Lex Fridman: belief

Transcript context

…Yeah. So I’ll give a practical answer and a more abstract answer. I think the practical answer is look in practice, models get used by all kinds of different customers. And so you can have this idea where the model can have specialized rules or principles. We fine tune versions of models implicitly. We’ve talked about doing it explicitly having special principles that people can build into the models. So from a practical perspective, the answer can be very different from different people. A customer service agent behaves very differently from a lawyer and obeys different principles. But, I think at the base of it, there are specific principles that models have to obey. I think a lot of them are things that people would agree with. Everyone agrees that we don’t want models to present these CBRN risks. I think we can go a little further and agree with some basic principles of democracy and the rule of law. Beyond that, it gets very uncertain and there our goal is generally for the models to be more neutral, to not espouse a particular point of view and more just be wise agents or advisors that will help you think things through and will present possible considerations. But don’t express strong or specific opinions. OpenAI released a model spec where it clearly, concretely defines some of the goals of the model and specific examples like AB, how the model should behave. Do you find that interesting? By the way I should mention, I believe the brilliant John Schulman was a part of that. He’s now at Anthropic. Do you think this is a useful direction? Might Anthropic release a model spec as well? Yeah. So I think that’s a pretty useful direction. Again, it has a lot in common with constitutional AI. So again, another example of a race to the top. We have something that we think a better and more responsible way of doing things. It’s also a competitive advantage. Then others discover that it has advantages and then start to do that thing. We then no longer have the competitive advantage, but it’s good from the perspective that now everyone has adopted a positive practice that others were not adopting. And so our response to that is, “Well, looks like we need a new competitive advantage in order to keep driving this race upwards.” So that’s how I generally feel about that. I also think every implementation of these things is different. So there were some things in the model spec that were not in constitutional AI, and so we can always adopt those things or at least learn from them. So again, I think this is an example of the positive dynamic that I think we should all want the field to have.…

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