John Schulman: belief

Transcript context

…Yeah. That does feel like something that's missing from today's systems. People haven't really pushed too hard on this middle ground between large-scale training—where you produce a single snapshot model that's supposed to do everything like a deployed model—and on the other hand in-context learning. Part of that is that we've just been increasing context length so much that there hasn't been an incentive for it. If you can go to a hundred thousand or a million context, then that's actually quite a lot. It’s not actually the bottleneck in a lot of cases. I agree that you'd probably also want to supplement that with some kind of fine-tuning. The capabilities you get from fine-tuning and in-context learning are probably somewhat complementary. I’d expect us to want to build systems that do some online learning and also have some cognitive skills, like introspecting on their own knowledge and seeking out new knowledge that fills in the holes. Is this all happening at the same time? Is it just a new training regime where all these things can happen at once, whether it’s long-horizon or this kind of training? Are they separate or not? Is the model smart enough to both introspect and act on longer horizons so that you get adequate reward on the long-horizon tasks?…

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