Tim Scarfe: belief

Transcript context

…Sure. Sure. They sure. Actually, I take that back. You're completely correct. They didn't crack it, but the path to it is more well understood than it was before, and there's no reason to think we don't we can't crack that, by the way. That's some kind of magic that's happening. Like updating a neural network live while parsing it is a potentially solvable problem onto exactly why humans do it. Remember, humans are primates with dog shit wetware, and we can get machines to do. So far, machines are undefeated in their ability to get ahead of humans in every domain we've ever tried to really push for long enough. And so once especially get updated live learning, what you're going to get and even before them is incrementally more true deep understanding. So for already for concepts like history, concepts like reasoning by analogy, GPT 5 understands more deeply than 99% of all humans. I've conversed with lots of humans and the degree to which understanding at a real depth is shown is not in evidence. And as a matter of fact, it's it's evidence against it. And I'll also say something because Jared is here, I'll give him this 1. When you are talking to some people that you think know things, this is facetious. It's gonna get clipped out of context. I'm being a cunt for comedy, but there's some some tiny grain of truth here. When you talk to people that think really understand a relational map of a concept, because they're embodied and because they've been in the space, and then you start really pressing on the fine points. What about this? What about this? What about that? You get your analogy of the blender graph, where it's like, oh, you're vibing all of this and you know about 3 things about it and the rest you're vibing live. Humans do that. So what I would say to push back is the assumption that humans have this deep understanding is itself often not an evidence and often the evidence is against it. And so when AI vibes its way to understanding, I would say I would say that humans do that too. Like for example, hallucination. People are real pissed that AI hallucinates. Have you ever spoken to a human being? Half of the analogies I used on this very podcast just now talking to you are fucking hallucinations. They're close enough to the truth to be cogent and have some amount of information transfer, but they're not perfect. I'm not deeply in touch with some real causal map of the world. It's all simulation all the way down. And so that idea that we need grounded understanding or whatever that is, I'm highly skeptical of. I think there's something there, but the more capable AI gets in the way that humans are able to really really live update their knowledge, the more capable AI gets in doing that, the less of a moat we have, and then at some point like, now it fucking understands, bro. So a few things you said there. First of all, GPT 4.5, that was a vanilla dense model. The reason why GPT 5 is is faster is it's a mixture of experts. So it's roughly the the same size model, but now you only have a fraction, let's say 10% of the parameters in play during inference. And also it wasn't a reasoning model. It was just a vanilla model and that's why it was so slow. But I don't want to come across like I'm really skeptical. Are a massive AI podcast. Love technology and I'm really interested in AI. Part of the reason for the skepticism is I'm very worried about doomerism and we'll come back to that in a minute. I think that there are principled reasons why we don't need to worry about superintelligence and recursive self improvement. But I do think that intelligence is something that we can create in a computer. But we need to distinguish intelligence as a property of adaptive matter, which is something that happens as part of the process of evolution in the real world from the types of algorithms in computers that we might say are intelligent. AI is operating in the real world. And I can prove it with 1 step. Unplug the reactor from the data center, no more AI. AI is absolutely in the real world. Its data stream was fed to it by layers of abstraction in the same way, much the same way that your data stream arrives to you as representations 8 neural networks deep from your op your optical nerve going all the way back, and by the by the time it hits the back of your occipital lobes, you're not seeing the fucking world. None of us see the world. It's abstraction all the way down. You can't process the photon density at your eyes. Full stop. And so arguably AI actually gets way more coherent data than we do, and so when it's thinking in its data center, it not only is truly embodied in the sense like as much as your brain is it just doesn't have effector arms, but like getting a data center effector arms is like a it's a nominal problem. Right? As a matter of fact, does have 1 because I have a Tesla of which I'm very proud. Yes. It's a it's someone's like, oh hello, can I talk to you? Oh my god, no. The unwashed masses. I have a Tesla. Tesla like Elon says it. It's an s but he says it like a z. My Tesla drives itself. It fucking drives itself. It's a chip inside the car. How the fuck does it know what's around? Cameras. It is already a an intelligent organ and an end as a suite of effector organs, sensory effector and intelligent. But his thing is, it drives it better than I do.…

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