Inside CVC | Stuart Russell on Why Europe Lost the AI Race and the Poisonous Logic of "But China"

Stuart Russell co-wrote Artificial Intelligence: A Modern Approach, the textbook that trained a generation of AI researchers, and founded UC Berkeley's Center for Human-Compatible AI. He joins Philipp and Steve on Inside CVC straight from the floor of SuperReturn in Berlin, where AI was the only topic anyone wanted to talk about.

Stuart takes apart the "but China" argument used in Washington to block AI safety regulation, and explains why China's rules are actually stricter than Europe's. He walks through why a Paris AI startup can raise three hundred thousand euros while the same company gets one hundred million in California, and why that funding gap, not regulation, is what pushed a generation of British and European AI pioneers to Stanford and Silicon Valley. He also breaks down why no country, including the US, actually has a sovereign AI stack, and shares a randomized trial showing chatbot users gave up on problems twice as often once the tool was taken away.

If you sit on a board or in the C-suite weighing where AI belongs in your business, this conversation will sharpen the questions you ask before you approve the next deployment.

Transcript

Introduction

Welcome to Inside CVC, the podcast that brings together leaders in innovation and capital investment to explore the trends shaping the business of corporate venture capital. I'm your host, Steve Schmith and together with Philipp Willigmann we're speaking with corporate investors, entrepreneurs and ecosystem builders. Driving the Future of Innovation Inside CVC is brought to you by you Path Advisors, helping corporations and startups unlock sustainable growth through strategic partnerships. To learn more, visit UiPath dot com. That's the letter U dash path dot com. And to catch up on all of our episodes, search Inside CVC on your favorite podcast platform or visit upath.com forward slash podcast. In this episode, we sit down with Stuart Russell, UC Berkeley professor and co-author of Artificial Intelligence A Modern Approach, the textbook that taught a generation of AI researchers. In this conversation, we talk about why Europe's own AI pioneers built their careers in Silicon Valley instead of at home, Why China's AI regulations are already stricter than. Most people in Washington realize why no country, including the U.S., actually controls its own AI stack. And the real questions board should be asking before their next AI deployment. We close on the bigger question, Stuart says. Nobody has really answered yet. What gives people a purpose in a world where AI does most of the work? Here's our conversation with Stuart Russell.

Interview

Stuart welcome to inside CVC. Thanks for joining us. How are you doing today? I'm doing well. Thanks for having me. Yeah, absolutely. AI, such an interesting topic lately. You're hearing about the topic in the lens of competition between countries, between regions, everything from national security to innovation to supply chain. And largely those conversations recently have been pitted the United States and China. I'm curious, is that framing helping us or is it pushing governments and corporations in some ways in the wrong direction? Well, I think competition can be healthy, and to a large extent it has been a driving factor in China's policy. I mean, it's very clear going back to twenty seventeen. That China identified AI as a national priority, He invested a huge amount of money, or at least what seemed at the time to be a huge amount of money, one hundred and fifty billion dollars they said they were going to invest in in AI, both, you know, creating facilities, startups, university research and so on. And, um, you know, and I think it's been quite successful. China now has a thriving, uh, research activity in AI, um, as well as a good degree of penetration of AI into its general digital ecosystems, uh, you know, the WeChat world and Alibaba world and so on, um, as well as lots of active startups and some companies that are doing a pretty good job of creating AI models, especially open source models, which has been really become a Chinese specialty. Really. Um, I think more generally though, and in the AI safety community, which I belong to, um we call this the but China argument where in any reasonable suggestion for, you know, a regulation around safety, like AI systems should not convince children to commit suicide. But China, right. You know, they'd say, well, you know, if we have this regulation, this is going to slow down American progress, and then China will win the race and the world will be dominated by Chinese AI systems. And and by implication, Chinese communism. Mhm. Um, they don't actually explain how having AI systems not convince children to commit suicide, uh, really hobbles American industry. Um, and also they don't explain that China already has much stricter regulation even than the European Union on AI systems. And yet China competes quite successfully. So the but China argument is, is really a poisonous, uh, argument in, in this context, it's, it's, it's based on a false premise that China doesn't regulate. Um, and it leads, it's used as a justification, uh, to do things that are really self-destructive for, um, for the AI ecosystem in the West. thank you, Stuart, for for joining us. Such a pleasure to have you. as we think about, all the advanced economies outside of the United States, is there a risk that a lot of other economies become dependent on a handful of AI companies for critical parts of their economies. And what does that actually mean for economic and strategic autonomy? Uh, that's a great question. And I think to many countries, um, you know, and I spoke with, uh, India's chief scientific advisor late last year on exactly this topic. You know, so many countries view this as the likely evolution of the global economy, that as AI systems become more capable, humans won't be competitive anymore. Um, and, uh, a large fraction of the economy will consist of, uh, American AI systems and their robotic extensions, uh, doing most of the work. And at that point, your domestic economy just becomes a vassal state of the American AI companies. So that's what they see as the likely path. At least the path that is being laid out for them. Um, and this is very explicit in the White House's AI action plan that the world will be dominated by American AI systems. And that's it. So that's not a very promising future. It when you when you think about it from the point of view of an individual citizen, whether they're in the US or in, uh, you know, global South or even a European country, um, almost all the paths to self-improvement, as we have traditionally conceived them are going to be cut off. Right? There won't be. Oh, you know, I could, I could go to law school and become a lawyer and and be successful and have status and, you know, economic security and so on. That path may not exist anymore. And all the other paths you can think of may not exist anymore. And, you know, we can argue about universal basic capital or universal basic income or various other kinds of redistribution. But that doesn't solve this basic issue that there in some sense, what we used to call, you know, a future, right? If you talk to your kids, you say, you know, think about your future, you know, work hard, do this, do that. That, that concept just goes away. I mean, given what you just said, We all, you know, we were born in Europe. Uh, you are one of the most acclaimed experts, professors, researchers on the topic of AI. Um, I can probably name a handful of others who come from Europe. Europe produced so many remarkable experts and share of people who shaped AI research on it, innovators who built amazing companies and careers. Why has Europe struggled to retain this, this talent Why has everybody left? That's an interesting question. I mean, I can say how it was for me. Um, you know, so I, I was an undergraduate student at Oxford in, in physics. Um, had, I had, I decided to stay in physics, I think I would have found a future in Europe. Um. I probably would have gone to Cambridge to do my PhD and then ended up in Geneva working at CERN or something. You know, I wanted to be a particle physicist. Um, and if I remember correctly, at that time, uh, funding for physics in the UK was at thirty times the scale of funding for computer science. Um, the UK government back in nineteen seventy three had more or less banned AI as a, uh, as an academic research topic. Um, they allowed it to continue at Edinburgh and Sussex, but they basically said there's not going to be any AI funding. Um, they changed their mind in the mid eighties. That was already too late for me because I had by that time decided to go to Stanford. Um, so when I wanted to, to look into studying AI at the PhD level, this would be nineteen eighty one when I started thinking about that. Um, basically people told me to go to the US that that's where there was active research where government, the government was funding that research. And, um, so I went to Stanford to do my PhD and I think that happened to, to many computer scientists, Even those who perhaps became originally, uh, did their PhDs in the UK and then became professors at Edinburgh, for example, like less Valiant moved to Harvard. Um, Geoff Hinton moved to Toronto. Uh, and as you say, right, the, the three people who won the Turing Award for deep learning, Geoff Hinton, Yann LeCun and Yoshua Bengio are basically one Brit, two French people. Um, and so entire generations of British, uh, computer scientists just upped and left because the opportunities in the US were so much greater. And that was primarily to do with the US government funding basic research, which then created Silicon Valley and the economic Opportunities. Um, and it's, it's a pity. I think the UK government was very much behind the times not understanding really the impact of, of computing generally. Um, I think the same was probably true for a lot of European governments. There was a sense that computing was just not a respectable academic discipline. It had nothing like the status of physics or mathematics or molecular biology or history or comparative literature. Right. Where, where people would view Europe as a a center that, you know, was at least comparable to the US and in some of these areas actually superior. But computer science was always sort of a poor stepchild in UK and Europe. Let's turn the lens a little bit back to Europe. And the notion of sovereign AI, national security, right? All of those issues that a lot of which we've already talked about or touched upon on, on in our conversation. AI is a big space. Um, when you think of sovereign AI in that tech stack, what are the things that actually matter? Is it the models? Is it the compute? Is it the talent? Is it the cybersecurity, the data, the protection of it? What matters most? Well, I think first of all, it's important to understand that pretty much no country in the world has a sovereign stack. You know, this is something that I would say policy makers fixate on, right? We must have, you know, a sovereign stack. You know, I heard this in India. I heard it in Europe. Um, but the US doesn't have a sovereign stack, right. The, the machines that make the chips come from the Netherlands. The the actual making of chips happens in Taiwan. Um, you know, the only area where the US has something close to a monopoly is, is in the design automation. So companies like cadence and Synopsis, um, both of which just to wave my Berkeley flag came out of Berkeley. They have, uh, about a ninety nine percent share of the global market. Um, so, so that means there really isn't a country that has a sovereign stack. If you look at what, what are we worried about now? Are we, um, you know, if I'm the government of India, I'm worried about the AI systems, not about whose chips they're running on. Um, and I think from a strategy point of view, it would not make sense for India to try to replicate ASML and TSMC. Um, that would take I should not even sure it's possible, but it would certainly take at least a decade and, and perhaps hundreds of billions of dollars to do that. Um, by which time the industry is already left you far behind. So I, I feel that the right place to start is at the top end at the consumer facing systems and the B2B systems. Um, and, you know, to a large extent, this is what China is doing. They also face restrictions on the ability to import, uh, the chips to train the AI systems, which is, which is a problem and also restrictions on access to the Western hyperscalers. Um, but They'll, you know, they're focusing on how do we get these AI capabilities? And in China, a lot of that is open source. How do we get it to actually be beneficial to our economy? Um, and, and also useful to our citizens in, in, in a way that, uh, in some sense, um, contributes to their well-being rather than just viewing citizens as, um, you know, a source of income for the, for the, for the AI system providers. So I think if I, if I was a European country or India or another country in the global South, I would be concerned that my citizens have, you know, their only option for gaining access to the benefits of AI is to get a subscription with an American software company. Um, so you know, how, how you get compute, you want to start an Indian, uh, consumer facing AI model, you know, maybe one that is more adept in, in the various Indian languages. Um, the right thing to do would be to, you know, to get compute from, from one of the existing hyperscalers and just view compute as a utility now. Right. And, um, it's difficult to provide it more economically than the hyperscalers are already providing it. Um, you know, it can be done, but it's not the first order of business. Once you have a model up and running, you know, that is good and is contributing value for your citizens. Then you could think about migrating it onto a domestic platform. But I think you start at the top and you and you work down using the revenue that you generate. It's maybe in a decade or fifteen years time it will become economical. But there's still there's just the physics that you can't overcome. Let's turn a little bit to our audience and lots of board director CEOs naturally thinking a lot about this topic. What questions should those business leaders be asking before approving any major AI deployments? If you're in the boardroom, if you're in the C-suite. What are the questions they need to be asking? So let me preface this by saying, you know, my experience, uh, in, in the boardroom is somewhat limited. I am a director of a large Canadian insurance company in tech financial. Um, and so I have experience of that process as they've rolled out AI in various functions in the company. So so first of all, AI is not one technology. I think it's very important. Most media articles you read about AI sort of just view it as this sort of unitary, um, commodity in some sense, right? You buy a certain amount of AI and you stick it in your company. It just doesn't work like that. Um, and what I've always told people when they, you know, when they ask questions about how could I use AI in my company, I say, well, first of all, you have to get the data plumbing, right. That that's, and that's going to give you ninety percent of the value, right. If you get the data plumbing right, you can build even very, very simple machine learning systems or. decision support systems on top of that. But if you don't have the data plumbing, right, if you, you know, then there's nothing the AI system can do. It's going to be operating on incorrect or out of date data, or not even have access to the right data at all. Um, and so, you know, that process has been unfolding, but it's still remarkably slow even in companies that you might think, you know, would have solved this decades ago. Uh, you know, like big banks, for example, know, you know, the, some of the big banks are operating with literally hundreds of legacy systems that are all siloed from each other where the data is, you know, essentially the same fact is duplicated twenty times across twenty different systems and gets out of sync and, you know, has errors and, and all this stuff. So, so these are very basic things that from, from academic computer science, you know, we all think, oh, this is, you know, this is just, you know, the table stakes for, for being in the digital age, but it's still, um, it's still not a solved problem in many, many companies. And, you know, you can see this by, you go look at the scale of the, the meetings where people talk about interoperability, for example, data integration, schema integration, all this stuff. It's still a huge concern. Um, so that's the first thing, right? Get, get all that in place and you will start to see the benefits just from that even without the AI, right? Um, and then, um, then you have to figure out, well, what, what are the things that AI could be used for? And you have to understand what what are the options, right? What are the kinds of AI that exist? And everyone thinks that, oh, large language models do everything. So I'll just shove them in everywhere. But they don't. Right. If if you want to do something with your supply chain, for example, then you're talking about a very different family of algorithms, um, uh, based on work in planning, uh, and other people, you know, working in operations research, for example, that goes back for decades and comes with mathematical proofs of correctness. Um, and, uh, you know, that's a completely different technology from large language models where, you know, you could ask the large language model, you know, how can I get, you know, Twelve hundred units of of this product to, to this warehouse by next week. And it'll just make stuff up. Right. It'll, it'll move units of product that don't exist using, uh, using transportation systems that could not possibly get it there in a week. And, um, and so on. So, um, so what people's experience so far has been, is of trying large language models in lots and lots of possible niches within their company, lots of places in their workflows and, and so on that plausibly could benefit from that type of system and finding in some cases that the level of error, uh, is intolerable. And in some cases finding that, that it works. So developing this experience and sharing best practices is is happening all the time. And I think this is something that, you know, a year ago, we weren't really thinking about very much is what effect does having AI systems as part of the operation? What effect does that have on the humans and their understanding about what's going on? And we're starting to see complaints that we just have no idea what's going on. Um, and in some sense, in some cases, we don't even trust the AI systems to tell us what's happening. And a kind of a cognitive de-skilling effect where people very quickly come to rely on the AI system to do the cognitive work for them, and then they've lost context and they in some sense even lose the ability to do that kind of thinking anymore. And there were some nice experiments done recently with, you know, proper randomized controlled trial. And you have one group that's using a chatbot to help them with some simple math problems, and another group that's just doing it without the access to the AI system. And then they switch to a mode where everyone has to do it without the AI system. and the people who had been using the AI system give up twice as often. After ten minutes of using the AI system. Wow. So it's you know, that's a huge effect, right? Psychologists usually, you know, they're trying to get, you know, a quarter of a standard deviation or something like, you know, and it's, but this is a huge effect. And, you know, I think people will redo that type of experiment in lots of different ways and we'll understand more about what's really going on. But, um, you know, this is. Also showing up anecdotally, you know, where I talk to people who, you know, they start to worry about their friends or their direct reports who can't function without the AI system, you know, and there are, there are a lot of sort of, you know, tick tock parodies and, and other videos going around where people say, why did I come into the store? Uh, you know, and they're asking their AI system to remind them why they went to the store and, and, uh, and so on. And these sort of, you know, comedy videos about, you know, the human cognitive dependency, but it's real life that people are really losing the ability to function as human beings. Stewart. I know we, um, a little bit over time, um, I don't know if we can try and experiment because I was, I would love to ask you one last question if that's okay with you. Sure. Um, so you obviously talk a lot and think a lot about the risks of AI. Uh, and, um, you alluded to, you know, what happens with humans. Um, maybe as our last questions, if you, if you look ahead, right. And if you envision, you know, work and a positive society evolving, What, what, um, what should, you know, corporate leaders, governments, you know, what humans, people who are building with AI, what, what should they do and have in mind to ensure that we are, you know, moving into a positive future, uh, which is supporting humanity and, um, hopefully making us all better, better humans, uh, and not moving into the gloomy side of things. Many people also talk about. Yeah. So, uh, so I think the premise here is that the, um, the worst case scenarios that people describe, you know, the loss of human control, uh, over AI systems that are essentially much more intelligent than human beings. You know, let's say that we find technical solutions to that problem. So then we have this question how do we coexist in a positive way with AI systems that are much more intelligent than human beings? Um, and I think this is the second most important question, right? The first, most important question is, uh, how do we not lose control to AI systems that are much more intelligent human beings? So let's assume we solve that one. I have asked many, many experts, you know, I've convened workshops with not just AI researchers, but economists, science fiction writers, futurists, and saying, basically, give me a description of a world where we have superintelligent machines that can do pretty much everything we currently call work that you would want your children to live in. And so far, I would say that the, you know, the chorus of answers is, is not deafening, right? There's actually very little in, you know, in think tank white papers, you know, all the way to science fiction. There is very little in the way of a concrete description of how this would actually work. And even in the most positive science fiction scenarios, the the problem that's unsolved is, is the problem of human purpose. And I don't think this is an unsolvable problem. And it might just be that we are still operating with blinkers created by some sense, thousands of years of channeling, of human effort, our education system, our culture into essentially productivity, right? That in order to have a functioning civilization, people and to have a reasonable standard of living. And, you know, physical security and so on, people have to be channeled into becoming productive members of society, whatever you want to call that. So what happens when that requirement goes away? Because the productive members of society that we used to need are replaced by AI systems. What what is the process that replaces that in terms of giving humans purpose So if you're really going to deliver value to people in these kinds of roles, we need much more understanding than we have now of human psychology, child development, all these things, right? You know, and it's, it's a little bit different if you think about like an orthopedic surgeon, right? They're, they're very good at fixing broken legs and arms and all the rest. Um, but only because we have, you know, thousands of years of investment in medicine and, and the underlying science and a huge volume of experience in how to do this. well, but we don't have that for human psychology and just human flourishing in general. Um, and part of the difference is that human minds are much more variable than human leg bones. Mhm. And, uh, so it's a very different kind of science, um, perhaps much more difficult. But I think, um, we need therefore to change the focus of our, uh, scientific research establishment, our education process, our professions and so on, you know, and it's, it's urgent because the changes are going to be very rapid. And our, our education system in particular is notoriously slow at adapting. And, uh, and so we've got to figure out what the The destination looks like, you know, what is this world that we want to reach, right? You know, not a world of ninety eight percent unemployment with two percent of incredibly rich and wealthy people. You know, then they have a layer of servants who serve them, and then everybody else is essentially. Redundant. And that's not a world anyone wants. So what is the world that we want? And then how do we we have to take steps to get there. But we can't take steps until we know what the world is that we want. So I think you're asking the right question.

Closing

That's it for this week's episode of Inside CVC. Thanks for listening and we'll catch you next time.