Sholto Douglas: belief

Transcript context

…That's really interesting. Does that look to you like chain-of-thought? Or what would you imagine as these models get better, as these models get smarter? What does the synthetic data look like? When I think of really good data, to me, that raises something which involved a lot of reasoning to create. It's similar to Ilya's perspective on achieving super intelligence effectively via perfectly modeling human textual output. But even in the near term, in order to model something like the arXiv papers or Wikipedia, you have to have an incredible amount of reasoning behind you in order to understand what next token might be output. So for me, what I imagine as good data is data where it had to do reasoning to produce something. And then the trick of course is how do you verify that that reasoning was correct? This is why you saw DeepMind do that research for geometry. Geometry is an easily formalizable, easily verifiable field. You can check if its reasoning was correct and you can generate heaps of data of correct trig, of verified geometry proofs, and train on that. And you know that that's good data. It's actually funny because I had a conversation with Grant Sanderson last year where we were debating this and I was like, “fuck dude, by the time they get the gold of the Math Olympiad, of course they're going to automate all the jobs.” Yikes. On synthetic data, there’s a thing I speculated about in my scaling post, which was heavily informed by discussions with you two and you especially, Sholto. You can think of human evolution through the spectrum of getting language and so we're generating the synthetic data. Our copies are generating the synthetic data which we're trained on and it's this really effective genetics, cultural, co-evolutionary loop.…

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