00:09:13 · Brett Caughran, Khe Hy
[00:09:13] S0: a [00:09:13] S1: big one, right? Like scheduled tasks on Claude code require your computer to be on. So it's kind of running locally. So basically, you're still running all these things locally. But because you can access them from your mobile phone, your iPhone, it doesn't feel like you're constrained anymore. Like, you could just be on the go just doing your work. So that's I said a lot there, but let me, I'll pause and let you react and ask questions. [00:09:44] S0: No, it's interesting. I mean, the one fundamental question I have is what does this mean for Wall Street people who've been working to build fluency in the tools? I mean, a core part AI training on Wall Street the last eighteen months has been, really last nine months has been Claude fluency. And even a core part of your practice has been helping firms get up to speed on Claude code and Claude co work. Yeah. What does that mean now? Like what's the sort of the next step as Claude gets decommissioned in some of these instances?
00:10:19 · Khe Hy
[00:10:19] S1: Yeah. I don't think it's going to get decommissioned. At least right now, I think people are just adding because a lot of people kept chat GBT. So they're just turning Codex on and giving people two options. But the thing is that like It it requires, you know, now assuming all the systems are equal, then it's like, what system are you designing? Right. And so I think that what I've noticed in my training was a lot of the training, I'd say until like two months ago was just like, show me how to use these things, show me use cases and so on. In the last two months though, as people have gotten more fluent, it's like, it's what's the art of possible, right? Like what is possible? So, you know, I think of one of the great advantages of Claude Code and co work was the ability to schedule tasks. Right? Because that's where you could say like, Hey, every morning, you know, I'm a consumer analyst. I'm tracking inflation. You know, go to Dick's, go to Amazon, go to, you know, to adidas.com and pull up the price of, you know, Stan Smith's sneakers and save it in an Excel spreadsheet, right? And so the original, like the, so that was the first time you could start doing these things, like really automating agentic work was with a trigger. And so then the question though is, well, what do we wanna automate? Right? That's actually a harder question than how do I automate something? Yeah. And that once you figure out like, well, you know, some folks will go through the way back machine to look at like the number of stores that existed across, you know, going back quarters at a time using the internet archive, using the Chrome extension, right? All Codex does, like, Codex makes the idea... What do they say that? Execution is cheap. Ideas are precious or rare. And so now the execution, I would argue, has just been flattened, right? It's it's got an even... You have no excuse now. Maybe before you had an excuse like, Well, I'm not gonna scrape because I don't know Cloud Code or, you know, we haven't set up the Chrome browser or what have you. You don't have an excuse anymore. What data are you gonna go look up in your remit, in your landscape that was not possible for? And what are you gonna do with it? Right? And so I think it's the, you know, kind of it's like another the abstraction is moving up to like, what's the hard problem to solve now?
00:13:05 · Brett Caughran, Khe Hy
[00:13:05] S0: Yeah. Yeah. That's where I like I would chuckle. I would chuckle when I watch these YouTube's on like setting up an OpenClaw or an Hermes agent. And it's like all of this futuristic approach. And then at the end, the use case was creating a podcast summarizer or something that's like, all right, like, okay, how was that? Totally. And the sort of like tension to me in long running tasks and people would, oh, I have 20 agents working overnight. And I would be like, On what? Me, we're not even at that point for... Certainly, I come at this from a institutional grade, bottoms up fundamental research perspective where the individual components of that process are still not good enough getting closer. But just let an agent rip for twenty four hours on doing a full deep dive due diligence process. It's just going to be this long running slop chain [00:14:02] S1: that [00:14:02] S0: burns a lot of tokens. Yeah.
00:14:07 · Khe Hy
[00:14:07] S1: No, I hear you. And one of the things I'd be curious how you're approaching this with your clients is I feel like we're ending this era where I call it the faster horse era, where you do something like an earnings preview and it would take you a couple hours And then you used AI to make it better and faster. Like, you built a faster horse. And the faster horse era of AI, it's easier because you just like, you look at everything you do, and you're like, how do I make it faster and cheaper? But then there's another question that's like, now that you have this resource that's different and very powerful and has a different cost structure, what are you going to do with it? Right? So like one of the things that's like interesting with private equity you know, they're constantly receiving SIEMs, right? Confidential information memorandums. These are like deal opportunities, right? And so they could have hundreds. You know, the firm could see dozens of them a week, maybe more depending on the firm size. So, you know, in the old days, what you do, an analyst would review it and like, this is worth spending more time on. Let's take it to phase two. Right? But these are also, like, really information rich documents and pretty, like, opaque private markets. What if you could have a bunch of agents that go in and, like, basically scrape every KPI from every SIM and categorize it for you? Like it never made sense to do that before because you'd have to pay an associate, you know, a $100 to do that. It's not a good use of their time. But if it costs you, you know, 20 of Fable tokens, maybe it's worth doing, right? Maybe it's worth it. So I'm curious, like how people are if you're seeing that, like faster horse to like expanded art of possible.
00:16:14 · Brett Caughran, Khe Hy
[00:16:14] S0: Yeah. I love this question because I think like, you know, as a Wall Street user, we stare at this blinking context window and it's actually like harder than it would seem to figure out to enact that context window to be helpful in your investment process. Worse than that, I think the knee jerk reaction of most investors is to use it the wrong way. Is this sort of, hey, I used to read the 10 ks, let me do a 10 ks summary now and speed read through that, the artisanal process of comprehending the company. And then sort of, you know, or the expression layer, like here's a bunch of notes, create a bold thesis for me, right? It's sort of an AI slop thesis. So I think I don't know what it is about AI, but it causes probably late, Fred, like it helps elevate the sort of lazy predispositions in all of us that sort of just want to like sit on the couch and eat Cheetos versus apply real rigor. And it makes makes the sort of transmission mechanism to laziness much easier because it sort of looks good enough. [00:17:19] S1: Yeah, yeah.
00:17:20 · Brett Caughran
[00:17:20] S0: That to me is a very important force to fight against. Yeah, yeah. So I think it like we think about it as sort of five layers of like the triage layer, the execution layer, the decision decision, decision support layer, the sort of rigor enforcement layer, and then the idea tracking layer. So like, it really is to me like building a system around it. To your specific question, the decision support layer, you know, something like that where you could build a headless RMS to sort of go into all of the prior operating memorandum, all of the prior deals, pattern recognize the sort of the best deals, the worst deals, errors you've made in the past, and have that in a dashboard while you're sort of doing the evaluation of a new operating memorandum. That to me is like that decision support layer is the real... It's not some AI slop summary, right? It's, hey, go and actually run and analyze, you know, sort of a real institutional grade process on this OM. And part of it is, you know, pulling in the lessons of history. I think to me what has been so exciting, and I think we're at a real inflection point on these tools on Wall Street is the fable and soul class of models have gone from answer engines, prior LMs were like Q and A, give me an answer, to actual work mechanisms. To me, you can build with Fable and Soul is so much more impressive than what you could build with Opus or GPT 4.5, etcetera. So the ability to actually, you know, ramble Max into a context window and have it build the first version of, you know, some architectural thing is I think that we were sort of going through this diffusion mechanism where more people are realizing how powerful this latest class of models and probably then sort of looping that back into the work you're doing with Codex. Like you probably just ramble into Codex and that decision support layer on the OM analysis in a way that's like much more frictionless than it was six months ago.