“I have yet to find an AI technologist who is convinced that a kill switch is something that they can just code up in a couple of weeks and implement.”

think the idea is you slow down the R D a little bit, but what you do is you give the alignment, which is behind the frontier models, a little bit more time to catch up. But I definitely want to come back to what Emily was saying about alignment, because I think there's two importantly different dangers here. One is that the agents wind up going rogue, spreading over the internet and doing a whole bunch of extremely damaging unintended things. The other is that the agents wind up being controlled by North Korea and doing exactly what North Korea wants them to do. So I feel like nuclear weapons to take a very salient. Analogy are extremely aligned with the interests of the person who holds them. The point is, I am only going to drop a nuclear bomb on a city if I want to drop that nuclear bomb on the city. It is doing exactly what I want it to do. It might not be good for humanity, but it's the person in control of it is in control of it.

That's an interesting point because one of the arguments is that one of the worst case scenarios here is that becomes not the case. But the one area around which I see all these politicians on both sides of the aisle actually coalescing around is the idea that there needs to be a kind of master kill switch.

you tried turning the AI off again to save humanity?

Or just turning it off? Like, I don't think it can be turned off. I think we've already reached the point at which it can't be turned off. You know, there's enough open source models installed on enough computers in enough different countries that you can't turn it off.

A lot of what we're talking about was informed by this really good piece from The Verge called Inside the Suddenly Explosive World of AI Safety. Everyone should just go and read it. That talks about meter and what's the other lab? Redwood?