Erik Brynjolfsson on the Exponential Mismatch Remaking Work and Politics

In early March 2020, with only a couple of dozen confirmed COVID cases in the United States, Erik Brynjolfsson posted Albert Bartlett's line that the greatest shortcoming of the human race is our inability to understand the exponential function. He had used the quote before, usually in an optimistic register, as in *The Second Machine Age* with Andrew McAfee, to talk about Moore's Law. This time the doubling time was two or three days, and he was scared. Flying home from his last in-person conference in Las Vegas, he was the only person on the plane wearing a mask. A flight attendant put her hands on his shoulders, touching him all over, and asked whether he had an anxiety disorder. He said no, it was because of COVID. Within weeks there were thousands of times more cases.

Linear intuition is what a body learns. Walk for ten minutes and you go about ten times as far as you walk in one. Double something ten times and you do not get ten times as much; you get a thousand. Twenty doublings is a million, thirty is a billion, and after that the numbers stop being graspable. The world is becoming more exponential mainly because of digital technologies, so intuitions are out of whack more often — wonderful for compute, dangerous for viruses. You can learn the math, Brynjolfsson says, but brains learn from experience. Hang around Silicon Valley and AI labs and the curve starts to feel normal. He still underestimates it. Talking about GPT-3, he realized how fast natural language had moved.

The catch, which he underlines, is that most of the world is not exponential. How humans learn, how organizations change, how institutions adapt — those do not improve on a doubling clock. Exponentially improving, or exponentially more dangerous, technologies then smash into skills and governments that barely move. That mismatch, he argues, is at the root of a lot of social trouble, including growing inequality and dysfunctions in political and economic systems. Elon Musk is one person who has internalized the curve, mixing first-principles math with a life spent watching Silicon Valley compound. Brynjolfsson's own first shock was as a graduate student in the late 1980s, plotting computer power by industry and watching every curve take off. Extending those plots was not rocket science. The orders of magnitude arrived anyway.

People talk about Moore's Law as a single skyward line. Brynjolfsson unpacks it as a stack of revolutions: milk one process, then find the next. No exponential continues forever. If something cannot go on forever, eventually it will stop — a line he calls profound and widely ignored. Curves become S-curves or end some other way, sometimes catastrophically. COVID saturates the pool of people left to infect. Chip generations hit physics. The remarkable fact is that humans keep laying a new S-curve on the previous one, generation after generation, with new materials and processes. He does not think anyone has a really good theory of why. A beginning of one is incentives. When the rest of a system doubles every two or three years, the lagging component becomes a bottleneck, and whoever can fix it reaps huge returns. Resources automatically flood the slowest part.

Some version of Moore's Law will continue, he thinks, along other dimensions. One he nicknames after Jonathan Koomey, his college roommate: energy use falling by a factor of two. For most people a new iPhone that is twice as fast is nice; a battery that lasts twice as long is more valuable, and a lot of chip progress now is about reducing energy, not raw speed. Deep nets are highly parallel, so GPUs and then tensor processing units each add another 10x or 100x on top of silicon. OpenAI's work on faster training is the algorithmic layer of the same story. Image recognition, in his accounting, is at least three things multiplying: faster chips, a flood of digital data that did not exist when photography was chemical — he took maybe fifty digital photos the day before the taping — and techniques that have improved a couple of orders of magnitude even though the core idea of deep nets is decades old. A hundredfold in compute, data, and algorithms is a millionfold when you multiply.

Fridman raises the GPT-style worry that self-supervised models will eat all digitized human knowledge and then starve. Brynjolfsson treats data as another of those bottlenecks critics always use to declare the exponential over, and that history keeps walking around. Training already uses less data; capture is still rising, especially in China; some systems simulate, including self-driving stacks that drive in virtual worlds and game agents that exhaust option trees. He does not know the exact bypass. He expects more efficient algorithms, more capture, or a source nobody has named yet.

Autonomous vehicles, Brynjolfsson says, were sold as a yes-or-no. That was too simple. The useful frame is a continuum of assistance, from guardian-angel features of the sort Toyota talks about to full autonomy. Highway driving from Los Angeles to Phoenix in good weather on straight roads is close to a solved problem. Boston in the snow, crazy roads, intersections that are not right angles, and guesses about what other drivers will do is game theory and an understanding of human motives. Some of those settings could take decades. He is humble enough not to referee Waymo against Tesla. What has convinced him is the Level 4 idea of running only in well-mapped areas and expanding those islands until they connect, rather than going out in the wild and hoping.

Waymo had just opened driverless rides to the public in Phoenix. A year earlier they had added a second safety driver because the first one fell asleep; now the safety driver was gone. Remote observation is in the stack — customer service can tune into the car — and he bets they could remote-control it even if that is not an official function. The long tail is what is proving harder than early people expected. You can handle 99.99 or 99.999 percent of cases and still meet something never in the training data. Humans more or less work around those. They also kill about 30,000 people a year in the United States and maybe a million worldwide. People will not tolerate that death toll from a machine, so the edge cases have to be much better.

Fridman's criticism of Waymo is that safety talk crowds out the question of whether anyone loves the product. A perfectly safe vehicle nobody wants to sit in is still a failure. Brynjolfsson points at Musk's Roadster. Early electric cars were underpowered and no fun; Musk built something that went zero to sixty faster than almost anything else, a marketing move and a technology move at once. Too many AI projects, he thinks, are unimaginative. They ask what we do now and how a machine can do the same thing cheaper. Jeff Bezos did not put a robot cashier in a bookstore. He deleted the bookstore, stood up warehouses and delivery, and had people order on a screen. That reinvention, not labor-saving copycat automation, is the historical pattern of general-purpose technologies.

A century ago, factories ran on steam engines or waterwheels. Power moved through pulleys and crankshafts that could not be too long without breaking under torsion, so every machine clustered around one giant engine, stacked in multiple stories to shorten the distance. Small steam engines were thermodynamically unattractive. When electricity arrived, managers pulled out the steam engine, bought the biggest motor General Electric would sell, and changed almost nothing else. It took a generation of managers retiring or dying before anyone asked why they still had to do it that way. Electric motors could be large, medium, or small. The management literature fought over group drive versus unit drive. Unit drive won, factories spread onto a single story, and layout followed the workflow of materials instead of who needed the most power. That is the assembly line. Productivity then jumped. Paul David documented the history in research papers. Michael Hammer's slogan for the computer age was the same lesson: don't automate, obliterate.

During the reinvention, measured productivity often slips. With Chad Syverson and Daniel Rock, Brynjolfsson wrote *The Productivity J Curve*: a downward dip while people create knowledge and intangible assets that appear on nobody's balance sheet and in no GDP line, then an upswing. Self-driving cars are the live case. Hundreds of billions have been spent; basically no chauffeur or taxi driver has lost a job; consumers have little to show. Firms are betting, he thinks justifiably, on the upward arm of the J. As an economist he is disappointed that the past fifteen years of productivity growth were lower than the fifteen before that, the 1990s and early 2000s, despite eye-popping technology. He personally hopes for an upward tick as soon as the next year. Anyone who expected a one-to-one map from cool demos to higher productivity was always going to be let down while the factories were still clustered around the old engine.

Fridman would have expected Wikipedia, Twitter, and the rest of the connective tissue of the last fifteen years to show up as a productivity boom. Brynjolfsson's answer is that most of those goods have a price of zero. GDP is price times quantity of things bought and sold. A zero price contributes, to a first approximation, zero. Hours spent on little screens and big screens never get priced. That does not mean they create no value. He gets a lot from Wikipedia, cat videos, and podcasts even when he does not pay. Simon Kuznets, who built these statistics in the 1930s, said GDP was a measure of production, not well-being. Almost everyone forgot the disclaimer and now treats last year's 2.3 percent as how well-off we are.

With colleagues he is building a parallel account he calls GDP-B, for benefits. If you get value from Zoom, Wikipedia, or Facebook, it counts even at a price of zero. Productivity is just GDP divided by hours worked, so any undercount of GDP is an identical undercount of productivity. The hard part is measuring what people would have paid. Cash registers record the price; they do not record willingness to pay. The team runs massive online choice experiments, now with hundreds of thousands or millions of people: how much would I have to pay you to give up Wikipedia for a month, or your phone? Some questions are hypothetical. Some are enforced, which is expensive. Pay someone three dollars to stop using Facebook and see if they do. Some people quit for ten dollars; some will not quit for a hundred. The answers trace a demand curve. Women tend to value Facebook more than men; older people a bit more than the young, perhaps because the young know other networks he cannot name. A first paper is already in the *Proceedings of the National Academy of Sciences*.

Why have the big networks barely tried letting people pay to drop ads, YouTube aside? He teaches digital business models. Zero marginal cost plus competition pushes prices toward zero; ads, bundles, donations, and volunteerism are the other knobs. Economists, he thinks, underestimate donations — Wikipedia, NPR, this podcast's sponsors. Theory says a flat, wide demand curve, generic news, tends to favor advertising; a steep curve for a small, high-value audience favors charging. Two-sided platforms grow one side and harvest the other, which is Facebook's user-then-ads machine. A mix can work, even a slider for how much advertising you will take, if the incentives are compatible. Fridman rants that newspapers make subscribing so many clicks that he simply does not. Brynjolfsson is with him. Loading a new iPhone can cost an afternoon, a bigger deterrent than price. Make buying painless, he says, and people will give you more money.

Netflix's *The Social Dilemma* and Tristan Harris sit in the background as they turn to networks. Brynjolfsson is worried that people are afraid to try other business models, and that some of the current ones push bad choices. Daniel Kahneman's System 1 is the fast, reptilian click-and-retweet brain; System 2 is the frontal cortex that is supposed to be careful and does not get used enough. Social products, he thinks, tend to exploit System 1, which runs on sex, violence, disgust, anger, and fear — useful for some purposes, a poor way to organize a society. We have connected billions of brains. The most important job of that infrastructure, arguably, is to favor truth over falsehood. The way it has been designed, not necessarily on purpose, does the opposite.

MIT colleagues including Deb Roy put that fear on the cover of *Science*. False information on Twitter spread further, faster, to more people. Not because people like lies. Because people like what is shocking, the "can you believe this" post, not the mundane thing everyone already knew. The most unbelievable things are easier to invent if you are not constrained by the truth, so falsehoods carry higher emotional valence and travel. Algorithms are not the whole story. YouTube has been accused of amplifying extremist content, but in the Twitter study, even after bots and automated tweets were stripped out, lies still spread significantly faster. "It's just the problems with ourselves that we just can't resist passing on the salacious content." He still blames the platforms. Design can add friction or it can make retweeting effortless. Dinner with Jimmy Wales convinced him that Facebook, Twitter, Wikipedia, and Reddit are making choices that make false news more or less likely. Academia, far from perfect, still tends to amplify an important discovery through citation. He is disappointed that the heads of the networks have not tried as hard to amplify truth, in part because revenue models reward growth and speed, which is not the same as finding what is true.

Fridman offers the darker thought experiment: maybe the last few years revealed that there is a lot of money in dismantling the idea of truth itself, until ideas are toys unconnected to reality. Brynjolfsson says that future is possible and not inevitable. He is not a technological determinist. "Technology doesn't shape our destiny. We shape our destiny." German rocket scientists at the end of World War II supposedly said their job was to make the missiles go up; where they came down was someone else's department. That is not the right attitude for engineers. People will profit from a world where facts are not checked and fame is orthogonal to truth. One reason the last few hundred years went better, he suspects, is the scientific method, a tool for favoring what survives tests. Societies and organizations that embraced it did better than those that did not. He wants people to keep embracing not only truth but methods that lead to it. Econ Twitter, for all its nastiness, sped up the clock of learning about new NBER papers. The underlying premise of these networks is still amazing. The cesspool is a design and a habit problem, not a fate.

A tactic Brynjolfsson has come to take as seriously as outright lies is nitpicking: find an extreme nutcase on the other side, amplify them, and make them look typical. You are not literally lying. You take an idiot ranting on the subway, a KKK member, an Antifa member, someone twelve people would have noticed, put video in front of tens of millions, and watch anger do the rest. He has caught himself getting angry and wanting to tell friends. A friend who studies Russian information campaigns described operators playing both sides — some pretending to be BLM, some white nationalists, throwing epithets — not to make one side win but to heat everyone up and seed distrust. The tools are being used that way, and it has been super destructive for democracy. The people who run the platforms, he says, have a social, moral, and personal responsibility to design toward truth and constructive communication, not to "shut that stuff down" in the sense of turning speech off.

Fridman extends the worry to a single sentence, even a joke from three years ago, lifted out of context and used to define a person. Brynjolfsson's reply is that we all need to be more forgiving. In his twenties he went through his friends and found that every one of them held at least one opinion he completely rejected, nobody except, of course, himself — and he is sure they thought the same of him. Leadership has to walk that line in public; Barack Obama, Fridman notes, has been eloquent on cancel culture because saying the wrong word infuriates people. Platforms should slow nitpicking. Users have to learn, too. "If we're idiots about using it, nobody can design a platform that withstands that." Fire cooked food and burned people; later came extinguishers. Cars came, then seatbelts. Digital tools will need both better technology and better habits.

Looking at the rest of the century, Brynjolfsson calls himself a techno-optimist, not a singularity person. Technologies already in the can — vision, voice, problem-solving — can take ten, twenty, thirty years to diffuse, so even with no new inventions there are decades of progress in wealth and, especially, health care. He does not see the end of work soon. Child care, health care, cleaning the environment, scientific work, artistic creativity, even folding laundry are still things machines do not do nearly as well as humans. Society will restructure and reskill for ten, twenty, thirty years or more. Fifty or a hundred years out, all bets are off; one or two hundred years, he thinks it even likely that machines do most of what people do economically better, leaving humans to be human. Shame on us, he says, if we treat that abundance as a problem. Virtual reality good enough to replace waking hours at screens is possible; Nick Bostrom's simulation argument is on the table. Truly superhuman AI is, by definition, able to think of things he would not. Alignment is then the worry, including the fact that our values disagree and that his own 1860 self would look barbaric to him now. Exponentially improving biotech, nanotech, AI, and the old-school nuclear arsenal are a non-trivial branch where we destroy ourselves. That is one reading of the great filter: a trap still ahead, or, more hopefully, a filter already behind us because most planets never get intelligent life at all.

The nearer question is jobs. We are far from artificial general intelligence, he says, and already have human-level or superhuman narrow systems. With Tom Mitchell he wrote "What Can Machine Learning Do?" in *Science*, interviewing AI experts for a rubric that scores any task's suitability for machine learning, then applying it to O*NET, the U.S. government's map of about 970 occupations and their tasks. Radiologists have twenty-seven. Reading medical images is a core one; machines have gotten as good or better, with *Nature* papers to show it. Radiologists also administer conscious sedation, do physical exams, and explain results to doctors and patients, and machine learning is not up to those yet. The occupation continues, restructured. No occupation in the data set was one where machine learning ran the table. Almost none was one where it had no significant foothold.

Machines tend to be good at routine problem-solving, mapping inputs to outputs when there is a lot of X and Y data. They still struggle with fine motor control, emotional intelligence, human interaction, and thinking outside the box. Give them a well-structured task and they shine; even asking the right question is hard. Picasso, shown an early computer, was unimpressed: all that does is answer questions. On a scatterplot of the 970 occupations, lower-wage jobs tend to have more machine-learning-suitable tasks. Cashiers already watch scanners recognize apples and oranges. Some high-pay jobs, including airline pilots, also load heavily on suitable tasks. Economists, paid fairly well, mostly did not land in that category — a result that fell out of the data rather than being drawn to flatter the authors. As technology improves the plot will move. Taken as it is today, it is likely to worsen income inequality.

Andrew Yang, then a presidential candidate, made universal basic income a centerpiece and had read Brynjolfsson's book. Brynjolfsson likes the thoughtfulness and the values. On UBI itself he was originally a fan, became less of one as he studied it, and has been coming back a little. The economist's simplest move, when people lack income, is to write a check. Sociologist friends convinced him that does not hit the core. Voltaire said work solves three great ills: boredom, vice, and need. A check handles need. People need meaning and something to do. When steelworkers and coal miners lost jobs and were just given checks, alcoholism, depression, divorce, and drug use all went up. Sitting around collecting money did not make them happy. Maybe that is how they were raised; maybe it is innate. They needed to feel wanted.

The second objection is that we are far from the end of work. Cities need cleaning; robots cannot do most of it. Child care, health care, mental illness, aging, roads — there is plenty that requires at least a human component. Write those people off, or find a way to keep them engaged. He still wants more buying power at the bottom. Super-hard-working people have seen wages barely grow for twenty or thirty years while the super smart or super lucky made billions, and he does not think those people need hundreds of billions as incentive. Most would do the work for a billion, or for nothing. Would Bill Gates have founded Microsoft at a 70 percent top tax rate? We know he would have, because that is what the top rate was. Brynjolfsson's favorite tool is the earned income tax credit: supplement wages of people who have jobs and give employers a reason to hire more. A minimum wage can discourage employment; an EITC can encourage it. If the employer can only pay ten dollars, the rest of society kicks in five or ten more, buying power rises, and entrepreneurs start building for that market. Someone suggested a conditional basic income whose condition is learning new skills. He likes that too. The hard question for a laid-off worker is what to learn next. Better platforms could map existing skills onto demand and draw the path from A to B.

Asked what he would fix in the tax system, Brynjolfsson does not call for a reinvention. A lot of it is Economics 101 that worked in the twentieth century and got forgotten: invest in education and infrastructure, welcome immigrants, make the tax system more progressive. Top rates used to be much higher; in many cases they are now lower than rates on poorer people. He wants more Pigouvian taxes — tax the bad, not the good. Right now the United States taxes labor and capital, and a basic principle is that if you tax something you get less of it. Work and investment are things we still want. Pollution and congestion are not. A carbon tax is a no-brainer among economists, Republican and Democrat, including Greg Mankiw, who chaired George W. Bush's Council of Economic Advisers, and Dick Schmalensee. It could raise hundreds of billions, which could be recycled through an EITC so the overall system became more progressive.

Congestion pricing is the one that kills him as an economist. Sitting in a traffic jam is a complete waste. The cost is taken from him and everyone around him and, in his image, dumped in the ocean. Charge a congestion tax, as Singapore does, and traffic flows; the same money can go to health care, infrastructure, or food. He and others have been invited there to talk. None of this is radical. Most economists would agree point by point, and the economy would be more efficient and fairer. R&D is as close to a free lunch as the field has. His erstwhile MIT colleague Robert Solow won a Nobel for showing that most living-standard gains come from technical progress; Paul Romer later won one for showing that investment in research and human capital can speed that progress. The distinction that matters is applied development versus basic research. Tesla improving a self-driving stack can capture the value. Fusion, core AI, and biotech ideas leak; they are classic public goods. The U.S. government used to invest much more and then slashed it, eating the seed corn. People pocket money today and, five, ten, twenty years later, are a lot poorer than they would have been.

The pandemic, Brynjolfsson says, is horrible in death and economic destruction, and also the largest reorganization of work he has seen since World War II mobilization. In the previous twenty weeks he saw more change than in the previous twenty years. Before COVID, a little over 15 percent of Americans, about one in six, worked remotely. That rose to roughly half. It was not even. Information and professional workers who live in data could go home. Manufacturing workers who live with people or physical things could not, and they were much more likely to be laid off. Will it stick after a vaccine? His hypothesis is that a surprising amount will. He moved academic seminars to Zoom and was surprised how well they worked: a broader audience, including Europe, and a more egalitarian room because chat let shy graduate students speak and answer one another in parallel, a higher-bandwidth seminar. Bandwidth and email scaled up gracefully, a bright spot among institutions that failed. Economists call the question hysteresis. Boil an egg and it stays hard when it cools. The remote-work system will not snap all the way back.

Fridman names the people the microphone misses. Voices with Twitter followers can still function remotely. Millions whose jobs were smashed do not have those followers. Reading *The Rise and Fall of the Third Reich*, he hears how Depression-era suffering turned into destructive politics, and he wonders what this quiet suffering becomes in five or ten years. Brynjolfsson says that worry is the center of his. There is a moral piece: most Americans, he thinks, want shared prosperity, and people have fallen behind not only because of COVID. Median income barely moved for a couple of decades while the top one percent skyrocketed, partly from how technology was used and partly because the political system shifted wealth toward powerful interests. There is also a political risk. People who feel they got a raw deal wanted to smash the system in 2016 and 2018; demagogues harnessed that destructively.

The analogy is trade. Almost every economist thinks voluntary exchange makes both sides better off, and that trade can still create losers. The formula was free trade plus compensation for the hurt, so a bigger pie could make everyone better off. The 1980s and 1990s delivered more trade and did not deliver the compensation. People felt the bargain was renege, and he thinks they were right. The backlash, especially under Trump, is bad for living standards and intelligible as betrayal. Technology has the same shape. It can make the pie bigger and still leave a majority behind. Economists say grow the pie and adjust the system so the rich, the middle, and the poor can all be better off. Again we are not doing that. Again people say the system is not working. Instead of fixing distribution they start saying technology sucks, throw rocks at the Google bus, blow it up. Almost exactly two hundred years ago the Luddites smashed looms and spinning machines for the same reason. The imperative, moral and national, is not just prosperity. It is shared prosperity.

After more than thirty years at MIT, Brynjolfsson has moved to Stanford to run the Digital Economy Lab, closer to the people inventing AI, and, he admits, for the weather, which greeted him with fires, an orange sky, and a heat wave. Asked what is beautiful about universities amid the usual criticism, he reaches for revealed preference: watch what people do, not what they say. He could do other things and he stays. Magical is the right word. He would do the work for free. He likes research, conversations with students and colleagues, being around the smartest people he can find and being challenged. A university is full of people who feel that way, and he is lucky to live in a society that pays him to put food on the table while doing it. He hopes someday everybody can have a job like that. Academia is a kinder, gentler version of the world, with slack cut, harsh debates and petty politics included, most of which he tries to avoid. The lab's graduate students, undergraduates, and postdocs each have a piece of work he could not even understand. That, to him, is the joy.

Fridman asks for books. From his teens, Hermann Hesse's *Siddhartha*, a reminder not to get too wrapped up in material things. For career, Robert Heilbroner's *The Worldly Philosophers*, lively portraits of Adam Smith, Karl Marx, and Keynes that helped make him an economist; a bit dated, still a fair whirlwind through the first couple of hundred years. More recently, Max Tegmark's *Life 3.0* — Fridman's first podcast guest — which makes profound things joyful and intelligible. Andrew McAfee's *More from Less* made him optimistic that rising living standards can land more lightly on the planet because of digitization. James Clear's *Atomic Habits* is the self-help book he actually uses: small changes with atomic-scale impact, a pun Clear leans into.

The meaning of life is a ridiculous question to put to an economist, Fridman says, and also to a human being. Brynjolfsson learned a lot from his son Luke, nineteen, who has always loved philosophy and once, at seventeen, made a sharp case for hedonism. It did not strike the father as right. He is something of a utilitarian, greatest good for the greatest number, and still finds that too shallow. Real happiness does not come from seeking pleasure. If you focus on being happy, it does not work. You have to be doing something bigger. Look straight at a dim star and it disappears; look a little to the side and the more light-sensitive part of the retina picks it up. Happiness is like that. Find another goal, some meaning, and you end up happier than if you go squarely at pleasure. He is not an evolutionary psychologist, but he guesses brains are wired for more than pleasure because we are social animals wired to help others. When we do, or at least feel we do, the reward systems kick in more deeply than they do for something selfish and shallow. We are all, he says on the way out, part of an invisible college. It does not matter so much anymore where we sit.