Invest with AI
00:27:02 · Doug Clinton
[00:27:02] S0: Yeah. We've we've built so much stuff that we no longer use, and, literally, some of it has lasted a month, two months. And I I think, like, with the emergence of the coding tools, that has accelerated it, you know, because it is so easy right now to to write... Not even just, like, basic code. I mean, to write pretty decent code, like, very functional programs, even for people who aren't, like, high level engineers, is is pretty easy if you know how to use the the systems well. Now writing something that scales to an organization of maybe a 100 people, a thousand people, that is a different sort of challenge. But I I think that we should assume that that reality continues to be persistent, which is technology will evolve really fast. So build your systems as kind of quick as you can to work really well for the paradigm today. Take advantage of everything you can today, but be ready to be nimble and don't have a lot of, you know, sunk cost in something where you say, well, we can't change that because we just spent $2,000,000 building the system. But there is this better thing. It... I mean, it's almost like NVIDIA chips. Right? Like, you buy your h one hundreds and then the b 100 or b 200, right, is way more efficient. You save more money, you know, per token that you create by just investing in the better chips. So you just have to be prepared to do that just like the data centers are doing with their with their infrastructure.
00:28:25 · Khe Hy
[00:28:25] S1: Yeah, a funny story on this. There was an AI vendor that was very hot on, it's kind of an enterprise search in 2025 before the MCP takeoff moment, which I would say is like December of last year. And everyone was into this vendor. I don't want to name it. But I was into this vendor. I did tons of calls with them last year. And then when the connectors hit, it went dark. No one wanted to talk to this vendor at all. Last week, a friend texted me. He's like, hey, I actually stayed on that vendor, and they have incorporated some of these elements of knowledge graph, different semantic search. Permissioning is a big deal in the knowledge graph. And it's really good. That's what the front... What the text told me. So even in the vendor cycle, you see it. It's wild. Like, I would not wanna be building and selling on that side of the fence.
00:29:21 · Doug Clinton
[00:29:21] S0: Yeah. I agree. And I think that's actually a great point, too, Kay, because if you think about in the investment space, the ability to control your own destiny too. This is just a philosophy that we have at Intelligent Alpha. But to the extent that we can sort of build something ourselves or use open source and modify ourselves very quickly, we've tried to very much do that and not rely too much on third party vendors because of what you just described. Like, you know, the vendor could be awesome when you install them day one, and something new could come out. And and by no fault of their own, just the fact that they bet on a certain idea or a certain technology, maybe they fall behind a little bit. And six months later, maybe they're awesome again. Like, it's just really hard to not be able to sort of control your own destiny given how fast things are moving. So that's the other piece for us is just philosophically kinda how do you think about how quickly and and what kinds of investments you wanna make in building your technology infrastructure, and is it something that you wanna have a lot of control over? Or or are you okay with relying on good third party vendors for that?
00:30:23 · Brett Caughran
[00:30:23] S2: Doug, have you thought about model model routing? You know, sort sort of feels like every few weeks you get a new model and, you know, the OpenAI models were ascended. We need this sort of h one twenty six has been all about Claude. Now this sort of conversation shifting back to Kimi and DeepSeek and the open source models and toke... Token efficiency argument, etcetera. How have you thought about evaluating that systematically? And when do you make the decision to plug in a new model, sort of pivot pivot your routing structure?
00:30:52 · Doug Clinton
[00:30:52] S0: Yeah. We do a ton of benchmarking internally. We've actually started to publish some of this work as well externally literally today. So it'll be a few days, I guess, maybe prior to when this gets launched. But we launched something called the IA 500, which is a nod to the S and P 500, obviously. But it's a it's a stock picking benchmark where we have a quarterly rebalance structure just like most indexes, and we test 10 of the top models. Top meaning, you know, most influential. It would be all the names that you would expect to kinda be in that group. And we look at their ability to pick a portfolio of about 200 stocks, And then we actually ensemble all of the picks from the 10 models into a single portfolio of about 500 stocks. And it's performed quite well versus the S and P 500 over the past year since we've been kind of putting this data together. So you can check it out at ia500.com. But the the kind of answer to your question, though, Brett, is we're always doing these tests on an ongoing basis to see what models sort of stand out where. And what I would tell you is this. There, in our view, there is a separation that is notable between the closed source models and the open weight models that we see on most of our benchmarks. In general, I would say when we're talking about investment tasks, not just stock picking, the the closed source models and GPT has consistently been the best one that we test in this section. They always have a little bit of outperformance as a group relative to the open weight models. So we do see some superiority there. And I know, like you said, Claude has been, probably the hottest one this year. And, this is a contrarian take, but but I do think that GPT is probably a better model overall for the stock related work that we do. It generally is our top performing model in most of the things we look at. So that's piece number one. On the open source stuff, I actually think that, to your point, is is really interesting because Kimi's had a moment. Nematron is is actually, I think, quite good as well on the American side. GLM, I know, was was hot before Kimi came out with k three. I think we're gonna see kind of the same thing there in in this open source space, which is just like the closer...
00:33:17 · Doug Clinton, Brett Caughran
Or the the closed source models, very rapidly sort of pushing the boundary up but maybe not quite hitting the closed source. That's sort of my bet is that we'll see a persistent gap between closed and open, mainly because I think a lot of it is, you know, based on distillation, how fast some of the the the open source stuff is moving, which is totally fine. But if you really wanna be on the frontier, and I think you do want that in the investment space where 1% edge is important, I think you really do have to spend more time looking at the closed source models. [00:33:53] S2: Yeah. Okay. How have you thought about this? I mean, you've spent a lot of time with Claude specific training, Claude coworking in particular, which has been sort of the default UI for, you know, mid to small size funds. Super powerful, particularly when you sort of set it up with the right connectors and the right skills. If funds have adopted Cloud Desktop, Cloud Cowork Code, etcetera, and you see this model evolution, what's next? What's the counter punch that evolution?