Mike Israetel: belief

Transcript context

…Because we think that this thing is incredibly dangerous and it will want to do bad things to us. But it doesn't have the affordances. You can code value functions. You can hypothetically have agents haven't thought this through too much, but you might have agents that are smart enough to do damage, real serious damage, but not smart enough or able to introspect on their own abilities to do damage and where that leads. They don't have the time horizon for it. That's a real serious problem. I think the only way we mitigate that problem is having agents and architectures organizations that are smarter than those agents and can outplay them. And that's already a thing in the real world. Cybersecurity versus hacking. Yeah. 1 1 thing that baffles me is that people think like, you know, a lot of people thought the Internet was gonna collapse under its own weight from hacking, and it just never happened. I mean, how many gigantic corporations or governments have ever gotten hacked like catastrophically? I mean, it's like a handful, maybe 0. And so it's the same network security problem, but with more capable actors. If you wanna pit humans against AI, it's gonna hack every network because it's gonna get smarter than us eventually. It's we're not the seat of intelligence of the universe. But if you have humans aligned with good AI against bad humans and bad AI, and bad AI is only initially at least gonna be made by not so great humans. And then it's the same problem. It's like you know who wins good guys or bad guys versus who wins good guys or the bad guys with guns versus who wins good guys or the bad guys with nukes. 1 thing that I think is a fucking terrible idea is to regulate our own AI down on capability fearing it will kill us and do nothing about Chinese AI, and then you guarantee that the bad guys get the good stuff. It's like a pacifists are amazing people on vibes. But like if you're a pacifist in The United Kingdom and you succeed in the arming The United Kingdom, it will be The United Kingdom Of The Russian Federation the next day. And then you just you just have to do the same thing again and you're probably not gonna talk Vladimir Putin in a disarming like you did like the Tories or whatever party you guys have here. So it's 1 of these things where is AI new and totally different and totally amazing? Yes. Are any of the game theoretic problems fundamentally different from a network security and physical security architecture perspective? No. I think they're roughly the same. You know, many of your clients and customers, for example, they they can now go and chat GPT. And they can say, I'm I'm taking, you know, TRT. I'm taking this supplement stack. I'm taking all of this stuff. And it'll do something that was very difficult before. So before, might have gone on examine.com. Or they might have…

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