Ce Zhang: evaluation

Transcript context

…Yeah, so that's the biggest- It's a big investment. Yeah. Okay, well, I mean, it looks like it's paying off, so. And then high level, I will confess or admit or mention for the listeners who are also similarly skeptical, I did not used to care about long contexts because I was like, you know, 30K is enough, 100K is enough, right? I'm not, you know, modeling DNA sequences or anything like that. Why do I need long context? And I mean, first of all, I'll throw that open to you. But second of all, I think what Mamba did for me was change that perception of that. It's only about a long context. The only reason you want sub-quadratic architectures is for long context. Actually, that's not true. And it's also just more efficient to train, period. Right? I'll just leave that open to you. Like what's the motivation that people should keep in their heads? There are multiple things, right? So one thing is that, I mean, the moment a model can do for long context well, so it often means that it's kind of cheaper. Yeah, so I mean, that's why it's kind of long. I mean, in principle, transformer can do long context. It's just very expensive. So I think what those like state-based models trying to do is try to push the size of the state, right? Like as small as possible. That's why it's kind of long context, right? And try to kind of like decouple this like quadratical dependency, right? To make sure you can have a much better execution pattern. One direct consequence of those is you can do long context really cheaply, but on the other hand, also introduce a whole bunch of benefit even you are not doing long context. Right? So I think that's actually probably equally important. Because data gets smaller, you can do really large batch size, right? You can actually be very faster. Right? So yeah. And another thing is like, one of the hypothesis that we have is, like in Stripe Hyena, it start to have a hybrid architecture, right? It has part of it has like state-based model and part of it is still the transformer. So different component probably deal with different things kind of better. So maybe by putting them together, by thinking about how information propagate, over this whole horizon of this context, you can probably get an even better quality model than transformer. Right? So I think that's why we are kind of invest a lot of things, on those models. Not only for the context, which is very important, but also for a whole bunch of benefit it could get. Yeah. How should people treat the distinction between Mamba and Stripe Hyena? Like what's the point of releasing these two as separate models? Is one like sort of the together proprietary one and then the other is like the more open research one?…

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