All-In with Chamath, Jason, Sacks & Friedberg
00:52:47 · David Sacks, Jason Calacanis, Chamath Palihapitiya
[00:52:47] S1: China. I'm confused. [00:52:48] S0: Yeah. [00:52:49] S1: Haven't you been saying for like the past year that open models are gonna eat Anthropic's lunch? Well Have you changed your tune on that? [00:52:56] S0: In the face of new evidence, you would have to say that they are lowering the prices and the opportunity in the enterprise is is so great that they're not being dramatically impacted by it. You will see anecdotally startups and enterprises exploring it, but it's obviously not enough now that we know their numbers that it's slowing them down. It's not slowing them down yet. Not debating revenue traction. It's still the biggest risk factor, but if you're [00:53:28] S2: I'm not debating revenue traction at all or business quality. What I'm saying is there are moments where there's systemic risk in markets where everybody goes risk off. And I think that what I am saying very explicitly is I've not seen in the last fifteen or twenty years the last time we came close to this was in the GFC, where you have a narrow part of the economy represent basically all of everything that's happening in the capital markets. And all that means is that everybody has to pay a lot more attention. It doesn't matter what side of AI trade you're on, whether you're long it, whether you're short it, I don't think that's important. I'm saying something more structural, which is about 10 companies collectively to fund one thing will be competitive with global governments on borrowing. I don't know what the implications of that is because I have not seen that.
00:53:28 · Chamath Palihapitiya, Jason Calacanis, David Sacks
[00:53:28] S2: I'm not debating revenue traction at all or business quality. What I'm saying is there are moments where there's systemic risk in markets where everybody goes risk off. And I think that what I am saying very explicitly is I've not seen in the last fifteen or twenty years the last time we came close to this was in the GFC, where you have a narrow part of the economy represent basically all of everything that's happening in the capital markets. And all that means is that everybody has to pay a lot more attention. It doesn't matter what side of AI trade you're on, whether you're long it, whether you're short it, I don't think that's important. I'm saying something more structural, which is about 10 companies collectively to fund one thing will be competitive with global governments on borrowing. I don't know what the implications of that is because I have not seen that. [00:54:17] S0: To the point, Sacks, that you made, in challenging me, which is great, their revenue is slowing down dramatically. You know, their quarter their over quarter revenue is decelerating. So, obviously, it's on a [00:54:28] S1: big number. So We don't know q three yet. I mean, the scuttlebutt on x, and who knows if we can totally believe this, is that things did slow down in q three a lot, but some of that may be attributable to the fact that they eliminated 99% of distillation. Right? So they voluntarily took a big hit to the revenue [00:54:45] S0: From people stealing from them. Yeah.
00:54:45 · Jason Calacanis, David Sacks, Chamath Palihapitiya
[00:54:45] S0: From people stealing from them. Yeah. [00:54:46] S1: Yeah. To prevent the copycat models from from basically distilling them. So there was that. And then I think the scuttlebutt is that since they cut their token costs considerably, which was done pretty recently, now the demand has reaccelerated. So Yes. [00:55:03] S0: Jeb and Sacks in full effect. Yeah. [00:55:06] S1: Yes. It seems like they're back to resuming their previous growth after a couple of hiccups in q three. But I think the thing that's really impressive is just that the way they're able to cut their token costs, it's not just because they're giving up pricing power. Right? It's not that they're losing margin. They're actually gaining margin. The reason they're able to do that is because all the algorithmic improvements [00:55:29] S2: Efficient. [00:55:30] S1: They're making in order to make their models much more efficient. [00:55:33] S0: Yes. [00:55:34] S1: And Now here... Yeah.
00:55:34 · David Sacks, Jason Calacanis
[00:55:34] S1: And Now here... Yeah. [00:55:36] S0: Go ahead. Just just to dovetail with that, Sacks. What's happening with the models is even more impressive. The... They're lowering the price of tier two and tier three models. Right? Your your Sonnets, your Opus. Those are getting cheaper, but enterprises are saying that's enough for me. So as they make that cheaper, then businesses get more value, they solve more problems out of it, and they don't feel the need to be on the latest greatest fable. If you're all hollow, maybe you need that to do science. If you're a mathematician, you need it. If they can keep making their second and third run models this much cheaper, 10 x cheaper each year, the Jevons paradox is gonna get even more pronounced because enterprises will use it more. Token cost matters. Right? We're sitting here with the token max and cost.
00:56:19 · David Sacks, Jason Calacanis
[00:56:19] S1: What we're is that... I mean, we we keep seeing Jevons paradox play out at every level, which is as you cut the cost of tokens or more and more use cases, enterprises use it more and more and the the net result is that the total demand just keeps going up. And the total revenue and market size keeps growing. Also, just to reiterate a point that you made earlier in the program, the new applications that are coming all consume a lot more tokens than the ones that we started with, right? Like agents use a lot more tokens than prompting. And cyber defense is gonna use a lot more tokens than just doing, you know, a pen test once a year, right? You're gonna need continuous monitoring of agents, attack surfaces, all that kind of thing. So all the new applications are gonna be using exponentially more [00:57:08] S0: 50 to a 100 number of tokens. Yes. [00:57:11] S1: Exactly. So the price will keep going down, but the market size keeps increasing. I think the biggest risk factor to Anthropic, I do think it is some version of the open source models, maybe not catching them, but sort of saturating intelligence at a low price point for [00:57:27] S0: Commoditizing it.
00:57:27 · Jason Calacanis, David Sacks
[00:57:27] S0: Commoditizing it. [00:57:28] S1: Commoditizing it for things like coding or, you know, where there is a big chunk of the market, that is still a risk. I would say sort of a second order risk is that in order for Anthropic to stay ahead and to keep making their models more and more efficient, they have to keep using more recursive self improvement, right? You see this in the charts that they published, that they are using their latest models more and more to discover these compute multipliers, to discover these new algorithmic efficiencies. So they have to kind of push the limits of what they themselves call risky activity. I don't know if I fully believe in the Doomer story about RSI, but they themselves say that RSI is very risky, but they have to keep pushing RSI in order to keep delivering these algorithmic improvements in order to stay ahead of the open models. And I just wonder how sustainable that is. At some point, do you get a breach with the doomers that becomes irreconcilable? Do the doomers want to basically shut down RSI or basically slow it down to the point where then the open models truly do catch up? I don't know the answer to that question. It's sort of just a more nuanced take on this idea of open models catching up.