00:20:20 · Khe Hy
[00:20:20] S1: I think one of the big ones, believe it or not, is just, I would say, kind of a failure of imagination. And what I mean by that is that you're kind of flash with this, you know, you're presented this blank flashing prompt. And it's like, this is, you know, 200 Nobel Prize winners at your disposal. Go nuts. And what happens is people kind of anchor to the existing mental frames that they know. And that's like Google and chat. So immediately, you start off on chat. So that would be kind of the first challenge, which we kind of help people shift into that agentic mindset. And then you treat it like texting, where it's, you know, one sentence prompts instead of a two hour prompt, which, again, prompting has shifted. But it's not like an iMessage. It's much more like a brief in many of these cases. So I think that's the one of the biggest ones is to kind of like unlock what's possible, right? And so, you know, I joke with people. It's like, you now have a data scientist in your pocket, right? But what does that mean? That means you need to have the data that a data scientist works with. You need to understand the language to communicate with a data scientist. But more importantly, you need to look at your problem space and understand what a data scientist could be useful for. And if you never worked with a data scientist and you might have at some of the firms you have, that is a completely different world. Right. And so I really kind of think of it as the expansion of kind of what's possible. That would be one thing. The second thing is usually just like, how do all the pieces fit together? The data, the prompting, the custom instructions, the memory, the statelessness of an LLM, projects. You just kind of like... So I would kind of put that in kind of like marrying the workflow with the tooling and the features. And by the way, there's like, you know, Anthropic's been on a tear, but they've launched a new two new features a week since January of this year, and we're in April. So that'll be another. And the third is and this is actually a harder one is, I think I saw this quote the other day is, so much knowledge work is just vibes in spreadsheets. And people don't actually know what they do for work. It's the vibes. They're like, okay. You know, it could be anchored around thing like quarterly earnings. Like, there's an... There's a huge...
00:23:02 · Khe Hy, Brett Caughran
They know what they do there. But if you ask them, like, what's your screening process? Like, oh, like, sometimes I look at 13 Fs. Like, we'd go to these idea dinners. There's not, like, an actual... It's Vibes, right? Sourcing happens on Vibes. And this actually isn't an LLM question. This is much more of a process... I don't even... I'm not a... I'm not, like, a formally trained consultant, but it's like process excavation exercise, for lack of a better term. [00:23:34] S0: Yeah. [00:23:34] S1: Or maybe workflow mapping, right? And I think that, you know, it's okay to run with Vibes when, you know, without LLMs because you can kind of, like, flow back and forth and you can kind of work in that kind of osmosis, a little bit reactionary way of doing things. But with LLMs, you need to be clear on your intent. You need to be clear on your steps. You really need to, like, deconstruct things. And that is that's the hardest thing with my clients because every analyst is every firm is different and every analyst is different. And mostly, you know, that is kind of it's a very kind of human exercise of, like, extracting out knowledge and then remapping it towards an LLM.
00:24:22 · Brett Caughran
[00:24:22] S0: That's such a salient point. And I think that's where, you know, I'm spending a lot of my time, and I completely agree. It's almost investing after you've done it. It's why it's hard to learn, right? Because really great PMs will have a process, but the process leans more towards an intuitive sort of tacit understanding than anything that's overly robotic, right? Almost no PMs have a 24 page checklist of, If I look on an idea, I go check off every single step. They've done it hundreds of times. And like a master playing a piano, they look at an idea, hit 72 Bloomberg functions, update the model, call the man. And they just sort of do it out of muscle memory such that the workflow is tacit, not explicit. That's really hard to build an agent around. You kind of have to slow down and document that process. And that's really hard to do because investing is a non linear process. That's what we've done for four years at Fundamental Edge. We've tried to create a linearity out of it to bridge that knowledge of the junior analyst to the PM. It's not perfect. But I completely agree that that's a big part of the journey of enabling agents in your process, documenting your process. Also documenting LMs or wonderful pattern recognition engines, like the mental models in your mind that you come back to over and over. How do you codify those such that an agent can think the way that you
00:26:09 · Khe Hy
[00:26:09] S1: think. And have you found because right now, let's say there's a 20 step process and LLMs might be good at three of the 20 steps. Have you found that to be frustrating for people that are like, it's kind of like the junior person, Oh, I'm just going to do it myself. You know, how do you how do you find that tension? Because I found that people, you know, you go in for the first time, like, I want this to, like, pull all this data, do this, do this, send me a report that does this. And then, you know, send me a Teams message every time the stock's down, like, x percent. Right? And, like, woah, woah, woah. Like, let's pick the one, like, thread here that is this... That lines up with the strength of the LLM. Like, how have you kind of... Or has the expectation versus reality, it's like the Tim Urban kind of delta, how has that manifested with in your client work?
00:27:04 · Brett Caughran
[00:27:04] S0: It's definitely a journey. I think no consultants or vendors do themselves a favor by over promising around the capabilities of LLMs. This has been rampant amongst the vendor community to say, Hey, our tool can do anything an investor can do. Splashy demo, adoption, disappointment, churn, right? It's been this sort of cycle for three years. So we try to really be very transparent around, we have a red light, green light, yellow light framework of what's possible today, what's sort of close, maybe, maybe not, and then a red light. For example, an LLM cannot one shot a complicated three financial statement model today. It can take a model that's five years out of date, update it, resegment it, find the three key drivers, key in the thematic leverage, and then run automated research around those ideas and feed those findings back into the financial model and compare my model to consensus. That's sort of like one of the holy grails of agentic systems. It's possible today. So what we do is we try and chunk down that process to like the 17 steps you would need to sort of have the pillars of running one of those holy grail movements. And we just experiment. The frustrating thing with AI sometimes, and exciting thing is something that doesn't work today could work in three weeks. Or sometimes what works on Tuesday doesn't work on Thursday because Anthropic decides to throttle compute. So it's this messy evolution, this messy journey of what's possible today. Now, what we try to do is find those workflows today that pass that moment. Hedge funds are hard to impress. And chatbots, for the most part, were not impressive. You're starting to see more workflows where the outputs are impressive. So I'd say like, as you're thinking about any investor adopting these tools, you can't do it all at once. You can't rewire your entire investment process. This is likely an eighteen month, nine to eighteen month journey. So find those individual workflows where you have overlay. The And for the first period of time, process it. Don't change the way you're doing it. Just have an agent run-in parallel as sort of... We use the analogy of bumper bowling. Something, have bumpers behind that process that can catch any process, mistakes that you might make. So I think it's a slow journey.
00:29:50 · Brett Caughran
Finance is not going to flip the switch overnight the way that engineers have with this much more complicated game than building apps with coding agents. So the challenging part of why we try to provide value is this takes a lot of hours just to stay up with this stuff, right? Like, I don't know what I'd be curious actually, what's your learning journey? Like you're sharing all of this with the world via your blog and social. Like what does your week look like in terms of what you're reading, watching, you know, listening to to stay up to date?