Tim Scarfe: evaluation

Transcript context

…It's not alive. Why would you say it's not intelligent? It's solving real world problems at speed. So intelligence is about adaptivity. Right? So the neural network in that Tesla model, it is frozen. A bunch of frozen weights are not intelligent by definition. Now the thing is, I I I wanna I wanna sort of be a little bit, you know, open minded here. I mean, we were just filming at the Diverse Intelligences Summer Institute, and I'm a big believer that there is a huge, as you were saying, there is a huge diverse space of possible intelligences. And the folks at the Santa Fe Institute, way they think about this is that intelligence is something which is represented mostly by adaptivity, but also representation and inference. So I'm amenable to the idea that bacteria are intelligent, viruses are intelligent. I'm also an externalist. I like looking at language and culture. You can actually think of it as being an organism. It's an adaptive organism. Viruses have the ability to delete strategies very, very quickly. And knowledge we're talking about knowledge being non fungible. Knowledge decays very, very quickly. And that's actually a good thing because it's the way that we can adapt our strategies because things that don't work die off. So we're very adaptable. So then the question is, don't believe for the functionalist reason I gave you before, I'm basically an anti functionalist. I don't believe that you could upload your mind into a computer. Even if you did a particle by particle simulation, I don't think that would be intelligent. But that does not imply that I don't think we could build an artificial life simulation which was intelligent. So I I interviewed this guy called Blaze from Google a few weeks ago. And he basically built this adaptive Turing machine where you have all of these different tapes and they could merge parts of the Turing machine tape with each other. And he noticed that after several generations, you know, 100,000 generations, that they started preserving themselves. You have this like self preservation behavior. And the reason why we have self preservation and living systems in the real world is because obviously the ones that kept around were the ones who had this drive to preserve themselves and that led to life and agency and intelligence. So there's this kind of evolutionary track. And intelligence is about the accumulation of course grained knowledge to adapt to the world. So that's the real world intelligence. But then could we build something like that in a computer? Yes, I think we can. But the important thing is that this is a process of evolution. We're only going to do this through neuro evolution methods. What we're not gonna do is you're talking about like this this LLM understanding stuff better than anyone else. That's bullshit. Right? Like, it's it's like a database. If you put a model into a database, right? So you just you sort of defined all of the tables and you gave them names that represented the domain. Would you say that the database understood the domain? Of course, you wouldn't. It would I say that the database understood the domain. I think what you and I are gonna crack into which we probably couldn't discuss on a podcast, we'd have to sit down and draw things out together, is a mutually agreed upon complete definition of what understanding means. Because I'm gonna say understanding is a spectrum. You're going to say it has a few critical things that must be there. I probably am inclined to agree with you, but I don't know how much traction we're gonna get on that until we have like a real vibe session with each other. What I will say to your point is updated live learning is an unbelievably useful feature and to get human level intelligence is a non starter has to happen. But you can be artificially intelligent without exhibiting all human traits and qualities. And so we already have artificial super intelligence in many regards. For example, a computer that can just answer Jeopardy questions is in domain specific artificial super intelligence. If you look at it like that, the Tesla self driving software actually does update live within a tiny context window of that truck is right there, and now it's here. Now it's here. I need to go around that truck because it has a world model that knows roughly what a truck is and knows that that somebody can come out that side and they need to go around it. And so after it leaves the truck behind, it doesn't know shit. So for example, my Tesla will have it on hurry mode a lot, you know, because then you could just dial up how fast it's going. And always it's trying to get into some fucking lane to get you another 2 seconds, which is the mode I turned on, so I'm not upset at it. But a lot of times it like doesn't even look super far ahead, even though I know the camera system can do that. It's it's also non reasoning model, so like it's gonna go into this lane and I'm like, the fucking lane ends in a mile, it should know that, but it doesn't know that. And then it just rediscovers every drive that I have that that shit is coming up. The thing is visual reasoning models are already a thing. They're starting to be a thing. And when we get those visual reasoning models with big context windows and somewhat lossy, but decent long term memory. Now we're getting real fucking close to what you would call understanding. And then we're 1 algorithm away, updated live learning from what you I think would catalog outside of the embodiment problem real close to understanding. And the 1 thing you said that really strikes out at me is if you point by point represent every, I would say, like brain function. Is neuron like, do neurons reason at a sub subcellular level unclear?…

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