“It certainly seems like a very high-stakes bet. You're betting that one kind of AI will get better at a faster rate than a different kind of AI. I'm not sure we have any prior evidence that can really give us a good sense of whether that's likely to be the case.”

…if AIs become very powerful. We won't have ways to be able to measure how good they are at things, how deceptive they are, whether they're lying to human beings. It'll just be very, very difficult to do any oversight over them. And so that's the theoretical future that people are worried about. And

…right now. But if we have a super capable AI researcher, maybe AI will be able to crack that problem first or at least make significant improvement in that area, which I don't know if I buy that argument, but yeah, just putting that out there as something that kind of the labs have put forward.

Yeah, certainly. And I want to talk a little bit about kind of reactions to this report. The AI Futures Project, which is the group that previously put out AI 2027 in AI 2040, wrote a blog post

responding to the report,

explaining kind of how this updates their timelines