Shawn Wang: evaluation

Transcript context

…Yeah. No, that's great. I mean, I think in the start small just meant cheap. I think today small is actually a more nuanced discussion, you know, that people weren't really having before. Yeah, we can keep going. This is a slide that I smiley disagree with Sarah. She's pointing to the scale SEAL leaderboard. I think the Researchers that I talked with at NeurIPS were kind of positive on this because basically you need private test [00:14:00] sets to prevent contamination. And Scale is one of maybe three or four people this year that has really made an effort in doing a credible private test set leaderboard. Llama405B does well compared to Gemini and GPT 40. And I think that's good. I would say that. You know, it's good to have an open model that is that big, that does well on those metrics. But anyone putting 405B in production will tell you, if you scroll down a little bit to the artificial analysis numbers, that it is very slow and very expensive to infer. Um, it doesn't even fit on like one node. of, uh, of H100s. Cerebras will be happy to tell you they can serve 4 or 5B on their super large chips. But, um, you know, if you need to do anything custom to it, you're still kind of constrained. So, is 4 or 5B really that relevant? Like, I think most people are basically saying that they only use 4 or 5B as a teacher model to distill down to something. Even Meta is doing it. So with Lama 3. [00:15:00] 3 launched, they only launched the 70B because they use 4 or 5B to distill the 70B. So I don't know if like open source is keeping up. I think they're the, the open source industrial complex is very invested in telling you that the, if the gap is narrowing, I kind of disagree. I think that the gap is widening with O1. I think there are very, very smart people trying to narrow that gap and they should. I really wish them success, but you cannot use a chart that is nearing 100 in your saturation chart. And look, the distance between open source and closed source is narrowing. Of course it's going to narrow because you're near 100. This is stupid. But in metrics that matter, is open source narrowing? Probably not for O1 for a while. And it's really up to the open source guys to figure out if they can match O1 or not. I think inference time compute is bad for open source just because, you know, Doc can donate the flops at training time, but he cannot donate the flops at inference time. So it's really hard to like actually keep up on that axis. Big, big business [00:16:00] model shift. So I don't know what that means for the GPU clouds. I don't know what that means for the hyperscalers, but obviously the big labs have a lot of advantage. Because, like, it's not a static artifact that you're putting the compute in. You're kind of doing that still, but then you're putting a lot of computed inference too.…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.