00:10:50 · Raj Shah, Khe Hy
And so there's this really interesting period of time where if you've done the math on what shares are being issued, what's the pro form a balance sheet, who's running the business, what does the growth look like, there's a great opportunity to get in there early transformative m and a situations. But LLMs are not necessarily gonna be great at that yet if they're relying on SEC filings and the... That type of data in identifying those situations for you. What it does hold the promise of is I should be able to put a slide presentation about a merger into the LLM and have it produce the pro form a financials and flag for me. Hey. By the way, this company is now, after a result of this deal, the cheapest among its 10 peers. That's pretty interesting. Synergies look really reasonable compared to the OpEx at, you know, a, b, and c peers, those types of things. And so we're really excited about it. I could go on and on about the screening and idea generation front. Bottom line is I think it's still the most exciting part of of of this whole AI in investment management or hedge funds world, but it's also the part that I think needs the most most work. [00:11:51] S1: Can I ask you about the the screenings or the two prompt question? Such a good, like, entry, you know, like, AB test right there. Can you say a little bit more about how that is set up? Is it, you know, a really good ChatGPT prompt or skill that you've reused? Or is it... You've mentioned third party software. Like, how are you actually running that prompt?
00:12:14 · Raj Shah
[00:12:14] S0: So we've we've done it a couple ways. And it's it's a great question. Originally, we started doing... We started using some service providers that that purport to have these types of tools. And and and a lot of ways, they're they're really good. I shouldn't say purport to have them. They they have these types of tools. Portrait is a great one, and I I know David's been on with Brett before too. Portrait's a great tool that can take a lot of your natural language queries and turn them into both quantitative and qualitative screening, and then they'll go and scour and find those. And that's great. But we found, you know, I should say, I I think LLMs almost always... They're they're always going to try to find the shortest path of the shortcut to getting you the answer that you've asked for. And so I think part of the problem in why you get different results is that when you put this criteria in, it's sort of scouring only part of the investable universe. If we do a similar screen on Bloomberg, for example, using EQS. EQS is like the OG screener. It's actually very, very good. It's in... It's almost entirely quantitative. But if you run that screen over and over, you will always get the same results back. Like, I have a very high degree of confidence. When you run an EQS screen, it's it's it's it's finding everything, and everything is accurate as much as you've put the data in. So what we've started to do because of some of these sort of LLM native screening tools have had these inconsistencies is we'll sort of start with Bloomberg. We'll start with the quantitative to know that we've scoured the entire investable universe of these ideas. We'll spit that data out. Bloomberg's kind of been the sleeper, I feel like, in the AI wars for hedge funds right now. Because obviously, they sit on all this data, but their their models have generally sucked, but they're they're coming from behind here. Bloomberg's now because their language, because it can dump into Excel, you can take those screens out of Bloomberg and then throw them right into somebody else's product and say, look, this is the universe I want you to start Don't worry. I mean, to pick on portrait, let's say AlphaSense, don't worry AlphaSense, don't worry Portrait about burning a bunch of tokens running this screen for me. I've already picked from 1,500 names, picked 500 for you.
00:14:21 · Raj Shah, Khe Hy
Now go do this qualitative screen for me. And so I think tying a lot of these things together helps helps on that idea generation front. [00:14:28] S1: And, Brent, we talked a little about this with Ying on the podcast. It's like the separation of the deterministic query from the nondeterministic query. And this is... Raj is giving a perfect example of... You've kind of hacked away into it. Right? But eventually [00:14:42] S0: Yeah. [00:14:42] S1: Imagine that those emerge. [00:14:45] S0: Yes. And they... And we've we've we've hired a couple enterprise and college students this summer, and this is their core focus is, hey. We know there are some strengths here. There are some strengths here. Like you said, deterministic, nondeterministic. Let's actually just build an internal solution where we can weave all of these things together. And and they've only been at it for, like, a week, and it's showing a lot of promise. It's also showing some obstacles because it's funny. Some of these resources that we're using that we're we're paying for really would rather prefer to be walled gardens where you're doing all of your research, and they make it really hard to sort of leverage the best of what they offer and and tie it in with the best of, you know, what somebody else offers.
00:14:45 · Raj Shah, Brett Caughran
[00:14:45] S0: Yes. And they... And we've we've we've hired a couple enterprise and college students this summer, and this is their core focus is, hey. We know there are some strengths here. There are some strengths here. Like you said, deterministic, nondeterministic. Let's actually just build an internal solution where we can weave all of these things together. And and they've only been at it for, like, a week, and it's showing a lot of promise. It's also showing some obstacles because it's funny. Some of these resources that we're using that we're we're paying for really would rather prefer to be walled gardens where you're doing all of your research, and they make it really hard to sort of leverage the best of what they offer and and tie it in with the best of, you know, what somebody else offers. [00:15:22] S2: How... Sort of interesting conversation there too, of like, how are you thinking about the UI here? Right? It's like in the past to try a bunch of these chatbots, you had to have six logins and what's the new dashboard as we sort of transitioning from chatbots to agents? How are you thinking about the UI and how all these things plug in together? It still feels to me very disjointed. Have you done to sort of turn this into a new seamless operating system?
00:15:54 · Raj Shah
[00:15:54] S0: It's terribly disjointed. It's a huge problem. In some cases, you're like wrestling with PDFs. In some cases, you're wrestling with spreadsheets. In some cases, it's like natively text on a website where like your best option is to like export it. And this is... I mean, we haven't... The answer... The short answer is we haven't built it yet. It's it's it's like goal number one for the next two weeks for, you know, our our enterprising college students is to figure out the best way to put this together. Part of the interview process we had given them was actually to go filter through a bunch of names and, like, do some preliminary screen for us. And one of them came back with a pretty interesting output that they use Claude code to create, which was an interactive website. And it had pulled in using MCPs, using, in some cases, agents to, like, navigate with a mouse, some third party website and pull that data in. I think that's ultimately probably the way we're gonna end up consuming this is... But I'd like to be... I don't know. I'm a, you know, now an old hedge fund portfolio manager. I prefer using my keyboard over my mouse. So I'd rather not have to click a 100 times to go get to the data I want. But I think the reality is it's probably gonna be something like, you know, a very simple website where every day it's filtering in. I... We originally thought of emails too, like, just email me when you find things. We have this concept of we internally, we call it the meta screen, which is like, we have a dozen screens. We probably have 50 screens. If there's a company that shows up on, like, seven of those screens, like, slap me in the face with it. It's it's on the meta screen. Like, we should absolutely be looking at that scenario. And still though, if it emails it to me, like, I'm probably still gonna get it. Like, I won't I won't get to it till later. You know? Like, it's just gonna get lost in the box. So I have to establish this workflow of that. Web page has to be one of the first five things I do every morning. Just go check that site, anything that's popped up that looks really interesting. So that's what we're theoretically playing around with right now. But, you know, this is this is evolving. And I know... Maybe it'll be an app.
00:17:45 · Khe Hy, Raj Shah, Brett Caughran
[00:17:45] S1: That on that point, Raj, the the Claude code project that that the intern built, you, you mentioned it. Yeah. But just to call it out for others, the computer use has gotten really, really good. Like, especially on Codex where it full on just takes control of your computer and clicks through Bloom. Yeah. Like, and it's funny because some some people believe it's like, why even bother with APIs and all that? Because computer usage is getting so good that we're just gonna, like, give AI this human thing, but they're good enough at it. So it's interesting to hear how you're... [00:18:17] S0: No. It's totally true. I think it's a matter of speed. Like, if you have a direct API connection, obviously, you can just move faster. But you're absolutely right. Like, it's it's contextualizing these pages. It's clearly remembering how they work so that the second or third go around, it's... It knows exactly where to go to pull the information it needs to. It's scary, but it may be the way we sort of to Brett's... The way Brett, you kind of have to hack it together. Can this... So for now, [00:18:40] S2: I don't know... Could Codex Computers say, go into my Bloomberg, run BQS, download the Excel, go to AlphaSense, pull in expert network calls? Does it have that capability?