Max Bennett: belief

Transcript context

…Yeah. So on on that then, I mean, people say that language models, you know, they they confabulate and, they they don't, preserve epistemic factualness. But you could also argue the same thing about us. Right? So we actually confabulate everything. We don't really have goals. We just kind of, generate these post hoc confabulations, and then we explain our behavior, and we kind of pretend that that was what we wanted to do, that we had beliefs and so on, but we just kind of make it up as as we go along using this kind of active inference. So I guess the the the question is, like, we do think of ourselves as being like, even though we are emotional and subjective and, like, you know, like, we believe in religion and lots of things that that we presumably made up, but but we have Wikipedia. We have objectivity. Even though it's an illusion. Right? There's there's no such thing. Know, even, like, general relativity, it's not as objective as as we think. Like, if you keep asking why and why and why, it just kind of disintegrates into incoherence. But there seems to be some objective structure which is preserved. And how is that explained given that our brain simulations are, you know, they don't seem to select for for truthfulness? I think the desire to say is are humans better or worse than ChatGPT is almost like a a red herring, where I look at ChatGPT as like an alien. It's like an alien brain. And there are certain things it does that are clearly better than us. I mean, information retrieval in chat GBT blows a human away without a question. So, like, in many ways, it's way better than humans. And but but there are certain things that human brains do that ChatGeeBT does not have. And if we're trying to build human like intelligence, there's certain inspiration we can garner from from human brains. So I think, there's a component of our model based sort of rendering a plan and then executing that plan that has a level of explainability that is unique relative to a system that is sort of just iteratively predicting the next token. But we also do the same thing that ChatGPT does. When we make model free choices and then you say, why did you do that thing? What we engage in is exactly as you're describing, a post hoc explanation. Because I didn't render a plan. I was just walking down the street, and you say, did you move your your foot there as opposed to 2 inches to the right? What I'm gonna do is render a post hoc explanation of why I did that, but I didn't really think about it. I'm just explaining it after the fact. So so it's definitely the case that humans have that component, but it means there's also another component which we which I would argue is is unique and important, which is our ability to pause, render a plan, and then execute against that plan. But the key thing that's that I I think is the dividing line between these models and us is the ability to render hypotheses and make interventions in the world. That's the, like, that's the key thing. And so it it's not the case that our brain has the true, objective state of the world in our head. I don't think that's, there might be components of objective truth in chatty bitty that it contains that we don't have. And I think in its information retrieval, it has probably, in some ways, more aspects of reality than I do, in terms of it having read all of Wikipedia and answering questions about biology that I don't even know. But there are also components of the world that the human brain has rendered and contains that ChatGPT does not because of our ability to make hypotheses and intervene and learn the causal structure of the world. And I think that is the the dividing line. But it's I wouldn't I wouldn't say it's because the human brain knows the objective state of the world, and ChatGPT does not. Max Bennett, it's been an absolute honor to have you on MLST. Everyone at home, you need to buy his book immediately. It is a a wonderful, wonderful book, Max. You did such an amazing job of bringing all these things together, and we've now spoken for 4 and a half hours going through the last 3 chapters. And my god, it's been an honor. Thank you so much.…

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