LLMs are (still) mostly powered by imitative learning, not RL