How Can LLM RL Work Despite Information-Theoretic Inefficiency | Dark Hacker News