00:49:40 · Apoorva Mehta, John Coogan, Jordi Hays
[00:49:40] S3: That's right. I mean, you know, if you think about it, like, Instacart is the original agentic company. Yeah. Right? You literally do send an agent to go to the grocery store for you. Right? It is just a human agent. It is a human chopper who's doing that for you. Right? And... But yeah, I agree. I think that there's, you know, a lot of improvement to be made in the overall interface and I think that companies like Instacart that have this, you know, you know, they've solved this extremely hard problem underneath can build those interfaces and, yeah, very excited about about seeing what that trend looks like. [00:50:18] S1: Do you think we need a new word to differentiate between an agent that you task individually? You give it a task, it goes off and does it and it comes back and it's sort of an asynchronous process that comes back to you versus something that you design that runs continuously forever. I mean, people kind of throw around continuously running agent but when I think about... [00:50:44] S0: Or yeah, proactive. [00:50:46] S1: Yeah. Like there are agentic systems that will like go push order order the burrito button. [00:50:51] S3: Sure. [00:50:51] S1: And that's a lot less interesting than when do you get the agent that actually knows where you are, what you're eating, what you like to eat and Yeah. And and is just basically handling your lunch order every day on an ongoing basis. And maybe there's Yeah. [00:51:07] S0: Some Yeah. Like I'm I'm sure you guys thought about this like putting sensors and fridges to Yeah. When
00:51:07 · Jordi Hays, John Coogan
[00:51:07] S0: Some Yeah. Like I'm I'm sure you guys thought about this like putting sensors and fridges to Yeah. When [00:51:12] S1: You're out of milk. You're out [00:51:13] S0: of something and it automatically orders it. But but a lot of these efforts haven't... The hardware side hasn't really worked.
00:51:21 · Apoorva Mehta
[00:51:21] S3: I... Look, I I don't wanna comment too much on Instacart. I'm not... I'm no longer the the the CEO, and I think we have an incredible team there. But, like, look, like, you know, I think about agents all day, you know, for my new company. You know, we are... We are... We do have, you know, very, very, very long running agents that are Yeah. You know, continuously, you know, looking for information, getting better and, you know, in doing that in a way that is, you know, not only token efficient, right? But also, you know, in a way that is, you know, where every single day that you run it, you realize that it is just better. You're getting better results for lower cost, lower time. And so, you know, the way I think about this is that, like, you know, we're going to have lots and lots of long running agents in our lives and, you know, if you think about, like, the... What the the analog is for for humans is that, you know, when you and you onboard an employee, right, and, like, they now have... They... You know, you're onboarding them with all the context that they they they need and then over time, as they get more and more content, as they as they do more and more projects with you, they get smarter and smarter, they get better and better at that job, right? But... So, you know, if you think about that, that is effectively, you know, what we're thinking about when we think about a continuously running ongoing agent, right? And so, you know, know, hardware, things like that are just ways to plug in more context for that agent. And so, we're going see that in the consumer world but I think we're going to see that far more in the enterprise world because there, you know, you know, you're just willing to pay more for, you know, faster information, better information that gives you, you know, a massive competitive advantage as a as a as a company. You know, as consumers, I don't think anyone's really figured out how to charge people for this. Right? Like, I think, you know, we saw that Elon tweet recently where he was like, the revenue didn't make sense for, you know, the the
00:53:32 · John Coogan, Apoorva Mehta
[00:53:32] S1: Rock bottom. [00:53:33] S3: You know, for for the margin that we were spending on it or something like that. So I do think this is going to be something that we're going to see over the next few years. That gets solved. But so far, it's far more interesting in the enterprise setting. [00:53:48] S1: Yeah. Yep. How are you thinking about agentic trading AI hedge fund activity spreading among other institutions that have frontier model capabilities? Like Google for a long time, I mean, they're spending a lot of cash but they've had a huge balance sheet And there's a pretty small treasury team, I believe, that moved that money around. They've always had the opportunity to do something more aggressive or differentiated with that. Do you think that that will take off broadly? Is this gonna be a new trend? Or would they partner with a fund on something? Do you think that that plays out as the capability becomes more diffused? [00:54:26] S3: Yeah. I mean, you know, so this is what's new. Right? This is what's new about, you know, AI based trading. [00:54:36] S1: Mhmm. [00:54:36] S3: Right? The reality is that, like, you know, have people been using machine learning models you know, algorithms to trade before? Yes. Right? Like that's not new, right? There's like, you know, Renaissance exists, for example, right? Like, you know, but the the you know, historically, the primary way they've been doing that is, you know, more what's called stat arb. [00:55:02] S1: Mhmm.
00:55:02 · John Coogan, Apoorva Mehta
[00:55:02] S1: Mhmm. [00:55:02] S3: Right? This is more, you know, understanding, you know, statistical, you know Patterns. Patterns, right? And, you know, basically making a call based on those patterns to say, hey, you know, these patterns are going to repeat in somewhere or not, right? And that's primarily the way ML and AI has been, you know, used. But for the first time, you can have, you know, reasoning, right? Like, know, what Warren Buffett would do, right? When they're analyzing a company, they're saying, hey, what are the moats? Right? How strong are these moats? Like, how like Mhmm. How durable is this growth, etcetera, etcetera. Like, these are very complex questions that are not, you know you know, you know, they're not necessarily, you know, stat arb. They're much more fundamental, you know, research, right? And and that's what AI trading and and, you know, AI agentic, you know, you know, capital allocation allows for. Yeah. And and so now, your question is, okay, well, is Google going to do it? I'm not sure. Uh-huh. These things are still exceptionally hard to build. Mhmm. And whatever you think is simple, right, chances are those things just get priced in and that that alpha just decays very fast.
00:56:22 · Jordi Hays, Apoorva Mehta
[00:56:22] S0: Yeah. So so are you thinking about hardware at abundance? Because I imagine you'll see the same... Like, if you're now introducing, like, reasoning into this sort of algorithmic trading, you might see the same sort of arms race that that we saw in sort of traditional high frequency trading playing out with this sort of new iteration as in like, if I can basically... If I if I can get tokens like faster and like something happens in the market and my... And I just have faster chip, I can probably get a better entry or exit or or whatever. And so do you think that's something that, like, you're gonna need to have, like, compute on prem? Or is that, like, kind of would be an adjacent strategy? [00:57:07] S3: Yeah. So there's several different strategies, you know, to... There are actually probably like several 100 different strategies here to to to play. And now, you know, maybe at at a high level, right, when I was leaving Instacart, there's sort of, you know, two things that I wanted to do, you know, in my next company. Number one, I wanted to be... To have the potential to be a trillion dollar company. And, you know, I mean, obviously, it's been inspiring to see what anthropical is doing. Five... Sound effect for [00:57:40] S0: Sorry. The trillion dollar