Robert Wright: belief

Transcript context

…ms to be, first you understand the mind, then you just put the unders- you know, you design a, a, an AI in accordance with that very clear understanding. Of course, that's not what wound up happening. And I, the, the misunderstanding I had perfectly reflects this kind of 180-degree difference between the expectation in those days and what wound up happening. So in the piece, I describe a neural network, or at least what I think is a neural network, and it had in fact been put forth as a model, what I describe, by a guy who'd collaborated with Hinton on, he had co-authored a paper. But he was a psychologist, and w- the, the neural network he was describing was kind of his idea about something that would work. And, and the, and the key thing is that each node in the network would represent the specific sense of meaning of a word. So a word like throw, which can either mean host, as in throw a dance, or hurl, as in throw a ball, there would be a node for each of those. And I, I, I, uh, I, I won't go further in describing the model, but the key thing is that my assumption was, you know, obviously the people who set this model up will have to translate their understanding of the meaning of the words into the machine. You, you can't do AI that generates language without somebody making some kind of connection between the words and the meaning. Well, of course, it turns out you don't have to do that, and the machines, I would say, and I, I may describe some things in the book in a way that some people in AI would, um, differ with, but I think you could say that the, the machines which turn out to use the vectors as a means of representing the meaning of words, they in a sense discovered, in quotes, that meaning is a property of words. Nobody told them that. They just said, "Here's some gibberish, predict..." We just said, "Here's some gibberish, predict the next gibberish." And, you know, and they, you know, implicitly recognized in the course of their training that in order to do that, you have to represent the meaning of words. So that, that, that kind of... The misunderstanding I had is exactly, I think, analogous to the larger misunderstanding that we'll be, we'll be putting our understanding of the human brain into, uh, the AIs. And I think it's important for laypeople to understand this because what it means is that, you know, you can just keep feeding data into these things along any number of dimensions, visual data, olfactory data, audio, and the machines will reverse engineer, I think, I'd put it this way, will reverse engineer cognitive functionality that's in the human mind. And, and I would add, and Nathan, I'm curious as to what you think about this, whether you think this, this thing I'm gonna say would be accepted by AI researchers, and if so, is duly appreciated, which is that, you know, the training of a large language model, I think, is at least as much a process of natural selection, of evolution, as of learning. So, for example, we presumably in our brains have a mechanism for representing the meaning of words. We still don't know what it is. much a process of natural selection, of evolution, as of learning. So, for example, we presumably in our brains have a mechanism for representing the meaning of words. We still don't know what it is. Some people, by the way, psychologists, had long posited a mechanism that would be quite analogous to what we now understand goes on in large language models. Uh, but in any event, I think it's pretty safe to say that that is a product of natural selection, right? Now, also in the course of the training, the machine becomes conversant in a specific human language. Well, that's more a product of human learning, right? During, uh, you know, as, as the organism is developing. But I think what laypeople need to understand, and I'm, again, I'm, I'm curious as to your view on, on, as to how this would hold up in AI circles, but is that these things basically, in a certain vague sense, re- recapitulate evolution. Recapitulates, in a way, a misleading word, because the cognitive functionality is not the exact same mechanisms the brain uses, but, you know, I think close enough in a lot of, uh, cases. It's kind of doing millions and millions of years of evolution in a few months, and you can, you can do a lot with that, and we have not begun to, I think, exhaust the potential of that. I mean, as so far as words go, we're close maybe, but There's a lot more to do. Do you, do you think people in the field would agree that, yeah, it's, it's a lot like i- i- it's really, evolution is maybe a better term for many purposes than learning? Yeah, it's a great question. I think there's a couple different senses, and I'm not sure I have gripped all the senses that you mean when you use that term. You certainly do hear people talk about pre-training broadly as sort of being analogous to evolution in the sense that there's y- there's sort of this question of like, well, why are humans so sample efficient, right? And the models need so long to train.…

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