Liron Shapira
Host of Doom Debates, the #1 show for high-stakes debate about AI extinction risk.
“Ten IQ points a year. It's like, okay, you've now surpassed all geniuses and you keep going. I agree with Eliezer Yudkowsky on this. The precondition for losing control really is just having a bunch of robustly 200 IQ AIs. They can siphon off more and more power and never give it back. It's a one-way ratchet.”
Doom Debates! · AI Research & Frontier Labs · November 2026
Context
Discussing whether Anthropic's AI-driven research speedup can continue, Liron Shapira argues that even a linear climb in capability would pass human genius quickly, and describes how he thinks control would be lost.
Transcript
…is also going to yield more than twice as many billions. And so if this cycle continues, you are going to get, at the very least, a linear march upward in capabilities. And given that human IQ is not exactly a super high ceiling here, if you have a steadfast linear march, speak for yourself.
Speak for yourself.
But yeah, I mean, look, 10 IQ points a year. It's like, okay, well, you've now surpassed all geniuses and you keep going. And I can tell you, me personally, I agree with Eliezer Yakowski on this. The precondition for losing control really is, in my opinion, just having a bunch of 200, robustly 200 IQ AIs, I do think they can go off and just kind of have power in a way that's irreversible. Like they can just siphon off more and more power and never give it back. It's just kind of like this one-way ratchet. And, you know, if 200 IQ doesn't cut it, okay, fine. So keep marching up to 300 IQ. Like by that time, combined with their advantages of like, you know, speed, the ability to back themselves up, the ability to coordinate, to work 24-7, to just, you know, be relentless. I think that humanity, you know, unless they love us and they want us to persist, I just, I think we're the weaker force in that battle.
No, I think, look, if you look at this math stuff that was generated, you know, largely by these unreleased models. And I just learned that not nearly all the all the math magic depended on swarms of agents. Most of the problems OpenAI solved was just a single LLM that was given, I think, three…
Right. On something that did
on things that people had worked on for years, or if not decades, unsuccessfully. I mean, and of course, math is just math. And people say, oh, but the proofs are sloppy. You know, it reminds me of when they were saying, oh, yeah, it generates video, but the people have six fingers. And you're…
More from Doom Debates!
