Michael I. Jordan: preference

Transcript context

…You should ask the language model builders, because all they're doing is predicting the next word and there's not any thinking about uncertainty quantification in doing that. And you can graft in ideas, but they're dubious. They're often putting up dubious prior in or or and and so you can you go to the statistician and and that's what people have done and they've said, okay, I can just treat it as a black box and I can put conformal prediction around it. It's a nice method, doesn't require a lot of assumptions. So yes, that's true. But it makes a lot of there's an exchangeability assumption. The data, if you scramble it, it's the same. And so while I think all of that's really crucial and important, I tend to think more about the broader context. So I gave an example in that article that you mentioned of a duck who goes to a lake and this is a statistician duck. So it's kind of calculated that over the last year there tends to be twice as much grain on that side of the lake than on this side, 2 to 1 ratio. So now the next day I need to decide on the duck which side of the lake I go to. And the Bayesian duck who has those probabilities, would then do the maximal expected value and they'd go to the left side of the lake with probability 1. But the actual ducks don't do that. They go to probably 2 thirds to that side of the lake and 1 third to the other side. They're hedging. But it's not just a hedging thing. Hedging would just do occasionally going to the other side of the lake. They're actually getting the right ratio. And so the explanation is that you weren't thinking about the context right of this uncertainty. It's not just you, the individual duck. Probably you evolved in a world where there are many ducks. And if all the ducks went to the same side of the lake, obviously you've missed out on a resource. And so is there an algorithm that allows many ducks to cooperate here? Well, if they all have that same uncertainty, then they can sample with probably 2 thirds and go to this side versus 1 third. And that's actually a Nash equilibrium of the bigger system. So the right way to think about uncertainty there is that in the context of the population, what should be how should I use my uncertainty? Another kind of uncertainty, that's kind of the economic side. Another uncertainty in economics is the 1 I've alluded to, information asymmetry. You know things I don't know. And you have expertise I don't know about. But we're gonna work together, and I'll maybe give you a contract, a menu of options. But even if I interact with you for a while, I still might not know. There's things you're going to know that you're not going to give away to me. And maybe you'll hedge, you you'll lie a little bit, so I don't know about that. That's not just sampling. That's a different kind of uncertainty. And then finally, there's what I like to call providence, you know, that's more like a database kind of uncertainty. know about that. That's not just sampling. That's a different kind of uncertainty. And then finally, there's what I like to call providence, you know, that's more like a database kind of uncertainty. If I want to do a medical operation and you're a doctor and you look at the data for people like me, here's the, if you do the operation this way, the probability of survival versus this. And I look at that and say, great, but now you tell me all that data was gathered 10 years ago. And I'm gonna say, okay, my confidence interval should go up. All right, well, classical statistics, could talk about that. In fact, I'd be more of a Bayesian to think about that. But it doesn't. It's just sort of the data is the data. And it should be in a bigger system that as data is flowing around, it should always be tagged with metadata about how old it is and that should be quantitatively brought into the uncertainty quantification. We're not doing anything like that right now. And so the poor LLMs, are basically doing none of the above, have to strike out a little bit in all these directions if they're gonna start to do like what humans do. We are pretty good at getting these, with a little bit of providence. Oh, it's old data, I discount that. We get a little bit of context. Oh, there's a social environment here, I should just do the same thing, should randomize. Oh, there's some sampling uncertainty and so on. We put all that together almost seamlessly. And then we do this in a social context where if I don't know how to get from here to the other side of town, I will ask someone who looks Danish. I know something about how to gather more data and so on. So the poor LLM has none of the above. And so what should it say when you ask how sure are you? And all it's doing to the best of my knowledge is that it's just, well, in the past someone asked a human on the Internet how sure are you of that equation you just wrote down? And someone said something, oh, I'm very sure because of this or that. And I think it just mimics that those kind of assertions, but that's not reasoning under uncertainty. And if we did have epistemic, you know, quantification, what would be the the main uplift from that? Is it is it about, I I know I don't know something, so I'm going to kind of lean in and try and do more epistemic foraging in that area?…

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