%20(150%20x%20150%20px)%20(5).png)
Invest with AI
Brett Caughran and Khe Hy lead a deep dive into AI for investing, joined by guests at the cutting edge of the field. Our goal is to be your Sherpa through a rapidly changing landscape by distilling what's working, what isn't working, and how you can leverage AI in your own process. Follow along as we tackle AI's biggest challenges and opportunities, one episode at a time.
Invest with AI
Canary CEO/Ex-Tiger Global PM: How Bad AI Is Leaving Alpha on the Table
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
Joe O'Donnell ran the short book at Tiger Global for almost a decade before leaving to build Canary, an AI intelligence platform now used by some of the largest hedge funds and mutual funds in the world.
His take: because so many investors are misusing AI, there’s more alpha available today than there has been in a long time. He, Brett, and Khe get into the 20+ page investment reports his agents write on their own, why summarizing an earnings call is "lossy" in ways that quietly inflate your conviction, and the one line he uses to spot a finance-AI company that's already lost.
If you're deploying capital with AI in the loop, this episode could help you avoid expensive mistakes.
Timestamps:
[00:00] Intro
[00:38] — Running Tiger Global's Short Book to Founding Canary
[01:15] — Why 2023 Was Too Early for Institutional AI
[03:28] — AI Is Only as Good as the People Who Build It
[04:57] — Buffett & Druckenmiller vs. 1,000 Junior Analysts
[06:25] — The Layer Cake: How Canary Is Actually Built
[10:40] — What a Model Upgrade Actually Changes
[16:21] — Is AI Judgment Real Yet?
[17:16] — The 20-Page Investment Report an Agent Writes Alone
[19:22] — Why AI Summaries Are "Lossy" in Dangerous Ways
[24:46] — Can You Just Build Canary With Claude Code?
[29:28] — Super Analyst: A Junior Analyst Across 4,000 Names
[32:32] — What Fine-Tuning a Model Actually Takes
[36:07] — Why "AI for Financial Services" Already Lost
[39:08] — Untraining AI: The Excel Problem
[46:53] — Your Proprietary Data, Headless Canary, and MCP
[51:07] — Advice for Funds Starting From Zero
[54:52] — Why There's More Alpha Available Than Ever
-----------------------------------------------
Want to actually build these workflows yourself?
The AI Accelerator is Fundamental Edge's 6-month cohort for investors who want repeatable AI workflows. Learn More below:
https://www.fundamentedge.com/ai-accelerator
Watch the full podcast series on our site: https://www.fundamentedge.com/invest-with-ai
Follow Invest with AI on:
Spotify: https://open.spotify.com/show/033xcEEovVViS7hIYwNuGZ
Apple Podcasts: https://podcasts.apple.com/us/podcast/invest-with-ai/id1896918892
But we have explicitly launched within the last two months uh AI idea generation agents that deliver fully baked investment reports. There's no such thing as AI for financial services. To deliver institutional great insights is quite difficult.
SPEAKER_02Great. Welcome to the next episode of Invest with AI, where we explore the intersection of fundamental investing and artificial intelligence. I'm really excited to have Joe O'Donnell here today from Canary Data. Joe, could you start with just a brief overview of your background?
SPEAKER_01Yeah, sure. Happy to do it to that. And uh thanks for having me. Um so the quick background is I used to be a professional investor. Um the majority of my career in that world, I spent uh running the short portfolio at Tiger Global. Um I did that for almost 10 years and left uh at the beginning of 2021, hung out for a while, and then started working on Canary uh in 2023. And uh we've been at it ever since. Um and Canary basically builds uh the very, very, very brief summary, which I imagine we'll get into in more detail. Um, Canary basically builds intelligence for professional public markets investors. So our customers include many of the largest uh hedge funds and mutual funds in the world.
SPEAKER_02What's um, you know, even back in 23, you were much more forward-thinking than I. I played with these tools in 23, and I'm like, these just like I couldn't really do anything, you know, institutionally useful for the hedge fund research process at that moment in time. What was sort of your your conversion moment where you where you swallowed the AI pill back at that point?
SPEAKER_01Yeah, I mean, I wish I wish I could uh take credit for being extremely forward-thinking and thoughtful about it. I think you know the problem statement that we started with at Canary was basically that there's an enormous set of things that should exist for professional public markets investors that don't. And you know, the the early uh sprouts of AI, you know, GPT2, et cetera, I think provided a roadmap for what might eventually be possible to do. But at least at that time, a lot of the things, practically speaking, were just not institutional grade. And so we found ourselves doing to build things like data products or intelligence on top of those data products was more of a hybrid model. We used AI to the extent of its capabilities, but we were very cautious and we continue to be cautious to not overuse AI. Um, I think it can be uh an exciting crutch. Um, but you know, uh to deliver institutional great insights, it's quite difficult. And back then it was very difficult. And so I think, you know, the the cocktail of sort of more deterministic code and just manual processes versus AI was, I think, heavily weighted towards the former in those days. Um, but we kind of saw where everything was going. And so um, as AI has improved and continue to improve, we've shifted more of that mix towards AI, and I think you know that that's likely to progress uh hopefully significantly more in the future.
SPEAKER_02That makes sense. And um, you know, that we we can count on a hand or two um you know former hedge fund investors who've actually moved to build product at firms like Canary. Uh and one of the sort of the conversation points that comes up a lot is the the value of domain knowledge in building AI product. Can you talk to us about that concept? Like what elements of building Canary um have required that domain experience from from your time sitting on the desk?
SPEAKER_01Yeah, I mean I'll make a I'll make a very uh bold claim to start with, which is that um AI in general, I would say for this particular use case is only as good as the people who build the AI and train the models and things like that. You know, nobody wants the average invest public markets investor's opinion about anything. You know, mathematically speaking, the average public markets investor destroys value. And I would say an even stronger form version of that might be the 75th percentile um public markets investor, at least maybe conservatively doesn't add that much value. Everyone's in search of the last few percentile. I'd sort of equate it um to uh to a golf tournament. You know, there's every every professional golfer, you know, PJ Two order takes 250, whatever, 300 shots over the course of four rounds. And the difference between uh you know the person who collects all the gold and the person who nobody's heard of is a few shots. And um, but that ends up being all the value. And I think that's that's certainly the case. There's just the power law distribution with respect to public markets investing. And so I'd actually say the stronger form version, it's a necessary requirement to build good AI for public markets investors to have actually done the job at an extremely high level for a long period of time, because the simple problem statement of what do you actually build and how do you actually know what a good result is can only be determined by people who actually know what the correct answers are to those things. Um, and so yeah, I think another, another way, another thought experiment, um, at least that that that I thought of this morning, uh, which I think is kind of interesting, would be, you know, if you were if one were to compare, um, you know, put on uh have a contest, you know, an investment contest. On one side, you put whatever Warren Buffett or Stanley Druckenmiller just by themselves with, I don't know, Google, something like that, um, and the SEC and something like that, um, against a thousand junior investment makers, and they get to do whatever they wanted. Um my strong suspicion would be that Buffett and Druckenmiller would win. And um and that's it's all around judgment, how to build a portfolio. And you know, AI tools can be can be useful in a lot of different ways, but um, the judgment layer is ultimately what determines whether investors have good results or they don't. And so, in order to build um the right kind of tools, the judgment layer has to be a vital ingredient in all of it, and it can't be built without having that level of expertise, in my opinion. And you know, there's a the you know, I think going back to what you're saying about how there's only a couple companies um in our industry that are designed that way, I think honestly there's an adverse selection bias to it. You know, the world's best investors are usually busy being the world's best investors. They're not they're not foolish enough to start a startup. So um I think it's a it's an odd group that uh I guess we're we're a part of on that regard.
SPEAKER_00Joe, I'd love, could you talk a bit about for uh non-technical folks, the architecture of Canary, like the data where the LLM sit in, the deterministic piece, the judgment piece. I'd love to hear that from you.
SPEAKER_01Yeah, I think we think about hundreds of microservices that interconnect in in various ways. I think there's there's different, I mean, there's there's there's a few different orthogonal ways that I think about at least, but one is just um more obvious, which is we build data sets that don't exist. So just starting from that. I think you know the investment world, um I would say if that if there's still a data advantage that exists in the investment world um across the mainstream data sets, it's limited and declining. Um but there's a lot of uh niche data sets that don't exist and by virtue of limited distribution can still maintain enduring value for investors for at least as long as they have limited distribution. And so we build as many of those as we can think of. And some of those are horizontal in nature and apply to all industries, and some are vertical specific. And a lot of them tend to be needle in a haystack type data sets where we might look at you know millions of data points and find three cool things per year. And that's wildly inefficient for an investment firm to do, or for anyone trying to build a data business around that particular thing, but it's extremely inefficient for us to do, especially at scale. And so that's one thing we do. Uh the next thing that we do is we build intelligence on top of those data sets and the more mainstream data sets. Um, basically, so the more basic way to think about would be given data point X, what would a talented investor think about it? I think that context layer is incredibly important and non-intuitive in a lot of different ways. And so that applies to small categories, but it also applies to big categories like uh an annual report. You know, there's um the sort of the old school way to look at an annual report, we just look at the annual report. Then there was, you know, blacklining and stuff like that. And then there's um today, you know, there's Claude, tell me all the interesting stuff in these two reports, make no mistakes or whatever, something more elegant than that, obviously, but something along those lines. Um and then there, I think there are some people that are trying to build um models around that. But getting that right at the level of a world-class investor is super important. And there's a lot of big um sort of more mainstream data sets where we build intelligence layers on top of that. And and only then, after those first two layers are built, can you build useful services and tools that sit on top of it? And so what we do on top of that is sometimes we'll build vertical specific things. So, for example, we have an AI forensic accountant that can unpack and does unpack uh across 10,000 companies or so all of the various uh accounting manipulations that certain companies engage in and provides a clean PL versus the company's reported results versus consensus, et cetera, stuff like that. So these sort of specialized roles, um, we build AI that just does those roles at the level of a world-class practitioner. And then we also build increasingly just investment analysts that take all that information in and help people do research, help people generate investment ideas in um increasingly specialized ways. And so, you know, it's sort of a layer cake where all of these things are necessary to generate results that are actually useful for investors in terms of bending the curve on their results. I think one thing um on the other side, one thing that we shy away from is doing things that we think are just going to be that either already are or are very likely to be subsumed by foundation models. Um, you know, I think um we're we're a startup, it's it would be foolish for us to try to compete with Claude, for example, uh, on anything that Claude is likely to do very well. Uh so what I would put in that category would be everything, you know, broadly speaking, productivity tools. So things that are not bending the curve on revenue, but they're certainly bending the curve on cost or you know, human cost or just actual cost, um, you know, build me an Excel model with one click or something like that. Um, Claude is wonderful at that with no plugins at all. And fast forward 12 months, it will be even better. Uh and 12 months from there, I I think it will be perfect. Maybe, maybe even sooner.
SPEAKER_00Could I uh the layer cake analogy is a powerful one? When you go from like an Opus 4.6 to an Opus 4.8, like which parts of the layer cake become more powerful? Like how has how has your product changed as the models have like step functioned?
SPEAKER_01It's so um it's a it's a sensor, I think, in some ways, and it requires deep analysis and QA at every single layer of all of those things, which is uh an unfortunate reality of what we do. Um, but what I'd say is that the scope of what's possible in AI increases, and sometimes it's a step function large increase. I would say the first time we really saw that was way back uh for with Gemini 2.5, um, where we started seeing that that level of functionality where we could build sort of more performant, more horizontal um investment tools. And um it really depends. So it's model specific, and we don't um often use off-the-shelf models for the stuff that we expect world-class results from, because as I was saying before, they just don't have the ability to deliver that last few percentile of results without a lot of fine-tuning. And so um, we're finding just that fine-tuned models that we build ourselves, not because of anything related to costs, which I think is a big topic these days, um, but but for performance purposes. You know, our goal is to deliver the absolute best results. Um, and the costs are just what they are. And so um, you know, there's no uh trade-off that we are that we think we can participate in with respect to that. And so um sometimes we'll a new model will come out that a lot of people will be excited about in different industries, and maybe sometimes in financial services as well, that where we'll see literally no improvement at all, and sometimes a degradation, um, and other times it will be different. But it's really nuanced and um and I would say increasingly um only applies to certain things that we do.
SPEAKER_02Joe, one of the um one of the things I I see a lot of uh AI finance firms run into is the is a sort of um dichotomy of uh of what to build and how to build it. And the how to build it is sort of getting um a little bit more it's not not simple, but more simple than it was from an infrastructure perspective. I'm curious to understand how you think about the the how to build, and now you have that advantage coming from the seat. How do you map out products that you think will be institutionally useful, particularly when you you know public equities is a really sort of combinatorial complex practice. What one investor does from an investment process can vary quite differently from from uh from from another.
SPEAKER_01Yeah, I mean, I think it's um we basically try to meet investors wherever they are. So we're increasingly building flexibility in our platform. There's a headless version of Canary that uh we have an API, we have an MCP, for example, where if people want to integrate what we do um into their own platforms and their own processes, we're happy to facilitate that. Um and we explicitly don't track what any of our users do at all. But um, you know, from conversations with them, I think some of them will take certain parts of our product because they value those things and use them. And some people will take those exact same parts and use them in a different way. And some people will take just the data sets and that's all they want from us. Sometimes they want the intelligence layer on top of that. Sometimes they want the entire platform. And um we're happy to be flexible in terms of how we deploy that and/or how people use us. And I mean, every every uh company, I presume, could could say this to a certain degree, but if we help people make a few better investment decisions per year, it's pretty easy to justify um having a canary subscription. Um but the other side of it, I would say, you know, to maybe answer your question in a different way, is um one is just the obvious thing, which is just I've faced these problems a lot, and and my teammates as well have faced these sorts of problems a lot in their investment careers, and we know the landscape of what exists, and so we're just basically trying to fill in the blanks. Um, and there's a there's an enormous amount of work to do there, an enormous, enormous amount. I think you know, we have a lot of great engineers and we build product and ship really fast, but the degree that to which our backlog builds uh gr is is is still greater than the rate at which we ship. Um, and so there's a lot of things, and sometimes it's a just like I was saying before, sometimes it's just a very specific data set around something that maybe two years ago would have been totally inefficient to ever try to productize. Um and sometimes it'll be um you know sort of a uh um a specific tool around um something like forensic accounting, like I mentioned before. Um and then um from there, I mean, I think there's trying to think of the best way to put this, but there's a lot of stuff that sits outside of the scope of what investors demand. You know, we think we can provide things that you know, just sort of from our own brainstorming sessions are gonna be valuable to people, whether people have asked for them or not. And then we'll put them out in the world and um see if people like them or not. And so far we've been pretty successful at doing that. I think there's there's a lot of different failure points uh that exist in the public markets investment process that we're trying to solve. Um and there's an interesting just starting point from the problem statement of most public markets investors do not succeed at their job of generating alpha. Um and so uh uh try to help people mimic their existing investment processes and make it more efficient. I don't think we'd be doing our job at the level that that we we can.
SPEAKER_02And what one of the you know, one of the concepts I've been pulling the thread on is this concept of AI judgment. And the sort of early consensus was that large language models are a you know stateless uh next token generation machines designed to produce fluent sounding pros. They have no native judgment, but there's been some movement in the evaluation of you know the human plus machine, the Phil Tetlock approach, actually using AI to actually apply better judgment. And you've talked a little bit about using the platform to make better decisions. Can you talk to us about you know using a tool like a canary to actually make better decisions?
SPEAKER_01Yeah, I think so. We're starting to see early rewards on this. I I think it's hard to know exactly what the uh efficient frontier will will look like on that. But we explic, I probably should have mentioned this before, but we have explicitly launched within the last two months uh AI idea generation agents that deliver fully baked investment reports. You know, their output is, and I think my the analysts that used to work for me at Tiger Global were very, very talented. But the output that those analysts generate is on par with the output that my very talented analysts would generate. Um we've only launched a few of them, and you know, the the future is uncertain in terms of actually how uh how performant they will be at, you know, what we'll see how their portfolio does over time. But the output there is a fully baked 20, 25-page investment report backed by primary research, including expert calls and things like that. Um and at least um ex ante, what I'd say is the work product is excellent. And so perhaps that's the early signs of judgment being something that is within the capability set of today's AI models. I think you know, in order to get there though, as I mentioned before, there's a lot of work that needs to be done. You can't just simply if we were just starting from scratch with um no infrastructure built up, no proprietary data sets built up, no intelligence layer built on all those other data sets, I think it would be very unlikely that we could build AI analysts that were actually good at doing analysis. Um but because we've made those investments and because of the industry expertise that I have and other people in our in our on our team that have, um, we have been able to build things that at least look, act, and sound like really good investors. And um I think we've started with um sectors and investment styles that we know very well. Um and so perhaps uh that will be an enduring and scalable advantage across all sectors and all investment styles. I guess time will tell and we'll we'll do our best to build products that that do those sorts of things. Um but I think going back to your your original statement around um judgment, what I'd say is that without all those things built up, um AI judgment is a very lossy thing at every layer, and it's it's lossy in pernicious ways. I think say something that I think most people do these days would be something like you know, summarize an earnings call for me or something like that. Um if if one were to compare the actual earnings call to the AI summary of that earnings call, or if it or an A-B test of take my best analyst, have them look at the earnings call, summarize it, and then compare it with any model or um uh you know system of experts or whatever, they're not equivalent um off the shelf. And I question what's actually achieved by doing something like that, if it's lossy. You know, I think um, and that's just one example, there are many other ones, but you know, at least in my own investment career, any of the investors that I know, no one ever looked back on a year that didn't go as planned and said, you know, if only we had read more earnings calls or something like that, or if only we had looked at more ideas. It actually almost always was much more simple than that. The year didn't go well because uh a few times a year we had a big decision to make and we made the wrong decision. And um probably that one another reason why we could steer away from productivity-based things because those things um I think including today's models, all the best ones, um they're lossy. And what would I tell one of my analysts um from an efficiency perspective about take the example of earnings calls again? Read the earnings call. Read the transcript, listen to the earnings call. Um, that is the least lossy thing. And uh AI sounds super smart, and it is increasingly uh smart, but um, it's still lossy and um in many things that people use it for, it sounds like it has good judgment and doesn't, um, or it's just has different judgment. And um those things can be dangerous. Helps people helps investors build conviction when they actually should have lower conviction if they're using that for um for for their investment research.
SPEAKER_00Is it fair to say that there's kind of like a hierarchy of judgment? So, for example, you know, one judgment could be like what data sets to use, and another set like layer of judgment could be should we Even be in the sector. And like it, and then the lossiness angle is kind of a new one that I hadn't really kind of put into the like the layer kick of judgment. But maybe pulling on that thread of you know earnings calls and the lossiness, like where is the judgment strongest? You you gave an example where it might be too lossy, but where where does it work well, the judgment layer?
SPEAKER_01Yeah, I think uh specifically for AI, I would say that I mean there's there's tons of utility in generalized investment research for AI. So I don't want to say that it's um it's it's very clearly very useful. But I would say that the the overwhelmingly best use case for off-the-shelf AI tools is um information gathering purposes. Um couple orders of magnitude better version of vertical search. Um and there's tons of value that can be generated from that, you know, just finding things you want to know about have otherwise found and then you know doing deeper research as a talented human or whatever from there. Um, you know, that kind of stuff I think is is is very, very useful, um, especially across enormous and increasingly large data sets that exist in the world for investors. Um for judgment purposes, you know, what to work on. Yeah, I think I think for sure, I mean we've built um to pat ourselves on the back, we've built some very useful idea generation tools for that that use AI to help people discover new investment themes or things that align with their thinking on specific topics that incorporate our proprietary data sets and have more generalized um search and judgment tools to help people figure out what to work on. Um and then we're you know, as I mentioned just now, we're building basically the stronger form version of that right now. Well, no one, you know, for I take take idea generation. Nobody really wants a list of you know 14 companies that meet their investment criteria. Um what they really want is of the 14, which ones are good investments? That's the ultimate goal. And so that's the purpose that I would say our idea generation agents are effectively the final boss version of investment idea screening. Instead of having 14 companies on a list, you get somewhere between, whatever I call it, zero and three finished investment reports done at the capability level that one would expect from a good analyst. Um that's the right way to screen. And um and so yeah, I think I think it it really depends on um the use case. AI for judgment-related purposes, um I would say to a large extent is still an unsolved problem.
SPEAKER_02Joe, I'm with you on the the chatbots being not that helpful across the institutional stack and primarily good for information gathering. But I think there is this belief now that with clawed code that funds can build a whole new stack of internal tooling that they can they can deploy clawed code to build a live earnings preview or an idea generation agent based on their own historical trades and idea uh ideas and thesis work, et cetera. Um I'm still sort of I still wonder whether that's a prototype or whether that's production grade to actually deploy institutional capital. Could could you talk a little bit about that movement, where you think that could work, and where you think that approach may have issues?
SPEAKER_01Yeah, I mean, I think you know there's nothing against that again against the laws of physics for an investment organization to build canary inside of their own organization. I'm sure you know most investment organizations have a resource set and a team of uh extremely talented people that could go reconstruct a large percentage or all of what we do. Um the level of effort required to do that would be extremely high. And I suspect that some people will decide to do that and other and others will not. Um I think what uh what we found at least is that in order to say, whatever, our our our firm's investment process is X, in order to codify that into software, it's non-trivial. I can't everyone kind of knows what it is when they when they when they think about it or when they look at an individual investment, but but then when you actually try to turn it into some sort of deterministic logic that sits in a software program, um, it's not so easy. You know, people plug in a set of things that they think they like, and the output from whatever the take your idea generation example um is nothing like what they actually think they what they actually do like. Um and then there's this iteration process, you know, sort of like um automated driving where you can get you can, you know, whatever. I maybe this is harder than understanding it, but uh you get sort of 90% automated in a really easy way, and then they'll the next 10% end up taking five times as long as the first bit. And yeah, some people will make that investment and maybe they'll build great products and that that's great. Um I think uh in order to get to that level, it still requires a lot infrastructure that I mentioned too. So do you have the right data? Um, which requires building integrations with who knows how many data providers, but that by itself is not sufficient. Well, then what is what do each of those individual data sets actually mean? And every single one of those needs to be fine-tuned, every single one, at least as far as I can tell. Um, there's nothing that's sort of just obvious. Um, you plug in data set X, and then a foundation model just intuitively, right now, understands it at the level of a really good investor. Um, and in order to do that next layer, well, you got to get your really good investors that should be working on investments to go train all those data sets. And that's an ongoing process. Um, and then you can build the emergent sort of, you know, what I maybe I would call microservices on top of all that stuff. And then you can build that emergent layer of investment analysts. Um it can be done for sure, but it requires um, well, I can just speak for ourselves, it requires years of investment um and many thousands of hours of fine-tuning and data collection and cleansing and aggregation and stuff like that. Um and I think every firm has to decide basically what the demarcation line is between what they want to do themselves and what they want to outsource to someone else. Um and there's yeah, every different answer.
SPEAKER_02And and how's that work today? Because, you know, um where you may get pushback is the is the argument that while while correct that most public investors don't have alpha, no public investor believes they don't have alpha, right? And so I think a lot of the mindset with agents is I want to clone my process. I want to have a digital, I want to have a digital analyst that acts and behaves like an like a senior analyst or a junior analyst on my team. And and to that degree, agents have been interesting as a flexible substrate to be able to actually train that that analyst. How do you work with funds? Like if I gave you an analyst training manual and said, Joe, make this make this agent behave like a junior digital analyst, how far away from that reality are we today?
SPEAKER_01What I would say is for our products, um, you know, we have a product that basically does the job of a junior analyst. Um everything post-idea generation, get up to speed on a company, understand the key investor debates, what are all the key questions, how to do the research, automate all the research and generate answers. That's called superanalysts. So that was our that was our name first before Alpha Sense um independently came up with the exact same name for a product uh a few weeks later. But um the uh we have a product that does that effectively across all listed companies that refreshes every day and does all the research that one would expect a talented um investment analyst to do. It integrates with our own transaction data. For example, we have our proprietary transaction data set that we built um from scratch. Um it integrates with web traffic data, app traffic data, point of sale data, foot traffic data, Google Trends. It can build you a scraper, it can do an expert call. And it just does this autonomously across um right now, I think close to 4,000 companies and refreshes every day. And our customers who use that product wake up with an email if there's new research that uh to be conducted, wake up with an email with all that research already done. Um, and so you know, that's um we don't ask questions. We don't say, What kind of research do you want me to do? We just do the research and we put we deliver the final result, and you know, customers who don't care about it can discard it or they can take some insights or they can disagree with the analysts, just like a normal investment firm would do. Um, and that's something that we already have have built today. I think um as you as you move up the stack, uh the it becomes increasingly more difficult. Yeah, the holy grail is idea generation and maybe after that portfolio construction. Um and as I said, yeah, that's sort of uh a problem we're working on, and and uh we've started to hopefully solve it, at least from the idea generation level. But um our and also worth mentioning on that same point, our idea generation analysts already do incorporate firm zone preferences into that thing. So the very first thing that happens when someone onboards to one of our idea generation agents is they get a questionnaire. You know, what kind of things do you like? What don't you like? If you want, you can send me some research that you've done in the past, and then we'll construct uh an idea generation funnel around those preferences as part of the solution, and then our own models as another part of the solution. Uh, and then similarly, there's check-in points throughout the investment memo generation process where a user can say, follow up on this point, or I think you might be wrong about this, et cetera. And our models will take that into account. But they don't only do that, they take that into account, they also have their own ideas and combine the two and do further research. And so um that's you know, I I I have no evidence whatsoever to suggest that's the best model. But all I can say is that so far our um our customers like it.
SPEAKER_00Joe, could you talk a bit? Uh you mentioned fine-tuning a model. And again, for like a non-technical lay person who knows about AI, what is what does that process actually look like for you guys?
SPEAKER_01Well, it's horrific. So you know, AI sounds really, really uh obviously is neat, but the actual process of of getting an AI model to operate in your image is is just horrifically bad. It requires um a ton of manual effort. A ton. You know, I mean to take um, you know, just to go back to the disclosure statement example, um, in order to get a model that can parse through a disclosure statement at the level of a very good analyst, it requires very good analysts going through whatever, annual reports manually and saying, here's what I would say, and then they look at what the AI says. And um, you know, through fine-tuning, reward functions, etc., all that kind of stuff, over time, you hopefully get uh an equivalence. But the only way to do that is by uh a huge amount. I cannot understate how much manual effort to actually make that work. And one of the reasons for that is is somewhat obvious, but another reason which is not is is the is the thing I was talking about, uh the long tail of results. I think um the rarer there there's this there's this um relatively large Venn diagram overlap between um very interesting insights and things that only happen very rarely. Um and so you that needs to all be codified. You need to have a sufficiently large sample set that you've actually fine-tuned to actually allow the AI models to understand when they see something like that and what it means. Um and um it just requires an enormous amount of management. And that's just that's just enclosures. You know, that's there's there's there's hundreds of those, literally hundreds of those examples uh that are required to get performant fine-tuned models, uh performant defined as what a good investor would think about this particular particular thing. Um it needs to be equivalent. My view is that is let's let's say um I don't want to pick on anyone in particular, but um let's say you could get a result that was 95% as good as a good analyst for task X. Some some tasks, by the way, that's fine. But um tasks that are centered around making the right decision, that is not fine. Um and if you if it's not possible to get to 100% equivalence, then the default, what I would at least recommend for professional, for public markets investors, the default setting for that would be do it yourself. Do it the old do it the old-fashioned way. Because as I was saying before, AI is a the the date one of the dangers of overusing AI is higher conviction when you didn't deserve it. And you never know what the missing percent is or how big the percent is or what what even category it's in, um, or whether it's a false positive or false negative without doing the A-B testing. And the only way you can't you can avoid doing the A B testing is if you have AI models that um you rely on, either that you built yourself or that you buy from someone, hopefully like us, um, that where you're convinced it is doing the job at the equivalent of you or your investment team.
SPEAKER_00And is it fair to say so someone like an anthropic is hiring you know a lot of former bankers, former PE folks, investors, and they're doing a kind of a mass, like a production line version of what you described. And then Canary's just you know, alpha in this situation is like the people doing those evaluations and the fine-tuning are just more in line with a hedge fund analyst, like if they're more specific. Is that a fair analogy?
SPEAKER_01I think that's part of it. It's it's this multiple-layered thing. So um I think the you could throw infinity investment bankers at a problem that applies to public markets investors, and you would not get to the correct level of of accuracy. And I also think there just tends to be adverse selection. Going back to what I was just saying about what it actually empirically means to go train a model. Um, this isn't a job, you know, no one's leaving to invest in banking. I suspect at least, maybe there's a counterexample, but no one's you know, the head of TMT banking at Goldman Sachs isn't leaving to go train models for anthropic. There might be a former Goldman Sachs investment banker that is training models for anthropic. My suspicion is it's not their best bankers. Um, that may be maybe wrong, but that'll be my guess. And so uh, and I think for certain industries, that's fine. You know, if they're if they're more formulaic uh industries or lower leverage industries, totally fine. You don't need um the absolute top people doing it to get to results that um can be impactful at at through all stacks of all parts of the organization. I do not think that's the case for public markets investing. And um, hiring this is an industry where if one is a very talented investor, uh it can be very lucrative. And um the I guess I've seen some very large salaries thrown to uh our compensation packages thrown to some of the AI engineers. And so perhaps one day the those companies will do the same for uh to train for public markets investing, but I haven't seen it yet, so we'll see. Um and uh but I think until that that happens, it's unlikely that um for this particular use case that will be satisfied. I'd say more generally speaking, whenever I see a firm that says we do AI for our financial services, um to me that's that's they've already lost. I don't that that doesn't mean it. There's so many different things that that could possibly mean that um, and they're all very specific with lots of nuance. Um there's no such thing as AI for financial services. You know, we build specifically for public markets investors, and and I would say more broadly the ecosystem around public markets investors. But I think we you've actually started to see this without naming names, but you've seen um certain other companies specialize in investment banking, for example, which makes sense. Uh it's it's um building AI for investment banking seems to me to be a really interesting business model. But um building AI for financial services means um you probably don't understand the problem statement.
SPEAKER_02I did it, I did a um uh test on this exact point, Joe, with AI and Excel. And um AIXL will build a great investment banking model and and your balance sheet will always balance. But almost none of my models do I actually forecast out the balance sheet to balance the balance the balance sheet in the forecast period. So it's it's funny. You actually have to like uh I totally agree with you, you have to untrain, you have to untrain the the AI to actually make it approximate what a public public market investor does. Um, bringing me to the question on Excel. I mean, uh to your point, there hasn't been a great software stack for public market investors because of this sort of combinatorial complexity of what we do. We've sort of been stuck with a Bloomberg Launchpad and Microsoft Excel to, you know, to be the workbench, the analytical workbench. How do you see uh Excel shifting uh both in terms of ingress-egress with AI tools? Um, but how do you think the operating system of the public market investor will shift in the coming months and years?
SPEAKER_01Um It's an interesting question. I I suspect it will be different for different investors based on their strategy, based on their own preferences. But what I'd say one thing that seems obvious to me is that a lot of those things will just be abstracted away. I mean, what do people use Excel for? Um, well, one is building a financial model and forecasting key KPIs and stuff like that. Some of it is more bespoke analyses. I would say in the category of bespoke analyses, a lot of that stuff can be abstracted away or already is to a certain extent. I mean, you know, our super analyst product, for example, does a lot of work that would have traditionally been done in Excel without opening Excel. And the answers to those questions are just emailed to our users and available in our web app and our MCP. Um, and so there's just no need to get in, you know, open anything new to to get to the um the same results. I think um you know, build build me a model, that kind of stuff. Um I I think if it is very likely that some version of people interacting with Excel for that explicit purpose of you know forecasting EPS or whatever. Um I don't see uh it doesn't seem like there's a real problem with doing that in Excel. Um and I don't think it again depends on the model. If if um if someone has a coverage universe of 500 companies, for example, it's different than a concentrated portfolio where they own six stocks. Um so perhaps there's there's more of a uh an automation case in the former case than than the latter case. But those as I was saying before, um if the foundation models themselves haven't already solved most of those problems, it's only a matter of time. And that time frame, who knows? Uh we'll see. But I would personally be surprised within 24 months if um that's not a solved problem. But I would argue, you know, somewhat facetiously, that already is a solved problem. You know, the old the old version of build me a model was email a cell site analyst and say, send me your model. Um and that's the same thing. Um and there are products like Canalus that existed or still exist that um did that sort of you know the old-fashioned, more deterministic way. And everyone's got a different thing. You know, for me personally, um, I had a different kind of heuristic, which was if I was really relying on some nuanced understanding of a financial model, it probably wasn't that good of an investment. Um, you needed to build a model um somewhat on a backwards-looking basis to understand um risk reward and stuff like that. But most of my investments, I already knew whether they were good investments or not before I built a model around them or built a nuanced model around it. So I think it'll just be different for different people. And um Excel by itself, I don't hear a lot of people complaining. Maybe I I don't talk to a lot of people, but I don't hear a lot of people complaining about using Excel for those kinds of purposes. Um, Bloomberg perhaps is another story.
SPEAKER_02What um we you talked earlier in the um sort of agenc analyst about um you know, obviously the data piece, but curious about the where you're at on integrations uh to other data, even simple things like sell-side research, pulling in the right news stack, and obviously pulling in expert network calls from players like AlphaSense have continued to be a challenge to just even get all of this external data into one environment. You've talked about creating your own data sets, but you can can you talk about this integration challenge of pulling in these other publicly available semi, you know, semi proprietary. Data sets where the barrier has been more commercial than technical.
SPEAKER_01Yeah, I mean, so for us, we basically what we solve for is the best solution. So if there are data sets that don't exist and we think they're useful, we build them. If there are data sets that do exist, and there's a huge amount of um single or few product data vendors out there that have enjoyed a uh uh sort of a cartel-like situation for for a long time around their data sets and they've acted accordingly, raised price, not innovated, et cetera. And you know, to the extent there are there is a vendor like that that our customers value where we can uh build a competing product, then we'll do that. You know, uh an example I mentioned for that is transaction data, where we've built our own credit and debit card transaction data set that um you know, according to us, uh and and you know, a number of third parties that have also validated the data is more performant um and more accurate and faster than um the existing mainstream products in that in that category. And um, we saw an opportunity there. It's obviously an important data set, and so we just decided to build a better version of it. But we're also very happy to partner with people on that front too. And um I don't feel like I need need to own all of the data layer in order to build good products, just like an investment firm doesn't need to build their own data sets. You know, if my ultimate goal is to basically build AI investment analysts, um, I don't necessarily need to own all the data or any of the data if I can find vendors who are willing to work with me in in a variety of capacities. And so that's what we're finding is um that there's a lot of people that are willing to collaborate. And um, you know, if we if they if they aren't, and or uh if the data set doesn't exist, then we'll build it and or we'll we'll iterate on on existing stuff that does exist, but we think we can build better versions of it. Um but what I would say is is our um our model today and which what what I suspect we'll continue in the future is going to be some combination of all those things. So um, you know, one advantage that we have by virtue of selling many to one, or sorry, um, you know, we buy, we buy, we buy wholesale and we can sell retail is that we can actually provide a really interesting economic solution for a lot of data vendors where um they can come work with us, we can write them a big check, um, and then we can amortize those costs across a larger customer base. Um and so yeah, it'll be a hybrid. I think there's some people that are more waltz gardened than others, and we'll see if it works. Um, I think uh my suspicion is for most people it will not. But um my suspicion is is that for um a lot of things we'll we'll partner and if we can't, we'll you know, and and it's valuable, then we'll build our own.
SPEAKER_02And how are firms thinking about their internal data? I mean, even getting to zero zero data retention at uh at Anthropic and Cloud has been a challenge for for some smaller firms, and there's always this sort of back at back, back of your mind concern that you're putting in proprietary information to a foundation lab. Are firms you know inputting the historical trading, earnings preview thesis work into your system? How do you how do you help them get confident? You get the IT teams confident in the sort of safety, security, compliance, auditability of that. And and how important has that been to really superpower these tools for for clients?
SPEAKER_01Super important. Um yeah, I think this is basically the one of the main reasons why we built the headless version of Canary. So you know, the Canary, the Canary web app is effectively what we view as the best way to interact with all the things that we've built. But we're not um we'll also be flexible enough to allow users to do whatever they want. And so if they have their own internal system, you know, for Canary in the Canary web app, you cannot upload your own investment notes, for example. Um, but if you wanted to use all of the canary data intelligence layer tools, analysts against your data, you're more than welcome to do that using our API andor our MCP. And so our view is that what is likely to be the quote unquote winning operating system um for this part of the world is you know, Claude or something like Claude, um, whether it's you know GLM 5.2 or 7.2, whatever the future versions are, or or Claude, or um, you know, whether there's some sort of enterprise tier where um those data retention compliance issues um become become less severe. I don't know. But um we we don't we don't seek to be the operating system for those sorts of things, but we do want to make sure that all the things that we do can be really easily integrated into whatever operating system that firms choose to build against. And so um, and yeah, empirically what I'd say is even today um our users have a hybrid model. Well, they'll come to Canary um frequently to do a lot of analysis, and then they'll do a lot of analysis outside of Canary, typically via MCP.
SPEAKER_02Gotcha. And so how how does that work in in practice? I have Claude, I'm doing an earnings preview, or I'm doing uh you know, a scan for short ideas, and it's pulling in the sort of underlying analysis that's done in the canary platform through through through the API as well as the other integrations and connectors that I have connected to to um to Cloud. Is that is that right?
SPEAKER_01Yeah, so I mean I what I'd say there's there's a lot of different examples for how people can use it. And some of them would just be the data stream itself. You know, just I want to look at um, I don't know, something really basic, insider buying and selling data. Great, we can help with that. Um, or if you wanted to ask, you know, say, tell me what Canary thinks about tell me what the Canary Superanalyst thinks about insider behavior at Company X. Well, then you have our intelligence layer that gets pulled in via our MCP and we'll tell you that. And then you could, for example, ask, well, show me all the supporting and refuting data points in Canary and you know, XYZ, other MCPs that I have, plus my investment notes or whatever that support or refute these theses. Um, or you could use it to cross-pollinate it against the various data sets that we have. So, for example, um, one thing that I never talk about, but we do, is we have just KPIs. So in our MCP, but also in Canary that typically just sit beneath our analysis layer. Um, but so for example, if you wanted to say, you know, map out Duolingo DAUs as Duolingo themselves have have disclosed, and then compare it to um app usage data and Canaries transaction data. Yeah, type that into Canary, let it burn, or uh sorry, into Claude using Canaries MCP, let it go burn some tokens, and then it'll provide you an analysis that cross-pollinates across all those things. So um yeah, there's there's a lot of different things. I'm sure I'm leaving off a bunch of stuff. I mean, my users often I imagine don't share their favorite uh and best use cases for for our stuff, nor should they. So um, but yeah, those are those are some ideas.
SPEAKER_02So Kay and I, between us get a handful of conversation, handful of questions a week or a month from firms that sort of ask us for advice of, hey, we want to start this journey. And FDE, you know, for deployed engineering is a big, is sort of a big concept now. Like, what do we do? Like, how do we go from where we're at, which is sort of interested, to actually AI native in the investment process? What would what are your thoughts on the FDE model? What are your thoughts on what what advice would you give, you know, public equity investment fund in making that journey?
SPEAKER_01Yeah, I mean, I think um this is another thing that is likely to um be different across investment firms. I I think it's gonna be a very heterogeneous environment, uh maybe forever. Um it starts with that thing I mentioned before about just what do you want to do yourself versus what do you want to um outsource to someone else? And you've got to draw that demarcation line first. Uh you know, I think my um subjective biased opinion on that would be internalize the vertical search model, um, where people can query your investment notes um and a bunch of other data sets that are exclusive to the firm. Perhaps there's a creation of um you know one-pagers that get circulated in advance of investment committee or things like that, whenever a new idea is being discussed, things, whatever, whatever the demarcation line is. But you know, the the vertical search capabilities and perhaps the automation layers, what I would call the more sort of cost-centric automation layers, can be things that firms do themselves. Um, and then there's a decision point to be made for what I would say the higher order um judgment-related tasks. What do you want to do with that? And um, some firm, as I as I said before, some firms will decide that they should do it, they should do it themselves, and great. Other firms will want to outsource all or part of that to vendors, hopefully like Canary. Um, but everyone's gonna have a decision point there. Then the second layer is um is just integrating. I mean, there's this integration task that I'm sure both of you are very familiar with, which is just okay, well, um, blocking and tackling. What are all the data sets that we need? Uh, and there's typically a lot. But then there's that next layer on top of it, which is which is which I mentioned also before, um what do they all mean? It's not, it's not, it's it's necessary but not sufficient to have um connectors into all the data sets. But what do they all mean? That's a separate problem. Um, if you're trying to build the intelligence layer. And uh and then I I would say effectively, you know, it's the stacking that I mentioned. People can outsource it uh or they can or they can insource it. And um a starting point, uh I think I think a very bad starting point would just to be to get whatever an enterprise subscription to ChatGPT or Claude and start trying to use that. Um the other side of it is uh having someone that can teach them what the actual limitations are, capabilities and limitations are of all the models. Because without understanding the where the the the talent frontier is, a lot of bad decisions can be made um using using them over over or under using the models. Um and so it's not you know anyone can can can get a subscription to a foundation model, um, but to actually make those sorts of things enterprise grade, wherever the cutoff is, it's non-intuitive and requires a lot of thinking before deployment.
SPEAKER_02Yeah, makes a lot of sense, and that that resonates uh for sure. What um this was great, Joe. What what did we miss?
SPEAKER_01Um yeah, I mean, I think there's there's a lot of talk. The one other thing that that I at least hear from people quite a lot is you know, what does the investment firm of the future look like? Um and uh I have a few thoughts on that, and who knows whether I'm I'm right or not. Um I say empirically that the changes have been um more more around the edges than than transformational so far. But what I think um is is is likely to be the case is that investment firms are much more organized around people with good judgment as opposed to a sort of pyramid hierarchy of junior analysts, mid-level analysts, maybe portfolio managers, and then you know, the big boss on on top, something like that. You know, the the I would say this is this has always been the case, but it will be increasingly the case that having great judgment is the determinant of great investment returns. I think that's one thing. Um the second thing is I I could I could see it being uh somewhat counterintuitively likely that there's a lot more investment firms that sprout up um as a result of the AI revolution. Um the cost of doing business will just go down dramatically. Um and as a result, I could certainly see it being the case. You know, the economics today of running a, I don't know, a $25 million hedge fund, not great. You know, no one's no one, uh I shouldn't say nobody, I would say very few people have a $25 million hedge fund and that's where they want to stay. They start a $25 million hedge fund because they want to want it to grow and whatever. But it certainly seems possible to me that in the future, um that the economics of that could be totally fine. Um and so I don't think we're gonna hit some sort of um hyper-efficient market situation where the stock prices of every company in the world are exactly right at all times. My strong suspicion is that actually for now, um the amount of alpha available to the world is perhaps greater than it's been in a very long time, um, primarily around the misuse of AI. Um and uh I my strong suspicion is that the amount of alpha on a normalized basis for a long time will not decline.
SPEAKER_02Yeah, I fully, fully agree with that. Um well this was really uh really interesting, Joe. Thank you so much. Uh thank you so much for for for being with us and um I look forward to to staying in touch and following the the canary journey. So thank thanks again. Thanks for having me to be.