Here is something I'd like to see: You give the machine the formally specified ruleset of a game (go, chess, etc), wait while the reinforcement learning does its job, and out comes a world-class computer player.