“Approximately two years ago, reinforcement learning was not a major component in the training of large transformer models like GPT-3, Grok-1, and possibly GPT-4.”