Speaker unverified: belief

Transcript context

…Arguably, that's alignment, you know, we want models to like humans. Sometimes it's good. Yeah. But you also want humans to actually say the truth. Disagree. Yes. Exactly. To challenge you. Definitely. If what you're thinking is not factual. There was also this cool paper by Cohere and the University of Edinburgh, which was human feedback is not gold standard, I think. And where they actually established super interesting things such as humans prefer models which are over assertive. And if you have the choice between an answer which is false, but given super assertively and an answer which is right, but not as assertive, humans naturally will say the assertive but false answer is a better one. So basically, Arenas are not giving you factuality, which should be a super important aspect of LLMs, I think. That's the same with everyday life. You know, people just trust the person saying the thing assertively, even though it's false, and then actually try and figure out what the truth is. So yeah, I think you mentioned that, you know, it's like a more social experiment. I think it's a good point. Like the same biases that people have interacting with humans, they kind of put in the models themselves.…

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