“The RLHF training process is considered very safe because the model's objective is narrowly focused on producing text that a human will approve, without any other goals or concerns about the real world.”