Samir Patel on Using AI Intelligently Within the Investment Process

the primary purpose of this podcast is to educate and inform the views information or opinions expressed by hosts or guests are their own neither the show nor any of its content should be construed as investment advice or as a recommendation to buy or sell any particular security

Security-specific information shared on this podcast should not be relied upon as a basis for your own investment decisions. Be sure to do your own research. The podcast hosts and participants may have a position in the securities mentioned personally through subaccounts and or through separate funds and may change their holdings at any time.

Welcome everybody to a new episode of This Week in Intelligent Investing. Great to have you with us again and great to welcome a special guest, Samir Patel. the founder he runs escaladin capital i've known samir for a long time really respect his investing and his writing he's done a lot of work on

AI in value investing and that's what we want to focus on today. I think we're going to learn a lot. I'm really excited for the conversation. Samir, welcome to you and welcome to my co-host Elliot Turner as well. Thanks, John. Thanks, Elliot. It's great to be here. Pleasure to be speaking with you and thanks so much for the opportunity to share my thoughts with your audience. Yeah, absolutely. So, you know, maybe you can tell us.

how you're thinking about this, some of the highlights of what you found, and we'll take it from there. Yeah, absolutely. And I think it might help if I just really briefly explained how this all got started. You know, I think as value investors, we all try to be reflective and think about our mistakes, right? And so one of the biggest mistakes I think I've made in my career is with regards to COVID.

And, you know, it's a tendency of value investors. I think Lisa Rapuano talked about this with regards to the global financial crisis many years ago. that often you can be so focused on the micro, right? And we have this habit of being like, okay, you know, the macro is not really affecting us. We're focused on our own companies, their situations. But then every so often, there's this sort of event.

at a macro level that's so important that it can overwhelm those micro factors. And so I think, you know, one of the lessons I learned from COVID is, you know, obviously there's no plan. for your business when revenue goes to zero right it doesn't matter how good your business is if your business is shut down

That's not something you really plan for. And so during COVID, basically, the only important analytical issue became for many companies, you know, when will COVID end? Does the company have the balance to get through that, etc.?

And, you know, frankly, we were slow to respond in those early days when if we'd spent a little bit more time on it, we might have realized, okay, we own some great businesses, but they are going to face disruption from COVID. And maybe instead of reflexively doubling down on price weakness, we should actually be selling.

focusing on other businesses. So with AI, that's kind of how I came to AI, where... You know, I hadn't really paid too much attention to AI before probably late 2024, start 2025. And like many people, you know, I think when that first blender bot or whatever it was called from Facebook meta came out, you know, I played with it.

I had a fun time with it. I showed my mom, my wife. I'm like, hey, look, this thing can write a decent high school compare and contrast essay. Like, that's kind of fun that you just tell a chat bot to write me something on. you know, war and peace or whatever it might be. And there you go, you get this, you get this essay. But obviously, over the past few years, we start to see AI, you know, particularly with GPT 3.5, and some of the subsequent model releases, becoming more and more capable.

And so I realized at some point, you know, it makes sense to step back from my everyday looking at bottom up fundamentals of. value investments and to spend some time just thinking about AI, right? Because this was kind of when there was a lot of hype and maximalism, which there still is.

But I just wanted to investigate in sort of a dispassionate way, you know, a number of questions. One is how is this going to affect my companies first and foremost? And then the second is, you know, how am I going to use this in my business?

And so that's where that's where a lot of this came from. I spent a lot of time on it. You know, I think there's a lot of surface level analysis. And I think particularly once you get deeper into using some of the different tools, particularly if you're using it through APIs.

where you have more of an ability to customize and deal with, for example, the hallucinations and some of those things. It gave me, I think, a really broad view of not just how it's going to affect our companies, but also I became increasingly convinced of its potential to really...

help us from a value investing standpoint. And I know that's been a topic of some discussion recently. Guy Spear publishing the article asking, is value investing dead? And I kind of come out on the other side of that, where I think this is one of the greatest things to happen to value investing in a long time.

I'll kind of stop there and let you guide where you'd like to go. Yeah, that's a fascinating introduction. And I think, you know, I'd urge everyone to read Samir's Q4 2024 letter, which is a really deep dive into the way you're using. AI right now. And I'd imagine there's some evolution since then, but I've always been intrigued by this notion. I've seen it called Kasparov's principle or Kasparov's law that like a human with a machine is better than merely a great machine.

And so maybe to start, when you think about using AI, you're talking about using it, not it replacing. certain key functions so like how do you make sure one of the things that when i look at ai there's just so much information there's so much data there's so much we could do with it so how do you filter out the vastness of it and get specific and focused? Yeah, I think those are great questions. So I think let's start with your Kasparov question, which I think is fascinating.

John knows that I'm a big fan of mental models. I think I've given a couple talks over the years at conferences and things like that. And so I think about AI through... that lens of mental models and so you think about let's start with one for example trait adaptivity right and so it's this idea

popularized by a lot of people as you don't judge an elephant by its ability to climb a tree. Different people, different machines have different traits that makes them better, more or less suited for certain things. And so I recently read a book called The NVIDIA Way by Tay Kim, which is a little bit about the history of NVIDIA.

And one of the analogies that jumped out to me is this idea of CPUs versus GPUs, right? So for a long time, CPUs dominated workloads because they kind of think a little bit more like humans, you know, sequentially, logically. You know, AI is built on GPUs, as we all know by now, right? NVIDIA. And it's massively parallel. So when you think about investing, you know, what...

AI is not able to do. What that GPU is not able to do is sort of that reason. And I think we're all aware of this, that even the reasoning models don't really reason causally or from first principles or from truth the way that a human might. They're essentially very advanced pattern matching engines that can sometimes mimic reasoning in many ways, but they don't reason. And so that's how I fundamentally think about the divergence, the counterfactual or the corollary to that.

is humans are very, very slow, right? So if I want to read a 10K or a proxy looking for some basic information, say, how has management been compensated in this business over the span of 10 years? the company's guidance compared to the actual results and where has that differed over a long span of time.

To read that sequentially as a human would take a long time. AI doesn't read sequentially. AI sort of reads massively and in parallel. So I think the opportunity, as you discussed, as I mentioned in my letter, really marries human judgment.

right, which I don't think that the AI currently has the ability to replicate, I don't see the path to it getting there, at least based on LLMs, but using AI for some of that sort of mass data processing, and we've always had the ability to do that with spreadsheets with tools for the quantitative data. But what's a lot more interesting sometimes is to be able to go through things like conference call transcripts and...

things like that and extract the insights and the information much faster than a human would be able to do. And then that leaves more time for you to spend your own judging. Yeah, so maybe talk about that a little more, right? You now have more time.

to spend on human judgment so like what does that mean what have been the biggest shifts in your time allocation and cognitive energy now that you've had like a year over a year working with ai in this way Are you turning over more rocks, turning over your portfolio more? What does it mean? Yeah, so I think the portfolio turnover is obviously a little bit more dependent on market conditions as far as the research side of it.

I think what it does is two things. So I like to talk about and I think there's, you know, there's always another mental models, this idea of trade offs or opportunity costs, right? I think there's always a little bit of this concern that I hear when I talk to other people about like, well, okay, if I don't do the work myself, am I going to miss something? Right. And I think that's a little bit of a red herring where. For example, there's lots of kinds of analysis you can do on a company.

But most of us don't have the time if we run reasonably diversified portfolios to do all of that exhaustive analysis on every single company, right? There's always going to be some additional proxy or 10K or some additional detail industry report that you could look into. And so where I have seen the biggest application is, I think of my, if I can just step back, I think of my research process almost in three phases. So the first step is simply orientation.

Right. The first step is you come into a new business, you come into a new industry, you don't know any of the jargon, you don't know any of the history, and you're just trying to get a lay of the land. It's like if someone drops you off in a new country you've never been in before. You don't know the language, you don't know the customs, you don't know the culture. Your first step is just simply trying to get some sort of high-level overview.

And what I used to find is that oftentimes when I was doing my research, I'd be reviewing the same document, whether that was a 10K, a transcript, whatever it might be, a news piece, two or three times. The first pass was simply, like I said, that orientation of trying to figure out what is this company? What do they do? What are the very basics?

The second step is then once you have that, trying to develop some sort of more detailed understanding. It's like, okay, I understand at a very high level. Now let me dig in on the key issues. And then the final piece is sort of that due diligence piece, right? Of like, okay, there's certain issues that it's very important, but I know the exact details, the exact numbers I really dig in on these issues. And so I would say that AI...

is probably decreasingly applicable with a caveat in those three buckets, right? So it's amazing for that first pass. So I'll just give you an idea. You know, there was one company that was a peer to a company that we've invested in. I was interested in them, had seen them for a while, saw the stock go down on earnings. And I said, you know what, let me dig in.

And very quickly, you know, it would have taken me previously two or three days to get to this point. But within three or four hours using AI, I figured out, OK, there's enough red flags in this business that it kills it for me.

Right. I can just move on to the next idea. And you guys know how valuable that is as a portfolio manager or as an analyst where, you know, not spending two or three or four days or a week on an idea that's not a good fit for you allows you to go and spend that time on something else. So what I would say is that AI is enabling me both.

to get to kind of the meat faster, right, to get through that first step and to really get to, okay, what are the core issues? Then, of course, there's data extraction, there's, you know, being able to go back farther in time with some of those things I talked about, like looking at many years of data as opposed to a few years of data. because you just simply can't read 10 or 15 years' worth of 10Ks or proxies, right? It's just difficult to have the time to do that.

And I think the final piece of it is then you start thinking about, well, if I don't have to spend so much time reviewing securities filings, what are sources of information such as expert calls, things like that, human interaction, talking to other shareholders, more time.

spent with management or the board of the company to learn from a different angle that the AI can't replicate, right? Because the AI just simply doesn't have access to that data. So where I've been you know, where what's been interesting is initially I thought, okay, this is going to allow me to get done, you know,

five or 10 times as many companies. I think what I'm finding is that could be true, but I'm choosing to reinvest a significant portion of it in depth and understanding whether it's existing portfolio companies or a new company that I'm getting into, doing a bit more exhaustive.

work, including, again, having more time for expert calls, having more time for sort of human scuttlebutt type of research as opposed to just the desktop research. Yeah, maybe let's pick this apart, like in some of the pieces, you know, that early learning curve on a given company.

Talk about the tools that you use. Like, have you automated processes to kind of tap into AI and build that initial impression on a company? Or do you approach each project kind of as a discrete? Yeah, it's a mixture. And I've played around with a number of different approaches. I'm very fascinated by AI automation to the extent that my wife's actually launched an AI automation business. That's a whole separate topic. So I do use many AI automations.

As far as researching companies, no, not really. I do have a prompt library of standard prompts that I use. But one of the things with automation is obviously it takes time to build the automation. And then you kind of have to decide, okay, is that worth it? And then, of course, there's going to be more of a...

One size fits all output, whereas the way that I like to learn about companies is a bit more interactive, right? Like I might ask one question, but then you want to follow a question on that same topic or you see something of interest and then you go to that specific source material. And it's a back and forth between. sort of the AI generated portion and the actual source material. One thing that I think is very important is using the right tool for the right application. And so with

AI, obviously one of the large problems is hallucinations, right? As I mentioned, it doesn't reason causally. And in many cases, it can either get things wrong, it can misinterpret things, which is one set of problems, or it can simply make things wrong. So initially, you know, we tried using chat GPT or Gemini or the public LLMs.

And we found, or even, by the way, things inside of like AlphaSense, right? And we found that what's better is if you take the securities filings, you put them in something like Notebook LM, which largely doesn't hallucinate, it can misinterpret things. but it doesn't really make things up. And having that ability whenever there's a statement to immediately directly click to the source and review the source material, which I find very important, right? Because...

Obviously, and we're not even talking about getting to like, okay, the numbers you put in your model in terms of, you know, the cash versus the debt, but just even that understanding a lot of times, if there's something interesting, you want to be able to click through and dive deep on that topic.

And so it's really interesting because there's all these tools I have built, but just something like Notebook LM, I think gets you a lot of the way there in terms of being able to query. And it becomes, to your point about the Kasparov. you know and chess and everything it becomes more about asking the right questions right which is one of the reasons it's a little difficult to automate because for any given company the right question

isn't going to be the same. And the right question for the company might not even be the same at different points in that same company's evolution, again, depending on what the current issues are that's important from a valuation perspective. Yeah, that's super interesting. You raised a lot of...

Material that I'm fascinated by right now, when I got your letters from John, I was in the middle of this book, Co-Intelligence by Ethan Mollick, and he talks, basically the entire book is about how and why we should start using. AI as early as possible because the more you use it, the better you'll be at it. And this notion that, you know, how you ask the question will determine a lot of what sort of value you're able to take out of the AI that you're using.

And he also talks about hallucinations. And he says that hallucinations are actually kind of interesting. Everyone's focused on the risks of hallucinations. But if you know that AIs hallucinate and you understand why they hallucinate.

you could actually use that to your advantage, i.e. when you're looking to pursue something creative, when you're doing the steps that involve brainstorming, when you're doing things that are far more open-ended in nature rather than kind of specific and closed-ended.

then hallucinations are actually to your advantage. And you can find creative material that you can't actually create yourself. I thought that was fascinating. And then Notebook LM is like, it's guardrails around the material you're working with. It's very specific. set of materials so like you mentioned doing more expert calls with some of that free time or at least going deeper in some companies one of the things i started doing was creating these notebook lm folders on

industry verticals or sub-industry verticals or companies. And I've now done like, I don't know, hundreds, thousands of expert calls, and I've started organizing them by topic and then kind of scraping. asking questions of it like what's evolved what's changed what holes might i want to learn more about what other experts should i seek

Curious. I'd love to hear more of some of the specific tools into the weeds of where you're taking this. Do you have a preferred model when you're getting started early on? Talk me through some of the process. Yeah, it's almost hard to structure, right? Because there's so many different directions you can go. I think for a moment, I'll just start by sticking with the idea of picking the right tool for the job.

And, you know, again, back to mental models, Charlie Munger, you think about man with the hammer, the gallbladder surgeon, you know, whatever analogy you want to use. And I think. Again, to your point, that creativity, the idea of trait adaptivity, the creativity or the hallucinations can be good or bad. And so that's why originally, you know, I tried using chat GPT deep research, you know, to try and see what that would be.

I found I was spending so much time fact-checking or wading through inaccurate information that I largely don't do that anymore, except for very specific things, where I think it actually can be helpful. And I got this idea from my friend Gary. silver ring, Gary Resurus, is this idea of using it as a devil's advocate, right? To your point, leveraging that creativity of, okay, well, here's all of my research that is validated, is from good source materials. You know, what could I be missing?

And then you have this sort of very sophisticated pattern matching engine that can go out and might find some angle that you haven't thought of.

Right. And then you go through it quickly. And as opposed to like when you're researching a company that you don't know and you don't know what is and isn't true, you're like, OK, I understand this issue. You don't have to worry about that. Don't have to worry about that. Oh, maybe this is something I need to look into further that isn't covered in my research or that isn't covered in my analysis.

So I think that's very powerful. To your point about Notebook LM and the ways it can be used, I do also think there's a lot of... And I think a lot of organizations are looking at this, but there's a lot of data or there's a lot of things written down and the ability to go through that, you know, just as an example, when I was building some of these automations and I wanted to figure out, okay, what kinds of businesses work well.

for us when I built the letter screening tool. I built the rubric for that largely by asking book LM to look over our old letters and our old write-ups and say, what makes a business that AskLatin wants to invest in? And of course, you and I could name many of the factors off the top of our head where it's like recurring revenue, things like that.

Being able to do that without forgetting or without having to go through all of this, even being able to go through when you've written down notes on a company from years and years ago.

Or, you know, in the case of like my website, being able to go through all these book reviews I've written, all these mental models I've put together and just being able to query against that internal knowledge base to refresh your memory on something that you may have forgotten where I don't know if you have this, but I'll.

the time, I'm like, oh, there was this statistic I saw somewhere. I know in one of my research documents, I talked about this phenomenon, but I have hundreds of these things. I don't know where it is or where to find it. And it's amazing to be able to very quickly get back into that.

mind space. So I'll stop there. And I think you wanted to go deeper into maybe specific tools or applications. Yeah, I mean, one of the things that there are so many different models, right? And by the way, just knowing Gary, at least a little bit. It's so not surprising that he used devil's advocate as his best example. That totally resonates. You know, like, do you think you get different output from one model versus another? Do you feel that...

When you ask questions, you know, that perhaps you might be leading AI to an answer for you just with how you ask questions. So how do you avoid doing that? And are there kind of like structured series of like. fail safes that you have to review the output to make sure that yes even in you know you use notebook lm for certain things where you don't want to risk hallucinations but when you're using something else you know might you

Like, how do you make sure that you're not dealing with a hallucination of an answer rather than a high quality, like impactful information? Right. And I mean, my answer is largely I don't. Right. And that's why, you know, so. If you think about any process that you're implementing, you always have to think about the counterfactual. You always have to think about A-B testing. So just as an example, one thing I will do.

is if there's a company I know very, very well, like let's say it's a portfolio company, there's a little bit of this, actually I'll back up, there's a little bit of this Dunning-Kruger effect, right? And one of my mental models is this idea of product versus packaging, where if you see someone in...

a nice suit and they sound well-educated, you're more likely to take them seriously, right? As opposed to if someone dresses like I do on a daily basis, right? In a ratty old t-shirt from 10 years ago. You know, that's the thing about AI is it's the proverbial person in the suit because it always gives you a very polished, sophisticated, grammatically correct, well-structured output.

But I've had instances to what you mentioned, where I'll run, I'll ask it to generate, you know, a report or what have you on a company that I know very well, and, you know, be able to then actually benchmark, okay. What is it getting right? What is it getting wrong? What's kind of the error rate? And that's why, like I mentioned, I no longer use LLMs or even largely, you know, things like what are embedded in AlphaSense.

to do research, right? It may be helpful more for finding sources. I think that's one of the really helpful things. I like to call it automating Google, whether it's perplexity, whether it's chat GPT.

you know, being able to query in natural language and find, you know, again, another mental model base rates, right? So you think about like, okay, what is, you know, I was recently researching a British travel company, what are historical statistics on the percentage of people who book travel this way versus that?

Right. And trying to find those statistics yourself on Google, like, you know, all of us as value investors have done that sort of thing. It just takes a lot of time, right. Finding the right source, you know, learning, thinking about how to. translate what you have in your head into a keyword, sorting through a bunch of sources versus using AI, you can very quickly surface high quality sources.

and then review those yourself, right? But the AI is sort of automating that Googling process for you. And to your point about creativity and hallucinations, it might come up with 10 or 50 or even 100 search terms that I might have not thought to make for that same idea. So, no, I think my answer is that I would not use public LLMs for serious research.

a workflow and you were to use a rag type system or you were to be able to use the API and turn the temperature down. That's a different thing. By the way, random side, I don't know why chat GPT, them and I. Claude and all these don't have like a built-in temperature dial that you can use.

Because you can do that via the API. But it's really frustrating to me that sometimes, like you said, sometimes I want a query and I might want the wackiest off the wall answer, you know, think about like, okay, here's an ingredient in my pantry, I want to make something new for dinner.

Right. What are some cool things I could do? You know, right now I've got a lot of basil growing in my garden. It's, you know, I got to use it up before the first freeze, you know, give me lots of ideas for what I can do with basil. Well, their hallucinations are great, right? You need, you need some real creativity, you know, versus.

you want a factual answer to a question, you don't want it to be hallucinating. So that's why I personally tend to use notebook LM or those types of systems where you don't have that risk of hallucination. And then like you said, you also...

I do think it's still very important, at least I find, to review the source material yourself. It's just helpful to, if you think about it, I think of it as like Control-F on steroids. This ability to like, okay, if I want to learn about how has a company's margin evolved, you can get the very quick high-level summary of all the initiatives that have impacted margin, the reasons it's gone up, down, sideways over the past.

two quarters, two years, two decades, whatever timeframe you want to look at. But then it's citing right there, notebook LM, all of those source documents. And you can go see, okay, on this transcript three years ago, here's exactly what they said. And even with a highly reliable tool like notebook. book LM, make sure that it's not misinterpreting, because particularly with spoken material, it can still confuse. If the person didn't speak in a grammatically perfect way, it's not smart.

Right. It's just doing pattern matching. And so it may generate some sort of unusual outputs that you look at yourself. You're like, oh, this is what the person is trying to say. It's very clear to me. But the machine doesn't understand that. Yeah, that's really interesting. It makes a lot of sense. I love that citation feature in Notebook LM because you know exactly what it's talking about. It is helpful for data mining and figuring out where things that might...

have some imperfect recall in your brain. You don't exactly remember where it comes from, and you can very specifically drill it down once everything's sorted nicely in Notebook LM. Right. Sorry, just to answer your other question that you asked about, you know, which model and I think, you know, honestly, all of them are pretty good. It really comes down to personality, you know, and then there's just some sort of workflow things of like, you know, at one point, my wife and I were using.

We had subscriptions to chat team between perplexity, Claude, and Gemini. And at some point, it just becomes too much, right? Because you kind of want your information consolidated in one place. I'm always going to have Gemini because of Notebook LM, but also because I just use Gmail. And so naturally, as part of the Google ecosystem, you have that.

I do have a chat GPT pros inscription. And then there's perplexity where, you know, just a small example. I think a lot of people don't really know how to use chat GPT versus perplexity yet. Where, because perplexity does the real-time web search for those things I was talking about.

statistics, so on, you know, and returning sources, perplexity is often much better for those sorts of things versus if you want to write code, you want to, like we discussed, play devil's advocate, something like that. chat, which is more helpful. So I've been using GPT-5 a lot. I really like it. I do use Gemini for some things.

But I would also say that I'm very, oh, I forgot about Grok. You know, I used to use Brock a lot more and then that's kind of fallen off. But I would say that it's just one of those frog things where it seems like every time, you know, whoever's the latest model seems to be the latest and greatest and you switch over to that.

For sure, you know, very excited to try out Gemini 3 that's supposed to come out in a few weeks. And I'm sure I'll find use cases for that. But I think it just boils down to personality. And, you know, as an example, one thing I seem to use Gemini for a lot. is summarizing my internal research. Writing my investor letters used to take me a huge amount of time because I have all this research, but obviously you're not going to deliver.

your LPs, a 50-page document full of internal notes on every company in your portfolio, right? So taking what I've written internally and just shrinking that into something that I can present to LPs and say, hey, here's how this investment did this quarter.

why we like it, what we don't like. And that's something I use Gemini for extensively because I find that Gemini just seems to be a little better than Chachi because that's the task of writing the summary that kind of meets the criteria that I think should be in there with. less editing needed. Yeah, that's super interesting. I think for the listeners out there, it's important to just take note that it doesn't really matter all that much which AI you use, right? It's all about how...

good you are at structuring your questions and how curious you are and i think that's something that all investors could appreciate that's what we're here for at the end of the day right asking the right questions and looking in the right places One of the use cases I thought was really interesting that you'd pointed out was looking for opportunities. You'd pointed out something about like analog semiconductor industry where like AI will drive these.

third and fourth order impacts like that to me taps into this idea of creativity using AI where it's fairly open-ended and you talked about making connections where they might not have been obvious so like beyond traditional screening Like, how do you use AI to identify these non-obvious investment opportunities? And maybe are there any like major light bulb moments you've had in these kinds of like thematic high level queries?

Yeah, that one's interesting. I mean, so the what you're referencing, that's actually that came from an expert transcript where it was a funny analogy, but the expert basically. That's something to the effect of, well, if your refrigerator is going to be AI powered and order you milk and you run out of milk, it's going to need a sensor to determine when the milk has run out.

And so therefore, that's just an example of as you think about AI translate physical world, you know, inference being not just something we're doing through an LLM, but something that's being done by.

You know, whether it's factory automation, whether it's robots, whether it's, you know, whatever it may be, those are all going to require mechanical components. And so, yeah, looking at the second order impact is interesting. I mean, I would say, you know, I've been a little bit disappointed by that. And I'd certainly love to hear from you. you or from any listeners if I'm missing something. But I have tried. I have tried asking AI those kinds of questions.

And I feel like AI is sort of very good at delivering almost consensus, right? Because it has its training data, sort of the whole corpus of human knowledge. If you're doing web search, it's the stuff that's already out there. And so I know that there are people who are talking about, oh. you know, chat GPT made this scientific discovery or it solved this and diagnosed my illness or things like that. And there may be areas where that's.

possible. But if you think about, you know, sort of like the Peter Thiel model of the world, right, your zero to one versus your one to n. I feel like the AI is very, very good at the one to end where if something is already out there and there's some pattern out there that matches what you're putting into it, it can return that to you in terms of that true, you know, and that's creative in a sense, but that's also.

more just like knowing what's out there in terms of the true zero to one coming up with a new idea. So far, I haven't really found a way to use AI for screening other than, again, that stuff you're talking about where you're looking for a qualitative.

factor that you might previously not have been able to identify, but you kind of have to already have the documents or have to already be asking the right question. So, you know, is there room to do that? I've played around with tools. I know people who've been building that. I know people who've been working on that. I personally haven't.

found a good way to do that yet, except for, like I mentioned in my letter, taking other people's investor letters that are public, putting those into AI through an automated workflow, and having it screen for ideas that match the kinds of businesses. that we want to invest in. But no, in terms of the true creativity, I haven't seen success there quite yet.

Yeah, I don't have good anecdotes myself either. Still pretty early in this journey. And, you know, I think not enough time has really played out to see if anything truly stands out. experiment with it and tinker you'd mentioned reading the nvidia way and i thought it was very interesting i've been very involved in life sciences researching companies making some investments to my detriment thus far in some ways but uh some

Some wins, some losses, more frustrations than anything else. And, you know, I thought it was very interesting at the very end of the book, Jensen is kind of asked, where does this all go next? And he's like, our most profound impact is going to be in. Life sciences, biotech, drug discovery and whatnot. So I've started this big project that I used AI.

to get started in asking creative questions on where I should look and have now done dozens of expert calls, which inevitably will lead me to writing a white paper on this and maybe a couple investment ideas. I'm not sure. they are investment ideas yet, but they're ideas. So that's interesting. And I've learned a lot. I've learned that, you know, Nvidia is way more verticalized in the life science space than they are in any other area. Like BioNemo is literally one of the more profound models.

So they're a model builder. They're like an open AI, but they're also helping companies build better instruments and make their tools better and using a layer of AI. in cell analysis that formerly had not been possible before because gpus are being added to each instrument in certain areas like definitely kind of going into more depth than necessary and something like this but i i think it's interesting it's a it started with a book

It then went to AI and has been in many tangible places in my mind and in my efforts of late. So I think, you know, TBD overall, but I do think that's interesting.

It is. And there's so many, I mean, there's so much of the thing is, you know, like when we say AI, you have to be careful because like I, when I say AI, I usually mean an LLM and that's what most people. But of course, when you think about protein folding or the things Google's working on, there are, of course, many other applications of AI or machine learning outside of LLMs, which I think you're referring to some of them.

Absolutely. But it's actually both is one of the interesting things I've learned because people are building these like multimodal agentic models that are. semantic in nature. So someone who formerly was a bench scientist really good at analyzing cells can now have the power of a statistician in their hands and could get those insights with language rather than numbers.

And so there are all these different areas. There are protein language models, which I never had heard that phrase before, a PLM instead of an LLM. Yeah, I think it's kind of the tip of the iceberg. And I think one of the interesting takeaways is that a lot of people don't really know where it's going. Kind of like investing, but you start getting specific ideas. I wanted you to share one of your stories that I found fascinating.

from the letters which was you know i i love the phrase the first ai powered proxy contest tell us about this because that was that was awesome that was really interesting Yeah, yeah. So that's, that's definitely interesting. So I'll give you the real brief version, which is, you know, we're not professional activists, but we're willing to stand up for our client interests when they be. And at one of our portfolio companies, Astronova, we were not satisfied with how...

Results were trending. In particular, they'd made this disastrous acquisition that had just gone from bad to worse, was losing lots of money, and they didn't seem to be taking real concrete steps to address that. You know, and so we engaged and, you know, we tried to talk to the board and, you know, I won't get into all the details because all that's public. But long story short, our presentation to ISS, which, you know, that deck is very important.

We largely built it using AI. I forget the name of the specific tool, but there was some slide creation tool that we used. I think there's a couple of them we tried to do. a few different ones and settled on one. And then the images, even in the images in the deck were generated by ChatGPT or Gemini and just having high quality images. I mean, just as an example, there was one slide where we talked about the differences between the culture.

business in the US and the culture of this business in Portugal, that they And there was this really funny image we generated of like, you know, these very happy Portuguese people doing business in a much more relaxed way, you know, sitting on the beach and looking at reports as opposed to, you know, people in New York kind of staring at reports at midnight.

in a skyscraper so yeah so that was interesting and you know we got there's people out there who would have charged us 50 100 200 000 to put together that iss deck and you know using ai we were able to do it

you know in a week or two at much lower cost you know basically free other than my own labor and of course even the inputs to that right like when we're gathering all this data being able to use notebook lm to go back like i said 10 years of proxies and like okay here's you know one of the things we mentioned

for example, is what are the STIF targets and how has management performed relative to the board's own targets? And why hasn't a change been made if these targets are consistently being missed? you know, but being able to go back and pull all that data very quickly. So, you know, that's an interesting kind of example of as look, I'm a one man shop, you know, I don't have an army of analysts, you know, we do have, you know, 95 million in AUM, whatever it may be today based on the market.

but we don't have infinite resources. And so being able to use AI kind of as this, I talk about it as force multiplier, right? And I think it's particularly interesting for smaller funds. where if you're trying to be capacity constrained and go after a specific strategy like we are in the microcap space, but then being able to use AI as this kind of what I call this army of infinite interns, where you then...

you know, you are able to level the playing field a little bit versus a $500 million fund or a $5 billion fund that can just throw a bunch of junior analysts at a problem, right? If you want to learn about something, you want to learn about an industry, you tell a junior analyst.

Hey, go out and get me all this information. And of course, you know, a junior analyst isn't always going to get it right. They're not always going to have the judgment or insight. That's where the senior analyst or the PM or the head of research comes in. But I think that's how you can really think about AI is like if I had infinite army.

of reasonably competent interns or first or second year analysts, what would I have them do? And kind of had it do that, right? With that same idea of, of course, you're never going to take, I mean, even if I get an idea from other principal that I really respect, a fund manager who I really respect, I'm not just going to.

blindly buy that idea right i'm not just going to say oh they think the stock is cheap this person tweeted about it they wrote it up on dick they wrote it up in their letter let me go you know put money into it no of course you go do your own research you verify what they said you think about it independently you

etc etc but it's still helpful to have that starting point and that's what ai can really be so yeah so the proxy contest was just an interesting example of the you know and unfortunately for me my business is pretty simple so i don't have the need to do this a lot. But then you think about a lot of businesses and just how much time people spend in meetings or dealing with email or preparing invoices or etc, etc. And the extent to which AI workflow automation can handle a lot of that.

right and take it away from the human plate and allow you to focus on what is another is another aspect so the proxy contest was a really interesting uh yeah i don't know if anyone else has done that yet but it's it's fascinating to think about how that could

make, for example, not that I hopefully have to do this again soon, but how it can make small cap activism easier or cheaper by taking out some of those steps and being able to generate some of that content that you previously would have had to hire a PR firm or someone to do.

yeah i mean let me ask you this would you have done that were it not for ai and what you knew you could do with ai Yeah, I mean, I think you might be giving me a little bit too much credit in terms of planning, you know, and again, not to get too deep. But I think when we started, we didn't necessarily, you know, a lot of proxy contests or a lot of these things get resolved early. And we were originally just trying to have a conversation with the company.

You know, and that's not how it worked out. And we ended up having to kind of take it almost all the way to a boat. So, you know, I don't want I don't want to suck at this foresight that how everything was going to be. But it's certainly I mean, let me put it this way with the resources that we had available to us, I think are. efforts would have been a lot lower quality were it not for the ability to generate a lot of this stuff using ai do a lot of the analysis using ai

And going back to the point about expert calls, you know, one of the things we did that I think a lot of other shareholders really respected, we did almost like, I think, 30 expert calls. I have to look, it was like somewhere between 25 and 35 conversations we had with people in the industry who knew that business and that. industry very well, and we presented a lot of that material.

But just being able to do those calls, again, if I'd been having to make that slideshow myself, right? And of course, I still, it's not like I just looked at the AI slide and said, this is good enough. It's just having that framework to start, right? Having the formatting laid out and the images. And it's like, okay, I'm going to change all the text.

Next, I'm going to make it four bullets instead of three or two bullets instead of three or whatever it may be. But having that starting point was tremendously helpful and allowed us to spend time and resources elsewhere. Yeah, I get the sloppiness of the journey, but I think it's an interesting point to consider that were it not for AI, you running a firm by yourself, you now had a tool at your disposal.

That was pretty valuable as a manager in getting a good outcome. Exactly. Yeah. I mean, as a first-time activist, we got endorsement from ISS for two of our five candidates.

glass lewis endorsed all five which is you know not not very common and i attribute a lot of that to again being able to if it wasn't for we wouldn't have had such high quality material to put out there if it wasn't for ai at least not with our resources or we would have had to show we would have had to decide whether it was worth spending

you know, extra hundreds of thousands of dollars, which, you know, I'm not, I'm not made of money. So that would have been a, you know, quite the decision. So, you know, a lot of AI conversation is very... High level. I like that we've been specific, but one of the things that I really like that you did in your letter was talk about a very specific star goal. And that goal was increasing your watch list to 150 names.

And then inevitably 250 names. So I'm curious where you are in that journey. Are you finding that inachievable goal or is it a moving target? And, you know, how do you feel about the depth?

breadth and quality of increasing the number of names in your coverage area. Absolutely. So maybe just for the benefit of... readers who didn't read that or listeners who didn't read that letter or who haven't read about our strategy in general, we've always tried to build this watch list, which is certainly not unique to us, right? But the idea is that rather than always going and looking for ad hoc ideas,

just having this list of companies that you know well, that you like, that you think have a good future, and then just waiting for whatever combination of circumstances or market sentiment to take the valuation to a level that you're comfortable paying. initiate a position. And so we've always done that.

And I think just the challenge is as a one-man shop, you know, if you're trying to keep up with your existing portfolio, I mean, you've caught me during earnings season, right? I've had, I don't know how many companies that haven't reported the last handful of days. And then you're also trying to handle your trading, you know, like with all the...

carnage in the market today, and you're also trying to handle new ideas, we would put names on our watch list, but then what would happen is There's names that, so for example, like JLL and CBRE, the real estate brokers, we'd research those in quite some depth in 2017, never come back to them.

Right. And so we have them on our watch list and we have a price, but who knows since then how many acquisitions or dispositions they may have made, what may have changed in the competitive landscape. Maybe they are now dominating the industry. Maybe some new competitors have come in. Who knows? And so that name may have been on our watch list. but it wasn't really very useful because it's, you know, the information's way out of date. So I think there's a couple different ways where AI helps.

And our goal is to not only increase the number of companies on our watch list through the research process we've been talking about, but also then the ability of AI to do a very quick... update where if you have your thesis written out, right? So say that you have your research on any company, we'll just go with CBRE or JLL.

Right. And we've identified what we like about the business model, what we don't like about the business model. And then you just put that thesis into, say, a notebook LM, along with the transcripts and the earnings reports and things since that period, the 10K, you know, whatever interesting. there may be. And then you basically ask, what did I get right? What did I get wrong? What has changed? What hasn't? And it allows you to do this sort of qualitative update where

Again, it's not just all about one of the things we found that we would miss on our watch list is because we were only monitoring the price. You know, we'd see the companies whose stocks go down a lot, but a lot of times if a stock goes down a lot, that's because they're, you know, where there's a smoke, there's a fire, right? There is.

there is market sentiment issues, but there's also, hey, there may be this company struggling. What we would often miss is what can sometimes be even more interesting, which is maybe this company's business is getting way better, but the market's just not recognized. right for whatever reason but we just wouldn't see those because again it is very mechanical in terms of we basically only review it if the price went down to a certain level so

I think that AI's ability to do that qualitative work, and again, it's not like you're going to make an investment decision off that, but if I want to know, hey, here's this company I looked at 10 years ago, what's new with this company? Is their business more or less still the same, or has something interesting happened? Again, getting that.

quick report. Here's what was validated. Here's what's invalidated. Also, just by the way, getting that feedback, which is enormously helpful as an investor, because if you think about like Philip Tetlock and super forecasting, right, one of the ways to be a good forecaster is to actually evaluate how did your forecast

do? Why was it right? Why was it wrong? And so going back to that idea of internal data, there's all this research I may have done years ago, and I may have said, okay, I think this company has a great competitive position. And maybe I look at it 10 years later, and that business is completely eroded.

Well, that says that there was something missing about my analysis at that time that it didn't capture. And that tells me something, even if I never invest in that company, that's useful information for the future. So with regard to the specific target... You know, that's absolutely still our goal. I mean, I think, like I mentioned, we've been investing a lot of time in depth versus breadth, right? So I think, you know, as with any KPI.

You don't want to get overly, you know, you think about the mental model of incentives, right? So if I just wanted to put companies on the watch list, what I would do is I would find a bunch of companies that are really easy to research, that are not complicated.

regardless of whether or not they're businesses we like, regardless of whether or not they're, you know, the kinds of businesses that will ever trade at evaluation that we find reasonable. And you just go ahead and run research on them and learn about the business and put it on your watch list. But obviously, you know, there's...

plenty of cases in the past where we put a business on our watch list and eventually decided, you know what, the quality of this is marginal or there was just too much debt or et cetera, et cetera. And so I think. It's almost equally valuable in terms of killing, like I mentioned with that idea earlier, where there was a company we researched and it did end up on our watch list. But we basically decided, look, unless...

they change this or unless they fix this technology platform that they've built through all these acquisitions that hasn't been integrated. We've seen this play out badly too many times. We're not walking into a company that needs three years to fix all its technologies. Right. So we were able to get there in four hours and then that allows us to spend time with something.

I think it's up to every investor, you know, where they choose to reinvest that time and what they find valuable in their own process. But that's kind of what we've been doing. The watch list is absolutely a target, but I'm trying not to be too...

You know, recently, we've been focused a lot more on depth and breadth, right? There may be other market conditions or other times where we're focused more on breadth. But I also don't want to say, okay, well, I'm not going to spend a week on this really interesting business just because I could go spend.

Another four days on four other businesses, but we're less likely to invest in those at this time. Maybe I'll just share a few. Sorry to jump in, Elliot. Did you want to continue? I wanted to ask one more question that went off that, if you don't mind.

Go for it. Yeah. Yeah. So, you know, are you relatedly increasing the number of positions in your portfolio? Because if you increase the top of your funnel, might the bottom of the funnel similarly grow? Yes, absolutely. And that's an evolution we've been on for a while. And so originally, we were very, very concentrated. We're still concentrated.

But we've gone from a target of 10 to 15 stocks to more like 15 to more like now 20 is our target. And of course, that's not a hard and fast number. Maybe sometimes it's 22, maybe sometimes it's 17. But 20 is kind of a good midpoint of the range. That was driven both by...

Again, just some of the idiosyncratic volatility in the world we've experienced, whether it be COVID, inflation, but some of these macro things where you can have a really good microthesis, but then there's a change in the world from a regulatory perspective, a pandemic, whatever it might be that you don't anticipate.

over time, you know, that there's probably more than just five, 10 really good ideas out there. But then to your point, there's also just this mathematical analysis of if I own 15 stocks today. And that's from, you know, if you think about just your selection bias, right? Like how many companies were we really working on before? Was it 50? Was it 100 per year in whatever level of depth?

And so we're picking the 10 or 15 best out of that list. But now if we can work on 150, 200, 300, whatever that number is, these are fake numbers. right just for just for mental math but you're you know and you could get really deep into like okay what's your hit rate or etc etc but if you assume some sort of normal ish distribution of ideas you know if you look at twice as many ideas, even if the incremental idea you look at is half as likely to be

good, like you should come up with ideas that are better than your existing ideas, or at least as good, which to your point would suggest that yes, you can achieve some of those benefits of diversification without giving up the upside because you're actually just finding other ideas that are just as good or better than the ideas you already own. So trying to become more diversified and...

You know, also trying to it just gives you a broader the other secondary benefit is it just gives you a broader insight into again, we all do some of this where, okay, even if we don't invest in a sector, you may read some of the earnings transcripts or some of the research.

of these bellwethers, right? Just to know what's going on with the economy or what's going on with kind of the broader space. And so I think also just covering a broader universe gives you more of those insights naturally that may help with elsewhere in the portfolio.

Yeah, maybe I'll share just a few ways that I found AI helpful. So one is helping you to analyze financial statements. So if I notice that something might be off in... financial statements of a company i can say this doesn't how does this actually work together and often i'll learn some intricate details of accounting that'll explain it and sometimes it'll

also tell me well that's kind of a red flag so if you're if you have kind of specific questions about financial statements gemini for example with deep research can go in and analyze financial statements of a company through time or whatever you're interested in. And I found that quite helpful. You also learn as you do it.

and might uncover some you know red flags that maybe you wouldn't really know whether they're serious or not another thing that i found helpful is just sort of the learning kind of the evergreen component let's say i want to you know learn about how did carl icon invest in netflix back in the day you can basically have uh gemini give you the whole

case study take you back in time tell you where netflix was trading you know kind of set it all up for you the way you you you want to consume that kind of a case study and maybe learn something from that so i think it's great for you know case studies or learning from other investors that for whom there isn't a book out there or you know something existing you can sort of almost create your own personalized a booklet

on any topic that you want. And then one other thing that I found really helpful is I listen to a lot of audiobooks or just audio in general. There's an app called 11 Labs Reader. where you can upload any PDF and it'll do a really good job just reading it to you as an audio. So that's... That's also very helpful just from a process standpoint. If you spend a lot of time in the car or working out or what have you, you can actually have documents read to you quite well.

Why do you like that app specifically, John? Well, I'm familiar with Eleven Labs. They have a really good sound studio, like an audio studio, where you can actually generate... audio and all kinds of different voices i mean it's really advanced ai and so the reader app that they have is is free and i it does the best job from what i found in terms of

you know, just reading it in a very human-like way. Fascinating. Yeah, that's actually, that's something I've been looking at more as well because just not. even from a productivity standpoint, but more from a health and general wellness standpoint. I spend way too much of my life behind a computer screen, which I'm sure that you all do as well. It's a peril of when your job is to...

read documents all day long. But that's one of the ideas that I've been trying to work on a little bit more where it's like, okay, if I have materials to review, but can I use notebook LM to turn it into a podcast and then I can do some yoga or I can.

you know, do a workout or I can just go for a walk, but I can be listening to and consuming that information in kind of a slightly different format. So I haven't heard of 11 labs or tried it out, but I'll have to check that out. Thanks so much for the recommendation. Yeah, and the other thing I'd mention is in Gemini, you can create these gems, right, which are basically like templates that I found very helpful if you do some task repetitively.

within your process you can use gems within gemini to basically create and then over time refine that part of the process that prompt if you will i found it really really helpful Yeah, those are I use a lot of Gemini gems. I think you can do the same thing. My wife's written about it with chat GPT project instructions. And I think that's like a nice halfway between building a full automation.

you know, which takes time and is a little more inflexible versus, you know, at the same time, if you find yourself copying and pasting the same prompt all the time, super helpful to be able to just kind of save it and reuse it. I actually did that for... A couple of things. I mean, one is like my summarizing, like I talked about my investment letters.

You know, summarizing my research for investment letters. A second is even just like when I was putting my website back up, which I did recently, like refactoring the HTML to remove like references to broken plugins and, you know, all things like that. But basically for each web page, there was kind of.

standard processes needed to go through and so just saving a prompt that says hey here's some html do x y and z to it and gemini and then it would so yeah those are those are super helpful yeah and maybe one that i kind of

figured out wasn't really helping a lot, which is, you know, perplexity has this comment browser and it can tie into your Gmail and draft responses to incoming emails. You got to be careful about the security with that. Yeah, exactly. That's one concern. I mean, that's where I'd wish that Chrome and Google would have it because they already have my Gmail. So I'd rather it all stay with Google, you know, because they can steal my information already.

But let's put that to the side. You know, what I found was I'd end up with all these draft emails in my, in my. drafts folder and it would create like additional stress because i'd have to i'd feel like i need to now go through all those drafts and if i don't go through them and a lot of them are for emails that don't really need a response at all

And so it wasn't really the kind of time saver that I was hoping it would be. And actually, Gmail offers something where it's going to, you know, on a case-by-case basis, you can ask it to. draft an email you can kind of it has a chat window where you can tell it what to what to say and it'll say it for you and then you can approve it before you send the email so i found that more helpful actually

So it's really funny. I'm glad I'm not alone. It's funny you mentioned that because one of the most surprising or counterintuitive things that's happened since I adopted AI, and I can't remember if I wrote about this in my letter or just on my website. My wife basically had to stage an intervention for me because.

A lot of people think about, oh, well, okay, if AI can write 80% of code or 70% of code or whatever the fake statistic is, are there going to be 80% less software developers? I've always kind of thought, well, I think you're just going to create five times as much software.

You know, in the same amount of time, we're using the same amount of resources. And there's we could go into like the Berkshire textile mills and sort of the economic rationale for that. But just on a personal level, you know, it's this there was this idea. you know that i had initially where i was like oh well with ai if i get x amount of research done then i can

free up time. But it's almost this paradoxical, like, you know, before, let's say I eat dinner, and then I have an hour or two productive hours left in the day. And I might say, you know what? I'm not really going to be able to make that much progress. I might be able to go through one expert transcript or I might be able to go through, you know, this specific portion of the business, but I'm just going to leave it for tomorrow versus now. It's like you could complete half a recent one.

expected you know in an hour or two and so there's this there's this behavioral or this psychological pressure of like well if i if i just sit down on my computer for another 30 minutes i could like you know hash out this entire issue. So paradoxically, AI was supposed to make me less stressed and save me time. And I think I just reinvested all of that.

into more of a kind of like your email draft. Because now, before you were like, well, I can't answer these emails anyway, so who cares? And now you're like, well, but there's 70 pre-written drafts waiting for me. Yeah, I'm kind of Jared's Paradox. Exactly. yeah yeah definitely well this all of this is evolving so quickly i'm also with you samir on your point of basically having subscribed and tested all of the different llms and then

found that, hey, you don't really need all of them. I'm basically using Gemini now because the MOI email is also on Gmail and so forth. And I guess I trust Google a bit more than I trust.

some of the others so that does come in because it's scary how much data these llms kind of accumulate over time on us and especially if you give them hooks into into different systems it's really a ton so yeah i guess that's that's that we've learned a lot any any parting words or anything else elliot you want to mention

Well, I just want to pat ourselves on the back because we had an investment conversation about AI without asking if AI is a bubble or we should be buying NVIDIA stock. i don't know that seems unique in the podcasting landscape nowadays i think maybe one closing theme that i was looking to ask you about samir though is getting tangible with ai investments and it doesn't have to be

a very extended piece, but you had this firm rule, do not bet against AI. Are you making any specific changes to your investing? Are you betting on AI in any specific way other than through process development? Anything on that might be an interesting way to close this investing podcast on AI. Yeah, that's the million dollar question. And I'm, of course, in this very unique position where I'm very bullish on AI as a tool.

Right. But then you look at the markets and you look at this bifurcation or divergence between it seems like all of our value stocks or anything that's not related to AI. No one cares about or is ever going to care about anything tangentially related to AI.

trading at a massive valuation i mean i think you know i think if you think about it through a value investing framework right i think you're always looking for sort of these hidden beneficiaries for the second order impacts you know to your point the point we talked about earlier about the refrigerator and the milk sensor

So, you know, just a couple ideas. I mean, one is, you know, thinking about, okay, what does inference enable, right? And ironically, you know, if these companies don't get a return on CapEx, that might actually... better because then if they put all this money into capex they have to use it you know inference gets cheaper and cheaper doesn't really matter if the vendors are making money right but how can what what businesses benefit using it

Whether it's businesses that kind of have proprietary data, like we've just been talking about, where you're able to then go and do things, whether it's being able to assist with when you just think about like... you're building a website or you're putting up an eBay listing and all of a sudden now AI takes a lot of friction out of that. So I think those are interesting to work through. I haven't really spent a lot of time.

the NVIDIAs and so forth of the world, just because, again, as a value investor, I think it's a little hard. With a small microcap mandate, I certainly don't think I necessarily have an edge versus the people who are spending a lot of time on that, a lot smarter than me, actually. resources than me. So I'm trying to focus on, hey, what are these second and third order beneficiaries that may not be up now?

But as AI continues to propagate throughout the world over the next three to five years, that's going to be a tailwind for other parts of the economy as well, even though right now it's mostly concentrated in these large AI vendors specifically. Great. Well, I think we'll leave it there. Samir, Elliot, thank you so much. We learned a lot today. I hope the listeners enjoyed it as well.

Really appreciate you taking the time, Samir, to come on and share what you learned with all of us. It's been a real pleasure. I appreciate the opportunity and also the recommendations. I mean, I'm always interested in talking to other people. You know, it's not like I haven't been on good ideas. So I loved hearing how you're using AI slightly different from me. And hopefully if any listeners have good ideas, they'll share those with us. Take care, everybody. Talk soon.

Thank you for listening to This Week in Intelligent Investing, brought to you exclusively by MOI Global, the research-driven membership organization.