“The recent evolution of large language models occurred in three stages: pre-training for grammar simulation (GPT-3), supervised fine-tuning for useful work (InstructGPT), and reinforcement learning to surpass imitation.”