Blaise Agüera y Arcas: belief

Transcript context

…a few questions, I guess, hidden in there. You know, 1 of them is, you know, do I think of of LLMs for instance of of today's, you know, sort of frontier models as being less than or different than in some basic way, you know, our our brains, what are those gaps? So first of all, I mean, they're obviously very different. I mean, we know their their their architectures are different, they're trained in a very different way. The fact that, you know, for me, the remarkable thing is actually how convergent a lot of a lot of their properties are with those of brains despite all of that. You know, the fact that you find internal representations and many of them that that that surprisingly resemble, you know, ones that you can measure in human brains, it's brain score type measures of of Machen Shrimp and and Co, you know, or or, you know, sensory modalities, you know, in humans can be reproduced remarkably well even by models trained on pure language, which is really remarkable. You know, speaks to how much is encoded in language and how much of what is encoded in language is a reflection of architectural properties of our brains and umwelts and how much of that is then reconstructed essentially by those models. Now, the question of, you know, what we draw first when we draw a picture and how that all works. I mean, remember that, you know, image synthesis models like CLIP or what have you are working in pixel space, to begin with. And, you know, diffusion models, by the way, you know, work very differently from various other kinds of models. I mean, we now know that, you know, you you can drive a robot with a transformer. So if you give 1 of those robots a paintbrush and you say or, you know, know, a pen and you say, now draw, what it will draw is gonna be very, different from what you get from a diffusion model that starts filling in pixels. And for that matter, all of that is different from what happens in your own head when you're visualizing something. So, you know, I think a lot of this is is not so straightforward to to analyze because of all those differences in the way that IO and the representation space works. I do think that today's models are highly compositional. I mean, even with a lot of those original, you know, image synthesis models, the fact that you could say, you know, a teddy bear at the bottom of the sea playing with the speak and spell or whatever, it'll do it, you know, tells you that that they they can compose. Again, are there capabilities like ours? No. I mean, there there are definitely places where they're better, places where they're worse, places where they have surprising gaps. So it's different, but but I wouldn't say that there's a fundamental lack of composition there at all. places where they're better, places where they're worse, places where they have surprising gaps. So it's different, but but I wouldn't say that there's a fundamental lack of composition there at all. I think if if anything, the biggest gap between transformer based models and what we do is actually narrative memory, right, or being able to form long term memories and and and that way have a kind of persistence of a self over over long periods of time. They they don't have that yet. I'm conflicted. You're pointing to this universal representation hypothesis. I think Chris Ola popularized it with some of his visualization experiments. And it it's true. The representations are very convergent. And and other things lead me to to believe that the models produce these kind of superficial imposters that they give you exactly the right answer, but for the wrong reasons. And 1 of the hints of that is when you do variations on on the input, it's it's not robust. There's the there's the Turing machine argument as well. So that, you know, these LLMs are finite state automata, but they can access tools which are Turing complete. So, you know, perhaps we could say the system is Turing complete, but I don't believe that ChatGPT is is effectively searching the space of Turing machine algorithms. It hasn't been trained to do that, But it is surprisingly robust with the ARC challenge. It can actually Yeah. You know, it can it can do really well, especially if you do some evolution and do do some refinement and so on. So it feels like we're we're knocking on the door, but it but there's something missing. I think that in many…

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