Close All Tabs - AI Will Create Enormous Wealth — But Who Benefits? with Erik Brynjolfsson | Dream Machines

Episode Date: August 17, 2026

AI is advancing at a breakneck pace, but how quickly will that actually transform the economy? Dream Machines hosts Alexis Madrigal and Robin Sloan talk to Stanford economist Erik Brynjolfsson about what the data can tell us so far, why previous technological revolutions took decades to reshape everyday life, and whether this time will be different. They discuss early signs that AI may already be affecting young workers, what happens to the career ladder when machines take over junior tasks, and where the biggest challenges — and opportunities — lie. Guest: Erik Brynjolfsson, economist and Director of the Stanford Digital Economy Lab Read the Transcript here⁠ ⁠Watch the Video here⁠ Credits: Producers: Anayansi Diaz-Cortes and Derek Lartaud Sound Design: Brendan Willard Executive Producer: Jen Chien Editor-in-Chief: Ethan Toven-Lindsey Special thanks to Chris Egusa, Annie Fruit, Paul Lancour, Matt Morales, Vivian Morales, Zaldy Serrano, Xtine Tiñoso, Hazel Tesoro and Alex Tran. Support for the production of Dream Machines comes from the Krishnan Shah Family, Dorothy Marsh, and other generous KQED members Learn more about your ad choices. Visit megaphone.fm/adchoices

from KQED. Support for KQED podcasts comes from Star One Credit Union. Give your savings account the love it deserves. When you keep your money with Star 1, you keep more of your money. Star 1 credit union in your best interest. Minions don't wait, especially for Wi-Fi. Summon's same-day Wi-Fi from Xfinity and get the same price for five years. And seeing minions and monsters only in theaters.

Xfinity. Imagine that. Restrictions apply not available in all areas. Learn more at Xfinity.com slash same-day Wi-Fi. Hey, it's Morgan, reminding you that we have this cool new podcast series guesting in our feed, running all month. It's called Dream Machines, and it's a podcast about AI. Here it is. I'm Alexis Madrigal. I'm Robin Sloan.

And this is Dream Machine. It's a podcast about how AI works, also about how it feels. And, of course, we make it right here in San Francisco where it's all happening. Nobody really disagrees at this point that AI is going to transform the economy in some way. But one of the questions that is still really open and hotly debated is how fast that is going to happen. And also the question of like, how would we know that it's being transformed?

You certainly read a lot of headlines and see a lot of stock prices go up, but like what's actually happening out there in offices? We even have the economic data that we need to make sense. Yeah, exactly. It's sort of the question of like, what would be the canary in the coal mine? And as we were looking around for answers to these questions, turns out there's a paper called Canary in the Coal Mine. Out of the Stanford Digital Economy Lab.

Yeah, that's right. And it's the work of a scholar there named Eric Brinjolfson. He was at MIT for 30 years before coming to Stanford very early in his understanding of the potentially transformative effects of AI and has become one of the folks who is really trying to shine a light into the kind of mysterious recess of the economy. so that the rest of us can know, like, what is happening and how fast? Yeah, because it really matters the time scale, right? If we have this incredible transformation in three months or in three years or in three decades,

it will mean tremendously different society. And we need someone who can tell us, like, where are we on this chart? So here is our very own canary in the coal mine, Eric Brunyolfson. There's a lot of, as you know, definitions of AI, you know, superintelligence AGI, you know, just like advanced AI. I'm quite taken by the definition that you used in sort of a transformative AI. Yeah, transformative AI. You laid this out in a research agenda, you know, for kind of the community last year. And there's a bit more to it.

But the core of it is, as you said, we'll know it's transformative AI because, it'll have this significant economic effect. It'll accelerate economic growth from sort of the baseline that we've become accustomed to for a long time now, maybe around 2%. It'll multiply that by three or five times. Almost without talking about AI specifically, it'd be really interesting to have you dramatize what that means for an economy. You know, if that did happen. Transforming of AI. Transforming AI and that growth rate.

Like how would that change our day, our week, our year? our jobs? Well, these ways for me to think of it is by analogy. And so we don't know for sure what's going to happen going forward, but we can look back at, say, the industrial revolution. And in the second machine age, Andy McAfee and I wrote about how we went through this big transition from an agricultural society to an industrial society. And we call it the second machine age because now we're having in the early stages of a second big transition like that. The first one was about machines doing what our muscles could do and animal muscles.

And that took growth from being sort of growing very, very slow, tens of a percent per year, to growing about 2% per year, which may not sound like much, but you compound it, and now we're like 30 or 50 times richer than our ancestors were a couple hundred years ago. I think that now we are able to use machines

to augment our minds, our brains, and I think that's going to be at least as big as what the Industrial Revolution did. and it'll transform society at least as much, not just in growth rate, but also like, you know, just what we do in our daily lives. Think of the way farmers versus people in factories or modern society, how different that is.

Using that analogy, I feel like I should probably know this, but during the industrial revolution, did people's lives change in a decade in terms of the kind of work they did, how they worked in the world, the kind of goods and services they had access to? Was it more of a generational change? Like, what was the pace of that?

It was actually pretty slow. In fact, there's something called Engels paused, this period of like several decades, 40, 50 years where there wasn't much of an improvement in living standards. You can read Charles Dickens, you know, actually life could be pretty miserable, even worse than it was for the people living in agricultural society. Right, like life expectancy declined, people got shorter.

Yeah, all that smoke and soot and kids working in factories and stuff, you know, it was pretty miserable for the majority of the population. But under the surface, things were changing, and you're beginning to have this productivity gain. But it took a while, and eventually it started taking off. This time around, it's also going to take longer than I think some of our friends here in San Francisco and Silicon Valley think,

but certainly a lot faster than the last time around. Let's go back into the historical example. I think it's actually really useful for people to think about this, right? You have this general purpose technology of electricity. General purpose technology, yeah. The term of art. I love it. And also, we were talking about this earlier.

It had not occurred to me that it's the twin GPTs. That seems a little uncanny. Honda, we keep using, yeah, so my friend and former student, Daniel Rock, he wrote a great paper with some folks at Open AI called GPTs are GPTs. Yeah, it's generative pre-trained transformers and general purpose technologies. It used to be when I said GPT to economists, we all knew, we were talking about general purpose technologies, but the AI people have pretty much stolen that acronym for us. Yeah, and for you turn into a global brand. So be it. Yeah, but general purpose technologies is pretty much what drives all economic growth.

There's all these inventions. but there's just a handful that really raise our living standards. The steam engine was the first really powerful general purpose technology, electricity, computers, and now AI, which is kind of like the mother of all general purpose technology, I think so. You know, Demis Hasabas has this, I was over at Google Deep Mind a few weeks ago in London. They had this mission statement. They want to, let me see if I can get it right,

they want to solve intelligence and then use that to solve all the other problems in the world. So, you know, modest little mission statement. But, you know, there's a lot of truth to the fact that if you really could solve intelligence, there's so many other things you could solve. That is, of course, if you trusted that that's what people who had solved intelligence would actually do with this intelligence. Well, that's a very good question. There's a bunch of things you can use it for. And also, to be fair, as I've come to think harder about the problem,

intelligence is not the be-all and end-all. A lot of very smart people think and wish it was. But when you get into the real world, you know, just look around us. There's lots of PhDs in the economy, and they don't, like, rule the economy. You know, you go into a – I was at the Starbucks across the street just before coming here, and you walk in there and, like, suppose – Seven PhDs just right in there. Yeah, imagine. Imagine. You pull up with a bus and you say, okay, good news, you know, we've got Einstein. We've got a bus of 200 Einstein. They're going to come help you.

I think the manager over there'd be like, oh, well, like, do they know how to push a broom? I mean, what are they going to do? I love that. Because that kind of argument, just even that playful vision gets into the real country, details of like what happens in an economy. It's not just, you know, sugar and eggs, whip them together and voila, you get growth and productivity. It's really crunchy. And the thing I like best about your work is that it's kind of ongoing attempt to dig in

and sort of look for signals. You had a paper called Canary in a coal mine, which feels like exactly what we need most right now. We need those kind of early signals. You know, they might be warning signals in some cases. They might be like hopeful bekins. we're going to steer toward. The Canary's metaphor is a little dark at times.

We didn't necessarily mean it that way. We meant it in the more general metaphor is like an early warning signal. And it's very uneven. And that's part of it. You know, going back to the intelligence point, you know, I just want to finish that earlier point about these general purpose technologies,

they tend to take a long time to play out because even when you speed up one part, there are other parts that are bottlenecks or weak links that take longer to emerge. But that gets to, you know, your question about the Canary's paper where we did find that there are certain jobs that are already beginning to be affected and others that are actually going the other way that are becoming more valuable.

That's interesting. I mean, so one of the findings, at least as I read it, you can correct me, if I'm wrong, is that if there is an impact, it seems to be maybe as you'd expect on entry-level workers, the people whose contributions can perhaps most seamlessly be replaced by Claude or Chad GPT or whatever. I didn't necessarily expect that in advance. I should credit my co-auth Chandra and Rue Chen, who did most of the heavy lifting on this, and the folks at ADP, who provided all the data. And we went into it, you know, with a blank slate.

We just were open to whatever. Actually, the first cut at it, we looked at the top line. There's not much happening, actually. You know, the overall labor market, it wasn't that much happening. And we were kicking around like, okay, maybe we write a paper that all these headlines, these newspaper articles are all kind of overblown. And then, you know, Barat and Rui looked in a little more deeply, and they said, wait a minute, there's like,

There's one group that's really being affected. Early career workers, especially in the most exposed occupations. So one of the things you can do with, you can take all the jobs in the economy, 950 according to the Bureau of Labor Statistics, and each of them you can break down to a bundle of individual tasks. And once you do at that fine-grained level, you can really make headway. It's hard to say whether AI will like replace a radiologist, well, some people say, but it's much easier to look at a specific thing. Can it read a medical image?

and that's one of the 26 things a radiologist does. When you break it down to those tasks, and by the way, the paper that did that the best was this GPTs or GPT's paper we were just talking about, when you break it down that way, you can rank all the occupations. And then when you also look at the age, you find that the combination of the most exposed occupations

with the youngest workers had about a 16% decline in employment. So there's a couple threads to pull on there, and I want to do them both. One is the young workers, kind of the entry-level workers, and just the whole idea of like careers and experience and everything else. But the other one is the data. My perception is that right now we as a country, a society, maybe a planet, probably don't have as much data as we would like.

Oh, no. It's a tragedy. I spent a lot of time with our friends in Washington who are like working in the statistical agencies and they're having their budgets cut. It's getting, you know, there's more and more need for better data and less and less resources going into it. I think it's incredibly foolish. One of the things we're doing at Stanford is trying to help with that. We create something just last month called the Stanford AI Economic Indicators that is like a dashboard that brings together all this data from a lot of private sources as well as public sources so people can look at it in one place. Instead of all these sort of dueling anecdotes, you know, everyone can see the data.

The timing around this technological change of these major changes feels to me like it should be a deflationary narrative about how quickly AI is going to be deployed. But I'm kind of hearing from you, like, sort of the opposite of that, that you feel like there are, that it is, that we are going to deploy this much faster than in these previous instances, even though historically we haven't been able to do that, that it's just there's all this inertia across so many different sectors. And it also sounds to me a little bit from the data board that, like, that is what's happening. It is taking longer. Well, I think both of those things are true. What I would say, you know, I wrote about the need for these complementary investments. I wrote a paper called about the productivity paradox, about the first wave and then about this wave. And most importantly, a paper called the productivity J-Curve with Chad Severson and Daniel Rock.

And they all make this point that, you know, just because you have amazing technology, it doesn't translate into productivity, business changes, transformation of the economy. That said, that said, I think it's happening a lot faster this. time than with the industrial revolution or with electricity, which also took like, you know, 30 years. There's just a lot of structural reasons why it's going faster. For one thing, you know, the internet has been built out so we can just go from zero to 100 million users of chat GPT and like, you know, what was it, 60 days. Now it's a billion. Yeah, but like they're just going super fast. And, and a lot of the cognitive work, you can, you know, like software, you can do it a lot faster now than you could. It's still, there's still a lot of bottlenecks, which, you know, so I find myself

sort of between these worlds. When I talk to most economists in, you know, in New York or Washington or in businesses, you know, they see all the structural barriers. When I talk to the guys at the frontier labs, they're like, oh, we're going to have RSI, recursive self-improvement. It's all going to happen super fast. And I point out of bottleneck and they're like, oh, AI will solve that. And I'm kind of, I'm kind of between them. I put myself, you know, they're like probably two orders of magnitude apart from each other in terms of rate of speed, and I'm at the geometric mean, like one order of magnitude. And so I do think it's faster than most people in the rest of the world are ready for. And that's why we create that statement

about we must act now. I do also at the same time think that most of the technologists, they haven't really spent as much time in big companies as I haven't realized like how hard it is to get them to change. Can I ask this, though, knowing that you have a company that is essentially trying to accelerate the change? Yep. Is there an argument to be made that, Letting this take some time is actually a good thing for society because this kind of disjuncture in a labor market or in just the value of intelligence or any of the ways that we might describe this transformation is actually kind of a good thing to let it settle in more slowly. I agree with your core point. But what I would say is that we need to speed up our understanding. We need to speed up our adaptation to it.

We need to speed up our reskilling and what, you know, be prepared or preparation. At the same time, I'm very sympathetic to the idea that those core capabilities, you know, there was just another statement that come out, these statements are all coming out about, what they call it, pacing. Pacing. Ah, yeah. Pacing the frontier. Exactly, which is, you know, but that's on the capability side. So the way I think about it is that there are these two lines.

One of them is skyrocketing, which is the capabilities. The other one is our ability to adapt to it, which is barely moving. And that gap is where most of the big problems and challenges and opportunities lie over the next five or ten years. That's where all the action is. I tell all my economist friends, you should be focusing on that gap. My part of it is to close the gap from the bottom and speed up our understanding. Other people can think about the technology. Support for KQED podcasts comes from Star One Credit Union.

Give your savings account the love it deserves. When you keep your money with Star One, you keep more. of your money. Star 1 Credit Union in your best interest. Minions don't wait, especially for Wi-Fi. Summon's Same Day Wi-Fi from Xfinity and get the same price for five years. And seen minions and monsters only in theaters. Xfinity. Imagine that. Restrictions apply not available in all areas.

Learn more at Xfinity.com slash same-day Wi-Fi. You know, we're talking about sort of the acceleration of an economy, but there's also this simultaneous thing that maybe is different. Like, maybe there's a historical analogy for this, maybe there's not. I think of it almost as like the weirding of the economy. And this connects to a couple of things that we've talked about. You know, the idea that suddenly there's no such thing as an entry-level worker. Instead, you can only be, you know, a senior software engineer with an army of, you know, AI bots.

And you're really good at them. But it raises the question of like, where do senior engineers or senior people of anything come from anymore? Also, you know, again, to make that industrial evolution analogy, it's interesting to consider that when you replace your you know, horses going around a post with a steam engine, even though that's new and kind of radical, you fundamentally do understand how that works. Whereas when you replace your software engineering team with much of AIBots, you probably don't understand how it works anymore. Do these differences, I mean, do these matter? Should we really be looking closely at these? This, this incident of weirdness? I mean, those are, those two parts of your question are so different in my mind.

Sure, sure, yeah. I mean, they're both weird. I'll put them on that. So let me do them, you know, one at a time because I think they're both interesting to talk about. So on the, on the junior versus senior, this is a real problem. Like we described earlier the data that there's less demand for these young workers, but where do the middle managers come from? Where are the senior workers come from? And you need them. And you need people with that kind of judgment.

And we used to have this grand bargain where people come in sometimes do some kind of boring scutwork. But in the process, sort of by osmosis, they would learn how the business ran, learn how to be a lawyer or a doctor or investment. bank or whatever. Now they don't have that opportunity. And I think part of the answer has to be we have to like consciously and explicitly train them. I was talking to some folks at InfoSys. And, you know, their junior people are very much in the bull's eye of not being needed as much. But they tell me that they're still hiring a bunch of them, but now instead of having them do some of that boring work that AI could now do, they are explicitly training them with an AI system. Actually, AI can be a really good tutor and it can help them learn faster. So,

that's one approach to the answer. I think it's going to be, you know, kind of a new social contract that we have to think about because we don't only just abandon this whole generation of people. I mean, I'll tell you a little bit of a sad story, or hopefully it has a good ending. A student came to me in my office a couple of months ago graduating from Stanford, pretty good school. And she said, I don't have a job. My friends don't have jobs. Is my generation doomed? And I was like, whoa. I mean, you're a Stanford student. You should be a mindful optimist. I know, I'm a mindful optimist.

You're graduating. And what I tried to say is, look, I don't want to sugarcoat it. There's a bunch of jobs disappearing, like you just said. But on the other side, you can use these tools to do things you never could have before. You've got superpowers where you can do vibe coding and create all sorts of software that would have been impossible five years ago or even one year ago. And you needed to lean into those new possibilities. There's more startups being started than ever before because a lot of people. people are seeing this opportunity to create things they couldn't have done before.

So, you know, that's what I tried to teach in my class. I have a master class that also does this. But I really want people to, you know, I understand the downside, but I think there's almost too much emphasis on that. And there should be more of a leaning in to AI allowing you to do new things you never could have done before. They're harder to see because a lot of them didn't exist before. But that's where the opportunity is.

And that's the part I want to speed up is the transition to those new. new opportunities. I don't want to just, like, ostify everything and try to freeze everything in place. I don't think that's the strategy. We need to be nimble and have that more flexible opportunity to create new jobs, new opportunities. Another recent historical analogy to this is, you know, self-driving cars where we essentially see them being, generally speaking, safer drivers than human drivers. But when they do make a mistake, it tends to be sort of a novel mistake or the kind of the way the system breaks down is. We could not have anticipated, you know, power outage.

It's like an alien intelligence. Yeah, it is. They respect the cones too much. Yeah. You know, yourself driving cars. Exactly. They worship the cones. But they don't worship some other things.

And, you know, as an economist, actually, you know, this can be a little dangerous and off-putting. It's also an opportunity. Like, in economics, gains from trade happen when there's two entities that are very different from each other. If they're identical to each other, there's not much room for gains from trade. So I actually mostly encourage the folks at the frontier labs to lean into making the AIs really good at things that humans are not good at. And let us be good at the things we're good at. I think too often they do it the other way around. They're trying to smooth the edges and make them good at things that's hard for machines and easy for us, like buttoning a shirt.

And ultimately, the fact that it's an alien intelligence means that we can lean on it to do some amazing things. but there'll still be a role for humans, which I think is important. Like, I think it's good to not replace all the things that humans are doing. I wrote this paper, the Turing trap, where I basically argued it's a mistake. It's to do what Alan Turing said, which is make AI that's a perfect imitation of humans. We should make it different so we each have something to contribute. You know, it's funny. We can flip this around, you know, almost perfectly here on the podcast.

And in other conversations, Alexis and I are cautious and often critical of AI, particularly the industry, but we're also quite enchanted by the spaces inside these models, you know, these mysterious high dimensional spaces and their capacity and the things they seem to be able to organize and then kind of cross-connect in ways that humans can't, certainly not at that scale. And it makes me think of, you know, the old ancient, almost economics debate between central planning and sort of, you know, action at the edges. And you, you of course, know, this well, but for folks listening in the podcast, there's a few different ways you can organize an economy.

You could have, you know, now. Yeah, you can have robert and Alexis deciding exactly how much to make of everything, you know, all fashionable t-shirts and cool. What a great world that was bitch. And, you know, this has the benefit of coherency and you can actually have a plan and execute it. It can be aligned with the values of your society, at least supposedly in theory. All these things.

But then way over on the other side, you have the idea of, you have the sense that the real information about how the world works and what people want and what the problems are are, of course, where are there at the edges, you know, in people's lives, in their kitchens, in their businesses. Sort of the Hayekian view of like the great censor of the market. Now, it does seem, or that was, the story, you know, at least up until 2020, 2023, it does seem like maybe we have these machines now with a capacity that could change that balance a little bit. So first I'll ask you, like, do you see some of that potential? Oh, for sure. No, this is a big concern. Like, let me just say, You've heard I'm pretty optimistic and excited about the productivity potential and creating an enormous amount of wealth.

One of the things that I'm most worried about is it could be very badly distributed where everything gets really, really centralized. And the core reason for that is what you just brought up. You know, Friedrich Hayek wrote this amazing paper called The Use of Knowledge in Society, which, you know, articulated what you just said, that most useful knowledge is, like, widely dispersed in the economy. Like, you know, does this, do people in this neighborhood like peppermint ice cream, or is this trillions? truck half empty and maybe we could put some more stuff on it or whatever. There's all this, like, detailed information. And he argued that there's no way a central planner, even guys as as smart as you two, could make all the decisions in the economy. There's just too much of this detailed information. And that was totally true for the 20th century and up until recently.

But Zoe Hitsig and I have written a paper called AI's use of knowledge in society where we argue that, wait a minute, you know, you could actually go take these trillions of parameters and make all sorts of decisions. You could use the Internet of Things and other techniques to bring data and bring it all to say Bentonville, Arkansas, to pick an arbitrary city, and you'd be able to know all sorts of information about what's happening and make decisions. And it's actually beginning to happen. You know, when you look at the data and big centralized retailers are out competing those mom and pop shops, they know more about what people want in each neighborhood and which trucks are empty. They have all that detailed information that high exit would be

impossible, which is great for efficiency, but it may not be the best thing for freedom and democracy. We need to think hard now, while we have some optionality, how can we design a world where we maintain our freedom, maintain our decentralization of decision-making, and that if we go too far down the path of disempowering people, it may be very hard to reverse course later. And for what it's worth, everyone's noticed, it's kind of beginning to happen, right? And we better like take it seriously because we go much further down that path. It'll be too late. Yeah. I did want to ask you, I mean, this is from your book, Second Machine Age, of course you co-authored.

Andy McAfee. And you have, to paraphrase Martin Luther King Jr., the arch of history is long, but it bends towards justice. We think the day to support this. We've seen not just vast increases in wealth, but also on the whole, more freedom, more social justice, less violence and less harsh conditions for the least fortunate and greater opportunities for more and more people. I think like there have been, the majority of my life, I think I would have more or less agreed with this. I think in the last 10 years, my own kind of priors have been challenged on this. Not just you.

Yeah. No, we wrote that in 2014. I'm glad that I just checked and it's getting like, it's got more and more sales and citation. So I'm glad that it's got legs that way. And that was sort of maybe like the peak of this end of history. Obama era optimism. Exactly. But speaking of Obama, you know, he modified that quote and he said that it bends, you know, I'm going to misquote him, but the gist of what he was saying was, you got to push on the arc. You got to push on it. It doesn't happen automatically. We don't just sit back and watch it happen. And that's very much, you know, you mentioned I was, I call myself a mindful optimist, you know, that the arc of history bends if and only if we push it. And I think lately it's been going the wrong way. It's also long, like Martin Luther King said. So, you know, it's not going to be monotonic where it always improves. every month or every year. So I think it's fair to say we've had some backsliding and some bad things who have happened. I'm still optimistic, maybe a little less optimistic than I was in 2014,

but I'm hopeful that we work harder. But this is exactly why I do what I do. It is I'm not here to predict the future. I'm here to say, here are some possible futures and here are some levers that matter. And we need to push on them because most of us do want not just more abundance, but also more freedom and shared prosperity. Can I keep pushing on this one a little bit? Just because I think there's even some people in this building here at KQED who their biggest concern about AI. They have many environmental this and other things. But it's really about particularly being here in San Francisco, you really see it, the concentration of wealth.

And then our city politics has been taken over by tech wealth. It's also the concentration of power that goes along. Those tend to go together, don't they? Yes. Yes. Yeah. And I guess the maybe more difficult version of this question is do you think that AI as a result of what the frontier models need, the amount of capital that you need to deploy to do these things? Is this a kind of technology that inherently leads to higher concentrations of wealth and therefore power?

I don't want to say inherently because I don't want to be in the prediction business. I want to be in the design business. But there's a lot of strong forces. And there's two sets that are particularly obvious to me. One is, you know, the scaling laws and the fact that the AI itself tends to work better at large scale. That's why they're raising hundreds of billions, trillions of dollars to build bigger and bigger data. And borrowing it. And, you know, there's just this amazing thing that Dario Modi and others noted that you make the model bigger and it works better.

So they say, let's make it even bigger. And it just keeps working. So that tends to lead to kind of a bit of a winner-take-most outcome. Actually, I've been somewhat surprised how many frontier labs there still are. It hasn't all just concentrating to one singleton. But who knows where that's going to go? It certainly you have to be pretty big to be a player in that. But the one that I'm more concerned about is the other like 90% of the economy,

that, you know, whether you're in retail or manufacturing or health or wherever, you know, having decentralized information may not be as competitive with having it become more centralized. And that could also lead, and it has been, if you look at the data, there has been more concentration. So we need to think harder about how do we decentralize it. I mean, part of one of the answers that I would put forward is this idea of pushing AI to complement humans rather than substitute for them. But we need to look at all levers. Thinking about that discussion about centralization, about kind of where the power, where the money resides, is that change in the capital versus labor share of the economy a warning sign? I mean, if that continues, yes.

Yes, this one I'm not going to hedge on. So capital is inherently much more concentrated than labor. You know, a cool thing about labor, even though it's pretty uneven. Like, we all have basically one brain and like some of them, you know, maybe a little smarter than others. but, you know, it's distributed through the economy. You know, no matter how smart you are, you can't run a whole Fortune 500 corporation, so you delegate stuff. And the whole economy has all these delegated decisions to all these different brains.

But with capital, you can concentrate it much more. And empirically, you know, the income from capital is much more concentrated in a small abstraction of the people. So, you know, we were saying earlier that economic power begets political power. So that's something I worry about. I mean, what would that world even look like? You know, like I've looked at some other countries that have, you know, labor share that's like much lower. Yeah. Labor share of income is much lower.

And it's not a good set of countries. They're like resource-cursed countries. It's Saudi Arabia. It's like mineral mining countries. Exactly. And you get these like gated communities and these people who have a lot of wealth. I was talking to somebody from a Latin American country. She's very wealthy.

And she said, you know, all the wealthy people in my country are in prisons. I'm like, what do you mean they're in prison? She said, well, the prisons of our own construction. Like, we can't go outside of our houses. When ever I get into a car, I have guards on either side of me. It's like one of those dystopian science fiction movies. What interventions could we at least start to prepare for now economically? You know, obviously there's a whole suite of politics and, you know, who we elect and how we organize ourselves as society.

But just thinking about economic policy, what are, what would be some smart things to start thinking about right now? So I think we have some levers to push back against this. And that's, you know, why I care about it. So I did write this paper, The Turing Trap, about how it's a trap to have AI that just replaces and imitates humans. So a couple of things we can push back on. First off, there's a lot of economic incentives right now that I think mistakenly steer us towards favoring capital over labor. Like in the United States and most countries, for that matter. Tax policy.

Tax policy is much, you know, capitalist tax much cheaper than labor. So you have two brilliant entrepreneurs, you know, one of which says, I'm going to make a $100 billion company with, you know, thousands of employees. and the other one says, I'm going to make a $100 billion company with, like, no employees, but lots of robots. The U.S. government says, oh, the first person, you know, Alice, you're going to have to, like, pay a lot more taxes. Your whole organization is going to pay a lot more taxes, and the second one's going to pay less. And, you know, the first rule of taxation is, like, whatever you tax more, you get less of. So we're basically putting our thumb on the scale saying we're going to get more capital intensive and less labor intensive. Like, for most of history, maybe that didn't matter that much.

it wasn't that much leverage to do things different ways. Now we really have the potential. And there's a lot of other tax things you can do. My friend Daronasamoglu has written a lot about this, Pasquale Estrapo. So those are some things that we can do, like, you know, that are really hardcore. I'm an economist, but I've come to think that actually culture

and the way people think about it is more important than the, like, hard dollars. So here in San Francisco and Silicon Valley, I run into a lot of people who have this mindset that the goal of AI is to replace humans. And that's just wrong, I think. And when they, all the benchmarks that you see being published, almost all of them are geared towards, like, how well can this machine by itself do the task?

With Andy Hopped, and we've developed a new set of benchmarks, we call them Centaur benchmarks, sort of like part human, part machine. And the idea is to say, hey, guys, think about not how well a machine by itself can do it, but how can a human and machine together do it? it. And in many cases, in most cases, the human machine can do better than the machine by itself or the human by itself. But it requires a different architecture. You know, Doug Engelbart years ago's, you know, talked about how we should make machines that amplify humans. Steve Jobs called it, you know, bicycles for the mind. And that philosophy has kind of been lost out a little bit.

I want to revive it more. And if we design machines more to augment humans and to complement what we're doing, we're likely to keep people in the loop. And ultimately, I think that's going to not just be fairer. I think it's going to create a lot more value than trying to get the machine to do everything by itself. Eric, talking about your early apprehension of the curve we were in and kind of the trajectory of AI, let's talk situational awareness for people on the ground level. So outside of Open AI headquarters here in San Francisco, imagine you're somebody who works at one of the Kaiser hospitals.

great day-to-day job, how are you going to know that it really is taking off, that you're inside this exponential and that this prediction of a transformational AI era happening pretty fast rather than predictably slow is true. Like day-to-day, week-to-week, what should I be watching for? Well, you can come to the AI economic indicators and we'll have a monthly update of all those metrics.

You know, I'm not joking about that because I just see so many anecdotes and stories. And honestly, it's a little frustrating because, you know, every month there's hundreds of thousands of jobs destroyed, hundreds of thousands created. And if you're a reporter with an angle, you can definitely find anecdotes that support your story. And you'll have some man in the street who tells you what, you know, what the people are. And I've been reading those, and I just don't know how to aggregate them. So, you know, I'm a data person. I'm a statistician.

I'm an economist. So, you know, getting the stuff aggregated, I think, is a way to do it. The problem is that most of our indicators are kind of lagging indicators. and we need more so forward-looking leading indicators. And we're trying to invest in creating those. But I think that would, that's my answer. Promise me when the day comes that you're like, uh-oh, you're just going to put it all in one big blink tech.

It's just going to have like a little siren. Yeah, yeah, yeah. Well, we do actually have these color-coded. Like the transformation tracker, we've got like the 12 different metrics and we color-code them by like which ones are moving in the direction. Right now, only two of the 12. are moving in that direction, you know, above a significant level. They're all kind of moving a little bit.

So, yeah, we can, we can, you can take a look at that. I just want, is it happening. Dot, Stanford. Yes. All right, you give an idea. We'll put a little sign-up sheet that, like, if you want, we'll send you a text message. You know, it'll be like, all 12 indicators.

459 on Thursday. Okay, we're officially hitting the singularity. Well, okay, until then. Yes. Until next Thursday. Exactly. Thank you so much. Thanks for your time.

Such a pleasure. Yeah. Wonderful. KQED's Dream Machines is made by humans and hosted by me, Robin Sloan and Alexis Madrigal. Our series is produced by Anianz-Cortez and Derek Larto. Sound design by Brendan Willard. Jen Chien is the executive producer and Ethan Tovin Lindsay, our editor-in-chief. Support for the production of Dream Machines comes from the Christian and Shaw family,

Dorothy Marsh, and other generous K-K-K-Evon. QED members. Special thanks to Chris Agusa, Annie Fruit, Paul Lancour, Vivian Morales, Zaldi Serrano, Xteen Tinoosso, Hazel Tesoro, and Alex Tran. And of course, thank you to the close all tabs team for letting us visit their feet this month. You can find Dream Machines right here on the Close All Tabs feed every Monday in the month of August. If you'd rather watch than listen, the video podcast will be up on YouTube. Search KQED News or go to YouTube.com. slash at KQED News. And close all tabs. We'll be back on Wednesday with a brand new episode. Support for KQED podcast comes from San Francisco Opera. This fall, dive into a world of political intrigue with Simone Bocanegra and Mary Queen of Scots. Discover a life of Parisian luxury in Monon and laugh along with the marriage of Figuero. Learn more at sphopra.com.

Minions don't wait, especially for Wi-Fi. Summons Same Day Wi-Fi from Xfinity and get the same price for five years and seen minions and monsters only in theaters. Xfinity, imagine that. Restrictions apply not available in all areas. Learn more at exfinity.com slash same-day Wi-Fi.