Tim Scarfe: belief

Transcript context

…thing in the world and see if my hypothesis is correct. And so that's very different than what's happening in a language model where its understanding of the world derives solely from its input data. Versus in a world model, my understanding of the world comes from the delta, the difference between what I hypothesize is gonna happen in the world and then my actual experience of it. And this distinction really matters the more we are gonna start offloading our cognition to these systems. Because, for example, everything that ChatGBT knows is on the basis of its input data. Mhmm. And so that means if false information or wrong information is in the input data, ChatGBT is gonna know that information. There's there's absolutely no hypothesis testing embedded into ChatGPT versus our true AGI agent that 1 will will be invented. What it would do is it's going to hypothesize aspects of the world, and it's going to test its own hypotheses. And so if you give it false information, if it reads articles about how the Earth is flat, it's not gonna just start talking about how the Earth is flat. It's going to say, okay. Well, that is incongruent with my model of the world. I'm gonna now render some tests where I could differentiate them, and I'm gonna perform those tests, and then I'm gonna conclude that the world is not is not flat. So so, when 1 says that Chachapiti does not have a world model, I think some people misinterpret that as suggesting that it's just dumbly looking at the statistics. And I think that's not at all what we're saying. In order to correctly look at the statistics of language, clearly, it's built up a very rich and complex model of the text that it's seeing, and that's how it's able to predict the next word so well. But it's not what most people mean when we say world model. Yeah. I mean, couple of things on that. Yeah. I think I think people conflate the machinations of of language models with how we represent them statistically or or abstractly. Because if you look at lot of papers, they they actually represent it like a probability, you know, like a joint probability distribution. And, of course, you know, the way that language models work is completely different to that. But you are bringing in some very interesting things. So first of all, we are agents in the world. So the the agential lens is quite interesting. We interact with the world, so we're not just learning from observational data. It's quite interesting, actually. I was talking with Nick Chater. You know, we said, oh, why is it that in our everyday experience, we see, we experience the world in in 4 d color? And he said it's because it's interactive. So in your experience plane, you can actually seek new information. Right? You can saccade your eyes. You can get new information, and you can you can touch things. And when you're doing future or past simulations, you don't have that interactivity. So there's something about interactivity, which is really important. But it even then, right, you know, we could you know, how far how far could you go? So a complete 1 to 1 simulacrum of the world wouldn't be a particularly good model. And in in physics, there is no causality. Right? So it's just dynamics. So causality is actually something which emerges very, very far up. So we're talking about a model which is an approximation of the real world, which may or may not include causality. It probably would because we're it's an interactive model, and it has this kind of agential map. But I guess we're just kind of drawing the line somewhere, and we're saying, well, that is a world model. And yeah. Well, I think the it's I'm actually I'm not sure if even if we rendered let's think about it. If we rendered a perfect three-dimensional map of every particle in the universe, and and that was the input data to some infinitely large model, I still would argue that it is learning something different than a model that is given some form of agency where it can hypothesize rules and then test its own rules. Now given infinite time, it is possible that those will converge because given infinite time, every possible hypothesis I could conceive of will end up showing up in the training data. So eventually, I'll see the training data of every possible experiment I could run. Yeah, if time is infinite, I guess you could suppose that happens. But what's so different is there's this dramatic dimensionality reduction that happens when you show me something uncertain, and then I can conceive of specifically the tests I wanna run to map the the uncertain thing to my mental model of the world. And so that's a very different way of learning about things. It's not just input data and then and then self supervising on predicting one's own input data. It's building a model that can I can simulate possible outcomes and then hypothesis test those outcomes? And so I think, you know, when we look at even this is not uniquely human at all. If you look at, you know, the way a rat would deal with something novel in its environment, it's drawn to the novel thing, and it explores the novel thing until it feels like it understands it. And then it will move away. And so when you show a child an object that is perplexing, they will touch it, turn it around, try and understand it until they feel like they built a model of it. That simple act is doing something very different than the self supervision we see in most AI models today because I see something I'm uncertain about, and I'm volitionally gonna create new training data for me. I'm I know the training data I want now. I wanna see what happens when I pick it up, and I turn it to the left, and I turn it to the top. A convolutional neural network doesn't do that. So the way we teach CNNs to understand rotations in 3 d objects is we manipulate the training data ourselves. We we audit, we take data, we take imagery, and then we rotate a bunch of different ways so that we are the ones curating the dataset to teach it these things. Yeah. But that's different than the way we learn about things. So I think this is a key aspect that's missing from AI systems today that we're that folks are working on, but that's something we're gonna have add in.…

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