Episode 25: Nicklas Hansen, UCSD, on long-horizon planning and why algorithms don't drive research progress
Generally Intelligent
00:00
Predicting the Future
The future is an interesting problem because I don't think anybody knows the answer. Maybe you could have something that predicts only specific things about the teacher that are themselves useful for predicting the reward. Because that's really what I care about. But of course, a valid constraint then is, if you're learning reward functions, will actually generalize to new tasks. Yeah.
Transcript
Play full episode
