Unsupervised Environment Design

Josh: It makes me think about unsupervised environment design, which is Mingxie who we talked to recently on the podcast. He perturbs the environment in order to give reinforcement learning agent the like kind of optimal or most difficult next thing to force it to generalize. And I could imagine like if you could find a good way to identify what is desirable versus undesirable information then you could do unsuper supervised environment design such that you're like perturbing the environment toward tasks that would allow you to retain the desirable information but forget the undesirable information so like that. Josh: You can definitely imagine, well, I guess it's not easy to design these things, but for things