“The term "RLHF" is a metaphor for all language model adaptation; the reinforcement learning component itself is a small part of the process and less important than the data and infrastructure.”