Tim Scarfe: evaluation

Transcript context

…Yes. I think that totally makes sense. The there's some interesting ideas in AI here, actually, where there's this concept of knowledge distillation, in AI, where, 1 way in which you can have model a teach model b the things that model a knows. 1 way is you can wholesale copy the parameters of model a. Of course, that's totally biologically implausible. There are aspects of of parameter copying, which is the components of our brain that are genetically hard coded. That is a version of parameter copying. But for other applications, it's not feasible to just copy parameters or it's maybe not desirable. So knowledge distillation is saying, Okay, well, we can have a set of data that we give model A. And we either look at the outputs of model A or the layer before the outputs so we can see sort of more richness in its representation of of the input you give it. And then take those that data, that almost labeled data to model b and then train model b on it. So that's, distilling some of the knowledge through almost training model b to try and act similarly to model a. And so that type of information transfer, I think, does occur in nonhuman primates, and that's imitation learning. However, it is not nearly as, rich because what happens in non human primates is it's primarily grounded in just the actions that I'm taking, which is much less rich than I can share not only the the data of what you see me actually do, but I can also include in the data I'm transferring to you things that happen only in my mind. And that that opens the door for much more transference of of, and the word you use, the computations that I'm performing. Yeah. So I think it's absolutely absolutely true. Before, we were, learning in the physical world. So we were learning from physical things that we were directly observing, and now we are learning from imagined actions. But there's a bit of a latent component to language as well. So for example, someone might come up to me and say, blue swirly thing is over there. And I'll say, well, I don't know what you mean about the blue swirly thing because I've I've never seen 1 before. So there's this kind of, inference process. And this is where it starts to get really interesting because there's a diffusion. Right? There's a kind of, there's a message passing that happens between all of the different agents, and it's filling in missing information. So even though many of the agents wouldn't have seen anything like what we're talking about, sometimes it can be filled in with subsequent interactions with people, and sometimes it can just become a kind of latent category which can be filled in later. So there's this real diffusion process going on, which I think is quite difficult to articulate. Part of what's so interesting about language is it's still an area of such controversy amongst cognitive psychologists, linguists, even AI people. And so much is still, you know, unsettled about it. There are still, you know, debates today. I mean, there's debates today about whether language is primarily a tool for thinking or communication. And and most people were, you know, Chomsky is the most famous proponent of the idea of language for thinking. So, you know, and he has evolutionary arguments that language initially evolved not as a tool for communication, but for our own process of thinking and then later was accepted or used for communication. That's a minority view. And then other people argue, which I'm more amenable to, that language was primarily used as a tool for communicating. And these ideas actually are reemerging with language models because, the way language models learn about the world in some sense is language becomes the reasoning tool itself, which is more Chomsky like. Even though, you know, there's a lot of, you know, I think the the success of language models, I think, in a lot of ways discredits a lot of Chomsky's ideas, and we can talk about that. But but interestingly, the fact that we're using language as the fundamental mechanism for reasoning and thinking is actually somewhat Chomsky like. Versus language as communication, the idea is language is a a condensed set of tokens that I'm passing between minds. But the goal, the real communication I'm trying to share with you is what's going on in my mind. In other words, the mental simulation, the more mammalian component here. The rendered 3 d world is what I'm trying to transfer to you. And I condense it into this code that then you, you know, reverse engineer back into a mental simulation. And theory of mind, 1 reason why language might might be so rare in the animal kingdom is mentalizing theory of mind, which is relatively rare in the animal kingdom, is a prerequisite. Because in order for me to reverse engineer the language code you've provided me, I need to be able to infer what you might have meant by what you're saying and reason about why you would have said this and what knowledge you have and and etcetera, etcetera. So so yes. So I think language is is intended to cue to another person to render something in their mind. This is also where, you know, teaching is so important is is such a key aspect of language learning because, we can infer, you know, what declarative labels is this person aware of. And when they're confused, then you can start you have to start trying to iterate to understand what are they confused about that I'm saying that I can disambiguate for them. So there's also a disambiguation process where you ask follow-up questions when you feel like you don't fully understand what's going on in someone else's head.…

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