Tim Scarfe: evaluation

Transcript context

…We have intelligence. We don't have full human intelligence yet. Intelligence in my view is not 1 or 0. Intelligence is a spectrum. You agreed with us earlier. We're talking about viruses and bacteria. They have some kind of intelligence. I think we're scaling mount intelligence with these frontier models. And I think we're absolutely not there on every human capability yet. I think by around 2030, we'll be there with every human capability, which also means with many human capabilities and many non human capabilities, we will be so far beyond that, I'll say this, the transformational change to society will be massive by the late 20 twenties. In a few ways, I'll make predictions right now. Drug discovery is gonna be turned completely on its head because you'll be able to have a 99.9 percent chance of getting the right drug off a render alone, which is fucking nuts. Another 1 is, I just don't think people will have to drive their own cars much past 2028. And like, that's the thing. And kind of to your point, it's gonna happen. And the next day, and Sam Altman has said this, everyone's gonna, yeah, of course fucking cars drive themselves. Like, do you understand how difficult it was to get here? Do you understand how hard it is for a car to drive itself? I I I know, but I'm a little bit skeptical about these kind of utopia views of of the future. We will have many and I call these things coarse graining. We a new technology. And by the way, in many of these cases, like drug discovery, I'm actually interviewing John Jumper tomorrow who won the Nobel Prize. He's at Google DeepMind for his AlphaFold paper. And this is brilliant stuff, but you and I both know how things work in the real world. So we now have this incredible AI that can identify all of these different candidates and so on. And then you have all of these bottlenecks. You have to do clinical trials and RCTs and so on and so forth. So it's just not quite the secret unlock that we think it is. And there are some real technical challenges around doing this continual learning thing in particular because the way we understand things is very, very structured. The way that we do continual learning on neural networks runs rough shot over all of the structured knowledge that, you know, that should be in the system. So there's a lot of technical challenges there. I also wanted to talk about the agency thing here. So 1 of the reasons I make this argument is that Doomers, for example, they have this perspective that we are basically building our successor. Now, I read that Ray Kurzweil book, The Singularity is Nearer. And term singularity, I think it originally came from John von Neumann. And he's famous for inventing cellular automata. Smartest man of all time. Yeah. He's the smartest man of all time. And apparently, he wrote about this idea in private, but he never published it. And then there was this book in the early '80s by Verna Vinge, like A Fire Upon the Deep, where this notion of singularity came. And of course, Ray Kurzweil picked it up and started writing about it from the nineties onwards. But to Kurzweil, interestingly, it's a slightly different notion. It's because I colloquially thought, oh, the singularity just means when we build AI, which is super intelligent and recursively self improving. But he was coming at it more from a transhumanism point of view, which is that basically we will transcend. We will upload our minds. We'll start putting things in our brains and so on. And our intelligence will just recursively self improve. And we'll move to the next level of our evolution. And in a sense, I think, even though I don't agree with it, because I think even with Neuralink, there's bandwidth problems. I think we've already got all the world's information at our fingertips. And as I was just saying, it's not actually as helpful as we thought it would be. But in a sense, I think enhancing ourselves is a more realistic vision of the future than building an agentic AI, which is a completely separate thing. Because the arguments I was saying before that life is a property of matter, agency is something which comes about as this self preservation drive. an agentic AI, which is a completely separate thing. Because the arguments I was saying before that life is a property of matter, agency is something which comes about as this self preservation drive. Intelligence is something which then comes about as a form of modeling the environment and acquiring and accumulating information over time. What we've built at the moment are systems that do this behavioral cloning. They just look at the data that we output and then we can get them to role play agency. But it's not agency, it's not the same thing. So I think there's a fundamental difference there.…

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