The AI Economist: The Skill You Need to Stay Employed in the Age of AI

I've been studying economics of technology for about a quarter century, and I haven't observed a period where there is this pace of change.

Is artificial intelligence overhyped, or are we fundamentally misunderstanding this technology?

I don't think it's overhyped. I think it's very dangerous to not be using AI. Where we will overestimate is how fast that will happen.

30% of the entire US stock market is in seven companies. If a bubble burst, what happens to the economy? The cognitive work that we've all built careers around can now be done by AI at a fraction of the cost in a fraction of the time.

The skills that gave you dominance before AI may not be the same skills that give a person dominance after AI.

Is the future of work more about skills than it is jobs? And would you say that that is the most important skill for the future of work?

Yes. And yes.

What is this telling us about how this AI disruption in the workforce is going to unfold?

People's jobs will be to AI. So for the last couple years, AI has been everywhere, and all of the major tech companies, they're telling us that this technology, it's going to change everything: jobs, how we live, healthcare. But then we're starting to hear a growing list of critics tell us AI is overhyped. It's not going to live up to its promise, and it's not going to be this transformative technology that we think it's going to be. And then most recently, the study from MIT showed that 95% of companies that have invested in generative AI have gotten zero return. So, is artificial intelligence overhyped, or are we fundamentally misunderstanding the evolution of this technology?

I don't think it's overhyped. I think the impact that this technology will have on society is very hard for us to understand. You misunderstand all the things we'll have to do in order to take advantage of it. You think about how surprised everybody was in November when they first saw ChatGPT, when Deepseek displayed its capability. The stock market moved about a trillion-dollar movement in January of this year, and that was because people were surprised. In that case, we were surprised at its performance relative to its cost. It seems very hard to imagine that it's not going to be transformational to absolutely every nook and cranny of the economy. I suspect where, um, where we will overestimate is how fast that will happen.

So, do you think over the long term this technology could be more transformational than the internet and electricity?

Certainly, um, more than what we've seen so far from the internet. So, in other words, if you look at a a graph of GDP growth over time, let's say since the year zero, it is sort of goes along, um, until about 1750, the beginning of the industrial revolution, and then it just shoots up like this. And that was a result of mechanization. Uh, it completely transformed our output of human civilization, our wealth, and prosperity. And when I think about the transformation and I'm anticipating from AI, it will be much more like the industrial revolution than the internet. So, if you think about the industrial revolution, and you write about this in your book, and if we were to draw some patterns from electricity to AI, you write that 20 years after the invention of electricity, which was one of the most transformational technologies of the 20th century, three comp 3% of companies had adopted it, and they didn't even really see some big economic wins. So, what happened that electricity went on to become this massive transformative technology, but it started for two decades kind of where AI is? Is it hyped? Is it going to live up to its expectation? Can we draw parallels from what we've seen?

That's it's a great question, and it's all around something that in economics we call co-invention. So, in the case of electricity, the original value proposition of electricity is it will, it will reduce the operational cost of a factory. So, for example, you might be running a factory, and let's say you have oil lamps, and someone says, "Hey, you could you be using electricity instead?" And so, and maybe it will shave off, you know, 1 to 2% of your operational costs. Nobody wanted to tear apart their existing factories to bring in electricity. So, that meant the only ones who were willing to try electricity were entrepreneurs building new factories. And even then, most of them said, "No, no, I'm going to just stick with what I know." Um, but a few said, "I'll try electricity." And in the beginning, um, they got a very small productivity lift. And keep in mind, all of this is about productivity.

So, if somebody has, if one person could produce 10 things, and then that same person, because of electricity, could produce 20, we're saying that you could double productivity, just for anybody who missed that economics class.

Yes. Exactly what it is. The mental model, uh, you can have is a factory before electricity would have, let's say, steam or a water wheel outside of the actual building, and that water wheel would turn a long steel shaft that would be inside the factory, and the steel shaft would have wheels on it, and the wheels would have pulleys, and the pulleys would be attached to the machine. So, as the water wheel turned, or the, you know, steam turned, they would turn the shaft, the shaft would turn the wheels, the wheels would pull the pulleys, and the pulleys would power the machine. So, once they brought in electricity, the, the entrepreneurs are walking through their new factory floor with electricity and say, "Wait a minute, why do, why are we still building these factories with these big thick timber columns?" The reason we used to have the big thick timber columns was because we needed them to support the big heavy steel shaft. We don't have the steel shaft anymore, uh, because now we have cables with electricity. So, they could get rid of those big, uh, timber columns. The cost came down. Then, after a while, they said, "Wait a minute. Why are we still building these factories in multi-stories? Because that's expensive to build that way. We used to build it that way because the steel shaft could only be a certain length because they're so heavy, and you needed every 10 ft, you needed a support column. We don't have the shaft anymore." So, they said, "Okay, wait a minute. We can start building these as single-story factories, which are much cheaper to build. Land is cheap." So, they did that. Cost came down again. But now everything's on the same level. We can completely redesign the factory floor and so reorganized the flow of people and materials and machinery. In some cases, you know, uh, productivity in those factories, once they got redesigned the workflow, went up by 3, 400% productivity lift, sometimes 500%. The point is, electricity itself, when everything else stayed the same, only had a very small productivity gain, but then all these other co-inventions that came along. So, in the beginning, the difference in productivity between an electrified factory and a non-electrified factory was very small, but as time went on, and they kept doing all these other co-inventions, the difference got bigger and bigger until the difference was so big that if you were not an electrified factory, you couldn't compete, and you're out of business. And that co-invention process takes time. We are just in the very early innings, the, the very beginning, where people are, for the, you know, just starting to introduce AIs, um, you know, into their, into their businesses. They haven't even begun any of the co-invention process. Uh, and so that's, you know, that's that's why I think we're getting the kind of results we're getting, like you mentioned the MIT study.

Yeah. It isn't making sense to me. It's it's as if it's 1997, and we're evaluating whether the internet is overpromised. Most of the ecosystem hasn't been invented yet. So, of course, any studies aren't going to check out with a positive result. And even then, it's still so hard to understand who are going to be the winners and losers in this world. Maybe you were Amazon, but we couldn't, and we're going to do it right. Or maybe you were the 50% of everybody else that flopped. But what you said was really interesting, that the companies had to build from the ground up with electricity, and those became the winners. So, if you're a company that's just slotting in AI right now and you're expecting that to be the game changer, it's probably not going to work out. But we're probably going to see a crop of companies that we don't even know who they are yet, and they're going to build from the beginning, AI-first, and that's when we're going to see the next generation of companies that we might see those 50, 100x productivity grow from. So, so yes, I agree with all of that, with the exception that, um, I wouldn't rule out some of the existing companies. You know, when the internet came along, you gave that as the example. Uh, it's true that a lot of the winners of that were companies that were born in the internet era, like the Amazons and Google's. Um, but another winner was Apple, and Apple, um, you know, had been around, and so they were able to adapt. Another winner, uh, you know, arguably, although it took some time, is was Microsoft. So, I, I think there's certainly an advantage that AI companies have first, because they don't have all the the the baggage, uh, of things that they have to redesign. Um, but the incumbents do have some advantages, some very significant advantages. They have a customer base already. That's probably the biggest one. And especially for AI, the reason that having an existing customer base is so valuable for AI is because AI is the first tool in history that learns from use. So, having customers that are that use the technology, uh, generate the outputs that are required for feedback loops to, uh, that the AI can learn from. And so, if AI is the first technology that learns from use, and it is, that means a company can't afford to also not be using it right now, because the AI is getting better every time you're using it. So, if you're an organization that says, "This doesn't work," and you write it off, it's a bigger disadvantage over the long term.

Yes, I think it's very dangerous to not be using AI.

And so, and we really don't know who the winners are going to be. It could be new companies that no one's seen. Who's going to be the the Googles of the future? Or it could be who we're looking at right now.

Yes. And there was a really interesting article in 1999. The Wall Street Journal writes this article warning that the internet hype looked a lot like the electricity bubble of 1880. And they were right. The dot-com crash happened the very next year. But something else happened. Both electricity and the internet didn't just meet hype. They exceeded it far beyond even what the most optimistic early adopters thought these technologies would do. So, if AI exceeds its promise, and we it delivers on a lawyer does something in 10 minutes that took them a week, a doctor has superhuman diagnostic accuracy at a fraction of the cost. Aren't we looking at a technology that could fundamentally expand what's possible in the economy and transform economic output in a way we haven't seen before?

Okay. And I really, uh, like the way you phrased that, because where I, what I thought you were going to land that sentence was, if the lawyer can do what used to take him a week in 10 minutes, then are we heading into a world with, uh, almost no lawyers? Like, in other words, wiping out that industry. But you didn't. You talked about it. You instead landed on economic expansion. In economics, we use a term elasticity, or elasticity of demand. And the idea there is that as price falls, the demand for that thing increases. And depending on the slope, um, it either increases a little or a lot. If the demand for legal services stayed exactly the same, then we would need far less lawyers. But if it becomes much cheaper to have legal services, will demand more legal services? And the question is, will it increase so much so that we need more lawyers in the future than we do today? Or does the planet have some finite capacity for how much legal services we need? And the demand elasticity for different types of jobs will be different. And so some things, you will imagine, uh, will need more of them, and some will need less of them. But what's for sure not true is that as the cost comes down to do services, whether they're healthcare services, legal services, teaching services, whatever, that this is not just a zero-sum game. That because the person can do a thing faster, we'll need less of those, um, services.

And you can see it really easily with healthcare. If healthcare drops to a fraction of the cost, more people are going to get healthier. It's not that there's some maxed-out peak where we stop wanting those services. We'll just start to do and expand more things. But something did happen with electricity and and the internet. A bubble did burst. So, even if AI turns out to be the most transformative technology humanity has ever seen in the long run, is it possible that a bubble still bursts in the short run? Because we're seeing a lot of those indicators right now.

Absolutely, that could happen. And, you know, one place that people are are pointing to as a potential bubble bursting is in the enormous amounts of capital that are currently going into data centers. And so, in other words, we're building a lot of a lot of capacity. Those are people making bets. They're making bets that there will be demand for that much both training AI models and what are what's called inference, which is using the AI models to make predictions. Many, many billions of dollars are now being invested around the world in data centers, and some people are raising the question, um, are data centers the new railroads? So, you know, with railroads, the countries that were early in investing in railroads, um, they almost all experienced a financial crash, uh, in the in the beginning, because there wasn't enough use for railroads. Um, but eventually, what we got infrastructure being built out, and over time, you know, we built up cities along the railroads, and hotels, and businesses, and and cargo transportation. Eventually, those railroads, uh, generated a very high return on investment, but it took a long time. And so the question is, is the timing of data center investment getting out ahead of the timing of the the actual use for the reasons that we just described? We talked about, um, with the factories.

And so it's just data centers that could lead to a crash. You don't think people are going to get spooked that the techn that AI is going to take a bit longer than people hope for, and as a result, people start to flee and get antsy? I mean, 30% of the entire US stock market is in seven companies. The Magnificent Seven, I think it's Apple, Amazon, Alphabet, Meta, Microsoft, Tesla, Nvidia. If a bubble bursts, what happens to the economy and to those companies?

First of all, when we think of what does it mean for a bubble to burst? It means people's expectations have changed. So, in other words, the price, the stock price reflects people's expectations of future earnings. So, that means something's changed in people's expectations of what the future will be like. And in the case that you described, um, with those seven, it would have to be that there was some significant change in belief of, uh, that was directly relevant to to those to those seven, as opposed to, for example, um, what the returns would be to the Ford Motor Company, uh, that might be using AI and like putting autonomy in its factories, um, or having AI capabilities in its cars. And so, in terms of just this the speculative capital, my intuition would be there's just a, the enormous delta, the change in capital flows into data centers, seems to be where it's most extreme. Um, but, you know, it, it could happen, and it did happen in January, uh, with the Deepseek. Um, now, that didn't, you know, that seemed to be a very temporary blip, but, um, it's not impossible.

And do you have any estimation as to when we would maybe look at a bubble bursting, or we can't, we can't make those predictions? We don't know.

It's very hard to make those types of estimates. In other words, what you're doing is betting against the general intelligence, uh, of the market. Um, given how, especially right now, because things are changing so fast, you know, I've been studying, uh, economics of technology for about a quarter century, and I haven't observed a period where there is this much, uh, this pace of change. So, and so it makes it very difficult to make bets. But of course, everyone investing has to, you know, every investment is a bet.

And so, would you think it's fair to say, expect some ups and downs? Expect there to be more hype, excitement with the technology. Expect people to also get kind of disillusioned with the technology. But over the long term, this technology is still not overhyped and it will change the game. But expect the ride to not be straight linear.

Yeah, that's exactly what I would say.

And so, let's say all AI progress stopped tomorrow, and there's just no more advancement in the field. Do the current AI systems still hold enough capabilities to fundamentally disrupt the workforce and how we live?

I would say that the, that the capabilities we have today, with no further progress, uh, certainly have the capability of having a very significant impact in every industry. In other words, we've hardly scratched the surface of deploying what we already have. And so then, when you compound that with the fact that the slope of of, um, improvement is so steep, and we get to this point of uncertainty.

Because my fear is that, I mean, GPT-5 came out, and it was underwhelming for a lot of people. And so there was a lot of speculation, we're starting to hit a wall with this version of generative AI, with this architecture, and we're not going to see too much progress. And so I have two concerns. One is that people start to check out from thinking they need to pay attention to this technology, because they see a headline, "Progress is slowing. It's not going to turn out to be what it, what people promised it will be." And then they think they don't need to pay attention for their own job. Or a business thinks, "Yeah, this is overblown. I'm going to turn away." But you're saying, even if we stop today, there is still so much to grab from the current systems that it's still going to disrupt either your job or your company if you're not paying attention.

Yes.

And do you have any bets on AGI? And AGI is artificial general intelligence. So, this is an AI system that could be just as smart as anybody at everything. And that's what companies see as the holy grail. Everybody is positioned towards AGI.

Yeah. So, I don't really understand the line in the sand that people have drawn around AGI. When people say "smarter in every category," I think what they mean is, you know, for example, able to answer technical questions. Like when, um, Grok-4 was released by XAI, Elon Musk's company, one of the key benchmarks that they use was something called HLE, Humanity's Last Exam. And that is predicated on a large number of very deep technical questions, like PhD-level questions from many different fields, chemistry, physics, biology, and so on. And the idea of this benchmark is that an AI that's able to do well on this would be not just smarter than you or I, but smarter even than a Nobel Prize winner. Because a Nobel Prize winner might be able to score reasonably well, let's say if they were Nobel Prize in chemistry, in the chemistry questions, but not necessarily in the physics questions, or the math questions, or the Latin questions. And my view was, okay, let's imagine we had that kind of oracle. What would that do, uh, to us today? In fact, I was sitting with with a couple of my colleagues, and we were saying, like, what if, what if someone creates that genius-level of intelligence? Who would it even replace? Like, in other words, if you go like, pick any large company and say, "How, how many Nobel Prize winners do they hire?" Like, do they employ? Like, most companies employ zero. Few might employ one or two. And think about like a baker, or a, you know, a retail store, or a manufacturer. How many geniuses do they hire? How many do they need? And then we sat there and said, "Well, what organization does employ a whole bunch of PhDs?" And we were thinking, and then we just looked at each other, said, "Wait a minute. It's ours." And ours is arguably quite dysfunctional. Uh, and I don't mean just, you know, ours, like every university. In fact, I would say one of the slowest adopters of AI, in terms of how to effectively use it for teaching and education, has been universities. And so it's not at all obvious that having a bunch of geniuses, um, you know, packed into an organization is going to, you know, all of a sudden transform. So, for me, the, the, uh, AGI line is, um, is not so obvious what the implications are. In other words, it's not obvious that it's more than just a continuation of what we're already doing, like an increasing capability, uh, that can do more and more things. And all the people that have been, most of the people have been talking about AGI have been talking about AGI in a box. Meaning, like, in, um, Dario's, uh, the CEO founder of of Anthropic, has an essay, "Machines of Loving Grace," and he describes achieving this this sort of level of AGI bar that he describes as a country of geniuses in a data center. And the reason he describes it that way is because they can do anything you and I can do on the computer, but they couldn't for this glass of water. Uh, so there's no physical instantiation. And the reason for that is just that, because robotics is so far behind the intelligence that we have, uh, online.

Speaking of jobs, I mean, the disruption is already happening. We're starting to see it in the numbers. Uh, there was a Stanford study released by some of your colleagues called "Canary in the Coal Mine," and it showed that specifically for young workers, so if you're 22 to 25, and you're in an AI-exposed field, so this is finance, marketing, computer science, accounting, audit, there was a 6% decline in their employment. And then there was another study, I think it was called "Seniority Based or Seniority Biased Technological Change from Generative AI," and it showed that for new college grads, there was a 22% decline in employment in the workforce. So, what is this telling? What is this telling us about how this AI disruption in the workforce is going to unfold?

There is a fear that the things that AIs are good at are things that new employees typically did. And so that the first place that AIs are are biting with respect to displacing human workers is at the entry level. There are a few things about that "Canary" paper. Uh, one is that overall employment went up. Two is that although the number of employed for at the youngest age went down, the wages didn't. Now, that's very curious for for a labor economist. When demand falls, wages should fall. So, that raised the question, well, why didn't wages fall? And part of this was potentially what that paper was measuring. It was measuring people working at reasonably like mid and large firms. And so, for example, software developers, if there was an increase in young people going to work for startups rather than the the mid and large-sized companies that were, uh, included in that study, they would fall out of the data set, but they're still being employed. They just, they're being employed in places that aren't being captured in those data. The conclusion that people jumped to, which was, "This was the first evidence of our greatest fear, which was AIs were coming into the workforce, and they were starting at the bottom, and they were going to start working their way up." And maybe maybe that's what's happening, but it's still too early to tell. Like, the fact that that paper got so, like, everyone's talking about it is just, you know, reflection of how little data we have. It's all so new. It's just a very short time window, um, you know, since effectively, like, this whole effect that they're measuring was a couple years before the release of ChatGPT in November '22, and then a couple years after. And so we're still very early. And that's why they call it "Canaries," and they had a question mark at the end of the title, you know, "Canary in the Coal Mine?" Question mark. Is in other words, they were saying, "This isn't definitive." And so I think we shouldn't ignore it, but I would also say it's not conclusive. Um, and it could be that some of those people, um, or, or in fact, many of those, the people that look like they were were missing, uh, or we're dropping out, uh, were going to work somewhere else. Um, and especially young people in software development, that's not too hard to imagine.

So, all of the discussion about software development being not the best career to go into anymore because AI is going to slowly creep into that territory. You would say maybe that's actually overblown, and if somebody has studied computer science and they've just graduated, not to fear because we're not seeing in the data where these people are getting hired, but it's because they're going to be in startups in in places that just aren't in or aren't in the data where we're measuring jobs. So, somebody shouldn't be super fearful if they graduate as a computer science.

Well, all I'm saying is that's a that's a possible explanation. So, it's just too early for us to know. Um, but what I would say to people in software, or any field, also, all AIs are computational statistics that do prediction, and sometimes it feels like there's more like there's, you know, a ghost in the machine, it feels alive because they can communicate in natural language, but all of that natural language is is produced using computational statistics. So, the whole thing, whether you're using AIs for for for language, or using AIs for for vision systems, you're using AIs for control systems to control robots, it's all statistics that does prediction. And what AIs cannot do is they have no judgment. They have zero judgment. And so no matter what field you're in, if you're in software development, in medicine, if you're a a baker, um, or an artist, the thing that we have, the machines don't have, is judgment. And so, just to, uh, give listeners a mental model of judgment, I teach at a business school. Uh, we're sitting in it right now. If you came to the business school 40 years ago, the primary subject that people studied was accounting. And, you know, they would also do finance and marketing and so on, but accounting at that time was the major, the the dominant field. And for accountants, most of what they did all day was arithmetic, adding, subtracting, and so on. In fact, there was a homework assignment that people would get. They say, "Go to your phone book and tear out page 47 and add up all the phone numbers." And the reason that was a homework assignment is people had to practice their adding. Like they would just add up all the numbers, carry the one, like all the that, you know, the longhand form of of of arithmetic. They spent a lot of time building the muscle to be good at arithmetic because 80% of their job was doing arithmetic. Then, all of a sudden, along came spreadsheets. And now it didn't matter if you were sort of mediocre at arithmetic, or you were excellent at arithmetic. The machine was better than everyone. It was superhuman. It was the the arithmetic version of AGI. And you might imagine if these things came along, and now, in an in an instant, the machine can do all this arithmetic perfectly. Never makes a mistake. That we wouldn't need any more accountants, or very few. But yet, there are accountants all over the place. There's so many accountants out there. And the question is, why? Like, what are those people doing, uh, that, given that the spreadsheets can do all this stuff? And the answer is, they're applying judgment. They are no longer doing the addition and subtraction, but they are deciding what numbers should we give to the spreadsheet, and then the spreadsheet does this magic, and then what numbers, like, how should we interpret the output? And so they are applying their judgment. Don't want anything. They have no wants. So, where do they get their direction from? People. A key thing a person needs to do when they set off to to build something is they have to want to build a thing. And then they have to want well, it to be able to articulate what what its capability should be, and how it should function, and how it should interact with it with a person, and so on. And so, and then when it actually does its job, they have to be able to assess it and say, "Okay, you know, I can either make this this, uh, tool work faster, or it can take longer, but give a more detailed answer." And so, like, all those are trade-offs. Judgment. So, the thing that I would encourage every student, whether you're in computer science, uh, you know, software development, or, um, or any field, is, you know, as you're developing your trade, recognizing the difference between what's effectively prediction and what's judgment. And the more we develop, people develop their judgment muscle, the more they will be able to contribute to the, you know, overall production of stuff, um, and work alongside AIs.

So, judgment becomes, you have access to the best supercomputers in the world. What do you ask them? And when they tell you a result, is that good? Or which answer do you even go with? So, you're basically evaluating the output of supercomputers, understanding how to apply that. And those skills will become more important, potentially, than anything that you're doing right now in your job. So, doesn't that also mean somebody who is excelling in marketing today, and they're maybe running the numbers, knowing how to which campaign to pursue in a world with a supercomputer, that might not be the same best person with the strongest judgment to handle how AI comes into the marketing department now? So, there may have been someone that's not as great with the numbers, but they're better at judgment. They're better at understanding trade-offs, and that could be the new person that's best at marketing in the AI age.

Yes, that's exactly that. And and that was one of the key points in our book, "Prediction Machines" and "Power and Prediction," was that the power who has the power in different organizations, uh, may significantly shift for exactly the reason you say that this, that the skills that gave you dominance before AI may not be the same skills that give a person dominance after AI. So, you could even be somebody with great judgment in marketing, and then you could move to finance because you don't really care about the numbers, but you're great at assessing trade-offs. So, you become even better in that department, which means org charts could look really different because the skills change. Do you think so? One, is the future of work more about skills than it is jobs? And two, would you say that that is the most important skill for the future of work?

Yes, and yes. Yes, skills will trump jobs, and yes, in our view, judgment becomes the number one skill.

So, if the future of work is less about jobs and it is more about skills, and you're a new grad, you should be doubling down on building judgment and understanding whatever it is that you studied. Figure out how to direct AI systems in that field. Figure out how to weigh what an AI gives you. Is that good? Could you ask for more? And maybe you actually surpass people that are, because if we go back to that study, the "Canary in the Coal Mine," and you had mentioned that wages did the same, and that employment actually grew in those same departments, or in those same occupations where new hires weren't getting hired at the same rate. And even what was fascinating is that even if you were just 30 years old, you've been in the workforce just a few years, your employment was stable or continued to grow. So, it was just the new hires. But if you've come out of college and you can't get a job at some of those big organizations that...

Or you choose not to get...

Or you choose not to get a job, that's a whole thing, too. That's a different type of a journey. Would you recommend or advise new grads, double down on judgment, start understanding what AI means in your field, and then move towards a startup or start building your own experience? Do kind of put together your own apprenticeship, in a way, and start to patchwork yourself into your career.

It's a great, it's a great question. I think a significant shift in education that we'll start to observe is the difference between learning from reading versus learning from doing. And when I say doing, the key thing I mean by doing is making decisions under uncertainty that have consequences, where the person who's making the decisions, um, owns the outcome. And the reason is that when you own the outcome, then you feel the pain of a bad decision. And so that creates a loop where you make a decision, there's an outcome, the outcome is either good or bad, and you own the outcome. And the reason I think that's very important is because that loop creates judgment, because you start getting far more attentive to the trade-offs, and trade-offs are the essence of of of judgment. And so you can think of judgment as as, um, occurring at two levels. First is in just preferences, like, what do I want? And so, do I want, do I want to build a thing like this, or do I want to build a thing like that? And then, as I'm, um, making decisions, I'm weighing different outcomes. And that weighing of trade-offs is a, is the, the second form of judgment. And so the best, most salient way to develop that muscle is to actually make decisions. And so we have created an education system that's largely about reading. You and I both participated in a reading group, and a member of that reading group is Rich Sutton, who recently, uh, was awarded the Turing Prize. And he recently wrote this essay that we are shifting from an era of data to an era of experience. And his point was that that these AIs, um, in order to get over the next hump in terms of the next level of intelligence, is they can't get that much smarter from simply reading. They have to have experience, meaning take actions that have outcomes that generate feedback. And I think it's the same for people.

So, if you're already in the workforce, what should you be doing today? I mean, even if you think your company hasn't talked about AI, you feel pretty comfortable in your job, what does that mean you need to be building and doing to build that judgment skills? Because you're seeing everything that's in a book or simply on the internet, expect AI to do it, because it's probably read it, and it's read it more in depth and more times, but it hasn't done any of the actual doing, but people have. So, if you're a marketer right now that's been working for 10 years, or in finance, you've been working for 10 years, or or a sales rep, what should you be doing in this moment to start preparing for AIs that will inevitably step into your department?

People will need to become comfortable with with a much higher velocity and propensity to ask "why." You know, if I asked you those in terms of just let's say branding, um, this call, I've got questions, and I'd say, "Why did you pick that?" Um, you know, "Why did you pick that name?" And like, "What were the two other names that you considered for this?" And then you would say, "Well, I picked this one because maybe it appeals to this kind of of demographic, or because..." And then I'd say, "Well, why do you care about that kind of demographic?" And I would be asking you why. And the reason I would keep asking why is because in your answer, you would be implying trade-offs. "Well, I wanted something that would appeal to this kind of person." And then I would say, "Well, that means you're that you're making a trade between this kind of person or that kind of person." Uh, or that you said, "I want these kinds of conversations. I want them to be sort of, you know, the conversations to be a, um, representing the types of questions that my listeners will have." And then I'd say, "Okay, well, that's a trade-off between, you know, having that focus versus a different." But each thing I'll be asking you why. And as you answer my questions, it will force you to be thinking about the trade-offs that you're making. And so you would have answers to all of that. People's jobs will be to direct the AIs by applying their judgment. And their judgment is is reflected in their reasons for why they will will choose one thing over the other. I think that is the most important piece of advice on the future of work because we hear everybody has to learn how to use AI, and that is that AI is going to be like a computer. When you show up at work, we expect that you can operate that. We don't ask that anymore. But what is part two? Post asking an AI question, what are the actual ways that you continue to differentiate and compete in the job market? It's not working with AI, because we all are going to have to do that. It's what you have just summarized. And I think that that's absolutely vital. If we're going to, I want to understand the structure of the knowledge economy itself. Because if you are a lawyer, or you are an inside sales rep, or you work in finance, the modern economy was built on the assumption that the cognitive skills required to do your job are relatively scarce. And now we're seeing AI systems be able to do those same cognitive tasks for pennies. So, what happens to the structure of the knowledge economy and to all of the people in it when the cognitive work they've built careers around, that we've all built careers around, can now be done by AI at a fraction of the cost in a fraction of the time? I'm, I'm going to sort of pull a thread from your first question through to this question, which was your first question was about, is this hype? That if you take healthcare, and the way you and I receive healthcare, like it's really bad in terms of how expensive it is, and the quality of care in many cases, not all, but in many cases, quality care is such enormous room for improvement that it's just hard to fathom how much better it could be relative to what it is now. Uh, like, in other words, I think it'd be much, much, much better. And by better, I mean cost-adjusted better, so that it is not just quality care better, but much cheaper, and therefore much more accessible. So, much would have to change for that to be true. That yes, we will have to reorganize, uh, the way everything works. Yes, we will have to have a, a totally new division of labor between people and machines in order to be able to provide that kind of service at that low of a cost. And so the number one recommendation that I have now for organizations is to create the systems inside their their companies to enable experimentation. Because nobody knows, like when you ask me about the org chart. Yes, it'll be different. If you, if you would, you know, ask me the next "why" that I felt was coming down the line, which is, "Okay, how will it be different?" Answer is, no one knows. So, nobody knows how hospitals are going to be different. Like, you know, people, and I think amongst the best at this are science fiction writers. Um, you know, they are very good at imagining, okay, if we have these technical capabilities, what would the hospital of the future look like? And they, you know, have to paint the picture as a science fiction, right? The one thing we know is it should be drastically different than it is now. And so how do we get there is through a whole series of experiments. Every employee adds value by asking "why." Like, that is their job. Um, and because the machine can keep doing stuff, but it never has a preference. It only does what it's been told to do. So, we aren't, at least in the short term, looking at a bunch of layoffs and a rapid decline in employment, but it almost becomes more on a micro level. So, if you aren't able to build that judgment skill in an organization, and you're not great at working with AI, your job might be at risk. But overall, as the cost of doing cognitive tasks falls, we'll probably just use more of them. So, it might be who's in organizations may change, but companies are still going to be needing people to drive these machines that don't have any desire, and to decide the output, if what the machine gives you is actually good and worthwhile, and how you implement it.

Yes. And when you say, um, cognitive tasks, you, in both in both our books, "Prediction Machines" and "Power and Prediction," we write about two core cognitive tasks: prediction and judgment. And so, while the AIs are are getting better and better at prediction, uh, they have, we have made zero progress on AI having judgment.

And what's an example of a prediction, say, in healthcare or in marketing, that somebody would be able to understand?

So, a a prediction is in healthcare. Let's say I have a a lesion or a mole on my arm. I'm not sure if it's cancer. Uh, I can take a, you know, my phone and take a picture, and an AI evaluates the image and predicts cancer or not cancer, the same way that a doctor would look at and predict cancer or not cancer. Um, except the AI has been trained on millions of images, and the doctor went to medical school and got trained on only thousands of images. And the judgment is, if the doctor says, "Well, we can do this treatment. Here's the benefits. Here's the the risk of the treatment. What do you want to do?" Weighing those trade-offs is judgment. And so, and that might depend on my age. It might depend on my, uh, like how much I like sports, or how much I do this, or how much my lifestyle. And so, um, it's up to me to weigh the trade-offs and make a decision, or it's up to my doc, if my, if my doc is needs to be there with me to help me weigh off the the trades and can ask me questions and and infer from from the things that I'm saying what the right trade-off is for me. And then, if we were to zoom in on an actual, a micro-skill level for a particular occupation. So, if you are a lawyer, or a writer, or even a consultant, being able to write well is a barrier to enter that field. If you can't write well, you're just not even in the running at all. So, now that we have AI systems that can write better than most people, does that mean we're going to see more opportunity for people to become lawyers who can think really well, but they can't necessarily write? Or does the competition in journalism and and legal fields actually become more fierce, because you can no longer lean on being a good writer? You have the the competition moves upstream to how well you can think. So, now it's about how do you think through that case? Are you able to scenario plan and almost wargame what your opponent's going to do? And the thinking becomes more competitive, because the writing, in some ways, has been automated.

Yes. Don't think it'll be any any less competitive. It's just the skill that becomes the ba, the basis for competition, uh, shifts. And the example you gave there, let's say in law and writing, I think is a, it's, it's a very good one for a broader point of redesigning the factory floor, just like we talked about earlier with electricity. You know, there's a cartoon that's gone around the interwebs, um, two-frame cartoon, and in the first frame, somebody says, "Oh, uh, you know, I have these ideas. Is I'm going to use an AI to draft a three, three-page email." And then the second frame is the person said, "Oh, I just got a three-page email. I'm going to use AI to summarize it down to a couple points." And so people look at that.

They and they joke, um, because everyone knows that there's some of that's going on, but it actually hides something that's really foundational for the factory floor. So step one is, you, you just sort of type out your your thoughts. The AI then takes those thoughts and predicts the essence of them. Then you receive the email. Traditionally, you would read the email. Now, if two hours later you went to talk to your producer about that email, you would convey the few points in the email. You could not remember the specific word sequence that they sent you. You would not be able to recite the three pages that they sent you.

And so that begs the question, we currently have a factory floor where it goes from an idea, maybe jumbled, that may be clear in my mind, to then right now, we put into, uh, an AI. Then the AI determines the, you know, predicts the essence, and then based on the essence, it predicts the sequence of words. Then we send the sequence of words. Then an AI reads the sequence of words and then goes back and and summarizes it. A lot of those steps can be collapsed because, in essence, the only thing that you cared about that you then wanted to talk to your producer about were like the key idea. You didn't need all the extra stuff.

And so when we're thinking about these very, the very beginning of this, um, discussion, you asked about the economic impact. The long-term big economic impact is going to be a result of that. It is going to be a redesigning of the factory floor. Whether the factory floor we're talking about is in law and the way we, um, conceive of ideas and communicate the ideas, uh, whether it's in, um, you know, in when we talked about medicine.

If you go to see your doctor, you walk in and a typical, uh, patient-doctor interaction, let's say seven minutes. And in that seven minutes, the doctor asks you some questions that maybe they, she, she puts on her stethoscope. She might, you know, uh, listen to your heart rate or or, you know, do some test, and then your seven minutes is up. And while your doctor's talking, she's taking a few notes. At the end of the day, all the patients have gone home. The doctor sits at her desk and she will fill out her charts. Then they, she's charts. Uh, they probably go to India, and some, uh, people there receive the charts overnight. They read them and then they convert them into reimbursement codes. So they write down the reimbursement codes, uh, of of what happened in that patient-doctor interaction so that the, the doctor in the hospital can get re, reimbursed. Then they send those codes back to the hospital, uh, in the, in the US, and then the hospital then, uh, sends out the reimbursement codes to whoever the payer is, like Blue Cross Blue Shield, Medicare, whoever. And then they send a payment. All of that process originated from data that was created in those seven minutes.

Now, today, we are building AIs. For example, there's a number of companies who are building AI tools for doctors. They say, "Hey, you know, uh, you can use our tool so that the AI can generate your doctor's chart." So rather than sitting there for two hours at the end of your day when your patients have gone home, uh, the AI can do that, and it'll take it from two hours down to fifteen minutes. And then you take those charts and then you'll ship them to India. And then in India, now they've got AIs that will read the charts and will take a task that used to take an hour and make it three minutes to identify what are the reimbursement codes and so on. But they are each AI's just increasing the efficiency of that step, but they're not changing the factory floor.

But you can see that everything's just from that seven-minute interaction that ultimately we will collapse all of that process. And right in those seven minutes where that information is being generated, um, you could imagine the reimbursement occurring at at the minute the patient walks out the door at the end of seven minutes instead of that whole chain. And so, um, those types of jobs that are in the, you know, in a workflow that can be collapsed, uh, they will be reoriented towards, for example, things like audit. Is this like a legitimate claim? And, um, is this following an appropriate process? And so, which will, which is tied to liability. So, you know, who's who's liable for what. But that is a completely different, um, emphasis of skills than the ones that, uh, that that are the basis of the jobs in the current factory floor.

>> So humans are actually going to have to bring more to the table in a world with artificial intelligence because if you, for example, with the writing example with law, the competition moves upstream and it becomes more about the thinking. In a world where you were the person transcribing a doctor's notes, now you're the person deciding who, who could be liable here? Is this claim legitimate? All of the skills actually become a bit more challenging in a world with artificial intelligence. So when people say we're stepping into a world that's going to look like Wall-E and nobody's going to be thinking for themselves, that's actually not true.

>> No, there'll be a lot of thinking. I mean, applying judgment, um, requires a lot of thinking. I suspect there will be a some transition period where there will be some jobs that feel like less thinking in between the time where AIs are very good in the digital world but are very weak in the physical world. Um, so there'll be that window of time where AIs are doing a lot of the sort of sophisticated prediction tasks, and because they're so poor in the physical world, there'll be a lot of jobs of just sort of implementing what AIs want to do. An example of that are Uber drivers, where before you had to know the city. You had to be knowledgeable about the city to drive a taxi. Now you can put your brain on autopilot. Um, but that's just because, uh, we don't, you know, that the that the physical implementation of AIs, uh, is still, you know, far behind. Once that catches up, then we will really be in a world where our job is is judgment. Um, in, you know, everywhere where we've got people, is people applying judgment.

In the examples you gave, it also means that AI could lead to less inequality in the workforce because it means somebody paired with an AI could do higher-order work than they're currently doing today.

>> Well, um, >> depending on, >> it depends on judgment. In other words, >> what we don't know yet is will judgment be more evenly distributed than prediction was, or will it be even more skewed? If it's more skewed, then there'll be even more inequality. If it's more, uh, uniform, then there'll be less inequality, and we just don't know yet.

>> So what does that tell us about what we should be learning in school? So if we can easily see judgment and experience are key for the future of work. What should colleges be doing? How should they be reorganizing themselves? I mean, what conversations are you having here?

>> Trade-offs. Uh, so much has to do with like rather than learning facts, learning trade-offs and therefore always asking why. So, you know, in other words, uh, let's say in history, rather than memorizing the facts, uh, was this, was at this point in history, given these things had just, you know, had happened, was this, was a better decision to do this or do that? Why? And every time you ask the question why, then you're forced to think about the trade-offs. Everything you need an answer for because that is how, um, the role that we'll play in guiding the AI. And so it feels very much like we're heading into a world where the reason we have to, um, be so good at why is because the AI does all the work sort of up to that point and then and then stops. So, we're all going to become executives, miniature chief executive officers of a bunch of AI systems in the same way you would ask your team, why did you pick that? Why, why is this your presentation? We all have to do that for each and everything we do.

So, when you think about college, then, how easy is it for a college to be redesigned this way? I mean, is college still worthwhile? But how and what college teaches needs to fundamentally be reorganized, or are we going to get to a point where it's not going to be the investment that it once was?

>> Yeah, it's a great question. So definitely it needs to be reorganized. Uh, that feels, uh, for sure true. What's less obvious is is the current incarnation of university the best way to teach this skill? We don't know. We don't know whether, um, whether it's the best way. Uh, and one of the reasons is if the best way to learn judgment is by actually making real decisions that have consequences where you own the outcome, universities aren't aren't the best designed for that. Um, so, but we don't know, we don't know what the best way is of of teaching judgment. Uh, I, you know, I'm putting this out there as a, um, as an one of many ideas of how we teach judgment is through is through active decision-making. Um, but what I will say is that universities did adapt in the case of accounting. So in accounting, if you would have said fifty years ago to an accountant, your main part of your job is always going to be asking why, accountants would have said that's absurd. Um, but now that's all they do because the machine does all the arithmetic. And

>> We've talked a lot about the cognitive workforce, but a few years ago you were on the Sanctuary podcast and you asked a question, and I actually want to rephrase it back to you to hear your answer. So, you said humanoid robots will become the largest market in history. But what will be the triggering event that will happen that will cause the penny to drop so that the rest of the world sees that too? How would you answer that? What will be that mark?

>> Yeah. So, so first of all, for listeners, Sanctuary is based in Canada, in Vancouver, makes, uh, is was one of the first companies in the world to focus on humanoid robots, general, what they call general purpose robots with humanlike intelligence. And then a few years later, after Sanctuary began working on this, um, Elon began his work on the Optimus project, and Elon would tell Tesla shareholders, this is going to be the biggest market in the world. It will dwarf the automobile market. Um, and everyone, you find everyone finds that very, just Wall Street seems does not process it. Um, in other words, you know, when you introduce Optimus, it didn't do anything to the share price, uh, of Tesla because it just, it feels, um, so science fiction. When will the penny drop? I think the penny will drop when there is a first implementation of a generalized capability in the physical world that feels similar to what people experienced in December of '22 with ChatGPT. There was something about ChatGPT because it was general that you could ask it anything, and you know, it would make mistakes and it would, but it had a reasonable attempt at anything, and at that point, people sort of started to realize, wait a minute, this feels like, uh, we've entered a different category of thing. Right now, the humanoids are very limited in their in their capability. In the videos that you see online today, uh, they do extremely narrow tasks. So people watch them, they're curious, they're interested, but until they, you see one, uh, that's like in a room like this that is able to just do a bunch of different things on command and you haven't been given a menu in advance. So if they say Ocean Aid, you can ask it to do these eleven things, you'll ask it, but I don't think you'll be that impressed. It's when you ask it to do a thing that has not been told to you and it just goes and does it, that all of a sudden you think, "Wait a minute." I like, we've, we've entered a new category. And so that's, that's when I think the penny will drop, and it has to do with generality.

>> And you're confident that moment's coming. So when people look at robots and they think it's taking five minutes to pick up the coffee mug, who's hiring this thing? You're saying no, no, no. There will be a ChatGPT moment in human robotics and everything is going to change.

>> Yes, there absolutely will be. I mean, at this point, it's just, it's now turning the crank. It's engineering. It's getting, it's just making, in other words, there's already a base level capabilities, and it's, and it's just the physical part of it is slow, and that's just engineering. It's just going to get, uh, faster. The fingers will get more dextrous. Uh, they'll have more degrees of freedom, and their cognitive capacity, like their library of things they can do. You know, the way that we do language is we predict the most likely, the, so-called next best token, the next best word when you're forming a sentence or a paragraph. Um, and so you generate a sentence or a paragraph by predicting word after word after word. And when you and I receive it, it feels like a human-generated sentence. There's all these things we have, all these jobs we do with so many different physical tasks, but those physical tasks are like, um, think of them like a book. There's so many different books and essays and news articles and blog posts, but all of those blog posts and books are a resequencing of effectively thirty letters and characters, punctuation marks. So there's a very small number of symbols that that can just be resequenced to produce all these different books. What's happening in robotics is they're training robots to do a small number of verbs. Pick up, place, speak, look at, read, so on. Um, and once you train those robots to do those verbs, then it becomes a job of predicting what's the right sequence of verbs to complete the task. And there's so many issues of implementing that in the physical world that each one is a very significant job, you know, to to overcome, but the process is, oops, the process has been, um, you know, is underway, and I'm sure we will hit a series of of, you know, unexpected snags as we go, but I can't imagine that this is not going to happen.

>> And how long have you have, where to put a time frame on it?

>> I would say that we will have a, I, what I will call a factory-grade general robot. So let's call it humanoid, um, in ten years. And and by factory, what I mean is it's able to do a, a general, a wide range of tasks, but in a controlled environment. And then I would say we will have a, um, generally capable robot for an uncontrolled environment, like walking down the street or in your house, in twenty years.

>> And that's the robot that can do any physical job a human can do that we see today. It could also do.

>> I wouldn't say any, but I would say a lot. And so do you think if we're to, >> And again, it doesn't have judgment.

>> Okay. >> It doesn't have judgment. >> We're still, we're, we're still, whether it's a physical robot, it's an oracle in a computer, we are still guiding these systems. We are the ones with the desires, the wants, the evaluation capabilities. And so my final question is, is there a potential future at all where we could be looking at a post-work world? Do you, as an economist, think that's not out of the equation? We don't know when. We don't know exactly what that would look like, but I can't say with one hundred percent certainty humans will always be the main entity in the workforce. So there is a, um, a venture firm called Bloomberg Beta. And, uh, we, we, we were hosting a a conference a few years ago, and one of the partners there, a fellow named Roy Bahat, um, and I, I asked a question like yours, and we were talking about the future of work and so on. And he said, "We're already there." So he would point to you and I sitting having this conversation, and he would have said, "Do you think in 1915 anyone would have called what you and I are doing right now work?" Um, he would have made the point that, uh, we are in a post-work, um, and it's just, we, we don't notice it, and it's just such, it's sort of a long arc transition. And so, you know, one of the things I think about is, um, did you watch that TV series Downton Abbey?

>> Okay. So, in Downton Abbey, you know, there's the Lord Grantham and the family that lives upstairs, and then the servants, uh, you know, who, uh, work downstairs. And as the series goes along, they're, you know, they're living their lives and they're doing their things. You know, they don't work. The Lord Grantham and his family, none of them work, but it doesn't mean that their life isn't easy, and it doesn't mean they don't compete. They are, they are, they are competing for other things. So, in other words, there's always something scarce. They're competing for status. They're competing for affection. They're competing for, um, you know, power, uh, amongst, uh, you know, their class of people. Uh, they're, they're competing for, uh, recognition amongst the the charitable class and so on. So they're not being paid for it, but they're competing. And there will always be scarcity. And as long as there's scarcity, there will always be something that we would view as, um, as work. And, you know, a distinction here is, uh, are we paid for it? And do we need to be paid for it in order to to do it? But I don't think that, uh, you know, even if AIs are able to do all kinds of things, that we'll all be sitting around with nothing to do. Uh, we will always be in some form, um, working towards whatever our objectives, our goals are, and competing for scarce resources.

>> AJ, it has been a pleasure. Thank you so much.

>> Thanks, Shenade.

>> Thanks so much for joining us for this episode of I've Got Questions. If you've got questions, we'd love to hear them. Send us a message on our website. And if you found this episode interesting, we would love for you to subscribe to the channel and share it with someone you think may also like it. All right, we'll see you next time. I've Got Questions was created by me, Chenade Bal. The show is produced and edited by Tara Cuts and Sandra Itinan, and executive produced by Pa Piers Torres. Artwork by Corey Vincent at Field Studio.