etn.
00:49:46 · Fabian, Kosta, Sam, James Lowe, S5
[00:49:46] S6: Thank you. Appreciate it. [00:49:47] S0: It's all for fun, guys. Really appreciate it. Can I bring [00:49:49] S1: Anish to the show? And then we're here. [00:49:52] S0: And then we wrap. Thank you very much. See you soon. Thank you so much. We love to do it again. [00:49:56] S4: Of course. Legend. [00:49:57] S0: And I love the shoes. So cool. [00:50:19] S1: Great story. [00:50:19] S4: That's a story for the show. [00:50:21] S6: To see you. [00:50:22] S0: Here he is. How are we, sir? [00:50:23] S1: I love this jumper. [00:50:24] S0: Hey, dude. [00:50:25] S4: We're just in [00:50:25] S0: today. Style [00:50:26] S1: pod by the a16z team right now. Although we do have the we do have the AC. [00:50:31] S6: To be [00:50:32] S0: fair, yeah, we're actually starting but We're in the [00:50:34] S1: Yeah. Yeah. [00:50:34] S0: Dude, the timeline loved this. I've been telling everyone I threw a photo up last night. I woke up, and it has a 155,000 views. And everyone's everyone's lovely. Everyone's saying, where's Gabe, actually? So Gabe's fans. He's everywhere, man. [00:50:46] S1: Yeah. I'm sure he's on a plane somewhere. [00:50:51] S0: Percent. Five in the next trillion dollar founder. [00:50:53] S5: He's he's just got a printer in his backpack. He's printing the turbinates out one at a time.
00:50:53 · S5, Kosta, William Loveday Powell, Fabian
[00:50:53] S5: He's he's just got a printer in his backpack. He's printing the turbinates out one at a time. [00:50:57] S0: Yeah. Just... He's probably throwing them out the airplane at this point. Very good. Thank you so much for joining us, Anish. It's always a pleasure to sit down with you. I'd love to start by just asking, I would love to get your take on the market right now in general. Just and I'm not just talking about like capital markets. I'm just talking about consensus around AI. We're seeing the market is reacting to general things that are happening in the space. I think we saw rumors of maybe Anthropic slashing their valuation. Saw OpenAI cut their funding round. [00:51:26] S2: By [00:51:26] S0: 10 bill. By 10 bill. I'm seeing takes online around. Maybe this is a slight correction, but we are only 1% into the journey of what can possibly happen here or there. I'd love to get your thesis, just high level. Where do think things are at? [00:51:39] S5: Yeah, it's interesting to sort of frame this against what was happening five years ago, you know, because if the closer you look with this product cycle, the more substantial all of this stuff is. [00:51:48] S6: Mhmm.
00:51:48 · Fabian, S5, Sam
[00:51:48] S6: Mhmm. [00:51:48] S5: Whereas I think five years ago when you... You didn't wanna look too closely. You know? Right? It was like, we had specs. We had all of this sort of distortions that felt like they were transient. And here, actually, the sustainability and the sort of substantiality feels really, really high. So I think that there's gonna be this choppiness and noise, and I would not wanna have to be my job be, you know, predicting memory prices in three months. Like, I think that's really difficult. But, certainly, I think the trend that's occurring is irreversible. And there's actually two huge breakthroughs that have happened in the last few months that I think are under discussed. You know, one, of course, is everything that happened with agent swarms or, as Sam would prefer, agent teams. [00:52:24] S1: We have this written down a lot, mate. [00:52:26] S5: Yeah. I think that, like, there's a white hat version of that that's super interesting. What Michele sort of called another scaling law last night. And then the other thing is this paper that was released on Monday about having models edit the weights of other models Yes. Which gives you continuous learning opportunity. And it was released, and there was enthusiasm, and then it was criticized. But I think it's a bigger deal than people appreciated.
00:52:26 · S5, Kosta
[00:52:26] S5: Yeah. I think that, like, there's a white hat version of that that's super interesting. What Michele sort of called another scaling law last night. And then the other thing is this paper that was released on Monday about having models edit the weights of other models Yes. Which gives you continuous learning opportunity. And it was released, and there was enthusiasm, and then it was criticized. But I think it's a bigger deal than people appreciated. [00:52:49] S0: Let's talk about that. We Jacob Lauritzen on, the CTO of Legora. He did a post around how he basically one shotted an internal use case for them. It did, 800,000 views. This was off the back of the Spaceway programming paper that's published a couple days ago. Right. Is is it hype? Is it is it is it a new means in which we can scale this thing? Like, what what do we [00:53:07] S5: I I think so. The smartest people I've been talking to have been able to not just replicate, but extend it. Now, look, there may be some constraints, like, may be opposing ideas and models can't hold both of them in their head. Like, you can't be highly conscientious and highly open. Mhmm. Right? Because that sort of implies that the same weights are pointed in different directions, but you may be able to do things like get models that are able to have high entropy and non verifiable domains, beautiful art and prose, but then high precision and verifiable domains where they can also write great code and be mathematical. Yeah. And we've never been able to kind of, reconcile those trade offs before.
00:53:07 · S5, Sam, Kosta
[00:53:07] S5: I I think so. The smartest people I've been talking to have been able to not just replicate, but extend it. Now, look, there may be some constraints, like, may be opposing ideas and models can't hold both of them in their head. Like, you can't be highly conscientious and highly open. Mhmm. Right? Because that sort of implies that the same weights are pointed in different directions, but you may be able to do things like get models that are able to have high entropy and non verifiable domains, beautiful art and prose, but then high precision and verifiable domains where they can also write great code and be mathematical. Yeah. And we've never been able to kind of, reconcile those trade offs before. [00:53:40] S1: What happens then to, if you look at the whole stack in terms of frontier models, open weights, all of that kind of stuff, like where does that leave us in terms of the frontier? What does that mean for the kind of you know, who wins, I guess? Where is the value being accrued in that right? [00:53:58] S5: Mean, I don't think it's zero sum, right? The frontier, basically, we've got a somewhat, you know, maybe a duopoly, though there's a few other, obviously, sort of Google's been growing in fits and starts, and SpaceX, I think, is actually more recently relevant. But they've maintained pricing power for their Frontier models. Right? And in fact, Mythos, the the legendary Mythos, the too dangerous to release model [00:54:17] S0: Europeans hate Mythos.
00:54:17 · Kosta, S5, Sam
[00:54:17] S0: Europeans hate Mythos. [00:54:19] S5: Very funny. But, like, there's a lot of aura associated with it, and it's essentially infinity dollars per token. Right? You can't even access it same with Daybreak from OAI. So despite that we've seen these really capable open weight models, Frontier is still maintaining pricing power, and we're just seeing the whole Pareto curve get filled out. [00:54:35] S1: Okay. That's really interesting that you mentioned aura in terms of, like, the fact that the, defensibility is because of their aura kind of thing. Why is that? Yeah. I guess like was that a... Do you think that was like a deliberate strategy just to garner aura through mythos? Like what? Yeah, what what is that? And and can that hold as a moat, I guess, for the Frontier Labs? [00:54:58] S5: Well, it is... It... It's very funny. Because the reason I think this is happening is if you look at the total buyer set for all of the AI spent today, which is, you know, probably a $175,000,000,000 conservatively, it's 10,000,000 people. Yeah. And all of them are on x. So it's very funny because it's sort of... A lot of the choices that are being made, these huge capital allocation decisions are kind of downstream of x and what those conversations are. And I think, you know, many people on x, like all of us, are hyper AI pilled. So this, like, idea, this fantasy that this model, there's some crazy intelligence that is being held back from us. [00:55:31] S1: Yeah.