I've already considered eg training neural nets to evaluate MuZero-style learned representations of hidden states. I have predictions about why that would fail. Do you have something else in mind, or should I explain some reasons that class of methods wouldn't work?

→ View original post on X — @esyudkowsky,