Dwarkesh Patel: belief

Transcript context

…Not exactly. I don't remember which models were available for fine-tuning. Assuming we had 3.5 available for fine-tuning at the time, you could have made something decently close. I don't think you would have been able to do just one iteration of fine-tuning with purely human-written data. You'd want to do several iterations. If not you're not going to do RL, which we did, you’d want some kind of iterative supervised fine-tuning where humans edit the model-generated outputs. If you train on human-generated data, even if it’s really high quality, it’s just hard for a model to fit that data perfectly because it might be something a model is capable of outputting. You need to do something iterative that looks a bit more like RL. If you’d done that, you could have gotten pretty close but it would have been non-trivial. We also had another instruction-following model trained with RL, released a little before ChatGPT. If you put a chat wrapper on that you would’ve gotten decently close but that model had some differences in strengths. That model was good at writing and poetry but it wasn’t as good at knowing its limitations, factuality, and so forth. Stepping back from 3.5, I think I heard you say somewhere that you were super impressed with GPT-2. Compared to your expectations in 2019, has AI progressed faster or slower than you would have expected? Faster than I expected since GPT-2. I was pretty bought into scaling and pre-training being a good idea. But when GPT-2 was done, I wasn't completely sold on it being revolutionizing everything. It was really after GPT-3 that I pivoted what I was working on and what my team was working on. After that, we got together and said, "oh yeah, let's see what we can do here with this language model stuff." But after GPT-2, I wasn't quite sure yet.…

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