WEBVTT - Why the US Must Engage China on AI Safety Before It’s ‘Game Over’ 0:00:02.720 --> 0:00:08.840 Bloomberg Audio Studios, podcasts, radio news. The fact is China 0:00:08.880 --> 0:00:11.399 has cutting edge AI labs. There are only six months 0:00:11.400 --> 0:00:14.920 behind the frontier in the US. And if they get 0:00:15.160 --> 0:00:18.080 a really powerful model that can do cyber hacking, for example, 0:00:18.680 --> 0:00:20.919 and they put it in their systems and their systems 0:00:20.920 --> 0:00:23.600 are open source, it's game over. This is just going 0:00:23.640 --> 0:00:26.160 to proliferate over all around the world. So that is 0:00:26.200 --> 0:00:28.560 a very bad outcome. So let's try and be creative 0:00:29.200 --> 0:00:32.160 about at least trying with China to talk to them. 0:00:32.200 --> 0:00:35.279 You're going to need both sides to slow down. But 0:00:35.360 --> 0:00:37.920 my point about the discussion in China, which I've came 0:00:37.960 --> 0:00:40.800 into touch with when I was there, is that there 0:00:40.840 --> 0:00:44.159 is at least a discussion and openness to think about safety. 0:00:44.200 --> 0:00:46.599 The door is opened, just a crack, and so it's 0:00:46.680 --> 0:00:49.080 up to American negotiators to go there and try and 0:00:49.240 --> 0:00:50.560 give it a nudge and open it more. 0:00:59.400 --> 0:01:02.280 I'm Stephanie Lander's head of Government and Economics at Bloomberg, 0:01:02.560 --> 0:01:05.319 and this is Trumpnomics, the podcast that looks at the 0:01:05.360 --> 0:01:08.640 economic world of Donald Trump, how he's already shaking up 0:01:08.680 --> 0:01:11.839 the global economy. Modern Earth is going to happen next 0:01:15.640 --> 0:01:18.839 last week we looked ahead to this week's summit between 0:01:18.880 --> 0:01:22.399 the US and Chinese presidents and recording this on Tuesday, 0:01:22.480 --> 0:01:24.759 May twelfth, and by the time you listen to this, 0:01:25.040 --> 0:01:28.759 President Trump should already have touched down in Beijing ready 0:01:28.800 --> 0:01:32.920 for what feels like a high stakes encounter on Thursday 0:01:33.360 --> 0:01:37.360 with President Chi Jinping. Even though the expectations for concrete 0:01:37.360 --> 0:01:40.800 policy outcomes, as we heard last week, are firmly under 0:01:40.800 --> 0:01:43.679 control now. Along with trade, one of the big issues 0:01:43.760 --> 0:01:48.040 hanging over that conversation will be US China competition in 0:01:48.120 --> 0:01:52.600 advanced technologies and the race to dominate AI. Tesla's Elon 0:01:52.720 --> 0:01:54.840 Musk and Apples. Tim Cook are in the group of 0:01:54.880 --> 0:01:58.120 top executives Trump decided to invite with him to China, 0:01:58.200 --> 0:02:01.400 and there is some mild speculation that President Trump might 0:02:01.560 --> 0:02:06.160 further loosen the export controls on US cutting edge semiconductors 0:02:06.200 --> 0:02:10.440 for Chinese companies. But the safety of this race, protecting 0:02:10.480 --> 0:02:15.079 the world from the misuse of these increasingly powerful tools 0:02:15.520 --> 0:02:19.160 both countries are now capable of producing, well, that doesn't 0:02:19.200 --> 0:02:22.080 seem to be high on the agenda. This week's guest 0:02:22.120 --> 0:02:26.400 on Trumpnomics thinks that is a big and potentially dangerous 0:02:26.600 --> 0:02:30.320 missed opportunity. Sebastian Mallaby is a senior fellow at the 0:02:30.320 --> 0:02:33.760 Council and Farm Relations, author of multiple award winning books, 0:02:34.040 --> 0:02:39.040 and most recently of The Infinity Machine. Demisseivis Deep Mind 0:02:39.360 --> 0:02:42.880 and the Quest for Superintelligence. I sat down with him 0:02:42.919 --> 0:02:45.400 in our New York studio a few weeks ago now 0:02:45.600 --> 0:02:48.480 to talk about his new book, The Future of AI 0:02:49.000 --> 0:02:52.040 and the Role of China, which he has recently visited. 0:02:52.840 --> 0:02:55.560 I think you'll find what he had to say fascinating, 0:02:56.040 --> 0:03:06.440 whether hopeful or pretty alarming. Well, i'lli that to you. Sebastian. Welcome, 0:03:06.480 --> 0:03:08.880 Thank you very much for being here in the flesh 0:03:08.919 --> 0:03:12.160 in the New York studio, which is always fun. There's 0:03:12.160 --> 0:03:15.520 so many themes coming out of your book, The Infinity Machine. 0:03:15.840 --> 0:03:17.920 I wanted to start with something that we talk about 0:03:17.960 --> 0:03:21.080 a fair bit on Trumpanomics, which is the impact of 0:03:21.120 --> 0:03:25.760 AI on the economy, society, I guess especially jobs. But 0:03:25.880 --> 0:03:28.640 I think given the themes of the book, and I 0:03:28.639 --> 0:03:30.799 know things that you've been thinking about since it came out, 0:03:31.240 --> 0:03:34.160 will move on fairly quickly to the question of safety. 0:03:34.440 --> 0:03:36.240 How worried would we be about the pace of the 0:03:36.280 --> 0:03:39.960 improvement in AI's capacity, but also how likely is it 0:03:40.040 --> 0:03:43.800 that governments or anyone else will be able to bend 0:03:43.840 --> 0:03:48.480 it to their will? Humanities will So first thing first, 0:03:48.520 --> 0:03:51.040 we had a conversation with darren Os and Moglu about 0:03:51.080 --> 0:03:54.080 it recently. I mean economists. I've noticed they tend to 0:03:54.120 --> 0:03:58.280 divide into those who say, AI is going to have 0:03:58.280 --> 0:04:02.119 a very radical impact on business models, on the economy, 0:04:02.360 --> 0:04:06.600 on society, but that's going to take a long time 0:04:06.720 --> 0:04:09.760 and we will have time to prepare. There's another group 0:04:09.760 --> 0:04:11.800 who say, no, no, this is all going to happen 0:04:11.880 --> 0:04:16.279 much sooner, but it will probably be less seismic. You 0:04:16.360 --> 0:04:20.160 had all these years researching this book really up close, 0:04:20.320 --> 0:04:23.000 looking at the evolution of this technology. What's your best 0:04:23.040 --> 0:04:23.640 guess on that? 0:04:24.440 --> 0:04:26.719 Yeah. I think the second view, which is that there'll 0:04:26.760 --> 0:04:29.240 be an explosion in the power of the models and 0:04:29.240 --> 0:04:32.479 it will immediately affect the economy, is put about by 0:04:32.600 --> 0:04:36.080 the technologists who are just focused on the algorithms and 0:04:36.160 --> 0:04:38.080 on the models and what the base model can do, 0:04:38.839 --> 0:04:43.160 and it completely underestimates the rollout challenges. If you think 0:04:43.200 --> 0:04:47.120 about this idea of recursive self improvement, which preoccupies a 0:04:47.120 --> 0:04:49.400 lot of people in Silicon Valley, once you've got a 0:04:49.480 --> 0:04:52.520 machine that is good enough to write the code for 0:04:52.600 --> 0:04:55.440 the next iteration of the machine, and then it recursively 0:04:55.480 --> 0:04:59.239 self improves. It'll just explode upwards. And that is true 0:04:59.279 --> 0:05:02.160 so far as the computer science goes. It's not remotely 0:05:02.200 --> 0:05:04.280 true as far as actually applying it to the real 0:05:04.320 --> 0:05:07.920 economy goes, because you need to build the computing capacity 0:05:07.960 --> 0:05:10.880 to serve those models. So you need all the chips, 0:05:11.040 --> 0:05:12.920 and there's a shortage of chips. You then need all 0:05:12.960 --> 0:05:16.240 the power to fuel all those chips. The people who 0:05:16.240 --> 0:05:18.400 are in the center of this a talking about having 0:05:18.400 --> 0:05:21.159 to launch data centers in space. That just gives you 0:05:21.200 --> 0:05:23.080 a sense of how difficult this is going to be. 0:05:23.480 --> 0:05:26.480 And that's not to say anything about the institutional frictions 0:05:26.839 --> 0:05:30.520 of integrating AI into say a law firm or a 0:05:30.560 --> 0:05:33.480 bank or whatever. I mean, that's a non trivial challenge. 0:05:33.600 --> 0:05:35.240 You've got to get all your clients to sign off 0:05:35.279 --> 0:05:38.120 on the confidentiality of their information and so forth. So 0:05:38.160 --> 0:05:40.240 I do think that rolling it out such that it 0:05:40.279 --> 0:05:43.159 makes a difference in terms of job displacement or productivity 0:05:43.800 --> 0:05:45.719 is more kind of a decade plus than it is 0:05:45.760 --> 0:05:47.280 a kind of three to five year thing. 0:05:47.520 --> 0:05:50.360 And on your point about the exponential improvement, I guess 0:05:50.360 --> 0:05:52.400 the other thing and the colleagues who are much more 0:05:52.400 --> 0:05:55.840 focused on this than the economics team have also said 0:05:55.839 --> 0:05:58.120 this to me that when you think about sort of 0:05:58.160 --> 0:06:00.520 pie chart of all the kinds of time asks that 0:06:00.640 --> 0:06:04.839 humans do that are part of the workplace, that exponential 0:06:04.880 --> 0:06:08.240 improvement is really only in a certain subset of those tasks, 0:06:08.240 --> 0:06:10.080 and there are other areas where they're really quite far 0:06:10.120 --> 0:06:12.640 behind that are quite important for jobs. Yeah. 0:06:12.760 --> 0:06:15.080 Another way of saying the same thing is that there 0:06:15.080 --> 0:06:18.160 are sectors which are probably going to be quite difficult 0:06:18.640 --> 0:06:20.880 to disrupt, and the public sector is pretty slow at 0:06:20.880 --> 0:06:24.279 adopting tech, and that's a pretty big chunk of some economies, 0:06:24.360 --> 0:06:29.760 right construction, healthcare. There's a whole slew of economic sectors 0:06:30.320 --> 0:06:33.320 which don't seem to be very right for disruption at all. 0:06:33.800 --> 0:06:35.480 And then the ones that are right for disruption the 0:06:35.560 --> 0:06:39.440 archetypical people working on a screen doing knowledge work and analysis, 0:06:39.440 --> 0:06:42.760 which is very similar to what AI does. As I say, 0:06:42.839 --> 0:06:46.800 you've got to get the customers okay with it, and 0:06:46.839 --> 0:06:48.200 that's quite a challenge. 0:06:48.800 --> 0:06:51.039 Railways is the example that's often used, and it takes 0:06:51.080 --> 0:06:54.200 a long time for business models to adapt to that. 0:06:54.520 --> 0:06:56.800 If I think of something like the smartphone, which is 0:06:57.040 --> 0:07:01.599 extraordinarily new by historical standards, less than a decade old, 0:07:02.880 --> 0:07:05.440 day to day, all of us have been transformed by that. 0:07:05.600 --> 0:07:08.000 So is it comparable to that that it's not necessarily 0:07:08.040 --> 0:07:10.720 having a big change on the economy or on productivity, 0:07:10.800 --> 0:07:13.720 but day to day we're all being affected really quite quickly. 0:07:14.320 --> 0:07:15.760 Yeah. I think there's a lot of truth of that. 0:07:15.760 --> 0:07:18.880 That people will be using it. They already are using it. 0:07:18.880 --> 0:07:22.920 It'll occupy quite a lot of their headspace, but it's 0:07:22.960 --> 0:07:25.400 not necessarily making them more productive. One of my favorite 0:07:25.440 --> 0:07:30.560 stories about the anthropic agent agentic system, which was out 0:07:30.640 --> 0:07:33.640 in January of this year. This is before Mythos, but 0:07:33.720 --> 0:07:36.240 the earlier one. People got very excited about, Oh, this 0:07:36.320 --> 0:07:39.800 agent can organize my email, it can create a bespoke 0:07:39.920 --> 0:07:42.520 new summary for me in the morning. And so a 0:07:42.560 --> 0:07:46.440 friend of mine was implementing all these productivity enhancing hacks, 0:07:46.880 --> 0:07:49.440 and finally, after doing about four of these, she said 0:07:49.440 --> 0:07:53.240 to the agent, would you please now sync up my 0:07:53.480 --> 0:07:57.960 calendar such that you check that the train timetables on 0:07:58.080 --> 0:08:00.320 average are not too late, or some sort of extra 0:08:00.320 --> 0:08:05.120 productivity thing to which Anthropic answered or Claude answered, Are 0:08:05.160 --> 0:08:07.280 you sure you're not just giving me this taskause you're 0:08:07.280 --> 0:08:10.920 trying to avoid your own work, you know? So, I mean, 0:08:11.960 --> 0:08:15.360 I do think there's a difference between the excitement around 0:08:15.400 --> 0:08:18.280 it and the adoption of it to actually make you 0:08:18.320 --> 0:08:22.080 more productive. And then even in science, Alpha fold, which 0:08:22.320 --> 0:08:25.760 was the deep mind system published in twenty twenty end 0:08:25.760 --> 0:08:28.160 of twenty twenty for which de mister Sabis got the 0:08:28.200 --> 0:08:32.360 Nobel Prize, which unraveled protein folding. This was said at 0:08:32.360 --> 0:08:35.959 the time to be the breakthrough for faster drug discovery, 0:08:36.200 --> 0:08:39.080 and it probably will be over the longer term. But 0:08:39.280 --> 0:08:41.600 now we're six years later, there isn't a single drug 0:08:41.600 --> 0:08:44.520 that's been discovered thanks to alpha fold. Not to say 0:08:44.520 --> 0:08:46.840 that there hasn't been progressed, because now there's a much 0:08:46.880 --> 0:08:50.679 more sophisticated build out of alpha fold, with other aspects 0:08:50.679 --> 0:08:54.480 of the drug discovery process being automated by AI. So 0:08:54.480 --> 0:08:57.080 I think on a kind of ten to fifteen year horizon, 0:08:57.280 --> 0:09:00.760 yes it will massively accelerate drug discovery, but on a 0:09:00.760 --> 0:09:01.960 six year horizon hasn't. 0:09:02.080 --> 0:09:04.080 That's a really interesting example. There is a wonderful one 0:09:04.120 --> 0:09:07.920 of the good documentaries about that whole period. It's almost 0:09:07.920 --> 0:09:09.760 a sort of companion to your book or a part 0:09:09.800 --> 0:09:11.840 of your book. But there's a wonderful bit of footage 0:09:11.880 --> 0:09:15.320 where they have decided to just identify all of the 0:09:15.360 --> 0:09:17.840 proteins and then and just put them onto the Internet 0:09:17.880 --> 0:09:20.720 and have them allow researchers anywhere to access them. And 0:09:20.760 --> 0:09:24.120 they're watching in real time the uploads from all over 0:09:24.160 --> 0:09:26.480 the world on a map of the world, and it's 0:09:26.600 --> 0:09:28.360 very moving because you're sort of seeing in India and 0:09:28.400 --> 0:09:30.280 Africa or all these researchers a lot more than they 0:09:30.320 --> 0:09:32.400 thought in the space of a few hours uploading all 0:09:32.440 --> 0:09:35.400 this information. But it's fascinating to hear that they have 0:09:35.520 --> 0:09:38.040 not actually produced any new drugs on the back of it. 0:09:38.200 --> 0:09:40.280 I mean that goes to something and we move into 0:09:40.360 --> 0:09:43.920 the broader sort of policy and safety territory a little bit. 0:09:43.920 --> 0:09:48.120 But Darrenus and Moglu was on the show recently and 0:09:48.880 --> 0:09:51.079 I was checking in with him because we talked to 0:09:51.120 --> 0:09:55.440 him over time on this very issue, and his focus 0:09:55.840 --> 0:09:59.959 and his advice to governments had been to think about 0:10:00.080 --> 0:10:02.760 out you can't alter the nature of the technology, but 0:10:02.880 --> 0:10:05.760 you can change the way it's adopted, and that it 0:10:05.840 --> 0:10:07.920 was adoption was going to be the main thing that 0:10:08.040 --> 0:10:15.000 determined how jobs were affected, whether you're enhancing productivity of 0:10:15.120 --> 0:10:18.959 humans or merely just allowing businesses to replace humans. I 0:10:19.040 --> 0:10:21.400 later have a conversation with Jason Furman about that, the 0:10:21.480 --> 0:10:23.600 former head of the Council of Economic Advisors, and he 0:10:23.600 --> 0:10:25.920 thought that was just completely wishful thinking that you can't 0:10:25.920 --> 0:10:28.240 tilt the technology even to that extent that you can 0:10:28.360 --> 0:10:31.520 encourage it to be more complementary to labor as opposed 0:10:31.559 --> 0:10:33.959 to just substituting for labor. So what do you think 0:10:33.960 --> 0:10:34.320 about that? 0:10:35.160 --> 0:10:36.880 You know, I think in a race dynamic where you 0:10:36.880 --> 0:10:40.439 have multiple labs in multiple countries producing this stuff, the 0:10:40.480 --> 0:10:43.400 idea that you're going to control all of them in 0:10:43.440 --> 0:10:48.199 some direction that makes the technology complementary to labor, not replacing, 0:10:49.120 --> 0:10:51.120 does strike me. I'm kind of on Jason's side on that. 0:10:51.320 --> 0:10:53.120 The other thing to say is that the labs are 0:10:53.120 --> 0:10:55.960 going to create what the consumers want, and if the 0:10:56.040 --> 0:10:59.960 consumer's meaning in some cases the business consumers, the enterprises 0:11:00.880 --> 0:11:03.320 have an opportunity to replace labor, they want to take that, 0:11:03.480 --> 0:11:06.640 and if they don't, their competitors and some other country will, 0:11:06.679 --> 0:11:10.600 And so that the laws of competition and capitalism, I 0:11:10.640 --> 0:11:12.360 think are very much on Jason's side. 0:11:12.559 --> 0:11:14.679 I think part of that is this dynamic which is 0:11:14.720 --> 0:11:18.480 also very relevant to the book of the race with 0:11:18.559 --> 0:11:22.360 other countries, particularly the race with China. But if you're 0:11:22.400 --> 0:11:26.000 a country, if you're the UK for example, or could 0:11:26.040 --> 0:11:28.199 be any European country that's not necessarily going to be 0:11:28.200 --> 0:11:31.880 at the forefront of the technology, but clearly does want 0:11:31.920 --> 0:11:35.040 to be at the forefront or close to the front 0:11:35.520 --> 0:11:38.760 in adoption in a way that also helps the economy, 0:11:38.880 --> 0:11:42.720 helps their workers. Is there any advice from your book 0:11:42.840 --> 0:11:45.760 or from the research that you've done that you would 0:11:45.760 --> 0:11:47.080 have for government who are trying to do that? 0:11:47.160 --> 0:11:48.760 Yeah, very much. I mean, I think there is a 0:11:48.760 --> 0:11:52.400 lot of commentators, especially in Britain, draw exactly the wrong 0:11:52.480 --> 0:11:55.760 lesson about the story of deep Mind. So the classic 0:11:56.040 --> 0:12:00.000 complaint is deep Mind is a British company founded in London, 0:12:00.000 --> 0:12:03.120 and main founder is British. She's very patriotic about it. 0:12:03.240 --> 0:12:06.199 He refuses to move to Silicon Valley, and it turns 0:12:06.200 --> 0:12:09.199 out to be the most consequential sort of pioneer AI 0:12:09.280 --> 0:12:12.880 lab in the world. That's amazing, But these idiots sell 0:12:12.920 --> 0:12:15.800 it to Google in twenty fourteen and it becomes a 0:12:15.840 --> 0:12:18.960 subsidiary of an American company, and that's terrible, and I 0:12:19.000 --> 0:12:21.920 think that's wrongheaded in the sense that really the sale 0:12:21.960 --> 0:12:25.120 to Google by deep Mind in twenty fourteen was a trick, 0:12:25.240 --> 0:12:28.319 one could say, to get the American parent to pour 0:12:28.400 --> 0:12:31.240 nearly a billion dollars a year of research and development 0:12:31.240 --> 0:12:34.320 money into London, with enormous spill of effects for the 0:12:34.400 --> 0:12:37.320 tech ecosystem in London. So I think that's a win 0:12:37.480 --> 0:12:40.760 for Britain. The company stayed in London, the founder stayed 0:12:40.760 --> 0:12:44.560 in London. That's great for Britain. What was terrible for Britain, 0:12:44.679 --> 0:12:48.360 and about which Britain should beat itself up mightily, is 0:12:48.400 --> 0:12:50.400 that My book also tells the story of deep Mind 0:12:50.440 --> 0:12:54.000 trying to help the National Health Service, and both Demis's 0:12:54.040 --> 0:12:56.400 mother trained as as a nurse, his co founder, me 0:12:56.480 --> 0:13:00.480 Stuffer Sulliman, his mother trained as a nurse. For this reason, 0:13:00.480 --> 0:13:03.319 they really wanted to help the NHS. They were willing 0:13:03.360 --> 0:13:05.520 to do it pro bono for the first five years, 0:13:06.000 --> 0:13:09.400 and they started helping and rolling out AI algorithms that 0:13:09.440 --> 0:13:13.520 recognized early signs of blindness, early signs of cancer. This 0:13:13.679 --> 0:13:16.040 was just super positive, helpful stuff that was going to 0:13:16.080 --> 0:13:19.200 save the NHS time and help h based patients, and 0:13:19.280 --> 0:13:23.600 it got blown up by a stupid backlash over privacy 0:13:23.679 --> 0:13:25.959 or privacy depending on which side of the Atlantic youuron, 0:13:26.520 --> 0:13:28.400 which I mean, I really kicked the tires on this. 0:13:28.679 --> 0:13:32.920 The complaint about leaks of patient data were fictitious and 0:13:33.320 --> 0:13:37.560 so Britain because of this political climate of suspicion of tech, 0:13:37.960 --> 0:13:41.280 particularly suspicion of the subsidiary of the American tech company, 0:13:41.800 --> 0:13:45.640 created such a backlash against this collaboration between the NHS 0:13:45.679 --> 0:13:49.840 and deep Mind that both the doctors and the computer 0:13:49.920 --> 0:13:53.800 scientists at deep Mind got cold feet and the collaboration collapsed, 0:13:53.800 --> 0:13:56.920 and Britain could have been really the cutting edge country 0:13:56.920 --> 0:13:59.640 in the world on medical AI and they blew it. 0:14:00.080 --> 0:14:03.040 Putting together the two pieces of what you said, they 0:14:03.080 --> 0:14:05.360 are related because one of the reasons that there was 0:14:05.360 --> 0:14:09.560 a lack of trust was the foreign ownership, the US ownership, 0:14:09.600 --> 0:14:14.360 and particularly the association with major tech companies in Silicon 0:14:14.440 --> 0:14:18.760 Valley who have not necessarily earned our trust. Many people 0:14:18.800 --> 0:14:21.960 would say, over the last twenty years, they've made promises 0:14:22.000 --> 0:14:26.120 about privacy, about sharing of information that have been broken. 0:14:26.160 --> 0:14:30.040 Every single tech company has broken them, including Google. They 0:14:30.160 --> 0:14:34.160 have changed their terms on which even they were developing AI. 0:14:34.600 --> 0:14:37.480 In different cases there were promises even that had been 0:14:37.480 --> 0:14:41.920 built originally into the deep Mind Charter around safety and 0:14:41.960 --> 0:14:44.920 other things, which then got overridden as they needed just 0:14:45.000 --> 0:14:47.880 more and more money from the parents. So it seems 0:14:47.920 --> 0:14:50.200 to me, yes, in principle it would have been great 0:14:50.240 --> 0:14:53.800 to have the UK on the cutting edge, but also 0:14:53.920 --> 0:14:57.480 completely understandable that having given tech companies a benefit of 0:14:57.480 --> 0:15:00.320 the doubt of various other times and rather regressed that 0:15:00.320 --> 0:15:01.840 they wouldn't want to do it in the case of 0:15:01.880 --> 0:15:03.240 this very sensitive subject. 0:15:03.560 --> 0:15:05.880 Yeah, that's all true. At the same time, it would 0:15:05.880 --> 0:15:08.080 have been nice if the critics looked at the specifics 0:15:08.160 --> 0:15:11.240 or what was going on in this collaboration between deep 0:15:11.280 --> 0:15:13.040 Mind and the National Health Service, and if they'd look 0:15:13.080 --> 0:15:15.840 more carefully the specifics, they would have seen that this 0:15:16.000 --> 0:15:18.560 was positive for Britain. And if you just have a 0:15:18.600 --> 0:15:21.280 blanket view that anything to do with American tech is 0:15:21.360 --> 0:15:25.320 ipso facto a matter of suspicion because there's been bad 0:15:25.360 --> 0:15:27.920 behavior in other instances, I'm not in favor of that. 0:15:28.040 --> 0:15:30.360 I think that Britain did miss an opportunity there, but 0:15:30.400 --> 0:15:32.480 I think the issue that arises out of this is 0:15:32.520 --> 0:15:35.760 that it's in the interest of the tech companies, especially 0:15:35.760 --> 0:15:40.600 now with this super powerful technology AI, to get governance right, 0:15:40.640 --> 0:15:45.720 to be proactive in inviting governments with democratic legitimacy to 0:15:46.280 --> 0:15:49.560 hold sway over the technology. Because if you just take 0:15:49.600 --> 0:15:53.720 it unto yourself as a youngish startup and you say 0:15:53.840 --> 0:15:56.720 this is how healthcare is going to evolve, I'm doing it, 0:15:56.720 --> 0:15:58.760 it's not legitimate and people won't accept it. 0:15:58.880 --> 0:16:01.120 I mean, some of this gets to that sort of 0:16:01.120 --> 0:16:05.520 broader question of our degree of control and also the 0:16:05.560 --> 0:16:08.480 degree to which any of the people involved in this 0:16:08.760 --> 0:16:11.920 can shape the outcome in line with what they themselves 0:16:11.920 --> 0:16:14.920 want when they're in the competitive context. You have a 0:16:15.000 --> 0:16:17.400 very striking phrase in your book towards the end, as 0:16:17.720 --> 0:16:21.080 inventors dream of shaping the technology that they create, often 0:16:21.120 --> 0:16:24.640 the technology shapes them. The technology plus the business, political, 0:16:24.720 --> 0:16:28.040 and geopolitical currents that it unleashes. There's a bit in 0:16:28.080 --> 0:16:30.880 your book where it starts to feel I felt like 0:16:30.920 --> 0:16:33.680 it's some of the historical analysis that you see of 0:16:33.720 --> 0:16:37.360 the start of World War One. You have everyone individually 0:16:37.720 --> 0:16:43.400 in your story had potentially the right incentives. Was very 0:16:43.400 --> 0:16:48.320 focused on being careful, identifying risks ahead of time, managing 0:16:48.360 --> 0:16:52.240 them ahead of time, not inflicting things on society before 0:16:52.280 --> 0:16:54.840 we were ready for them. The end result has not 0:16:54.960 --> 0:16:56.160 been that at all. 0:16:56.840 --> 0:17:00.000 That's right. The central character in my book, demister Sabez, 0:17:00.680 --> 0:17:04.040 is a good person, has good values, but whether he 0:17:04.080 --> 0:17:06.960 can actually do good even if he is good, is 0:17:07.000 --> 0:17:11.000 the central conundrum. And he really wanted to make things safer. 0:17:11.160 --> 0:17:13.280 That's why he fought a secret three year war with 0:17:13.359 --> 0:17:17.680 Google trying to force them to put more safety guardrails 0:17:17.720 --> 0:17:21.679 around AI. That's why he suggested to Rishie Sunak the 0:17:21.760 --> 0:17:25.680 idea of a Bletchley Park AI safety conference, which happened 0:17:25.720 --> 0:17:28.800 in late twenty twenty three. So he does care about safety, 0:17:29.200 --> 0:17:33.640 but when other labs are willing to go fast, there 0:17:33.680 --> 0:17:36.840 isn't much a single AI leader can do. And what, 0:17:36.960 --> 0:17:38.840 of course that points you to is that the collective 0:17:38.880 --> 0:17:41.560 action problem needs to be solved by government. 0:17:41.760 --> 0:17:43.880 But what we've seen in a lot of these cases 0:17:44.000 --> 0:17:48.480 is the researchers involve the companies involved say, yes, we 0:17:48.520 --> 0:17:50.919 think it's very important to have a collective agreement. But 0:17:51.040 --> 0:17:54.920 if a government comes along with concrete rules, whether it's 0:17:54.960 --> 0:17:57.199 the I'm sure you probably feel strong about but the 0:17:57.240 --> 0:18:02.160 EU Internet Safety Rules or AI Safety rules. President Biden 0:18:02.240 --> 0:18:05.480 had a rather different set of rules. There tends to 0:18:05.480 --> 0:18:08.960 be the response, how could governments possibly understand this technology? 0:18:09.040 --> 0:18:11.600 We want some agreement, but not this one. So how 0:18:11.600 --> 0:18:12.240 do you solve that? 0:18:12.560 --> 0:18:15.159 Yeah, look, you're right that you know at the moment 0:18:15.320 --> 0:18:18.800 there is lots of state level initiatives to regulate AI 0:18:19.640 --> 0:18:22.560 in the US, and for example, open AI is actually 0:18:22.560 --> 0:18:25.359 funding the opposition to those regulatory efforts. So there is 0:18:25.440 --> 0:18:27.159 hypocrisy sometimes. 0:18:26.680 --> 0:18:28.919 And after the fact they often say yes, actually it's 0:18:29.000 --> 0:18:30.760 quite good and we've adopted it, but they never want 0:18:30.800 --> 0:18:31.480 it in advance. 0:18:32.119 --> 0:18:34.879 Yeah. On the other hand, there is a counterexample, which 0:18:34.960 --> 0:18:38.159 is that in the Biden administration when they set up 0:18:38.200 --> 0:18:43.640 the AI Safety Institute, and it included a mandatory requirement 0:18:43.760 --> 0:18:46.520 for all the frontier labs to submit their models to 0:18:46.560 --> 0:18:49.919 this institute for review. The institute didn't have a veto, 0:18:49.960 --> 0:18:52.199 which it should have done in my view, but still 0:18:52.320 --> 0:18:56.080 it was mandatory to submit. And I've spoken to the 0:18:56.240 --> 0:19:00.720 Biden people who created that executive order and they say 0:19:00.800 --> 0:19:04.119 that there was zero resistance from the labs. All the 0:19:04.200 --> 0:19:07.760 labs work more in the mode of saying, look, you 0:19:07.840 --> 0:19:10.120 need to be aware something's coming down the pike. It's 0:19:10.200 --> 0:19:12.479 very powerful. You need to get ahead of it. And 0:19:12.520 --> 0:19:14.600 when they did get ahead of it, or they started 0:19:14.600 --> 0:19:17.680 to at least move on it, there wasn't resistance. That's 0:19:17.680 --> 0:19:20.439 what the Biden people say, So I think it depends. 0:19:20.440 --> 0:19:23.080 It's a bit like with Google and deep Mind and 0:19:23.119 --> 0:19:25.760 the National Health Service in Britain. There can be good 0:19:25.760 --> 0:19:29.679 instances of technology company behavior. And again, if we paint 0:19:29.960 --> 0:19:31.639 this is true of anything, like if you paint the 0:19:31.720 --> 0:19:35.320 Chinese as uniformly bad, you can never collaborate with them 0:19:35.320 --> 0:19:38.240 on anything, and that's actually a bad outcome in itself. 0:19:38.840 --> 0:19:40.320 Same with the tech companies. 0:19:40.119 --> 0:19:43.119 Mentioned China in a US context and the group of 0:19:43.160 --> 0:19:45.640 companies that you're writing about in your book. I could 0:19:45.680 --> 0:19:47.399 totally see how you have a kind of circle of 0:19:47.400 --> 0:19:51.240 trust among those companies and shared incentives. But the thing 0:19:51.280 --> 0:19:53.760 that I think God in the way then and gets 0:19:53.800 --> 0:19:56.480 in the way often in practice, is this fear of 0:19:56.560 --> 0:20:00.439 China the outsider, which is not necessarily part of this system. 0:20:00.720 --> 0:20:03.160 And if you want to not be applying the breaks, 0:20:03.160 --> 0:20:04.760 if you want to not have these kind of control, 0:20:04.800 --> 0:20:06.159 all you need to say is what China is going 0:20:06.200 --> 0:20:06.919 to get ahead of us. 0:20:07.119 --> 0:20:10.159 Yeah. So what I think about that is that the 0:20:10.200 --> 0:20:13.320 West really needs to remember what happened in the Cold War. 0:20:14.000 --> 0:20:16.240 What happened in the Cold War in the nineteen fifties 0:20:16.440 --> 0:20:20.520 was you had simultaneously very moments of extreme tension, the 0:20:20.600 --> 0:20:24.399 Hungarian uprising, Soviet tanks going to Hungary. You have the 0:20:24.400 --> 0:20:27.320 Sewis Canal crisis at the same time, which prompts the 0:20:27.320 --> 0:20:30.800 Soviets to actually threaten the West with nuclear attack because 0:20:30.840 --> 0:20:34.360 of the Sewis intervention was viewed as imperiodists by the Russians. 0:20:34.880 --> 0:20:38.080 And simultaneously with those events you had the creation of 0:20:38.160 --> 0:20:44.000 the IAEA, the Institute for Atomic Energy and forgetting them 0:20:44.040 --> 0:20:47.000 but basically counting all the nuclear material and trying to 0:20:47.000 --> 0:20:50.119 prevent the nuclear material from going into loose nukes. Then 0:20:50.160 --> 0:20:53.960 in the sixties you get the Cuban missile crisis, enormously 0:20:54.040 --> 0:20:56.600 tense moment, and six years later you get the Nuclear 0:20:56.640 --> 0:21:00.160 Non Proliferation Treaty. So you can do what Reagan said, 0:21:00.160 --> 0:21:03.760 which is sort of trustpor verified, do arms control, even 0:21:03.840 --> 0:21:06.399 if there is competition with China. And when I went 0:21:06.480 --> 0:21:09.240 to China in March, because my book came out there 0:21:09.320 --> 0:21:12.320 first and spent eight days talking to AI leaders, both 0:21:12.359 --> 0:21:15.639 in academia and in the private sector. It was stunning 0:21:15.720 --> 0:21:19.119 to me how often they raised the subject of safety spontaneously. 0:21:19.560 --> 0:21:22.199 They would say, these agents are becoming powerful. We have 0:21:22.240 --> 0:21:25.040 a window of opportunity before they are too powerful to 0:21:25.160 --> 0:21:28.919 make them aligned with human incentives. The government should be 0:21:28.920 --> 0:21:32.840 warning people about dangers with these models. And in fact, 0:21:32.880 --> 0:21:36.720