WEBVTT - Karen Hao on the Reckless Race for Global AI Power 0:00:02.720 --> 0:00:07.200 Bloomberg Audio Studios, podcasts, radio news. 0:00:10.320 --> 0:00:13.360 These companies have consolidated a historic amount of economic and 0:00:13.360 --> 0:00:16.320 political power, just like the empires of old. You do 0:00:16.400 --> 0:00:19.520 not need to unlock the benefits of AI through a 0:00:19.640 --> 0:00:23.920 corrosive approach. So why are we not holding these companies 0:00:23.960 --> 0:00:25.239 to those higher standards? 0:00:26.360 --> 0:00:30.360 Karen Howe, who says the AI industry is a reckless 0:00:30.520 --> 0:00:34.360 race for global domination, I think you do have a 0:00:34.400 --> 0:00:36.640 problem with the people at the top. Right. 0:00:36.880 --> 0:00:40.480 We often obsess over the cast of characters at the top, 0:00:40.600 --> 0:00:42.960 who's the good guy, who's the bad guy, who's the 0:00:43.320 --> 0:00:46.040 less bad guy? And what I'm saying is that's not 0:00:46.159 --> 0:00:49.920 the only option. We shouldn't have to live with companies 0:00:49.960 --> 0:00:53.479 that are so vastly powerful, where one person at the 0:00:53.479 --> 0:00:55.120 top is able to decide all these things. 0:00:57.200 --> 0:01:02.040 From Bloomberg Weekend. This is the Michelle Hussein Show. I'm 0:01:02.120 --> 0:01:09.679 Michelle Hussein. I don't know if you at the point 0:01:09.720 --> 0:01:15.200 that you tap a prompt into Chat, GPT or CLAWD, Gemini, 0:01:15.400 --> 0:01:19.080 Microsoft co Pilot, whatever you use, if you ever stopped 0:01:19.080 --> 0:01:22.160 for a moment and wonder how the data got there 0:01:22.520 --> 0:01:25.560 in the first place. I suspect you don't. Few of 0:01:25.640 --> 0:01:30.560 us do. These AI tools have been revolutionary in the 0:01:30.640 --> 0:01:32.880 last few years in the way that so many of 0:01:32.959 --> 0:01:35.920 us have changed how we work or how we learn. 0:01:36.440 --> 0:01:40.560 But Karen Howe, who's talked to me for this episode, 0:01:41.040 --> 0:01:44.240 wants us all to step back and to really think 0:01:44.360 --> 0:01:48.720 about what the process involves. That is, not only the 0:01:48.840 --> 0:01:53.760 land used for the data centers, the water, the energy involved, 0:01:54.160 --> 0:01:58.559 but the actual process of gathering and preparing the data, 0:01:58.640 --> 0:02:01.720 where it comes from. Who does the work in putting 0:02:01.720 --> 0:02:06.000 it together. Karen trained as a mechanical engineer and then 0:02:06.080 --> 0:02:10.280 did work in Silicon Valley before turning to journalism. She's 0:02:10.320 --> 0:02:14.160 written a book called Empire of Ai, and it's that 0:02:14.480 --> 0:02:17.640 use of the word empire to describe the top AI 0:02:17.800 --> 0:02:20.920 companies that I ask her to explain right at the 0:02:20.960 --> 0:02:24.639 start of this It started, I think, when she spent 0:02:24.720 --> 0:02:28.840 a few days inside Open Ai observing them. They were 0:02:28.840 --> 0:02:31.400 not a well known company at the time she wrote 0:02:31.400 --> 0:02:36.080 about them in a less than flattering way. Today, it 0:02:36.200 --> 0:02:38.840 is safe to say that there is no love lost 0:02:38.919 --> 0:02:42.720 between Karen Howe and the top execs of the AI world, 0:02:43.120 --> 0:02:45.240 and you'll get a sense of that in my written 0:02:45.280 --> 0:02:49.000 notes that accompany this conversation. As ever, you'll find them 0:02:49.000 --> 0:02:52.640 at Bloomberg dot com. Forward slash Michelle, you'll get the 0:02:52.760 --> 0:02:57.480 backstory to this conversation and what I followed up on afterwards. 0:02:59.160 --> 0:03:03.040 Karen and I in London, where there was something on 0:03:03.080 --> 0:03:06.040 the table that caught her eyes she came into the studio. 0:03:06.560 --> 0:03:09.240 It's a different way that I make notes ahead of 0:03:09.240 --> 0:03:14.200 a conversation, with a spider's web style diagram that reminds 0:03:14.240 --> 0:03:17.359 me what I really need to cover. So that moment 0:03:17.560 --> 0:03:20.600 between us is where this begins. 0:03:21.280 --> 0:03:23.200 I love this mind map. Do you do this? Oh? 0:03:23.240 --> 0:03:24.320 Yes, I do do this. 0:03:24.320 --> 0:03:25.440 This is awesome. 0:03:26.320 --> 0:03:28.920 I do this, but I wish my writing was a 0:03:28.919 --> 0:03:29.800 little bit Nita. 0:03:30.320 --> 0:03:33.320 Oh it's so me. I can just about. 0:03:34.000 --> 0:03:37.280 I can just about, but no, it helps me. But 0:03:37.360 --> 0:03:40.520 I never did it before this show began. 0:03:40.880 --> 0:03:45.280 I'm always so fascinated by the format in which people 0:03:45.280 --> 0:03:47.160 take notes, because it really is such a window into 0:03:47.200 --> 0:03:50.960 how they they think. Yeah, and I'm such a linear 0:03:51.080 --> 0:03:53.680 thinker that diagram confuses me. 0:03:54.080 --> 0:03:56.160 Interesting, So what does this reveal about me that I 0:03:56.200 --> 0:03:58.920 think in a I don't think in a linear way, 0:03:59.000 --> 0:04:00.120 in a non linear way. 0:04:00.280 --> 0:04:03.080 Way I think, which I think is probably a lot 0:04:03.120 --> 0:04:04.200 more information done. 0:04:04.360 --> 0:04:08.040 Essentially it's what enables recall because you know you don't 0:04:08.040 --> 0:04:10.480 want to be referring to it too much as whatever 0:04:10.520 --> 0:04:13.800 helps imprint it in your in your brain. So at 0:04:13.880 --> 0:04:18.440 least I feel it's hope question very good. You're reading 0:04:18.520 --> 0:04:23.960 upside down now? Yeah? Okay, Well, Karen, I would love 0:04:24.040 --> 0:04:28.800 to start with your own words, because you write in 0:04:29.080 --> 0:04:31.560 Empire of Ai that over the years you found only 0:04:31.600 --> 0:04:35.080 one metaphor that encapsulates the nature of what these AI 0:04:35.560 --> 0:04:40.120 power players are, and that is empires not engaged in 0:04:40.200 --> 0:04:43.479 overt violence and brutality, as we know from history, but 0:04:44.400 --> 0:04:48.080 they too seize and extract precious resources to feed their 0:04:48.200 --> 0:04:53.440 vision of artificial intelligence. Now, that is a real attack 0:04:53.800 --> 0:04:56.960 on the companies behind some of the defining tools of 0:04:56.960 --> 0:05:00.000 our age. What is it that's made you feel so strong? 0:05:00.720 --> 0:05:04.919 These companies have consolidated historic amount of economic and political power, 0:05:05.000 --> 0:05:08.919 and they are becoming fast the supreme powers in the world, 0:05:09.000 --> 0:05:11.080 just like the empires of old. And they do this 0:05:11.240 --> 0:05:16.560 through the dispossession of resources, labor value from the majority. 0:05:17.240 --> 0:05:19.560 They claim to resources that are not their own, the 0:05:19.640 --> 0:05:22.680 data of individuals, the intellectual property of artists, writers and creators. 0:05:23.320 --> 0:05:27.440 They exploit an extraordinary amount of labor through both the 0:05:27.480 --> 0:05:31.280 workers that they use and exhaust to produce their technologies 0:05:31.320 --> 0:05:34.200 as well as the workers whose jobs get automated away 0:05:34.279 --> 0:05:37.880 by the deployment of these technologies. So there is this 0:05:38.040 --> 0:05:40.720 category of work that is involved in the production of 0:05:40.760 --> 0:05:45.200 these technologies, including content moderation, including the labeling and cleaning 0:05:45.320 --> 0:05:48.680 of training data. That's called data work. That kind of 0:05:48.760 --> 0:05:51.760 work is now the number four fastest growing work in 0:05:51.800 --> 0:05:55.000 the US. And the types of people who are doing 0:05:55.040 --> 0:06:00.960 this work are scientists, lawyers, college graduates, PhD graduates, and 0:06:01.279 --> 0:06:05.520 they are essentially providing their knowledge and their expertise to 0:06:05.600 --> 0:06:09.400 the training of these models, but under deeply exploitative, poorly 0:06:09.480 --> 0:06:12.840 paid conditions where they are not working nine to fives. 0:06:13.279 --> 0:06:17.040 They are in their homes waiting for work to arrive. 0:06:17.120 --> 0:06:18.760 They never know when it's going to arrive. They never 0:06:18.839 --> 0:06:20.640 know when it's going to go away. They're being pitted 0:06:20.680 --> 0:06:23.480 against each other. You know, I've been covering this kind 0:06:23.520 --> 0:06:27.520 of work for almost eight years now, and it's the 0:06:27.560 --> 0:06:30.720 same story again and again with every single community that 0:06:30.760 --> 0:06:34.080 it hits. You see the same exact patterns of exploitation. 0:06:34.360 --> 0:06:36.280 Are you saying that we should be thinking about the 0:06:36.320 --> 0:06:39.719 exploitation of labor at the point that we're using clawed 0:06:40.080 --> 0:06:40.800 or chatt GPT. 0:06:41.120 --> 0:06:45.240 Yeah. Absolutely, So other industries have had similar problems. The 0:06:45.279 --> 0:06:48.320 fashion industry also engaged in a lot of labor exploitation 0:06:48.800 --> 0:06:52.839 and a lot of environmental degradation. The answer was never 0:06:53.040 --> 0:06:56.520 let's just get rid of clothes altogether. The answer was 0:06:56.560 --> 0:07:01.640 always let's hold ourselves to higher standards in producing clothes 0:07:01.720 --> 0:07:03.200 without those problems. 0:07:03.400 --> 0:07:05.479 It's still work, though, isn't it. And there are so 0:07:05.560 --> 0:07:07.880 many communities where in different parts of the world there 0:07:07.920 --> 0:07:12.160 is an absence of work altogether. I just want to 0:07:12.200 --> 0:07:16.080 ask you whether you think that there are definitely challenges 0:07:16.160 --> 0:07:19.760 in this industry and your work really details them and 0:07:19.800 --> 0:07:22.720 deconstructs them. But is there an aspect of this that 0:07:23.080 --> 0:07:26.440 is about the challenges of our age that it's not 0:07:26.480 --> 0:07:30.520 necessarily more exploitative than what's gone before. It's the problems 0:07:30.520 --> 0:07:31.080 of our time. 0:07:31.320 --> 0:07:33.000 I want to go back to what you were saying 0:07:33.000 --> 0:07:37.200 that it is work though, because what is happening with 0:07:37.560 --> 0:07:41.880 this US case now is that the people are actually 0:07:42.360 --> 0:07:45.680 people that are highly educated and could have gotten full 0:07:45.720 --> 0:07:48.880 time employment, but the reason why they do not have 0:07:48.960 --> 0:07:53.640 access to better work now is also because of the 0:07:53.640 --> 0:07:54.960 AI industry. 0:07:54.560 --> 0:07:55.400 Taking away jobs. 0:07:55.600 --> 0:07:58.240 It's because of the way that the AI industry designs 0:07:58.280 --> 0:08:02.040 its technology and sells it as a tool for automating 0:08:02.080 --> 0:08:06.880 knowledge work. And so there's a shrinking job market for 0:08:07.040 --> 0:08:10.200 the unemployed population in the US twenty five percent of 0:08:10.200 --> 0:08:15.480 them our four year college degree holders. That's a historic record. 0:08:16.240 --> 0:08:19.040 And so it's not actually just giving work to people 0:08:19.040 --> 0:08:24.120 who don't have work. It's actually degrading the work opportunities 0:08:24.160 --> 0:08:28.520 that exist right now, taking knowledge workers out of a 0:08:28.640 --> 0:08:33.520 full time employment kind of situation into a gig work 0:08:33.679 --> 0:08:34.400 kind of situation. 0:08:34.600 --> 0:08:37.320 What do you say to a country like Malta, which 0:08:37.360 --> 0:08:41.720 has just signed a major agreement with OpenAI where everyone 0:08:41.760 --> 0:08:43.800 who lives in Walta is going to get free access 0:08:43.800 --> 0:08:45.800 to chat GPT plus for a year. They're going to 0:08:45.800 --> 0:08:47.280 go on a course to learn how to use it. 0:08:47.520 --> 0:08:49.480 And the government of Malta is very proud of this 0:08:49.640 --> 0:08:52.360 agreement because it says it puts Malta at the full 0:08:52.440 --> 0:08:54.360 front of the digital age. 0:08:54.880 --> 0:08:58.480 I've met a lot of policymakers from developing countries, from 0:08:58.520 --> 0:09:04.720 smaller countries that do see themselves as they can only 0:09:04.760 --> 0:09:07.760 participate in this revolution if they offer themselves up for 0:09:07.840 --> 0:09:10.680 sale just because you know, the government is proud of 0:09:10.679 --> 0:09:13.280 this partnership and they think that this is what gives 0:09:13.280 --> 0:09:15.400 them the seat at the table, that it's actually leading 0:09:15.400 --> 0:09:19.359 to their citizens having a more dignified life. 0:09:19.200 --> 0:09:24.760 But their government would see it as equipping them for work, 0:09:25.040 --> 0:09:29.160 or access to knowledge or everything else that they will 0:09:29.280 --> 0:09:31.679 need to use large language models for. 0:09:32.320 --> 0:09:34.240 I would argue they are being exploited, and I would 0:09:34.240 --> 0:09:37.120 be curious what the multi people think, not what the 0:09:37.160 --> 0:09:40.480 multi government thinks. And the reason why I think they 0:09:40.480 --> 0:09:43.320 are being exploited is because they are having their data 0:09:44.040 --> 0:09:48.480 harvested to train the next generations of models. And that's 0:09:48.600 --> 0:09:51.480 you know, that's why these companies are going out and 0:09:51.480 --> 0:09:55.839 striking these partnerships. They need more and more and more data. 0:09:55.920 --> 0:09:58.080 They need more and more and more data centers to 0:09:58.120 --> 0:10:01.280 produce their technologies. And one of the challenges when it 0:10:01.320 --> 0:10:04.760 comes to searching for different sources of data is that 0:10:05.120 --> 0:10:08.560 they have run out of Internet data. They've run out 0:10:08.559 --> 0:10:12.240 of high quality Internet data, and so now one of 0:10:12.280 --> 0:10:15.680 the best sources of high quality data is straight from 0:10:15.679 --> 0:10:18.760 the user. And so if they can strike these whole 0:10:18.840 --> 0:10:21.120 of country partnerships, I mean you get a whole of 0:10:21.160 --> 0:10:24.880 country database to upload freely all. 0:10:24.720 --> 0:10:26.880 Kinds of thing like everything that's if I was part 0:10:26.920 --> 0:10:29.480 of it, it would be everything that's on my phone, which 0:10:29.520 --> 0:10:30.800 is which is. 0:10:30.840 --> 0:10:33.960 Every Yeah, everything you used. You know. Some people think 0:10:34.080 --> 0:10:37.520 of this as just oh, well I was using Google. 0:10:37.720 --> 0:10:39.440 I was, you know, saying a lot of things to 0:10:39.440 --> 0:10:42.520 Google and personal data as well. But I mean, I've 0:10:42.559 --> 0:10:46.440 never seen people upload their medical files to Google to 0:10:46.600 --> 0:10:50.440 ask for an interpretation of it. But people do that 0:10:50.480 --> 0:10:57.120 with CHATGBT, with Gemini, with Claude, and that's intimate medical information. 0:10:57.640 --> 0:10:59.720 Is there no positive side of it for you, Karen? 0:10:59.800 --> 0:11:02.440 Like that uploading of medical data, if it's done with consent, 0:11:02.640 --> 0:11:05.679 isn't it part of the progress of science. You might 0:11:05.720 --> 0:11:08.400 get something that troubles you diagnosed, or you might be 0:11:08.480 --> 0:11:12.360 part of helping wid diagnosis. 0:11:12.800 --> 0:11:15.960 What I have problems with is the fact that it 0:11:16.080 --> 0:11:20.040 is in the hands of deeply powerful companies that have 0:11:20.200 --> 0:11:24.119 absolutely no accountability to the people that they're actually affecting 0:11:24.200 --> 0:11:26.920 around the world. And this is why I call them empires. 0:11:27.120 --> 0:11:30.160 That's the governance structure of an empire. You at the 0:11:30.240 --> 0:11:32.920 top are able to make hundreds of decisions in a 0:11:33.000 --> 0:11:35.640 day that then have cascading effects on billions of people 0:11:36.000 --> 0:11:39.079 around the world, and there are no formal mechanisms by 0:11:39.120 --> 0:11:43.760 which those billions of people can actually challenge, contest, provide feedback. 0:11:44.120 --> 0:11:47.839 That's why we've historically moved from empires to democracies, because 0:11:47.840 --> 0:11:54.200 democracies are these institutionalized ways for people to say when 0:11:54.280 --> 0:11:57.160 they like something, when they don't like something, and how 0:11:57.240 --> 0:12:01.080 they want to collectively govern and self determine their future. 0:12:01.200 --> 0:12:03.720 Are you actually saying that the power of these companies 0:12:03.760 --> 0:12:05.520 is greater than the power of governments. 0:12:06.160 --> 0:12:10.760 They are definitely increasingly becoming that way, which. 0:12:09.960 --> 0:12:13.160 Means even the US government, the most powerful country in the. 0:12:13.080 --> 0:12:15.400 World, I would say so. I mean, when there was 0:12:15.440 --> 0:12:18.160 the recent spat between Andthropic and the Department of War, 0:12:19.320 --> 0:12:23.520 the Department of War used the nuclear option to try 0:12:23.559 --> 0:12:26.240 and get a company to fall in line, which is 0:12:26.400 --> 0:12:29.959 to threaten to declare this company as a supply chain risk, 0:12:30.400 --> 0:12:31.000 which would. 0:12:30.800 --> 0:12:33.040 Affect its work right across the US government. 0:12:33.600 --> 0:12:36.880 The company Anthropic still did not fall in line. 0:12:37.240 --> 0:12:40.120 But isn't that because it did well in another way. 0:12:40.240 --> 0:12:42.040 It took the moral high ground, and maybe we can 0:12:42.080 --> 0:12:44.839 talk about that a bit later, but it got lots 0:12:44.840 --> 0:12:49.760 more users from the pr of that momentolutely. 0:12:49.120 --> 0:12:52.120 One in the situation, both pr and in the SPAT. 0:12:52.120 --> 0:12:55.079 What I'm saying is this was a moment that demonstrates 0:12:55.559 --> 0:12:58.480 that these companies can, in fact have more power than 0:12:58.520 --> 0:13:01.479 the US government. They are will to actually go against 0:13:02.000 --> 0:13:06.000 the US government because they know that whatever penalty comes 0:13:06.440 --> 0:13:11.080 at them from the government might actually be counterbalanced by 0:13:11.240 --> 0:13:14.640 the other benefits that come directly to the company. 0:13:14.240 --> 0:13:17.360 By profit elsewhere. So tell me, then, Karen, what is 0:13:17.400 --> 0:13:19.440 the answer, Because I know there's a call to action 0:13:19.520 --> 0:13:21.760 in your book. You clearly think that things can be 0:13:21.920 --> 0:13:25.800 and should be done another way. What are you hoping for? 0:13:26.120 --> 0:13:28.600 And do you think a figure like the Pope, like 0:13:29.360 --> 0:13:34.079 a unique figure way outside the sort of government policy realm, 0:13:34.480 --> 0:13:36.600 is the kind of person who might provide some kind 0:13:36.600 --> 0:13:37.560 of roadmap here. 0:13:37.720 --> 0:13:41.000 I think it's important for all leaders in society, whether 0:13:41.040 --> 0:13:44.520 religious or otherwise, to be thinking about these questions and 0:13:44.559 --> 0:13:48.360 thinking about how to ultimately protect human agency at a 0:13:48.360 --> 0:13:51.400 time when the way that Silicon Valley conceives is this 0:13:51.520 --> 0:13:55.240 technology is in fact threatening it. And we've seen Popelio 0:13:55.520 --> 0:13:58.760 be quite strong on these issues and identify early on 0:13:58.840 --> 0:14:02.760 in his papacy that this is a particular issue where 0:14:02.800 --> 0:14:06.800 he is deeply concerned about the pathway of technology development 0:14:06.960 --> 0:14:10.319 undermining human dignity. And to go to your question of 0:14:10.400 --> 0:14:12.079 what is actually the answer here, I mean, one of 0:14:12.120 --> 0:14:14.280 the things that I always say is AI is like 0:14:14.320 --> 0:14:19.120 the word transportation, transportation refers to everything from bicycles to rockets, 0:14:19.200 --> 0:14:21.560 and in the same way, AI refers to a very 0:14:21.640 --> 0:14:26.240 vast collection of different technologies, some of which are designed 0:14:26.640 --> 0:14:31.360 kind of like rockets. They're very resource intensive and they 0:14:31.400 --> 0:14:34.360 exact a lot of cost to produce, and so in general, 0:14:34.920 --> 0:14:38.240 we only deploy rockets for a very narrow subset of 0:14:38.240 --> 0:14:41.240 transportation needs. We never go around and say every single 0:14:41.240 --> 0:14:43.760 person should have a rocket and use that rocket for 0:14:43.880 --> 0:14:46.480 every single one of their transportation needs. That would destroy 0:14:46.520 --> 0:14:49.640 the environment, and it would also just be really inefficient 0:14:49.760 --> 0:14:53.880 and probably not actually get people to places faster because 0:14:53.920 --> 0:14:56.080 of the amount of time it would take to boot 0:14:56.160 --> 0:15:00.120 up the rocket. And so what I suggest is in 0:15:00.440 --> 0:15:03.440 my book and throughout my other work, is that we 0:15:03.520 --> 0:15:07.920 need to shift our portfolio of AI technologies to better 0:15:08.000 --> 0:15:11.320 reflect the kinds of specialized AI that we in fact 0:15:11.400 --> 0:15:16.520 need to tackle real societal challenges, things like improving healthcare 0:15:16.520 --> 0:15:20.320 and accelerating drug discovery. Deep Minds alpha fold it is 0:15:20.360 --> 0:15:22.760 an example of a system that I call a bicycle 0:15:22.880 --> 0:15:26.360 of AI. It is a system that predicts with high 0:15:26.400 --> 0:15:29.480 accuracy how an amino acid sequence will fold into a 0:15:29.480 --> 0:15:33.600 protein structure. This is really important for accelerating the understanding 0:15:33.640 --> 0:15:37.239 of human disease and then designing drugs to target those diseases, 0:15:37.640 --> 0:15:40.360 and it won the Nobel Prize for chemistry. This is 0:15:40.400 --> 0:15:44.360 a technology. Even though it's called AI and Chatchibt's also 0:15:44.400 --> 0:15:47.760 called AI. Deep Mind's alpha fold is one that is 0:15:47.920 --> 0:15:53.000 fundamentally different from how chatchebt is produced and how it operates. 0:15:53.360 --> 0:15:57.680 It uses very little data, it uses very little computing resources, 0:15:58.200 --> 0:16:01.360 and there's so hewpful and there's no need for labor 0:16:01.440 --> 0:16:04.880 exploitation either, and it's still able to provide a lot 0:16:04.880 --> 0:16:07.680 of benefits. So my question is, when you look at 0:16:07.680 --> 0:16:11.360 the portfolio of AI technologies, wouldn't you want to pick 0:16:11.400 --> 0:16:14.040 the ones that have great benefit very little cost. Why 0:16:14.040 --> 0:16:15.760 would you want to pick the ones that have very 0:16:16.040 --> 0:16:17.200 great costs? Right? 0:16:17.240 --> 0:16:21.440 But these are value judgments, right, And something that you 0:16:21.600 --> 0:16:23.840 think has no value, someone else might like. Let's take 0:16:23.880 --> 0:16:27.680 AI in defense. We know already that many armies use 0:16:27.760 --> 0:16:31.520 it for targeting and the use of weapons. You and 0:16:31.560 --> 0:16:34.560 I might see that as wrong. Others might see that 0:16:34.600 --> 0:16:37.560 this is necessary defense. It's part of the defense of 0:16:37.600 --> 0:16:39.520 your realm, and if you're not going to do it 0:16:39.600 --> 0:16:41.520 through the tools of our age, then you're going to 0:16:41.600 --> 0:16:45.640 fall behind and the bad actors, terrorists, bad governments, whatever 0:16:45.680 --> 0:16:46.360 are going to get you. 0:16:47.480 --> 0:16:51.400 Yeah. I think that's a particularly controversial issue where people 0:16:51.400 --> 0:16:54.320 fall in different places. But what I'm arguing is there's 0:16:54.480 --> 0:16:58.080 many different ways to produce AI. You do not need 0:16:58.120 --> 0:17:01.760 to unlock the benefits of AI through a corrosive approach. 0:17:02.280 --> 0:17:05.840 So why are we not holding these companies to those 0:17:05.920 --> 0:17:08.240 higher standards. 0:17:24.720 --> 0:17:28.359 I'm conscious that both anthropic and open AI are heading 0:17:28.400 --> 0:17:31.720 towards public offerings of their shares. Is that a moment 0:17:31.760 --> 0:17:36.399 that you think could bring greater accountability, that once a 0:17:36.480 --> 0:17:39.040 company is public and it's got all kinds of investors 0:17:39.080 --> 0:17:42.679 and it's got annual general meetings, that a level of 0:17:42.720 --> 0:17:45.840 scrutiny that is lacking comes in a useful one. 0:17:46.600 --> 0:17:50.040 I do think it could be. I am hopeful that 0:17:50.160 --> 0:17:54.720 it could be, because when companies become public, they do 0:17:54.840 --> 0:17:59.560 institute more government structors. But I Laigh this critique on 0:18:00.040 --> 0:18:03.760 existing public companies as well. I call Google, Amazon, Microsoft 0:18:03.800 --> 0:18:07.119 empires of AI as well. I do think that ultimately 0:18:07.240 --> 0:18:11.320 governance needs to come from the public and from publicly 0:18:11.359 --> 0:18:14.439 elected officials as well, not just the market forces and 0:18:14.520 --> 0:18:15.480 not the founders. 0:18:15.520 --> 0:18:18.040 It sounds like you don't have faith in the actual 0:18:18.040 --> 0:18:21.840 people running these companies. A lot of your book is 0:18:21.880 --> 0:18:25.439 about open Ai, and Sam Mortman in particular. Tell me 0:18:25.480 --> 0:18:27.840 the story of how you got to know him and 0:18:27.920 --> 0:18:30.320 the company back in the beginning of this work. 0:18:31.160 --> 0:18:35.520 I started covering AI in twenty eighteen for MIT Technology Review, 0:18:35.680 --> 0:18:38.520 and that is a publication that focuses a lot on 0:18:38.680 --> 0:18:42.240 fundamental AI research. So I was reading scientific papers every week, 0:18:42.280 --> 0:18:47.280 talking with researchers at universities and corporate labs and open Ai. 0:18:47.520 --> 0:18:50.240 You know, it became part of my radar because at 0:18:50.280 --> 0:18:53.920 the time it was a nonprofit fundamental AI research lab, 0:18:54.000 --> 0:18:56.320 and it was one of the only ones. There was 0:18:56.560 --> 0:18:58.640 open Ai, then there was Google, then there was deep 0:18:58.640 --> 0:19:02.280 Mind and just a few others, but that was kind 0:19:02.320 --> 0:19:06.320 of it. And so in twenty nineteen I proposed to 0:19:06.960 --> 0:19:09.800 open Ai that there were a lot of changes happening 0:19:09.800 --> 0:19:12.480 at the organization. They had just created a for profit 0:19:12.680 --> 0:19:15.439 arm they'd just gotten a big investment from Microsoft, that 0:19:15.520 --> 0:19:18.360 maybe they wanted to reintroduce themselves to the public, and 0:19:18.920 --> 0:19:21.199 I suggested that if they were interested in that, I 0:19:21.200 --> 0:19:25.160 could profile them. So I went to the corporate offices 0:19:25.400 --> 0:19:28.760 and embedded within the organization for three days, interviewed a 0:19:28.760 --> 0:19:32.119 lot of employees and executives, including Sam Wltman. Sam Wtman 0:19:32.200 --> 0:19:34.159 was actually not very important at the time, so no, 0:19:34.840 --> 0:19:38.440 because he had just officially become CEO. He was previously 0:19:38.480 --> 0:19:40.640 just co chairman and didn't really have much to do 0:19:41.080 --> 0:19:44.120 with the organization at all. And so they set up 0:19:44.200 --> 0:19:47.720 interviews for me with Greg Brockman, the chief technology officer, 0:19:47.760 --> 0:19:50.880 in Ilios Setskaver, the chief scientist, who were the main 0:19:51.040 --> 0:19:53.480 runners of the organization's day to. 0:19:53.520 --> 0:19:55.480 Day and they didn't like the profile you wrote. 0:19:55.640 --> 0:19:56.960 They did it, they they did not. 0:19:57.440 --> 0:20:00.159 You said at that time that there was a misalign 0:20:00.240 --> 0:20:03.760 between what they were actually doing and their promises. What 0:20:03.840 --> 0:20:05.119 was it that made you say that? What did you 0:20:05.200 --> 0:20:06.320 see that led you to that. 0:20:06.760 --> 0:20:11.160 At that time? The misalignment that I saw was they're 0:20:11.160 --> 0:20:14.040 called open a eye for what they said was a 0:20:14.080 --> 0:20:19.440 commitment to transparency, and this organization was already deeply secretive, 0:20:19.920 --> 0:20:23.119 and I noticed this culture in the way that employees 0:20:23.200 --> 0:20:27.199 were very very nervous about what they said to me 0:20:27.560 --> 0:20:30.800 and where I was allowed to go within the offices. 0:20:31.080 --> 0:20:34.640 So that seemed already quite bizarre. And they also said 0:20:34.680 --> 0:20:37.880 as a nonprofit that they would never commercialize their technologies, and. 0:20:37.840 --> 0:20:40.000 Now, of course they're fully for profit, which is the 0:20:40.040 --> 0:20:43.000 source of the case that collapsed between Elon Musk and 0:20:43.080 --> 0:20:46.239 Sam Ortman. Is Elon Musk then the good guy in 0:20:46.280 --> 0:20:47.480 this for you? 0:20:47.520 --> 0:20:52.320 No? No? I mean, here's the thing. I think we 0:20:52.440 --> 0:20:56.320 often obsess over the cast of characters at the top, 0:20:56.400 --> 0:20:59.520 and we're always trying to think, like, who's the good guy, 0:20:59.560 --> 0:21:02.600 who's the bad guy, who's the less bad guy that 0:21:02.680 --> 0:21:05.520 I'm able to settle with? Because that's all the option 0:21:05.600 --> 0:21:08.359 set there is. And what I'm saying is that's not 0:21:08.520 --> 0:21:11.680 the only option set. We shouldn't have to live with 0:21:12.119 --> 0:21:16.320 companies that are so vastly powerful where one person at 0:21:16.320 --> 0:21:18.200 the top is able to decide all these things. It's 0:21:18.200 --> 0:21:21.479 a governance structure that's the problem. It's not that I 0:21:21.480 --> 0:21:24.040 think the problem will be solved if we just swap 0:21:24.160 --> 0:21:26.439 out one of these characters for one another. Or for 0:21:26.560 --> 0:21:27.240 a new person. 0:21:27.320 --> 0:21:29.159 But I think you do have a problem with the 0:21:29.200 --> 0:21:32.199 people at the top, right or certainly the small number 0:21:32.280 --> 0:21:34.080 of people at the top, And you kind of know 0:21:34.240 --> 0:21:37.560 the world that they came from because you were part 0:21:37.600 --> 0:21:39.720 of that same world. You went to MIT work for 0:21:39.760 --> 0:21:42.879 a startup. This could actually have been the industry of 0:21:42.920 --> 0:21:45.960 your life. It could have been reporting from the outside. 0:21:46.040 --> 0:21:48.159 So I get the impression that you do have a 0:21:48.200 --> 0:21:49.400 problem with these characters. 0:21:49.920 --> 0:21:53.680 Yes, I do have a particular problem with the culture 0:21:53.720 --> 0:21:57.240 that they're steeped in and the ideologies that they have, 0:21:57.480 --> 0:22:00.879 because it's an imperialist ideology. I mean, you know Sam 0:22:00.920 --> 0:22:04.320 Altman said before he stepped into his position as CEO 0:22:04.320 --> 0:22:07.080 of Opening Eye, who's the president of y Combinator, and 0:22:07.520 --> 0:22:10.720 he said, I'm going to invest in ten x more 0:22:10.760 --> 0:22:13.320 companies year after year until I invest in every company 0:22:13.400 --> 0:22:18.520 under the sun. And there's this kind of sense of 0:22:18.600 --> 0:22:20.800 all the people that are steeped in that Silicon Valley 0:22:20.840 --> 0:22:22.840 culture which I was a part. 0:22:22.640 --> 0:22:24.960 Of, which was pretty young at the time, though maybe 0:22:24.960 --> 0:22:28.080 it was hubris foolishness. 0:22:27.840 --> 0:22:31.080 Except that it's echoed through the way that he approaches 0:22:31.119 --> 0:22:34.720 Opening Eye strategy. Lets scale the models tenx more every 0:22:34.840 --> 0:22:38.879 year until we have reached the scale of the human brain. 0:22:39.040 --> 0:22:42.120 You know, like He's always thinking in these kinds of terms, 0:22:42.160 --> 0:22:43.800 and all of them are thinking in these kinds of 0:22:43.880 --> 0:22:46.800