ChatGPT and InstructGPT: Aligning Language Models to Human Intention
Deep Papers
Do You See Other Major Applications Just Skipping the First Step?
RLHF helps you get more fine-grained tuning of model behavior, whereas supervised fine-tuning collecting demonstrations can kind of more drastically shift the model behavior. For instance, let's say you have some model that just like to start off with it sucks at generating summaries and so getting a bunch of ranking feedback between different really shitty summaries is maybe not the most useful. Instead, what you might want to do is collect some examples of really, really good summaries and then have your model kind of try to imitate that for a bit. And we have some results on this in a different paper, but it's still a very open question actually.
00:00
Transcript
Play full episode
Remember Everything You Learn from Podcasts
Save insights instantly, chat with episodes, and build lasting knowledge - all powered by AI.
