Demis Hassabis, Google DeepMind and the Battle Over AI Safety

20:12Kara Swisher:They will use their collective knowledge to discuss the news, international affairs, climate, culture, sports, lifestyle, fashion, and wellness. Plus, The Guardian wants to emphasize they're not billionaire-owned, meaning they're free to report the whole picture without interference. Go to theguardian.com slash stateside to learn more and listen wherever you get your podcasts or watch on YouTube starting May 13th.

20:37Kara Swisher:Now, he had hoped there'd be one collective effort to develop AI because there wasn't a lot of AI development. As you said, it was a winter of AI at the time. And he had hoped there'd be sort of a Manhattan project with Google and DeepMind at the center. And, you know, one of the early interviews I did with Sam and Musk when they were starting OpenAI was their worry that DeepMind and Google would dominate. Not just them, but possibly Mark, who they had thought little of at the time in that regard. And so he was shattered after the first ethics and safety group meeting in 2015. 2015, Musk hosted it and then teamed up with Sam to start OpenAI when they were getting along.

21:17Kara Swisher:And it was a rival. And it was started. He was testifying honestly for a second when he talked about the original reasons. I happened to write that story when that happened, when OpenAI was created. So talk a little bit about the fallout of the meeting and how it changed his approach to AI development and the way he ran DeepMind. Yeah. I mean, so what happened, as you say, is that there was this safety oversight meeting, which was supposed to be for DeepMind to get useful feedback. And it turned out the people at the meeting were not giving useful feedback. They were kind of learning how DeepMind was doing stuff and then - Collecting information.

21:55So in that meeting, there was both Elon, but also Reid Hoffman, who was one of the funders of OpenAI. And so the reaction was from Google, forget the idea of safety oversight boards. We've tried that once and it was terrible. So we're not going to do that again. And Demis's reaction was, well, we can't give up on the idea of a safety oversight board. It's essential for this technology. And so we're going to have a fight about it. And so now this is, you know, 2016 is when this really gets going. And there's this thing called Project Mario, as they called it internally at DeepMind, where for three years, they retained, you know, armies of lawyers and investment bank strategists and people like that.

22:41and there were term sheets running to 50 pages that flew back and forth between the negotiators in Mountain View and the negotiators in London. And I was shown some of these things.

22:51Kara Swisher:He tried to spin it out. He tried to spin it out. Yeah. I mean, he thought of spinning out because that would be the threat that would force Google to agree to the Safety Oversight Board. And Reid Hoffman promised a billion dollars to finance a spin out. They went to see Joe Tsai, the co-founder of Alibaba. in Hong Kong to try and get some money out of him. It was a real thing. And, you know, what happened after three years is that Demis caved. He was so exhausted by this three years of negotiation. And, you know, you get another long term sheet from the lawyers and he would hold his head in the hands and say, I don't want this.

23:30Kara Swisher:They were never going to let him do it. I don't want this part of my brain to expand, is what Demis wound up saying about legal documents. And so the upshot was, you know, he didn't get what he wanted. Google was never going to do it. I kind of developed a weird respect for Sundar in learning about this process because he was very canny. You know, he knew he didn't want Demis to spin out. He didn't say it quite directly, but he, you know, hid behind his chief counsel and other people who played bad cop on his behalf. And at the end of the day, he got what he wanted, which was to keep Demis inside the tent, not spun out, so that he would be the secret weapon.

24:07Kara Swisher:Right, because he did understand what was happening in Silicon Valley with the competitors, right? Google had to be at the forefront of this. And Sundar, who's a very lovely affect, is really good at playing rope-a-dope, I've always found, you know, in niceness. But let's talk about this idea of what Demis wanted to do here. He says he was naive to believe in a singleton scenario where AI developed collectively, but he also saw himself at the center of the effort. Now, you could say that's egotistical, and it does take a certain amount of narcissism that you should be in charge of all AI development.

24:40Kara Swisher:I have problems with, I am more with the collective, but at the same time, it always devolves into either one person or just a few people developing this thing. But talk about this idea, because they were never going to do it once they understood the financials here, the amount of money and the amount of power they would. Was that naive of him or egomaniacal? He says he was naive. Yeah, I think it was naive of Demis to believe that AI would be developed by one single lab on behalf of all humanity. I mean, humans are just disputatious and tribal and jealous and they don't do that. Greedy, mostly.

25:16That too, probably.

25:17Kara Swisher:I like that you said probably there. No, I said probably because what strikes me, Cara, about this whole field is that the normal cynical explanation that people are doing it for the money is weirdly less true. Because AI is such a weird technology. It attracts people who are basically doing it for power, for ego, for... Power is, yes. It's Promethean. You're right. It's not about money. So, you know, in hindsight, why did he not see this at the time? And I don't think he was doing it for the money. He is a true believer of all these people. And of course, there's the glory. He wants a Nobel Prize.

25:54Kara Swisher:That's pretty great. And he has curiosity. But should he have been less naive? Or explain the naivete, because it seems obvious to see that they were not going to do this. Yeah, I mean, I think there's, you know, there's some, you know, mitigating circumstances that you can cite that make it a bit less crazy to have believed in this single thing. That would be simply that when he was beginning, you really could fit all the AI believers in one conference hall. And so there was a sense of a single community at that point. And then, you know, he starts DeepMind, and it's going really, really well.

26:29and he does this Atari model, which is just an astonishing agent that can learn by itself how to pay dozens of different video games. And that impresses everybody. And then he gets the Google checkbook behind him. And it feels like he's the only game in town. And if there's a second game in town, it would actually be Google Brain in Mountain View. And so that is the same company. And, you know, open AI just wasn't a thing then and didn't really become a thing, you know, until, let's say, 2019. Until recently. Yeah. Yep. So I think it wasn't totally crazy at the time to think, well, I can't see anybody else on the horizon who's doing this, so it's going to be me.

27:14But it's still crazy, still naive, still, you know, immature to think that when you're confronted with a godlike technology, there won't be lots of acolytes trying to do it. and there'll be sectarian splits and you can continue that metaphor.

27:30Kara Swisher:Right. Now, we've heard Altman, Musk and even Mark Zuckerberg claim they care about AI safety. I believe none of them, only to watch them prioritize growth and their own egos of time after time. Is he different? Does he remain different? There's almost no incentive for tech companies like Google to make AI safety their top priority. For example, Google is eager to supply AI to the Pentagon after agreeing not to do so when it bought DeepMind. So while he understands, Demis understands the risk of developing AGI, does he have any more power to contain them in the face of the pressure, both national security, financial, just to win, to be the dominant technology of the next era, etc.?

28:11Yeah, I think Demis has gone a long journey, right, since founding DeepMind 16 years ago. So, you know, at the beginning, he thought there might be a singleton scenario. Then when that started to break down because OpenAI, you know, launched a competitor, he still hoped to get oversight for his technology, as we discussed with Project Mario negotiating. And then when that didn't work, he said, well, at least I'll make AI good for humanity by doing it for science. And that's when he did the protein folding system. And then, you know, along the way, although interestingly, he didn't ever tell me this.

28:43I only know it from other sources. He was the one who said to Rishi Sunak, the British prime minister, you know, let's have a global safety discussion at Bletchley Park in 2023. And that happened. And so I think he had a reasonable track.

28:59Kara Swisher:Explain why Bletchley Park. That's where. Well, that was where during the Second World War, the German code was broken with. By Turing. With sort of the early kind of precursors of AI. It was very symbolic. Exactly. Exactly. So I think, you know, up to 2023 and that summit, he had a pretty good track record, Demis, in, you know, walking the walk as well as talking the talk on safety. I think, though, that once he got into the race full on to do chatbots in competition with, you know, OpenAI and then Anthropic, he's done less and less, at least visibly. And it strikes me, for example, that by his own logic, open source models are dangerous, and yet Google releases them.

29:43He wanted to deprive the military of AI, and yet Google now supplies them. And his sort of rationalization for this is simply that there's a race on. Chinese labs are part of the race. If he were to step back, if he were to quit, it would make zero difference to outcomes for humanity, because there'd be six other labs doing it anyway.

30:07Kara Swisher:Sure, but he wouldn't be part of it, like, in some ways, right? Yes. Look, and I think that gets to a super interesting debate about, you know, do we care about the gesture or do we care about the outcome? Right. I mean, I think his power had waned in that regard because people caught up, even if they're not as brilliant as he is. Now, he reminds me a little of Nikola Tesla, ultimately, you know, the greatest mind, and Edison just ran right over him. In 2024, he shared the Nobel Prize in chemistry for his work using air to predict structures of proteins, as you noted. Amazing thing. It will lead to the creation of new drugs and vaccines to fight diseases, an amazing thing, and an astonishing use of AI in service of humanity.

30:44Kara Swisher:I think one of the best examples, and not marketing, that the rest of them foist upon me on a daily basis. And it could be where AI makes people's lives better. Talk about this and what it can mean for medicine. I know a lot about this. I've done a lot of reporting on this. But it really was a moment. And it has taken off, whether it's mRNA, all kinds of stuff, drug discovery, you know, the quickening of drug discovery and ideas. Talk a little bit about this. Yeah, well, I know your CNN show gets into this, right? And so you probably could talk about it more than I can. But I mean, certainly in the creation of AlphaFold, this system, I mean, that in itself is an amazing story where Demis at Cambridge as an undergraduate had been told by a biologist friend, oh, there's this conjecture in structural biology by a Nobel laureate from the 1970s called Christian Amphinsen, that if you stretched out an amino acid chain and you looked at the DNA sequence on the chain, the sequence tells you how that chain will fold itself up like a self-executing origami model into an intricate, beautiful shape, which is a protein.

31:59And we have proteins in our bodies, and there are proteins in plants that get the basic building blocks of nature. And you can tell this intricate structure just by looking at the code. And so it became a sort of grand challenge in biology. Who can create the computational system that will just look at the code and then guess the shape? And from the 1990s, there were teams in different universities competing to do this. And DeepMind decides after winning Go, defeating the Go champion in 2016, well, that was a... That's the game. The game of Go. Yeah, exactly. That was just a game. But then having solved that problem, which involved this massive combinatorial complexity, because Go has all these different permutations.

32:40Kara Swisher:They do, but boring in comparison. Boring in comparison. So why don't we move on and leap all the way to predicting the combinatorial complexity of these possible shapes. One strand of amino acid could fold itself up in billions upon billions of different shapes. So in that sense, it was similar to Go. And having solved Go, they felt that their computer science skills had reached a point where maybe protein folding was crackable. And so they set out to do this in 2016. By 2018, they'd beaten all the universities. They had the best model. But it wasn't good enough that a pharmaceutical research team could simply use the prediction from computation to create a drug.

33:25So then the question was, should we push on and try to really make it that accurate? And the view of the leader of the team within DeepMind was, forget it, boss. We can't do this. This is impossible. Don't send us down a blind alley. We should just declare victory. We're the best team in the world. And then we should move on. And Demis was like, no, I want you to actually solve the problem, not just be the best. And so they had a bit of a fight about this. And Demis said to me, well, I was being unreasonable. but I wanted to be reasonable in my unreasonableness. So I listened into the team's discussions for a while.

33:58And when I heard they had lots of ideas that were just flowing naturally, which they hadn't yet tested in the lab or in the computational lab, I figured, well, there's more stuff they could try. And so then he pushed them. He switched out the team leader, put a new person in charge. And then in 2020, they succeeded and they did predict the shapes of proteins. And then the kind of the badass move was to open them so so just like, you know, for free, any scientist in the world can effectively do a Google search and get the shape of the protein that they want to work on.

34:32Kara Swisher:We'll be back in a minute.

34:46Kara Swisher:support for the show comes from net suite every business is looking to see how ai can work for them but while the possibilities are endless the risks are as well and you can't sit it out because your competitors are already making their move but with net suite by oracle you can put ai to work today. NetSuite is a top rated AI cloud ERP trusted by over 43 ,000 businesses. It's a unified suite that brings your financials, inventory, commerce, HR, and CRM into a single source of truth. That connected data is what makes your AI smarter. It knows how to intelligently automate routine tasks, how to deliver actionable insights, and can even help you cut costs and make fast AI powered decisions with confidence, giving you total flexibility.

35:29Kara Swisher:From software and IT services to healthcare, equipment manufacturing, financial services, and many other great American industries, NetSuite delivers a customized solution for your business. If your revenues are at least in the seven figures, get our free business guide, Demystifying AI at netsuite.com slash on. The guide is free to you at netsuite.com slash on. netsuite.com slash on.

36:01Kara Swisher:One of the things that's interesting, the parallel you make in the book and it's frequently made about the development AI is to J. Robert Oppenheimer, who we discussed earlier in the creation of the atomic bomb. It's one that Demis and others in the field welcome, but Oppenheimer is also, not just because I saw the Chris Nolan movie, I read a lot about Oppenheimer, he saw the bomb as an evil thing. it kicked off the nuclear arms race. He became a scientific exile during the Red Scare. How does he feel, Demas feel about this? Because right now, as you note correctly, there's a huge backlash to not just data centers in the US, but AI in general, the polling with the public is, you know, only Trump has worse polls at this point.

36:39Kara Swisher:But AI is pretty disliked. So talk a little bit about that. Yeah, I mean, I think he feels slightly that that's why you have to talk up the optimistic side of the AI story, I think he's a little frustrated with Dario, frankly, for going on TV and saying, within five years, 50 % of entry level jobs will be gone, you know. And to propose to kind of say that kind of thing, without proposing the policy remedy, you know, Demis sort of points slightly a finger at Dario on that. But I think, you know, I'm not sure I agree with him, frankly, I think, you know, Dario has a good case to say, hey, we've got to call it like we see it.

37:18And that will wake people up. And politically, maybe there'll be a stronger policy response if we do that. So I'm not sure Demis has any particular solution to this, you know, popular fury.

37:29Kara Swisher:Does he see himself as an Oppenheimer? I mean, Oppenheimer went through a pretty rough period before we liked him again, right? Or when he got finally fetid at the end of his life. But there was an exile happening. There was exile happening. You know, I think Demis embraces the kind of scientific glory, the leading of the Manhattan Project, the heroic story of, you know, going off into the wilderness and focusing on nothing but the scientific mission. I mean, there's something about Demis, which is almost, somebody called him a warrior monk. And off he goes, you know, in his mind's eye to the desert in isolation.

38:04And that just is a super appealing self-image for him.

38:08Kara Swisher:Yeah, nothing narcissistic about that, but go ahead. That's the Oppenheimer bit that he wants to identify with. He doesn't identify or doesn't talk about, you know, the idea that he would be ostracized and, you know, there would be kind of a political... Yeah, Dario's taken the flack for that at this point, right? Yeah, yeah. And I think there's an interesting debate about whether, you know, if you look at Dario's principal stance on the Pentagon using AI for weapons. It was good in terms of raising public awareness of the issue. It didn't change the outcome. So if you're a Demis and you look at that, you could say to yourself, well, it didn't change the outcome.

38:48So what was the point of that? I'm much smarter. I'm going to go do this behind closed doors. I will talk to politicians as I did with Rishi Sunak and suggest things. They can take the credit. What's the point of going public if you don't change how the world works?

39:01Kara Swisher:Absolutely, except he didn't change how the world works either, but that's another issue. So we're in the middle of an explosion of growth around AI, and you write, the estimates of achieving human-level AGI, again, much disputed by 2030, now appear slightly conservative. Talk about how Demis looks at this, because I get a different answer, different people, and I don't think they even, I think they're just guessing. I feel like at this point, they're just making it up. But talk a little bit about, you get very different numbers from different people. You do get different numbers. I mean, if you talk to people at Anthropic, they're really saying 2028.

39:36And what they say is that by 2028, there'll be recursive self-improvement, meaning that the models will be coding the next model. And once they've done coding the next model, the new model will recode the next one after that. And so you'll get this vertical acceleration in the capacity of the systems. At that point, you're done.

39:55Kara Swisher:So they go from dolphins to humans. At that point, you're done. The race is over. and they say 2028 in written material. I think actually privately, they even think it could be next year. So they are super near term in their prediction. Demis would love to kind of nudge his prediction beyond 2030, 2031, 32. And as I was saying, that's partly because his definition is, can it be Einstein? Which is obviously the most expansive definition. Then there are people in between who say, well, it's about when most economically valuable human tasks that you could do in front of a screen could be taken over by AI.

40:33And then people might put, you know, a 2030 kind of number on that. So I think that's the range.

40:38Kara Swisher:Right, right. And you just, of course, are seeing a lot of the layoffs from Meta and many others. So every episode, we get a question from an outside expert. We have two for you. Here's the first one. Hi there. I'm Adina Friedman, the chair and CEO of NASDAQ. Based on all of the time that you've spent with Demis, what do you see as the role of enterprise adoption and realizing the potential of AI. And if you do see it as being significant, what are the key obstacles that enterprises need to overcome in order to be able to achieve its potential within the enterprise? So I just referred to the layoffs, so go ahead.

41:10Yeah, so I think the revealing thing here is that the answer to the question of what does Demis have to say about enterprise adoption is zero, nothing. And that shows us the separation between the builders of the models and the real world users who are going to actually put it, you know, change the way corporations function, all that stuff.

41:31Kara Swisher:Although many of the builders are talking about it, Anthropic would be the one who seems to be making the most headway in that area. Yeah, I mean, I think Anthropic is very good at sort of going for enterprise tools. So coding first, cybersecurity second, and then having four deployed engineers. This is the kind of Palantir model who help corporations to adopt this stuff. So I agree with you. But certainly in a big company like Google, the people who are doing the Palantir stuff are in some other division of Google. And Demis is really just focused on building the product. And to the extent that you have this separation, and you're right, Cara, that it's not as clear in a smaller company like Anthropic.

42:11But I think the separation maybe does tell us something about what could go wrong here, right? That, you know, one set of people are just focused on building something and they're not really thinking about how it's going to be used.

42:25Kara Swisher:Yeah, and then another is actually actively thinking. And we're starting to see more companies, for example, justify layoffs by pointing to AI. Sometimes that's not the case, but they still are using it as an excuse. Cisco laid off 4 ,000 people earlier this month while announcing record revenue and earnings. Meta laid off 8 ,000. Goldman Sachs estimates AI a limit around 16 ,000 net jobs a month over the last year. Now, some of these, as I said, companies are using as a convenient scapegoat for too much hiring they did during COVID, which I think most people, especially the tech companies did, talk about mass joblessness.

42:57Kara Swisher:And is that something that Demis thinks about? And if so, what's standing in the way of that? And he said that he thinks the companies who are placing developers with AI, quote, show a lack of imagination and a lack of understanding. Is he being, again, naive about corporate incentives around tech? And some people are moving too fast forward, and it's not going to work, obviously. But talk a little bit about this. Yeah, I think there's sort of two extremes which are wrong. One extreme is to say there's nothing to see here. There's not going to be any big jump in unemployment. We've been here with previous kinds of technologies and the labor market always adjusts.

43:34And we had the internet and actually unemployment was very low throughout that period. And I think that's naive because AI is not just another technology. It's more powerful. It's more general. It's directly competing with human cognition. and it can be scaled if you have enough data centers, you know, up the wazoo so that you replace, you know, tons of people. So I think we can't just draw comfort and be complacent just because of the technology history. On the other hand, it's also way too simple to say, now that this system can do all this stuff, you know, humans are done. Because, you know, new jobs will emerge to some extent.

44:16I mean, you know, it's partly that there are tasks that humans are better at dealing with humans sometimes. So if you think about the sales team, enterprise sales team in a big company, where the job is basically to go schmooze the humans who are your customers and forge a bond with them, I think humans have an edge at that. So I think there are a lot of tasks where humans will remain superior. And to the extent that the economy grows because of AI, there'll be more of those tasks. So I think we should be aware of both extremes. I mean, one number I like. It's unclear. I like to cite this statistic, which is that no technology has ever driven economic growth per capita at the frontier more than 2.5 % a year.

45:02So people who tell you that AI is going to like double the size of the economy in 15 years of smoking something, it's just not that.

45:10Kara Swisher:Ah, abundance. You're not an abundance person, neither am I. The argument, I'm so tired of that. The argument around rushing to develop, obviously, A.G., I think, has largely centered on beating China. But in many interviews with these people, I call it the Xi or me argument, like we got to do it or China will. And I get it. But in a recent op-ed, and I really appreciated it, that you wrote in the New York Times, you called for the U.S. to negotiate a safety pact with China because we cannot beat them. And not just falling behind, but there's certain things as with nuclear energy, as with everything else.

45:42Kara Swisher:If we're falling behind, what incentive does China have in signing a safety pack? I think there's huge incentives for them to do so. And it's not just to keep them from being ahead of us. It's that there are commonalities here as with, as I said, nuclear energy or nuclear weapons or cloning. There's a lot of stuff that is global. And oddly enough, I ran into Tony Blinken last night and he said I had urged him when he first got his job to do a global pact around safety and with China especially. And he's like, well, I didn't get that done. I'm like, no, you didn't. And then I walked away. I was rather intent on getting them to stop, you know, to pay attention to this.

46:22Kara Swisher:But talk a little bit about this, this sort of Xi or me and what happens with China, because you are correct. They are doing astonishing work here. and we had been ahead, obviously, the U.S. or democracies and now up at all democracies because Demis is in London. But talk a little bit about this race and also the need for a safety pack, which I think is critical. Yeah, yeah. So, you know, China does everything fast. So my book came out in China before it came out in the U.S. And I went to China in March for the launch and I spent eight days basically going around talking to AI leaders, both in academic labs and in industrial tech companies.

47:01And what struck me was how often they brought out the topic of safety. And this was interesting because that is not what you hear from people who put the chip export controls in place in 2022 in the Biden time. So those people, several of whom are my friends, and I've talked to them a lot about this, they said to themselves, credit to them, the scaling laws are real. AI is going to be super powerful. And they said this before ChatGPT came out. So we're going to do something. We're going to prevent the bad guys from getting this AI. But their definition of bad guy was China. And in my view, there's a more important set of bad guys who are criminals, who want to do cyber attacks, terrorists.

47:47Individuals. Rogue states, you know, random bad people. Rogues. Just rogues. Individual rogues. Exactly. And they missed out that whole category. And by making China the enemy and the rival and saying, you cannot get AI and we're going to deny you the chips and say you won't have NVIDIA, no NVIDIA, no dice, you'll be left behind. You know, the mistake was to make China into the enemy and lose maybe a chance to talk to them about what about non-proliferation of this stuff. Maybe we should just say, you're a technology superpower, we are a technology superpower. The way we're going to avoid some catastrophic war between us is the same way we did it in the Cold War with the Soviets, which was parity, mutually assured destruction.

48:31Parity is actually good for balance and stability. Whereas when you're talking about random rogues using this stuff for bad ends, the Cold War lesson is that's where you need the IAEA to keep track of the material and the non-proliferation treaty. And my friends, you know, say, who were in the Biden administration say, well, you know, yeah, but you can't talk to the Chinese because they don't care about safety. I went there and they did talk to me about safety. They say, you can't trust the Chinese. And I go, you think it was easy to trust Khrushchev? You know, I mean, this was the Soviet leader who banged his shoe on the table at the UN and put missiles in Cuba.

49:09Not an easy guy to talk to. But in that period is when we created the non-proliferation regime. for.

49:15Kara Swisher:That's right. And they have their interests of their economy being destroyed by a rogue something is just as high as ours. Totally. It's just really kind of, kind of, I was always like, not so sure I want either of you. You know, I'd like a global safety group that is, includes companies, includes legislators, includes China, includes the US. So that we have, it's a very similar to nuclear. Exactly. We, we trust cruise drop. That's a great way to put it. I thought that piece was terrific. Thank you. Now, we have a second expert question for you. This is from Kent Walker, president of global affairs at Google Unalphabet and the bad cop, apparently.

49:54Kara Swisher:Let's listen. Hey, Kara. Hey, Sebastian. In reading the book, I was struck by how well, Sebastian, you captured Demis' vision of using AI as a way of solving intelligence to then solve everything else. It was a scientific tool to help us address some of our biggest problems. The challenge is that polling in the U.S. suggests that America is one of the least optimistic countries about AI, much less optimistic than China is. Maybe we've had too many debates between AI accelerationists and AI doomers, but how would you suggest we go at the challenge of creating a grounded optimism, recognizing the challenges, but also encouraging people to put these tools to work to benefit themselves in their everyday lives and to benefit our society?

50:36Thanks.

50:37Kara Swisher:That is a great question. One, Google, you should stop being so aggressive, but go ahead. Yeah. So, I mean, we referred to this a bit before, but I think that medical breakthroughs that delivered AI-generated or AI-assisted drug discoveries that really saved people's lives, that would be a great thing to change public opinion on whether AI is good or not. I also think that we ought to put in place preemptively before we totally need it, things like wage insurance, retraining schemes, active labor market schemes to help people get other jobs. Do that early. Don't wait around until it's obviously, you know, essential, because at that point, public opinion will be so mad at you that, you know, it won't make any difference.

51:22Kara Swisher:There'll be riots. There'll be work riots, just as they were before. Yeah. I think a useful statistic here is that in the 12 years from 1999 to 2011, the total number of job losses as a result of China renting the WTO was 2 million in the US. 2 million is nothing. 2 million is like the amount of labor market churn you get in an average month in the United States. And yet, you know, the backlash against China, the perception of the China shark, the fury about globalization politically was absolutely massive. So imagine if you got an AI shark that was way, way bigger than China. The reaction will be, you know, as you say, it'll be people in the streets.

52:08You know, I mean, with the Industrial Revolution, what happened was you got revolution politically, you got Marxism, you got a whole lot of turmoil. And that was a revolution which took place probably one-tenth the speed of the AI revolution and may arguably have been smaller.

52:26Kara Swisher:Absolutely. I think they don't even understand the anger, the rage that it's building. And it's manifesting itself. I think a lot of the AI anger is a manifestation of affordability and nervousness about the future. And now it has become that. It has become that. I don't know what you think, but I think there's some merit to this idea of Trump accounts, where you would put shares in AI companies or other stakes and distribute that democratically to all young Americans. Sounds good. No, he doesn't seem interested in doing... He does them, but he does other things more. The stuff that he does that's populist, I'm not against the Trump accounts.

53:05Kara Swisher:I don't like him calling them Trump accounts or Trump Rx or whatever. He has to put his fucking name on everything. But, you know, but that idea is correct. You know, UBS, some version of this or more, like where are the jobs and do some really serious government studies and where we work with companies. And, you know, this recent AI advisory board has nobody on it except for business people. There's no way they're going to come to any conclusion but their own. and they could use critics. They could use someone like you, someone like me, someone like an academic, someone like, you know, they're just refused to have wide ranging points of view here.

53:41Kara Swisher:That is probably going to be the biggest problem, I think. I think it's also relevant to Ken's question that, you know, the record in the last couple of years, well, during the Trump period, is that all the energy on legislation has come from the states. And the AI companies, the tech companies, have lobbied actively to stop that stuff from passing. And they have a decent argument in that a patchwork of state-level stuff would be way less effective than a federal thing. But then that's why you need federal intervention, right? Which they are also trying to stop, which is why they're sitting in the front row of the Trump inauguration.

54:17Kara Swisher:Yeah, I don't quite, I think that's a bit tough. I know you like to be tough, but I think that in the Ben Buchanan period, when Ben Buchanan was the czar in the White House, what he said on the record publicly is that whenever he talked to the labs, they were willing to support what he was trying to do in terms of setting it up an AI safety institute, in terms of then requiring labs to share their models with the safety institute before they were released. You know, clearly that regulation should have gone further, but he says that actually the labs were encouraging him to go down that path of regulating.

54:56And I think I don't have any reason to disbelieve that. So, you know, I think...

55:01Kara Swisher:I disbelieve that. You disbelieve? But look, with the Trump administration just as starting to announce now, the same thing, the exact that they rejected, correct? Now, they do that with a lot of stuff, but their most recent noises have been exactly what that Biden executive order said back then. Yeah, I agree. So Trump... Which is crazy. The Trump people have done... They're like, this is our new idea. And I'm like, well, that was two years ago, three years ago. Bletchley Park. 100%, completely agree with that. But that's a description of the Trump administration's U-turn. I think the position of the AI labs, which is what we were talking about earlier, I think they actually have been open to sensible federal action.

55:38They just hated the state action.

55:41Kara Swisher:Well, they're getting the state action because initially they lobbied against any kind of tech regulation, right? And so they find them, they should be supporting it. They shouldn't have kneecapped Amy Klobuchar, for example, or they've been kneecapping politicians for a long time. And I think people don't trust them. I think that's where it is. And then when you have any manner of them swanning around looking like Daddy Warbucks everywhere across the world, it's not great. It's not, you know, Elon hasn't helped. Jeff Bezos certainly hasn't. And neither is Mark. I mean, I think the imagery, you don't have Demis is there, or, you know, Dario is a hero and like, he's fine, but like, that's it.

56:17Kara Swisher:Like, that's the problem. There's no heroes. There's a lot of people who look like they're in it for the money. So what is your explanation? I'd love to get your view on this. So why is it that Demis is not particularly famous in the US? I mean, I've been... He doesn't want to be. He doesn't want to be. He could be. He could be all over the place. He could be Dario Modi-ing everything because he deserves that spot. But, you know, I think he's a true wonk, probably. And so he's, science is everything. But, you know, history is littered with people whose sciences weren't as savvy. And Nikola Tesla wasn't as savvy as Edison.

56:53Kara Swisher:Edison was a PR person. He did all many of nefarious tricks against Tesla. Like, you know, I don't think he's that much of a victim the way Tesla certainly was, if you read a lot about what happened there. But I think he likes to be pure. He likes to have that image of himself as pure. That's my, I don't know him very well, But that's what I can see. But he certainly deserves it. He certainly deserves the attention. You know, Jeffrey Hinton likes to take up space, that's for sure. But you could see him out there more. But no, here he is with your book. So obviously he's talking to you. So last question.

57:23Kara Swisher:Today's AI race is exactly the scenario he had hoped to avoid. But in an excerpt in The Atlantic, you wrote that Demis has, quote, come to see salvation paradoxically in his own career advancement and securing personal influence. We have concentrated so much power in the hands of just a few people. After years of talking to many of them for the book, what do you make of his conclusion and what does it mean for the rest of us who are worried about concentrated power? I know I am of anyone, even if they were the angels. I don't feel like they should have this many small amount of individuals to have this much power over something so important.

57:58Kara Swisher:Of course, it's happened in history, but it's never ended up well. And this is quantumly more powerful than any like railroads, trains, the telegraph, stuff like that. Are you worried and should the rest of us be worried about this concentration of power, even if it's in the hands of people that are better than others, but a lot of them aren't? Yeah, yeah, yeah. Yeah, look, I think in a way the story of Demis Sassabis is of a decent person, a good person, who wants to be good but can't actually have a good impact because he's inside this race dynamic, inside a big corporation. And whatever he may personally think about safety, there are these larger forces that are driving him on.

58:41and if he were to quit tomorrow and take a professorship in theoretical physics of printon it wouldn't change the race it wouldn't make the world better it wouldn't do anything very constructive i'm not sure i want him to do that because you know it would just be a gesture and i prefer to focus on the outcomes so what this points to is that we do need government intervention you know the only way to counter the power of this handful of individuals who are kind of collectively very powerful, but individually impotent in terms of the big social questions. You know, you need government to take a serious intervention.

59:19And, you know, why not? We have drugs in the United States, which are approved by a regulator. And if they're dangerous, they're not allowed to be released, right? Why would an AI model be released without being tested by the government first. Why would we allow, you know, open weight models that can be ripped off by anybody, abused by anybody, used for a cyber attack, and you wouldn't be able to kill it with a kill switch? It's crazy, right? So I think what we need is far more active government action, precisely because AI needs to be governed. Right.

59:53Kara Swisher:So where does that leave the rest of us if AI goes the way that social media did and the government fails to act? I mean, obviously, different governments are doing different things, but there's no coordinated global effort at all. And in this country, forget it so far. And of course, we did nothing to slow down the relentless and often poisonous and toxic pace of social media. So what has to happen? What do you see happening? What does Demis see happening? I think he's got to the point where he's trying a little bit to, you know, do politicking behind the scenes and suggest that, you know, more action is needed and it has to be collective and the government needs to lead that.

1:00:33But whether he's putting real energy into it, I can't tell you for sure. And so I think it's just, you know, that history we talked about before of Project Mario and trying to get safety oversight. It's almost like he's burnt out on trying to solve the governance problem. And so he's just heads down on the technical problems, which is a regrettable place to end up. But, you know, all I can say is that, you know, I mean, that's the most optimistic thing that's happened in the last few months, was when Anthropic released Mythos, and the Trump administration freaked out. And they called in the heads of the banks and said, you guys are going to have hacks that empty your bank accounts unless you can harden your cyber defenses, because this stuff is for real.

1:01:17And they basically did a 180. And now you know what, that Anthropic had this list of 40 responsible tech companies that were going to get the model first before anybody else got it, and then they were going to roll it out to others. Those subsequent waves have been frozen because the whole decision-making around the release of the models has been requisitioned by the Trump administration, and they're sitting on it, and they're saying nobody's allowed to have this. And so they've gone from being laissez-faire to maximally interventionist. It shows you how freaked out they are. And maybe that's a hopeful sign for action on regulation in the future.

1:01:53Kara Swisher:Unless they're trying to do it to make more money. Who knows? Yeah, possible. How can we take all those bank accounts and get them for ourselves? Cara, I was trying to do this sort of optimistic, you know, ending riff. I understand, Sebastian. You just burst my balloon. Oh, I don't know. You know, intelligence has its limitations, but stupidity and greed are infinite. So that's my feeling. I'm sorry. But I think the moral of the story is that tech just needs to be scary enough to force the government to act. That's really not the way. That's not the government I want, but fine. I'll take it in some ways.

1:02:26Kara Swisher:Anyway, this is a marvelous book. I really appreciate it. And thank you so much for your time. Of course. Thank you. I really enjoyed it.

1:02:38Kara Swisher:Today's show was produced by Christian Castro-Wassell, Michelle Alloy, Catherine Millsop, Megan Burney, and Caitlin Lynch. Nishat Garwa is Fox Media's executive producer of podcasts. Special thanks to Bradley Sylvester, Madeline LaPlante-Duby, and Julia Sharp-Levine. And thanks again to American University's Kogod School of Business for hosting this event. Our engineers are Fernando Arruda and Rick Kwan, and our theme music is by Trackademics. If you're already following the show, you win a Nobel Prize. If not, you're just in it for the money. Go wherever you listen to podcasts, search for On with Kara Swisher, and hit follow.

1:03:13Kara Swisher:Thanks for listening to On with Kara Swisher from Podium Media, New York Magazine, the Vox Media Podcast Network, and us. We'll be back on Monday with more.