One Step Forward (progress pathology with Carl Benedikt Frey)

We've had some technological progress for the past 20 years, and what is also extraordinary is how global the productivity slowdown has been. The global productivity slowdown is very real and it's happening at the time that we were very excited about exponential technologies.

What you need at the technological frontier is space to experiment. The beautiful thing about the decentralized economy is that people can take very different bets and pursue different technological trajectories.

Hey, Paul, how's it going?

Hey, Vass, all good, how about you?

I'm good, I'm going to rip from your playbook and just start with a light, easygoing, kindhearted question, which is, are we even allowed to believe in progress, like rely on it, or is presuming its inevitability now considered a dangerously optimistic position?

Yeah, it's a good one. It's not light, but there are a lot of people talking about abundance and there's techno optimists, there's lots of people talking about things that I think would fall into that category. There's lots of footnotes, there's lots of caveats. So, it's a turbulent discussion, and I think the jury's out as to how well we're actually progressing.

Well, I think our guest today is going to tell us more about that, and I'm kind of wondering if maybe we just need to set the bar a little bit lower in our minds, right? The comfort version is that invention happens. It's going to happen, it's inevitable, prosperity follows, and then eventually everyone's going to get a better phone, access to better medicine, maybe even a government website that works.

Yeah. I wonder if in any of these books and articles and things, they have an example of where a government website really nailed it, really, really worked. It almost seems like an oxymoron or something that can't work, it's just the kind of innovation that's impossible. Our guest is even going to be looking at harder questions about, is progress the default setting or is it actually almost the exception of the rule, and we perk up once in a while, but it's not necessarily the direction overall?

Well, we're going to learn. And it's really important to look back at history, and I really appreciated this read and it's really well sourced, I was learning a lot from the footnotes. The impact of future technologies on progress, it's all dependent on so many things. Investment, governance, competition, openness, education, and the capacity of institutions to adapt. There are a lot of elements, supportive elements that need to be in place to facilitate progress. And if they're not, progress will be slowed down, if it happens at all. And that's why today we're delighted to have Carl Benedikt Frey on the show. He's the associate professor of AI and work at the Oxford Internet Institute, and the director of the Future of Work Program at the Oxford Martin School.

Professor Frey, welcome to Policy Prompt.

Technology is central to your work, but so is the idea of progress, and one of the themes of the book is that the two are not necessarily the same thing or that they're not the same thing. So, let's start with a basic question, what do you mean by progress? Is there a core definition you work from or does the meaning change across contexts?

So, in the book, I do primarily focus on technological progress, although it's very much about the interaction between technology and institutions, and those interactions being what in the end of the day determines economic outcomes. But I think there's a broader concept of human flourishing, which is not at the heart of the book, but it's very much related. And I think that encompasses several different things. And so, one thing that we do know from the literature is that income and reported life satisfaction, those are very strongly correlated. And over the long run, productivity growth is, in the end of the day, what determines people's incomes. And so, technology translating into higher incomes is a key component of human flourishing, but so is meaning. And there is no strong link between meaning and income.

Although, technology does many other things in terms of transforming our life, it's not quite clear that that has led to a higher sense of purpose and meaning. I think many people working in technology have a sense of meaning and purpose, but it's not clear that the users of technology share that same sense of meaning and purpose. And so, that's part of human flourishing as well, but it's not something I discuss in any depth in the book.

Well, let me just follow up on that by, you mentioned human flourishing, there's human wellbeing, there's living standards, there's so many things that we look at, humans at the center of it, is that changing, that we're going back a little bit more to thinking about things that really matter to humans, and perhaps focusing less on GDP and wealth per capita and things like that, and looking at some of these other things about how are people doing? Is that part of how you're defining progress that we need to actually change some of the metrics that we're focused on?

Well, I think it's been clear to academics really from the inception of GDP as a measurement that it's not really capturing human flourishing. So, it's trying to measure the size of our economy and how that is growing, or hopefully growing, or contracting over time. I think there is a sense that we need a wider range of metrics to then capture how that growth is being produced. And so, looking at the emissions that are created as a result of that is one important component, we see that the most prosperous country arguably in the world, the United States, has had not very couraging developments in measures, like life expectancy. So, clearly you can see GDP per capita grow, but life expectancy in certain groups decline. And so, yes, I think there's a sense that we need a broader range of metrics, but that doesn't mean that GDP per capita or productivity and these measures of our material standards are unimportant.

I love that meaning came up so early in this conversation. Carl, one of my jokes with Paul is that he's always asking me these huge massive questions out the gate. I'm going to bring it back to the book too, and very early on in how progress ends, you open with this question of whether China can achieve the Soviet dream of authoritarian technological leadership. It also questions if either China and the USA, or the USA, are actually on steady pathways towards progress or are headed towards stagnation. Are these economies too big to fail and keep the world stable, or do we need to recognize a fragility that we maybe didn't appreciate as much before?

I don't think any economy is too big to fail. So, you can accumulate a lot of debt as a big economy, you can stretch your military ambitions beyond your capabilities, and you can see very large empires like the Roman Empire in the day fracture, and you see the Soviet Union collapse. So, no country, no empire is too big to fail, I think that's very important to remember. Now, I'm not suggesting in the book to be clear that either the United States or China is about to collapse, but the reason that the book is called How Progress Ends is that we've had some technological progress for the past 20 years, the internet and the personal computer are very promising technologies that transform the economy and transformed our lives.

But if you go back to the '90s and you think about what those technologies did in terms of connecting the best talent around the world, giving us the world store knowledge in our pocket, streamlining the research process enormously, and if you think about the economic consequences, which essentially a decade-long productivity upstreak, mostly confined to the United States, that's quite extraordinary to me. And what is also extraordinary is how global the productivity slowdown has been. It's been slowing down to varying degree, the United States has still been doing better than Europe, for example, but the global productivity slowdown is very real, and it's happening at the time, but we were very excited about exponential technologies like computers which have been empowered by more slow and now more recently AI.

I thought it was fascinating to see in the book, taking back to the Cold War era, even the early phases of it, of is the Soviet system going to be more dominant here or is the American system? And it was interesting how you unpacked that. I had forgotten how many American economists, like Paul Samuelson said, "I think the Soviet economy's going to outperform over the coming decades," and then they were proven wrong. One of the central theses I think in your book is to say that every system is going to have its limits. Even if it's performing well, it's going to have to adapt at some point to new conditions, new parameters, and therefore, if it doesn't, I think you went as far as saying it will fail.

So, that sounds like a constantly adaptive system. Are we in one of those moments now, with, I think a lot of people are arguing now with AI that it's so disruptive as a cognitive technology, a general purpose technology that's cutting across everything, that we're again in one of those major moments of adoption and adaptation is really going to matter here to sink or swim. What do you think of that thesis? And maybe you can talk more about your broader thesis about that all these models have to adopt regardless of how well they do at a certain point.

Yeah. So, key theme of the book is that the institutions need to adjust as technology changes. And so, as we mentioned earlier, the Soviet Union was doing very well for some time, at least on paper economically, of course, a lot of that growth focused on heavy industry, it didn't really benefit consumers, it didn't really improve the standard of living for Soviet citizens the way that you might expect a more open decentralized economy. But nonetheless, growth was rapid for around four decades, the Soviet Union grew at around 6% per year, and then it slowed down and eventually collapsed. And so, the book tries to grapple with why is that? And I think the reason that the Soviet Union grew at more than respectable rates for a long period of time is not that the Soviet Union was a overly dynamic economy, but it was one that was set on catching up to the United States.

And so, it took ample advantage of the Ford Motor Company's open door policies, sent several delegations to Detroit to study what was going on in the Ford factories, drew in talent from the United States, both talent that was committed to the communist course, but also engineers that were there for financial capitalist reasons, if you like, and they contributed to setting up the Soviet automobile industry as one example. And during the age of mass production, the Soviet system worked reasonably well in the sense that Soviet bureaucrats could hold factory managers accountable by benchmarking performance across factories. And that works well when a technology is mature and established, but when something is new, when you're dealing with novelty, what are you benchmarking against?

And so, as you're dealing with novelty and as the capacity for monitoring diminishes, it also invites cronies. And perhaps more importantly, what you need at the technological frontier is space to experiment and having people pursuing different technological trajectories. So, if you were an aircraft engineer in the Soviet Union, for example, you could go to the Red Army and ask it for funding. If they declined, maybe you had two or three other options. If they declined, your idea would die with you. And that's quite different from the American system of decentralized finance, where Bessemer Venture famously declined to invest in Google back in 1999, they probably regret it today. But it also illustrates that Google wasn't a safe bet at the time. AltaVista and Yahoo were dominating such. And so, to know if something will catch on, somebody needs to take the risk and invest in it.

And the beautiful thing about the decentralized economy is that people can take very different bets and pursue different technological trajectories. And so, what I argue in the book is that kind of decentralization is really important at the technological frontier, whereas scaling is really important to harvest the benefits of those technologies. A key question when it comes to AI then is scaling what we need or is [inaudible 00:15:04] experimentation what we need, or perhaps even some combination of both? I think if scaling was all that we needed, consolidation would be the way forward because if you consolidate data and computer resources, that's conducive to scaling up AI. I think we are beginning to see diminishing returns to that approach. I think the progress that we've seen over the past couple of years has come from trying out new ideas, and we can discuss those.

And more importantly, I think the way forward will probably require having AI that approaches the compute and data efficiency of a human that can learn from a reasonably small set of examples and adjust to a changing world in real time. That I think will probably require another breakthrough or two, and what that means is that that will in turn favor competition and decentralization, both in terms of having more players trying different thing, and both in terms of giving people within organizations that experiment with the use cases of these technologies more decision-making autonomy.

So, with those elements in mind and the breakthroughs necessary or on the horizon, would you hypothesize that China's surveillance state will create better AI by producing more data or will it facilitate a worse kind of innovation because of more narrowed experimentation and dissent?

So, if the world was just a static distribution of events, then the task would just be collect all the data, scale up data resources and computes, and that will eventually take you there. And if that was the way forward, then China will probably have some advantage in terms of data collection capabilities. But I don't think, as we discussed earlier, that is really a feasible way forward because the world is changing all the time. And I think if we look at the Chinese system historically, what has elevated it close to the technological frontier and even at the technological frontier in several technologies like solar and EVs is actually decentralized competition, but with Chinese characteristics.

And so, a key difference between China and the Soviet Union is that the Soviet Union was essentially ran out of Moscow. Every key industry from the railroads to steel were managed centrally. In the Chinese system, on the other hand, provincial governors and city mayors have much greater autonomy. And so, what China has created is essentially decentralized political system within the one party state, where you have members of the CP competing for promotion inside the party by meeting growth targets. And so, that has created fairly extensive laboratory for both technological progress, but also institutional change. And so, when China began to open up in '78, different provinces took different approaches. Beijing continued to pursue more top-down industrial policy, whereas the South set-up special economic zones and integrated into global markets.

And Deng Xiaoping saw the results of that firsthand during a Southern tour in 1992, and decided that we should permit that kind of experimentation within China more broadly. And so, it is that decentralized competition often also facilitated by local government has driven China to the technological frontier, not a centralized Chinese surveillance state along the lines of the Soviet Union. But I think there are open questions about whether that decentralized competitive approach can be sustained, I think there's a tendency towards more industrial policy being centralized in Beijing, and if you look at artificial intelligence in particular, progress there is overwhelmingly come from private initiatives and often entirely new firms, like DeepSeek. And I think one advantage that Chinese firms have is that as long as they serve national objectives, it's easier for them to enter and scale, because in China rules and regulations are enforced selectively, and that's both the strength and the weakness.

And so, whereas new firms may not face the same compliance costs that American, or for that matter, European firms face in setting up and scaling, the larger they become, obviously the more mission-critical they become, and the more they have to invest in political capital to have a seat at the table when political priorities change, which it tends to do from time to time. And so, American advantage has been of having a system of secure private property rights which protects firms from the state. China doesn't have that. But what America has and Europe certainly more of, is protective regulation for incumbents. And those are two different parts of the rule of law that functions quite differently, one part potentially favoring China, one part of it most certainly favoring the West.

It's interesting to think about these companies entering the ecosystem, vying for position, et cetera. When you look at the electric vehicle market right now, it's very interesting to see what's going on in China versus the US. In China, I understand there's more than 100 players in the EV or plugin hybrid market, and maybe as many as 20 significant sized ones. In the US, you've got a very small fraction of both. I don't know how many total players there are, but there's a very small number of major players, like the order of four or five and not many others. And so, somehow that's evolved I think to your point, that when the direction is clear and the technology is established, the state-centric model can really push forward quickly. Of course, not all those companies are going to survive.

But I wanted to go back to your point about Google coming out of nowhere, very quickly from nothing to very dominant. Same thing happened with Amazon. And now in the AI space, we've got OpenAI, Anthropic, that have come up very quickly, largely out of nowhere a couple years back. But have we entered a phase where things happen faster now in these new technology scalable spaces, where you can have your moment in the sun, but it's actually only a couple years long, and you're no longer like an oil company that established for a good run and then faded? Anthropic may thrive or it may be gone in a year or two. Are we somehow in a more volatile space, or does it just appear like it is because we're in the middle of it?

I think that it is pretty clear that we're in a period of very intensive competition, and I think it's pretty clear that not all of the companies that are investing hundreds of billions into AI infrastructure are going to find those investments profitable in the end of the day. And so, if we are on this trajectory of just scaling up LLMs, well, first of all, I mean, even if we reach [inaudible 00:23:38] and open source models are just six months behind, or whatever it might be, that is not the viable business model for any of these firms. And even if it turns out that AI has more of a moat than it seems to have at the moment, not all of these firms are going to thrive. And so, I think we will inevitably see some consolidation in the foundation model layer, and then there's another open question, which is whether LLMs themselves are the future of AI or not.

And so, to be clear, I'm not a computer scientist, I don't have special insights to that question, but it's far from clear whether it's large language models, whether it's small language models, whether it's world models, whether it's something else that are going to be the path forward. And right now it seems that American technology companies are very heavily invested in one particular approach. And so, that is another aspect I think worthwhile looking at, and seeing what is actually happening in other domains is going to be no less important than keeping track on the foundational model layer.

You mentioned antitrust earlier, and I wondered how important you thought antitrust has been to at least America's innovation story, not just as consumer price policy, but also as industrial policy towards technological progress.

I think it's been absolutely essential historically to American progress. I think if you go back to the late 19th century, what Mark Twain famously called the Gilded Age, America was a society of small government and large enterprise, and it also has the political system which were essentially [inaudible 00:26:01] mobilized voters around the prospects of gaining office rather than policy issues. And so, that created very fertile grass soil for cronies, and so the Pendleton Act, which created the first meritocratic civil service in the United States was absolutely critical in dealing with that, and then having a government capable of regulating American tech giants, and punishing anti-competitive behavior was absolutely essential as well. But it was also crucial. I think that rather than size per se being the issue, it was an issue of using antitrusts to pursue anti-competitive practices. And so, that meant that firms could continue to scale as long as they serve consumers.

And so, that balance has always been very important, that you can grow big as long as you invest in innovation, as long as you invest in new technology and you serve consumers, but once you overstep and once you engage in cartel building or erecting barriers to entry, then there are tools that can be used to prohibit that. But clearly it's by no means a perfect tool, and it's been enforced selectively historically. And I think that the downward trend in antitrust enforcement over the past couple of decades has been harmful for entry in American innovation.

Policy Prompt is produced by the Center for International Governance Innovation. CIGI is a nonpartisan think tank based in Waterloo, Canada, with an international network of fellows, experts, and contributors. CIGI tackles the governance challenges and opportunities of data and digital technologies, including AI and their impact on the economy, security, democracy, and ultimately our societies. Learn more at CIGIonline.org.

One of the big questions that percolates a lot around right now is the impact of AI on economic productivity. You've weighed in on this topic a few times, and you just hinted to your views about the limits for large language models potentially progressing further. But how would you describe where you stand on how big these productivity gains are likely to be? You compare it to say computers or something else, you could compare it to electricity, or pick an example you want to give, but how big are you seeing this? And let's assume that it includes perhaps some elements beyond just generative AI, like reinforcement learning, like alpha fold for protein folding, and some scientific AI use and things. But what's your core case on where we're heading from an economic productivity point of view?

So, I think that crucially depends on what AI is being used for. So, whether it's used for automation or whether it's used for developing new products and new types of industries. And so, think about it this way, if all we had done since the 1800 was automation, we would have productive agriculture, we would have cheap textiles, but that will be about it, right? We wouldn't have vaccines, antibiotics, rockets, airplanes, computers, et cetera. So, most improvements in material standards comes from doing new and previously inconceivable things. And I think a key reason why a productivity growth was so strong between 1920 and 1970 roughly, was that the second industrial revolution, electricity and the internal combustion engine, created a lot of new products and new industries.

And so, the car industry was the largest industrial undertaking that the world had ever seen, it created a host of industries producing the components that came to making the car, it created industries producing the machine tools to make the components, gave rise to road commerce, gave rise to tourists. And in addition to that, you have electricity which produces a range of new industries standing behind basically every electrical appliance that you have in your home. And so, the spawning of these new industries was really important for that growth. The early days of the Industrial Revolution, which essentially centered on the mechanization of textiles, didn't produce anything like that productivity upsurge. Productivity in Britain really took up with new applications of steam, like the railroads and steam ships and so on from 1830 onwards, and I think when you look at the computer revolution, I think that's probably the closest parallel to artificial intelligence.

It's certainly done a lot to automate work in manufacturing, it's aided offshoring to places like China. It has most certainly created new sectors in jobs, those have been very clustered to a few places, like the Bay Area, but they're not nearly on the scale of what we saw during the second industrial revolution. And so, I think with AI, most applications that I can think of are about process improvements, they're about doing things that we are already doing a little bit faster. I struggle to think of any new products that I've seen that will likely lead to new industry on anything like the magnitude of the car industry. Although it's clearly early days, and let me be clear that it's absolutely feasible that we will have new industries growing out of artificial intelligence. But right now it's mainly about process improvements.

And a key difference to the computer revolution is that the computer revolution essentially in many ways predictable, it was computing, that software that has been programmed and so you would know what the machine would do in any given contingency, and the internet also basically automated a lot of downtime, just sitting around waiting for information to arrive. A book that you might have ordered, or saving you a lengthy trip and a search through the library. AI is automating the production process of knowledge work itself, but it also makes mistakes. And even though hallucinations seem to be going down for various reasons, you still need in most instances human verification. And that verification is costly, and that verification is also becoming harder the better AI gets, because the greater expertise then needed to actually spot the mistakes.

And so, if you think about the productivity gains from AI, it's the time saved in the production process then minus the cost of the verification. And with the computer evolution, you didn't have that minus the cost of verification in that way. And so, I think that the AI revolution will match up most closely to the computer revolution. I do suspect that its productivity gains are likely to be a little bit more muted, as long as we don't achieve something like superhuman AI that doesn't make mistakes.

Interesting. And the cost of verification point I think is worth underlining, as you say that was not the case for computers. I mean, there was a bit of verification, but certainly nothing like what we're talking about here. Imagine a case where you've got 15 quickly generated reports or assessments or whatever they are, who actually verifies those, or at least does a peer review type exercise to say these are good or not? It becomes a bit of a bottleneck there, where the cost of that verification goes up significantly, whereas the generation of the base material is very, very low cost. So, you can't imagine a bottleneck forming there. Let me take the side of some other economists though, just to probe a little bit further, who would be pointing to the US growth in labor productivity, which you do acknowledge in the book. And I guess pointing to AI as the driver of that, not necessarily total factor productivity, but labor productivity.

Is that simply, in your mind, due to this massive scale of investment that is huge in the US, in the AI, and the labor mobility, which is very impressive in the US market that is getting some benefits here, if it's a little bit superficial? Because a lot of economists are pointing to those numbers as starting to show positive impacts from AI.

Yeah, great question. So, first of all, let me just take a minute to credit Christian Catalini, co-author, who has written an excellent paper on the point of verification.

Yeah.

So, I wrote an op-ed comparing that to the computer revolution, but research behind that is not [inaudible 00:35:46]. On the point of labor productivity, so I think it's an open question to what degree that has to do with AI. I think that is pretty clear that if you look at TFP growth and adjusted measurements of that, taking into account working machines harder and so on, we haven't really seen much of an upsurge productivity over the past quarter. And so, the much hope for productivity boom is yet to be identified in the productivity statistics. That is not to suggest that there will not be a productivity upsurge driven by AI. So, I do think that there will be a productivity upsurge, but I think how significant will it be? Even more importantly, how sustained will it be? Because it matters if we see a couple of years uptake in productivity or if it's a decade or several decades, that makes a huge difference.

And so, that is still an open question. As I said earlier, my best bet is that it's going to be roughly on par with the computer revolution, probably a little bit less. I think it may be a little bit quicker to materialize just because the rollout of the technology has been a little bit quicker, but probably not a little bit quicker is the restructuring that many firms will have to go through. And so, AI is two things at the same time, it's this consumer good that anybody can use for everything from solving problems in the household, to having questions about mental health, to work related things. And so, that's on a lot of people's phones and a lot of people probably will use it for work without their employer even knowing about it.

But I think the big gains will come when firms begin to reorganize their own processes around the technologies. And so, with electrification, took around four decades for engineers to figure out that rather than replacing the steam engine as the central power source of the factory with an electric motor, you could actually sequence the machinery and have a natural flow of production, and give every machine its own electric motor, which gave rise to mass production. So, that process took much longer, and that process, even if it's shorter than with electricity, it's still going to be in the decades rather than the years.

You've been clear that the book isn't a prediction, it's a framework for thinking about what happens next and how we're going to get there. What are some of the most common mistakes you've seen people, and by people of course, probably government, state-led policy make when they try to predict technological direction or progress or even adoption?

I may have some responsibility for this as well. So, we wrote a paper back in 2013, looking at the potential scope of automation. And we argued back then that roughly 47% of American jobs are exposed to automation technology, and so that was broader than AI. It was about AI, it was about mobile robotics like driverless cars, it was about [inaudible 00:39:27] in industrial robotics. And we also tried to consider the potential for task simplification in that paper. So, a key bottleneck to automation is that really when you automate something, the technology does exactly the same thing that the human does. And so, the electric washing machine doesn't do the same thing as a laundress, right? We didn't automate that by building a robot that would walk down to a well and then perform the motions of hand washing, and then walk back to the house, and then hang the clothes to dry.

We invented the electric washing machine. And so, we tried to, very imperfectly, consider that. And so, the virtue of that, the approach was that at least we were clear about it's about automation, it's not just about exposure. We were considering probability that a job is fully automatable. Since then we've seen a number of studies with some improvements, I should say, but often less clarity about what exposure in their vocabulary actually means. Exposure augmentation or is it automation? And I think there's a tendency out there to look at these studies and say the exposure is X, Y, and therefore we need to worry about this particular set of jobs, without really considering whether exposure is automation and augmentation, the first button.

The second relevant question is obviously, well, some jobs will be displaced, but what are the new roles that are likely to emerge? And that's just much harder. And so, we're more likely to focus on the automation of existing things because maybe we can say something about that. Whereas, my great-grandmother would never have anticipated that her great-grandson might be a hot yoga instructor or something like that. And as a result of that, there's been much more focus on the potential for scope for automation rather than looking at A, actual adoption, and then B, what is that technology that I'm actually being used for? And so, there is an emerging literature that's increasingly interested in adoption, but I can use my computer to do very different things. If I've adopted a computer, it doesn't even say that much.

And so, I think what we do need is much better measurements for what technology is being used for, and also as well the type of new activities that are emerging in real time as a result of technological change.

Last question before we just open it up to you to see if there's anything you wanted to say but didn't get a chance to do so. There's a lot of talk about middle powers, this whole middle powers concept, and Sweden of course is one of them, and Canada is one of them, but so is Thailand, and Argentina, and South Africa... So, almost everyone is a middle power in relation to the United States and China anyway. And so, what's their realistic goal and ambition here in these fast-moving technologies? And probably not even just limited to artificial intelligence, you could put quantum computing in there and perhaps satellite technology and things in there. Are they going to be able to keep up at this technology frontier, or are they really facing a sovereignty challenge here by the power of these technologies and not really having sovereign capacity?

Or have we seen that in the past in history, where technology's moved across countries, they obviously started in a few places and then diffused, are we seeing a similar pattern now? Is it somehow different because algorithms can scale immediately as opposed to hardware? What's your view on how much of an opportunity there is here for countries that don't have a lot of these cutting-edge technologies to get in the game, or is there a high risk from them being captured or left behind?

I think if you look back from the beginning of the post-war period, most of the growth around the world has been the consequence of the diffusion of American technology, including in Europe. And so, in the United States you begin to see a productivity upsurge in the 1920s that continues to be quite strong, even during the Great Depression, you then see a bit of a collapse during the Second World War, and then you see a resurgence after World War II, which lasts up to the 1970s. But the later part of that, the post-war miracle is not an American story. And so, all of Europe essentially shares it to a very good degree, and it's in large part also a result of a transfer of American technology, the American mass production system also through Marshall Aid. And all of that was fine as long as the relationship between Europe and the United States was intact.

But that's also not all of the story. Because what's also surprising in the case of Europe is that it's not so much that it's not leading innovation in digital, but it's also even failed to catch up. It hasn't closed the gap in digital the way it's closed the gap in mass production. And so, you sort of have to ask yourself why that is, and what we should learn from that. Now, you're obviously talking about middle powers and that ranges from Canada to India and Sweden, and so what these countries should do is probably not necessarily the same thing. But if we stick with Europe for a minute, basically those who think that Europe should do more targeted industrial policy in artificial intelligence, saying that we need something like an Airbus for AI, and Airbus is clearly a very good example of successful industrial policy.

But at the same time, Europe also had a lot of initiatives to catch up in semiconductors. So, France had an initiative, England, Germany, those didn't work out very well. And so, what's the difference? I think a key difference is that aircraft technology was fairly static at the point when Airbus was established. So, the jet engine had already been developed almost two decades earlier. So, Airbus was basically catching up to a clear target being Boeing. Semiconductors are much more dynamic technology. So, you're catching up to the previous generation. Well, you're already behind when they caught up because the technology's moving so fast, and AI is even faster moving like that. And so, I think some of the middle powers, if you look at Europe as a whole, as a middle power, I'm not sure if that's a right way of looking at Europe, but we have to ask ourselves, what are the preconditions that are holding Europe back?

And I think part of that is a lack of a harmonized single market for services and digital, and that has been extensively discussed in the Dräger report and other places. Part of it is that experimentation in Europe is more costly. And so, hiring and firing is more costly, Meta can just pivot from the metaverse to AI, that's much harder for firms like Nokia or SAP and so on to do. And then the sort of regulatory burden of entry in Germany or really anywhere in Europe is just higher because of rules and regulations like the GDPR. And then on top of that, continental Europe has a more bank-based system, which is better at funding physical assets rather than intangibles and more closely tied to incumbents rather than promising startups.

And so, I think those things together makes it harder for Europe to reach the frontier. And so, Europe needs to think about all of those things to get its act together, and other middle powers should arguably be part of that as well, because you need scale in order to be able to do this. And so, middle powers need to work together in that world. And so, trying to establish those preconditions I think will be important. And embracing, frankly, open source and/or open weight models when it comes to AI, focus more on the application layer, investing in research and alternative approaches to AI, which are less capital and compute intensive than large language models, I think all of those things are a part of the story of what middle powers should do more broadly.

Thanks very much for the conversation.

There is so much that's in flux as we consider what the future of progress is going to look like and feel like, and how you strike that balance between, again, optimizing the enabling conditions for technological progress versus maybe some of what Canadians are feeling right now, because very recently our AI strategy was released. Thinking about the role of the state in facilitating technological adoption is different from the role of the state in facilitating progress itself. But I know we had a bajillion other questions we wanted to ask him, was there anywhere where you thought we were going to go that we didn't or a question you were dying to ask that we just didn't get to because he's so amazing?

Well, I think we didn't get into it all was many of the examples of the technologies over the ages, and the book really does go through a whole set.

Yeah.

And that, I think for people that are interested in that, they'll find it really quite rich, about just all these examples of technologies that we've gone through. But I had a couple of big takeaways in addition to what you were just saying. The first one relates to that idea that at some point, a general purpose technology, some technology that's really important, reaches a, let's call it a plateau. He called it a static level I think once or twice, where it's got a critical mass, let's say, of power, impact, disruption, potential for change. And then, people go and countries focus on it and they really start to adopt it and engage, and then you get all these new ideas and these spillovers and things.

And he was basically saying we haven't hit that with AI yet, which is one of the reasons why Europe can't do its Airbus equivalent, of saying, okay, this is what we're going to do on AI. Because the moment you start that, it's like, well, we better do this. And you saw that through even the EU AI Act, they had to keep revising it because things were changing. So, even on the governance and regulatory side, it's really tricky. But on the industrial policy side, it's kind of like, when do you grab onto this? And Canada's going to face the same thing. And then the other major takeaway was this idea that the production cost of the raw material, let's say, is almost a zero marginal cost of just producing reports on the value of locating a hospital here versus there or on any topic.

You can imagine you can produce dozens of these reports that have facts and citations and are generated by AI, but who actually reviews those and verifies those becomes a very scarce commodity here, and increasingly scarce as we're kind of flooded by these AI generated things. So, we're going to see a premium on verification. He mentioned Christian Catalini and others who wrote this paper called Some Simple Economics of AGI, which is an excellent paper, and it gets to the heart of this question of verification becoming the next thing that's really going to matter here. And we're very focused on that at CIGI as well.

Well, I'm looking forward to seeing more related work from CIGI and also just starting to extrapolate some of the lessons from this book as we look to Canada's future technological progress.

Talk to you soon.

Policy Prompt is produced by me, Vass Bednar, and CIGI's Paul Samson. Our supervising producer is Tim Lewis with technical production by Henry Daemen and Luke McKee. Show notes are prepared by Lynn Schellenberg, social media engagement by Isabel Neufeld, brand design and episode artwork by Abhilasha Dewan and Sami Chouhdary, with creative direction from Som Tsoi. The original theme music is by Josh Snethlage.

Please subscribe and rate Policy Prompt wherever you listen to podcasts, and stay tuned for future episodes.