00:00Alex Turner, a former Google DeepMind research scientist who also left the firm for similar
00:05ethic concerns. Alex, great to talk to you this morning. And look, there's this big question
00:10about where regulation in the industry will come from. Will it come from the industry itself?
00:15Will it come from politicians? Considering what we're hearing from Trump and from China,
00:20do you think that this is an industry that in any form will slow down and have regulatory
00:24forces stop this march towards what many fear? I think there's a good chance that this will
00:32actually happen. I think this is an idea that kind of shocks first contact, but it's really not in
00:39anyone's interest to allow the technology to proceed at such a breakneck pace. It endangers
00:46Trump. It endangers China. It endangers everyone from my home state of Iowa to the big cities on
00:53the coast. So I think there's a real chance that we can reach an accord here. How are you reading
00:59this? Because Anthropic, OpenAI, SpaceX and all of the big firms are in unison when it comes to calling
01:03for the slowing of the pace. And to your point, it's really for the good of everyone. Do you read
01:07this as a sincere strategy or maybe just a bid for regulatory capture? So I'd like to preface,
01:14I have no equity in any of these companies. I left Google DeepMind months ago at a significant
01:22financial cost. I've been writing about these issues since before big AI was even a thing. Many
01:27people have, including some of these CEOs, have been warning at some point the technology will be
01:33progressing so quickly we may need to slow it down. There will be risks. There will be very serious
01:38risks. So I think a lot of this is very sincere. I think at the same time we need to
01:43be careful in how
01:44we actually implement this so that some kind of regulatory capture doesn't happen. It's very
01:51possible that people say good things on Twitter and then actually in the negotiating room they're
01:56trying to get something that's suspicious. Alex, I just wonder through what form this slows down?
02:00Because if you look at Dario's proposal, for example, it doesn't really get into details about
02:04slowing down necessarily. It talks more about outside forces coming in, giving some sort of
02:10you know, observatory lens that so it's not just anthropic calling for its rivals to do the same
02:15thing. So what does it actually look like to slow down in this context? I mean, there have been
02:20proposals. You can run this stuff in a sandbox, but we know that it can escape the sandbox. So how
02:25does
02:26this actually take form in a way that prevents the worst from happening?
02:30The way AI advances is depends on compute. It depends on GPUs, on TPUs,
02:37and compute has a very narrow, narrowly bottlenecked supply chain. There's been a good
02:45amount of research on how to track compute, how to verify in a potential treaty that compute limits
02:52are being respected. So we know how much compute is entering the market. We know where it's going.
02:58We know whether it's going to projects that are using it in an authorized fashion. So I think what it
03:03would look like is tracking compute, restricting it, still allowing massive innovation and advancement.
03:11Under the proposals I've heard, there would still be what would look like to the American public
03:15very fast advances from AI, but restricting it so that it doesn't suddenly jump to levels that we don't
03:22know how to understand or control.
03:24And then there's China, Alex, because Amodei even acknowledged that this plan wouldn't really work unless
03:29the superpowers in the whole world agrees to slow down. And of course, we have China
03:34saying, criticizing the dire warnings about AI risk, calling them overblown and fear mongering.
03:39How do you square those two?
03:42So I don't really know how much to read into what China says here. It's possible this is more of
03:48a
03:48bargaining position rather than the CCP's sincere belief.
03:53But what I do know is that we don't need to trust China to have a good deal. If China
03:59and the US come
03:59to an agreement, we don't need to trust that they're going to abide by it. There are ways we can
04:03verify
04:05and track compute through the supply chain and punish defection by any party.
Comments