“Reinforcement learning, where models generate their own training data, is a key technique to overcome the bottleneck of finite text data available on the internet.”