Uncapped with Jack Altman
00:27:33 · Jack Altman, Ev Randle
[00:27:33] S0: How do you feel about the closed versus open source stuff? We've talked about this a bunch. And obviously, we are exposed to a lot of the open source world through Fireworks and other things. What we've seen in general, I think, has been a rise in open source this year. At the same time, obviously, the frontier labs have gotten massive. And the price per token is very different. The use cases are similar but different. So anyway, what's your latest on open, closed stuff? [00:28:09] S1: Yeah. Mean, if if you'll if you'll entertain me, I I I feel like we probably have a similar synthesis of what we feel now. But I think what would be interesting to do is, like, maybe we steel man each side of the argument. Great. And then like that... But I think once you have the steel man for each argument... Because I think there's a sort of like a lot of like... There's a lot of talk online about open source versus closed source, which I think is sort of off track Right. From the reality on the ground. So I think if you steel man both sides, you can actually understand, okay, well, what's the synthesis that probably blends the two steel man? So I'll do closed source. You do open source. [00:28:44] S0: You go first.
00:28:45 · Ev Randle
[00:28:45] S1: So the steel man of the closed source argument... And what we mean by this is why would a company get inference from a closed source frontier API rather than from an open source API inference provider like a Fireworks using Kimi K3 or any of these other models, US or Chinese open models. And people say, well, look, intelligence is fungible, or it's sort of a commodity. And if we take that seriously, that idea that a unit of intelligence is this fungible thing, There's a lot of history and a lot of good literature about who wins those types of markets and... When you're when you're optimizing for the cost of a fungible output. And it's... Oftentimes, the winner is something we call a low cost producer. And so the defining feature of one of these low cost producers is that you are structurally, as a company, you have a lower cost to deliver a comparable outcome. And so you can either charge lower margins, or you can charge lower prices at the same margin as competitors, or you can earn higher margins than your competitors at the same price. And so two famous examples in tech. SpaceX is probably the easiest one to understand. SpaceX has the lowest cost to them in order to launch a kilogram into space Yep. In the world. No one can launch a kilogram into space as cheaply as SpaceX can, and that gives them a pseudomonocaly on launch, at least at at the same level... Unit level of of capability. A previous one in computing is AWS. They had massive aggregated demand that created economies of scale. They were able to amortize a lot of investments over a massive customer base, which gave them lower cost relative to competitors in the CPU cloud business. So for AI, the equivalent here would be how do you become the low cost producer of a unit of intelligence? So we could say this is like a given level of capability, latency, context length, and reliability for some task. Like, who can produce that task at a given level of capability the lowest, reliably? And I think proponents of the closed source argument would say the leading frontier labs have the best models in the world, many of which are not released to the public or not released via API, and they're using those models and pointing them at their infrastructure to improve their serving efficiency and co design their serving infrastructure with the actual models that they're training.
00:31:09 · Ev Randle
They have the most aggregated demand. They're the largest, and so they have economies of scale. The data center space that they're buying is cheaper for them on a per unit basis than anyone else because they're the largest purchasers of compute. And if all of these are true, if you have economies of scale, if we are in some weak form of RSI right now where the models are actually designing actual inference efficiencies at any rate that is compounding, that means that you can have a low cost producer of intelligence in the form of an AI lab. And that means that this argument that open source equals cheaper is a misnomer and is not necessarily true, because at the same level of margin, if this was true, a FrontierLab at any point of the intelligence curve, like the capability curve, would be cheaper than a inference provider serving on an open source system. So that's sort of the economic argument for closed source, and then I'll let you take open source.
00:32:08 · Jack Altman, Ev Randle
[00:32:08] S0: Okay. So I mean, there's... I'll start with the economic piece. And obviously, as you know, I mostly mostly buy it. But economically, one difference is that because the... You know, to... There's the model cost, and then you gotta do the inference on it. The model, in theory, has some training that's been amortized into the price. And so to the extent that the labs want to be profitable or have enough margin that they make up for it, they do need to do something to make up for the training costs versus if you're using Kimi. You don't. The model is free. So there's no amortization there. And then, obviously, you still need to run the inference. And it's probably right that on a per inference unit basis, the labs are still going to have the advantage over the inference companies. But maybe in theory, that's made up. The other thing is, do the models care to take their margins down to that level? Is that actually worth it, or is that really expensive for them to do? And so [00:33:09] S1: Then they don't have as much money to train.
00:33:11 · Jack Altman
[00:33:11] S0: They don't have as much money to train. And also, it's like if they're really focused on the real frontier, and they're like, you know what? Use your open source for the taxes. Use it for recommending which iPhone you should buy and should you get the Duo or not. And we're going go solve biology. We're going go do robotics. And it's just not worth it to go optimize two models ago to go do the basic stuff. And so there's some there's basically some argument along the lines of we're going to get saturation on many tasks. And as open source of the future is better than OPUS 5.5 and Astra are today, that's going to be good enough for most tasks. And the frontier labs are not going to care to go reduce their margins, optimize two models ago. They just need to keep their eyes on the frontier. And so maybe there will be some space for that. But what I would say probably is more than the money pieces. And I wish I actually had because to your point, the specific numbers on the margin are where it all turns on. So I don't have those on hand. But I think more important is, do you want to own your own intelligence? And this is like the Alex Karp stuff of just like, you cannot trust as an enterprise, you cannot trust any company to own all of your data and to be fully reliant on any company. And you need to own your own models. You need to own your intelligence. And I think there is an argument to be made there. And I think for certain industries, that trade is going to be worth it. And particularly as the open source models get good enough, even if you're not saving any money, you're in control. And that's got something.
00:34:52 · Ev Randle, Jack Altman
[00:34:52] S1: Then even think about the branding of let's say all your tokens are Anthropic tokens, and all of a sudden, Anthropic enters your space. And then your customers are like, well, what model are you using? You're like, 100% of our models that we're using is Anthropic. We're like, well, then why don't we just go to Anthropic? Yeah. If you're Eli Lilly no special sauce. [00:35:12] S0: If you're Eli Lilly and all of your data lives in Anthropic [00:35:15] S1: Yeah. [00:35:16] S0: And then Anthropic's like, hey. By the way, BTW. We got some drugs coming. You're like, Okay, well, don't feel so good now. So I think there are some reasons that people won't want to do that. And then I think there's situations where you can post train on open source, and that can be useful. And if you've got a lot of data, that data can outperform the extra intelligence that you might get from the next model. And so you'll see some of that, too.