21 Jul 2026 · 35m · Erik Brynjolfsson, Sam Ransbotham · 26 quoted facts
Stanford economist Erik Brynjolfsson argues that AI's results depend on human choices, not on the technology alone. The better question is what we will do with AI. He says organizations, skills and institutions are the real bottleneck, and that the value comes from new products and ways of working, not just cutting headcount.
The tension is between fast technology and slow institutions. Productivity has barely risen so far, which he explains with a J curve: costly reinvention comes first, and gains arrive later, as with electricity. Early data shows young workers in AI-exposed jobs losing employment, but those who use AI to augment their work gaining. He closes by calling himself a mindful optimist and warning about concentrated wealth and power.
Brynjolfsson says AI capability has outpaced expectations, but businesses and political institutions have changed far more slowly than he expected. Productivity growth is barely above where it was before the AI boom.
His earlier research found that spending on new processes, skills and business models is about ten times the direct spending on computer technology. Because these assets are invisible in standard statistics, companies and economists underestimate how much is needed.
ADP payroll data shows workers aged 22 to 25 in the most AI-exposed occupations have seen a 16 to 17 percent relative fall in employment. Older workers and those in unexposed jobs like home health aides did not.
In the payroll study, groups using AI mainly to automate work saw falling employment, while those using it to learn new skills saw growth. He warns that measuring AI only by headcount cuts, the 'Turing trap', misses new products and quality gains.
Factories saw almost no productivity gain for up to thirty years of electrification, until they redesigned layouts around separate motors. He expects AI to be faster, but still to follow a J curve of cost first and payoff later.
Brynjolfsson says it depends on how AI is used. In his team's payroll data, young workers in jobs where AI mainly automates tasks lost employment, while those using it to learn new skills saw employment grow.
He points to a 'J curve': companies must first spend on new processes, skills and business models, which lowers measured productivity before gains appear. He compares it to electricity, where factories waited up to thirty years for payoffs.
It is Brynjolfsson's name for using AI only to imitate and replace human workers, as the Turing test frames it. He says this leads to narrow cost-cutting, such as a CFO demanding headcount reductions, and misses chances to create new products and better quality.
Brynjolfsson says a radiologist does about 26 distinct tasks, and image reading is only one of them. As AI made reading images cheaper and better, radiologists became more valuable for the rest, and employment grew.
Brynjolfsson says it means expecting a good future but actively working to build it, rather than assuming it will happen or that doom is inevitable. He describes AI as an 'amplifier of intention' that needs human agency behind it.
Stated by the speakers. Each quote matches the transcript, which shows what was said, not that it is true.
In 2012, Erik Brynjolfsson and Andrew McAfee rode in a Google self-driving car from Mountain View to San Francisco, after writing Race Against the Machine.
“Actually, it was 2012, right after we wrote Race Against the Machine, we rode from Mountain View up to San Francisco.”
Erik Brynjolfsson states that the accident rate of the self-driving cars he rides is 90% lower than that of human drivers.
“The accident rate is 90% lower than with humans.”
Erik Brynjolfsson reports that his earlier research found investment in intangible complementary assets was about ten times larger than direct investment in computer technology.
“we found that the investments in those intangible assets are about ten times bigger than the direct investments in the computer technology.”
ADP, described as the world's largest payroll processor, shared its payroll data with the Stanford Digital Economy Lab for the Canaries in the Coal Mine study.
“ADP, which is the world's largest payroll processor, they shared the data with us here at the Digital Economy Lab”
Daniel Rock and colleagues created a taxonomy ranking about 18,000 distinct tasks by their exposure to large language models.
“have created a taxonomy where they rank all of the tasks in the economy, you know, 18,000 distinct tasks by how much they're exposed to LLMs”
The Canaries in the Coal Mine researchers examined about 750 occupations and divided them into quintiles by AI exposure.
“So we looked at about 750 occupations, we divided them into quintiles.”
Erik Brynjolfsson says the youngest workers in the most AI-exposed quintile of occupations show an employment decline of about 16 to 17 percent relative to other occupations.
“now it's more like a sixteen to 17% employment decline relative to the other occupations”
In the study, workers using AI mainly to automate work had falling employment, while those using it to augment and learn new skills had growing employment.
“The automating group had falling employment, and the augmenting group, the ones who were learning new skills, had growing employment.”
Geoffrey Hinton predicted that AI would replace radiologists because they read medical images.
“Jeff Hinton famously said, Hey, you know, reading medical images, that's what radiologists do, that's going to be replaced by AI.”
Erik Brynjolfsson says that as AI made image reading cheaper and better, radiologists became more valuable for their other tasks and net radiologist employment grew.
“as reading images got cheaper and better because of AI, it actually increased the value of radiologists for doing those other tasks, and net employment grew for radiologists”
Jevons paradox describes cases where making something cheaper increases demand enough that employment rises even as efficiency improves.
“some people call it Jevons paradox, where making something cheaper increases the demand for it, and you end up having more employment rather than less, even as it gets more efficient.”
Erik Brynjolfsson says at least half of goods and services in the economy have elastic demand, where small price drops produce large increases in quantity demanded.
“And at least half the goods and services in the economy are on that side of the curve.”
Paul David and others documented that factories saw essentially no productivity gain for up to thirty years while electrifying.
“Paul David and others documented this very carefully that for literally up to thirty years, there was essentially no productivity gain in factories as they were electrifying”
Factories eventually redesigned around separate electric motors for each machine and laid out by the flow of materials, rather than clustering around one central steam engine motor, after which productivity rose sharply.
“Instead of being clustered around a central motor like a steam engine, they had separate motors for each piece of equipment, and they distributed the work and they laid out the arrangement based on the flow of materials.”
Erik Brynjolfsson's company WorkHelix helps companies identify opportunities for value creation with AI.
“My company, WorkHelix, is very involved in showing companies how to identify the opportunities for value creation.”
Erik Brynjolfsson recounts that a CFO told him she required every division to report headcount reduction from AI as its return on investment.
“that's why I'm demanding that every division tell me how much headcount reduction they're getting from AI”
Erik Brynjolfsson calls the focus on using AI only to replace or imitate workers the 'Turing trap', after Alan Turing's imitation test.
“I call it the Turing trap because I think that it's a trap to only focus on using AI to replace or imitate workers.”
Sam Ransbotham cites a Science article associated with Neil Thompson saying about 90% of the most important new AI models now come from industry.
“they've got a science article that said something like 90% of the most important new AI models are coming out of industry”
Scaling laws for large language models mean that more data, compute and parameters produce predictable performance gains, which has driven training costs into the billions.
“there are these scaling laws for LLMs where as you make a model bigger, if you have a model have more data, more compute, more parameters, you get a predictable improvement in performance.”
Erik Brynjolfsson says Anthropic and OpenAI are filing to go public at the time of the recording.
“Anthropic and OpenAI are filing to go public”
When asked what hardware he uses for his discoveries, Geoffrey Hinton reportedly pointed to his laptop.
“he pointed to his laptop and tapped it and said, This is this is the hardware I use”
Erik Brynjolfsson started his first AI company, an expert systems firm, in the late 1980s.
“the first AI company I started, I started back, believe it or not, in the late nineteen eighties.”
Sam Ransbotham says the US government put roughly a billion and a half dollars into non-defense AI research spending last year, about what Google spends.
“put like a billion and a half dollars or so into non-defense research spending last year, and that's roughly what Google is putting in”
Erik Brynjolfsson has a long bet with economist Robert Gordon on how fast productivity will grow because of technology.
“I have a bet, a long bet on that with Bob Gordon, and I've written about, you know, how I expect productivity to grow”
Erik Brynjolfsson says productivity growth is no faster than at the start of the AI revolution, perhaps slightly higher.
“productivity is not growing any faster than it was at the beginning of this AI revolution, maybe a smidge if you kind of squint”
Erik Brynjolfsson spent over twenty-five years at MIT before moving to Stanford in Silicon Valley.
“I loved my time at MIT. I was there for over twenty five years, but now I'm out here in Silicon Valley”
Brynjolfsson explains the lab's mission of closing the gap between fast technology and slower economic understanding and institutions.
“technology is advancing very rapidly, the capabilities are amazing.”
He rejects the framing of AI as something that happens to people and stresses human choice and values.
“It's what are we going to do with AI.”
He was too optimistic on self-driving cars, underestimated language models, and was disappointed by how slowly institutions adapt.
“I would have considered this AGI if you had asked me in 2012 or 2015.”
Payroll data reveals a sharp employment drop for young workers in the most AI-exposed occupations.
“early career workers, they had a noticeable fall in employment”
Jobs are bundles of tasks, and cheaper services can raise demand and employment, as in radiology.
“making something cheaper increases the demand for it”
General purpose technologies need costly complementary reinvention, so productivity gains arrive late.
“it took about twenty to thirty years”
Focusing on replacing workers and cutting headcount misses larger value from new products and better quality.
“it's a trap to only focus on using AI to replace or imitate workers”
Frontier models cost too much for universities, though basic research still needs academic and government support.
“No university can afford that, no nonprofit organization can afford that.”
He argues the future is not predetermined and urges attention to values and to concentration of wealth and power.
“AI is the greatest amplifier of intention ever.”
Today's guest has a bold provocation. AI isn't being held back by the technology, it's being held back by us. Curious how? Find out now.
I'm Erik Brynjolfsson at Stanford, and you're listening to Me, Myself, and AI.
Welcome to Me, Myself and AI, a podcast from MIT Sloan Management Review, exploring the future of artificial intelligence. I'm Sam Ransbotham, professor of analytics at Boston College. I've been researching data analytics and AI at MIT Sloan since 2014, with research articles, annual industry reports, case studies, and now thirteen seasons of podcast episodes. In each episode, corporate leaders, cutting-edge researchers, and AI policymakers join us to break down what separates AI hype from AI success. Today I'm talking with Erik Brynjolfsson, who runs a Stanford Digital Economy Lab and studies how people use technology in general, and now AI specifically.
He's trying to think about how technology is affecting economy and work. His book, The Second Machine Age, shaped a lot of this debate, and his team recently published Canaries in the Coal Mine, a research paper about what's happening to entry-level workers. Eric and I bump into each other a few times a year at the National Bureau of Economic Research, and he always brings up something I hadn't considered. No pressure, Eric, but I'm expecting the same today.
Okay, good to be here, Sam.
Welcome to the show. Let's start with the Stanford Digital Economy Lab. Can you give us a quick view of what the lab does?
Sure. Well, we study the digital economy. So I came out to Stanford and, you know, I loved my time at MIT. I was there for over twenty five years, but now I'm out here in Silicon Valley, the epicenter of the AI revolution. We're focusing on how AI and other digital technologies are changing the economy. Kind of the premise of the lab and of my work, my career, is that technology is advancing very rapidly, the capabilities are amazing.
At the same time, our economic understanding is not advancing nearly fast enough, our economic institutions, skills, organizations, they aren't keeping up. And my job is to try to close that gap.
It's those pesky people, the technology moves fast, organizations and people slow us down, I guess, is the summary there.
That's right, a lot of your work has has highlighted that too, and and we're doing what we can to keep up.
So looking back, I was thinking back on the race against the machine book,
That's right.
and you think about canaries today, what connects all of it?
Well, I think a lot of it, like the race against the machine book, that was the headline. The conclusion was that we should race with machines, not against machines. And that's what I've continued to emphasize is there's an opportunity for humans and machines to work together. It's not automatic. There's a lot of choices that we need to make, but the technology is enabling all sorts of new possibilities and trying to invent and discover those is a big part of what we humans need to do.
Actually, I like the choice part because I think it's so easy for us to slip into the vernacular of saying AI does this, technology does that, and I know you've been against that.
That's exactly right. I mean, it's probably one of the most common questions I get is, you know, what's AI gonna do to us? What's AI gonna, you know, make happen next? And I'm like, wait a minute, the premise of your question is wrong. It's what are we going to do with AI? This is an incredibly powerful tool, arguably the most powerful tool that we humans have ever had. And almost by definition, that means we have more ability to change the world than we ever had before.
So let's think about how we want to use that. It turns out that our our values, our choices matter more now than they did in the past because we do have this ability, you know, to make a really big dent in the universe for better or worse.
Yeah, man, for better or worse, I think that'll probably come up a couple of times as we talk here. You know, I was thinking about how much you got right in the early Race Against the Machine book, but that's not interesting as an academic. What do you think you got wrong? What were you surprised about? What's changed differently than you thought sixteen years ago or fifteen years ago?
Well, first off, since we're talking about what we got wrong, I should highlight that this is co-written with Andy McAfee.
All right, yeah.
So along with with Second Machine Age. So he's been a partner in a lot of my projects together. You know, so what do we get wrong? We were looking at some advances. I I'd highlight particularly in the second machine age. We started that with a ride in Google's self-driving car. Actually, it was 2012, right after we wrote Race Against the Machine, we rode from Mountain View up to San Francisco. And I gotta say, I thought, Okay, wow, you know, self-driving cars are just around the corner, you know, uh, metaphorically.
And I thought it would happen very quickly. It's obviously taken longer. I do ride around in self-driving cars a fair amount here in the San Francisco Bay Area, up in the city. They've had them for a while, and now they've come down to, uh, Palo Alto as well, and you can ride them all the way on the highway. But that's, you know, fifteen years later, so it's taken a while. And I think I, um, I was over optimistic about that.
That said, I think I underestimated how fast the technology would advance in other ways. You know, the way we you and I can talk to LLMs and, you know, whether it's ChatGPT or Gemini or or Claude and have them do work. I would not have expected that to have happened that quickly. I mean, basically, I would have considered this AGI if you had asked me in 2012 or 2015. So that's pretty cool that you can get really good advice from them.
We all know they still hallucinate make mistakes, but so do we humans. On average, they're really quite good. So that that happened faster, we got that wrong a little bit. And then the most disappointing part is, you know, look, I knew we were just talking about how, you know, human institutions change more slowly than technology, but oh my god, I didn't expect it to be this much slower. Um, and at times it's almost like it's moving backwards.
And so that's been pretty disappointing that our political institutions, our businesses, organizations, they aren't keeping up. You know, productivity is not growing any faster than it was at the beginning of this AI revolution, maybe a smidge if you kind of squint, but we're not really translating these capabilities into better business performance or the kinds of benefits that we all hope for.
I rode in one of the automated cars out in Palo Alto at the last meeting where I saw you at. It very quickly got boring. Now, I was fascinated, you know, for a few seconds and then bored very quickly. You touched on that would have been AGI a few years ago, or LLM stuff would have been AGI a few years ago, but it's not anymore. How much of this is sort of a boil frog? We're getting used to these technologies.
It takes even more to impress us. Is that the part of it? Is it a measurement part? Is it a J curve, which we'll come back to in a second, when we're all the above, none of the above?
all of the above. Part of it is, yeah, w- we, like you said, with the getting bored in the self-driving car, you know, the first few minutes are kind of scary and thrilling, and then it's like, exciting. And then, you know, within half an hour, you're like, Okay, I kind of get this. You know, these cars, they kind of drive like my grandmother, like very carefully and slowly, you know, which is, I guess, the right thing
That's great.
to do. Yeah, no, exactly. That's why they're accident
That's why you live in Boston, maybe.
rate. Exactly. The accident rate is 90% lower than with humans. And so I'm looking forward to the day that that, you know, the cars are all like that. Right now, they are a little, little slower, point to point. Sometimes when I'm in a rush, I I'll actually get an Uber because I know that the driver's more willing to to cut some corners, so to speak. But part of it is that, yeah, we just take for granted things that initially were eye popping.
I guess our brains are designed to get used to things. But the other two things that you said, I think are more important and more interesting. One is the measurement issue. Our economic metrics are just not designed to capture a lot of the benefits of AI, especially the benefits of free digital goods. So, you know, we've got a whole methodology for addressing that. You know, maybe we can talk about that later. Uh, and the other part is is the J curve, that there really is, and this is what we started the conversation about, there really is a gap and a difficulty in translating capabilities into real business change.
And it's just because there's so many complementary assets and investments that need to be made, many of them have to be invented, and we don't know what they are. These complementary, you know, co-inventions of new business processes, new skills, whole new business models and ways of working, they're hard. And we have, you know, entrepreneurs and managers and business school professors and consultants are all working on figuring them out. In earlier work I did with, you know, with Lauren Hit and Sunny Tambay and others, we found that the investments in those intangible assets are about ten times bigger than the direct investments in the computer technology.
But since they're largely intangible, largely invisible, we tend to not appreciate them as much and we don't measure them and and people don't even realize that they're there. But we're trying to make them more visible in our research and create an appreciation that you have to make those investments as well.
Yeah, actually, the measurement thing is huge and difficult. Remember, you and Avi Collis were talking about the measurement of value from Wikipedia and from Facebook and these other technologies. All this stuff is just super, super hard to measure, and when we don't measure stuff, we we just do a bad job, I think. But let's switch maybe to measurement because I think one of the cool stories about your canary in the coal mine study, I mean, I think your headline finding was that people aged twenty two to twenty five were the most exposed, lost about 16% of their employment.
But you're using ADP payroll data to find that. That's a very interesting signal of information. Tell us a bit about that study.
sure. So, you know, like everybody, we were seeing all these headlines about AI eliminating jobs or other headlines, AI is creating jobs. And, you know, any given month, there's hundreds of thousands of people who gain or lose jobs. So, a half-decent journalist can easily find examples of either one and and build a narrative around it, or just a a man in the street they can talk to. So you don't really to me, at least, I don't really know how to aggregate that, like, you know, what's the real story?
So we wanted to get large-scale data. And when we looked at the current population survey, you know, we didn't really see much. And when we looked at the top line of ADP, which is the world's largest payroll processor, they shared the data with us here at the Digital Economy Lab. You know, we actually didn't see much in the top line either. And I was almost thinking, okay, we'll write a story about how this is all just a bunch of, you know, hype.
But then we dove in a little bit more deeply, and some of the subgroups who are having these very large effects, like you just mentioned, people who are age twenty two to twenty five, early career workers, they had a noticeable fall in employment. But it really got striking when you subdivided it by how exposed they were to AI. So, uh, Elondu, Daniel Rock, and others have created a taxonomy where they rank all of the tasks in the economy, you know, 18,000 distinct tasks by how much they're exposed to LLMs.
Like, could an LLM help you write a memo? Oh, yeah, for sure. Could an LLM help you lift a box? You know, not really. And so if you rank every task like that and then aggregate them up to occupations, you basically have a score of whether or not this occupation is going to be affected by LLMs. So we looked at about 750 occupations, we divided them into quintiles. The most exposed quintile, you know, a little over a hundred occupations in this youngest age group had at first it was about a twelve to 13% decline, now it's more like a sixteen to 17% employment decline relative to the other occupations.
So it's really quite striking. We didn't see a fall in employment for the older workers, even in the exposed occupation. So there's, you know, you really have to slice it this way. And even when we controlled for other factors, like we controlled for interest rates, we controlled for the tech industry, we just can take the entire tech industry out if you want, or controlled for tech over-hiring. You can control for remote work, uh, you can control for education.
None of those things knocked out the core result. We continue to have this very noticeable decline. I could just say another thing, which is that, you know, as I mentioned, the older workers were doing okay. There are a couple of other groups that were doing okay. One is the people in the not exposed occupations, you know? So if you look at the other end of the spectrum, like home health aides, where, uh, AI is not affecting them nearly as much, they actually had growing employment, even for the young workers.
And then the most interesting result, the one I'm most excited about, was if you divide the way they're using AI into mainly automating and eliminating work versus augmenting and creating new skills, new opportunities, doing new things that you never did before. The automating group had falling employment, and the augmenting group, the ones who were learning new skills, had growing employment. So it was kind of a double win, higher productivity and growing employment. And for some reason, that doesn't get as picked up as much in the press when they write about our research.
But I think it's really important that we found that AI could be associated with higher employment for certain kinds of workers.
Yeah, actually, and I think that ties back to your how you use it, your decision making, your choices that we all make in that. You know, that doesn't surprise me, because if I think about a task and going back into this aggregation of task you alluded to, no job, I'm sure, runs 100% down the list of task and 100% down the list of non-automatable tasks. And the composition of that task is going to leave parts of those jobs more valuable, that's going to leave those parts more valuable.
Exactly. That's another really important point is that, you know, you can take every occupation, it's a bundle of tasks in O NET, like, you know, typical occupation has twenty to thirty tasks, like a radiologist has twenty six distinct tasks. You know, Jeff Hinton famously said, Hey, you know, reading medical images, that's what radiologists do, that's going to be replaced by AI. What he didn't factor in was there were twenty five other tasks that radiologists do, and most of them were not affected by AI.
And the net effect was that as reading images got cheaper and better because of AI, it actually increased the value of radiologists for doing those other tasks, and net employment grew for radiologists. And, you know, I'm not saying that always happens, but there's definitely cases where some people call it Jevons paradox, where making something cheaper increases the demand for it, and you end up having more employment rather than less, even as it gets more efficient.
Yeah, I mean, that's so I think it's the frustrating thing. I think everyone would like an answer, which is, oh, yeah, here's the clean answer. And what you've come across here is just a a gigantic it depends, and it depends on an increasingly complicated set of stuff.
Well, let me try and make it a little simpler. I mean, it is somewhat more complicated than the simple story of just, you know, you always eliminate jobs, or for that matter, that it always creates jobs. But as an economist, a really useful tool for me, and I'll get a little wonky for your readers, but it'll be useful time, is you can think of a demand curve. You know, when price falls, quantity increases. Demand curves are downward sloping, and so most of us intuitively know you make something cheaper, people will buy more of it.
But what really matters is how steep that demand curve. For some products, like, uh, I don't know, apples, you know, even if you made them ten times cheaper, you're not gonna buy ten times more apples. You know, maybe you buy a few more apples, but the demand curve is pretty steep. So as prices go down, you end up spending less. It's pretty intuitive, and most people think that that all goods are like that. It turns out that not all goods are like that.
There are lots of goods where as the price goes down, even just a little bit, the quantity goes up a lot. Like jet travel, you know, it used to be that very few people would fly in planes across the country, but now it's cheaper, and, you know, a lot of us fly around the country way too much. So there are goods and services like that. That's called elastic demand, or a flat demand curve. Whenever you have situations like that, you actually have growing demand for something, and you end up having more spending.
And at least half the goods and services in the economy are on that side of the curve. And that's really good news. It means that as we get more productive, more efficient, yes, some things shrink, but other things grow. I think there's kind of a a little bit too much of a bias in the public conversation towards the side that's shrinking. And, you know, one of the things I want to do with with your podcast is get more people to think about that side that's growing.
I teach a master class, which, you know, we basically show people how to identify and discover these new opportunities for wealth creation, how you can use AI to do new sorts of things, to create entrepreneurial ventures. And the more we do that, I think the better off people are going to be. They're going to help themselves, and they're going to help the rest of the economy by inventing new things.
Yeah, and that makes, you know, Apple is a great example. You know, like you say, we we quickly don't consume them when we get more of them. But, uh, I don't know of any company that says, Oh, no, we're pretty much all the IT projects that we could ever think about, we're doing them. You know what I mean? I mean,
exactly. There's it's like it seems like there's almost infinite demand for, you know, solving all sorts of IT problems. And, you know, when I look around the world, I don't see a shortage of problem, you know, poverty, the environment, healthcare, especially. These are all things where we could use a lot more resources, and, uh, I don't see us running out of opportunities to address those.
yeah, the fantasy that we would run out of those types of problems, uh, that's great, but I don't think that's gonna happen in the next the next couple of years, though.
Next couple centuries.
Um, let's go back to this J curve idea. First, explain a little bit about what is the J curve and, you know, how should people think about that?
This is another concept I think people find very useful, and that is whenever there's an important, powerful technology, we economists call them a general purpose technologies. You know, we used to just say GPT, but my AI friends, they stole They got that acronym. Yeah,
transformers, yeah.
generative pre-trained transformer, but AI is both. It is a generative pre-trained transformer, you know, like GPT five, but it's also a general purpose technology like the steam engine, electricity, you know, the internal combustion engine and cars. Those are the things that actually drove most economic growth, most productivity. They're the reason we have higher living standards than our ancestors, you know, a couple hundred years ago or even fifty years ago. AI is the most important general purpose technology of our era, maybe of all of history.
But the thing about these general purpose technologies is that the value really comes from the complementary assets. So, you know, with electricity, it enabled, uh, light bulbs and electric motors and air conditioning and ultimately computers and a lot of other things. And with AI, it's enabling all sorts of new business processes, and it's not just the the physical technologies, it's also these intangible assets that we talked about earlier. That's great. The thing is that these intangible assets take time to create, they're expensive, you have to invest in them.
You have to, whether you hire consultants or you do it yourselves, you have to reinvent how work gets done. During that costly period, you're spending more as you're reinventing your business processes, but output doesn't instantly go up. So, you know, mathematically, that means more input, no increase in output, productivity, at least as is conventionally measured, productivity goes down. So that's the downward part of the J curve. Then later you start harvesting that, you start using these new business processes to create more output.
Now output is going up, and that's the upward part of the J curve. So what we're seeing, you know, with most general purpose technologies, is that there's an initial lull where productivity is low or even negative, and then later it takes off. Now it's hard to see exactly where we are for AI. We're in the middle of it. But if you look back at earlier history, if you look back at, say, electricity, I map this out.
I wrote about some of it in my, uh, PhD dissertation. With electricity, believe it or not, it took about twenty to thirty years of time where companies were trying to reinvent how factories were organized. They were installing electric motors, and at first they did not get much productivity. You know, Paul David and others documented this very carefully that for literally up to thirty years, there was essentially no productivity gain in factories as they were electrifying, which is just remarkable.
But eventually, they completely redesigned how factories were organized. Instead of being clustered around a central motor like a steam engine, they had separate motors for each piece of equipment, and they distributed the work and they laid out the arrangement based on the flow of materials. When they did that, productivity started skyrocketing, like doubling and tripling productivity. But it, like I said, it took thirty years. So, you know, I don't think AI is going to take thirty years.
I'm pretty sure it's already, you know, beginning to happen. But at the same time, it takes longer than some of my technology friends, some of my AI friends expect. They think that as soon as you invent the AI, you're you're done. And the reality, a as you know, Sam, in your work and and in my work, there's a lot of business process redesign that has to happen. We're in the middle of that, and I'm doing what we can to try to speed that up.
My company, WorkHelix, is very involved in showing companies how to identify the opportunities for value creation. So if we can speed that up, it'll be great, but we're still going to have a bit of a J curve.
I'm thinking back on the electricity times. I'm guessing they didn't have the quarterly stock report pressures and the need for instant results that perhaps people deal with now.
Yeah, but I think the biggest thing, honestly, is is that they didn't really have all of our institutions, you know, not to pat ourselves on the back, but, you know, like business schools didn't exist, consulting books, the whole science of industrial engineering and management didn't exist. I went back and I read some of the books. I went to Harvard Business School Baker Library, and they had these old books, and it was, uh, it was pretty primitive how they were thinking about things.
Now there's a much bigger ecosystem for helping companies, you know, redesign their business processes. Uh, it still takes time and redesign, not just that, reskill their workforce. But my sense is managers are much more conscious of it right now and they're leaning in much more aggressively and they're sharing best practice. Like at WorkHelix, we have a tool that will scan through all of the opportunities and share them with the workers and it'll look at who the super users are who are really crushing it and then compare them to the average worker.
So it just speeds up the learning by probably ten times compared to what people would have done before they had these kinds of tools.
Yeah, actually, I really like that because that's again ties in measurement and some other things that, you know, we the annual performance review is really not helpful anymore. And the idea that you could get better measurement about what you're doing and how you're doing it helps. But let me kind of be antagonistic a little bit. I think I wrote something in Sloan Review, uh, a while back about rethinking AI objectives. I think you wrote something about Turing trap.
I think the first inclination that people have is to pass that Turing test, to make the machine do what the human can do. And that leads us to this incremental improvements of existing processes. It seem like measurement could accidentally feed into that as well, if we're not measuring the right things. What are you thinking about that short term versus long term?
Yeah, well, there's definitely a short term, long term. You know, there's your objectives. I was talking to a CFO a couple months ago, and she told me, Hey, like, we really need to measure the benefits of AI, the ROI. And I said, That's great, you should be measuring it. You know, tell me about it. And she said, Yeah, so that's why I'm demanding that every division tell me how much headcount reduction they're getting from AI. And I was like, Oh, wait a minute, you know, I I'm glad you're measuring, but isn't that a little bit simple-minded?
Because, yes, you know, there's nothing wrong with with cutting costs, but it's really missing the bigger opportunity of doing new things, allowing your workers to create new kinds of value. You know, there's a CFO getting paid millions of dollars, and, you know, the reason they're making big money is that they gotta think more creatively about value creation, not just having machines do what the humans are already doing and replacing them. You know, that's the classic Turing test is, can you make a machine imitate a human perfectly?
That's a very narrow way of thinking about it. And I think Alan Turing is a brilliant guy, but I think that mindset has done a lot of damage to how we use AI. My, you know, I call it the Turing trap because I think that it's a trap to only focus on using AI to replace or imitate workers. We should also think about how AI can augment and allow us to do new things, but that requires new measures, like how do you create new products and services?
How do you have higher customer satisfaction? You know, how do you improve quality, both the quality of the products, but even the quality of of the work life for the people working? These are all things that won't show up in your narrow replacement mindset. So this is something also that takes a longer time, like you said, the short termism tends to focus on just cutting costs and what you're already doing. But the ones that have the most competitive advantage, the ones that have the most lasting benefits, are the ones that think of going beyond that and doing new things, new business processes, new products and services.
It doesn't happen in weeks, but when you do achieve it, you have something that lasts for years.
Yeah, and you know, there's plenty of places where machines outperform us. You know, the easiest way to tell if this is a machine or a person and pass that Turing test is that I'm not gonna be able to do math very fast. Our objectives are not exactly the right the right standard, I guess. You mentioned it in the electrification that people didn't have the institutions around that. Neil Thompson with the MIT initiative, the digital economy that you you were obviously involved in before you moved to Stanford, they've got a science article that said something like 90% of the most important new AI models are coming out of industry.
We've had this massive swish from academia and public funding produces technology benefits to most of it coming out of technology companies and startups. And you mentioned Palo Alto and the the epicenter that's there. All right, I'm worried. Is that the right place for that stuff?
Well, it has, you know, benefits and costs. Part of the reason it's coming out of industry is it's gotten to be incredibly expensive to build these very large models. You know, there are these scaling laws for LLMs where as you make a model bigger, if you have a model have more data, more compute, more parameters, you get a predictable improvement in performance. And so they went from spending millions to tens of millions, hundreds of millions, billions, now now, tens of billions, even hundreds of billions of dollars on training these models.
No university can afford that, no nonprofit organization can afford that. So the companies are raising huge amounts of money a as we're speaking, you know, Anthropic and OpenAI are filing to go public. Uh, Google has a nice cash engine they can pour into this as this meta. So that's what's driving that. At the same time, what I tell my academic colleagues is, you know, your comparative advantage isn't spending money, more money on on computers, it's in thinking creatively.
Like, there's some brilliant ideas that happen just from like sitting in your office and thinking deep thoughts. That's still a good good strategy. You know, I had a conversation with Jeff Hinton a while back, and I asked him what kinds of hardware he uses to make his insights and discoveries. And he pointed to his laptop and tapped it and said, This is this is the hardware I use. And for those kinds of fundamental breakthroughs, it's not always a matter of having, like, thousands of GPUs.
I think academics need to lean into their comparative advantage. You know, that said, it's fair to say that more and more of the frontier research is happening through just a really expensive approach. In a way, it's a testament to how valuable AI is, you know, one of the reasons industry didn't do it before was not just that it was expensive, but it just didn't have that big a payoff. I'll share a a story that that, you know, the first AI company I started, I started back, believe it or not, in the late nineteen eighties.
That's how old I am. So I I started an expert systems company with Todd Lifbr. Expert systems are these rule-based systems. And, you know, we create a little bit of value for some banks and HR organizations, but the end of the day, it was nothing like the value that companies are creating today. That brand of AI just wasn't all that valuable, I have to be honest. It was only now what we have these very powerful neural networks that you really have a a commercial incentives that that's unlike anything we had before.
And so, in a way, it's a testament to the success of the field that industry is investing as much as they are.
Actually, I like that point. I mean, the US. government, I think, put like a billion and a half dollars or so into non-defense research spending last year, and that's roughly what Google is putting in. You know, the NIH and NSF budgets are just small. And so, as you point out, we are not gonna be able to outspend those, but get a little more
No, but but but, you know, but on that, I I think it's fair that that you wanna have both. I mean, it's the job of academia, it's a job of the federal government to invest before it's commercially viable. And so much of the long-term benefits of of RD, you know, they're not apparent, and and there's not a commercial incentive to invest in. You know, the internet, as you know, early space travel, you know, so many fundamental breakthroughs in biology that have extended our lives.
They happened in universities and government laboratories because they were not commercially viable. But then later people built on them, and you can't always tell in advance which ones are going to pay off. But if you do the fundamental research, the track record shows that eventually, some percentage of them are going to really make a difference in our living standards.
you know, that's the basic applied cut. Then we don't have to fund the applied stuff because the market will take care of that. And the basic stuff doesn't the invisible hand doesn't work quite as well in those, or it works slower than everybody would like. But so what should you know? I think you're gonna have a bunch of academics listen to this because they're curious what you say. What should people be looking at? What are the kinds of things that fall into that category?
You know, maybe your critique is that a lot of the studies that I'm seeing in academia are, let's say, you know, consulting projects for companies. What are these basic types of ideas that people should be pursuing?
Well, there's a lot of them in lots of different areas, and I don't wanna prescribe them because I I think in a way, part of the magic is letting people pursue their own ideas, even if there's nobody on the outside that sees the value of them. That said, you know, I'll give you my own list. I mean, I'm I'm focused on the economic side. I think that right now the biggest gap is that we don't have a good understanding of how AI is going to transform the economy.
We are creating incredibly powerful information processing entities. As an economist, you could think of a market as an information processing entity. You could think of firms and organizations as information processing entities. If you have a a thousandfold or a millionfold improvement in AI, it would be a miracle if firms and markets didn't also transform. And we develop new kinds of organizations to manage work, but we haven't figured those out yet. And so big part of my work is trying to understand what does the economy look like, firms, markets, and maybe some new kinds of entities as AI gets more powerful, as it aggregates information, as it processes information.
And then, you know, how are we gonna manage the concentration of wealth and power that could result? You know, I'm super optimistic about the potential of technology to boost productivity. I have a bet, a long bet on that with Bob Gordon, and I've written about, you know, how I expect productivity to grow. At the same time, I'm pretty worried that if we don't manage it right, it could lead to a great more concentration of wealth, and that tends to lead to a concentration of power, and that could hurt our freedoms if we don't figure out an economic system that balances not just wealth creation, but also shared prosperity.
Yeah, I think that's the I was looking up some some of your background before talking, just to kind of refresh. And I was, I saw in the phrase, uh, mindful optimist, and it kind of it kind of hit me. So what's it gonna take to get us to this positive version of the future versus the negative version of the future?
Yeah, well, let let me take a minute to define mindful optimist, cause I have a very specific meaning for that. Is there's a lot of people out here in Silicon Valley who are optimists, but I think they're kind of a lot of too many of them are kind of blind optimists. They're like, Hey, Eric, don't worry, it's worked out in the past, it'll work out in the future. Just chill. There are also a lot of people who are sort of what I call blind pessimists.
They're like, Oh my god, we're doomed. There's nothing we can do. I think both those groups make the same mistake. And it's the one we talked about at the top of the program, which is they think of AI as something that's gonna do things to us, as opposed to us having the agency to control the future. So, what a mindful optimist is somebody who sees a future and then works towards doing it, you know, is mindful about creating it, doesn't just assume it's going to automatically happen.
They're not like kids at at Christmas that just expect presents to magically appear the next morning, but instead they're more like kids who see a tree and and they see some boards and they're like, Hey, we could build a tree house there. And then they're like, Well, it's not gonna happen by itself. Let's figure out how to make it. And then they make that optimistic reality happen through their hard work and through their imagination. That's the kind of optimism that I would like to see.
And I I don't see enough of it. I see way too much just passivity. I gave the closing lecture to my class here at Stanford last Friday. And, uh, one of the things I told them was every time they hear the words AI, they should think of amplified intention. That AI is the greatest amplifier of intention ever. Whatever they want to achieve, AI can amplify it, but it's not gonna happen without that agency, without that intention.
And so that's a message I want to have for the broader world is like, you know, let's think about what our values are, what kind of economy we want to shape, what kind of, uh, shared prosperity we want to create. And I mean that for everybody, not just the technologists. And I've, you know, sending that message to them, but also to policymakers, also to citizens, to workers, to managers. Let's think about really actively using this. Over the next ten years, I expect the world is going to radically change, but there's no one inevitable future.
You cross the border between, like, say, United States and Mexico or other countries, you can see that, like, different institutions lead to very different outcomes, different levels of wealth and democracy and health. That's the same for the future. We can live in a lot of different possible futures. And one of the things I'm doing here at this digital economy lab is doing a lot of research to understand which paths are more likely lead to those beneficial outcomes.
I think that's a great way to end here, that with that we have not easy answers in either direction, and the gains are real, the harm is real, and which one wins is a choice that we're making now and something that we have control of.
More than ever. And one of the interesting things is, you know, this is the time for philosophers and for humanities and for economists and people who think about what kinds of values we want to instantiate in the world. Because with this greater power comes greater responsibility and more potential to change the world. So we really need to think consciously about what what kind of world we want to shape going forward. And I don't think there's been enough attention to that, but but the digital economy lab is in part focused on that.
Thanks for bringing the numbers and the nuance. We'll see you in Boston this summer.
Absolutely looking forward to it, Sam. Great talking to you.
Thanks everyone for listening today. We're off for summer break and we'll be back in the fall with an exciting lineup of speakers from Instacart, GoFundMe, Dropbox, and Honeywell. Please join us then. In the meantime, it helps our show a lot if you leave us an Apple Podcast review or a Spotify rating. If our show has really impacted your life and work, please drop us a line. You can reach us at SMR-Podcast at MIT dot edu. We'd love to hear from you.