Alessio Fanelli: uncertainty

Transcript context

…Albert Gu, yeah. Yeah, so what should we comment about it? I mean, I think it's useful, interesting, but at the same time, both of these are supposed to do really good scaling for long context. And then Gemini comes out and goes like, yeah, we don't need it. Yeah. No, that's the risk. So, yeah. I was gonna say, maybe it's not here, but I don't know if we want to talk about diffusion transformers as like in the alt architectures, just because of Zora. One thing, yeah, so, so, you know, this came from the Jan recap, which, and diffusion transformers were not really a discussion, and then, obviously, they blow up in February. Yeah. I don't think they're, it's a mixed architecture in the same way that Stripe Tiena is mixed there's just different layers taking different approaches. Also I think another one that I maybe didn't call out here, I think because it happened in February, was hourglass diffusion from stability. But also, you know, another form of mixed architecture. So I guess that is interesting. I don't have much commentary on that, I just think, like, we will try to evolve these things, and maybe one of these architectures will stick and scale, it seems like diffusion transformers is going to be good for anything generative, you know, multi modal. We don't see anything where diffusion is applied to text yet, and that's the wild card for this category. Yeah, I mean, I think I still hold out hope for let's just call it sub quadratic LLMs. I think that a lot of discussion this month actually was also centered around this concept that People always say, oh, like, transformers don't scale because attention is quadratic in the sequence length. Yeah, but, you know, attention actually is a very small part of the actual compute that is being spent, especially in inference. And this is the reason why, you know, when you multiply, when you, when you, when you jump up in terms of the, the model size in GPT 4 from like, you know, 38k to like 32k, you don't also get like a 16 times increase in your, in your performance. And this is also why you don't get like a million times increase in your, in your latency when you throw a million tokens into Gemini. Like people have figured out tricks around it or it's just not that significant as a term, as a part of the overall compute. So there's a lot of challenges to this thing working. It's really interesting how like, how hyped people are about this versus I don't know if it works. You know, it's exactly gonna, gonna work. And then there's also this, this idea of retention over long context. Like, even though you have context utilization, like, the amount of, the amount you can remember is interesting. Because I've had people criticize both Mamba and RWKV because they're kind of, like, RNN ish in the sense that they have, like, a hidden memory and sort of limited hidden memory that they will forget things. So, for all these reasons, Gemini 1. 5, which we still haven't covered, is very interesting because Gemini magically has fixed all these problems with perfect haystack recall and reasonable latency and cost.…

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