Carl Benedikt Frey on AI and Growth

Society progressed and evolved for thousands of years without any real significant technological advancements. So do economies really depend on technology to grow?

Well, I think the experience since the first Industrial Revolution has shown us that technology is very important to economic growth. And if you look across countries and you look at levels in income per capita, that is very strongly correlated strongly correlated with various measures of technology adoption. So I think we have good theoretical and empirical evidence to substantiate that claim. My name is Carl Benedikt Frey. I'm an economist and economic historian by background, and I'm the Associate Professor of AI and Work at Oxford University.

Benedikt Frey's latest book, How Progress Ends, looks at some of the side effects of technological and economic progress. He also has an article in the September issue of Finance and Development magazine on how the race to control AI could undermine its potential. He spoke to me from his office at Oxford.

Thank you so much for joining us on this podcast.

Pleasure to be with you.

So the underlying theme or thesis of your book hinges on this idea that technology is what makes or breaks societies. And you do argue that countries that aren't actively taking part in this technological evolution face stagnation. So what is the basis for that assumption?

Well, I think a key point of the book is to argue that there is nothing natural or inevitable about technological change and technological progress. If there was, the first industrial revolution would've happened a lot earlier in human history. If that was the case, most countries around the world would've adopted modern technology and would be rich and prosperous as a consequence. And more broadly, we don't see this convergence in income per capita that standard economic theory would predict. So there seems to be some bottlenecks, some binding constraints on technology adoption. In the book, I argue that it's inherently political. A key difficulty with technological change is that you constantly need to adjust your institutions as technology progresses. There is no end of history where you arrive at the right institutional recipe and you're done. Progress is constant work in progress. And often there are incumbent forces with a stake in the existing system, and they tend to resist. technological changes that threaten their rents. And that's a big part of the reason why we can see some economies sometimes growing very rapidly, but then to get to a more decentralized system that encourages innovation, as I argue in the book, you need more competition. And most institutions and people are not very interested in having more competition.

Okay. So the technology happens, going back to your sort of decentralization idea, technology happens when people are motivated to come up with ideas that serve our needs and wants. And among the case studies that you explore in this book about creating that space include California's Silicon Valley, of course. But more surprising to me at least is the ill-fated attempt at creating something similar in Russia. Maybe just explain the thinking behind those 2 approaches.

Sure. So I argue in the book that technological progress involves 2 very different phases. One is sort of more exploratory. So before you produce something, you need to explore and develop a new technology and product. And when you then settled on what it is you want to do, technological efforts become more focused on scaling and bringing that product to the market at lower costs. Effort shifts more from exploration to scaling, more from product innovation to process innovation. At some point, though, you run into diminishing returns, and the return to lobbying for restrictions on competition eventually becomes greater rather than investing in improving your product. And that's usually when stagnation happens. So if you look at some of the leading theories around the wealth and poverty of nations, You can broadly place them into 3 buckets: geography, culture, and institutions. Now, take the Soviet Union. It grew very rapidly in the 1950s, but it collapsed in the 1990s. Was geography of the Soviet Union very different in the 1990s from the 1950s? Answer is no. Was the culture of the Soviet Union very different? The nation very different in the 1990s? No. And if anything, it was probably somewhat more open and somewhat more influenced by Western entertainment, for example. And finally, when it comes to institutions, well, there were some attempts to institutional reform by Gorbachev, which went more in the direction, of course, of liberalisation. But on the whole, the Soviet Union was still very much a command economy. And so what I argue in the book, that if you want to understand the growth phase as well as the collapse, then you need to understand how institutions and technological change interact. You can grow quite rapidly for a reasonably long time just by catching up to the frontier, by absorbing technologies invented elsewhere and mobilizing resources and scaling Quite effectively, but eventually you run into diminishing returns. Something similar happened in Japan, but you know on smaller scale, it grew very rapidly up until the 1990s, essentially stagnates. And part of the reason for that is that you know while Japan was widely predicted to overtake the United States in the 1980s, and many worried that you know what had happened previously in automobiles, where Japanese producers were outcompeting their American counterparts. The same thing was now happening in semiconductors. And by many accounts, Japan was very productive in cutting-edge hardware, but it completely missed the transition into software and e-commerce. And part of that is that growth was very centered around large conglomerates that were efficient in terms of achieving scale, in solving coordination problems, in terms of sharing technological know-how between each other. But they also created very significant barriers to entry. And the US market and US institutions were just more competitive. So the antitrust case against IBM, for example, meant that IBM had to unbundle hardware and software, which opened up the market for new entrants like Microsoft.

Yeah.

And similarly, the breakup of AT&T meant that when the National Science Foundation eventually made ARPANET, the predecessor to the modern internet, available, it didn't just hand it over to a state company or to a monopoly on telecommunications, as happened in many other countries. And because AT&T had already been broken up, the internet could develop more organically through its users, and the technology wasn't stifled in the boardroom of one incumbent.

Yeah, or the government.

Indeed.

So setting out to write this book on a topic like this must have felt a little daunting given how broad the impact of technology is on our lives now. But I love how you kick it off with a quote that I think captures the complexity of it all. I'm paraphrasing here, but it goes something like, if a baron stole horses from a neighbor, the size of his cavalry would most certainly grow faster than if he had a successful breeding program. So why did you choose to start the book with that? I mean, what is the relevance there?

Well, I think it's a wonderful quote that illustrates so many things. One thing is that Technological progress is fragile and by no means natural, as I alluded to earlier. And in fact, once you invent something, once an idea is out there, it can be adopted by somebody else and used to outcompete you and erode the investment that you made. Whereas if you just take somebody else's stuff, you are better off in relative terms. So technological progress is something that is positive sum, but highly uncertain, whereas stealing is more certain and negative sum. And the key question, obviously, for any society is how do you get to positive sum innovation rather than zero sum transfer of existing resources? And I think this brief quote highlights that. That important dynamic very well.

Yeah, it's great. I mean, when I read it, I was thinking this is sort of like saying, why learn how to fish when you can just steal the fish? Because we've been hearing that all this time about development. So you talk about the importance of data in developing new technologies in the book and in the article that you've written for Finance and Development. And it's especially true I mean, the data thing is especially true in the case of AI. So who owns all this important data these days?

So one of the key points I try to make with regard to AI in the book is to actually push back a little bit against this notion that data is the new oil. So it is certainly true that that data is very important, just like oil is very important. But anybody who studied development economics also knows that there is such a thing as a natural resource curse. And oil-rich countries generally aren't very good at economic development. Now, if you go back to technology like the steam engine, in its early days, the steam engine was very cold and efficient. It was primarily used to drain coal mines, and even that, it didn't do particularly well. And so for the steam engine to be adopted in factories, on railroads, and on steamships, we needed very significant improvements to improve energy efficiency. And that was James Watt's separate condenser, was a key contribution to that. And what I argue in the book is that We still need a separate condenser moment with regard to artificial intelligence, because one of the key differences between AI and us is that AI needs vast datasets to function. We can learn from just a few examples, and anybody who has children knows that. Intuitively. And so if the world was just a static distribution of events, and you just had all of these events in your database, then you can certainly just brute force general artificial intelligence by larger datasets and more compute. But the world isn't just a static distribution of events, it's changing all the time. And that means that you cannot really know with existing methods to what degree the algorithm you're relying on will be reliable when circumstances change.

And you point out that an AI using these LLM or large language modules from a couple hundred years ago would have denied that humans could ever fly.

Yeah, indeed. So imagine you have data up until 1900, and you ask it, will humans ever be able to fly? So all the data it would have access to would be a long list of failed experiments and numerous deaths with people attempting to get off the ground. The only evidence you really had that flight was possible was data on birds, but even In the air, larger birds like ostriches are not able to get off the ground. In fact, there's no bird that weighs more than 30 pounds that can fly. And so presumably even that data would lead the LLM to infer that humans would not be able to fly. But with the benefit of hindsight, that would have been the wrong conclusion.

So we spoke briefly about Russia's attempt at creating a space for innovators in Zelenograd. And essentially Russia's kind of out of the picture in terms of any major technological advancements or innovations these days. But China is not. And it's hard to know exactly just how much China is putting toward developing AI, but there's no doubt that it's happening. And given all the data that that they are gathering, could China actually achieve what Russia failed to do back in the day in becoming a centrally planned technology powerhouse, do you think?

So I think as you sort of alluded to, China is not centrally planned. And in fact, a key difference between China and the Soviet Union from the onset is that the Soviet Union was organized along line ministries, which essentially means that across the entire Soviet Empire, the central ministry managed all the railroads or the entire steel industry. Whereas in China, provincial governments played a much larger role in managing their own local economies. And what that meant is that China, as it opened up, could essentially create a number of localized experiments, which it could then scale on the national level, the things that worked. And so China was much more capable of experimenting because of those differences in its institutional Blueprint. Now, in addition to that, China has opened up to the world. If you look at which Chinese firms that are most successful, it's foreign-invested companies rather than state-owned enterprises. It's mostly smaller firms rather than larger firms that are most innovative. So, in that regard, China is actually not that different from the West. What is obviously different is that the CCP and the Chinese Communist Party has a monopoly on the political sphere. And because the rule of law is not as predictable as it is In the West, building political connections in China is more important. And so if you're DeepSeek, for example, you can certainly fly under the radar of the CCP, and you might even be encouraged to do what you do as long as it serves national objectives. But when priorities change, which they often do, your entire business may also overnight be at risk. And so now for these reasons, it's quite hard also for the party to instrumentalize private firms for national objectives. And for that reason, the more national objectives you want to pursue, the more you become reliant on state-owned enterprises, which aren't very productive and innovative, or at least haven't been historically. And so So, making them a centerpiece of your growth model, I think, is unlikely to be a huge success. So, China is also facing some challenges in that regard, but it's certainly much more decentralized than the Soviet Union, where, for example, if you were an aircraft engineer and you had developed a new engine, You could go to the Red Army to ask for funding to build a prototype. If they turned you down, maybe you have 1 or 2 or 3 other options. If they turned you down, the idea was essentially dead. And obviously, if your idea didn't serve national objectives to begin with, it was no reason to pursue it to begin with. And obviously, that is not quite the case in China today.

So what about institutions in all this? I mean, I did a podcast recently with Simon Johnson, and his big concern was that AI is undermining these institutions that he says enables the technology to develop. What are your thoughts about that?

So I grapple with that I talk about that a bit in the book. And one key concern, if you look at the field of AI, is that it becomes more concentrated, focused on a few players. There are fewer institutions, for example, in academia that can keep pace with the resource requirements of the latest models. Yeah. More broadly, AI used to be a fairly broad church consisting of people taking very different approaches on AI. It's become fairly narrowed down on deep learning and generative AI in particular in recent years. Now, if you just look at the foundational model providers, there is significant competition among them. But at the same time, If you look at some measures and estimates of market concentration, Microsoft and OpenAI through their partnership controls almost 70% of that market.

Yeah.

And I think no less significantly, if you look at the industries that are using AI tools, they have become more concentrated as well. Why does that matter? Well, large firms are more likely to use AI to support scaling, to support process innovation, to make what they're already doing available to existing Consumers or new consumers at lower cost. And that is important, but process innovation and automation quite quickly runs into diminishing returns too. So think about it this way. If all we had done since 1800 was automation, we would have productive agriculture and cheap textiles, but not much else. We wouldn't have antibiotics, vaccines, rockets, airplanes, computers, et cetera. Most growth has come from doing new and previously inconceivable things. And those new things are more likely to be done by startups, which are facing growing barriers to entry. And so, we have a quite concerning trend even in the US, which is that the dynamism that usually made America special. So, if you look at the 5 largest companies by market capitalization in the US, they are 39 years old on average, and all are in tech. If you look at Germany, they're 100 years plus, and mostly stem from the Second Industrial Revolution. In France, they're 100 years plus old and focused on luxury goods. But even in the US now, there has been a decline in dynamism in startup formation. Those new firms don't grow as quickly. As they used to. And I'm happy to concede that Microsoft and Google and these companies are still innovating, but that's not very surprising considering the fact that they are accounting for a larger and larger share of the world's inventive talent. And so we have some empirical studies now that show that actually people who work for incumbents more broadly, used to be more productive in the startups where they previously worked. So we should be quite concerned about these developments in the US as of lately, because not only is there a risk that AI development will be monopolized, there's also a risk that the concentration of AI-using industries will push the direction of technological change further towards automation rather than development of new products and industries.

And so finally, I guess in that same vein, in terms of institutions perhaps getting in the way, I mean, how do we find this balance between oversight and regulation without getting in the way of progress?

So I argue in the book that pre-industrial China basically used technology to develop a bureaucratic state. to be able to raise taxes, and that way essentially undermined decentralized exploration, which the fragmentation in Europe fostered. And the key theme of the book is that you can use technology for surveillance and for monitoring, and you can create some Sputniks by doing that, but you're not going to create the broad-based kind of dynamism. Need for sustained growth and progress. And so paradoxically, by being an early adopter of technology, new technology for monitoring and surveillance purposes, which might give you a boost to growth in the short run, you might undermine growth in the long run. And I think that is what happened to China in pre-industrial times. And I worry that China and the US as well is at risk of repeating some of those mistakes.

Well, Carl Benedikt Frey is the author of How Progress Ends. Congratulations on this new book, and thank you for sharing some of it with us on the podcast today.

It's been a real pleasure, Bruce. Thank you very much for having me.

Again, that was Carl Benedikt Frey, the Dieter Schwarz Associate Professor of AI and Work at Oxford University. Check out his latest article on the battle for control of AI in the September edition of Finance and Development magazine. Go to imf.org/fand. And look for other IMF podcasts on Apple Podcasts or wherever you listen. You can also follow us on X. Our handle is @IMF. Thanks for listening.

@IMF_podcast.

I'm Bruce Edwards. Thanks for listening.