In 2024, you made a very important paper that basically talked about how AI was not being so impactful on the economy. Things have changed.
Daron Acemoglu, Nobel Prize-winning economist.
Daron Acemoglu, Institute Professor at the Massachusetts Institute of Technology.
His work has reshaped how we understand the role of institutions, inequality, and technology in economic prosperity.
One of the world's leading voices on how institutions and technology shape our future.
Now you're coming to the crux of the matter.
Okay. So when we talk of a productivity increase, there are two very different ways of doing it. One is let's keep GDP constant and we have less employment. That's an increase in productivity. The other one is we make workers more productive and keep employment the same or not much lower, and as a result GDP increases. Now we have made ourselves much richer, but even better if the workers are more productive in a creative, innovative way. We get new goods and services, new organizations, new things, better things.
Which one do you think is more likely to happen?
Oh, right now I think definitely less employment.
Bill Gates was proposing reducing the amount of work by paying the same amount.
Well, yeah. No, that's not going to work because if you've done that, you've actually made labor more expensive. So, what are you going to do to their incentive to automate? No, that's not going to work. That's a brainless, stupid suggestion.
Obviously, you do need global regulation for a global technology. And you know what's preventing that? Many things, but one of them is actually Silicon Valley itself. You know, if China did not exist, I say Silicon Valley would have had to invent it because they use China as an argument against any attempt to even bring the smallest amount of regulation.
And as a result, definitely worthwhile productivity gains, but nothing like what industry insiders or some techno-optimists were predicting.
For me, the hardest part to understand is because I think my productivity on a personal level has multiplied by three or by four.
Important, that's important, but I think my claim is that at the moment you are the exception. There aren't that many companies who are getting really amazing return. And by the way, you're not unique. I think there are several companies, some in Europe, some many in US, in China, which are doing exactly that, but that doesn't change things too much right now.
One of the most unfortunate things: this is a really transformative technology, indeed. We all agree on that, and we've left this technology in the hands of five or ten people. It can't really be optimal.
But what do you think is going to happen?
The alternative, I think, is very dystopian.
Professor, thank you very much for making time for us. It's an absolute pleasure. I've been following your work for years, and being able to talk to you directly, I think it's such an honor and opportunity. So thank you.
Pleasure is mine. Thank you.
Thank you. So, in 2024 you made a very important paper that basically talked about how AI was not being so impactful in the economy. Do you still think the same, or things have changed?
Things have changed, but some of the messages from that paper are still valid. But I'll give you my take, okay?
Things have evolved in the ways that I wasn't expecting and in some of the ways that I was expecting. So the bottom line of that paper is that it's hard to get really big productivity benefits from automation, especially when coupled with the fact that spread of new technologies is always slow, especially when they need to be integrated into complex organizations.
That's why I was predicting relatively slow diffusion and slow expansion of the tasks that are going to be automated with AI, and as a result, definitely worthwhile productivity gains, but nothing like what industry insiders or some techno-optimists were predicting.
Yeah, I think the paper talked about like a 1% increase in GDP worldwide.
In 10 years, within 10 years.
Which is like almost nothing.
Nothing like... it's sustainable.
No, well, I mean, you know, it is. Not many things give you a one and a half percent increase in GDP. So like tax policy, we talk about it all the time.
You need to improve taxes like tremendously to get one-tenth of that. So it's not like completely irrelevant, but obviously not like the productivity miracle that people were thinking about.
Yeah. Because at the time, I remember I started with AI in 2023, and then we heard messages from like Dario Amodei, from Satya Nadella, saying like this will be like 10% per year growth on GDP and like getting in a stage of like absolutely...
That's right. And my paper was a reaction to that because I thought that was just hype.
And that proved to... I mean, we did not have a 10%.
At least for the last three years, we haven't had that.
Now, where things have moved faster than I was expecting is with agentic AI. So my prediction in 2023, when that paper was written, was that agentic AI would develop very slowly, and that definitely seemed to be the case until 2025. And then with Claude, I think that was a real convex, accelerated introduction of agentic AI, where I think my predictions have more or less held up so far, but may not in the next two, three years because of this agentic AI.
The spread of AI has been slow. Many people are using ChatGPT or Claude or Gemini, but it has not transformed any occupation other than perhaps in coding, and as a result, productivity improvements have also been very small. Employment effects have been very, very tiny. So those are all up to 2026.
Now in 2026 we're seeing agentic AI improvements. The reason why things may accelerate is that if agentic AI is used to develop easier-to-use applications, so that even small and medium-sized businesses start adopting them, that would lead to an acceleration of AI, and both its negative effects on employment and some of its positive effects could come faster. It's still a question mark. I think there's still a lot of uncertainty, but agentic AI is certainly something that did not feature into that kind of calculation.
I think one of the things that impressed most for every one of us that has been into the field is that in the last six months, we have seen more advancement than in the last four years.
That's exactly, that's the agentic AI part.
It kind of responds to an exponential advancement of the technology. But shouldn't it be expected that the next six months are going to be again like...
My take is that the way that the technology... so if you say super-exponential, it creates the impression that it's predictable and steady, whereas the way I would interpret it is that it's sporadic and episodic.
So not much improvement in 2023, 2024, 2025 until the agentic models, and then suddenly a big burst.
And then it will kind of stagnate, and then something else.
I think we're going to have the spread of those agentic models. However, that's going to happen, and then the next breakthrough may come in one year, two years, three years, 10 years, 20 years. I think there's still some uncertainty there.
But you know, I am definitely not in the camp that says, oh, agentic AI as it is now is already AGI, artificial general intelligence, which industry insiders don't have a very coherent definition. But I guess the most common definition is AI models performing pretty much all human tasks at a level comparable to the very, very best experts.
Like doing math and physics and code being at the level of the top 1%.
We're definitely not there for one very simple reason. These models don't have judgment. These models don't have social interaction skills. They don't have context. And worse, they have zero capability to do anything physical.
But even if we had an AGI, that would not mean... and that's something that it was hard for me to understand, but I see it very clearly now, that would not mean that immediately we'll have the economic impact, because there is the bottleneck of humans itself.
I think if you take the definition of AGI seriously, it doesn't mean immediately, but it would have a very quick effect, because if it really can do all human tasks at the level of the best humans, you don't need the humans anymore.
Yeah. But there is this paper from Anthropic where they say what AI is really doing and what it can do. And there is a big difference between what it's doing in certain jobs or tasks versus what it could do in the jobs. Like in finance, there was something like it's doing 20% when it could be doing 75%. And the difference is not the capability of the model. It seems the technology is not the problem, but more like the utility of the technology.
I disagree with that. I disagree with that. So I think their claim that models can do 75% is very exaggerated, because the models don't have context. So humans understand context right now. You couldn't put AI doing 75% of finance without creating a huge mess, and humans would have to clean it up, and that's not the way to integrate AI.
So I don't find it surprising that they're not letting AI do more than 10% or 15% of tasks in finance. It would be irresponsible, dangerous, and probably counterproductive.
Again, AGI, I think properly defined, means the reliability of the models and their ability to understand context, like what are your needs? What are your risk attitudes? How is the world changing? For example, is the fact that we learn from data about US Treasuries that were very safe and now it's not safe. If we can immediately understand, the models can understand those as well, and they can also present the risk profile in a comprehensible way to humans of different abilities and different backgrounds.
If all of these can be done, then they should be able to do 100% of tasks in finance. And once that level is reached, sure, there could still be some human resistance, but it's going to be very quick to overcome that. But if you have that definition of AGI, which I think is the correct definition of AGI according to what they initially defined it, then we're very far from AGI.
So you still have doubts about the technology, not just about the impact, but about the technology itself. Do you use AI?
I do. I do. No, I mean, what do you mean by doubts? I think if the meaning of doubt is I don't trust anything from AI, no, absolutely not. I trust many things with AI. By the way, this was true even before generative AI. AI didn't start with generative AI. There were many things that you were doing with AI before generative AI.
So I think both Claude and ChatGPT, and I haven't used Gemini that much, but from what the others who have used Gemini say, they are quite reliable for a range of tasks at the moment, but you still have to verify and check certain things that they do.
Because for me the hardest part to understand is because I live AI myself, and my team and my companies, and we have multiplied our capacities, our productivity. It's insane. Like if we put it to my own person as a CEO of my companies, my group of companies, we have different... we have a visual production company, we have on top of that a real estate company, we have different companies in different areas.
But for myself, like things that used to take me much more time, especially researching and these kind of things, take me much less time, which frees me hours to do other things. So I think my productivity on a personal level has multiplied by three or by four. I can do three or four more things than I could do before.
And then if I check on my employees, for example, we just added like an AI that makes drafts of emails for customer service, and that was set up in 10 minutes and it manages to save two hours a day for some of my customer service employees. And the same thing in different areas, like for example a podcast like this one, we use an AI tool called Autopod, and it all of a sudden can create that each camera points at the person that is talking. So basically it saves us maybe like three, four hours of an editor.
And by the way, you're not unique. I think there are several companies, some in Europe, some many in US, in China, which are doing exactly that, that they have invested a lot in AI and they're doing things more efficiently, and many of us are saving one hour or two hours here, but that doesn't change things too much.
Like for example, I use AI for background research and literature review. Sometimes I use AI for asking for criticisms of pieces I write. I would have done the same things, probably slower, and at the end perhaps I would have watched one less game in the World Cup.
But I'm not sure that my overall output would have been tremendously different.
Yeah. That's the point then, because if we can do more, obviously we're not going to have an impact on the economy if we just have free time with this more that we do.
Coming to the now, you're coming to the crux of the matter.
So when we talk of a productivity increase, there are two very different ways of doing it. One is let's keep GDP constant and we have less employment. That's an increase in productivity, but it hasn't made us richer. GDP is the same, gross domestic product is the same. The other one is we make workers more productive and keep employment the same or not much lower, and as a result GDP increases. Now we have made ourselves much richer, but even better if the workers are more productive in the creative, innovative way, we get new goods and services, new organizations, new things, better things.
So I think that's the real promise of AI, and that's bundled in AGI. If AGI really arrives, yes, models are going to be more creative than humans as well, by definition, on average at least. But I think because AGI is very far, there is a real meaningful distinction between automation versus what I call pro-worker AI, making workers more productive in their task by giving them greater capabilities to do new things, new expertise, new tasks. I think that's a very important distinction.
Yeah, because so people can understand it, it's like you have two ways to take this, and you can just do the same you're doing and give more free time to people, and therefore...
Or unemployment, exactly. But therefore, like you get the result, or you can try to increase the result holding up the amount of work and the amount of employment.
Which one do you think is more likely to happen?
Oh, right now I think definitely less employment, because the AGI agenda is very much intertwined with automation, and that's why Dario Amodei, for example, is continuously talking about how this is going to destroy the entry-level job market.
I think again that's too accelerationist. I don't think the timelines are going to be as fast as what he's saying. But on the other hand, he knows the technology better than I. But you see the emphasis both from Amodei and others when they talk frankly that it's really a massive automation technology, and I think that is a problem.
Frankly, because look at Spain. I mean, Spain has already a huge youth unemployment problem. What happens if you stop hiring? You're going to have 75% youth unemployment.
It's not what we... it's not like what's going to happen. It's more like what's happening, because Dario said this famous thing about the next five years, 50% of your junior jobs will be gone. But then I saw a report from Adzuna, which is a job portal from the UK, that said that since ChatGPT came out in the market, they detected a decline of 37% on postings for junior jobs. So it seems it's already happening.
I always say that we see on the news these massive layoffs from Microsoft, from Meta, from Amazon, but the real thing that I think is happening is more like companies getting some employees that are a bit more productive and therefore not hiring new people.
You may be right. With the same amount of people, they can do more. You know, I emphasize, and I should emphasize, I always try to, including in the paper that you started with, everything about AI is uncertain. And by that I don't mean the future of AI is uncertain. What has happened is uncertain too.
So there are some researchers as well as some data sources which point to exactly what you're summarizing, which is that there has already been a decline in entry-level hiring in AI-exposed occupations. But then there are other people who claim, oh, there has been more hiring in AI-related areas. Both could be true for a while. Neither could be true.
I think the problem is that many of the trends that show slower hiring started before ChatGPT. So the timing doesn't quite square with a causal effect. And there may also be, some people have pointed this out, that there may be AI washing, that the tech sector expanded too much during COVID and in its aftermath, and now they need to shed some of those workers, and they are blaming it on AI. Possible. So that's why I mean, like, the past is uncertain.
Yeah. It doesn't help that if you blame it on AI your stock raises, and if you blame it on other things your stock goes down.
This is something really curious. Like before, if you lay off people, your stock was going down. It was a sign of your company not doing well, and now laying off people it means you're doing it right because you're automating it.
You got it. You got it right. Exactly.
So probably many companies are kind of relaying on that message to get the narrative on their side.
And that's something that we will see probably in the next couple of years, I assume.
Well, yeah. I mean, my guess is that uncertainty about the future will continue, but perhaps we get a better understanding of what has happened over the last few years, both in terms of the technology's capabilities, because while there is no doubt in my mind that the technology has really, really accelerated with the agentic AI, we don't even yet fully understand the capabilities of the technology. There's still hype there as well.
Yeah, absolutely. Talking about hype, Demis Hassabis, the CEO of DeepMind, he said, I think it was in Davos, he said that AI is going to be like the Industrial Revolution, just 10 times bigger and 10 times faster. Do you agree with that?
Well, I don't know what the metric is for 10 times bigger than the Industrial Revolution.
I assume it's more like it's affecting 10 times more sectors, because the Industrial Revolution was kind of like buried.
Then he's definitely right.
That's also likely, but you know, then there are many other aspects of the Industrial Revolution that I think, at least, I would be very surprised if AI is 10 times the Industrial Revolution. And there are some other aspects of the Industrial Revolution, and if AI is 10 times as fast, then we'll have a revolution, no democracy will be standing.
Exactly. So first of all, the Industrial Revolution was a real break with history. So there was no period of sustained economic growth, at least since Greek city-states, you know, about 200 years before the Industrial Revolution. So the Industrial Revolution completely changed the relationship between technology, science, knowledge and how what we do.
So another discontinuous change like that, perhaps, but I see AI much more continuous with digital technologies. I see generative AI more continuous with good old-fashioned AI.
But Demis Hassabis and Dario, don't they think generative AI is a game-changer? So there's some disagreement there.
Also, today we are amazingly fortunate that we live in a world that was created by the Industrial Revolution, where we have much greater levels of health, wealth, output, comfort. But the first three generations during and after the beginning of the Industrial Revolution suffered mightily. Life expectancy in British industrial cities fell to 30 years at birth. Can you imagine 30 years? Today, in the least healthy part of Africa, life expectancy at birth is about 60 years.
Real wages for weavers, the main sector affected by the new machinery, fell to one-third of what it was. So if we have 10 times that, what is that going to do?
Exactly. That's the point. These comparisons with the Industrial Revolution are very visceral, but they're also very imprecise. So that's why I don't know. But yes, certainly the Industrial Revolution, at least at first, affected a handful of sectors: textiles, coal, you know, machines like heavy machines like a steam engine, later spreading very slowly to other sectors. AI is impacting many more sectors.
Yeah. As far as I understand what the Industrial Revolution was, it was just explained in very plain terms, and probably it's more complicated, but it just made strength not be valuable anymore. So basically, before the Industrial Revolution, if you were able to lift up 120 kilos, you were employed and you had a very good salary. And if you could lift up 150, you were better paid, because that was a value that you had as a...
I mean, I think I would say it certainly made certain aspects of strength perhaps less valuable, but strength remained absolutely crucial. Physical dexterity became even more crucial after the Industrial Revolution.
Nowadays no one gets paid for lifting up weight.
Well, very basic thing. Well, I mean, you know, we have a great need for construction workers and movers.
They lift a lot of heavy...
They are supported by machines.
It's basically like I have the feeling that AI is doing the same thing with intelligence. It's not that we're not going to use intelligence in the future, but it's going to be less valuable and less specific.
It's also like different things, different aspects. So again, let's think of physical labor. What the Industrial Revolution did, which wasn't so much in demand, it increased the demand for children's labor, because the coal mines became very valuable and most of it was underground. So you needed little children with their little hands to go down into the dangerous shafts. So it made certain types of physical labor more valuable.
So I would expect that the same thing is going to happen. Certain kinds of intelligence probably will become more valuable. Certain kinds of intelligence may become less valuable. I don't actually see any evidence of that.
Even if we cannot compare what we have in front of us to what we had in the past, because obviously it would be a very basic mistake, it's the only reference we have. And what we have seen is in every single technological revolution there's been a phase where early adopters take benefit of it.
And then there is a phase where there's a transition where the world is changing, which is what you were saying about the Industrial Revolution. If I don't recall wrong, there was about 30 years where people dropped their salaries to try to compete with machines, and then they realized...
Or more, longer than 30 years. And also, look, I think perhaps, you know, people in this industry obviously don't see it that way, but a better comparison is the internet.
The internet was completely revolutionary.
And the internet affected every sector. But did we get huge productivity improvements from the internet? Not really. I mean, we got a sizable one for a couple of years. And then we have lots of new industries and things, but some of it has been badly developed, like social media. Some of it has been well developed, and lots of companies that moved first and overinvested in the internet went bust.
Mhm. So I think there are different parallels you can draw, and of course Demis Hassabis and Dario Amodei and Sam Altman think the internet was nothing compared to ChatGPT or generative AI.
I'm not sure I agree with that. I think the internet is completely revolutionary. Without the internet, we couldn't have any of the things that we've had over the last 20 years. Nothing that we have nowadays would work without the internet.
So, and that drives me to one of the biggest comparisons that we get when we talk about AI, and it's about the internet bubble, like the dot-com bubble.
That's why I was referring to.
What are the differences from what we're seeing today? Because first of all, I think people tend to confuse the stock markets with the reality of technology. I mean, one thing doesn't have to do with the other. And the proof of it is that if we travel back to the year 2000, we would never tell any company to not invest in the internet. We would say, like, yeah, obviously don't invest in pets.com, but doing nothing to invest in Amazon.
Exactly. It's a big mistake to relate the economic value of startups and everything that's related to the technology at the bottom, which was absolutely revolutionary and changed the world as it is, like AI is probably doing now. But on the economic part, like the value of Nvidia, the value of Palantir, all these things, there is something we can compare.
But as far as I know, and please correct me if I'm wrong, they told me that it's a bit different, it's not exactly the same. Like in the dot-com, we were investing and there was no spending, and now basically the companies, they're spending between themselves, but it's different in some...
I think there are so many differences and some similarities. On the negative side, I think the internet investments were much more durable. So the overinvestments that happened during the internet bubble were useful because later on they provided the optical cable network infrastructure.
Second, I think from the beginning, everybody understood what the internet was. If you look at those companies like Pets.com, Amazon, MySpace, some of them succeeded, some of them didn't. They had the right business model. They were trying to use the internet as a new platform for commerce and offering goods and services. It was clear from the beginning that's how the internet was going to be successful. The porn industry was the first one that understood that. That's why they invested a lot in the internet.
The things that, you know, previously you couldn't offer those things to the consumers. Now they could. The fraud industry also understood that you could do things like internet fraud.
Right away. With generative AI, it's still a little harder. How are you going to monetize? It's not clear, you know, given the investments that Google, xAI, Meta, OpenAI, Anthropic are making...
Absolutely insane. Insane.
Those companies need to generate something like a trillion dollars of revenue every year. How are they going to generate that? There's no business model to generate even a fraction of that.
Well, the business model at the moment is selling dollars for cents. No, because basically, if you pay for the OpenAI $20 thing, they give you like...