Plain English with Derek Thompson - How Superintelligent AI Could Upend Work and Politics

Episode Date: November 18, 2025

Many AI experts believe that some time in the next few years, we will build something close to artificial general intelligence (AGI), a system that can do nearly all valuable cognitive work as well as or better than humans. What happens to jobs, wages, prices, and politics in that world? To explore that question, Derek is joined by Anton Korinek, an economist at the University of Virginia and one of the leading thinkers on the economics of transformative AI. Before he focused on superintelligence, Anton studied financial crises and speculative booms, so he brings a rare mix of macroeconomic skepticism and technological optimism. They talk about quiet AGI versus loud AGI, Baumol’s cost disease, robots, mass unemployment, and what kinds of policies might prevent an “AGI Great Depression” and keep no American left behind. If you have questions, observations, or ideas for future episodes, email us at [e-mail]. Host: Derek Thompson Guest: Anton Korinek Producers: Devon Baroldi Learn more about your ad choices. Visit podcastchoices.com/adchoices

What's up? It's Todd McShay, host of the McShay Show at The Ringer and Spotify. We're building this thing up and I couldn't be more excited to be back, talking college football and everything NFL draft with the most informed audience out there. That's you. My co-host, Steve Mention, I will be with you three times a week throughout the football season with all the latest news, analysis, and scouting intel from around the league. For even more insight, subscribe to my newsletter, the McShay report to access my mic, drafts, big boards, tape breakdowns, and other exclusive scouting content you can't get anywhere else. It's going to be a great season. And I hope you'll be with us at the McShea show every step

of the way. Today, AGI. Last week, I started to watch the new adaptation of Frankenstein on Netflix. It is beautiful and lush and completely over the top, which I think is mostly par for the course for director Guillermo del Toro. But this is not a movie. review. There's something about the story of Frankenstein that seems to compel every generation to stage its own version. This is true going back to the beginning. When the original book came out in 1818 by Mary Shelley, it was a minor sensation with highly mixed reviews. It wasn't until five years later in 1823 when a theatrical adaptation at the English Opera House in the West End was a huge hit that Mary Shelley became famous and the book went into its second printing.

It's second of approximately infinity printings. To date, Frankenstein has been the subject of more than 400 movies, 200 short films, 80 TV series, and 300 TV episodes. The scientist who plays God and the creature who rises up to destroy his maker has become a truly modern myth. And I think I know why. Frankenstein was first published a few decades into the first Industrial Revolution. It was written in a state of wonder about new science experiments,

that combined a rudimentary understanding of physics with a fascination for bringing people back from the dead. In the late 1700s, Luigi Galvani, a physician in Bologna, believed that electricity emanated from the bodies of animals. To prove his point, he killed an obscene number of frogs. He cut up the frogs, he electrocuted them until he could see the muscles on their little frog corpses twitch. I apologize for the imagery.

But this movement that he inspired to electrocute, animals to make them move and make them seem to come alive, was called galvanism. In diary entries, published around the same time that she wrote Frankenstein, Mary Shelley wrote that she wondered if, quote, perhaps a corpse would be reanimated. Galvanism had given token of such things. Perhaps the component parts of a creature might be manufactured, brought together, and endowed with vital warmth. End quote. Every generation since 1818 has come of age in an era of technological marvels, productivity advances, and broad economic growth. It can't be a coincidence that this er-parable of technology was created just two generations into the birth of the modern world.

Frankenstein has been made and remade and remade and remade, I think, because it answers a question that ever since the first industrial revolution, we can't stop asking ourselves. are we, human beings, capable of creating an invention that rises up to destroy us? 200 years after Shelley wrote Frankenstein, we now face the prospect of an even more modern Prometheus, aGI or artificial general intelligence, which could be smarter than any human at any task. Here's Anderson Cooper on 60 Minutes asking Anthropics Dario Amade about the implications of such a thing. Half of all entry-level white-collar jobs? Well, if we look at entry-level consultants, lawyers, financial professionals, you know, many of kind of the white-collar service industries,

a lot of what they do, you know, AI models are already quite good at, and without intervention. It's hard to imagine that there won't be some significant job impact there, and my worry is that it'll be broad and it'll be faster than what we've seen with previous technical. This is a completely audacious prediction. The idea that AI could wipe out tens of millions of jobs by the end of the decade is ludicrous. But it's not just AI executives who are making this claim. Several economists who study AI most closely say it's not out of the question that by 2030 we could have a technology that is broadly seen as superior to white-collar workers at just about every white-collar task, reading, writing, analyzing data, making PowerPoints, writing software. The strongest piece of evidence to prove that this might be true

comes from Meter, a nonprofit base in Berkeley, California, which analyzes Frontier AI models' capabilities. According to one widely cited meter study, the length of tasks that Frontier models are capable of executing at a 50% success rate is doubling every seven months. If this trend continues, AI will, within the next year or two, be able to tackle projects that take the typical high-quality worker several weeks to complete. Now, of course, I think we should probably pump the brakes on the most dramatic forms of these predictions.

There's a huge difference between a tool that's capable of taking on complex tasks and a company that actually fires half its workforce because it's learned how to implement that tool. When I think about HGI, what I feel is not terror, but something more ambivis. skepticism, and even confusion. The skepticism comes from my suspicion that AGI is much further away than many of its boosters think. The confusion comes from the difficulty of thinking through how a human economy would incorporate a superhuman intelligence. What happens to prices and growth and employment in a world where AGI is cheap and employed humans are inferior at a range of valuable skills. It's really surprising to me how few economists have really started to think through

the implications of what AGI would actually mean for this economy. But we do have one of them on today's show. Anton Kornick is an economist at the University of Virginia, where he researches advanced AI and its economic implications. And today we talk about whether a real-life Frankenstein story is playing out in technology today. We try to think through some of the more complicated implications of a tech that is brilliant at solving math problems, but impractical for changing bedpans. We seem to be walking toward a future where many of the highest skilled jobs will be automated, and many of the so-called low-skilled jobs will be hard to replace with technology. The future will be weird indeed. I'm Derek Thompson. This is Plain English.

Anton Koranek, welcome to the show. Great to be on air with you. So in the last few weeks, we've done several shows on AI, and what happens if it goes wrong and turns out to be a big bubble? This episode begins with the opposite premise. What if AI goes right? If we take seriously the predictions of these frontier labs that we are within a few years or a decade of building AGI, artificial general intelligence, what happens to the economy?

So before you were one of the go-to experts on the economics of AI, you were a single, financial economist studying financial crises. So I actually wanted to begin with this. How much does this AI infrastructure build out remind you of things like the dot-com boom and the housing boom? It feels very much like that. There's a lot of speculative frenzy right now. It feels like if you are an entrepreneur and you say, I do AI, you can easily raise double-digit millions. and in the past, I would have said this has all the hallmarks of a speculative frenzy and a bubble. But at the same time, I believe that even though there may be some short-term frenzy going on, in the medium term, AI is going to be much more impactful and much more powerful than any previous invention.

Why? So if the bets that the leading AI lives are making right now, if they come through and we have artificial general intelligence, which you asked me to define before, so the charter of open AI, for example, would define it as something like machines that can perform virtually all valuable economic work. or I think Dario Amode described what he called powerful AI in an essay last year where he said this would be like a country of geniuses in a data center. And if we have any of those visions, it would utterly transform our world. But why should we believe them, right? You're talking about OpenAI. You're talking about Dario Amadee, the CEO and founder of Anthropic.

These are businessmen. They're raising money. They're spending billions of dollars more than they're actually bringing in. They need to existentially, to stay alive as companies. They have to persuade their investors, that they are working on something that is absolutely ginormous in its implications in order to justify the amount of capital that's going into building these machines. So of course they're going to say, oh, it's a nation of geniuses and a data center.

This is going to change the world. Why do you believe them? Why do you think they're right that we might have something like artificial general intelligence by the end of the decade? Right now, it's still a bet. And I think if you catch them in private, they are probably going to say as well that this is a bet. There is no certainty about this at all. But we have a number of indicators that suggest that we are, you know, on a curve, we are scaling these things and there are predictable relationships that tell us

if we put more and more computational power into these systems, they're going to get better. So that's kind of the first indicator. It's just extrapolating a curve and we know there's some risks with extrapolating things. Sometimes relationships suddenly stop and it won't work anymore. Then the second indicator is, you know, just my personal, lived experience, part of what I do in my research is I follow the capabilities of AI systems very closely and I write regular reports about how you can best use AI systems in science. And there, I'll say, I have just continually been blown away every time that I wrote another

piece on this topic over the past couple years because advancements have been so quickly. And then the third point that I think a lot of people in the space are making is we are talking about neural networks. And there is a proof of concept in nature that is sufficiently advanced and properly via a neural network, which is what we all have in our skulls, our brain can be generally intelligent. So at some level, all the artificial neural networks that we are designing nowadays are inspired by biological neural networks. There are some differences, and we run them differently, obviously, but the fundamental power of neural networks in biological brains and in silico is the same. And so in some sense, the bet that the frontier AI companies are pursuing is that, well, we see that there are biological neural networks that are generally intelligent. We are betting we can reproduce something like that, even if it looks a little bit different in silico.

You mentioned three reasons to think that AGI might be coming by the end of the decade. Just to reiterate for my own benefit, number one, you see that the benefits of scaling are continuing. Number two, your own experience of tracking the capabilities of these machines continues to rise exponentially. And number three, you think that there's a general property of neural networks that suggest that they'll achieve some kind of general intelligence, not just topic-specific intelligence. I want to zero in on the second principle here. What is the best indicator that we're on track for AGI so that five years from now we should expect that there's going to be this explosion of superintelligence? Yeah, there is unfortunately not like one single indicator. And part of the problem is our human intelligence is so broad.

We can do so many different things. You have a whole bunch of technical benchmarks. Let's see, for example, there's an ARC-EGI index that tries to do what you are suggesting that tries to basically measure our progress towards general intelligence. But then it leaves out some very basic capabilities that a 10-year-old, old child can perform because right now our brains are just much broader. They can do things like writing text like in language models or doing mathematical derivations like in reasoning models, but they can also steer our body to walk. They can smell the fresh air. They can listen to a podcast.

Our brains can do so many different things all at once. I guess the way that I look at it is that our specifically human capabilities have been shaped by a process of millions and millions of years of evolution to be just what they are, because that's what proved to be most valuable from an evolutionary standpoint. And that's why we, let's say, we can see very well. We can strategically plan ahead. we actually suck at math compared to a lot of machines because we didn't need to do that in our evolutionary natural environment. And we have this kind of very specific combination of capabilities and skills that all come

together in our brain. And so the question is, when will AI systems be able to master all of those that are economically useful? And maybe there's also a second question. do we actually want to pursue one system that can perform all those skills at once, or would it be just as valuable from an economic standpoint, if we have a whole bunch of separate systems that can do them individually and that we could perhaps even steer and control a little bit better?

I want to get a better understanding of what AGI would actually feel like if it existed. So this morning, I was working on a couple of different stories, and I asked ChatGPT to edit, copy edit, an essay that I wrote. I asked ChatGBT, BT to read a long science paper and summarize it. I asked it to do several math problems. I asked it to summarize a really complex, to me, field of psychology. It did all of this faster than any human being could possibly do any of it. And it did it across subject areas, right? It's a copy editor, it's a mathematician, it's a psychology researcher.

Why isn't that AGI? Why isn't the ability of these tools already to do things that no human can do across different subject areas? Why shouldn't we already think of artificial intelligence as an artificial general intelligence that somewhat sort of scrambles the terms of this conversation? Yeah, it totally does. And you know, if you had shown anybody, I think, what today's LLMs can do 10 years ago, they would have probably said, yes, this is AGI. This is exactly what we meant when we used that term.

But, you know, we have this kind of habit of shifting expectations. Whenever an intelligent AI system can do one thing, we'll say, okay, sure, it can copy at it, but it can't do this and it can't do that yet. So I think, you know, for practical purposes, these systems are already very powerful. Even if we stopped any capabilities progress today and we just spent the next decade rolling out the existing capabilities throughout our economy and throughout all our organizations,

we would have quite a significant level of economic growth just from that. but there are still a few areas in which today's models are clearly subhuman. One of them is they don't have the ability to dynamically learn. So whenever you open up a new, let's say, instance of your favorite language model, they have forgotten everything that you did in the past. They have not learned from that in their neural weights. You may be able to kind of put some of the conscious, context from past conversations back into its working memory, into its context window.

But that's really only a second best. And it doesn't make their intelligent as fluid at learning as if they adjusted their weights in a way that, for example, our brains do. And so that's one clear shortcoming and that limits the use and capabilities of models throughout the world. I was recently asked by someone at one of these larger frontier models to think through what might happen to the world if we got AGI in the next few years. And one way that I thought about it is that there's two scenarios that I see.

And I named those scenarios quiet AGI and loud AGI. So quiet AGI is a world where we get AGI, but nobody really knows that we have AGI. There's no newspaper article. There's no broad understanding. We develop something that is essentially Nobel Prize winning intelligent across every single domain of human endeavor and discovery. But there's no immediate announcement. There's no headline. There's no trillions of dollars worth of economic growth.

There's no clear productivity bump that the intelligence exists. But it takes a long, long, long time for it to actually result in changes the physical world. That's quiet AGI. Loud AGI is a world where it arrives with a bang. There's a headline that a large language model was given a week to solve the problem of pancreatic cancer, and we know that it did it. Loud AGI is a world where we invent something that immediately hacks the Chinese electricity grid, and like there's a total blowout to all lighting in Manchuria, and we realize, oh my God, holy shit, we developed essentially a, digital nuclear weapon that can take down other governments, right? Something like that is loud AGI.

Before we go and talk a little bit more about macroeconomic scenarios of what the arrival of this technology would look like, do you think the arrival of AGI will be more of a quiet phenomenon or a loud phenomenon when it is first created? You know, let's say a decade ago advances in AI. were a lot quieter, right? But nowadays, they tend to be on the louder side. I hope we're going to use it for productive purposes like cancer rather than offensive purposes. I think there's also a middle way and maybe that would actually be my preferred outcome. So it would be kind of a waste if we have these capabilities and it remains so quiet because we are not really using it for

anything. I would love to cure as many cancers as possible, right? And in that sense, I would love for the AI to be what you call cloud. So basically, we want to use the intelligence that we have as soon as we get it for all the productive, positive purposes that we can use it. And as you said before, there is some incentive to hype things a little bit because you need to raise a lot of capital for building data centers and so on. But the hyping itself doesn't really benefit the AI because it creates skepticism like what you expressed before. And it's not productive in itself. So I think our ideal aim would be to create AI that we do actually apply in lots and lots of instances across the economy or across many scientific questions.

Like, let's solve our energy problems by solving nuclear fusion. Let's solve all our medical problems, cancers and so on, so that we can all live better and hopefully live longer in good health. And it doesn't need to be super loud for that, but it does need to be impactful. But wouldn't you agree, Anton, that it's possible that we invent something that the leading labs call artificial general intelligence or superintelligence and various experts agree, yep, it's basically smarter than every human being at all of these different categories. But also, it doesn't immediately cure cancer. It doesn't immediately invent fusion technology. I feel like sometimes there's like a little bit of an unfair game that's played where we imagine the creation of this technology, and then essentially say, yada, yada, yada, every big question in human history is answered immediately. Well, what if we invent a technology that's smarter than all of us, but it's only a little

bit smarter than all of us? And so it actually doesn't understand how to solve pancreatic cancer immediately. It doesn't understand how to scale fusion technology immediately. It doesn't understand how to do many of the things that humans haven't figured out how to do, because it turns out there's a lot of problems of biology and physics that are just really, really hard. And so there's actually this long lag between the creation of this technology and the answering of these questions that we sometimes bundle into the creation of that technology, right?

Like, in a way, I feel like we're almost larding these expectations, loading these expectations on superintelligence that still just might be decadowing. decades and decades away, even if this technology is on pace to be extraordinary by the end of the decade. Does that seem like a fair frustration of someone like mine? Right. Yeah. It's reasonable. It's plausible. And I'll say, frankly, would I be surprised if it turns out that way? Not entirely, but it's not my mainline bet. And the reason for it is the following. You know, I spend my day-to-day life as a researcher. And what I perceive to be the greatest scarcity in making scientific progress

is precisely that we don't have enough minds working on all these problems that we want to solve. So let's say in economics, for example, we want to have more economic growth. We want to have less inflation. We also don't want to have deflation. We want job markets to function smoothly and, and, and, and, and. But we have only a very limited number of brains that are working on this. So if you tell me, you can suddenly give me a data center with a million genius-level economists, I personally would expect that there are so many problems to be solved

that they could make significant progress on. I'm sure if you talk to a biomedical researcher, they're also going to tell you, yes, I can give you a long list of problems that I know have a probable solution, but we don't have time and we don't have the capacity to tackle them. So if you give me that genius data center, then we could make a lot of progress in them. But I think what kind of the next step that this chain of thought brings me to is what are going to be the new bottlenecks? Let's say we are, for example, biomedical researchers. There's still going to be a bottleneck in that we are going to have to conduct wet lab experiments, right? We are automating some of that. There are some wet labs that are operating like a 99% robotic basis.

and we can accelerate that a little bit, but just having geniuses doesn't solve the problem entirely. It makes progress faster, but it doesn't make, let's say, every problem solved by itself. So in that sense, I agree with you. Right. Even in the case that, let's say, it's 2028, and some data center full of geniuses invents or identifies a molecule

that they say can cure late-stage panes, pancreatic cancer, you then need to test that molecule in rats, in monkeys, in people. You need to get it through FDA approval. There's an enormous process between the invention of that molecule in a data center and it's FDA approval and manufacture to deliver to human beings. And so that's what you're saying, is that even in a world where we get this kind of superintelligence, you're still going to have all kinds of molecular bottlenecks, let's put it, you know, human bottlenecks between the software invention of the technology and its application. What I want to do for the rest of our time together is just assuming that you're right, assuming that at some point in the back half of this decade,

something like AGI arrives, what would it actually mean for the economy and for workers and for life? So just to get you started here, to get the ball rolling. Let's speak. You have done as much thinking on this subject as anybody else, so we are going to speculate irresponsibly, but this is the most responsibly that you couldn't speculate irresponsibly. So at some point in 2028, let's put it, let's say, the Frontier Labs develop an AI system that if given the prompt, write a better version of yourself,

computer programmer, that machine will do it. They will be able to build a better computer program than any computer programmer that exists. And no one will know what to do with this technology more than the frontier labs. So it seems very possible that will be one of the first applications. Doesn't that suggest, Anton,

weirdly, that perhaps the first effect of AGI will be to disemploy the software programmers who built the AGI. That's indeed their very first objective. Yeah, I think that's the plan right now. Let's try to build one of those software systems. And by the way, if you walk around in Silicon Valley, you can hear more aggressive predictions on that in 2028 as well. But let's take 2028 is our median scenario. So that stage is

called recursive self-improvement, AI systems that can create better AI systems. And again, this question that you brought up before becomes very salient, is it going to be a actually really hard to create better systems. And are we going to be severely bottlenecked by, for example, compute availability, by the availability of server farms? Or if they're smart enough, are they going to be able to make very fast progress to create a system that's like an order of magnitude better than the predecessor system? And that's something that nobody really knows, right? We are speculating on top of the speculation. But, you know, let's say it's the next system that they create is like really significantly better than the previous one. And let's also say

that we kind of collectively decide we want to use those systems, not just to create a perpetual chain of ever-smarter systems, but actually to do something positive in the world. Like, for example, curing medical diseases. And so we'll use that successor system that is now as smart as a genius in kind of human microbiology. And let's create a million copies of it and try to attack all the diseases and known causes of death and better the human condition in that way. So if we decide that, and if it's really genius level, my expectation is we would make a ton of progress

in whatever field we choose to deploy it on very quickly. Now, another area, you were asking about economic implications, another area where these systems are going to be deployable because they are generally intelligent at some point would be in economic strategy, in running corporations, creating products, distributing products throughout our economy. So what you would expect is you not only have a country of geniuses in the scientific domain, but you also have a country of genius CEOs in a data center that can run corporations efficiently

that can decide where would it be good for me to deploy more AI systems, where do I want to deploy the humans, what's kind of the optimal mix of the two, and at the same time, you're probably going to run a whole bunch of copies of that system for the next successor system, which is going to be even better. So I think that's more or less what we should expect happening. and we would see rapid changes in the corporate world because of this extremely smart data center CEOs. We would see extremely rapid changes in the scientific domain because of all this geniuses working on scientific problems

and essentially anything where intelligence is valuable would see rapid progress. Now, so far we've only talked about the AI at some point the more intelligence you have in the world, the more the physical component is going to become a bottleneck. And that will bring us to the next question, which is robotics. But let me hand the mic back to you. I want to get to robotics just a second. But I want to pin us down on an implication of the story you're telling. You're describing a world where there's very sudden increases in productivity in some parts of the economy, but not all parts.

parts of the economy, right? Software programming becomes much more productive. Certain other white collar jobs become much more productive as we deploy superintelligence toward those ends. No one's going to be using super intelligent large language models to serve them food or to cut their hair or to take care of their sick parents, right, home health aids. And there's this idea in economics called Baumol's Cost disease, which says that as an economy sees productivity advances in one part of the economy, but not all parts of the economy, prices rise where human labor is still essential. So, for example, if it becomes cheaper over time to make clothes and food and electronics and GDP increases and productivity increases and wages rise, what you should expect and what, in fact,

we see is that it becomes prohibitively expensive to have a butler or a servant. It becomes very expensive to, say, attend a four-string quartet because you haven't been able to make a four-string quartet more productive, even at the same time as you've been able to make, say, the manufacturing of a shirt more productive. So Baumol's cost disease says that parts of an economy that are most dependent on human labor will, in a productivity surge, become significantly more expensive. In a way, Anton, isn't the scenario that you're predicting, a world in which we should expect to see very strange increases in prices in parts of the economy that are still dependent on human labor because of the rest of the economy is basically getting the humans wiped right out of it? I mean, how will that work at a price level? It frankly strikes me as like a very strange scenario. Yeah. So, Bomber's cost is whatever you're going to do.

can't automate becomes the bottleneck. And in some sense, if you look at our economy over the past century, you can see some of the signs of it, right? In the first half of the Industrial Revolution, we developed very powerful physical machines. We developed cars, we developed construction equipment and so on and so forth. And all of that needed brains to complement it and do. operated. And that's why especially cognitive work has become a lot more valuable. And over the past couple decades, the skill premium, the premium for being able to perform highly educated cognitive

work has been rising and rising. So what we are discussing now is that the exact opposite may be happening when we have transformative AI, that all of a sudden cognitive work becomes dirt cheap because the cost of running these AI systems right now is like right now it's several orders of magnitude cheaper than what a cognitive worker, let's say a copy editor would charge. Over time, probably these systems are going to get a little bit more expensive, but they're still going to be cheaper than what cognitive workers are being much now. And so what we should see is that everything that only relies on cognitive work, for example, What researchers do a lot of their time is going to become a lot cheaper.

And everything that involves interacting with the physical world is going to become comparatively more expensive. One of the godfathers of deep learning suggested already a decade ago that we should all become plumbers because that's something that AI won't be able to do for a couple years. And, you know, if you had followed his advice 10 years ago, I think, You would have done all right, actually. But maybe now would be the time to really take the advice on board, especially if the predictions of AGI within the second half of the current decade, so before 2030, turn out to be true.

Do you mind if I just quickly interrogate this premise? So I believe you're, I believe you're referencing Hinton, who said we should all be. plumbers, and he also said about a decade ago that certainly nobody should go into radiology, because one of the first things that AI is going to be able to do is read all of our cat scans and MRI scans and tell us exactly what's going on inside of our bodies, because it's just so good at surveying images. There are more radiologists today, Anton, than there were... Than in 2015. There's more radiologists today than there were five years ago, and those radiologists

are making significantly more money. I think the average wage of a radiologist in America today is over half a million dollars. It's one of the highest paid occupations in the U.S. Is it possible that rather than replace workers, AGI makes many white-collar workers more valuable because it makes them more productive and that rather than take the human out of the loop,

we end up just super empowering the humans in the loop. I mean, just one other historical example that might serve as a prediction here. Imagine if in the 1970s, we looked at spreadsheet workers like accountants and bookkeepers and said, well, someone's inventing Excel right now, and Excel is going to automate all of these accountants and bookkeepers. There's not going to be anybody working with spreadsheets.

future because spreadsheet work is going to be so, so efficient. Well, it turns out that what Excel did is not to destroy the spreadsheet worker, but rather to turn every single frigging white-collar worker into an Excel worker. I mean, who doesn't work with Excel these days? I don't even like it, and I find that sometimes I have to use it to make graphs or whatever. Why won't the same thing happen with super-intelligent large-language models? Why won't super-intelligent large-language models just be the excel of the future, something that every white-collar worker essentially has to work with that makes them more valuable at doing things that require a certain kind of new intelligence. Yeah, so people like Jeff Hinton, they're always ahead of their time, right? And his call was

clearly premature. Maybe it's going to turn out to be wrong, right? But I would probably categorize it as premature. Now, why am I, so first of all, what you are saying, again, is plausible, especially in the short term, I'm certain that the majority of cognitive workers is going to be augmented rather than automated. Now, the difference is if you do have the true HGI, then there's suddenly nothing left that only the human could do. So after the invention of Excel in your example before,

it still took us humans to take the data from, somewhere else, shovel it into the right columns in Excel, add the right equations and so on. And if you have an AGI, the whole premise of it is that it could do all of that and that there is no task left that it couldn't perform equally well as we humans. Otherwise, it wouldn't be an AGI. Now, you could say, well, maybe we won't have that by 2030, and that's certainly a possibility. But then we are saying, well, they were too aggressive. in their predictions, and they lost the bet of, let's say, AGI by 2030.

And maybe it's going to come by 2040. Maybe it's going to come never. But the whole premise of AGI is that it could do all those things and not just certain parts of your and my job, but it could do all the tasks across the board. One reason why AGI might not immediately destroy a lot of human jobs. is that there are many jobs that have a really significant

physical component to them. I'm talking about driving a car, serving food, as a home health aide, delivering a bedpan or helping an older person get out of bed and get into bed, manufacturing cars,

manufacturing whatever else. There's a ton of jobs that just require a physical component and large language models can be brilliant inside of a data center, but a data center is not going to serve anybody their pasta. In order to get beyond that physical barrier, we're going to need not just a moment of AGI, artificial general intelligence. We're going to need a kind of, you know,

aRI, you know, artificial robotic intelligence or, you know, general robotic intelligence. Is that coming, do you think? Are we also, do you think, at the precipice of a robotics revolution as well? Yeah, that brings us back to Baumol's cost disease. If you make intelligence plentiful and really cheap, then what you haven't automated yet, which is those physical interactions become much more valuable. And you know what? That means it's also becoming much more valuable to automate those. Now, I'll say AI has received all the buzz in the past year.

There have actually also been very significant advances in real. robotics. And I would see they have happened precisely because people have used modern foundation systems, the same kinds of models that are powering AI at large to give robots better brains and make them more capable. And that has kind of flown a bit more under the radar, but may be perhaps equally important as the advances on the pure cognitive domain. So, yeah, I do think robotics is advancing quickly. I think right now we are on a path where we are going to have generally capable AI systems before we have generally capable robots,

but the difference between the two is not going to last long, especially once you have cheap and plentiful AGI, it's going to become economically very valuable for those systems to have the ability to steer robots and to basically extend their value creation to the physical domain. Just so I can envision what you're talking about here, because I know that manufacturing plants are heavily roboticized. And I totally understand that self-driving cars are coming and you could think of that as its own autonomous robot. But when you're talking about a larger robotics revolution, paint me a picture of what that would look like in an average day, like a typical person

going about the world in the 2030s, as this robotics revolution that you're describing is taking off. How does their experience of the world end of life change in this scenario? Let's pick something tangible. Let's take the food services industry where there's like millions of workers. And imagine McDonald's has access to very cheap robots that can flip burgers, assemble them, deliver them to you, and you can go to any fast food joint and have your meal delivered within a much quicker time frame than today without any human in the loop. I think that's plausible within a tech leader. Is it your prediction that AGI and its attending Robotics Revolution

will destroy jobs in a permanent way or a temporary way? Because one scenario here is that if AGI can do the job of all the computer programmers and many of the paralegals and many of the marketing associates, and robotics can do the job of manufacturing workers and fast-workers, and fast food workers and drivers, well, we're talking now about tens of millions of jobs that are disappearing, and maybe one could say they're going to disappear forever.

Another prediction would say, well, in this world where all these jobs are disappearing, they're still creating gross domestic product, which is to say gross domestic income. The money is going somewhere, and that money, unless all of it is being saved, is going to be spent. And when money is spent, it's often spent on people, which means the jobs will be created in some other part of the economy, right?

Maybe we all become therapists and yoga instructors, because in the 2040s, that's all we want to do with other people, is just talk about our problems and do sun salutations. Is it your prediction that AGI destroys jobs in a permanent way or that there's a temporary period of job reallocation that ends in a few years with unemployment still being around 4 or 5%. Yeah, so there's two predictions that I'm very comfortable making. The first one is that the role of labor in our economy

is going to shrink and over time is going to shrink quite significantly. And the second one is that humans will indeed, for a very long time, want humans to perform certain jobs. So now to talk about things like unemployment numbers and so on, I think there is a wide range of possible outcomes. If everything happens very slowly, and if we decide that there's lots and lots of things that we want to be performed by humans,

even though machines could already perform it and could perform it perhaps on some measures better than us, then it's plausible that we'll have, you know, a transition with some disruption, but without, let's say, great depression levels of disruption. In many ways, I hope that our world is going to look that way. I'll also say in the very long term, if machines can do pretty much everything, I hope that, you know, the dream that people have already been talking about for centuries, that we may have to work a little less and enjoy life a little bit more would come through.

I'll also say it is absolutely plausible, though, that we may have very significant disruption, like at the level of the Great Depression with double-digit unemployment, 20% unemployment, I mean, maybe even greater than that, if the technology takes off very fast, if the disruption happens really quickly. And then one of the problems is you made the point before that, well, there's going to be people earning income who will demand the work of other people. If you suddenly have one third of your population who has no income, they can't demand the work of others. more are going to lose their jobs, and you could also have a negative downward spiral through exactly those same forces that you mentioned

that disrupts labor much more significantly. I think in the long term, there's going to be kind of three buckets of jobs in which there will be demand for humans. One of them is for spiritual activities. one of them is for performative activities. Like, let's say you want humans to perform in sports, because if a robot can run much faster than us,

that's not very exciting. And the third one of them is probably going to be to oversee what AI systems are doing. The scenario that you're describing is where my brain begins to break. And I think if I had to describe the way in which my brain breaks. It has a lot to do with the political system and the electoral calendar.

When I think back to the Great Depression, Herbert Hoover was punished for his perceived and real failures to allow the stock market crash to bloom into a national and even international depression. If there's an incumbent party that is seen as overseeing a period of technological change that disemployes 20% of the workforce, the opposition party is going to win on an anti-tech, anti-oligarchy message of shut down the machines.

The machines caused this disemployment, and now it's time to flip the off switch and bring back the golden age of America before there were machines destroying everybody's job. And I wonder how you, thought about the likelihood that the scale of macroeconomic revolution you're describing will essentially cause a political backlash that ends that revolution entirely? It's a big concern of mine. And you know, one of the main reasons why I'm describing this

scenario is because I hope it won't happen and because I hope that the people in power will recognize that once something like that, on the horizon will steer against it. And I also want to say if we do have a really significant backlash because some of this happens, it would be a shame if that backlash prevents us from using AI for all the possible positive uses that it could be employed for. And yeah, one of the great premises of technological problems, is that it makes us a lot wealthier as a society.

And that means that in principle, all of us could be winners. We economists, we call it a parade to improvement. Everybody could be better off if we take care of the losers. And I think it's going to be an important job for our political system to make sure that if we do have technology that can create so much more wealth, that nobody is left behind. Right, but you put your finger, you put your finger precisely

on the psychological and political problem here, which is that nobody wants to be the loser that has to be taken care of, right? Nobody grows up, you know, gets married, starts a family thinking, man, I hope I become an economic loser in a techno-macro-economic circumstance that requires the federal government

to bail me out forever. Like, no, people are the heroes of their own lives, right? They want to be thought of as winners, and they aspire to be winners. And that's why, you know, I have doubts about the way that artificial general intelligence will be developed, whether it's even possible, how to be rolled out. But in a scenario where it's essentially being deployed in the way that you're describing, I think that the emotional and psychological and political backlash to it is just going to be

absolutely immense because what you were describing, while simultaneously the single greatest technological achievement in human history would also be one of the most devastating blows to collective human esteem ever. You're essentially telling the assuas of the U.S. labor force and the human population that they have no place in the economy of the future, unless they essentially want to, you know, lead a yoga flow for four hours a day. And that's a very, very difficult place to put a lot of people. Let's say, look, if this begins to happen by the year 2028, then it's clearly inflecting the election cycle.

What are the ideas that you think need to be on the table in order to contain the kind of macroeconomic and psychological chaos that something like this could bring about? Yeah, I also very much hope that. this kind of disruption won't happen. And I don't think in a democracy we are going to let it happen. I also want to point out, though, that there are going to be very strong competitive forces that will push in the direction of just automating things because there's going to be so much profit on the table.

So this is not going to be an easy challenge to overcome, right? But ultimately, I think we need something like a no American left behind program to make sure that those challenges that you are describing are addressed. And that's the only way in a functioning democracy to have disruptive technological progress without the kind of major backlash that would completely stifle it. Anton Kornak, thank you very much. Thank you. Very much enjoyed the conversation. Thank you for listening. Plain English is produced by Devin Beraldi,

and we are back to our twice-a-week schedule. We'll talk to you soon.