Invest with AI
00:17:06 · Doug Clinton
[00:17:06] S0: Yeah. The... It's it's funny, Kaye. Like, the the terms, as you as you mentioned earlier, right, have have sort of evolved in, like, what is the hot thing at AI? I mean, it changes every six months ish or so, it feels like. And I would say this this concept of sort of the knowledge graph feels like very much the frontier right now and just trying to figure out, you know, how do you take all of the knowledge that you have as an organization or as an investment team or whatever your your your node structure is and make that entire graph usable by every agent that you have running on your platform. Right? And that may be hundreds of agents or thousands of agents depending on how evolved you are. But it's a really complex task to let... I mean, imagine thousands of different agents trying to pull different bits of context real time, creating new context, right, that other agents should then have access to. Doing that really well, I think, is going to be a source of alpha for people using language models or just agents in general, I think, to invest probably over the next... I'd give it twelve to eighteen months. I think that'll be a huge advantage before eventually it probably becomes a little bit more commoditized, whether that's because some third party vendor introduces a really easy to use system, which just doesn't exist right now that I've seen, Or, right, Claude code and codex make it so easy to just say, hey. Implement this, and it just works.
00:18:38 · Khe Hy
[00:18:38] S1: It's fascinating too because the... Sorry, Brett. The... Like schemas and ontologies, again, these kinds of buzzwords that no one was talking about six months ago, but it is this weird idea that you have this data, but you have this structure that oftentimes is just intuited by your process, by your analysts, by your PMs, by the industry. And that might be the gray matter that the LLM hasn't figured out yet. And just like watching the race to disentangle that, to categorize that, to label that, to, you know, to make it visible is just it's just really fascinating right now. And I guess probably early days to know if it's even effective.
00:19:25 · Doug Clinton
[00:19:25] S0: I think so. And it comes back to this idea of taste. You know, you hear the word taste often in AI, which I think is sort of this nod to longer term human value, this sort of instinct that we've talked about. I do think to some extent, like, creating these ontologies and these knowledge graphs, there is an element of, okay. Fine. Like, codex, point it at something, give it a really great creative loop, and it'll probably figure something out that's passable. But to make it a plus, I think it does need to have that element of taste from a human who really understands, like, this is what it means to generate alpha, not just, like, what the Internet has taught you as an AI model, but, like, this is why it's hard. Right? And this is kind of, like, maybe even the history of how how alpha sort of evolves in in a sense. And being able to take that taste and apply that to the ontology, that's where I think you get this unique value from humans coming in and taking sort of AI's baseline work, which is like B plus and then getting it up to that A level. And that's where I think you can really get some additional value from it.
00:20:30 · Brett Caughran
[00:20:30] S2: Could we drill into this this concept of knowledge graph a little bit more? Like, what does it actually look like in practice? I'd be curious, Kay, your perspective on like, what's the what's the sort of zero to one? Like, if a client comes in and wants to build a knowledge graph, what does that mean? Is that internal file structure? How do I get my data out of Excel files and OneNotes and PDFs and Outlook inbox and Slack? Like how do you, what's that actual process look like?
00:20:57 · Khe Hy
[00:20:57] S1: Yeah. I mean, I'm going to give you the training wheels version of this, but I think that it's a lot of the things that you just said where you have all of this, this data, but even something like earnings or EBITDA, that can mean 15 different things to 15 different people, even within a firm, within a subsector and so on. So if you go in and teach the model, like when I say, you know, for us, the concept of EBITDA means, you know, these adjustments, you don't never do this. You have to teach the model that. And by the way, you have to basically, like, give it a map that, like, these are the ways that we define all of these things. So once you give them the map and so this map can be just like a markdown file, right? And once you give them the map, then you have this data, right? You have, let's say, 10 companies with, you know, ten years of quarterly EBITDA across 10 different sectors. You have to basically say that like, well, the way we look at EBITDA in software is completely different than the way that we look at it in industrials. And in doing that, you know, when you want to grab that number, you have to make this adjustment. You have to look here. You have to do that. And so you're basically kind of putting it together so that when the model goes in and you're like, what's the EBITDA for Salesforce this year? It starts to go down this graph. It's like, well, I know that this like the way that Brett defines EBITDA is this. And I know that in SaaS, you look at it differently. You go here. And now I know that there's a Salesforce data here, but there's this other data I need to add back in or out. And so it's basically creating that like the term is like ontology, which is like, again, dumb person here, but you know, the definition, your definition of what these words mean so that you can translate it to the model. So then you give this, you give it all this data, again, markdown files, CSV files, skills, so on. And then you want to query it, right? And so there's all these different ways of querying it, right? You could do brute force agentic search, which is like command F EBITDA, but like still go through the EBITDA knowledge graph to get to the data point. Or you could use, like some of our other guests have discussed, semantic search through Rag and embeddings.
00:23:13 · Khe Hy
It's like, oh, okay, you know, you're asking me for profitability, which can mean a lot of different things. And how does that tie to this definition of EBITDA and Salesforce and da da da? You can then add, like, keyword search and, you know, different databases that, like, have sanitized the data and so on. So that's conceptually how I'm seeing folks do it. And I work with much smaller firms, but they're just starting with, like, a small group of names. And they're just, you know, we're gonna do the Salesforce ontology here. And then we might roll that up to the SaaS ontology over here. And just like see if it works, see if it just like speeds up the process of getting the right context at the right moment and with the lens that I'm used to looking at it with. And then you're getting dev shops that are coming in. Like the crazy thing about the dev shops coming in, because I've, like, sat in in a few of these conversations, is that a big chunk of the time is interviewing a PM. It's like, when you say EBITDA, what do you mean? Right? Because it's so baked into their heads, but you need a translator. And that's been a bit of a challenge with a lot of firms who are like, Wait, I thought you were going give me the answer. It's like, Wait, no, you need to tell me how you understand this concept and then I'll write code for you. And then a lot of people are well, I don't have time for that. Like, I thought you were just gonna do it for me. I don't know, Doug. What do