Give an AI a goal. Watch it achieve the goal perfectly and completely miss the point. Can you write constraints tight enough to stop it?