They're lossy approximations, and we should expect more powerful agents to expend compute on avoiding the losses.

→ View original post on X — @esyudkowsky