Will Bryk: preference

Transcript context

…Yeah. Right. Yeah. Um, so we're actually referencing a blog post that you wrote super intelligence and super knowledge. Uh, so I would refer people to that. And this is actually a discussion we've had on the podcast a couple of times. Um, there's so much of model weights that are just memorizing facts. Some of the, some of those might be outdated. Some of them are incomplete or not. Yeah. So like you just need search. So I do wonder, like, is there a maximum language model size that will be the intelligence layer and then the rest is just search, right? Like maybe we should just always use search. And then that sort of workhorse model is just like, and it like, like, like one B or three B parameter model that just drives everything. Yes. I believe this is a much more optimal system to have a smaller LM. That's really just like an intelligence module. And it makes a call to a search. Tool that's way more efficient because if, okay, I mean the, the opposite of that would be like the LM is so big that can memorize the whole web. That would be like way, but you know, it's not practical at all. I don't, it's not possible to train that at least right now. And Carpathy has actually written about this, how like he could, he could see models moving more and more towards like intelligence modules using various tools. Yeah. So for listeners, that's the, that was him on the no priors podcast. And for us, we talked about this and the, on the Shin Yu and Harrison chase podcasts. I'm doing search in my head. I told you 30 million results. I forgot about our neural link integration. Self-hosted exit.…

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