00:38:31 · Khe Hy
[00:38:31] S1: Yeah, yeah. Let me ask you this. A deadly problem. What if you did, like, what if you had a skill? Let's just use DraftKings for an, as an example. And you have email, you have something on MCP, you have, like, your internal notes, and you have your granola calls. What if you had a skill, like DIY Codex or Claude CoWork, and you have four skills? You have, like, draft, you know, granola scraper, you know, granola extractor, email extractor, and notes extractor. And you just say every at the end of every day, look at the emails from the last twenty four hours that mentioned DraftKings. And then this is where your skill comes in. You're like, I really think, like, Goldman really understands this name well. You might see, like, public betting, you know, betting data from, like, the state data, right, in, like, the body of an email and just save it on markdown, put some little metadata, who it's from, what's the date, DraftKings, save it as a markdown file in the folder called DraftKings. And you do that for three skills across three different kind of channels of data. At the end of a month, you've got 90 plus pieces of data that's pretty like again, it's curated because you created the skill. You told it what to look for. And these models are pretty smart, too. They can figure things out. And then whenever you have a DraftKings question, you just point your folder, you know, your coworker codex at that markdown folder. And you're like, hey, what's give me the bull bear thesis just off of my extracted knowledge, right? And you're going to have systems that do this, but people are like, we're helping clients do this. And I think who said it? It was Doug, who's like, you know, you get a few power users that like, this is my use case, and I'm going to kind of chop wood to make it work for me. And then the knowledge graph vendor comes in and like, Hey, we know that Brett in the corner has got these skills that every night he makes sure that it's this and that, and here's this markdown folder. And then the markdown, the knowledge graph vendor comes in and is like, yeah, we could do that, and we could do this across 100 names, across your cloud, across your Excel models, across... And so that's where you're kind of seeing this, like, nice interplay between kind of the tinkerer and the app player.
00:41:11 · Khe Hy
Like, I don't think they're in isolation from one another.
00:41:17 · Brett Caughran
[00:41:17] S0: Totally. And that's my sense. And you're closer to the technical backbone of how that's done. This sort of concept of knowledge graph is ascendant And right it's sort of Chuck supposes against the sort of first order use case of if I want to do a DraftKings earnings preview, I just ask Claude and it's web searching for all the right information. That's not a, that's not a one that's token inefficient. Two, you're going to get your reasoning on low quality context. You know, it's not an output you can trust. The skill layer is important too for the sort of like the breakdown. But if you can start to create this knowledge graph of internal external research and data, then actually like assembling that into an earnings preview with your structured skill, that's pretty token efficient. You can almost have that always on running earnings preview, ingesting your model architecture in to say, hey, based on what's happening, it looks like DraftKings is tracking ahead of revenue guidance by X amount. And you can see that like that's in a passive dashboard. That's not a chat. You know, that's not a chat interface where I have to think about, oh, maybe I should ask questions about this. No, I have that dashboard amongst all of my portfolio into a portfolio thing where it has sort of alerts of what's happening, hey, which companies are tracking ahead or behind, which key drivers are sort of improving, deteriorating, etc. That to me is like the real, not holy grail, but that's the, I don't know what you call it, the codex moment for public equity AI. Like that to me is within reach and more of an assembly problem rather than a technology Absolutely.
00:42:57 · Khe Hy
[00:42:57] S1: And I think that's the phase we're in now, right? Because even in the example you described, still a big problem there, which is the Microsoft three sixty five connector or you know, BIP sync or tamales, lack thereof, right? It's like, you know, in that example with a three sixty five connector, it's still running a search across your entire firms for like DraftKings, right? DraftKings earnings, DraftKings 2026, right? And trying to guess, you know, if there's, you know, 50 word documents on DraftKings, it's trying to guess which one do I wanna open, which one, you know, there's V7final, right? Like, is that the one with it? You know? And so that knowledge graph and the skills that build the knowledge graph is like, yeah, there's seven memos of, you know, Q2 twenty twenty six, but like, this is the one, but not only this is the one, these are the parts of the one that really matter to the analyses.
00:43:52 · Brett Caughran, Khe Hy
[00:43:52] S0: Yeah, yeah. Yeah, sort of like had this discussion, We had it before. It's like, we're almost sort of in the Napster era where like you can, with a lot of work and architecture and steps, can get something interesting and time and expense, you can get something interesting out of the back end. You know, are the tools obstructing a lot of those steps and you can all of a sudden now you can hit any button on your for those for those Gen Zers on the phone, you just have to wait three hours to download one song on Napster. Now any song at the touch of a button on Spotify. So if you believe the Elon Musk maximalist view of the world, you, the game hasn't even started for the fluency and power of these tools, but totally bring it back to something UK said is like, have to sort of really think about the right problem to solve. [00:44:41] S1: Yeah. Yeah. And like that, that's hard, man. Like, there's so many moving parts. Like, like, what's the right like that? What's the right problem to solve? And And it's not just, you know, generate alpha. Like, yes, that is a problem. But like inside that, there's so many subproblems. And I think that that's like the people that are really thinking through. Someone told me they're like, we used to hire where we're This might have been on a podcast together. I've had so many conversations, but it was like, we used to hire. Yeah, I think Thomas Lee said it. I forget. It was something about like looking for people that are really going to process.
00:45:20 · Brett Caughran, Khe Hy
[00:45:20] S0: Yeah, yeah. Yeah, this will be a conversation for another day, but some interesting movements on AI hedge fund approaches, too. And it's the exact I think it's the exact same question of, you know, one shot doesn't work or most likely won't work. But actually breaking down the, you know, breaking down the alpha generation process into the sub problems. You're starting to sort of see an architecture that I personally think could be quite promising. So things are moving. Things are moving. Things are moving. Think at a sort of accelerated pace. I don't think people most people I talk to don't realize how big the sort of mid summer inflection was and how important sort of improvement in foundation models, the sort of economic implications of open source models, and then what's happening in the connectivity ecosystem. Those three pillars to me conspire to you're really putting us at a really interesting inflection point in the application layer and AI usefulness more broadly for Wall Street use cases. Exciting and scary times we live in. [00:46:34] S1: Yes. Well, thanks, Brad. This is always fun. I like these one on ones. Awesome. [00:46:41] S0: Alright. Thanks a lot, Kay. [00:46:42] S1: Thanks, everyone. [00:46:42] S0: See soon.