“OpenAI's leadership recognized the need for a paradigm beyond pre-training and invested research efforts into reinforcement learning, even while the vast majority of resources were focused on scaling pre-training.”