“LLMs and reinforcement learning automate tasks that can be verified, whereas traditional software automates tasks that can be specified.”