Reinforcement learning is all you need, for next generation language models (yuxili.substack.com)