It would be cool if it could learn to recover from failures and try other approaches until it gets the job done, and in the end verify it was correct. The security implications are huge, though.