Dwarkesh Patel: uncertainty

Transcript context

…It was trying to convince him to break up with his wife or something. So this is an interesting example. Personas. Is Sydney Bing having this personality a feature versus another personality it could get locked into? And is that fundamentally what humans are like where in front of other different people, I'm like a different sort of personality? Is that the same kind of thing that's happening to ChatGPT when it gets RL-ed? I don't know. A whole cluster of questions you can answer. I really want to do more work. The sleeper agents is in this direction of what happens to a model when you fine-tune it, when you RLHF it, these sorts of things. Maybe it's trite, but you could just say you conclude that people contain multitudes and so much as they have lots of different features. There's even the stuff related to the Waluigi effects where in order to know what's good or bad, you need to understand both of those concepts. So we might have to have models that are aware of violence and have been trained on it in order to recognize it. Can you post hoc identify those features and ablate them in a way where maybe your model is slightly naive, but you know that it's not going to be really evil? Totally, that's in our toolkit, which seems great.…

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