to verify the accuracy of every piece of information and integrate that into the answer.
The best way to lose a customer—to lose a PM—is for the guy to ask about diluted share count or whatever, and for you to show an answer with either a hallucination or a wrong number. He’ll lose trust and be very unhappy. That’s also why finance is hard, and why most finance chat-retrieval software is a bit slow. If you want to do it right, there are so many different steps.
Let me just, in terms of my use case—my personal use case, to make it very selfish—I generally am not uploading much to Fintool. I’ll point it to something and say, “Hey, I’m looking at GE, particularly GE’s 10-K. Tell me XYZ.” Or I’m looking at Pfizer and want to understand its oncology program. Pfizer just had an oncology investor day; you can find the transcript here. Go look at that.
How much should I be uploading my personal stuff to ChatGPT, or whatever it is, when I’m working with AI?
I’ll say, from the perspective of a software vendor, we do everything so you don’t have to upload the data. For instance, we’ll process 20,000 podcasts for you, and you’ll never have to download the video or the audio.
Having said that, there are private internal memos, Excel files, and things that you own that might be good for the LLM to see. Sometimes customers tell me they have an interesting blog post that they think is valuable, and we’re going to capture that blog post. Or they have a piece of sell-side research published by one of the firms, and they’re going to upload that.
The clients who are uploading—are they getting better results, do you think?
When you upload, yes. For instance, I was chatting with the guy from the famous blog SemiAnalysis, which studies NVIDIA and companies like it and does very detailed research. They’ve been on the podcast a few times.
This guy is a genius, and his analysis is so valuable. When you upload it and ask questions about liquid cooling in data centers, his analysis matters in terms of understanding the market size and the technology. Maybe he has a piece saying the Abilene data center is close to completion, has 100,000 GPUs, and they want to expand to 200,000 GPUs. Then you have the information: “Oh, okay, they’re going to buy GPUs.” That’s buying pressure.
I want to say that the more information, the better.
No, it’s just—I’m surprised you said that because a lot of that information is in the public domain. I’m surprised that uploading it sounds like what people are trying to do is say, “Hey, I’m uploading something in the public domain, whether it’s a podcast or that post, and getting Fintool or AI or whatever it is to really focus on it when it does the analysis.” Am I thinking about that correctly?
It’s public. At the end of the day, we’re going to capture it. But let’s say you have an expert call. Let’s say you call the management team, you call investor relations, and you have a transcript of a call that isn’t available. It’s only on your computer. This is the sort of information you will upload.
Legal requirements aside, should I be recording my calls with management teams and uploading them when I’m looking for processing?
I’ve discussed this with so many customers who use recording software. They use Granola, they record on Zoom, and then on SharePoint or OneDrive they have all the transcripts.
So, again, legally, if it’s legal, I do try to get recordings of my calls with management teams. But one thing I worry about—I’ve had this issue with ChatGPT before—is that for a while I was doing a lot of legal analysis in it. I noticed that if you and I were in court and I uploaded 6 filings from you and 5 filings from me, then asked ChatGPT who had the better argument, it was always saying you because I had put 6 documents from you and 5 documents from me.
I worry that if I do a management call, it’s going to incorporate too much of my bias. I can’t help it: if I’m short a company, I hate the company, and my questions are probably going to be more pointed than if I’m long a company. If I’m long a company and it announces a really bad quarter, my questions are probably going to be more pointed as well. I’m worried it’s going to incorporate too much of my bias.
Obviously, there are things that are said—not that I’m getting MNPI—but they’re just different or phrased differently than they were in public. I’m worried about giving too much bias when I’m uploading that, if I’m asking the AI to interpret it in any way.
Yeah, you’re so right. ChatGPT is a horizontal product. It costs about $20 a month, and they’re trying to cover many, many use cases. Most of them are consumer-type use cases, and that’s why in every vertical you have big software vendors that leverage the same type of technology but do a better job of understanding the information and avoiding biases.
If you go to legal, you’ll find Harvey and Legora, which are big software vendors now. When you upload a brief and compare the arguments, it won’t say, “This guy is better because he uploaded more data than the other guy.” The next step is figuring out how to reduce the bias.
I see that in earnings calls. If you don’t do anything and just feed the earnings call to the LLM and ask, “Is it bullish for the company? Is it bearish?”—well, if the stock is down, the CEO is always pumping it up. “We’re going to make it. We have this new transformation. I’m cutting costs.”
You need these extra layers for the AI to understand: “Okay, be rational. Forget about his language. Look at the numbers. Put that into context. Look at the 5 earnings calls. Look at other companies, and then answer.”
If you have 5 more minutes, I’d love to pull on this just a little bit more. Let’s say I’m going to start recording my calls with management teams and uploading them to Fintool, ChatGPT, or whatever it is for AI analysis. How should I be thinking about my conversations with management teams in ways that would make the upload maximally beneficial for AI?
I might start doing this with the podcast because, obviously, I hope I’m having smart guests on. But if I’m having a conversation with a management team, how should I frame the conversation in a way that I could then upload it and make it maximally beneficial? Do I need to steer them toward providing numerical answers to everything to make it easier, or do I need to ask them to keep everything tightly defined? How can I get a good transcript or recording to give to AI?
I think now you have several recording software options.
I think Granola is the leading recording software. They do a great job of getting the whole raw transcript and creating notes. I guess, at the end of the day, it’s about asking small questions to extract some sort of additional information.
What types of questions would best feed the AI? Is it getting numerical answers, or would it be something like, “You guys said same-store sales were down 2%. Can you break it down in 4 different ways?” Just getting statistical breakdowns of things that they’ve already said. What would be best?
Something I learned talking about this with the IR person at Datadog—she’s extremely smart, and I think she was a PM before—is that she said, “I’m just answering all the questions with publicly available information.” But there’s a trick: sometimes people don’t ask about the company; they ask about the competitors.
If you’re in a Dollar Tree meeting and you say, “Hey, do you think Dollar General is seeing some sort of inflection point or impact from inflation?” as a way to get an answer, they might say, “Oh, yeah, they’re seeing a ton of pressure from inflation.” Then you can deduce that they struggle with that, too, or that their inventory is affected.
I don’t know. I think the heart of asking good questions to a management team and having them answer the questions is—I don’t know—it’s an art.
No, look, what you said there is interesting to me because I had not thought of it that way. Hopefully, they’re not giving you MNPI when you’re calling the IR team or the management team, but I hadn’t thought of that.
Like, “Hey, forget your company, McDonald’s. Forget you guys. I want to talk about Burger King. Their numbers looked a little bit different from yours last quarter. Let’s talk about Burger King and what they’re seeing that’s different.” Then, if they’re saying, “Oh, it’s got to be inflation. Burger King is seeing a lot of inflation,” you could probably guess McDonald’s has inflation on its mind.
That’s really interesting. Nick, this has been super useful. I probably need to have you guys spend more time with me. Dak actually showed me—you mentioned the workflows—and I’ve been spending a lot of time with some of the historical comp workflows. Again, that’s where I’ve really gotten a lot of benefit, but I probably need to spend more time on it. Any last thoughts before we wrap this up?
No. A last thought on AI is that it’s extremely early. I think most people form an opinion on the technology as it is today, and because, as human beings, it’s very hard to grasp what an exponential is, sometimes you try something and, 3 months later, it’s completely different. But you have the opinion of what it was like 3 months before.
You’re hitting the nail on the head. It’s funny you say that today because last night—or today—Google released its new video editor. People were saying, “Here’s the video that they were pumping out 3 years ago, and here it is today.” Everybody was mocking it 3 years ago, and today it looks like literal, best-movie-quality work.
I do have some friends who, when I talk to them, I say, “Hey, I’m spending an hour in AI every day, whether it’s ChatGPT, Fintool, whatever,” and they say, “I tried that 2 years ago, and it was worse than Google.” I’m like, “Man, 2 years—it’s just crazy how much better it’s gotten over that time.”
Yeah, that’s why you need to stay curious. It’s also hard because I’m always pushing my team, like, “Hey, use this new AI tool.” Then they’re like, “But it sucked last month.” And then you say, “Okay, I’ll give it a try,” and they’re like, “Oh, wow, I’m doing so much with it now.”
It’s just exhausting. Keeping up with AI news in general is exhausting, but at the end of the day, it’s worth it because if you can have an edge—if it does your job—that’s good.
The worry—I want to get mad at my friends when I say, “I’m spending a lot of time in AI,” and they’re like, “That thing is so dumb.” It’s like, “Dude, I literally just had a conversation with you where I said, ‘Hey, I’m worried AI is going to replace every quantitative finance job,’ and you’re saying, ‘AI is not as good as Google.’”
I think it’s probably a little bit better than Google. It’s interesting, as you said, to stay curious, but I’m worried it’s one of those things where, if you weren’t using email 10 years ago, you were just dead. Email was table stakes, right?
I’m worried AI is one of those things. It’s table stakes: if you don’t use it, you get your face ripped off, and if you do use it, the returns just come down even further because it makes the markets more and more efficient. It just gets tougher and tougher over time.
But the gap is so huge, though. I go into big firms, and I don’t want to name names, but some of them aren’t allowed to use ChatGPT at work. They don’t have it; they can’t even use it on their personal computers. Sometimes we get calls from CEOs saying, “I need AI in my firm,” but most of them don’t even know what it is.
I think, at least for now, there’s a huge gap in the market. Sometimes we see customers—I don’t know, Kennedy Capital, for example. Those guys are small-cap-oriented and are in St. Louis, Missouri. They are so AI-enhanced compared to some of the people I go see in New York.
It’s just a matter of having 2 or 3 great people there. They were like, “Okay, we need to get on this thing.”
Well, I’m going to be following up and making you introduce me to Kennedy Capital because I want to pick their brain and start talking to them.
Nick, this has been awesome. Again, Fintool was the thing that opened my eyes to, “Oh, my God.” You start putting things in, and things that took you a day before just spit out like that. Then you can spend a lot of time on other stuff.
I appreciate you coming on. We’ll have to do a follow-up in the near future.
Cool. Thank you.
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.