#168 Ian Bremmer on Regulating AI for a Safer Future

0:00Digital world order, it's increasingly run by tech companies that are sovereign in that space. And that's particularly true in AI, where the foundational models and what they do, what their inputs are, the data they train on, all of that stuff is determined. The rule set of how you as a human being, as a citizen, as a consumer will engage with them. Those things are determined wholly by the people that are running the tech company. I do believe that there is a very, very tall mountain to climb to have effective governance of that space. And unlike on climate change, where it started late, but at least it's comparatively slow moving.

0:43I mean, I know it doesn't feel that way because of all the things that are happening in the world on climate. But, you know, at the end of the day, like we still have years in principle to stay to 1.5 degrees centigrade. And if we don't, it'll be two. And we're kind of, we have the ability over decades to course correct. Hi, my name's Craig Smith, and this is Eye on AI. Today, I have Ian Bremmer with me, the renowned political scientist, to talk about AI regulation, the potential for a global regulatory framework, and the role of nation states in AI development. Ian also talked about who should be held accountable for AI harm, and the U.S.-China competition and collaboration over the technology.

1:26I hope you find the conversation as engrossing as I did. AI might be the most important new computer technology ever. It's storming every industry and literally billions of dollars are being invested. So buckle up. The problem is that AI needs a lot of speed and processing power. So how do you compete without costs spiraling out of control? It's time to upgrade to the next generation of the cloud Oracle Cloud Infrastructure or OCI OCI is a single platform for your infrastructure, database, application, development and AI needs OCI has four to eight times the bandwidth of other clouds Offers one consistent price instead of variable regional pricing and of course nobody does data better than Oracle.

2:21So now you can train your AI models at twice the speed and less than half the cost of other clouds. If you want to do more and spend less, like Uber, 8x8, and Databricks Mosaic, take a free test drive of OCI at oracle.com slash ionai. Okay, so Ian, let's start since I'm recording. Introduce yourself. I mean, obviously, I know who you are, and a lot of people listening will know who you are, but it's always good to have you introduce yourself. Sure. I'm Ian Bremmer, a political scientist, and I'm president of Eurasia Group. I'm interested personally about your journey, about how you got to not only running Eurasia Group, but making Eurasia Group as prominent as it's become.

3:20Can you just talk a little bit about how you developed the organization and your profile? Well, I mean, it's always been a misnomer in the sense that when I started it, it was focused just on Eurasia, but it wasn't a group. It was just me. It was Eurasia guy, which doesn't sound very credible as an organization. And as I started getting clients that wanted access to the research that I was doing as a political scientist, as a former academic, and I started hiring people. And I've done it now for 26 years. I started with nothing, and now it's a big company. Globally, we've got about 250 folks. It's just kind of a really cool political science story.

4:10For someone who really cares about trying to understand the world, that really matters to me because I think the world is becoming a much more dangerous place. You've spent a lot of time recently talking about AI regulation in particular. And I read your piece in Foreign Affairs with Mustafa Suleiman. And you guys talked about, in that piece, about developing a regulatory framework or kind of a stocked framework with various levels. piece. Does that represent your latest thinking or has your thinking evolved? In that piece, you talked about kind of a three-tier regulatory framework. I'd be interested in hearing about that.

5:00And I'd also be interested in hearing what you are thinking about AI development in general. with? I mean, look, it's evolved in the sense that the field is changing very quickly. There has been some regulatory efforts that have come about since Mustafa and I wrote that piece. And I think they've all been informed by it, actually, which is first time I can really say that in my career, which is sort of interesting, in part because there's so little out there. And also because Because I helped get the Secretary General of the UN interested in AI five years ago. And he's now launched a high-level panel on AI that I'm on the executive committee of and serve as a rapporteur.

5:51So I'm sort of writing up, you know, hopefully what will be the global direction of governance on AI. So, again, kind of the first time in my life that I'm sort of a policy practitioner, even though I don't in any way think of myself that way. Right. But but again, it's this is a very, very fast moving space. So that's my that's my kind of immodest rejoinder to how I want to how I would answer the question. It's changing. It's changing a lot because of my experience in the field as much as the field is changing. But yeah, look, I mean, I guess I want to say a couple of things to start. The first is a level of humility about what governance can do, given how fast the technology is moving.

6:38the utter lack of interest in any of the technology companies, the leaders of those companies in slowing down at all. They want to speed up because this is a very well-funded, extremely smart group of people that have competitors breathing down their neck. And so, I mean, they're not the bad people, but I mean, their incentives are overwhelmingly to get this out and faster, faster, faster, right? So you put those things together and the fact, you know, what I call a technopolar world, that when we talk about the digital world order, it's increasingly run by tech companies that are sovereign in that space.

7:25And that's particularly true in AI, where the foundational models and, you know, what they do, what their inputs are, the data they train on, all of that stuff is determined. the rule set of how you as a human being, as a citizen, as a consumer will engage with them, those things are determined wholly, wholly by the people that are running the tech companies. So I do believe that there is a very, very tall mountain to climb to have effective governance of that space. And I would argue the technology is moving a lot faster than the governance is, even though in the last 12 months, the governance has been a priority.

8:10It's been urgent. It's been, you know, on the top three, top five issues of all the major policy leaders around the world that you speak with. And I'm very heartened by that. And they are legitimately trying to figure out what to do. They are not coming to the table with preconceived notions of these are the equities that I must defend. So that's all to the good. But, but, you know, it's, it is, it is starting late. And unlike on climate change, where it started late, but at least it's comparatively slow moving. I mean, I know it doesn't feel that way, because of all the things that are happening in the world on climate.

8:51But you know, at the end of the day, like we still have years in principle, to stay to 1.5 degrees centigrade. And if we don't, it'll be two. And we're kind of, we have the ability over decades to course correct and have some impact on the world. You don't have generations on AI. You've got years. You've got years, a single years, right? So that's the backdrop. That's the context for talking about the geopolitics of AI, the governance of AI. Yeah. I talk to a guy periodically. Do you know Connor Leahy? He's a smart guy. He's young and a bit of a rebel. But he had this group, Eleuther AI. He now has a startup focused on alignment.

9:44But Luther AI developed the first open source large language model, GPT-J, and has gone on to develop larger models. And he's now very focused on regulation. And he says some very interesting things that I just wanted to hear what you said. because what you had in foreign affairs is kind of high-level framework. But in terms of specific regulations, he says, for example, the world got together and decided that human cloning is a bad thing or that we're not ready to go there. And it was banned globally through various structures. Is that possible with AI? Could you set a limit on how many flops per training session or how large?

11:00I mean, compute is an easy thing to track. uh i mean is that one of the things that one of these global bodies or or at least nation states could could do just outright ban something no well first so first of all um it's the right question to ask it's a very interesting question um and uh as a political scientist um i would start by answering it as if you were asking that question of nuclear weapons after the americans and Soviets already had a bunch and say, well, I mean, in principle, couldn't we just ban? Because it's obvious that we don't want them. It's obvious they're incredibly dangerous.

11:43We don't think they're usable. We're really worried about proliferation. We're really worried about what could happen in the world if we keep building them. As it turned out, we believe that at least once we were one in three at destroying humanity as a consequence of them. So it's a really bad idea to have them around, right? And yet, and we've got countries down with like Pakistan and North Korea that we clearly don't want to have nuclear weapons. It's utterly unacceptable. We can't do anything about it, right? And so the genie is out of the bottle, and we cannot wish it away. There is way too much money.

12:20There are way too many powerful companies with influence. It matters way too much for growth. There is already way too much proliferation. And the geopolitics are already way too competitive. If we don't do it, somebody else will strategically, never mind technologically, never mind from a business perspective. So for all of those reasons, absolutely not. You saw there was this one letter that was signed by a number of, you know, sort of AI, let's say, concerned enthusiasts and scientists. and it got, I think, several thousand people signing, a lot of whom actually mattered. It was only to limit one very small part of AI development.

13:06And it went absolutely nowhere immediately as anyone touching the field could have told you it was going to. So we've now had this high-level UN panel together for several months. We've put out our interim draft report. And that question that you just raised, and remember, this is 35 plus people, public sector, private sector, ministers, public policy types. It's a really good cross-section globally of people with expertise and power over and around AI all over the world. That question that you just raised has not taken up a moment of conversation over the months of our meetings. Not in any working group has not been discussed because it's not relevant.

13:55It's not credible. It's not anything that we think we could put muscle into. Now, keep in mind, when you're working at the global level, you understand, number one, the United Nations has no power, right? Number two, the United Nations has no money. What it has is legitimacy at the global level, a voice. It actually brings everyone together. So you can try to understand what the lowest common denominator is. You can steer it. You can create analytic ground truth that everyone can come together around, which is a really important thing to do in my view. And occasionally, especially if you have a crisis, you can raise the lowest common denominator, at least among a group of players, a subset of players, and maybe you can then broaden it out.

14:43Those are all worthy goals. The question that you just asked me, can you just stop it from happening? No, that is several factors of extrapolation beyond what anyone thinks is remotely possible. Even in, say, because this is not the level of compute we're talking, I'm talking about massive compute, something that no one has reached yet, but that is probably in the current trajectory necessary for AGI. Just say, have a treaty between the U.S. and China and Russia and France, maybe, that no training session or no AI system can go beyond a certain number of floating point operations per second, you know, and set it high beyond what's being done today.

15:46just as a limit? You don't think that would work? Oh, I mean, that's an interesting thought experiment for the future. So a couple of caveats before I try to answer that question. The first is, we don't know what the state and diffusion of compute will be like when that conversation becomes more relevant. In other words, how many actors will have access to it? is it going to continue to become logarithmically more expensive as it is presently projected to be using the same kinds of energy inputs, for example, that only a small number of states would have? Because if that's the case, then it's no longer a massively proliferating technology.

16:32It's something that only the U.S. government together with tech companies and the Chinese government together with tech companies can probably do, right? So you won't need a treaty because it won't be relevant. And then there's just a question of what mutually assured destruction looks like and what the Americans and Chinese decide to share. So it becomes a geopolitical question. That's one point. Second point is to what extent the advances in AI technology are likely to continue to be on the back of more and more explosive data sets, which have enormous amounts of garbage in them? Or is it on the basis of more finely trained data that is smaller, that you have full confidence in?

17:23And that's a very different kind of advance that will be massively more available at lower scale to governments, to companies, to rogue actors, you name it. Now, I also would say that the AGI question is, I mean, I am worried about so many other things and excited about so many other things in AI before we remotely get to whether or not an AGI is either technically or even philosophically possible. I'm much more concerned about people destroying the planet using AI than I am about, or destroying humanity, I should say, than I am about robots taking over, about artificial intelligence becoming sentient.

18:12I just think that what we have right now is an incredibly powerful, incredibly economically disruptive for good and for bad, and politically disruptive for good and for bad new technology that very soon everyone is going to have in their hands and be able to do stuff with it. And we are not ready for that. We're not ready for the economic upside, but I'm sure we'll figure a way. But we're also not ready for the economic, political, social, and security downside. And everyone with money in the field is going to take care of the upside problems. I'm not worried about not reaching the upside, that governance is going to stop the upside because the power is all there, the trajectory is there.

18:58I'm worried about the stuff that people aren't spending the money and time on. Because, you know, again, I'm a political scientist. I'm kind of that's what I'm paid. I'm paid to do. That's my concern. Yeah. Another question that I'm sorry to dive. Can I ask you, why did you start with AGI? Is it just because that's the really big, like high level, long term existential thing? Or are you more concerned about than I am. AGI I'm not concerned about. And I think as I mentioned AGI, it's as something that we'd want to control if it ever happens. We'd want to have controls in place to prevent it if we don't want it to happen.

19:42But I thought the idea of limiting the amount of compute that can be marshaled for any particular AI system as a practical way of guarding against not necessarily AGI, but just super powerful systems. I thought that was an interesting idea. The other thing, and I'm sorry to dive right into the specifics of regulation. But to me, that's more interesting than, you know, you have this body and it's going to be responsible for this and that body is going to be responsible for that. It's the specifics that interest me. Another big question that's been debated in the EU AI Act is the responsibility of model developers or model purveyors.

20:40Where do you stand on that? Because again, this is a frustration for me in the gun debate that gun manufacturers are free of liability and that in social media, the social media companies are free of liability. And there's been a lot recently about the plastics industry. I don't know where that debate started, but people saying, hey, you know, the plastics industry knew that this was going to be a problem and they just shifted it to the consumer with this whole, you know, you should be recycling. So is there should should these potential harms fall on the model developers? And that would be a way of of creating some some caution on their part.

21:41They certainly have to have some level of accountability. How much of it is solely on them? How much of it is shared in the industry? How much of it is with government actors? That's that's where the decision has to be made. But your question is obvious because, I mean, these companies, it turns out Americans are incredible capitalists when we talk about profit. We understand where the profit's being made. We know how to build business models that will create extraordinary world changing enterprises. And we make sure the shareholders get it. And that's driven a lot of growth. But when it comes to losses, we are the world's best socialists.

22:16We really want to make sure that we socialize those losses. Not on me. It's on the public. It's on the general public and preferably it's on the general public in the future so that we don't have to pay for it. Our kids do. Whether you talk about climate change and carbon and methane emissions, whether you talk about plastics, deforestation, biodiversity, social media, you name it. Right. And I, of course, is is the same thing that we have completely failed on social media, completely failed. And I think everyone knows this. It is addictive. It is horrible for our children. It is damaging them in ways that we do not know or understand.

22:55We would absolutely, if this was a GMO, if it was a vaccine, we would have testing regulations before we would allow them to be injected into the bloodstream, the consciousness of our kids. And yet here we're testing them real time on people, on societies, on democracies. And you see the implications. It is undermining American democracy. We are exporting tools that undermine democracy around the world. That is not the United States I grew up in. And, you know, we now have AI. And it's a much more powerful version of the social media challenge. You know, so, I mean, yes, I believe that foundational models need to be tested and that there needs to be transparency around that.

23:43There needs to be red teaming that should be done by folks that are either outside of the organization or if they're inside. There is some level of outside participation so that you know exactly what's being done. But not just in terms of breaking the model to have it do things that the company doesn't want it to do, but also in terms of testing the actual business model on society. Like we want to understand what it means when AI bots are released and people use them before they are released broadly. We kind of need to know what's the impact on society. And if the companies really don't want that, then there need to be some level of legal recourse to help ensure that they are accountable for damages that eventually might occur.

24:35My point here is not that there's a fundamental problem with like an advertising driven model for social media, or the monetization models that the AI companies will soon roll out everywhere around the world. Rather, I'm saying that all of the money that is made from them also means that those profits will need to pay for the social costs that come from those business models. And there are two ways to pay the social costs. You can pay them ex ante by by testing and trying to help ensure that those costs are limited or ex post once the costs have already occurred and you need to clean it up. And that would be true for a company that's dumping its waste and ends up causing cancer and a bunch of a bunch of people in a community.

25:28You know, as I saw growing next to where I grew up, like, you know, Woburn, Massachusetts and stuff like that. Or it's going to be true for AI. But those things are all these are all challenges where the business model does not yet account for public goods. that matter, that are part of GDP. That's a fundamental issue. By shifting responsibility to the model developers, to me, that would immediately clean up the problem because model developers would be terrified of putting something into the public space that could be abused or could cause real harm. Do you think that that's a serious consideration in the United States or in the EU?

26:23Well, even in the EU, where it would be most likely, you've seen the French in particular, but also the Germans and others pushing back very strongly against that, because they do not want their own existing unicorn companies and nascent unicorn companies to be cut off by a much less permissive regulatory framework than what their competitors enjoy with, you know, a big market and with a head start in the United States. In the United States, that's generally not the cultural orientation of how the sausage politically is made. I mean, the first thing that was done was bringing the corporates and the government together to create voluntary commitments that the big AI companies were going to follow.

27:15And then there was an executive order, which is not the same as legislation, but still very much a, what are the common sense things that we think we can regulate that the companies agree to. And so I think that the real answer to your question is it's going to be in between a full responsibility of the companies and some level of oversight, however inadequate in cooperation from the governments. But it will be better than what we've seen in social media, which has been hands off and nothing. It's been a disaster. I do think that the failings of regulating social media in any way have informed some of the concerns that everyone has around artificial intelligence, especially when you talk about the disinformation side, for example.

28:09Yeah. It would be more akin to what exists in the auto industry or the aircraft industry, that you have to ensure that your product is safe before releasing it. Is that right? I mean, to me, that seems like a simple. I don't think we're going to get there. I think that that is I'd love that to be the standard. It is not the standard. I'd love I personally I think that governments are a better position to make policy than corporations. They have you know, they have more people focused on it. They have the resources. They have the experience. There are lots of problems with governments. But when it comes to making public policy in the public interest, I would rather them do it.

28:53In this case, I think you need a hybrid model. You need the tech companies and the governments doing it together because the governments don't understand it. It's moving too fast and the companies have the resources and they're making all the decisions. So I think we're going to have to accept that the outcome is going to be made in larger part by self-interested parties. And you'll try to restrain that impulse by not having them do it by themselves. But that's where I think we are. Yeah. Is there an industrial model that you can point to where that kind of hybrid regulation has worked? I would argue climate change is probably the closest where, you know, the reason that you now have a COP process every year with 90 ,000 people attending this year, and it was private sector and public sector and academics and activists.

29:50And they are they have driven an awful lot of investment towards renewable energy. And the question is just how big, how fast. But we all know the end point. And that's because the hybrid model at base created ground truth, created an environment where one hundred ninety three countries around the world all accept that climate change is happening. It's driven by human beings. It's not cyclical with the earth. and that we are now at 1.2 degrees of warming. We are now at 440 plus parts per million of carbon in the atmosphere. We know where it came from. We know the implications. I think that if we could have a process where we can get to that on a, where the public and private sector together can create ground truth around the opportunities and the disruptions coming from AI, the state of the AI universe, which is very different from climate because you'll have to update it all the time, like continually.

30:52It can't be a once a year summit or report. But that I think would actually help quite a bit because any regulation structure you put in place, any governance structure is going to be wholly inadequate in a year, in three years. It's going to need to be inherently flexible to change as the technology changes in ways that we can't actually conceive of now. And yet most institutions and frameworks, when we set them up, they're actually quite rigid because we want to protect our equity since we're the ones setting it up. We don't want anyone else coming in and changing it so that we're not powerful anymore.

31:26Well, you actually have to have the flexibility. The creative destruction needs to be in the fabric of the architecture, which is a lot harder. Yeah, that's interesting. That's one thing that struck me in your foreign affairs piece, this idea of inviting the tech companies to the table. And certainly I agree with that in an advisory role or something. But in climate change, this is a good example. You know, every corporation, fossil fuel in the fossil fuel chain, you know, has responded at the margins. They've responded much more vociferously in their PR. I mean, there is just tremendous greenwashing.

32:26And I just, I think that at some point, that's the role of government. It has to draw a line and whether it stifles innovation or not, until the governments have the expertise to make informed decisions. I think relying on the industry to develop. I mean, that's, again, what happened with social media. It's, you know.