It is a mental exercise to re-frame the problem. Sure it would be inefficient but the point is every stage could be done with reinforcement learning from a high level abstract design without every dealing with 'code'.