“OpenAI made a bet more than two years ago to focus on applying reinforcement learning (RL) to language models, a decision that was unpopular at the time.”