Prefixrl

“Then we have this program called PrefixRL, which applies reinforcement learning to a really age-old problem... It comes up with totally bizarre designs that no human would ever come up with, but they're actually 20 or 30% better than the human designs...”