Yi Tay: belief

Transcript context

…between the experts. Yeah. Yeah. Yeah. I think as a architecture itself, the flop brand ratio makes it like worth it. Right. But I think the, the thing that's not very well understood is that, like, how does like MOE, like, like for me as a research question, is that like when you. Like, how does it, like, relate to capabilities and stuff like that, like, does this inductive bias actually, , for example, when you do, like, massive instruction tuning, I think there was this paper, like, Flan MOE or something, like, they showed that, like, , instruction tuning, I'm not, like, fully sure, I don't recall fully, but, like, when you do massive instruction tuning, like, MOE models are, like, they behave differently from a, from dense models and stuff like that. Like, I think, Okay, like, fundamentally, I just think that MOEs are just, like, the way to go in terms of, like, flop parameters. They show that they bring the benefit from the scaling curve. If you do it right, they bring the benefit from the scaling curve, right? And then, that's the performance per flop argument, activated params, whatever. That's, like, kind of, like, that's a way to slightly cheat the scaling law a little bit, right? By having more parameters, right? I think the more interesting thing is about, like, what trade offs do you make in terms of capabilities? Because of this new architecture. Mm. I think that's actually like the, the question that I, I think I, I guess all the frontier labs, they already know this, but nobody is writing papers anymore about this. So like, you just have to live with, with what? Like, but I think OI think I'm, I'm, I'm, I'm bullish about Moes. Yeah. Yeah. I had to, I mainly exercise for myself on reading research directions and what their asto asymptotic value is. Mm-Hmm. and I put OS pretty low because I think you have a good base model and then you upcycle it and it bumps you a little bit. And I think that's it. But like, I'm always seeking to invalidate my hypothesis, right? Oh,…

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