All you need for a good value function is high quality simulation of the task. Some domains have better versions of this than others (eg theorem provers in mat

All you need for a good value function is high quality simulation of the task. Some domains have better versions of this than others (eg theorem provers in math precisely indicate when you've succeeded) Incidentally, lean could add a search like feature to help human researchers, and this would advance ai progress on math as well