Invest with AI
00:41:09 · Kirk McKeown, Brett Caughran
[00:41:09] S0: I mean, the I'm a little torn on it, right? To be honest, because like there's a there's the hot take framework, which is like 10 years from now, stocks achieve singularity and they're no longer traded, right? [00:41:19] S1: Like there is that world that you could contemplate, whether or not like if there's a probability on all these things and like mediated by agents, like humans completely out of the loop. Yeah, more efficient, more efficient, more efficient mechanism for capital allocation. [00:41:34] S0: Yeah, like because here's the thing, right? Like, I actually come to a place where like the natives that we're creating are going into models either as weights or as inputs to react to, right? Because you could post-train a model using this data, um, and we're doing some work on that, but you could also sort of um like at the end of the day, the reasoning models have to reason against something. [00:41:59] S1: Yeah, right.
00:41:59 · Brett Caughran, Kirk McKeown
[00:41:59] S1: Yeah, right. [00:41:59] S0: And so the the thing they have to reason against is things happening in the world, right? And that is real-time transaction data, that is real-time claims data, that is sensors that are collecting traffic as it drives down the highway. Those the model needs to react to something, right? Yeah, there's a world where like the markets uh actually evolve to data element markets, data native markets, where you know, sort of the models are actually trading inputs that better inform the problem that they're trying to solve. And that's closer to ground truth. Like stocks are an abstraction of uh cash flow and what a com what somebody wants to pay for it. Right now, there's other components of like equity and what that means and the legal constructs. So I'm I'm munging a bunch of shit here. Oops, sorry. But um, you know, sort of uh, you know, I I don't think we're anywhere near there. I think we're I you know, I think I but like I do think you we will see increased volatility um because it assumes all the models are gonna be good too.
00:43:00 · Brett Caughran, Kirk McKeown
[00:43:00] S1: Yeah. Yeah, the exponential has been pretty steep. Like the log, you know, the lot the sort of trend in the log scale has been pretty linear so far. Um, you know, uh Dario's probably been more right than wrong on timing of some of these things. Um so it's a little little terrifying. I think we yeah, I agree we at least have to sort of contemplate that, contemplate that future as a note on the sort of distribution. Uh so that's the sort of the that's the long-term thesis, is like be part of the core infrastructure of sort of market market price discovery. [00:43:36] S0: That's right. And then just and then just play on sort of the growth in the AI economy and the need for agents to have real-timely private data that's permissioned to help make better decisions. So like as more decisions are being made, powered by data, we just want to capture that. [00:43:53] S1: Yeah. Do you think Claude can be sort of trained to be a good stock picker? Claude specifically. Um sort of the ascended or sort of the the the primary model today. Like maybe it's cle that maybe it's a close weights, maybe you have to sort of do it back in the post-training open open weight models. But I'm curious just your thoughts on LLMs fed with the right data as actually exhibiting investment talent.
00:43:53 · Brett Caughran, Kirk McKeown
[00:43:53] S1: Yeah. Do you think Claude can be sort of trained to be a good stock picker? Claude specifically. Um sort of the ascended or sort of the the the primary model today. Like maybe it's cle that maybe it's a close weights, maybe you have to sort of do it back in the post-training open open weight models. But I'm curious just your thoughts on LLMs fed with the right data as actually exhibiting investment talent. [00:44:18] S0: You know, it's I I would say large model, um, definitely. Uh, I'm not sure if LLM is going to be the best flavor or time series or tabular. Um, there's a there's a lot of conversation around different model types now, and you're starting to see prediction and forecasting models emerge more in tabular and time series, but there are also LLM players that are starting to come through. I do think, you know, sort of um systematic, uh, you know, sort of AI generated workflows uh will become as as as good as you know, strong analysts and PMs over the next couple years. I mean, you know, I think you can I think you can train it. Uh you probably want to keep human in loop. I would I would be more of a recommendation engine than a train than a than a than an actual stock pick engine, but like these these these these machines are pretty strong.
00:44:18 · Kirk McKeown, Brett Caughran
[00:44:18] S0: You know, it's I I would say large model, um, definitely. Uh, I'm not sure if LLM is going to be the best flavor or time series or tabular. Um, there's a there's a lot of conversation around different model types now, and you're starting to see prediction and forecasting models emerge more in tabular and time series, but there are also LLM players that are starting to come through. I do think, you know, sort of um systematic, uh, you know, sort of AI generated workflows uh will become as as as good as you know, strong analysts and PMs over the next couple years. I mean, you know, I think you can I think you can train it. Uh you probably want to keep human in loop. I would I would be more of a recommendation engine than a train than a than a than an actual stock pick engine, but like these these these these machines are pretty strong. [00:45:04] S1: Yeah, yeah. What um for people that want to sort of take a look at CarbonArc and explore how to use it in their investment process, what would be like the you know short Cliff Notes version of like how to think about interfacing with that tool to help drive better uh decision making in their portfolios?
00:45:04 · Brett Caughran, Kirk McKeown
[00:45:04] S1: Yeah, yeah. What um for people that want to sort of take a look at CarbonArc and explore how to use it in their investment process, what would be like the you know short Cliff Notes version of like how to think about interfacing with that tool to help drive better uh decision making in their portfolios? [00:45:24] S0: Yeah. So um, you know, sort of there are 450 discrete uh data assets structured there, about 10 petabytes of data with daily granularity. It's all permission for folks to you know plug in and play around. You can use our lenses product and and sort of our skills, uh, or you know, plug your Claude or you know, ChatGPT or Perplexity directly into it. Um, and you can curate natural language through that. Um, you can do one-shot questions of the different skills we've built on the way we've structured the data, um, or you can pull data out of the API directly and build your own time series models. Um, however you want to interface, we've set it up to be that way. It's all pay for what you consume. And uh, you know, sort of as as you sort of build into sort of uh the skills specifically, you can go in and plug into a brand or location or cohort, um do a one-shot query and get back a full report 20 minutes later that you know, sort of is better than anything I would have written. That's awesome.