Nathan Lambert: evaluation

Transcript context

…It's just one thing I do want to establish as a baseline for our listeners. We are updating all the weights. In some sense, the next token prediction task of training a language model is a form of reinforcement learning. Except that it's not from human feedback. It's just self-supervised learning from a general corpus. There's one distinction which I love, which is that you can actually give negative feedback. Whereas in a general sort of pre-training situation, you cannot. And maybe like the order of magnitude of feedback, like the Likert scale that you're going to talk about, that actually just gives more signal than a typical training process would do in a language model setting. Yeah. I don't think I'm the right person to comment exactly, but like you can make analogies that reinforcement learning is self-supervised learning as well. Like there are a lot of things that will point to that. I don't know whether or not it's a richer signal. I think that could be seen in the results. It's a good thing for people to look into more. As reinforcement learning is so much less compute, like it is a richer signal in terms of its impact. Because if they could do what RLHF is doing at pre-training, they would, but they don't know how to have that effect in like a stable manner. Otherwise everyone would do it. On a practical basis, as someone fine-tuning models, I have often wished for negative fine-tuning, which pretty much doesn't exist in OpenAI land. And it's not the default setup in open-source land.…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.