Hi again. I had already written about this later in my blog post (which is unrelated to this thread), but the point was that RLHF hadn't been applied to language models at scale until InstructGPT. I edited the post just now to clarify this. Thanks for the feedback!