How Should We Think About the Economics of AI?

A conversation with Erik Brynjolfsson

So, a break from political and policy madness. Along with everything else going on, there’s what looks like a revolutionary technology spreading through our society, which everyone calls AI, even though it doesn’t (yet?) look like the kind of artificial intelligence everyone expected.

However, in thinking about this revolution, and the economic role of technology in general, we have a problem: economists in general don’t know enough about the technology to make an informed judgment (and technologists aren’t great at economics, but that’s for another day.) So I wanted to talk about what’s happening with one of the few economists who has really invested in understanding information technology: Stanford’s Erik Brynjolfsson. Here’s our conversation; transcript follows:

TRANSCRIPT: Paul Krugman in conversation with Erik Brynjolfsson (3/20/25)

Paul Krugman

Hi everyone. This is my latest video conversation. And now for something, again, quite different. I wanted to steer away mostly from the current political policy environment and towards what is certainly one of the biggest stories, which is AI, but just generally technology. So I'm recording a conversation with Erik Brynjolfsson, who is one of the few economists who has really followed technology as opposed to just sort of plugging it into their models, which is what I do.

If you're wondering, I’m actually in a hotel room in Belgium. I'm committing the cardinal sin of a video interview with a bed in the background because it's a pretty small hotel room but here we are. Hi Erik.

Erik Brynjolfsson

Hi Paul, good to see you.

Paul Krugman

I've been meaning to get you for a conversation. There were some hitches there, but if it weren't for all the other things happening in the world, technology would be one of the biggest stories out there. Certainly very interesting things are happening. As I said, you're an economist who actually knows something about technology, which for lack thereof hasn't stopped some of our colleagues from having strong views. But why don't you tell us a bit about your background and how you came to be that guy?

Erik Brynjolfsson

Sure. Well, Paul, you probably don't remember this, but when I was an undergrad, you came and gave a guest lecture at Harvard. That's when I first got to know you. And at the time, I was kind of having one foot in technology and one foot in economics, trying to decide between the two. They asked me to teach a course at Harvard Extension School. So along with my college roommate Tod Loofbourrow, we taught a course on building expert systems and one on applications of artificial intelligence.

We started a company called Foundation Technologies in homage to your and my favorite science fiction author, Isaac Asimov.

Paul Krugman

Oh my gosh, I didn't know that. Yeah, okay.

Erik Brynjolfsson

So we were building these systems back then. And then I decided I really needed a graduate degree to get in a little deeper. And so I decided to go to MIT where I thought I could maybe do both AI and economics at the same time, but it didn't really work out and I had to choose. And I went more the economics route, but continuing to be really interested in technology and what I was doing. And I was lucky enough to have Bob Solow kind of mentor me. And you may remember back in the eighties, there was this puzzle of, as Bob Solow put it, we see the computer age everywhere except in the productivity statistics and I called that “the productivity paradox.” So we were wondering what all this technology was doing. And so my first real big paper was trying to address that puzzle about technology and productivity.

Paul Krugman

Okay, great. By the way, not everyone may know, but the foundation novels by Isaac Asimov involve mathematical social scientists who save galactic civilization. And when I was growing up, I wanted to be one of those guys and this is as close as I managed to get. And the productivity paradox is something we all talk about. Let's just talk for a second. Economists generally, starting with the work of Bob Solow—the great Robert Solow, sort of the founder of growth theory in economics—have always given a really crucial role to technological progress.

Erik Brynjolfsson

Yeah, I think what we all learned in grad school was that Bob Solow back in the 1950s, what he really got his Nobel Prize for was showing that, you know, we're not necessarily working more hours or even having that much more capital. That's not what's driving our higher living standards. It's mainly that we have much better technology than people did in the past. So technology is the main driver of living standards. And he just kind of stuck it in the model as this extra variable and off we went with better and better technology. Since then, a lot of people have tried to flesh that out a little bit in more detail. One of the things that are connected to AI is this class of technologies we call General Purpose Technologies. Economists just used to call them GPTs, but the AI people stole that acronym from us. But General Purpose Technologies are things like the steam engine, electricity, and I think AI. And that's actually what drives most of the progress in our living standards. So that's kind of behind my focus on that set of technologies.

Paul Krugman

You say, Bob Solow pointed out—I'm not sure when he first made that remark, 1987?

Erik Brynjolfsson

1987 in a little article in the New York Review of Books, yeah.

Paul Krugman

Not many young people realize how revolutionary IT seemed to be. When I was in graduate school, using the computer meant handing a box of punch cards to the high priests behind the glass wall at the mainframe and then waiting an hour for a printout to come back and you made one wrong key punch and you got this whole stack of hexadecimal. Anyway, you know, and then to the coming of desktop computing which is really...

Erik Brynjolfsson

Time magazine made it the machine of the year. I think it was 1982 because we were all being revolutionized by it.

Paul Krugman

Yeah. In terms of my own work, what it meant to actually do the work for a paper, it actually was revolutionary. Suddenly you could do all of your statistical work without having to go to the computer center at three in the morning.

But it wasn't really showing in the productivity numbers. What's your take on what happened to the productivity paradox? Because I have a version, and let me see if yours matches mine.

Erik Brynjolfsson

So, well, it went away in the 90s when we actually had roaring productivity. Productivity more than doubled to about 3% per year, which economists consider a pretty good rate. And part of that was the internet. Part of it was, I think, an even bigger part that's not as much recognized as large enterprise systems, these ERP systems from companies like SAP and supply chain management, customer relationship management systems, that started really changing the way business was being done.

And what’s really key is that these general purpose technologies rarely have a big effect all by themselves right away. It's usually after they start transforming the way business is done and companies rethink what they're doing. And that business process change, that re-skilling, that reorganization often takes some time before it translates into productivity.

Paul Krugman

Yeah, there's a classic paper, probably you cite as well, a Paul David paper on the dynamo of the computer, how it took decades before businesses figured out what to do with electricity. You have this innovation, and there was electricity everywhere, but it really wasn't doing anything until businesses figured it out.

Erik Brynjolfsson

Right. In the 1880s. He showed that companies started installing electric motors, but according to the records he went through, they didn't really have big productivity gains until about 30, 40 years later. The big change—to oversimplify a story—was that at first they just took out the steam engine, put an electric motor where the steam engine used to be, and nothing much else really changed. It took about a generation of managers to realize, wait a minute, electric motors can be light, small, medium, large, all different sizes, and you can lay them out by the flow of materials instead of having this one big engine in the middle of the factory. That new layout, which allowed for assembly lines and the flow of materials, that led to, according to Paul David, 100%, triple digit gains in productivity. But it really wasn't the electricity by itself, it was the combination of electricity plus new business processes.

Paul Krugman

Yeah. The six-story tall mill building with cramped corridors and overhead crankshafts and so on gets replaced by the sprawling modern factory with wide aisles. Around 1997, 1998, we were all saying, “Look, Paul David was right.” All it took was time and maturity, and we got this big surge in productivity. At least according to the numbers. I certainly thought we were going to see a much more sustained rise in productivity than we did.

Erik Brynjolfsson

Yeah, no, it then started fading around like 2005 or so. It dropped back down and since then it's been back down around 1.4% productivity growth per year, like maybe half of what it used to be. So that's been very disappointing and in a way it's a second productivity paradox. Along with Chad Severson and Daniel Rock, I wrote a paper called “The AI Productivity Paradox,” basically replaying some of those hypotheses and trying to understand why this new wave of technology

isn't translating into productivity gains.

Paul Krugman

Yeah, it's basically a decade of good productivity, you know, boring productivity growth that seems to be associated with IT, but it's like ‘95 to 2005 or thereabouts. It's really 20 years now that we kind of haven't seen, at least in the numbers, haven't seen those kinds of gains. And the iPhone came out during that period. Broadband internet basically comes out during that period. It doesn't seem to move the numbers.

Erik Brynjolfsson

Yeah, well, you said a couple of times, “at least in the numbers.” And so I think you're setting me up there to say something about measurement, too.

Paul Krugman

That's a really big issue.

Erik Brynjolfsson

Yeah, that's a big issue. Right now we're looking at screens. The average American spends about eight and a half hours a day looking at screens of various sizes. They're voting with their time that digital is a bigger part of their consumption than the time they spend on consuming atoms and services. So we're kind of transitioning to an economy where that's really important.

And GDP is an amazing invention, one of the great inventions of the 20th century, Paul Samuelson said. But one thing it doesn't do well is, it doesn't measure things that have zero price, with few exceptions. And so we're missing a lot of the digital revolution just by design. Zoom or Wikipedia or search, most Facebook value for end users, they aren't counted in the productivity statistics. There will be, of course, some electricity use or there may be some advertisers who benefit, but the consumer benefit is largely missing from our statistics. And as the economy becomes more and more based on digital goods and services, we are missing a bigger and bigger part. You mentioned the iPhone. Most of the stuff you consume on the iPhone, or a lot of it at least, you're not paying for.

And that means that it's largely invisible in the GDP statistics. And I should mention, just a slight technical note, productivity is just defined generally as GDP divided by hours worked. Sometimes you put other things in the denominator as well. But if you mismeasure GDP, you're going to mismeasure productivity just as much.

Paul Krugman

I can go back and forth on this, and I have, I guess. There's clearly a lot of unmeasured gains.

Watching videos of cute animals on YouTube or whatever is clearly a source of joy for some people. Watching actual musical performances is a source of joy for me. That's not going to show up in GDP at all.

Erik Brynjolfsson

Right. We're all listening to more and better music than we ever have in history, but that doesn't show up.

Paul Krugman

But the question always up for debate is, hasn't that been true one way or another for the enormous improvements in public health in the first half of the 20th century?

Erik Brynjolfsson

Definitely. Chad Stevenson wrote a good paper kind of arguing that yes, we've got mismeasurement, but also we used to have mismeasurement too. So you really need to kind of compare one mismeasurement versus another mismeasurement, which is hard enough just doing one of them, but doing two of them is pretty tricky. So that's one of the reasons that Chad and I wrote a paper together along with Daniel Rock to try to reconcile this. I think I've come around to the view that mismeasurement isn't the whole story or maybe not even the biggest part of the story. I do think that this reorganization is a really big part of the story, the reinventing of the economy.

The kind of reinvention that's going to be needed for the digital age is at least as big as what we saw in the industrial age. That said, I would love for us to do a better job on the measurement side as well. My team at Stanford, we've created a new measure of GDP, or of welfare I should say, called GDP-B which measures the benefits. It seeks to measure sort of the consumer surplus that's created by all these goods. So even if you pay zero for it as a consumer, you may still get significant benefit. And if you look at that side of the ledger, then the digital revolution is creating trillions of dollars of consumer welfare, even though we're not paying for it. So we are addressing that measurement side of it. One article about it just came out in the American Economic Journal of Macroeconomics this past month. But this new framework GDP-B, I'm hoping will be increasingly used alongside traditional GDP. As the economy becomes more digital, we'll measure not just what we're spending, but also the benefits we're getting.

Paul Krugman

Random thought here. A lot of these benefits, the unmeasured benefits, are kind of the consumer experience at some level. I'm sitting in Europe right now. The Europeans have been very upset that U.S. productivity has pulled way ahead of Europe over the past 25 years or so, even if a lot of the gains come from the consumer experience. People are absolutely as internet addicted, absolutely spending as much time on their mobile phones as Americans are.

Erik Brynjolfsson

So in a different paper, we did an international comparison. We picked 12 digital goods. At first I was surprised, but then it kind of made sense. What we found was that richer countries had a proportionally smaller share of their total welfare from digital goods than poorer countries or middle income countries. So, Mexico, Brazil, they actually had a bigger share of their total welfare from digital goods. And when I thought about it, it does kind of make sense. If you're poor, you're not going to be spending a lot of money on caviar, Rolls Royces, or even televisions. You're going to spend more of it on stuff that's free. You can consume as much free as anyone else, as long as you have a smartphone, which most people do. So it ends up being a proportionally bigger share of your welfare.

Paul Krugman

Yeah, actually if you think about that for individuals, Bill Gates, or whoever, gets probably no more or less pleasure out of stuff that's free than I do.

Erik Brynjolfsson

Yeah, exactly. Or then a working class person. And that's what we found. We looked at it both within zip codes within the United States, and we looked at it across countries. And digital goods tended to reduce inequality somewhat, just because they're available to almost everyone. We didn't have data from the very poorest countries. So I'm not sure it goes all the way down. I assume it doesn't if you don't even have a smartphone. But once you're at that basic threshold of being plugged in, then it tends to be a net equalizer.

Paul Krugman

Scott Bessent, the Treasury Secretary, just over a few days ago said, “The American dream is not having a flat-screen TV.” But you're saying in a way, kind of it is, or that a lot of the benefits really are a lot of economic growth.

Erik Brynjolfsson

Ha ha ha. Well, yeah, and that's why I'd love to get this GDP-B measure kind of systematized and have it start coming out every quarter. And then we could know for sure what's happening. It's entirely possible that as we switch around our consumption towards more digital goods, the numbers will look a lot different across inequality, productivity. Growth will look different if you're looking at the welfare contributions rather than just where we're spending money.

Paul Krugman

Okay, so you said, though, that you don't think the measurement issue is important, but that's not the key, or that there are bigger things explaining this?

Erik Brynjolfsson

Yeah, it's hard to know because nobody really did a good job of measuring the consumer surplus from, I don't know, the smallpox vaccines or whatever and things that were obviously incredibly important a long time ago. (And today.) So that comparison is a little tricky to make.

Paul Krugman

Yeah.

Erik Brynjolfsson

I spend a lot of time talking to companies and, actually going back to the 1980s and early 90s when I was a grad student, even then I would spend some time talking to senior executives. And I could see the business transformation that was going on then and now and how important that is. And in a follow up paper that Chad and Daniel and I wrote called The Productivity J Curve, we calibrated this and found sort of a J-shaped effect of this investment.

Paul Krugman

So an initial drop rather than a rise.

Erik Brynjolfsson

Exactly, because what happens is, at first you have to reskill your workforce, put a lot of time into training and people on their own have to figure things out. You build new business processes, even new products and services, new organizations. Like what we were talking about that Paul David found with the electricity. All of that is creating what I would call “an intangible asset,” but that's another thing that's not measured in our economy, a lot of these intangible assets. So you see a lot of effort but no measurable output. In other words, productivity goes down. More input, no output. Later, you start harvesting those gains. Now you're like, “Okay, we've got these things set up. Our workers are re-skilled. We've got new business processes. We've invented some new products.” And now you're harvesting that stuff. And now productivity is actually unusually high because again, you're not counting the intangibles. So at first it's down and then it's up. So it's a J-shaped curve.

Paul Krugman

That makes a lot of sense, that basically if you spend three, four years reconfiguring how you do your business, it's going to require spending a lot of effort. When we measure, we mainly count capital goods, production of capital goods as part of GDP but we don't count production of intangibles like organizational improvement as part of GDP.

Erik Brynjolfsson

That's right. And more and more of those are what's important. I was visiting Dell and, if they want to double the output of producing computers or anything, you know, they could build a second factory next to the first one. That would get counted. But if you put a lot of intangible effort into creating new business processes and re-skilling—so, get twice as much output out of the first factory—that would not be counted as an additional asset, even though from an economist perspective, it kind of has created an asset that you're getting a stream of value from.

So, we’re mismeasuring the digital goods that we're consuming, we're mismeasuring a lot of the intangible assets. And again, that adds up to trillions of dollars according to our best estimates. So, as the economy becomes more and more digital, more and more intangible, we're kind of missing more and more of what really matters.

Paul Krugman

You get all of this great stuff that seems to be happening in technology after the mid-90s that doesn't seem to be showing up in the numbers. And one answer would be that businesses are really spending a lot of effort on adapting. But then...

Stupid indicator, but if you go back and look at Dilbert cartoons from the early 1990s, the office work looks pretty much the same as it did until a couple of years ago. Was there really a lot of organizational change over that period?

Erik Brynjolfsson (23:22)

I think there was. A lot of it's under the hood, you know? Obviously the internet changed a lot, but it takes time. You know, when Amazon started doing online sales in the mid 1990s, it wasn't like the next day we all went to shop digitally. It's only just now, 25 - 30 years later that we're beginning to have digital commerce surpass traditional commerce in lots of areas, but you're also seeing the cloud infrastructure. And then most importantly, just recently, AI and agent-based systems are beginning to automate and augment a lot more of what's going on. But part of my mission is to shorten that. I've written a number of papers now to describe the kind of business process transformation that's needed, and I even started a company called WorkHelix that helps companies implement some of those changes.

Paul Krugman

So yes, you're trying to not just measure, but actually sort of make it happen.

Erik Brynjolfsson

I think so. You and I both know Bob Gordon. He's one of the greatest economists who's been studying growth and he was kind of a mentor of mine early on and helped me many times. And we agree on so many things, but one thing we disagree on is what's happening right now and in the near future to productivity growth. We made a bet. I bet that productivity growth was going to significantly exceed the Congressional Budget Office projections for the 2020’s.

He took the pessimistic side of that bet. We'll find out in a few years. But I'm trying to also, you know, actually improve my odds of winning by spending more and more time talking to managers and showing them how they can implement these technologies in ways that are more productive. And I'm pretty hopeful. Already we’re beginning to see some glimmers where companies that have installed AI have begun to have very measurable gains in call centers, software coding, sales… various internal large language models where they're beginning to see some big gains from it. As that spreads throughout the economy, I think we're going to have an upside surprise on productivity.

Paul Krugman

Obviously we hope you're right. So, let's talk about AI. I have colleagues who sort of bristle at even calling it AI and want to just say “large language models” (LLMs) or whatever. How would you describe what it is. Clearly something has changed. Clearly we have systems that do stuff that we weren't able to do before.

Erik Brynjolfsson

Yeah. Well, one of the biggest changes is the machine learning revolution. So back when I was doing AI, we called it that. I was building rule-based systems, expert systems, where I would sit down and interview experts and they'd give a bunch of if-then rules. “If someone's coughing, they might have a cold,” or something like that. We do it for banking and for other things. But you had to hand-code everything. Machine learning is different. With machine learning, you give the machine a bunch of inputs and a bunch of data on outputs. And then the machine learns. It figures out the relationship between the inputs and outputs. To make this work, you need lots and lots and lots of data. And the good news is that since the 1990s, we've digitized so much of the economy. We have a ton of data, text data in particular, but also images and other things. And we feed those into the machines, and they start finding patterns on their own. And they use neural nets, deep neural networks, these layers of neural networks that are sort of loosely inspired by the human brain. And they're beginning to really be very good, getting to human level or superhuman in more and more categories like image recognition.

And we've all played with, you know, Chat GPT and the other tools and they can write pretty decent essays. Maybe not the quality of your writing, but the quality of a lot of people's writing. And they're getting better literally month by month. And basically, under the hood, they're using these neural networks. The first wave was more trying to diagnose and pattern match, make predictions and classification. The new ones, the generative models, are actually creating new content, either text or images or even videos and other kinds of content now.

But what's different now versus before is a lot more data, like orders of magnitude more data. A lot more computing power, going beyond Moore's law and switching to GPUs like the Nvidia cells that allow them to do more faster, better processing than they could before, and some improvement in algorithms, those three things.

Paul Krugman

For viewers, Moore's Law is the—what was it?

Erik Brynjolfsson

Originally it was doubling every 18 months, then you kind of modify it doubling every two years, but it's sort of an exponential improvement in computer power that's been going on for about 50 years.

Paul Krugman

A single chip—if it doubles every two years, that's a factor of 1,000 in 20 years. It was amazingly stable.

Erik Brynjolfsson

Exactly. I mean, they call it a law but it's more like a mystery because there's no law of physics that says this has to happen. But they've just been able to keep shrinking the size of the transistors and the circuitry. And with that, you get faster and faster and more and more. It's beginning to get harder just to turn the crank on that dimension. But what they are doing is improving on other dimensions. To do large language models and other kinds of neural nets, you need to do a lot of matrix multiplication. And these GPUs happen to be very good at matrix multiplication, so they were able to piggyback on that, even though the underlying circuitry wasn't getting that much smaller. It was just a better design for these kinds of problems.

Paul Krugman

Much of this is as mysterious to me as it is to, I think, a lot of people, but I actually do know what matrix multiplication is. You have two big arrays of numbers that have to be kind of like matching sizes, and there's a process for multiplying them to produce a third array. This is something like PageRank, which is the original way that search engines were able to figure out what's important. And you're saying that a lot of this stuff basically involves some version of that.

Erik Brynjolfsson (30:04)

Yeah. A nice quirk of history was that these GPUs (Graphic Processing Units) were first used for rendering images on screens. You have to do a lot of matrix multiplication for that. And then some people like Andrew Ng figured out, wait a minute, we can use these for neural nets as well. We just change the code a little bit. And lo and behold, you were able to tap into orders of magnitude more processing power. And now they're making them special just for neural nets, but initially they were kind of repurposing them.

Paul Krugman

Okay, and so we have these redesigned chips with something like Moore's Law so data processing is still getting cheaper in general, but we also have specialized data processing in a form that is good for this. But should we be calling this intelligence? I mean, in some ways that horse has left the barn. We are calling it that.

Erik Brynjolfsson

You know, that's a good question. It's a little semantic. I don't think it's like a sharp line. So, I moved to Stanford a few years ago. I love Cambridge and MIT and all the intellectual horsepower there. But Silicon Valley is just a different world. Initially, I was in the economics group. And now I moved over to the computer science building, the Gates Building. And next to me is Chris Manning and a lot of the other giants of computer science. So I went and asked him,

“Can I call this intelligence? Can I call this understanding?” And he thinks it's okay to do that. These systems are basically answering questions the way an intelligent person would, sometimes a little better, sometimes a little worse. And I guess there's a philosophical debate about whether that's true understanding. But when it comes to solving problems and actually figuring out what the next step in this scientific process is, or understanding who committed the murder in this mystery story, or solving a little puzzle, or writing an essay that's kind of coherent about the Crimean War, they can do that. So I'm OK calling it intelligence.

Paul Krugman

Yeah, I mean, I have friends who say, this is just sort of super powered autocorrect. I'm on both sides of all of these debates, but my response to that is an awful lot of jobs—some erstwhile high skilled, high paid jobs—are basically like super powered autocorrect.

Erik Brynjolfsson

Absolutely. That's another good point. I had a talk with Demis Hassabis who came to visit Stanford a while back. He's the founder of Google DeepMind, got the Nobel Prize a few months ago, and he's been working on this basically all his life since he was a teenager. And he said that when he found out how well LLMs (Large Language Models) were working, he was actually a little disappointed because he thought matching human intelligence would take more than that. And he was like, “Maybe a lot of us, what we do is super powered auto completion, and that you can get a lot done with that.”

You know, from an economist perspective, that's kind of the point I focus on. If you want to look at how the economy is going to change, what we're doing, what exists right now is already enough to affect a big percentage, maybe half of all cognitive jobs. And what we are going to see in the next few years is going to only increase that ratio. So, we're already there in terms of having an economic revolution. I often use the term “transformative AI” just to emphasize: let's look at what's happening to business transformation, economic transformation, and leave it to philosophers and others to debate whether it's truly artificial general intelligence or something else. I don't think there's any doubt that it's economically transformative.

Paul Krugman

Okay, yeah. I mean, when did people stop saying, “Learn to code to make a living. Learn to code.” And now it turns out that that's one of the things that these models are really quite good at doing, right?

Erik Brynjolfsson

Hahaha. Yes, it turned out kind of by accident that they found that they were incredibly good at coding. The way they trained them was that they gathered all this data from the internet and the web, and they just scraped it all in there. And it turns out that there's a fair amount of code up there on the internet. And so these systems learn to complete code and understand. In order to do that effectively, you have to kind of understand the logic of code. And these neural nets were finding patterns there and were able to get very good at coding. It's one of the things they're best at. And so that's kind of a canary in the coal mine, our coding jobs.

And there's another interesting twist, which is that, what are these systems made out of? They're made out of code. So now we're getting, you know, the recursive self-improvement that people talked about in science fiction, 50 or more years ago. And that can lead to something like the singularity where you get these systems that are improving themselves, and then they're able to improve themselves at a faster and faster rate. I don't know if we'll have that kind of very rapid takeoff, but if you talk to the people in the big AI companies, they are using LLMs and coding tools a great deal now to generate a lot of the code that they use to make the next generation of large language models.

Paul Krugman

Oh boy. And then it kills all of us. No, I mean, there's a cynical side which says that the LLMs do have a tendency to fantasize a little bit. But give me your take on it.

Erik Brynjolfsson

Yeah, they hallucinate.

Paul Krugman

Search these programs for yourself and you discover that there are all these interesting papers attributed to me that neither I nor anybody else wrote. And there's the choking on its own garbage scenario, where as more and more of the data out there is generated by these models, it starts to just descend into a loop of self-replicating garbage.

Erik Brynjolfsson

Right. If most of the content on the web starts being generated by systems, can they learn from themselves? This is kind of a little bit above my pay grade but I talk to the people creating these systems, many of them are actually surprisingly optimistic that you can use synthetic data to train the models, that you can create additional chess games, additional Go games, and even additional essays. And coding actually is one that you can do and use to train models to get even better. You can have it write puzzles and then solve the puzzles and it improves itself. So they are making some progress now. We'll have to see how far you can go with synthetic data, but that could be a possible way of extending the runway as we're close to basically having fed all of the data on the World Wide Web into these models. So you need to have some other source of data or improvement. And that's what they're working on. Synthetic data.

Paul Krugman

Okay. Although it's all digital and it's just bytes, it turns out there's an awful lot of BTUs that are going into this. The energy consumption has turned out to be kind of alarmingly high. Is there a way past that?

Erik Brynjolfsson

Yeah. I think it's legitimate to be concerned about that. So one of the simplest ways to make these systems better—simple but not easy, I guess—is you just make them bigger. And every time you make them bigger, there's like a predictable improvement. People call these scaling laws. And so one of the reasons these companies are asking to build bigger and bigger data centers (OpenAI said they want to build something called Stargate that's $500 billion worth of data centers) is that, according to the scaling laws, if you just make them that much bigger, you're going to get that much better.

But of course, giant data centers require giant amounts of electricity. So they are re-starting up nuclear plants. Three Mile Island is getting started up again. They're building new small nuclear reactors. They're trying to get more and more solar. They're firing up coal and oil plants that otherwise they probably wouldn't have. So that has negative effects on the environment. Some people like Eric Schmidt have said, “Well, look. Yes, that's costly. But if we want to solve our environmental challenges, we need more intelligence. In the long run, it's going to be better to have this additional intelligence to come up with better solar panels and better processes for all of our industrial processes. And ultimately, the way to address that is to have more intelligence.”

The other thing, especially since DeepSeek, the Chinese model, came out a few months ago, is people are understanding that you can actually get an incredible amount of power from much, much smaller models, an incredible amount of intelligence.

And so that's a margin that more people are focusing on: can we use better algorithms to get more intelligence instead of just turning the crank and making them bigger and bigger? I don't know for sure how that's going to turn out. I don’t think anybody does. They're working on both dimensions and many other dimensions, really like six different dimensions where you can improve these models. So I think we all hope that we'll be able to get more intelligence without using lots more electricity.

Paul Krugman (39:58)

Okay. DeepSeek is sort of third hand, but doesn't use all the data in the world.

Erik Brynjolfsson

Yes. It's much, much smaller or is of a magnitude smaller. And interestingly, you know, it was not like one big breakthrough or “a ha.” It was a lot of small engineering tweaks. They got 3% improvement by doing this and 8% by that and 1% by that. Nvidia, who we mentioned earlier, has a whole stack of CUDA, this software that runs on Nvidia chips that most people use. They kind of bypassed all of that infrastructure and went down to the machine code level, which is

much harder. It’s much more painful to do that kind of coding. But if you do it, you can get huge efficiency gains. So this is the kind of hard work that they did to crank out more efficiency. And, now everybody sees that there's an opportunity to have AI systems that are much, much smaller, use much less electricity and still perform basically just as well.

Paul Krugman

That would be a huge relief if we managed to do that because the energy burden is kind of scary right now.

Erik Brynjolfsson

Yeah. Well, as an economist, I don't know the answer, but we're trying to think about what's the marginal value of intelligence? If we push the frontier of having a bigger model and get a few more IQ points or the equivalent, how valuable is that? Maybe the average store owner doesn't need, like, Einstein to come in and rearrange their shelves, but they need a pretty smart system to do that. And when I think about what's going on in the economy, I think not every business could benefit from having a bunch of Einsteins get off the bus and start telling them what they could do. But they could use a lot of sort of like reasonably smart people to help them out. Maybe over time that will change. There's also the elasticity of intelligence, that as people realize that really high levels of intelligence are available at very low cost, they'll reorganize in ways that take advantage of that.

We're in for some interesting times in the next few years as the price of intelligence falls dramatically.

Paul Krugman

I was about to say, as economists, we don't know anything about business, but you actually kind of got a foot in that world as well. Still, one big question we have here is, obviously, we've got something that is really, really impressive and novel and has startled a lot of people with how well it works. There are corporate valuations that are kind of based on how amazing the technology is. And one thing I thought we learned from the internet boom was that amazing technology doesn't necessarily mean amazing profits.

Erik Brynjolfsson

That's right. Well, let me underscore both parts of that. I do think that these technologies are really amazing and transformative, and they're probably going to create trillions of dollars worth of value. But much of that value, if we have a well-functioning market, will end up in the hands of consumers, not producers. Bill Nordhaus estimated that about 95% of the gains from innovations ultimately go to consumers, not to the people who create it. So there's no guarantee that any of these guys who invent the best new system are going to be able to cash in on it fully.

I think the internet's not a terrible analogy. You know, we saw lots of gains from that. But there was Pets.com which was a total joke, and there was Amazon that did pretty well, and there are lots of companies in between. I think this wave is going to be significantly bigger, both in terms of the consumer benefits and also in terms of the winners and losers. And lots of smart people are making bets that they're the company that's going to win. I think if you really do believe that the technology is as transformative as I do and as a lot of other people do, it's not implausible that some of these companies could be worth a trillion dollars or more, but probably not all of them. So, how to figure out which ones are on the winning side and which ones aren't? Well, we'll leave it to the market and others to sort that out over the next few years.

Paul Krugman

Yeah, I mean, back during the internet boom, the dot com and all of that, we used to say that investors seem to be betting on 10 or 15 different companies, that each one of them was going to be the next Microsoft and acquire this sort of spot, and they couldn't all do it.

Erik Brynjolfsson

Yeah. They couldn't all do it. At the same time, I do remember sitting around in the faculty lounge. I was visiting Stanford Business School, and Amazon was this little company competing with Barnes and Noble. Some of my faculty colleagues were saying, obviously, Barnes and Noble is going to crush Amazon now that they see that this is available and I was debating that. Sometimes the new companies do figure out how to capture all of the market. And then, it wasn't just books. Obviously, Amazon captured a lot more than that.

So yeah, we're at a similar stage right now. Even as we're speaking, the stock market is crashing for a lot of the—well, I wouldn't say crashing, but going down quite a bit for some of the companies that had the biggest bubble valuations a few months ago.

Paul Krugman

People were saying in the 90s that each company was going to be the next Microsoft, but then a lot of the companies that have gotten these huge valuations already are. They already have dominant market positions. So in some sense, it is almost like defensive investment. They're safeguarding their position, which doesn't seem to say that the valuation should go up.

Erik Brynjolfsson

Yeah, well, the folks at Microsoft have told me, and I don't think I'm disclosing anything that proprietary but, you know, if AI turns out to be everything it's cracked up to be, they'll be very happy they made the investment. If it turns out not to, maybe they're going to lose tens of billions of dollars. But that's something that they could afford if they had to. But they can't afford to be scooped. So when they look at the two sides of it, they feel like they need to be ready to make this investment.

Paul Krugman

Right. But that's an argument for them doing the right thing, not an argument that people bidding up the price of their stock are doing the right thing. Actually, with Nvidia, they've really emerged as a major player in a way that they weren’t before.

Erik Brynjolfsson

Yeah, Nvidia provides the chips, the GPUs that power most of these systems. And people build on top of those. And CUDA, the software that powers them, is what most people are using. And they continue to have a lead in that. People are trying to come up with other architectures that can work around that. Google has their own chips. They call them TPUs, Tensor Processing Units. But right now, Nvidia is benefiting from everybody trying to get a hold of those chips so that they can build bigger and bigger and better and better systems.

Paul Krugman

Yeah, the people who made money in the California gold rush weren't the gold miners, they were the people selling picks and shovels and blue jeans to the gold miners. And Nvidia is kind of in that position.

Erik Brynjolfsson

Yeah, that's where Nvidia is. And at some point, all that money that's being paid to Nvidia has to come from somewhere. And so the companies that are making those investments are betting that consumers and customers will be paying some multiple of that to them in order to justify the investments that they're making.

Paul Krugman

So how are you feeling about your bet with Bob Gordon?

Erik Brynjolfsson

I'm feeling pretty good. I'd say right now we're close to even. Maybe I'm a little bit ahead. But I've always thought this is sort of the second half of the decade thing. So I'm glad we survived the pandemic and everything else in terms of being remarkably stable. We're more or less at the pre-trend or even a little bit above the pre-trend prediction for productivity. So the past few years, we've navigated it pretty well when it comes to some major disruptions. And now what I see when I visit companies is that they roll out these systems for coding we were talking about, for call centers, for lots of other things, and they're getting 20, 30, 50% productivity gains in those particular applications. As that fuses through the economy, we're going to see an extra few tenths of a percent or half a percent or even a full percent of productivity growth. And that's more than enough to get me to win that bet.

Paul Krugman

It’s been hard with the pandemic and all of the stuff that followed. There's an awful lot of noise and dust in the air.

Erik Brynjolfsson

Just looking at the technology side of it, you know, there's a lot of other things. There's a lot of geopolitics, there's some bizarre macroeconomic things going on, and all of that, of course, could throw a wrench into things. But if you just look at the base prediction, I'm pretty optimistic about productivity and growth.

Paul Krugman

In the mid-90s, there were a lot of people out there who were sort of looking around and saying, “We think that we're going to have this productivity boom out of IT, and we already see it happening.” And I looked at the hard data and said, “I don't see it. I'm really skeptical.” I was totally wrong. History leads me to think that you're probably right. I certainly hope that you are.

Erik Brynjolfsson

Yeah, I'm more optimistic this time. Back then, you know, after I wrote that paper, the “Productivity Paradox of IT,” and then I gathered some data from 500 companies about what was happening, I saw some gains and I wrote a paper called “Paradox Lost” that basically showed that it was beginning to turn around. That was in 1995. And sure enough, even as that has spread, I think AI is bigger than any of the technologies we've seen before. As Andy McAfee and I wrote about in The Second Machine Age, you know, now we're not just augmenting our muscles, but our brains. And that's a really big event in history. I was going to say “human history,” but just history in general, to be able to do that.

Paul Krugman

We're almost obsolete already.

Erik Brynjolfsson

Well, for certain things. And so then I think we have a real challenge that economists need to get at, and that's what my focus is these days: how do we reinvent the economy for a world where machines can do not only most of the physical labor, but more and more of the cognitive work, and how do we figure out a system for not just having prosperity, but for having shared prosperity? Most people get most of their income from labor right now, and if machines are able to do more and more of that, we're going to have to think about other ways of distributing the benefits in ways that are widely shared. I think the default will be that benefits get more and more concentrated and both wealth and power get in the hands of fewer and fewer people. And if we don't want that to happen, we have to be proactive about kind of reinventing our system and our economy in a way that we have the benefits widely shared.

Paul Krugman

I think that's probably a good place to stop. Let's hope that the kind of wisdom that you're hoping for actually materializes. I'll take that bet. I'll take the under on that.

Erik Brynjolfsson

Okay, well we'll do what we can.

Hopefully it's not just a matter of watching it happen but working to make the good outcome happen. And thanks for giving me the chance to chat with you and I hope your listeners and all of us work towards that better future.

Paul Krugman

Thanks, this has been fun.

Erik Brynjolfsson

My pleasure.