24:32
Real, meaningful action will only happen when something serious enough happens. Either very serious financial losses, or significant human casualties. Simply put, the "Chernobyl" of artificial intelligence. Something has to happen that will change the political dynamics, but a way out of this situation is possible. I don't think it's appropriate to completely halt all AI development, but I think the US can focus more on security and shift the balance between reckless forward movement and security without losing American advantage. And secondly, it's not like there's no regulation at all in China. There are many regulations regarding artificial intelligence in China. They are very
concerned about certain aspects of AI, and I can't believe that the Chinese authoritarian regime thinks that real existential risks are in their interests. So I'm probably more optimistic that over time these two countries will be able to cooperate to some extent. But even before that, I believe it is possible to develop a system of soft regulation that simply shifts the emphasis on security a little bit without jeopardizing the US position. One technical question I need an answer to, Alex: to what extent, if America slows down a bit, will Chinese labs be able to leapfrog ahead, and to what extent are they really dependent on learning from American labs? So it's actually easier for the US
to maintain leadership than you might think. I believe there is a difference between America maintaining leadership and forging a path into entirely new, unexplored areas. The problem is that if America takes a break, we should at least expect China, using the approaches it has already chosen, to catch up with the leading edge. And this difference of three to four months seems to be quite significant lately. We may have only recently seen the release of models from China that could be called "Mythos" level. The open- source GLM 5.3 system from a Chinese lab may only now be reaching the level that Anthropic was at in March of this year in terms of potential damage in the
cybersecurity industry. This means that if America takes a break, it is not just a guarantee that China will not overtake us. This will not eliminate the risks of recursive self-improvement. The problem is that the American front line is already dangerous. We are already seeing real potential for harm. All American laboratories have models that they do not release to the public, but only provide to verified partners. We should expect this milestone to be reached, regardless of what American labs do. So the question is this: you know, this is the most accelerationist view you'll hear from me, do you need to develop stronger AI to deal with the damage from the slightly weaker AI that you've
already released into the world? And what is your answer? That's a good question. I mean, my answer is: It really is like swallowing a spider to catch a fly, isn't it? This is the whole logic of this technology. You need to, you know, you need to be ahead to prevent the Chinese from being ahead. You need to be ahead of the curve to prevent other people from having patterns that cause harm. I just don't see how it can be stopped. That's true. I mean, the way to stop it is to slow down, then slow down even more, and then slow down even more. And at every stage, you demonstrate through these valuable commitments that you actually do what you say you will do.
Regulation is by its nature difficult to administer and is clumsy. How much tolerance is there in this system for regulation? How possible is it to have regulation that is so finely tuned, as you say, to slow us down a little bit, but not too much. This is, well, Come on, come on, let's move on, I mean, this is, as you all know, I think this is the most powerful technology, certainly since the industrial revolution, maybe ever. But, nevertheless, history can help us understand some of these things. And if you look back at previous great inventions, whether it's the automobile or medical innovations, we have a pretty extensive
regulatory system. And we didn't just wake up one day and say, "Here it is, ready." It evolved in response to what was going wrong. It developed gradually. Mistakes were made. There were twists and turns. Things were held back by people like us who believe in not being too intrusive in regulation. And I think you can do the same here. Is there time for this technology? continue. Let's not be too fatalistic. I mean, as you know, I think that significant action will only happen after And by "quite serious" I mean either very serious financial losses or significant human casualties. I think it will happen pretty soon. You know, it's like " Chernobyl for AI" in short, but I hope it won't be that bad. But I think something has to happen
that will change the political dynamic. However, I really believe that a way out of this situation can be found. I don't think we should give up and say that because of the arms race we are inevitably going down a path where there is a 10% chance of human extinction. I think we really have an impact on this. Shashanka? I also want to remind people that we are not choosing between " doing nothing" and " draconian regulation," because the status quo is a kind of backdoor regulation. And what do I mean? Well, first, there's what the administration is doing. We have a de facto licensing regime for the most advanced models. At the very least, Anthropic and OpenAI should be reaching out to the government every week and asking, "Can we give this user
access to our most advanced models?" So, you know, just because it's not spelled out in an act of Congress doesn't mean that such regulation doesn't exist. And one more thing: while we are waiting for this to all end, the states are not waiting. The states are acting. We have bills that have already been passed and become law in New York and California. AI liability laws are being discussed in Rhode Island and other states. This is a patchwork of legislation that is taking shape below the level of federal debate. And if nothing changes in a year, it doesn't mean that there is no regulation. This means you have a patchwork quilt of adjustments that have already accumulated. And I think everyone in
Washington understands that perfectly well. There is a concern here because, as you say, Zanni, let's say you need some kind of " Chernobyl of AI" to get going. The likelihood of proportionate and sensible regulation after the " Chernobyl of AI" instead of perhaps not shutting down everything but something very harsh-is that it forces people to act, but also forces them to overreact. Certainly. Um, and the worse the accident, the more irrational or excessive the reaction can be. Again, if you look at the evolution of automobile or aviation safety, in the past we essentially did everything right. So I don't think it's
inevitable that it will lead to disaster one way or another. But I think it's especially difficult because everything happens so fast. Literally every few weeks, a new opportunity appears that is both exciting and frightening. And I think the next few years, or 18 months, are going to be really seismic in both the political debate and the development of technology. That's why I fear that something noticeable has to happen for the situation to change. It is also worth adding that American politicians have long failed to show themselves as competent adults when it comes to solving serious issues, especially over the last decade. I mean, the nature of American political debate is often just absurd. The problem here is that politicians need to
finally take their jobs seriously. And I think some of the polling results among voters in rich countries are the most interesting. When asked about their confidence in the government's ability to solve this issue, scores are much higher in most other rich democracies. Americans clearly lack confidence that their leaders can make a difference, and that partly explains why the polls have deteriorated so quickly. So, you know, now politicians have to finally do their job. This is important. Let's see if they can.
