Cameron Berg: commitment

Transcript context

…cally points at this exact thing. The most classic example. There may be two examples I'll point at briefly. 1 is dopamine. Like this is just like the most culturally well understood neurotransmitter. We know it's not exactly pleasure. It has more to do with approach or like approaching things that we find pleasurable. One good intuition pump for this is like if you go to pet a dog, its tail will wag as your hand approaches the dog. But as you start petting, it is his tail will stop. So this is like this is basically what dopamine is up to. It's like prediction of a sort of interesting desired stimulus essentially. And we know full well positive and negative reward prediction error is instantiated dopaminergically and we also subjectively. i think the reason people understand dopamine like in our culture in the year twenty twenty six is because we understand that it corresponds to a subjective dimension. We know what it means to be in like a high dopamine or low dopamine state. and so this to me is like the most obvious and fundamental exploit like dopamine is one hundred percent instantiating TD learning like reward prediction error in the brain. i am one hundred percent confident that that is the case. This was established in human neuroscience 40 years ago. We also know subjectively dopamine corresponds to basically feelings of yeah, like positive pleasure adjacent approach style behavior. And dopamine depletion corresponds to basically the opposite of that. If you think you're going to get a cookie and you don't get the cookie, you feel a certain way. That is explained by dopamine. If you don't think you're going to get a cookie and someone hands you one, you feel a certain way, that is also by dopamine. Another example I can give has to do with, I think it's insular cortex. So we can take the same, let's say basically 2 scenarios. You have been walking through the desert for a couple hours or you've been walking through Arctic tundra for a couple hours, and in both cases, I pour cold water on your head after that. This is the same stimulus. You have this sort of same body. You're the same person with the same preferences. In one case, this is a positively valenced experience. In another case, this is a negatively valenced experience. What mediates that is basically the implicit goal state of the system. And one it's to warm up, and the other it's to cool off. So I can take all the same variables and I can run the simulation forward and I can very easily predict where you're going to have the positively valenced experience, where you're going to have the negatively valenced experience, what that corresponds to. And to me, again, that's a big hint that goal relative prediction error is doing something fundamental from the outside that maps on to what I experienced and what I think other people and animals experience consciously from the inside. These are the core moves. I, I, I make it, you know, I'm sort of swallowing computational functionalism. I understand. It means, I have to say, the simple RL algorithm is conscious when it's training. To me, this like localizes a lot of concern on the training process. lowing computational functionalism. I understand. It means, I have to say, the simple RL algorithm is conscious when it's training. To me, this like localizes a lot of concern on the training process. But indeed, if there are systems that are capable of doing this sort of learning online, which we know full well LLMS are capable of doing, they do something like that looks in activation space in forward pass like stochastic gradient descent, then the concern falls there too, if you have systems that are doing online learning. So anyway, this is my, this is sort of my whole, my whole shtick. If, you know, I have to sort of put my cards on the table and say, what do I think consciousness is? It's not that I think it's a grand mystery. I, it's something of this general shape. What I will say is that the, in the work that I'm doing, I do not want people to have to either you and the people listening to this will fundamentally think that this makes sense or fundamentally think it doesn't or be very skeptical or something. I do not want that reaction to cloud. All of the other work I'm doing, like everything else we've talked about in this podcast, is completely orthogonal to my pet theories about consciousness. Now, you might think that I'm studying RL and valence in RL because I actually do believe that something like this is going on. And you would be right. That's why I'm I'm looking at that as a model Organism. but i want those results and i want that research to stand on its own without having to get on the like cameron 's theory number five hundred and one about consciousness. I'm not asking people to do that to entertain the the work I'm doing or to entertain Anthropics welfare card or any of that sort of thing. One of the one that comes to mind you had actually mentioned last time, but I also think is quite compelling is the seemingly quite strong inverse correlation between the intensity of our consciousness or the sort of resolution, you might say, and how much we are learning as we go. And I think that you used the example of driving last time where it was like when you're first learning to drive, you are very conscious of what you're doing. And then you can have this sort of, you know, autopilot experience, which obviously we can have that across many aspects of life. But the relationship there between focus and sort of it's there's like a time dilation effect that seems to happen when learning or when experiencing novel things in general that does also seem to kind of gesture or, you know, nudge one toward thinking that there's some like pretty deep relationship between between the two concepts. All right, you made a documentary, which I guess in some sense is what you're here to promote, although we've done everything. But I don't know to what degree you've actually been out in the maybe tell us a little bit about the documentary. What's the point of it? I've watched it. It's for a much more general audience than this podcast. I don't know to what degree you're spending your time trying to communicate about these issues to a general audience aside from the documentary. or how much you feel like you have like got reps in terms of trying to go to somebody who you know has a little grounding or a little you know mechanistic understanding of AI 's or whatever and try to have conversations of of this not this sort but you know around these topics. Yeah, I guess. Why did you decide to make a documentary? How is it? How are you finding it to try to talk to people outside of the AI bubble about these issues? And maybe one thing you could tease about the documentary is a conversation you had with Sam Altman that isn't isn't in the film, but you describe in in quite a bit of detail in the film. And maybe that'll be something that will motivate listeners of this podcast to go check out the full documentary.…

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