How Technological Progress Happens (and Ends) with Carl Benedikt Frey

Welcome to Geoeconomic Competition. I'm Francesca Ghirardi, your host, and in every episode I dive into the intricacies of geoeconomics with other experts. You might have noticed we haven't really done an episode on tech or on AI yet. So today I have Carl Benedikt Frey, Dieter Schwarz Associate Professor of AI and Work at the Oxford Internet Institute, but also the author of The Technology Trap and of How Progress Ends. 2 beautiful, fantastic books that I recommend you go buy and read. Thank you for being here with me today, Carl.

Good to be with you.

There are many very interesting things in your latest book, and one of the most interesting framings, at least for me, is that society needs 2 engines that often are treated as somewhat competing. And one is the decentralized exploration to generate new ideas. We could call it somewhat innovation. And the other one is centralized exploitation to scale them, adopt them, diffuse them. You pick the term. Can you walk us through how that tension works, and especially why getting the balance wrong has been fatal for so many great powers?

So if you run an assembly line, or if you run a factory, and every single worker on that assembly line questions every single procedure, you're not going to produce too much, right? But if nobody questions anything, you're also not going to invent anything new. And sort of that's the tension I'm trying to explore in the book. And the key point, as you mentioned, is that if you want to explore new things, a certain degree of autonomy, different people exploring different technological trajectories is usually important to come up with new ideas, new technologies. But when you then settled on what it is you want to do, and when you want to produce and make your goods or services available to as many people as possible, you want to cut costs, you want to exploit new technologies like factories that then leads you towards managerial hierarchies and greater centralization. And so both are needed for progress, but different stages of the technology lifecycle, if you like.

And we get to what stage we're in and who's doing this best in a moment. Part of your book is tracing the history and the experiences of some of these, let's call them actors, great power, whether it is Song Dynasty in China, the Dutch Republic, Victorian Britain, Soviet Union, and so on and so forth, but also modern Japan. And we get to that in moment. And you notice how many of these surged ahead and then stalled, right? So is there a common pattern in how the transition from the rise to stagnation plays out, and perhaps some lessons learned or elements we should keep our eyes open for?

So the case I make in the book is that moving between decentralization and centralization means that you need some institutional flexibility, and that is hard to do over the long run, because you need to change your institutions. And you often have incumbent interests that have a stake in the existing order of things. They are often going to resist reform and new technologies as well. But you can grow quite quickly for a fairly long period of time without being at the forefront of technology. And so if you take the Soviet Union, for example, we tend to look back at it as a failure now. And obviously, in many ways it was, but it was also growing at 6% per year over a period of around 4 decades or so, which is quite respectable. And it did so basically by taking technology invented elsewhere and scaling it domestically. And so it took ample advantage of the Ford Motor Company's open-door policy, sent delegations to Detroit to look at the Ford factories, learn from that, and build a vehicle industry in the Soviet Union. And in the age of mass production, the authorities could hold factory managers accountable by simply benchmarking performance across plants. And so it could that way determine if some factories were slacking off and not producing as much as maybe should compared to its peers. But at the frontiers of technology, that becomes much harder, right? Because when something is new, what do you benchmark against? In addition to that, what the Soviet Union needed for further growth as the mass production system petered out globally was something new. That something new was the computer revolution, to which Soviet contributions were essentially none. And a big part of that was that if you were an inventor in the Soviet Union, you could go to, let's say, the Red Army and ask for funding. If they declined, maybe 2 or 3 other options. If they declined, then basically your idea would die with you. That's quite different from the American system of much more decentralized finance, where Bessemer Venture famously declined to invest in Google back in 1999 and probably regret it today. But it also illustrates that in the late '90s, Google was not a safe bet, right? AltaVista and Yahoo were dominating search. And so somebody needed to take the risk to invest in order to know if Google would catch on. And the fact that Bessemer didn't invest didn't mean the end of Google because others stepped in. And so in a more decentralized system, you have different players that can take different bets, you can explore more technological pathways that way. And so that's more important at the technological frontier when it comes to pushing innovation, and where the opportunities for simply adopting technologies invented elsewhere diminishes.

Alright. And the other side of the spectrum is sort of the adoption part, which, you know, you've already mentioned in relation to the Soviet Union, you've already mentioned the limits. Perhaps in authoritarian states, other limits could be sort of the political purge and the political cycles where then the same people people who were praised 10 years ago that then put aside or even imprisoned, or we've seen that in the Soviet Union. We've seen that in China as well, most recently. But if I can ask you a question is, so where are we now? What are we in the phase where we should be betting into innovation, or should we bet towards diffusion, or is never really a single phase and there's a bit of both every time?

I think the answer is there's a bit of both every time. The second question is, who are we, right? So, if the focus on artificial intelligence, for example, and the European Union, you could argue that we have some catching up to do. At the same time, you could also make the case that AI, impressive as it is, may not be quite as transformative yet to have a material impact on economy. And so we're not seeing much of an impact on artificial intelligence either in the United States, or though it's harder to tell, in China either. So both are having productivity problems. So my view is that we need more innovation in AI to push forward. I tend to think that artificial intelligence needs its separate condenser moment, by which I mean that, you know, if you go back to the first industrial revolution and the proliferation of steam power in the early days, you know, steam engines were tremendously energy inefficient. They were only used to drain coal mines, and even that they didn't do particularly well. And so for steam to be applied to transportation, to rail, to steamships, chips and so on, you needed further innovation and you needed the separate condenser in particular for energy efficiency. With AI today, I think we need greater data and compute efficiency. And so humans are very good at picking up patterns from just a few examples. We don't need millions of trials. And so if the world was just a static distribution of events, then you could probably brute force AI, just throw more data and compute at the problem and you will get there, right? But the world is changing every time. And so the question always emerges, even if AI was accurate yesterday, is it going to be accurate today? And I think the key example here is Stuart Russell and his lab. Everybody knows about AlphaGo beating the world champion back in 2016. What few people know is that actually human amateurs using standard computers, exposing the best Go programs to positions that they would not have encountered in training, actually beat AlphaGo.

Wow.

beat those programs relatively easily. And so that also puts forward some fundamental questions in the sense that even if AI has achieved superhuman performance in a domain, when exposed to new situations, novel circumstances, that may no longer hold. And I think that resilience challenge still persists.

I have 3 questions. I'm gonna anticipate all 3 of them, but we're gonna go through one by one very, very slowly. So the first question is linked to what you've just said, and it is about the tech competition, right? So is innovation that important in the sense of having the breakthrough in your country, let's say the United States rather than China? Is that fundamental? Or at the end of the day, if indeed knowledge flows, and this is some— is also part of your book, part of the conversation in your book, if knowledge flows, then where the breakthrough happens is not as important as how you diffuse or you adopt that technology. And I'm asking this specifically because there is this sort of framework going around that the United States is innovating, innovating, innovating, but China is thinking a lot more about industrial implementation and how to turn AI specifically into productivity. Let's use that word. And this is sort of question number 1, very, very small question, right? And then question number 2, I'll skip the number 3, is if indeed then where the breakthrough happens is not fundamental. Then does that leave some hope also in the AI development for Europe, where no frontier company has been born, no frontier model, and so on and so forth? So these are my 2 questions, and we can go with the first one, which is US-China. How do you see that playing out?

So I do actually think that where the breakthrough happens matters, in large part because industries tend to emerge around those breakthroughs. The job hopping that then emerges from that then transfers a lot of tacit knowledge, gives rise to new businesses, and gives rise to dynamism more broadly. So I don't think it's quite as simple as, you know, you have a breakthrough and then, you know, that knowledge proliferates everywhere and every place adopts the technology. And if that was true, obviously every country, every society would be at the technological frontier and rich and prosperous, right? And so I do think that being at the frontier matters because that is also where global talent tends to go. But you can obviously undermine that advantage as we're seeing right now as well. Secondly, on adoption, so I think it's absolutely true that it's when technology is becoming adopted that it's beginning to have an impact on the economy. But I think adoption is also not one thing, right? Both you and I, we've adopted the computer, right? We probably use it for a lot of similar things. We probably use it for different things as well, right? And so So just by looking at computer adoption, well, it drives some differences in productivity across space. But in the end of the day, it's a question of what you use the technology for. Do you use it to develop new ideas, new businesses, etc.? Do you just use it to automate a few processes? And so I think the more relevant factor is dynamism. And I think dynamism is related to where these innovations happen, because that's where you get the inflow of talent. That's where you get new firms emerging, you get clustering. And so I think that part is actually—

Right.

And I forgot the second part of your question. Sorry about that.

It's my fault. I always anticipate too many questions. The second part is about Europe, right? And I think that you've already in part already answered that question. If there is no breakthrough in Europe, does that mean that Europe is going to lag behind for a very long time? If yes, what we can do about it? And if no, what we can do about it in the sense of where Europe is now and how— I'm not saying we should have a recipe, but what, according to your studies, what could be a way to catch up in in the current system?

So clearly, Europe has a tradition of being able to do catch-up growth pretty well. So it caught up to Britain in the 19th century, it closed the gap to the United States in the post-war period, but it hasn't really even closed the gap in digital. And so why is that? I think part of it has to do simply with the fact that the single market is less harmonised in digital and services than in goods. As a result of that, the market is more fragmented, and so the return to innovation in that space is simply lower. I think when it comes to trading services, language and culture matters more, and so that is inevitably going to be a bit of a drag on Europe when it comes to scaling, although they might have other advantages. I think the costs of experimentation are higher in Europe, in large part because of labor protection, which has some upsides, but it means that hire and fire is harder. And that means that it's more costly to experiment and pivot when a new technology arrives. So Meta can just pivot to the metaverse and then just reverse course and then pivot towards generative AI, whereas a firm in S&P will struggle for years to restructure. And for example, I think barriers to entry are higher in part because of rules and regulations like the GDPR, which introduced compliance costs which make it particularly difficult for younger firms entering the market. I think those things together probably can go some way in explaining why Europe hasn't caught up in digital the way it's caught up in mass production. So what does Europe do? Well, I think there are sort of 2 broad school of thoughts out there. One is sort of more related to top-down industrial policy, and you will hear Airbus as the lead example. And that is I think at the same time, it's also true that aircraft technology was fairly static at the time. The jet engine was the last sort of big breakthrough that had already been developed. And so Airbus was sort of catching up to a fairly static target being Boeing. There were numerous policy initiatives to seed domestic semiconductor industries, which essentially failed. I think part of the reason being that semiconductors are more dynamic. It's a moving target. And I think that's even more true of AI. I mean, we don't know if the future is going to be AI.

Yeah.

Future of AI is large language models, small language models, world models, some fusion of, you know, symbolic AI with something else. And so I think that's a process we will figure— something we will have to figure out through trial and error. And that means that we need to, you know, take a fairly diversified set of bets. And for those bets to be taken, I think we need to sort of solve some of the underlying challenges, such as the lack of a harmonized single market I'm very glad that I didn't even suggest it and a more unified or harmonized single market was your first answer.

It's something quite recurrent in several episodes. It's my personal thought, but I never try— I always try not to bias the question. So to the listeners, this was genuinely from Carl. It wasn't me or any of the other speakers. I'm going to let you go very soon. So I'll close with the last question, which is a little bit more broad and about AI. The book ends So we end on a note of uncertainty on what the impact of AI on progress could be and whether it will depend on how institutions foster their culture, openness, control. But if you had to place a bet right now, does the world end in a new golden age of progress, or are we at the beginning of what could be great stagnation? Now, I realize you don't have a crystal ball, but at the same time, it's a fun question to ask at the end of this episode.

So I think my base case would be something like the computer revolution, but a bit more moderated in terms of its impact on productivity. So with the computer revolution, what you got was sort of basically productivity upsurge over around a decade, mostly confined to the United States. I wouldn't be surprised if we don't see AI in the productivity statistics at all a decade or two from now. I think it will show. But I also think that certain challenges that we currently have will persist for some time. One of those challenges is verification, right? And so, you know, with computers, they were more predictable in terms of what they produced. They wouldn't fabricate sources and that sort of thing. They, you know, certainly had some challenges, but the error was usually human error. So with artificial intelligence, I think even as it becomes better, it's going to be harder and harder for us to verify the output because only Only fairly specialist individuals will be able to tell truth from falsehood, right? And so that means that we will have to learn with the algorithms in order to verify the outputs. And so in the end of the day, the productivity gains will be, you know, time saved doing the research or whatever it is you do minus verification. And so, you know, the computer revolution essentially gave us instant access to lots of knowledge, which saved hours, you know, researching the library or, you know, waiting for new material to arrive. And so it's essentially automated downtime. AI automates cognition and it still requires the human in the loop. And as long as that's the case, I think the productivity gains will be more muted than what we saw during the ICT revolution. But I still think that's the closest parallel.

And this is a great ending, especially because it's interesting. Well, the impact on the knowledge sector, if so we want to call it, is quite interesting how universities, think tanks, research centers will change. Both in terms of examination. I was reading something about, you know, a PhD thesis being submitted at a level that was completely different from the performance of the student throughout the 4 years that preceded. So is the output no longer going to be the way we judge the production of a student because it could change? But also how the organics, so how the staff is organised. Are we going to be needing a lot more specialised people with a lot of expertise that can check the content. Of course, now it's relatively early, but I find it quite funny that you start seeing on LinkedIn and around all these positions for code verifiers. So because now coding has been automatized so quickly and everyone was so enthusiastic about it, and then suddenly you have a whole new role, which is, okay, yeah, the machine is producing this code, but you'll have, you'll need a human being to verify whether that is right. So it is an open question whether all of these will create more productivity or just a different system, but I think it's a great question to leave the audience with.

Yeah.

So, Carl, thank you so much for taking the time to have this conversation with us. Thank you for joining me.

My pleasure.