Jeff Dean: belief

Transcript context

…Yes. So there's a sort of modus ponens, modus tollens thing here. One way to look at it is, look, LLMs have so much further to go, therefore we project orders of magnitude improvement in sample efficiency just if they could match humans. Another is, maybe they're doing something clearly different given the orders of magnitude difference. What's your intuition of what it would take to make these models as sample efficient as humans are? Yeah, I think we should consider changing the training objective a little bit. Just predicting the next token from the previous ones you've seen seems like not how people learn. It's a little bit related to how people learn, I think, but not entirely. A person might read a whole chapter of a book and then try to answer questions at the back, and that's a different kind of thing. I also think we're not learning from visual data very much. We're training a little bit on video data, but we're definitely not anywhere close to thinking about training on all the visual inputs you could get. So you have visual data that we haven't really begun to train on. Then I think we could extract a lot more information from every bit of data we do see. I think one of the ways people are so sample efficient is they explore the world and take actions in the world and observe what happens. You see it with very small infants picking things up and dropping them; they learn about gravity from that. And that's a much harder thing to learn when you're not initiating the action. I think having a model that can take actions as part of its learning process would be just a lot better than just sort of passively observing a giant dataset. Is Gato the future, then?…

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