[jhrmnn]

In math specifically, one could easily imagine a reward signal from some automated theorem proving engine

[cubefox]

Yeah. I went into some detail of how it might work here: https://news.ycombinator.com/item?id=38036986