“The transformational effect on large language models is primarily due to the human feedback (HF) component of RLHF, not the reinforcement learning part.”