00:03:27 · Joe O'Donnell
[00:03:27] S0: Yeah. I mean, I'll make a very bold claim to start with, which is that AI in general, would say, for 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 public markets investor's opinion about anything. Mathematically speaking, the average public markets investor destroys value. And I would say an even stronger form version of that might be the seventy fifth percentile public markets investor. And really, maybe conservatively doesn't add that much value. Everyone's in search of the last few percentile. I'd sort of equate it to a golf tournament. Every professional golfer, PGH or whatever takes two fifty, whatever, 300 shots over the course of four rounds. And the difference between, you know, the person who collects all the gold and the person who nobody's heard of is a few shots. But that ends up being all the value. And I think 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. And so, yeah, I think another another way, another thought experiment, at least that that that I thought of this morning, which I think is kind of interesting, would be, you know, if you're... If one were to compare, you know, put on... Have a contest, 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, and the SEC, and something like that against a thousand junior investment makers, and they get to do whatever they wanted. My strong suspicion would be that Buffett and Druckenmiller would win. And 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 the judgment layer is ultimately what determines whether investors have good results or they don't.
00:05:46 · Joe O'Donnell, Khe Hy
And so in order to build 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... You I think going back to what you're saying about how there's only a couple companies 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 I think it's a it's an odd group that, I guess we're we're a part of on that regard. [00:06:25] S2: Joe, I love... Could you talk a bit about for, nontechnical folks, the architecture of Canary, like the data where the LLM sit in, the deterministic piece, the judgment piece. Love to hear that from you.
00:06:38 · Joe O'Donnell
[00:06:38] S0: Yeah, I think we do hundreds of microservices that interconnect in various ways. I think there's different, I mean, there's a few different orthogonal ways that I think about at least, but one is just more obvious, which is we build data sets that don't exist. So just starting from that. I think, you know, the investment world, I would say, if if there's still a data advantage that exists in the investment world across the mainstream datasets, it's limited and declining. But there's a lot of niche datasets that don't exist. And by virtue of limited distribution, it 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 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. The next thing that we do is we build intelligence on top of those datasets and the more mainstream datasets. Basically, the sort of... 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 an annual report. You know, there's the... So 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... Today, you know, there's Claude, tell me all the interesting stuff in these two reports, make no mistakes or whatever. Something worked elegant in that, obviously, but something along those lines. And then there, I think there are some people that are trying to build 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 sort of more mainstream datasets where we build intelligence layers on top of that.
00:08:51 · Joe O'Donnell
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 a AI forensic accountant that can unpack and does unpack across 10,000 companies or so all of the various accounting manipulations that certain companies engage in and provides a clean P and L versus the company's reported results versus consensus, etcetera, stuff like that. So these are specialized roles. 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 increasingly specialized ways. And so, you know, it's sort of a layer cake where all of these things are necessary results that are actually useful for investors in terms of bending the curve on their results. I think one thing... On the other side, one thing that we shy away from is doing things that we think are just going to be... They're either already are or are very likely to be subsumed by foundation models. You know, I think, you know, we're we're a startup. It's... It'd be foolish for us to try to compete with Claude, for example, on anything that Claude is likely to do very well. 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. You know, build me a Excel model with one click or something like that. Cloud is wonderful at that with no plug ins at all. And fast forward twelve months, it'll be even better. In twelve months from there, I think it will be perfect. Maybe maybe even sooner.
00:10:41 · Khe Hy
[00:10:41] S2: Could I add the layer cake analogy is 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 your product changed as the models have, like, step functioned?
00:10:59 · Joe O'Donnell
[00:10:59] S0: It's so it's a it's sensor in some ways, and it requires deep analysis and QA at every single layer of all of those things, which is an unfortunate reality of what we do. 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 was way back for... With Gemini 2.5 where we started seeing that that level of functionality where we could build sort of more performant, more horizontal investment tools. And it really depends. So it's model specific and we don't 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 we're finding just that fine tune models that we build ourselves, not because of anything related to cost, which I think is a big topic these days, but for performance purposes. You know, our goal is to deliver the absolute best results and the costs are just what they are. And so, you know, there's no trade off that we think we can participate in with respect to that. And so sometimes 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. In other times, it will it will be different. But it's really nuanced, and and I would say increasingly only applies to certain things that we do.