Comment by erwincoumans

"making a new theory that is really useful is out of reach for the moment in my opinion."

Curious how this ages.

Recursive self improvement, self-play and multi-agent RL could make useful new theories, eventually.