Latent Space: The AI Engineer Podcast
00:51:04 · Shawn Wang, Jamin Ball
[00:51:04] S1: Like, all my zero days are gone. Like, I I had them for ten years. Like Yes. Whatever. Yes. But anyway, so so these are fun anecdotes or consequences of this. If you want to really do super Straussian and Machiavellian, if you're a nation state actor, what do you do to to to preserve your edge? Because, like, you cannot have the channels that you're relying on lock up. Yeah. Then you then then you send a Saudi decoy Yeah. To put the thing behind your TV. That's crazy. So that people... So that get distracted by that, so that your real spy Yeah. Gets in. [00:51:40] S0: That's just some war games. [00:51:42] S1: Yeah. I I I watch too much TV to to to think that what we are told is the full story. Yeah. Yeah. [00:51:48] S0: I think that's fair. Right? I think that's fair. Too much is at stake. [00:51:51] S1: Yeah. It's too much at stake. Okay. So coming back to the actual, again, finance, like, I really want to be a primer for everyone Yes. Who are are new to the economics of models and model labs and what what have you. One thing that have... That people have said is that the R and D per model. Let's say, like, what does what does Fable five... Sonnet five Opus five cost? Mhmm. Let's call it a billion dollars. [00:52:15] S0: Mhmm.
00:52:15 · Jamin Ball, Shawn Wang
[00:52:15] S0: Mhmm. [00:52:15] S1: That lasted me six months. Mhmm. Right? Because you just said the dep... The depreciation, it lasted last six months. Yeah. Yeah. And so isn't that worrying? Right? Like, about, like, well, these things are the fastest depreciating assets in history. Yeah. This is something that me and Saragol have have talked about before. Yeah. Yeah. Yeah. Yeah. And I think the the net is there's no, but, like, it's a qualified no Mhmm. Of, well, as long as my next model run is better than the than the than the current state of the art, we're good. Yeah. Yep. Yep. But it's a little bit of a house of cards if if the next one doesn't work out. Yeah. Which briefly affected OpenAI. Yeah. So a lot of people know by now that, you know, like, GPT four five, the the big the big run wasn't wasn't that successful. Yep. And that they, you know, they... They've they've since since fixed it. But, like, that slump Yeah. In a in a run or a failed run Yep. Can really, really screw with the the ROI on the thing. [00:53:08] S0: And look. I think there's a really good analogy to the trip industry. And I remember speaking to someone much older and much smarter than myself who did a lot of semiconductor investing in the eighties and nineties about all these chip companies. [00:53:22] S1: They've seen it all.
00:53:22 · Shawn Wang, Jamin Ball
[00:53:22] S1: They've seen it all. [00:53:22] S0: And they've seen it all, and they were highly skeptical. And sometimes you have industry vets that are highly skeptical, and they miss the next big opportunity because they never think the incumbent can be disrupted. But the point of view they had was it's very different to tape out a chip that is competitive with the leading chips at the time from building a machine that can continuously tape out chips that compete with, know, [00:53:49] S1: It's the Marvel Marvel factory. [00:53:50] S0: Yeah. The chip factory. And so he was like, this individual's point of view was maybe a chip startup can build one chip, but guess what? They're not gonna have the machine to constantly pump out a new chip every year and stay at the frontier from the chip standpoint. And that's why NVIDIA's king. It's not because they have the best chip now, it's they have the best factory. Process for chips. Exactly. Exactly. [00:54:09] S1: It's the machine that builds the machine that
00:54:09 · Shawn Wang, Jamin Ball
[00:54:09] S1: It's the machine that builds the machine that [00:54:11] S0: And so it's kind of one of these... I mean, this is the bear case on all these Neo Labs, which is, you and and you talk to many smart folks, what they might tell you is, hey. If you wanna really compete and be... Have like a soda model at the frontier, you can't you can't you can't go more than one order of magnitude of compute less than the labs. You just can't compete if you don't have, you know, one order of magnitude of compute of the big labs. And so as a neolab, you know, the question isn't can you build one model? It's can you attract compute and talent required to continuously pump out new generations of models every year to ultimately compete in your domain? And so I think for the labs, the reality is is I think it's very hard [00:54:55] S1: for
00:54:55 · Jamin Ball
[00:54:55] S0: anyone, unknown players, to catch up to kind of like the generalized frontier big models. I think you'll continue to see domain specific models or models or like new types of models, decision models like JEV. But I think the think the big labs, they are on the treadmill of continuously pumping out new models. Because guess what? If if Anthropic stops, OpenAI and Gemini will continue. And then you'll just kind of see revenue... You'll see revenue fall off, which is why I do think it's so important to stay at the frontier. Now what do I expect? And like, what have I seen from Anthropic now in the data? It's as their model families get better and better, they get more efficient, but they also capture more revenue. That could be because they're doing more economically valuable work or they're doing the work faster. But what you're seeing from now kind of black and white is, hey, as Anthropic has 2x'd their compute spend to build and train these models, they've 14x'd their revenue. And I'm not really sure what would cause that trend to stop. I think models are only gonna take on longer running tasks, more economically valuable tasks. And if you can't tell, I'm a hyper optimistic person, and so maybe I'm just drinking the Kool Aid too much. But I I do think we're nowhere close to hitting a limit of the dollar kind of like economic value that leading models can capture. I think what's maybe more interesting is, you know, OpenAI, you know, Luna class models, kind of also trying to capture and compete at the low end of the market with much smaller, much cheaper models, you know, and kind of span everything from Frontier to
00:56:49 · Shawn Wang, Jamin Ball
[00:56:49] S1: Small. [00:56:50] S0: Yeah. Distilled versions of their own models that are smaller and and, you know, cents per million in put output tokens, not dollars. [00:56:58] S1: Yep. Cosign all of that. I think that is a good way to think about it. I I I think, you know, fun... I'll just comment this because it's in my head. Six one Terra has been pretty pretty successful. Six one Luna. Like, that that whole series is very interesting. The approaches of the different labs is also super interesting. Anthropic to this day has no decision model. Mhmm. And they're like, well, we're good. Like, we we use TypeSave. Whatever. Yeah. Yeah. Yeah. And, like, well, they don't have a stand alone voice API. Right. They they they have a voice model. They just don't don't market it. Right. They don't... They've never done images. Right. Know, it's it's a it's a lot of enterprise focus and coding focus. Yes. And OpenAI has been a little bit more of, like, we'll do everything and Yeah. We'll try Sora. Shut down Sora. Yes. Remember when Sora was, like, the number one app [00:57:41] S0: in Yeah. Exactly. [00:57:42] S1: In The US. [00:57:43] S0: Yeah. Exactly.