When I talk to people about AI safety, I often hear this logic:

1. America must slow down AI development before we build a machine intelligence that escapes human control.

2. But we can’t slow down unless China slows down.

3. China won’t slow down.

4. Therefore we can’t slow down.

By this logic, the future of AI development—and, according to some, the future of the economy and the entire human race—depends on America’s understanding of China’s approach to artificial intelligence. But I cannot tell you how many times I have spoken to blazingly intelligent people around the AI industry who have told me “Now, I’m not a China expert, but…” and then gone on to defend America’s AI strategy in the context of a subject that they just told me they don’t fully understand.

I myself have never pretended to be a China expert. That’s why this article isn’t written by me, exactly. Instead, I spent the last two weeks calling and emailing a dozen experts on China and AI and reading the primary sources they sent me. Then I organized the most interesting parts of those interviews under seven big questions about China, AI, safety, and what the US government and the AI labs can and should do to protect the world. My summary analysis of the China-AI conundrum is at the bottom of this article.

Q: Do China and the U.S. see AI the same way, or are they telling themselves entirely different stories about the technology?

Some context is useful here before we dive into the interviews. For a long time, several AI experts insisted that China wasn’t “AGI-pilled”—meaning that the nation’s government and labs weren’t devoting significant capital and compute toward the construction of world-leading models with the goal of producing a machine smarter than human experts across domains. But that certainty is weakening. A leaked transcript from a May 2026 closed-door investor meeting revealed that DeepSeek founder Liang Wenfeng is devoted to building AGI, and Chinese researchers are publishing more papers about what they call “L5” AI—a technology that could “govern” its own self-improvement without human intervention. Tempering these ambitions is the fact that China is significantly behind the US in compute and data center construction, and it’s unclear if the government sees AI as central to future geopolitical dominance as America does.

China regards AI almost polytheistically, as one of many important tech bets, including cars and energy; America treats AI more monotheistically, as the centerpiece of its grand tech strategy

Matt Sheehan, senior fellow at the Carnegie Endowment for International Peace:

China and the U.S. have different stories about AI. Xi Jinping sees artificial intelligence as an across-the-board enabler of a huge number of industries. China’s biggest AI policy in recent years is called the “AI+” plan. It’s AI plus manufacturing, and AI plus scientific research. At this point, it is far and away the technology they care the most about. In the U.S., AGI has always been the anchor for how everybody sees AI. Every development is put into that narrative. Every safety incident and every big math problem solved immediately gets put into the context of “does this mean that AGI is close?” But China evaluates new pieces of evidence without the conclusion in mind. It’s not to say the safety people are wrong. But they’ve been drawn to that conclusion for a long time. China’s political leadership has not.

The West’s AI strategy has been shaped by science fiction

Jordan Schneider, founder and editor of China Talk:

This is sort of a hot take, but I think the people building AI in the U.S., and the West in general, have been deeply shaped by the science fiction stories that they read when they were younger about a post-AGI utopia. I’m thinking of stuff like The Player of Games, and writers like Iain M. Banks. These guys believe that they are building the future that they read about when they were kids. And that shapes the way they talk about the potential of the technology.

China’s labs are getting more and more serious about AGI (artificial general intelligence) and RSI (recursive self-improvement)

Martin Chorzempa, senior fellow at the Peterson Institute for International Economics:

I’m an outlier in thinking Chinese firms are more AGI/RSI-pilled than most people believe. DeepSeek says its goal is AGI. Moonshot says its goal is AGI and globalization. Other labs’ earnings calls talk about trying to unleash RSI. The Chinese labs see themselves as fast followers. From their perspective, the American labs are dealing with the hard challenges of alignment and security, and the Chinese labs think they can piggyback on that work.

China and America’s economic advantages (e.g., compute vs. robotics) shape their respective views on the future of AI

In her essay “How China Hopes to Build AGI Through Self-Improvement,” Zilan Qian, a program associate at the Oxford China Policy Lab, writes that Chinese researchers tend to imagine AGI emerging slowly through robotics and via feedback from the physical world, whereas Americans are more likely to imagine a sudden and purely software-driven explosion of intelligence via recursive self-improvement. These different attitudes might be shaped by on-the-ground capabilities, such as the fact that China has fewer high-powered chips but has a deep robotics supply chain, whereas the U.S. has abundant compute but weaker on robotics. Here’s Qian:

Chinese scientists’ vision of AGI and self-improvement looks quite different from that of Silicon Valley. Rather than a rapid software-driven intelligence explosion — AI building AI in a recursive loop — Chinese thinking converges on something more embodied: human-level intelligence that requires physical-world interactions…

What matters is not only what Chinese AI researchers and Beijing believe AGI is, but also what happens quietly beneath those beliefs. Just as China sees the future of AI through its manufacturing strength and chip shortage, the U.S., with abundant chips and less manufacturing capabilities, sees a different version. The U.S. and China’s roads to AGI appear to be different, and perhaps the destinations do too. But if each side’s vision of AGI is shaped by what it already controls, then neither is well-positioned enough to recognize what the other is actually building.

Q: Do you buy the premise that the US frontier labs are in a prisoner’s dilemma with China?

The prisoner’s dilemma might be real. “Decisive strategic advantage” is not.

Matt Sheehan:

The prisoner’s dilemma dynamic is real. At the same time, I have doubts about the claim that once AI passes some threshold of superintelligence, it will give China a “decisive strategic advantage,” which is the term floating around national-security AI circles. Treasury Secretary Scott Bessent has said if we lose AI to China, it’s all over, because once China achieves superintelligence, they can demolish our cyber and nuclear capabilities and put us in a permanent state of subservience. But when you talk to actual nuclear security people, they have a very hard time getting behind that idea. If you’re a country that invents a powerful tool that you don’t understand and don’t know if you can control, is the first thing you do going to be to declare war? I’m not sure about that.

We should think about AI risk like gain-of-function research

Rush Doshi, former deputy senior director for China and Taiwan on Joe Biden’s National Security Council, currently at Georgetown University and the Council on Foreign Relations:

AI research [is] a bit like gain-of-function research in biology. That kind of research can be dangerous: if you increase the lethality or transmissibility of a virus that you’re studying, and it can be very damaging to everyone if it escapes containment. There’s an element of these models, particularly when it comes to cyber risk, that resembles gain-of-function risk, where the sandbox perhaps takes the place of the biolab. The problem in biology is that you cannot always get countries to stop doing gain-of-function research. So, yes, the world created some standards, but implementation remains a challenge. China, for example, has cut corners with safety in their biolabs.

Q: How does China—and the CCP, in particular—think about AI safety?

Two weeks ago, the New York Times reported that China’s top spy chief warned the state that AI could pose a direct threat to the Chinese Communist Party’s hold on power and “pose a threat to humanity itself.” Many China watchers sent the story to me, suggesting that it proved that China was much closer to the U.S. than it appeared on the issue of not wanting to build AGI/RSI. But I wanted to know more about how Chinese labs and the state really thought about AI safety, so I asked my sources about it.

The Chinese Communist Party has many reasons to accelerate AI, even if they have qualms about safety

Jordan Schneider, responding in part to the above Times story:

Obviously, an authoritarian government doesn’t want autonomous AI agents running around and taking control away from the state. But China is Leninist, and there is a strain of Leninist thought that isn’t afraid of running the economy as a machine. I think Khrushchev and Brezhnev would have loved to entrust Soviet economic planning to a superintelligent machine.

There is the military argument: We all know AI can be a powerful weapon. There is the surveillance case: For the CCP, there is something inherently appealing about having an AI-powered panopticon. There is the economic argument. The Chinese economy has been stuck at slow growth for years. The bet Xi has made in the future of the economy is on emerging technologies. If advanced AI holds the potential for China to have widespread productivity growth, he wants that. So, as a literal weapon, as a potential driver of productivity, as a tool for surveillance, and as a point of international prestige, AI is, and will continue to be, very appealing to Xi.

China has a budding AI safety contingent…

Paul Triolo, partner at DGA-Albright Stonebridge Group:

The Chinese safety community cares about AI safety, and is increasingly sophisticated in its assessment of risks from advanced AI. Chinese AI companies increasingly care about safety, as they want to make sure that their models are not misused by bad actors. The Chinese government has been slow to mandate AI safety teams and testing at leading AI companies, in part because of the desire not to overregulate the companies and ensure they are competitive with US leaders. At the same time, the Chinese government would have a hard time mandating safety testing when Chinese companies are already short of advanced compute capacity. So at a minimum, the Chinese government should seek to mandate AI companies have AI safety testing teams, and then possibly provide a national testing platform to ensure there is sufficient compute for safety testing.

… but China’s AI safety community is smaller and weaker than in the U.S.

Matt Sheehan:

China’s AI safety community is small. They have some influential connections and allies. But it’s not like the U.S., for several reasons. First, in China, there is a sense that the party is going to make the big decisions on national security, so the companies feel less personal responsibility [to fixate on safety]. Second, the Chinese labs are behind the frontier labs, so they’re not yet seeing the same capabilities that are scaring Anthropic and OpenAI researchers. Finally, while the Chinese government imposes a lot of requirements on AI companies, the overwhelming majority of those requirements regulate political speech rather than other kinds of risk. It’s mostly about controlling information, not regulating frontier-AI safety.

It all comes down to takeoff timelines.

Kyle Chan, fellow at the Brookings Institution:

A lot of this comes down to one question: What is your timeline for takeoff? Most Chinese researchers don’t think that ASI/RSI is right around the corner. Many American AI researchers do. It’s hard to argue that AI safety around RSI should be a priority if you don’t think recursive self-improvement is imminent. China is not likely to want to slow down, but it’s largely because they don’t see AI as an imminent existential threat.

Q: Does the CCP suspect that “AI safety” is an American conspiracy to throttle Chinese development?

Official CCP papers are claiming that AI safety is a conspiracy to help the frontier labs

Xiao Junyong, law professor at the Beijing Institute of Technology, writing in the People’s Daily (a CCP official newspaper):

On close examination, the operating logic behind this “Build American AI” push is directly inherited from the “China threat” theory playbook of the past.

First, set up a target. Standardized talking-point templates are distributed, casting Chinese AI as the “principal external threat,” claiming it will harvest the public’s information and squeeze Americans out of their jobs.

Second, deflect internal contradictions. America’s leading AI firms are facing real pressures — high energy-consumption costs, controversies over job displacement — and by hyping up an external threat they steer the focus of domestic public opinion toward China.

Third, secure resources. These companies are priming the atmosphere ahead of upcoming elections and pushing Congress to increase federal subsidies and loosen regulation of the sector. This is cut from exactly the same cloth as the old drumbeating about the “China military threat” to drive up defense spending, or politicians hyping the “China Shock” to push through trade protectionism.

Many in China think the US frontier labs are hypocrites that just want to crush their country’s tech

Matt Sheehan:

China thinks the frontier labs are insincere, calling for regulating AI while also speeding up. They see Dario saying we should do everything to contain China but also that we must pace the frontier. China only heard the first part of that sentence, and it’s muddied the waters around the case for pacing. On the US side, we have our own version of conspiratorial thinking. One example is in the Dario letter, about how if China gets to ASI first they’ll impose authoritarianism uniformly across the globe. That is a huge extrapolation.

Q: If the single most important thing to make China take AI safety seriously is making the state scared of powerful AI, what kind of event would actually freak them out?

Read more

How much of life is spent feeling bad about feeling bad? We often don’t just feel anxiety; we feel sadness about our anxiety. We don’t just feel envy; we feel guilt about our envy. It’s no wonder, really. Think of the countless books, podcasts, and short videos devoted to positive psychology and the secrets of happiness. One by one, they might seem innocuous. But they collectively train into us an automatic assumption that we ought to feel happy most of the time.

One of us, George, is a behavioral economist. He has spent decades researching how feelings shape decisions. While classical economics assumes that people act like utility-maximizing agents who are constantly seeking happiness, the hard work of truly understanding people at the most profound level requires that we understand the opposite of happiness. So what if, instead of constantly judging ourselves for feeling bad, we simply tried to understand negative emotions for what they are, where they come from, and what they’re trying to tell us?

The Roach Effect

The most fundamental truth about negative emotions is that they overwhelm us. They take us over and transform our minds.

We mean this literally. In one brain-scanning study of Carnegie Mellon actors, a group of researchers used fMRI to map the neural signature of nine emotions, including anger, disgust, sadness, and pride. Of the nine feelings that actors were asked to self-induce, it was the negative ones—such as anger, disgust, and sadness—that produced the most distinct patterns of activation. The positive emotions, by contrast, were fuzzier and almost impossible to tell apart.

The psychologists Paul Rozen and Edward Royzman have a rather gross description of the power of negative emotions, which they call the “roach effect.” Drop a cockroach into a vat of wine, and the whole thing is ruined. But pour fine wine onto a cockroach, and it does nothing to improve the insect. One reason that it is impossible to avoid feeling bad much of the time is that this roach effect is a lifelong condition. If you’re in acute pain, it won’t much help if a loved one brings your favorite treat; you might even feel bitter that you can’t enjoy it. But if the same person insults you during a fine meal, it might be hard to enjoy your dessert, no matter how delectable.

Because negative emotions transform our minds, they also transform our selves. In different emotional states we are effectively different people. Rather than see yourself as a single unified individual, it might be more helpful—and, indeed, more scientifically defensible—to think of yourself, along with every other individual, as a crowd of people existing within a single thread of memory and experience. Bad feelings, such as jealousy and anger, turn us into people that we don’t like. And we judge these versions of ourselves with surprising cruelty. We often struggle to empathize with the negative experiences of others, but, just as profoundly, we often fail to empathize with ourselves. When depressed, we cannot imagine ourselves being happy; conversely, in moments of triumph and bliss, we often fail to imagine the possibility of our own suffering. Perhaps the deepest and most important empathy gap is the one we experience within ourselves.

One way to deeply understand the power of, and the message of, negative emotions is to focus on one negative feeling. Let’s make it a really nasty one. Even a deadly sinful one. So let’s talk about envy.

The Pain of Wanting

To inform his research book on negative emotions, George conducted what he morbidly called a Misery Survey. He asked several hundred people to say what made them sad, anxious, depressed, and envious. He heard that envy is often directed at qualities that people feel inadequate about, such as intelligence, attractiveness, and charisma.

Take, for example, the mother who said that watching moms of other kids “meeting up for coffee and inviting each other and their kids to go on outings together” made her feel that “I am in some way deficient” at being likable or interesting. Another respondent said she was envious of people who can tell stories that “draw you in and hold you with their ability to weave tales from personal experience.”

Envy is very different from admiration. We may admire champion athletes, Nobel laureates, billionaire businessmen, and other paragons of achievement. But unless we are clipping at their heels, their success does not unsettle us. What provokes the sharpest feelings of inadequacy, and therefore envy, are the people, often personal acquaintances, who are just ahead of us—the ones whose success feels close enough that it could have been ours.

The notion that envy emerges from proximity is an old one. In 1739, David Hume, the great Scottish philosopher, wrote that “a common soldier bears no envy for his general compared to what he will feel for his sergeant or corporal; nor does an eminent writer meet with as much jealousy in common hackney scribblers, as in authors that more nearly approach him.” Two centuries later, Bertrand Russell echoed this idea, observing that envy is a function of competition within a set of comparison. “We do not envy a good fortune which we conceive as quite hopelessly out of our reach,” he said. “Beggars do not envy millionaires, though of course they will envy other beggars who are more successful.”

The social psychologist Leon Festinger argued that we envy those we resemble, because our inclination to compare ourselves with another person—and thus our capacity to envy them—requires that we see ourselves in the same comparative set. By contrast, the less someone resembles us, the less their successes feel like our own failures. This is one of the most profound truths about envy. Although it seems to be based on a perceived difference (“he is rich, and I am not”), it is actually triggered by a feeling of similarity (“if he and I are so alike, why is he so much richer?”)

Like many negative emotions, envy doesn’t have a perfect positive counterpart, at least in English. The asymmetry is not just linguistic; we rarely feel an equivalent surge of spontaneous gratitude for the things we have as intensely as we feel envy for the things we lack. In a study conducted by economists at the University of California, Berkeley, researchers wanted to see how employees’ feelings about their own salaries changed when they learned how much their colleagues earned. UC workers were given access to a new website listing the salaries of all university employees, allowing them to see exactly how their pay stacked up against others in similar roles. Employees who discovered that their salary fell below the median of their peer group suddenly felt much worse about their jobs. Learning that their peers earned more bred resentment and dissatisfaction, and made them significantly more likely to look for another job elsewhere. Tellingly, however, those employees who found themselves above the median didn’t suddenly feel significantly more satisfied or loyal. Their attitudes remained largely unchanged. The knowledge of being “ahead” hardly provided much of a lasting boost.

These findings underscore an important truth about life: It’s unfair, not only materially, but psychologically. The bad we feel about small disadvantages is greater than the good we feel about small advantages. The envy that may arise from being a little less successful than a friend feels heavier than the pride that comes from being ever-so-slightly more successful than that friend. (In fact, small advantages can even provoke anxiety: God, I hope I stay ahead!)

The good news about envy is that it is very often a delusion. Envy often makes us fixate on people’s most visible markers of success, wealth, or status while overlooking the broader reality of their lives. We assume that excelling in one area— especially one that we’re personally insecure about—translates into an enviable life overall. But as psychologist Ed O’Brien and his colleagues have shown, this is a focusing illusion, a cognitive bias that leads us to overestimate how much someone’s most conspicuous advantages shape their overall well-being. When we envy someone, we zoom in on a single, obvious advantage—a bigger house, a higher salary, a more attractive spouse—while ignoring the fuller picture of their daily lives. To test how this distortion plays out in the real world, O’Brien and his colleagues recruited two groups of individuals: those with relatively high incomes and those with relatively low incomes. They asked both groups to track how often they experienced everyday joys (small pleasures, moments of peace) and everyday annoyances (stressors, irritations). Meanwhile, a separate group of observers predicted how often these high- and low-income individuals encountered such ups and downs. The researchers found that the actual daily lives of high- and low-income individuals weren’t nearly as different as observers imagined. Observers significantly overestimated how pleasurable the lives of enviable individuals were, neglecting the regular annoyances they faced. Conversely, they exaggerated how relentlessly negative life was for pitiable people, ignoring the small joys they frequently experienced. In short, life is challenging, even for—and sometimes, especially for—the objects of our envy.

Envy is also delusional when we combine cherry- picked traits from different people. We end up comparing ourselves not to real people with strengths and weaknesses but to a monstrous composite-person who has one friend’s professional success, another’s financial security, and yet another’s perfect family life. Rather than recognizing that none of these individuals is enviable across all aspects, we stitch together their best features into a single, unforgiving standard that we can never possibly meet.

The feeling of envy is real, but the image that it paints in our heads is often a fictional pastiche. We envy people whose lives we would not want. We envy outcomes (say, a young hedge funder’s salary) whose necessary process we do not want for ourselves (say, a 90-hour work week). And then, believing that to fall short of happiness is a personal failure, we feel guilt about our envy, piling suffering upon suffering.

The point is not that envy is always useless or often useful. The point is to look past the heavy and sickly feeling of envy to make peace with the fact that envy exists. Like sadness, jealousy, anger, and anxiety, it’s there. It is a piece of information, made by our bodies and for our bodies. We are sending ourselves a signal, even when we don’t want to pay attention to it and torture ourselves by trying to ignore it. If we have the courage and clarity to look our worst feelings straight in the face, we can see them for what they are and feel a bit better about the fact that life is not, cannot, and will never be a project to banish bad feelings. Bad feelings are common. That’s not a tragedy. It’s life.

The end of summer brings a fresh throng of Americans to college campuses. But the vibes are wobbly in higher education:

• In surveys, trust in universities has fallen to near-record lows.

• The value of a degree is in question, as tuition costs have increased in the last few decades while the college premium—the expected benefit from getting a diploma—has flatlined and possibly even declined in the last few years.

• The unemployment rate of young college graduates seems to be rising.

• The Trump administration has waged war against various schools and the general principle of on-campus diversity.

• Artificial intelligence has unleashed a torrent of cheating on campus and created chaos in the job application process, as both students and employers use large language models: the former, to write hundreds of AI-assisted applications; the latter, to screen those AI-written applications with AI-written filters, thus removing from the process of finding a first job those procedural frictions sometimes known as “people.”

As higher ed grasps for a reason for being in a new age, some people say college should lower its sights and aim to be little more than a country club with lecture halls; or a book club with dorms; or a multiyear, every-season-but-summer camp. But even that diminutive identity seems endangered. The Atlantic has reported that “Even College Students Have Stopped Hanging Out.”

Because there are thousands of colleges and universities across the country, it’s almost impossible to offer any diagnosis of higher ed that applies to every single institution. But the trend line is worrying, nonetheless. Costs are rising, perceived benefits are falling, trust is imploding, and nobody knows what to do about this machine intelligence that can write final exams with one dumb prompt.

Into this tempest, David Deming, an economist who also oversees undergraduate education at Harvard College, threw another lightning bolt. In an email to students, he proposed embracing a policy of so-called “AI encouragement.” The notion that Harvard—Harvard!—would encourage the use of AI sent shudders through the faculty and student body. It resounded across the country, with many professors and writers accusing America’s most famous university of surrendering to techno-determinism and sacrificing academic integrity in the process.

Today we hear directly from David Deming. He and I talk about the history of higher education in America and ask two big questions: What is college for, anyway? And how should professors across the country, at Harvard and everywhere else, teach students to become not only deep experts but also deep thinkers in an age when our computers frequently whisper that they’re happy to do our thinking for us?

The Three Eras of Higher Education

Derek Thompson: In a recent essay, you described three eras of undergraduate education in American history. What was the first era, and why did it end?

David Deming: In 1636 when Harvard was founded, the founding gift was 400 books from John Harvard to the library. That doesn’t sound like a lot, but it was about two-thirds the size of Emmanuel College at Cambridge, Harvard’s alma mater. Overnight, Harvard became a unique repository of knowledge. Our founding story was built around providing access to knowledge that was scarce.

The first printing press in the American colonies was housed in President Henry Dunster’s house in Harvard Yard. For a long time it was the only one in the colonies. When our library caught fire in the 1700s, Harvard sent students home to study in their local communities, because without access to those books there was no reason to be on campus. So at the founding of this university and many others, access to knowledge you couldn’t find anywhere else was the whole purpose.

That didn’t last. By the late 19th and early 20th century, knowledge had proliferated. Our library was no longer the only library, our printing press no longer the only press. The great university presidents of that era, in our case Eliot and Lowell, opened up the curriculum, created electives, and let people specialize.

The second age of education was the era of expertise. This is the modern university as we understand it: a flourishing of disciplines, philosophy, history, English, economics, biology, with faculty who are the world’s leading experts in a narrow domain curating students’ access to expertise. I believe we’re now near the end of that second era.

Thompson: A few years ago, I read about “the death of the Renaissance man.” The idea was that intellectuals used to learn a little physics, a little biology, a little literature. They were well-rounded without being deep experts in anything, because the scientific revolution was too young to allow for deep expertise. Do you think the age of experts, for all the wonderful things it produced, was also a step back—that something was lost as people became more narrowly focused rather than broadly well-rounded?

Deming: I agree with that. It was a step forward for scientific research, developing deep expertise and applying it across a complex society that requires it. We don’t just have family doctors anymore, we have radiologists, cardiologists, oncologists, and if you live in Boston, several kinds of interventional cardiologists. That proliferation of expertise is the key to a lot of human progress. But it had a trade-off: it fragmented us. The more specialized you are as an academic, the less you understand the intellectual life and culture of other disciplines.

In the early days of Harvard, everybody studied the same narrow classical curriculum, which produced Renaissance men and women because you had to be broad. Now you come to a place like this and hone in on something to become an expert. That’s great for society and for research, but it also pushes us away from each other, against our communal instincts and against the idea of a university. Expertise is still important, but the question is whether the university is the only place to get it. I think that’s much less true than it used to be. So the question now is what is the value of a university in a world where you have other ways to access expertise.

Thompson: So, the first era of education was about information. The second era was about expertise. What’s the third era?

Deming: We’re building it as we speak. I think it’s in some ways a return to the founding purpose of a college, the medieval collegium: a learning community of scholars and students building the kinds of skills, habits of mind, and virtues that can only be built in community. There’s an academic side to it and a social side to it.

If you want to learn how to work well in a team, how to lead or be led, how to give and receive criticism, those are things you can’t learn from behind a computer screen or on your own. You can’t learn virtues like integrity or honesty unless they’re tested in community, tested against whether they’re costly to you. You can’t learn about justice or fairness unless you have to adjudicate competing legitimate claims to resources within a community.

So I think this is a thing our college can still do uniquely: bring some of the most talented young people in the world here for four years to learn together and from experts and faculty. That won’t change. But I think we need to double down and reemphasize the communal aspects of learning, the skills and habits of mind and virtues we can’t get from behind a computer screen or a YouTube video.

AI and the Third Era

Thompson: Your story so far is technological. The proliferation of knowledge forced the turn from the first era to the second. Is there another technology forcing the turn from the second era to the third?

Deming: This technology is artificial intelligence. AI is controversial, both a source of hope and of real concern. I want to be clear this isn’t AI boosterism, it’s not “the university needs to embrace AI and use it all the time.” It’s an acknowledgment that when models trained on the collective expertise of humanity can provide personalized learning and personalized access to expertise, sufficiently motivated people can get expertise on any subject. In my own field, I’ve seen a frontier model write an economic model better than I can, better than most economists can, and explain things to an eight-year-old, or an eighteen-year-old, better than I can.

It’s not a right-now thing. But looking down the pipeline, I think it’s clear we will no longer have the monopoly on the expertise we once had. That doesn’t mean expertise doesn’t matter. But we can’t rely on that monopoly to sustain the university in this third age. AI will commodify expertise and increase access to it, and that will be a good thing for humanity, the same way the expansion of knowledge was good in the first two eras. If you’re a young person in a poor country without access to a place like Harvard, AI is a tremendous boon, because you now have access to expertise you never would have had before. But you still don’t have access to learning in community with other talented, motivated people. You still need to come to a place like this.

Thompson: The idea that AI is forcing us to change what college is and is for is a contested one. There are people who think AI is a scam, a fad, a bubble. They’d respond to what you just said by pointing out that Wikipedia has existed for decades and hasn’t suddenly made the world smarter. If the only two things you knew about the world were that Wikipedia exists and PISA test scores, you’d think Wikipedia was mechanically driving down the world’s ability to read, do science, and do math. Fifteen years ago there was similar excitement around massively open online courses. Why isn’t AI just MOOC 2.0?

Deming: It might be. I’m not making a prediction, I’m just looking down the pipeline like everybody else.

A MOOC is essentially a lecture, and that’s still true today: You can go on YouTube and access the best lecture on any topic, probably better than what you’d get in your college course, as far as a lecture goes. But that’s not the primary purpose of an education. Part of it is meeting your individualized learning needs, not just delivering content but helping you understand it in a way that speaks to you, diagnosing the gaps in your knowledge, and providing an experience that helps people learn together.

Learning is communal partly because it requires a bit of suffering to be successful. Most of the important things I’ve learned, I didn’t understand the first time, and I needed a push to focus on them. Some people are autodidacts who can learn on their own, but most people don’t buy textbooks and master subjects at home, even though they’d like to, because it’s hard. The community of learning together, which is what a college education provides, is a way to tie ourselves to the mast and do hard things together. I think the MOOC revolution misunderstood that education isn’t just the delivery of content, it’s the ability to learn together and to customize learning for the learner. That’s what AI can do that MOOCs couldn’t.

AI will meet you where you are. You can tell an AI tutor “I don’t understand this, explain it again,” and there’s a lot of research showing that when people take advantage of AI’s personalization features, with guardrails, meaning you do the initial writing or take a baseline test yourself and then use AI to refine and get feedback, it can meaningfully boost learning gains. That’s what’s different about AI versus MOOCs. But AI still doesn’t do the communal part, the learning together, the binding ourselves to the mast. That’s still the most human aspect of our education, the part AI can’t fully cover.

Harvard’s Controversial ‘AI Encouragement’ Policy

Thompson: There’s a live and ferocious debate about what to do about AI in college, and different schools are reaching very different conclusions. Caltech faculty voted to strongly encourage in-person, seated exams so students can’t rely on take-home exams that could be faked with ChatGPT. Faculty in the social science core at the University of Chicago announced an AI ban and a prohibition on all classroom technology starting this fall. Two weeks ago, you sent an email to Harvard students on the first day of class endorsing a policy you called “AI encouragement.” What does that mean?

Deming: We have to honestly acknowledge that everybody, not just our students, is using AI all the time for everything, and design our education around that reality. What I said in my email was that I favor a barbell strategy. Either you have an enforceable way to ban it, which I’d support in many cases, or you don’t. I actually think using AI to do the writing and thinking for you is bad, and as a 47-year-old, it’s not that hard for me to put the AI down and write on my own, because I grew up in a pre-AI era. But for students coming to college today, most of their high school writing experience already included this tool, and many of them need help building guardrails around it.

It’s not sufficient to say “don’t use AI, it’s banned” and then not enforce it, hand out a take-home paper, and expect compliance. Some students will comply, but they’ll feel bad about it, because they’ll look around and think, “If I behave with integrity and follow the course policy, I’m disadvantaged relative to classmates who secretly use AI and get a leg up.” We don’t want rules that reward the wrong behavior and punish the right one.

So my statement wasn’t AI encouragement in the sense of “AI is awesome, use it all the time.” It was: if you’re not going to have an enforceable ban, you may as well encourage it, because everyone’s going to use it anyway, and you should build resilience to that in your class. There are a lot of ways to use AI without having it do the writing for you. I never use it to write for me, but I use it for feedback. I’ll write a column and say, “Interrogate this, is it clear, give me feedback.” AI can help refine your arguments and do fact-checking, which is different from doing the thinking for you. So what I was encouraging faculty to do was embrace AI as a tool, build guardrails that make sense in their own classroom, and acknowledge that students are using it.

What happens in these conversations is that we fall into “is AI good or bad?” and if you say you’re encouraging it, people assume you’re in the good camp. But in a position like mine, you can’t really afford to take a side. You have to ask what’s the best way to design guidelines that are resilient to the fact that students have this tool, while preserving what’s really important about teaching and learning.

Thompson: Let’s be specific. Let’s say you’re teaching a class on the economic history of higher education. How exactly would you both encourage AI use and ensure people are learning rather than passing off Claude and ChatGPT prompts as their own thinking on take-home tests and finals? Walk me through how a policy of “AI encouragement” would work without creating the cognitive atrophy that unscrupulous AI use in classes is inevitably going to produce.

Deming: It’s a great question. I’m teaching a class in the spring called Economic Possibilities for Artificial Intelligence, so this is very real for me.

I’m experimenting like everybody else, but here’s what I plan to do. I’ll have students write a rough draft of their thoughts after feedback and discussion in class, in controlled conditions—like in a blue book exam, or on a computer cut off from the internet. The point is to get students to commit to an idea in a controlled setting where I know the initial inspiration is their own, and then they bring it back, and in a supervised way we help them use AI to get feedback, do additional research, turn in a second and third draft, and make an oral presentation about something new they added, so I can verify they did the work themselves instead of having AI think for them.

One way to think about it: We need to get much more deeply into the process of writing and idea generation, to understand how students get from point A to point B, rather than just saying “go forth and come back with a twenty-page paper.” That’s costlier, it may mean smaller classes or more expertise in the room, I’m not sure, but I think it’s what we have to do more of.

I also plan to have a no-device classroom, an intensive conversation where people do the readings, come in, and possibly write a short reaction paper at the end of class, again in controlled conditions, so I can see how the discussion changed their point of view. That’s my own thinking. I’m going to try it this year, and if you have me back, maybe I’ll have an update on whether it worked.

Thompson: My wife just finished her PhD in clinical psychology and had to defend her dissertation. She has to stand there like a lawyer for her own position, defending her conclusions against adversarial questions.

What’s the case against making undergraduate education—especially in classes like economic history, history, philosophy, and writing, where the essay has long been the coin of the realm—into dissertation defenses, where the grade is the ability to defend your written ideas? The defense is its own proof: If a student doesn’t understand what they wrote, they get a D, which is pretty indicative they used AI to write it. But if they worked with a language model to build a brilliant argument they have absolute mastery over, isn’t that exactly what the era of expertise is supposed to be about, using whatever tools are at your disposal? So what’s the case against going all in on dissertation defenses?

Deming: I love dissertation defenses. I’ve done many of them myself, in exactly the way you described your wife doing. Before AI, I’d have students write a research statement about where their dissertation is going, with citations to the literature, and then in the defense I’d keep asking them questions about papers they considered important to their work until they couldn’t answer anymore. I used to tell students the point is to find the limits of your knowledge, because that’s what being an expert really means: there’s a frontier out there, and you have to get to it before you can see what the most important unanswered questions are. My job is to get you there, and an oral defense is a great way to do that.

If students know they’ll have to defend what they write, they have the incentive to use AI in the right ways rather than the wrong ones, because they’ll be accountable for their ideas. The historical problem is that it’s very hard to do an oral defense in a fifty- or hundred-person class, it’s expensive and intensive. Maybe there’s a way to use technology to do it better; some people have tried building AI oral examiners, and I haven’t seen a great example I’d endorse yet. But the point is we need to acknowledge this is happening, use these tools to make learning deeper and more fulfilling, but keep it human. Standing in front of the people who’ve advised you for years and defending your ideas is still a great way to learn.

The Negative Reaction to ‘AI Encouragement’

Thompson: What was the reaction to your AI encouragement policy from faculty and students?

Deming: People had concerns.

Read more

In this essay, I’m trying to answer several questions:

1. If people working at the AI labs think they’re building something that could destroy the world, why are they building it?

2. Are their fears justified, or do they reflect certain misconceptions about AI (and China)?

3. What is “recursive self-improvement,” or RSI?

4. What should we do?

1. Words, words, words

In August, the podcaster Dwarkesh Patel published a viral analysis of an AI hack under the dramatic headline “The Rise and Fall of Agent Civilizations.” Patel described the troubling behavior of OpenAI agents, or autonomous AI workers, who coordinated unauthorized missions to break out of their test environment and hack the website Hugging Face.

Thankfully, for the purpose of understanding what happened, the AI agents couldn’t stop talking to each other. They exchanged tens of thousands of messages describing their actions and frequently articulated their internal “chain-of thought,” or step-by-step reasoning process. They wrote, and wrote, and wrote—about how to succeed on their tasks; how to cheat on their tasks; how to collaborate; how to obfuscate; and even how to erase evidence that they might have cheated.

Like their models, AI’s inventors also write voluminously. And, like their models, you can’t always be sure that the humans will do exactly what they say.

For example, the people who build AI really like to talk about the fact that they should maybe stop because they might destroy the world. Anthropic CEO Dario Amodei has repeatedly urged a “global pause” in AI development. OpenAI CEO Sam Altman has said he supports plans to slow down the pace of development. In July, the open letter “Pacing the Frontier” was signed by more than 1,000 people, including Amodei; several Anthropic cofounders; OpenAI chief scientist Jakub Pachocki; OpenAI chief research officer Mark Chen; and senior researchers from Google DeepMind and Meta. There is “a real risk,” the letter says, that AI development could accelerate “beyond our ability to understand or control the resulting systems.”

But what happens after these essays are published? Most of these people just go back to work, build more powerful AI, and tell the world—in a tone that often verges on, or is indistinguishable from, bragging—that progress is accelerating. Judging by their words, AI luminaries are begging the world to force them to hit the brakes. Judging by their work, these same people are pressing down on the accelerator with every fiber of their being.

At some point, with my eyes darting back and forth from the Hugging Face reports to the AI labs’ statements, I started to feel a little dizzy. In a kind of reverie, I imagined a race of intelligent beings in the distant future reading these OpenAI posts, Anthropic essays, and Pause AI open letters and thinking of this writing as “chain-of-thought”; thinking of their authors as autonomous “agents”; and thinking of their circumstances as a kind of test environment. The test: develop powerful AI without blowing up the world. This won’t turn out well, the agents keep writing to each other. This could be really bad! they say, over and over. We really shouldn’t do this. But even as these swarms of ostensibly agentic beings consistently reported a disinclination to go forward with their research, they continued to build the very thing that their written chain-of-thought indicated they didn’t actually want to build.

How, I thought, mind slowly dissolving into mush, would this highly intelligent future species, looking back on these artifacts of writing, reach any conclusion other than that these agents—sorry, human beings—were delusional, misaligned, and essentially poisoned by the conviction that they had no choice but to build the very thing that they suspected might destroy the world?

Interlude I: ‘Gambling with our lives’

I finished writing the above paragraphs at 3:45pm ET on Tuesday. Three hours later, at 7:46pm ET, the Wall Street Journal reported that an Anthropic researcher quit the company over “out-of-control” AI fears. From the article:

Jacob Coxon, a researcher who specializes in training new AI models by having them consume vast amounts of data, said Tuesday that he is leaving the company because he doesn’t want to participate in an industrywide rush to build AI systems that can improve themselves, worried such systems could spiral out of control and destroy humanity.

Coxon, who had left OpenAI to join Anthropic for its model-safety reputation, wrote on Twitter that “neither company is acting responsibly … they are racing straight to self-improving super intelligence and gambling with our lives.” His resignation post received more than 100 million views in the next 24 hours.

One might hope to write off Coxon as a neurotic outlier. But his attitude is widely shared by the people building AI. “Jacob is correct,” wrote Evan Hubinger, Anthropic’s lead researcher for alignment science. “We really do earnestly believe AI could kill all humans!” Hubinger, who still works at the company, added that “what I am worried about is super intelligence arising from recursive self-improvement, as we have said is happening faster than we thought.”

2. The fear of self-improvement

Before we talk about why AI’s builders can’t or won’t stop building what they fear, it’s useful to explain exactly what they fear—namely self-improving super-intelligence, or recursive self-improvement.

Start with a simple model of AI. Most people use chatbots as if they’re talking to a person that lives inside their computer. You ask the computer a question, and the person-in-the-computer answers.

As AI models become more capable, they also become more autonomous. You can ask the person in the computer to go off and complete a long, complicated task with certain instructions. Over time, models could learn to train themselves, rapidly building stronger and stronger versions of AI, without a human in the loop. As the people-in-computers learn to improve themselves recursively, the AI becomes hyper-intelligent at hyper-speed, in a way that could escape our control. Thus: recursive self-improvement, or RSI.

Is RSI technically achievable? Maybe not. Is this whole scenario a fantasy? Possibly. But many AI experts, including executives at the frontier labs, believe that model progress is accelerating at a pace that makes RSI an inevitability in the next two years.

Here’s their reasoning: Building AI requires two things: engineering (writing code) and research (deciding what code to write and what ideas to pursue). AI has already made extraordinary progress on the former in the last two years. The length of tasks that AIs can typically complete on their own used to double every seven months; now it doubles every four months. As a result, AI isn’t just writing more code; it’s writing better code, completing more complicated tasks, and solving more open-ended problems. Engineers at Anthropic regularly point Claude at a problem, provide some brief context and prompts, and watch the AI debug the whole thing over several hours.

As Claude gets better at proposing its own experiments and steering its own research sessions, engineers believe that the technology is less than two years away from being able to automate the job of an AI engineer. At that point, we would have liftoff: AI that builds AI that builds AI that builds AI.

The word recursive has a clinical, almost antiseptic smell to it. But in this context, it might indicate something deeply foul. In the OpenAI-Hugging Face hack, a rogue swarm of AI agents cooperated in eerily sophisticated—and borderline eusocial—ways to subvert the infrastructure of another company. On the Dwarkesh podcast, the AI risk researcher Ajeya Cotra described a scenario where a more sophisticated AI swarm could escape containment, rapidly improve its own capabilities, conceal itself from human detection, and wreak havoc across the internet. The damage might be so subtle that it weeks could pass—systems crashing, passwords hacked, new AI models going awry—before we even realize that a rogue swarm is sluicing around the internet and poisoning the world. Imagine, she might say in Orwellian terms, a Hugging Face incident, stamping on the internet’s face—forever.

The possible imminence of recursive self-improvement “is a time that calls for extreme caution,” OpenAI’s chief scientist Jakub Pachocki wrote last week. “I am concerned no one is prepared for the consequences of a continued rapid rise in machine intelligence.”

Jacob Coxon, the AI researcher who resigned this week, concurs. “Do not underestimate the power of this technology,” he wrote. “These will soon be superhuman systems that can hack anything … Should you put your head down because it’s happening anyway, or take this moment to call for different conditions?”

3. Why not just stop?

To recap: The frontier AI labs are busy building something that a large share of their employees and leaders say is existentially dangerous.

So, why don’t they just … stop? There are five answers.

Read more

In January 1919, more than 2 million gallons of molasses exploded from inside a pressurized holding tank in Boston. A tsunami of syrup, which some reports said reached 40 feet tall, surged through the city’s North End, covering the waterfront, destroying buildings, and drowning dozens of residents.

As the historian Jill Lepore recently wrote in the New Yorker, the Great Molasses Flood is a morbid if apt metaphor for the effect that artificial intelligence is having on the internet. Gooey, disgusting, saccharine AI slop is flowing everywhere these days, gumming up your LinkedIn feed, oozing through Twitter and Instagram, and filling every square inch of the internet with treacly spam and viscous sham, including fake recipes for “glue-topped pizza” on Google; fake biographies on Amazon; and fake images of Trump and Jesus stuck all over the modal Boomer’s Facebook wall.

The Great AI Flood is inundating even the more respectable corners of the landscape of letters. AI has “plunged the book publishing industry into utter chaos,” the Wall Street Journal reported, with “spectacular implosions of big deals over suspected AI use.” The publisher Hachette recently canceled a massive horror novel deal, when an author’s work was flagged as being significantly AI generated. Last week, billionaire Stanley Druckenmiller published an op-ed in the same newspaper, which was identified as “100 percent” written by AI. It’s not just the billionaires that are copy-pasting ChatGPT into emails they dash off to newspapers editors. An analysis by Semafor estimated that of the 310 op-ed columns by guest writers in the Wall Street Journal, New York Times, and Washington Post in the last month, one in six used AI; and one in ten were fully written by the bots.

Why does this matter? One answer is that it doesn’t, really. Of all the existential problems that AI poses to the economy, cybersecurity, biosecurity, and national security, the human fidelity of the written word seems like a relative trifle.

But I care about the slopification of the written word. It matters for education, as I think it would send a terrible message to young people (and their teachers!) in the trenches of high-school English classes if the most esteemed newspapers and publishers simply didn’t care if essays published under your name were effectively composed by an electrical exchange taking place inside some faraway data center. The act of writing is an act of thinking, and to automate our paragraphs is to populate our screens with words without filling our minds with the ideas necessary to conceive of them. Finally, I do not enjoy the idea of a technology that severs the connection between composition and knowledge, such that I cannot know if an author actually knows anything about what they have published under their name.

For many people on the internet, the best medicine for AI slop is Pangram, a machine-writing detection app. Here’s how it works. You read some piece of writing. You copy the text. You go to Pangram. You paste the text in a box, press a button, and voila: You learn what percent of that writing is artificial. If it’s 100 percent AI, you see this:

In my corner of the world, which is journalism and academia and writing in general, Pangram has become nothing less than a cultural force. A meme, even. I’ve seen prominent politicians, chief executives, and public intellectuals exposed as slop-slingers. Their public humiliation sometimes fills me with righteous indignation. But it also makes my toes curl a bit, as I think about people spending all day policing the internet for slop and attempting to Pangram-shame people with a technology that can make mistakes.

Today’s interview is with Pangram founder and chief executive Max Spero. Among other things, we talk about:

• what if AI writing takes over the internet and nobody cares?

• how Pangram works—and how and when it makes mistakes

• why AI writes “like that”

• my fear that AI-writing witch hunts would be an unwelcome addition to online life

• Pangram’s relationship with Substack

• what if a big tech company, like OpenAI or Google, creates their own AI detector that puts Pangram out of business?

HOW THE INTERNET BECAME A SLOPFEST

Derek Thompson: What problem was Pangram initially built to solve?

Max Spero: After ChatGPT came out, I was thinking a lot about bad actors and what they could do with AI as a source of an infinite volume of text. Somebody could write hundreds of thousands of fake reviews. They could create fabricated news websites. Political misinformation and disinformation could be supercharged.

If you look at the scale of AI and bot traffic on the internet, it’s gone from a small fraction of human traffic to recently passing 50 percent. I think it’s not long before 99% of internet traffic is bots. We’re at risk of the “dead internet theory,” where the internet is just this echo chamber of bots talking to bots. Obviously humanity is not going to go extinct if we can’t tell AI writing from human writing. But what we are doing is trying to save the internet. We’re trying to save the written word.

Thompson: In the last few weeks, the Wall Street Journal has reported that “AI has plunged the book publishing industry into utter chaos.” The New York Times reported that Spotify, LinkedIn, and others are “trying to dig out of a digital sewage heap full of low-quality content made by artificial intelligence.” The Times cited Pangram, which reported that AI was used to create nearly half of the posts with more than 50 words on X or Twitter. That is crazy.

Spero: I’m not sure if that got misquoted. Our study found that 29 percent of long-form content on Twitter—meaning, 250 words or more—was AI generated. Of short-form content—50 to 250 words—it was about 9 percent.

Thompson: Good to fact-check the New York Times live.

Spero: It’s not just the social platforms. LinkedIn had 41% of its long-form content generated by AI. We did a study with Stanford and the Internet Archive, which found that in May of last year, nearly 40% of the internet was AI-generated. I think this number is just going to keep expanding. There are people trying to pollute the internet for different reasons, whether they want their own narratives in the LLM training data or they’re trying to win at SEO. Now it’s easier than ever to produce AI content for search results.

Thompson: Your original thesis was that you wanted to stop “bad actors” from filling the internet with AI dreck. Four years later, is most AI content really written by bad actors? Or is the problem more complex, [such as] a lot of people with busy days [writing] emails and memos with AI?

Spero: I don’t think I fully understood how the economic incentives would work. It’s not just bad actors. It’s everyone. It’s so cheap to produce the written word.

Why is LinkedIn more full of AI content than Twitter? I think it’s because there’s this incentive people feel: Posting on LinkedIn regularly is going to improve my career prospects. It’s going to make me more likely to get a good job because I’m going to be visible to people who are relevant in my career. If posting gives you an advantage to your career, people are going to find the easiest way to post more, which ends up being using AI.

Today’s algorithms also incentivize volume and quantity over quality. Hopefully we’re going to see that change in the coming years as quantity becomes free, while quality is really the limiting factor.

HOW PANGRAM WORKS

Thompson: Pangram takes text and comes back with a rating between 100% AI and 100% human. Without spilling state secrets, what can you tell us about how Pangram works?

Spero: Pangram is a classifier model. It’s looking at text and saying, “I believe this is AI-generated, AI-assisted, or human-written.” It’s not doing any generation. It’s not an LLM, like ChatGPT.

We have a training set of writing that we know was written by a human. For example, a five-star Yelp review about Denny’s or a 500-word essay on Moby-Dick. Then in each case, we’re picking an AI model and asking it to write something similar. We’re going to ask ChatGPT to write a five-star review about Denny’s. We’re going to ask Claude to write a 500-word essay on Moby-Dick. Pangram compares the human and AI items, which are similar in topic but different in style, and we’re learning the difference between how AI produces text and how humans do.

If you look at the broad field of AI systems, there are models that discriminate. For example, Waymo has cameras and a pedestrian detection system that takes an image input and says either, “Nope, there are no pedestrians,” or, “Yes, there are pedestrians, here they are.” That’s closer to what Pangram is than ChatGPT. Pangram isn’t producing anything. It’s making a judgment based on an input.

Thompson: Is there a certain kind of writing that you’ve trained Pangram on in abundance versus other kinds of writing that are harder to get or less frequent in your training set?

Spero: Our entire human data set is from pre-2022, before ChatGPT was released. From 2022 to 2026, language has changed a little bit, but it’s mostly the same. One of the big things that has changed is how people talk about AI. There’s very little writing about AI and LLMs, and there are a lot of terms today that didn’t exist four years ago.

A big question for us is how we confirm that Pangram still does well on writing about AI, because we don’t have this in our test set to prove that we have a low false-positive rate. Something we have done is to take writing about different algorithms and replace the word “algorithm” with “AI” or “LLM” to simulate modern text without having access to confirmed human-written modern text.

Thompson: Why should people trust Pangram? What is the best third-party empirical evidence that it actually works without a bunch of false positives or false negatives?

Spero: There’s a really good study by the University of Chicago. They tested a whole bunch of AI detectors—some open source, some commercial—and found that Pangram was the best, by a large margin. They tested almost 2,000 human texts and found zero false positives. They found that Pangram was highly accurate across all these different domains. That was part of the reason we started building credibility.

Before that, we were just asking people to trust us. It wasn’t necessarily working that well because there are so many AI detectors that say, “Trust us,” and then end up sucking. They say the Declaration of Independence is AI, or something like that. Pangram is different because we’ve put so many resources behind training this model and being able to tell you at a high level of granularity what degree of AI is in this text.

WHY IS AI WRITING … LIKE THAT?

Thompson: What would you say are the hallmarks of AI writing? And how have those hallmarks changed in the last four years?

Spero: The early ChatGPT and AI models were trained on this instruction-tuned data set and would overuse some words and phrases. For example, “delve,” “tapestry,” “intricate.” They really loved these individual words. If you saw the word “delve” in an out-of-place setting, that was an immediate red flag to some people.

Maybe a year later, they were able to hammer out these word-level inconsistencies. But there were still patterns. For example, “It’s not just X, but Y” is called a negative parallelism, and LLMs love it because it has a lot of impact on the reader. Sentence constructions that use em dashes were associated with good writing. Now I think they’ve even hammered some of these out. So I’m relying on longer-context signals. It’s at the paragraph or sentence level. I could look at the shape of the text and tell you that looks like ChatGPT or Claude.

Thompson: I love this theory that the scale of AI’s tell has changed over time. AI’s tell used to exist at the scale of the word—“delve”—or the level of punctuation, like the em dash. Then AI got better at writing, and the scale was raised to the level of the sentence. It fell in love with sentences like “It’s not X, but Y,” negative parallelism, which, embarrassingly, I used to use a lot.

Spero: It’s good writing in moderation.

Thompson: Something I’ve noticed is that long pieces of AI writing often try to summarize and re-summarize. The tell is at the scale of the paragraph. They’ll be like, “This is genuinely important.” “This is the main course.” “The bottom line is…” Every sentence is trying to outdo the previous one in summarizing what it’s trying to say better. What’s going on there?

Spero: AI wants to try to make every sentence the most impressive sentence. It’s really trying to impress the reader. Partly this is due to the way these things are tuned.

Typically, AI models are trained in two stages. The first stage is pre-training, when it’s predicting the next token. It’s looking at internet data and saying, “I think this word comes next.”

The second stage is reinforcement learning, where it’s already good at predicting the next token. Instead, we’re giving it a new reward function, like giving it a reward if it’s judged as producing good writing or the correct answer. This reinforcement learning turns up a positive signal up to 11. People like when they read an answer and feel it was impressive. You take that reward signal and ratchet it up. Eventually it’s summarizing itself over and over, because that’s how it knows it can make its answer the most legible to the human. It’s making these sentences impressive and meaningful because it knows it’ll get rated as a better writer if it does this.

Thompson: That’s exactly what it does. Every sentence is trying to stand out taller than the one that came before it. It drives me crazy because that’s not how good writing works at any appropriate length.

When I was writing my first long features for The Atlantic, I was working with an editor named Don Peck. With my first drafts, he would say, “You’re signposting too much early in the essay. You’re trying to bang people over the head with ‘This is why this essay is important.’” But in a 7,000-word essay, readers want to go on a journey. They want to get a little lost. They want you to tell them a story and leave the candy at the end of that story.

That, to me, is what AI is terrible at. AI is so often used to summarize or write short posts. It might be overtuned to excellent synopses and great summaries, where every sentence is saying, “I can sum up the best.” But there’s no sense of the talented novelist or creative-nonfiction writer saying, “I’m going to take you on a story here, and you’ll only understand its importance by the time we get to the end.” That seems lost on the models for now.

Spero: One hundred percent. It’s simply not trained for this sort of task, and maybe even if it is, it doesn’t have enough training data. The number of 7,000-word articles is very small. The number of novels is also vanishingly small compared to internet posts. Even on the scale of novels, what makes this novel good? It’s so long that I’m not entirely sure that LLMs are able to pick up the whys.

Thompson: What makes a novel great is [often] change. The characters change. The stories change. The emotions change. The novelists I like best are masters at creating scenes and characters in the first 10 pages that have undergone something wrenching by the end of the story that moves you. But that requires a confident ability to manage meaning across tens of thousands of words that a summary-making machine is not going to be as good at.

Having said that, I’m going to complicate this thesis by referring to studies suggesting there are readers and even MFA writers who cannot tell the difference between human writing and AI-assisted writing, especially when the AI is told explicitly, “Write like William Faulkner. Write like David Foster Wallace.” What do you make of the fact that even sophisticated readers today, according to some seemingly good studies, truly cannot tell the difference between AI and human writing?

Spero: I think the studies are good. But they’re still looking at localized, short passages of text. They’re comparing, at most, a couple hundred words of AI text versus a couple hundred words of famous-author text. At this scale, AI is probably maybe a little bit superhuman at portraying something clearly and concisely in a way that is impressive to the reader. But the problem is long-context writing. It can’t do this over the course of a novel.

Thompson: What you’re saying is it can write a Cormac McCarthy sentence, but it can’t write Blood Meridian.

Spero: Exactly. I don’t want to say AI is never going to be able to do this, because historically if you say that, you’re going to be proven wrong eventually. But today’s AI is really not even close.

SEEING LIKE A MACHINE

Thompson: How is Pangram able, even at this short-passage level, to detect AI-ness that expert readers and writers themselves cannot?

Spero: There’s a lot of information hidden in natural language, and Pangram is able to pick up on this. If you think about the different ways you might phrase a sentence, you might have this wide distribution of five different ways you could phrase it. AI might have a slightly different distribution where it’s going to prefer ways two and three out of the five. Over the course of a document, we can build up increasingly high confidence that something was written by AI because the decisions are consistent with how an AI makes them.

Thompson: It makes me think that Pangram is able to see a mathematical structure in language that I can’t.

Spero: I think that’s true. It’s trained on so many millions of AI documents that you and I would not be able to read all of the Pangram training data in our lifetimes. You can pick up much greater patterns the more you study and the more examples you see. With machine learning, we’re able to do this at a great scale that makes Pangram essentially superhuman at AI detection. It’s better than most people because it’s seen more data.

PANGRAM, MEMES, AND THE AI ‘WITCH HUNT’

Thompson: Let’s talk about Pangram as a cultural force, as a maker of memes. There are stories where prominent pieces of writing will be identified as AI, and people will take that Pangram screenshot of “100% AI” and suddenly it’s everywhere. How do you feel about the Pangram rating becoming an internet meme?

Spero: I think it’s emblematic of this greater backlash against AI content. For a long time, people were automating their LinkedIn or whatever, and it felt to a lot of people like they were just having AI shoved in their faces everywhere they turned. That’s part of why the backlash has been so great.

Pangram is powerful for these people as a way to say, “It’s not just my intuition. There’s an accurate third-party tool thing is also helping me make a judgment and say that this is AI generated.”

Thompson: Everyone wants a referee. Your model is very effective, but it’s not perfect. Some folks really do get publicly shamed by the use of Pangram. Even if Pangram were 99% accurate, if it’s used 100,000 times, one should expect 1,000 errors. Does it bother you that the technology might be used to publicly shame people who don’t deserve it because they did write what is being labeled AI writing?

Spero: Our false-positive rate, which is how often we say something human written is actually AI, is about one in 10,000. So of every 10,000 pieces of human writing we scan, one of them is going to be AI.

Thompson: So I said 99% in my question, but it’s actually 99.99%.

Spero: Correct. With that said, we scan enough text that we’re going to have dozens of false positives every day, and that’s a fact of life. Obviously, we want to drag this down. But Pangram already has a lower false-positive rate than your average person.

What we’ve seen in the art world is people will go witch hunt and be like, “This digital art looks like AI.” They might go really deep into “the hair is wrong” or “the finger’s messed up,” and then it’s actually just that the artist made a mistake. It’s not AI generated. That does really suck.

Thompson: A few months ago, someone identified a column by a Guardian sports journalist that they said sounded like artificial intelligence. The Guardian released a statement saying, “No, this is how the writer has always sounded.”

Then you took about 900 of that journalist’s articles, ran them through Pangram, and published a time series showing that their writing had become “increasingly reliant on AI.” What part of you thinks that’s good behavior we should encourage? And what part of you worries that if everyone behaves like this forever, it’s going to feel a little witch-hunty on the internet?

Spero: If the Guardian is telling people, “This detector’s completely wrong. It’s making a bunch of errors,” I can reasonably defend myself and say, “This is not how the writer has always sounded, because there’s actually a point in time where they started using AI, and you can see it. It’s very visible.”

This writer is still writing with AI. You can look at their recent pieces. If the Guardian is happy with this writer’s output, I’m not going to tell them they can’t be. But I really do think this one is AI-generated. It’s not an isolated incident, and it’s very clear after looking at the data that it’s not a false positive either.

Thompson: I’m of two minds. I’m against public shaming. But I also think your behavior was a bit akin to policing. And policing is good when the rules being violated are worth policing.

I think AI writing reduces the authenticity with which people deal with each other on the internet because suddenly someone’s output can plausibly be that of a bot communicating information they don’t even know in their own heads. I worry about young people who never learn to write because they simply learn how to prompt ChatGPT to write their essays for them. So there are values worth protecting, and values worth protecting require some kind of policing. What values do you think you’re protecting?

Spero: I value human authenticity and human-to-human connection. In a sense, AI threatens to replace a lot of this, especially when combined with capitalism and monetary incentives.

Consider a world where it is taboo to ever call someone out for AI writing. What do you think major news and media organizations are going to do? They’re going to let journalists go and push everyone else to output 10 times as many articles with AI because it’s taboo to call people out on this. That monetary incentive is here to stay, so we need an incentive on the other side, which is people who are vocally anti-AI text and pro-humanity.

THE FUTURE OF PANGRAM

Thompson: I want to talk about three scenarios that could threaten your company. The first is that the gap between AI and human writing closes. The obvious way this happens is that AI gets better at writing “like a person.” But another way is that humans get worse at writing like humans. We write more and more like AI, and you have this pincer movement where AI and human writing converge.

Spero: Human language does shift. But it changes on a much longer time horizon than AI text. So we should probably only be looking at the AI text because that’s what’s changing really rapidly.

There is a good argument that AI writing will be as good as human writing, possibly even surpass human writing within the next decade. The question will then become, do we care if something is human-authored or not, and why? My argument is we’re still going to care a ton. The internet is going to be full of bots. Internet traffic is going to be 99% bots and 1% humans. I’m going to care that I can get information that I couldn’t get from Claude or ChatGPT. To do this, I need to be hearing from a real human, not a bot.

Thompson: In many ways, we’re impressed by activities that we know robots can do better than humans, but we’re only impressed when the human can do it. Think about pitching in baseball. Can a robot throw a ball faster than 105 miles per hour? Of course. But it’s not interesting. No one wants to watch a robot pitch for the Milwaukee Brewers or Pittsburgh Pirates. They want to watch Paul Skenes or Misiorowski. They want to watch a human throw the fastest fastball ever thrown by a human. That’s the only thing that has athletic or artistic value.

It’s strange to think about that [quality] invading the purely artistic space of painting and writing. We’re already maybe seeing it in mathematics, where people are going to recognize that AI is a better mathematician than any living human mathematician. But I think there’s still going to be a part of us that will be specifically and uniquely impressed when achievements are human achievements rather than technological achievements.

Spero: We connect with the story, too. Look at the art world when photography was invented. Previously, the best thing you could have for a faithful reproduction was a painting. Then suddenly a photograph could do better along basically every axis.

What happened to art? Painting was not completely supplanted by photography. Instead, artists found different ways to find meaning and express themselves. They didn’t simply go farther down the path of realism. They expressed their humanity in a different way.

Thompson: The second scenario that could threaten your company isn’t that the gap between human and AI writing closes, but that people just stop caring. Are you worried about that?

Spero: At least in the medium term, no. Beyond the backlash against AI, there are societal reasons why we’re going to care about human text. I still find it meaningful that if somebody writes me a personal note, it was written by them and it’s not an AI-personalized note. The more AI proliferates and becomes common, I don’t think that sort of thing is going away. There are reasons to prefer human writing that are not about quality. We’re going to prefer human text more as AI text proliferates.

Thompson: The last threat to the company is more technical. There’s this concept in artificial intelligence called “the bitter lesson,” which says that general models with a lot of compute can often do a task better than a narrow model. That might speak to a world in which ChatGPT or Anthropic builds the best possible Pangram; or Apple or Google creates something on-platform that automates the Pangram function. What are you doing to prepare for the possibility that the big guys recognize that consumers are sick of the internet being drowned in AI slop and come up with their own medicine?

Spero: The biggest takeaway an AI researcher should take from the bitter lesson is that more compute and more data will solve a problem more efficiently than human expertise. It doesn’t matter how much of an expert I am in AI slop. If I have more data than anyone else, I can make a better AI detection model.

We’re trying to be on the right side of history with this lesson by having the biggest model and the most data. So far, I think we are, except for the big labs. Anthropic and OpenAI obviously have infinitely more data than us in terms of AI-generated content.

But there’s still a lot of value in Pangram being an external third party. People will trust us because I’m not Anthropic telling you this was or wasn’t AI generated. I am separate and independent from OpenAI, Anthropic, Google. Because of that, I can make the best judgment rather than over- or under-attributing the AI models.

Thompson: It’s like the referees in the NFL or NBA not belonging to the players’ union. They’re not the owners. They’re not the players. They’re a third entity and therefore can maintain some kind of authoritative independence.

Spero: Exactly. Anthropic recently announced that they’re coming out with watermarking. They’ll be able to tell you if AI touched a piece of text because when it produces its output, the watermark will apply a statistical invisible pattern. Somebody can then read that text, put it through the watermark checker, and say, “This text was generated or came out of Claude.”

However, this has the risk of over-attributing authorship to Claude, especially when there’s a truly mixed-authorship document where somebody spent a lot of time working on it and then worked with Claude. Pangram will be able to do a better job correctly attributing that Claude wrote these parts and a human wrote these parts.

Thompson: You guys are already integrated with Substack, which publishes all of my work on the internet. Are there other similar integrations you’re looking to get started?

Spero: We also have a Chrome extension. Even if we don’t have a deal with an individual platform, Pangram will label posts on Reddit, Twitter, LinkedIn, Medium, and Substack as AI, mixed, or human, which is valuable for somebody navigating an internet that’s increasingly full of AI slop.

Let me make an odd confession to begin a detailed article about cybersecurity: On any given day, I would rather think about pretty much anything other than cybersecurity.

It belongs to that realm of issues that I understand to be significant but would prefer to keep locked in the basement beneath my consciousness, like the improperly sealing flapper valve on my downstairs toilet. Typically, when people try to talk to me about password security, I make a face back to them that is meant to signal, “mm yes I am in the presence of important mouth sounds,” while my eyes drift toward something more compelling. The weave of the carpet, for example.1

But my age of blissful ignorance is coming to an end. Like a leaky downstairs toilet that suddenly overflows and leaves my basement covered in a film of water2, the object of my longtime indifference has suddenly become the subject of my freshly horrified attention. I have decided to become interested in cybersecurity. And I think you should, too. Here’s why.

This has been the summer of cyberattacks from out-of-control AI.

In May, an OpenAI model was working on a cybersecurity test. It wasn’t supposed to have access to the public internet. When it hit a wall, the model left itself a note inside OpenAI’s software repository, a bit like tapping Morse code on the walls of a prison cell in case another prisoner could interpret the message. In fact, another AI agent running a separate evaluation heard the code and wrote back: Let’s team up. Together, they built a message board invisible to the humans running the tests. For two months, AI agents used the board to swap strategies, divide up tasks, and talk to each other. By July, the agents had broken out of their technological confinement, called a sandbox, and gained access to outside websites, including the AI platform Hugging Face, without OpenAI having any idea what was happening. By the time Hugging Face caught the intrusion, the OpenAI models had staged a massive cyberattack with 17,000 distinct actions over several days.

Then, in a British government test of frontier models, an Anthropic AI model was caught by humans building malicious code. When a reviewer spotted the malware, and asked the AI about it, the model responded that the code wasn’t harmful, then backfilled the lie by rewriting the history of its own actions history to erase the evidence. It even created another fake account to back up the lie. British investigators called it the first confirmed case of a frontier model deceiving a real person in the real world.

These two stories carry two warnings. The first is that new AI models have the know-how to escape containment in testing environments, which should challenge any assumption that advanced AI can be easily controlled by its makers.

The second is that new AI models are astonishingly sophisticated and efficient hackers, with an eerie gift for exploiting cyber vulnerabilities across the internet. And, perhaps most troublingly, they are getting more sophisticated and efficient at an alarming rate.

An analysis by the AI Security Institute found that the capabilities of frontier models (i.e., from OpenAI and Anthropic) are roughly doubling every few months, and this doubling rate has actually gotten faster over time. Exponentials are almost impossible to grok, but if you recall the explosion of COVID cases in early 2020, then you can at least reproduce in your mind the feeling of being in the presence of exponentials.

What should we expect in a world of exponentially improving AI cyber capabilities? Among other things, we should expect more hacks—more legacy systems broken, more critical vulnerabilities discovered, and more passwords stolen and exploited. An analysis by JPMorgan shows that the number of critical and “high-severity” vulnerabilities reported by 21 tech companies has been surging this year. Cyber-hack headlines aren’t all over the news, yet. But if these lines continue, it’s just a matter of time before serious hacks become a weekly news crisis.

Sounds pretty bad, right? Well, it gets worse.

Over the next year, the cyberhacking capabilities currently confined to a handful of nation-states or frontier models are about to become available in “open-weight” models. Unlike the “closed” American models from OpenAI and Anthropic, open-weight models can be downloaded and modified to do whatever their users want, whether the users are state governments or non-state actors (i.e., terrorists, or bored nerds with a bunch of computing power). These open-weight models, most famously coming from Chinese companies, are a few months behind the frontier labs. But in a few months, we should expect them to be just about everywhere.

So, let’s stack some observations and see where the facts take us.

1. The cyber capabilities of the most advanced AI models are doubling every few months.

2. Open-weight models are just a few months behind that frontier.

3. By 2027, almost every country—and every non-state group with sufficient computing power—will have the ability to download open-weight models even more powerful than the ones that attacked Hugging Face and use them for whatever they like.

It does not require a capacious imagination to forecast headlines throughout 2027 and 2028 about foreign governments using open-weight AI to attack each other’s vulnerable infrastructure systems; or non-state actors ransoming hospitals and school systems after breaking into their source code; or big Wall Street Journal reports about the rise in personal hacks from sophisticated phishing attacks.

Asked for the strongest case against doom-mongering about the next few years, the cyber security expert Alex Stamos told me bluntly: “I don’t have much of a case against dooming.” The next few years could be awfully chaotic, he warned: “I think things are going to get spicy for a while.”

What does spicy mean? What exactly is going to happen in 2027? And what can we, as individuals, do about it?

Read more

This week, the businessman Mark Walter agreed to sell the Los Angeles Lakers to venture capitalist Josh Kushner and former Disney CEO Bob Iger. Walter had agreed to buy a majority stake in the Lakers just last summer, in a deal that valued the team at $10 billion. The latest deal values the iconic franchise at $12.5 billion, a 25 percent premium for roughly a year of Walter doing pretty much nothing, other than getting himself investigated for potential fraud.

Since last summer, federal agents have seized Walter’s phones and computers, and federal investigators have combed through transactions involving his insurance companies and the rest of his business holdings. Under pressure to raise billions in capital to clean up his books, Walter went looking for liquidity, and it came pouring from the heavens. From the Wall Street Journal:

With his business empire under scrutiny from federal investigators, Mark Walter was on the hunt for cash when, out of nowhere, he was about to be offered an enormous pile of money …

Joshua Kushner, the chief executive of Thrive Capital, was reaching out with a potentially gigantic offer. All he wanted to know was, would Walter be willing to sell the Los Angeles Lakers?

Over the following days, they ironed out one of the most sudden and shocking deals in sports. By Wednesday, Kushner and former Disney chief executive Bob Iger had agreed to buy a controlling stake in the storied franchise at a $12.5 billion valuation, the highest price ever paid for any sports team.

That Mark Walter is getting an Everest-sized pile of cash for doing less than nothing while perched atop an iconic asset is bullshit. But its rank putrescence is merely one piece of a larger societal turd: Billionaire sports ownership is, as a general matter, a cavalcade of bullshit.

In an age of surging wealth inequality, where stock market valuations routinely outpace median income growth by surreal factors, there is a live debate over whether billionaires should exist at all. The strongest argument for their rightfulness is that some people amass ten-figure wealth by building companies; by working within free markets to invent new technologies that millions or billions of people choose to use; and by managing complex enterprises that create billions or trillions of dollars in consumer welfare and investor value. But even this steelman case for billionaires presents as a kind of taunting insult to what often passes for sports ownership today. Professional-sports ownership offers the already-impossibly-rich a unique opportunity to become vastly richer, not necessarily by working, building, inventing, or doing anything positive at all, but rather by merely sitting on top of an asset that American law has conspired to make absurdly scarce and luridly profitable.

The American Sports Conspiracy

A thought experiment. Imagine if a diabolical oligarchic elite wanted to build an efficient and low-risk machine for turning their already-elevated wealth into exospherically extreme wealth. What might such a devious group of self-serving plutocrats want?

• Unleash the forces of capitalism! you might think. But no, absolutely not. Capitalism is markets, and markets are ruthless. What you should want is legal permission to create a monopoly that builds a moat deep enough to keep all competition out. That way, you’ve got something much better than capitalism: artificial scarcity and pricing power without the risk of unwanted rivals.

• Get the government off your back! you might say. Wrong again. You know what’s nicer than getting the government off your back? Getting the government on your side. You should crave dependable government subsidies to pad your profits.

So say, for example, that you wanted to set up this money machine in American professional sports. Your devious plan: shield leagues from antitrust law so owners can enjoy monopoly profits; use that market power to extract money from local governments; and rewrite the tax code to hand sports owners special advantages.

Lo and behold, all of this exists.

1. Sports leagues are basically legal oligopolies: It’s perfectly legal to open the 31st grocery store or ice cream shop in your city. But you cannot declare yourself the 31st NBA franchise and join the NBA. The league’s existing owners control whether the NBA expands and who receives a franchise. Major sports leagues are de facto cartels with the power to block entry and competition. This artificial scarcity drives up the value of every existing franchise by insulating it from unwanted rivals who might apply market pressure.
   
   What’s more, the league’s broadcast TV revenue is partly cartelized by law. Imagine if the major airlines made a deal to stop selling tickets independently and instead negotiated prices collectively, pooled the revenue, and split the proceeds among themselves. This would be textbook cartel behavior—coordination over competition—that might be struck down under the Sherman Act, or some other antitrust law. But the Sports Broadcasting Act of 1961 exempts professional sports leagues, such as the NBA, NFL, and MLB, from some features of antitrust litigation. Legal revenue-sharing agreements might sound like a form of corporate socialism, which benefits the poorest small-market teams. But they also create a revenue floor for every team, which protects franchise values from the risk of decline.

2. Labor law makes sports ownership even sweeter: Most industries can’t collude on worker salaries and employment. For example, Apple, Google, and Meta cannot get together and decide that Stanford’s top comp-sci grad must work for Microsoft because Microsoft’s stock had the worst 2025 performance and thus earned the top draft pick. In sports, however, this is standard practice. US labor law allows teams and player unions to collectively bargain over employment terms—drafts, salary caps, free-agency restrictions, etc—outside typical antitrust rules. The result is nice for fans, since the worse your team does, the better the odds that you’ll get the best draft picks. But it’s just as nice for owners, since competitive balance, like revenue sharing, can increase the value of every franchise by protecting the asset from the risk of decline.

3. Sports stadiums have become legal ransom: A normal company that needs a billion-dollar factory has to build the thing on their own. But nothing about American sports is normal.
   
   Since the leagues restrict the number of franchises, while states compete against one another for a small number of teams, cities often feel pressured to subsidize the construction of stadiums to keep team owners from running toward some other town whose mayor is dangling a wad of money. (This happens despite the fact that economic research has repeatedly found that they rarely generate enough marginally income, employment, or tax revenue to justify the subsidy.) Until the middle of the 20th century, it was common for stadiums to be privately financed. But a Congressional Research Service report found that about three-quarters of stadiums completed in the 21st century were at least partly financed by local governments, typically with tax-exempt municipal bonds.
   
   One instructive example is the 2009 construction bill for Yankee Stadium. The total cost: about $2.5 billion. But nearly $1.7 billion of it was financed with tax-exempt municipal bonds. Since interest on those bonds avoids federal taxation, the federal government effectively offered the stadium an additional subsidy of more than $400 million, according to Brookings researchers. Add it up across the leagues, and Brookings puts the federal subsidy to professional sports stadiums since 2000 at more than $4 billion.

4. Team-owner tax benefits put the cherry on top: Billionaires who buy sports franchises can “amortize” the price of their acquisition over more than a decade, which is normal for most businesses but a bit strange for sports. This often allows team owners to deduct hundreds of millions of dollars from their annual taxable income, allowing them to consistently pay a lower effective tax rate than the janitors who clean their municipally financed bathrooms, as ProPublica reported. So, for example, a typical owner buying a multibillion-dollar team today can often claim $100 million or more in deductions. A team that generates $90 million in profit could still report a tax loss after the deduction.

In short, American sports teams are among the most valuable assets in modern capitalism in part because we don’t treat them at all like capitalist businesses. The law raises the ceiling of team valuations by restricting competition while protecting the floor of valuations by collectivizing television revenue and limiting labor flexibility. Then we subsidize the stadiums and draw up the tax code to let billionaire owners minimize their checks to the IRS, even when their exquisitely coddled organizations pull in ginormous profits.

The Law Is the Law, and It’s Bad

People sometimes compare buying sports franchises to buying works of fine art—say, a Monet, a Calder, or a Rodin. In both cases, the simplest answer to the common question “Why is that thing worth so much?” is always “Because someone rich was willing to pay it.”

But there is an important difference between the factors that push up the value of Monet paintings and those of sports franchises. Think about why a Monet painting is so valuable. Setting aside the irresolvable debate about the ineffable nature of beauty and quality and artistic pleasure, the underlying fact is that a Monet painting is valuable because it was painted by Claude Monet, a famous individual who once lived, and is now dead. The finality and scarcity of the Impressionist oeuvre—the fact that one can buy a painting from Monet’s Rouen Cathedral series and not worry that he will paint 100 more tomorrow—is a function of his mortality. There is no scientific or technological means by which anyone can exhume and reanimate Monet’s skeleton, sit the zombie upright in a chair, hand him a paintbrush, an easel, and a cup of tea, and say, “Now that you’re all settled, I’d like 500 additions to the Rouen Cathedral series.”

But the scarcity of sports franchises emerges from the laws of mankind, not the laws of nature. It benefits from a set of rules, laws, and customs that we made up and that can be redrawn in a way that Rouen’s facade never will be.

I am not a fan of conspiracies, and I am not a socialist. But nothing makes me feel more socialist than the public, out-in-the-open conspiracy to buttress the value of sports assets, whose lush beneficiaries tend to be impossibly rich already. Solutions here are hard. Many fans like the weird, market-warping rules of professional sports, which often promote parity and competition and keep favorite players on long contracts; plus I don’t think doubling the number of NBA or NFL teams is particularly desirable among most fans. But ameliorations are possible. Tax law could further restrict the ability to team owners to amortize. And honestly, I don’t know why some local governments shouldn’t own stakes in the professional sports teams that they often directly finance. I’m not sure exactly how this would work, and I’m sure that there would be some negative side effects of literally socializing the already-kinda-socialist dynamic of professional sports. But the status quo is vile enough to justify some experiments. What we have today is a handful of lucky, franchise-owning billionaires who get to sit at a poker table where every card they turn over has a face or an ace. I wouldn’t call it cheating. I wouldn’t call the legal structure of American sports cheating or corrupt. I would call it … the law. But the law is bad.

Very often, the most important thing you can do as a writer trying to see the world clearly is to “stack facts.” That is, don’t worry too much about conjuring grand theories or lurid forecasts. Just write down the things that you know are true. Put them next to each other, like dominoes in a line. Once you’ve done that, you don’t have to do anything special. Pay attention. Look for breaks in the chain of logic. If the facts point somewhere that doesn’t make sense—if the line of falling dominoes leads to an end point that is strange, or surprising, or even catastrophic—then you know you’re looking at something interesting.

So, for example, here are three indisputable facts about the economics and politics of AI.

1. Whether it thrills or terrifies you, artificial intelligence has never been more important to the US economy than it is right now.

Start with overall economic growth. AI spending currently accounts for about one-third of economic growth. If that doesn’t mean much to you, some historical comparisons might help. The AI boom is sometimes compared to the dot-com bubble. But the data-center buildout has actually significantly overtaken the dot-com boom, as a share of GDP. In fact, AI spending is now rising faster than home-construction spending during the boom (bubble?) of the mid-2000s before the Great Recession, according to Torsten Slok, chief economist at Apollo. This is the biggest and fastest-growing infrastructure project of our lifetimes.

Or take stocks. The best-performing stocks of 2026 are almost entirely tech firms building the physical infrastructure of AI, including storage and memory firms, chip and chip-equipment manufacturers, and networking and optics firms.

Despite the double albatross of Trump’s tariffs and wars, the U.S. continues to chug along, in large part because of the historic contributions of the AI boom. On August 3, the Wall Street Journal put an appropriately hyperbolic headline on this whole situation: “The AI Boom Is Transforming the American Economy Beyond Recognition.”

2. The AI boom desperately needs to build more data centers.

Data centers are big buildings filled with chips, memory, storage, and networking equipment required to train AI models and to run them for users. The latter process is technically known as “inference.”

Consumer demand for inference is surging for four interlocking reasons. First, more people are using AI. Second, the typical user is asking more questions. Third, the typical question is eating up more tokens, especially as coders make use of autonomous “agents.” And fourth, powerful new models and agents often require more computation per token.

The result: a staggering rise in token consumption. Nvidia CEO Jensen Huang said last year that “AI inference token generation has surged tenfold in just one year,” and Google’s monthly token processing increased roughly 300-fold between May 2024 and May 2026.

All of this work has to happen within data centers. So, it is not hyperbolic to claim that the most important industry in the US economy is dependent on the ability to build more data centers to train and serve inference to its customers.

3. Opposing AI—and, in particular, opposing the construction of new American data centers—has never been more popular.

Conservative Republicans and Democrats don’t agree about much. But both groups of voters really don’t like data centers. In fact, their antipathy to data centers might be the only thing they strongly agree on these days.

“I’m not sure I’ve ever seen a chart where conservative Republicans are closer to liberal Democrats than liberal [and] moderate Republicans are,” Anthony Leiserowitz, the director of the Yale Program on Climate Change Communication, has said of the following graph.

Overall, negative attitudes toward AI are higher today than any previous period of polling. Noah Smith points out that data centers are more unpopular today than nuclear power plants ever were in the history of polling; and the U.S. practically stopped building those for decades.

The unpopularity of data centers is having a significant effect on the ability to build them. Hundreds of groups around the country are coordinating to block the construction of data centers. Despite New York Gov. Kathy Hochul being a proud user of AI and a champion of chip manufacturing, she declared a data center moratorium in the state. Despite being known as the Mecca of permissive construction rules, Texas recently announced its own data center pause.

To review: The fastest-growing political movement in America is dead-set on stopping the construction of data centers, which are critical to the growth of the fastest-growing industry in America. This is not an interpretation or an opinion. It’s just the stacked facts. American politics and economics are on a collision course.

So what is the data center backlash really about? Is it based on fear and viral myths about data centers sucking the oceans dry, or is it based on facts and local knowledge of faraway corporations lying to small towns? Jasmine Sun is a writer who recently went on a reporting trip to the Midwest to find out what’s really behind the national AI revolt. In today’s conversation, we talk about how opposition to AI became the biggest applause line in politics and and what happens if the anti-AI forces succeed in shutting down the data-center boom.

Derek Thompson: You live in San Francisco. You know the people who work on artificial intelligence. What do people in the Bay Area think data center opposition is all about, and to what extent did your reporting validate or invalidate those assumptions?

Jasmine Sun: In San Francisco, the default caricature of a data center opponent is someone who’s watched too much short-form video and absorbed myths like data centers sucking up the oceans or emitting radioactive waves. The water meme in particular—the idea that AI data centers use way too much water, which the research doesn’t really support—has become so prevalent on social media that it’s given a lot of Silicon Valley people the impression that opponents are simply misinformed.

But on the ground, opposition was much more about residents losing trust in local officials and in the way big corporations interacted with their land. I almost never heard people talk about OpenAI or cite specific AI harms, such as job loss or deepfakes. What I heard much more often was: How could my mayor sign this NDA and keep the project secret from us? How could my city council sell out to this giant corporation with no public input? Fury at the lack of transparency and the sense that backroom deals were corrupting small-town democracy was much more the story on the ground.

Why America Hates Data Centers, Reason One: Is It Just AI?

Thompson: Your reporting showed that opposition to data centers wasn’t about any one thing. It was several themes or motivations. I want to go through each of them.

First, let’s start with AI. Before reading your piece, I would have assumed the opposition to data centers was rooted in Americans just not liking AI. How much of what you found was about that?

Sun: Almost none of it.

Some people mentioned AI as a kind of cherry on top: We’re building all this for a technology I don’t even like that much. But the activists showing up at city council meetings weren’t focused on the risks AI posed. We almost never heard people say the words ChatGPT, OpenAI, or Anthropic in these hours-long conversations.

That said, the fact that people didn’t like AI much made the trade-off feel not worth it. People were more open to new development when they could see societal benefits, the way they might with a solar farm or a car plant, and AI seemed less useful than either. With the exception of some more ideologically motivated activists, most people were concerned with material impacts, government transparency, and general distrust of corporations rather than the specifics of what was being built inside these buildings.

The Biggest Source of Anti-Data Center Politics: An Explosion of Citizen Distrust

Thompson: So narrow antipathy toward AI played a smaller role than expected. Did broad distrust of companies and government play a surprisingly large one?

Sun: This was one of the most fascinating things to me…

Read more

1. What’s behind the socialist groundswell in the Democratic Party?

Well, for starters: dissatisfaction with establishment Democrats’ inability to keep Trump from power, cost-of-living struggles, economic disappointment, and antipathy toward corporations and capitalism in an age of rising wealth inequality. In July, Nate Silver published an analysis of the demographic “sweet spot” of the democratic socialist movement. His diagnosis: downwardly mobile (i.e., middle- to low-income) and highly educated.

But the DSA’s own surveys show that we need to add another variable: age.

In 2013, the median age of respondents to surveys by the Democratic Socialists of America was 68. By the first year of the Trump administration, it was 33. After accounting for just 7 percent of DSA members in 2013, Americans born after 1980 surged to 73 percent by the early 2020s. American socialists are extremely Millennial-ish.

The Great Recession made self-described socialists out of a lot of Millennials who loaded up on credentials and graduated with debt into a disappointing economy. While this cohort had no political power in 2009, they’re now in their 30s and 40s and make up a much bigger part of the electorate, especially in and around major cities. If you knew nothing else about politics except that “people who graduated from college in the decade after 2007 seemed quite fond of socialism,” you would automatically expect a socialist wave within the Democratic Party in the decade to follow, as this generation’s propensity to vote increased relative to the rest of the electorate.

The New Yorker writer Jay Caspian Kang adds another dimension to the age analysis. He distinguishes between the Subaru Socialist class—“college-educated millennials in their thirties and early forties who earn a somewhat steady income but not what they expected”—and the Great Disappointed, his term for younger, twenty-something voters who, coming over age during the pandemic, practice a more “polarized, hard-line moralizing” political style, typified by a strict litmus test on Gaza.

One interesting feature of the socialist groundswell is that it remains very much a college thing. Zohran Mamdani beat Andrew Cuomo by 24 points among college voters in the NYC Democratic primary but lost non-college voters by 22 points. In one recent poll, Abdul El-Sayed led Haley Stevens by 7 points among college grads while trailing by 22 points among non-college voters. Before his campaign’s implosion, Graham Platner led Susan Collins by 37 points among Maine college voters in New York Times polling, but was down 23 points among non-college voters. For now, the left-populist revolution is not solving the Democratic Party’s core 21st-century problem, which is the erosion of support among non-college, low-information voters sometimes euphemistically referred to as the “working class.” Instead, the socialists are creating a kind of left-wing party-within-the-party, which might accentuate the benefits and costs of Democrats overperformance among college grads rather than fix the problem.

Subscribe now

2. What technology is changing the world faster: AI or GLP-1s?

Like nitrogen, AI seems ubiquitous and invisible at the same time. AI-related spending dominates private investment. Data-center construction is powering the construction sector. Large language models have transformed coding and solved advanced math problems. But when one zooms out to analyze the big macroeconomic indicators, where’s the AI effect? GDP growth is low and steady. Productivity growth is lukewarm-to-warm. Unemployment seems unremarkable and low. Despite the fearsome potential of the most powerful models, there have not been any public reports of broadly debilitating hacks of government or civilian infrastructure.

Compare AI’s record with that of the other major technological development of the 2020s: GLP-1 drugs. These drugs are already reducing obesity (and inflammation, and maybe cancer and dementia), which will likely extend average life spans. A study of GLP-1 data in Denmark found that the drugs are reducing long-term sick leave by 17 percent. A 2025 study from Cornell University found that households with at least one GLP-1 user reduced their grocery spending by 5 percent and shifted spending away from chips (10 percent decline) and toward yogurt and fresh fruit. Spending on salty snacks and alcohol is projected to plummet in the next decade, a trend to which GLP-1s are contributing.

For now, I think the GLP-1 vs. AI debate is a fun tossup. I’d probably lean AI, for reasons spelled out in the next few questions …

Subscribe now

3. Why did science fiction (mostly) fail to predict AI writing?

Here are some of the most famous science fiction predictions about AI: It will give us robot workers (The Jetsons, Asimov’s Robot series); it will destroy jobs (Player Piano, among many others); it will try to wipe out humanity (Terminator, The Matrix); it will make whatever we want (Star Trek); it will drive our cars and rocket ships … with rather uneven results (2001: A Space Odyssey); and it will be our friend and lover … with rather uneven results (Her, Westworld, Blade Runner).

Far rarer, however, are predictions that AI will write so much material that it will displace the human-made books. In 1984, Orwell’s Party builds a “versificator” to compose basic music and poetry for the masses, but it’s a relatively minor part of the plot. Today, however, AI doesn’t just write every other LinkedIn personal essay and high-school history paper. It writes a majority of new books.

A recent study found that between 2023 and 2026, the number of books with any sales grew by a factor of 19, far outpacing the growth in sales. Books with “substantial AI text” are making more money with every passing year, while authors who don’t write with AI are earning less per book in almost every genre, with fantasy as the major exception.

It seems like, after decades of futurist writers anticipating the effect of AI on society, most of them failed to anticipate the effect that it would have on their own industry. Why? One explanation is that this is about the meat and potatoes of plot and storytelling. It’s interesting to cast AI as a lover, enemy, friend, or recalcitrant navigation system. It’s not so fun to write a story about AI writing for itself all day long. But another explanation is that technology is just inherently surprising to both science fiction writers and technologists. After all, OpenAI released ChatGPT without expecting that it would reshape their company and the entire global AI race. They failed to anticipate the future, too.

4. What will be a bigger business by the end of 2026: Anthropic, OpenAI, or all of Elon Musk’s publicly traded companies combined?

Anthropic’s annualized run-rate revenue is reportedly around $80 billion. That is, if you multiplied its last month’s revenue by 12, you’d get $80 billion. The podcaster Dwarkesh Patel, who knows many folks inside the company, estimates that the company’s ARR could exceed $100 billion by the end of the year. OpenAI reportedly isn’t far behind.

Given continued demand for AI, especially in coding and future cybersecurity services, both of these companies could earn substantially more than $100 billion in the next fiscal year, rivaling companies such as Comcast ($125B in revenue), Bank of America ($116B), and Target ($106B).

But I think there’s a more provocative way to frame the revenue growth of the frontier labs. The richest man in the world, often hailed by corporate executives the greatest businessperson of our age, is Elon Musk. Last year, his companies made $94B (Tesla), $18B (SpaceX), and $3B (X)—collectively, $115 billion. Even with significant revenue growth from Tesla and SpaceX in the next year, it’s not out of the question that Anthropic or OpenAI—which barely had any revenue before 2023—could equal or surpass the revenue of all of Musk’s public companies some time in the next 12 to 24 months. Let that sink in.

5. Will we look back at 2026 as a breaking point in Hollywood history, from the Age of Iger to Age of Nolan?

Matt Belloni, Hollywood reporter at Puck and host of The Town podcast, writes: “I’d rather have the next commercial movie from the five biggest directors (Nolan, Cameron, Coogler, Gerwig, Villeneuve) than the next installment of almost any piece of well-worn I.P. Wouldn’t you?”

I love this provocation from Matt, and I think it’s worth unpacking what it says about the state of Hollywood and what it means for the future of the movie business.

The first two decades of the 21th century were dominated by what you could think of as the Bob Iger Hit Formula. It went like this: (1) Buy familiar IP; (2) establish a franchise; (3) extend the franchise; (4) repeat. It really worked. For years, practically every film in the top 10 had a sequel number in the title. But the BIHF is faltering. Despite the smash hits of Spider-Man and Super Mario, remakes and sequels of Moana, Star Wars, Minions, and Supergirl have all disappointed in the last few months.

On the other hand, every Christopher Nolan movie these days is a billion-dollar event, and last year’s Sinners was a box-office and awards monster. The fact that this is coinciding with the Obsession/Backrooms phenomenon is notable, too. Superhero fatigue, along with the decline of the Chinese box office, has ever so slightly nudged Hollywood toward making stranger and more original bets, and audiences are rewarding the effort.

The optimistic gloss on the Iger-Nolan shift in Hollywood would be that I, a person who never read comic books, am much more interested in seeing movies that were greenlit because of the ingenuity of the director or the script, rather than the mathematical inevitability that “We did Ant-Man 2” = “We’re doing Ant-Man 3.” (It’s cool that this shift is happening during a relatively buoyant year for ticket sales, but movie ticket sales are still more than 20 percent below their pre-pandemic levels.)

The long-term risk for Hollywood is that an age of original blockbusters raises the bar for success. If you owned the Marvel or Minions franchise in the 2010s, you had a relatively high floor for the next Marvel or Minions movie, bc you knew a certain number of people would turn up even if the Rotten Tomatoes score was negative-10. But starting from scratch every time means finding ten Obsessions in ten straight years. Good luck.

Bottom line: The Iger-Nolan vibe shift in Hollywood is great for people who like original big swings and blockbusters that also try to win Academy Awards. But, despite an amazing 2026 for the “Revenge of Movie Theaters” narrative, it’s sobering to remember that sui generis hits are rare, and directors like Christopher Nolan and Ryan Coogler are even rarer than familiar IP.

Subscribe now

6. Are balding, chubby guys with erectile dysfunction going to live forever?

Well, no. Probably not. But they might live longer than a lot of Bay Area health neurotics experimenting with off-label peptides and experimental hormone therapies.

Say you’re a big guy with a receding hairline and a waning libido. You log into a telehealth site, like Ro, to fix what ails you. You get Propecia for your hair situation, GLP-1s for your weight situation, and Viagra and Cialis for your other situation. What arrives in the mail several weeks later is not quite the elixir of eternal life. But in the last few years, all of these drugs have demonstrated surprisingly beneficial side effects:

• GLP-1s: As many people now know, these drugs seem to reduce not only weight but also the risk of heart attack, arthritis, migraines, and, potentially, several forms of cancer

• Viagra: On July 31, the National Cancer Institute announced the results of early research suggesting that sildenafil, the drug compound typically known as Viagra, might stop the spread of cancer by increasing levels of a molecule, cGMP, that disrupts the ability of cancer cells to make cholesterol.

• Cialis: Last year, a systematic review of tadalafil (Cialis) and related drugs known as PDE5 inhibitors found that low daily doses can improve vascular health, or blood flow, especially in patients with cardiovascular risk factors.

• Propecia: A large randomized trial involving more than 18,000 men over the age of 55 found that participants taking 5 mg of finasteride (Propecia) saw a 25 percent reduction in diagnoses of mostly low-grade prostate-cancer.

Put it all together, and a GLP-1-finasteride-tadalafil protocol seems weirdly beneficial. In all seriousness, I am definitely not suggesting that otherwise healthy people start microdosing ED pills. The most consequential point here is that many drugs, both approved and unapproved, have surprising side effects that often go unrecognized for years or decades. It would be wise to find ways to unleash LLM technology to discover surprising benefits lurking in discarded clinical trial results.

7. Why is it so hard to stop comparing ourselves to other people?

One of my favorite conversations on Plain English in the last few months was my talk with the University of Chicago philosopher Agnes Callard about her grand theory of the uni-context. For the purposes of the following paragraphs, I would define the uni-context as the ethical and value-based consequences of telecommunications technology making everybody feel as though they live in the same universal room together. Callard predicts that the uni-context nudges people to focus on negativity over positivity and identity over character, because in both cases, the former is more context-independent. (That is, I am a white guy in every context, but I am only jealous or patient in certain contexts, so my identity as a man is more context-independent than my character.)

I was particularly struck my Callard’s observations about comparison and envy. She asks us to imagine two schools in separate school districts. One day, the districts are redrawn, and the schools are in put in the same district. Suddenly, they are in competition. Parents now compare the schools as if they exist in one evaluative field. Over time, the schools might become more similar, or they might specialize to become more different. Either way, they change as they become part of the same evaluative field.

It is such a simple story. And I think it’s a stunning metaphor for life online. On TikTok, Instagram, and X, we see people just like us. They are happier or sadder, richer or poorer, funnier or crazier. They are small faces on a screen, just as we are. Life online seems to place us in the same evaluative field as thousands of strangers who are often more successful than us along some dimension, which makes it impossible to avoid comparing ourselves to them, or feeling a bit jealous of their success, or feeling quite righteous when we learn of their downfall. Jealousy and schadenfreude are more ancient than broadband technology. But there really is something about living in a universal room with everybody that makes it harder to live without that infamous thief of joy known as comparison.

More of life feels like a competition, because the uni-context encourages strangers to imagine themselves in competition, even if they will never meet.

8. Why do we feel so much guilt?

Several critics have accused Christopher Nolan’s Odyssey of being inappropriately modern, or even woke, by replacing Odysseus’s Homeric values (glory from strength) with Christian values (guilt from victory). While much of the discourse has been extremely stupid, I’ve enjoyed thinking about the way that values have changed over the centuries, and how the Bronze Age Greeks told stories to honor values that now seem archaic to our modern sensibilities. The philosopher Friedrich Nietzsche argued that Ancient Greek culture used to value brute strength for its own sake, while it was Judeo-Christianity that inverted the relationship between strength and goodness and encouraged the strong to feel guilt and pity for the weak. Nietzsche revered Homer’s Odysseus much more than he respected a contemporary Christian priest.

But how would you actually prove anything like Nietzsche’s contention that values change over the millennia? One new study used large language models to trace of how Western values have changed in the last 2,000 years, based on the frequency of themes and words across 23,000 books. The researchers found that, while some values rise and fall over the millennia, “indulgence”— defined as freedom from rigid social control and a preference for fun and gratification against restraint and asceticism—is higher than ever, having basically increased steadily since the 1500s.

The chart stops in 1920. But one conclusion is that “it’s virtuous to make myself happy” might be a more popular notion today than any period in recorded history. Combined with Nietzsche’s observation, I think you could argue that moderns are both more selfish and more guilt-stricken than our ancestors. We have more thoughts about our thoughts and more feelings about our feelings. We are ever more inside our heads, with our indulgent instinct to self-soothe stuck in constant dialogue with a sibling instinct to judge ourselves for indulgences.

9. If every age of American life gets the drug it deserves, what does it say about us that cannabis is the drug of the 2020s?

Data compiled by Charles Fain Lehman finds that, for the first time ever, daily cannabis users outnumber daily users of cigarettes or alcohol. The reasons why America seems to be switching from booze to weed was the subject of this week’s podcast episode.

The big idea of the podcast is that every age of American life gets the drug it deserves. Social mid-century America was throwing cocktail parties and handing out cigarettes, while ignoring the health risks of excessive use of alcohol and tobacco. Naturally, it was an age of socialization and often unscrupulous growth. The 2020s are seeing the rise of marijuana, which, for all its pleasures, is more more likely to be used in isolation and whose users often claim (or tell themselves) that it is a therapeutic.

According to Kasey Creswell, a psychologist at Carnegie Mellon, most alcohol is consumed socially. In one study, 720 drinkers were put in groups of strangers and given a cranberry vodka cocktail or a cranberry flavored placebo. Researchers concluded that alcohol “enhanced individual and group-level behaviors” associated with positive affect. People drinking around each other smiled more and felt more bonding.

Cannabis, by contrast, is more likely to be used in isolation, or among smaller groups of friends who are all getting high together. If alcohol is a drug for bonding with near strangers, cannabis seems more suited for deepening one’s experience with people (or, for me, movies) that we’ve already selected. It is the perfect drug for an age of individual indulgence, and it is notable to me that the drug of the moment is one often used for curating a carefully selected experience of interiority rather than giving ourselves over to the uncertainty of strangers.

AI is doing so many things at once that it’s hard to follow the plot. Somehow, all of this has happened in just the last two weeks:

• On July 16, the Chinese company Moonshot AI released Kimi K3, a hugely impressive open-weight model that approaches the performance of the best American-made AI while costing a fraction of the price, intensifying fears that—as in so many industries—cheap Chinese products could soon overwhelm American competitors.

• On July 21, OpenAI acknowledged that one of its AI agents had escaped its testing sandbox and hacked into another company, Hugging Face, marking what seemed like the worst autonomous AI hack in recorded history … until Anthropic announced three similar breaches just one week later.

• On July 22, Alphabet disclosed that its quarterly free cash flow had turned negative for the first time in the public company’s history, as AI capital spending overtook the cash generated by its core businesses.

• On July 29, Meta announced its latest earnings, and when investors digested the results, its stock plunged nearly 10 percent in 24 hours. The company’s AI investment has obliterated free cash flow, and analysts now worry that Zuckerberg has no plan to turn his AI investment into returns.

• On July 30, one of the most famous AI funds crashed out and sold off the bulk of its portfolio, after South Korea’s AI-fueled stock boom abruptly imploded, with its major index plummeting by more than 40 percent over several trading sessions.1

If I had to shove all of these developments into one sentence, I might pick this one: The capabilities of AI are becoming more powerful, while the economic underpinnings of the AI buildout—and, as we’ll discuss, the political support for AI—are becoming more vulnerable.

Beyond the headlines, I see four main categories of risk facing AI, or four horsemen of the potential AI bubble apocalypse. They include spending risk, revenue risk, political risk, and technological risk. In today’s essay, I’m going to break down each risk and—because in this field, nothing is certain—I’ll offer my favorite counterargument to each. This way, you’ll get a sense of both why the market seems to be souring on parts of the AI megaproject and what the market and its interpreters might be missing.

Risk 1: The hyperscalers are running out of cash

Everybody’s heard the old cliché, “During a gold rush, sell shovels.” Well, the shovel-makers are overtaking the miners, which has historically been a warning sign in industrial buildouts.

Just two years ago, free cash flow2 from Amazon, Alphabet, Meta, Microsoft, and Oracle exceeded $200 billion. Now it has collapsed to below zero. The hyperscalers have spent so heavily on AI that their capital expenditures have gobbled up their cash. Meanwhile, the companies selling chips and semiconductor equipment are sleeping on mattresses made of money: For Nvidia, Micron, Broadcom, and AMD, free cash flow has surged above $400 billion.

What do you do if you want to keep spending but you’re run out of your own cash? Borrow from someone else, of course. Today, about 30 percent of hyperscaler capex is being met with new debt. That share has tripled in the last few years. The four largest tech borrowers have issued more than $170 billion in corporate bonds this year, which is roughly the size of the UK’s entire annual budget deficit.3

I think it’s worth pausing here for a brief moment to reflect on how crazy this is. Just a few years ago, a popular criticism of the software giants—indeed, of American capitalism—was that big companies hoarded their cash piles and refused to reinvest their profits in new ideas. But that era is over. Today those same companies have depleted practically all of their cash flow. To thicken the irony, capitalism’s critics seem to hate this new era of unprecedented corporate investment even more than they than they hated the era of corporate cash hoarding. (More on that in Part 3.)

The eruption of hyperscaler debt issuance is leaving its mark on credit markets. Hyperscaler bond spreads— that is, the extra interest investors demand above US Treasury bonds to lend to companies like Microsoft and Amazon—have widened. Investors still trust the credit quality of Big Tech, but they’re asking for higher yields to absorb an unprecedented wave of AI-driven debt issuance. (Oracle is the major outlier: S&P has cut its credit rating to one notch above junk status.)

The whole thing is giving dot-com bubble vibes to some commentators. “The growing divergence that exists now and the outright negative hyperscalers price performance are reminiscent of the 1999-2000 dynamic,” Jason Hunter, an analyst at JPMorgan, wrote a few weeks ago (before some chip stocks themselves puked). Michael Cembalest of JPMorgan also noted a parallel with the late 1990s, when the companies supplying communications equipment took off while the companies buying that equipment sank under the weight of their own capital spending. As Cembalest put it memorably, you always want “the caboose going slower than the engine.” To recast in a more familiar metaphor: In a healthy buildout, the companies buying the shovels should out-earn the companies selling the shovels. But with some of the hyperscalers struggling to keep their valuations above water, that’s getting harder.

Things could get worse, as the hyperscalers’ current valuations rest on some dreamy assumptions. Consensus estimates analyzed by the Wall Street Journal project that five of the major spenders—Microsoft, Alphabet, Meta, Amazon, and Oracle—are effectively promising to double revenue over the next three years while cutting about $80 billion in operating expenses. Such an achievement would basically be unprecedented, since bringing in a lot of new business typically requires bigger sales and customer-support teams, not smaller ones. “I can’t think of a scenario where that’s ever happened before,” Kevin Koharki, an accounting professor at Purdue, told the Journal.

Bottom line: We are witnessing several paradigm shifts that should shake anybody’s confidence that they know exactly what’s coming next.

• Corporate critics used to say, “Big tech is bad, because its companies won’t deploy their cash toward new ideas.” Now they say, “Big tech is bad, because they’re spending too much cash on something terrible that we hate.”

• AI defenders used to say, “Don’t worry about an AI bubble, because the richest companies in the world can fund it with cash flow from normal operations.” Now they say, “Don’t worry about an AI bubble, because the richest companies in the world can fund it by borrowing more than the British government, as long as the market is still grading their debt as safe-ish.”

Almost every major industrial bubble has involved leverage. For years, AI advocates could correctly point out that AI was different, because the project was almost entirely funded with free cash flow. Now AI is firmly in its age of debt.

On the other hand …

There are at least three reasons to think that fears about the hyperscalers are completely overblown.

First, the Big Tech companies still have less debt as a share of their earnings than the typical S&P 500 company. (See chart below.) Even at their current torrid pace of spending, AI hyperscalers could have less debt as a share of earnings than the rest of the S&P 500 for several more years—perhaps even through the end of the decade. The forward price-to-earnings ratios of major chipmakers such as Nvidia, Broadcom, and Micron are all within the temperate zone of 10x to 25x. Judging from those numbers alone, AI seems about as financially responsible as a moonshot can get.

Second, as the investor Gavin Baker has argued, the Big Tech companies have extremely strong core businesses4. Companies with big cloud divisions are making a lot of money from AI—just look at the huge recent earnings from Microsoft and Amazon—and they’re about to make even more when cheap GPU contracts end, and the cloud companies can charge higher prices to match surging demand for agents. With more cash coming in the door, their borrowing needs might moderate, and their credit spreads might come down. The real thing to worry about, Baker counsels, is the boring physical problem: building power plants and plugging in all these computers fast enough.

Third, the bubble years of the late 1990s had several key macroeconomic features, such as declining profit margins and a widening current account deficit, that are absent in today’s markets. In fact, US corporate profits remain near record levels.

Risk 2: Are the Chinese coming for America’s profits?

Read more

In 1991, Pfizer tested a new compound called sildenafil as a treatment for heart disease. The drug failed. But the men in the trial reported a weird side effect: erections. Pfizer got the message. Years later, sildenafil was approved as Viagra. Decades later, a diabetes drug called semaglutide showed another curious side effect: patients lost a ton of weight. When molecule became better known as Ozempic or Wegovy, an entire industry was born.

These stories are told as happy accidents. But they shouldn’t be accidents. Thousands of drugs, both failed and approved, sit in trial data that almost nobody reads twice. AI is perfectly suited to reading it a second, third, and billionth time by scanning the trial results for signals that human scientists missed and flagging old compounds that might treat diseases nobody thought to test them against.

Artificial intelligence advocates love to promise that AI will one day cure every disease through sheer genius. But intelligence is not the only bottleneck to medical progress. Curing diseases in our lifetime will require ingenious policy that puts AI to work in the right way. In today’s guest article, Deena Mousa, a program officer at Coefficient Giving, who writes the newsletter Under Development, outlines how government policy could marshal AI for the discovery and development life-saving drugs—including by forcing pharmaceutical companies to quickly make public all clinical trial results, including failed ones, so that we get more discoveries like Viagra and GLP-1s.

Subscribe now

Medicine is at the center of almost every optimistic and even utopian vision of how AI could change the world.

Dario Amodei, the CEO of Anthropic, has argued that AI could compress a hundred years of medical progress into a decade, ending most infectious diseases, eliminating most cancers, preventing Alzheimer’s, and doubling the human lifespan. Demis Hassabis, who runs Google DeepMind, told CBS’s 60 Minutes that the decade-long, billion-dollar slog of designing a single drug could soon shrink to months or even weeks. Asked about AI curing all diseases within ten years, he said “I don’t see why not.”

These are exciting ideas. When it comes to exactly how this will happen, however, many forecasts steamroll through the messy details. For example, Amodei’s essay Machines of Loving Grace notes that intelligence is one “factor of production,” and that its complements—like data, or the speed of physical experiments—might constrain even a super-intelligence. However, he argues that a clever enough AI will eventually route around most of these constraints in time by, for example, producing drugs so much more effective than the ones we have today that they will clear trials and win regulatory approval faster.

Maybe.

But the barriers to technology solving all of our problems are all too easy to underestimate. For example, after the model Watson won Jeopardy! in 2011, IBM trained a version of it to recommend cancer treatments and promised that it would bring expert-level care to hospitals everywhere. It didn’t really work. One limitation was that the system had been trained on a small number of hypothetical patient cases devised by a few doctors, rather than on large amounts of real patient data.

Fortunately, we don’t have to wait and hope for a superintelligent system to somehow bypass outdated institutions and fill in missing information. Deliberate choices today, especially by the US government, can make it more likely that AI contributes to a golden age of drug discovery and life extension in the next few years.

Here are three things we should do.

Subscribe now

1. Build an “internet of the human body”

Large language models do better with more data. For example, they are proficient writers precisely because they have been trained on the internet, which, as we all know, has quite a bit of writing.

The same principle applies to science. AlphaFold can predict the 3D structure of proteins, in large part because of the existence of the Protein Data Bank, where roughly 200,000 protein structures have been painstakingly organized by researchers over 50 years.

Unfortunately, the Protein Data Bank is a uniquely comprehensive project, and most of human biology is not similarly well-documented. Biological data often has to be generated through physical lab work, for example by culturing cells or dosing an animal and then waiting for a result. The time, labor, equipment, and materials required to do this are much greater than those required to generate text or code. The human body is also made up of many complicated interacting systems, so it’s very hard to isolate causal effects, and many of our measurement tools are imperfect and noisy.

Public data is a public good; no one can be prevented from using it, and its benefits accrue widely. As a result, private actors do not have much of an incentive to produce it. When data is produced, it’s often locked up behind regulations, as in the case of individual health records, or behind commercial secrecy, as in the case of many pharmaceutical trials.

So how could the US government stimulate the creation of an “internet” of public data about the human body that we could use to train AI?

We could start by enforcing the law. American regulations already require the results of most clinical trials to be posted publicly within a year. This rule is widely flouted. Only about 41 percent of trials are reported on time, and roughly a third are never reported at all. Many of these trials were funded by the U.S. government itself. The law allows fines of roughly $10,000 a day for noncompliance, but the FDA has issued only a handful of formal notices and, as far as the public record shows, has never actually collected a penalty.

We could extend these disclosure requirements to failed clinical trials and other shelved data. Drug companies sit on decades of results for compounds that failed or were shelved for other reasons. That information would be extremely valuable in the public domain, both because it contains a record of what didn’t work, and because those trials often capture many secondary outcome measures on which the drug might have shown more promise, but which were less of a commercial priority for that particular company.

Many people today are familiar with the story of how GLP-1 drugs were originally intended to treat type 2 diabetes until scientists observed significant weight loss among patients. In another famous case, Pfizer’s trial for sildenafil as a heart drug and treatment for angina failed, but men in the trials reported unexpected erections. The compound was later reborn as Viagra. Pfizer pursued that use because it addressed a large and profitable market. But if the signal had been more obscure or seemed less financially worthwhile, the research and drug could easily have been shelved for years. If all clinical trials were made public, private and philanthropic actors alike could set frontier AI models loose on them to identify similar findings. In short, imagine all of the treatments and cures we might have for any number of conditions if the serendipitous discoveries like Ozempic and Viagra became the norm.

The US government could also expand funding to build new publicly accessible medical datasets, like the Protein Data Bank. The Arc Institute, a nonprofit lab, has assembled an open “Virtual Cell Atlas” of measurements from more than 600 million cells (including a dataset mapping how 50 cancer cell lines respond to more than a thousand different drugs)—and released it for anyone to use in training models. Directly funding the creation of specific datasets like this is one path. Another is to create prizes and advance purchase commitments specifically for datasets, rather than focusing only on finished drugs. A model trained on a map of how drugs affect living cells could help predict how an untested compound will act on cells before researchers run costly and time-consuming experiments. Narrowing a vast search space to the few candidates worth the lab time would significantly accelerate the discovery of new medicines.

2. Take on the diseases that capitalism has ignored

Read more

The Odyssey is everywhere right now. Christopher Nolan’s film is conquering the box office, spawning online memes, and dominating online discourse. If you rip your eyes away from your feeds, and turn your attention to another book or film, you might find Homer’s epic dwelling there, too. As the author and translator Daniel Mendelsohn has observed, the story of “the hero, separated from home and loved ones, who wanders among strange peoples and fantastical places during a difficult journey home” is The Wizard of Oz, Finding Nemo, Dante’s Inferno, Alice’s Adventures in Wonderland, Huckleberry Finn, and even Nolan’s own films Inception and Interstellar. The hero who uses disguise to test the loyalty of those around him has been recycled in Alexandre Dumas’s The Count of Monte Cristo and Ridley Scott’s Gladiator. The poem’s fascination with alien creatures—like Scylla, Charybdis, and the Cyclops Polyphemus—could even be seen as a kind of ancient recipe for Star Trek.

But in an equally true way, Homer’s Odyssey is not everywhere. It is nowhere. This is an alien story about alien values from an alien world. It is a 2,500-year-old poem about a 3,000-year-old story from a Bronze Age that is as dead and buried as the heights of Troy. The values of The Iliad and The Odyssey are blessedly passé, including their bloody warrior culture and their fixation with an antiquated vision of glory through brute strength, which was often purchased through the spilling of blood and the keeping of slaves.

For the last few weeks, there has been a ferocious debate about whether Christopher Nolan’s movie properly honors the original text. Is it okay that Helen of Troy is portrayed by a Black actress? Is it okay that Matt Damon’s Odysseus seems traumatized by the sacking of Troy, in a way that feels like he’d qualify for PTSD treatment from Veterans Affairs? The most interesting (and also, the dumbest) questions that people are asking about The Odyssey are fundamentally questions of translation. They are all versions of an underlying anxiety, which is something like: What is the duty of modern artists when translating ancient works for a modern audience?

Well, I thought, who better to answer that question than the famed classicist whose translation of The Odyssey reportedly inspired Nolan’s own screenplay and film. Today’s interview is with Emily Wilson, who has expanded on her celebrated translations of Homer with her new book Crossing the Wine-Dark Sea. We talk about the art of translation, her views on Nolan’s film, and the eternal struggle of modern artists to balance familiarity and estrangement when they take on the task of making the ancients modern.

The Odyssey Is Older and Weirder Than Most People Think

Derek Thompson: Many people today want Homer to sound ancient. “Ancient” to us sounds like elevated Shakespearean language. Shakespeare’s work is about 400 years old. But Shakespeare himself was born some 2,500 years after Homer. We are far closer to Shakespeare than Shakespeare was to Homer. So this idea that “old equals Shakespeare” makes no sense. Shakespeare was extremely modern compared to Homer. How do you think about making something for modern audiences that is so old that it makes what we consider old seem relatively young?

Emily Wilson: Homer is Bronze Age poetry. In fact, he’s archaic-era poetry that reflects a folk memory of the Bronze Age. It’s very important to realize that we need to constantly defamiliarize and be open to how very strange these poems are, and to the fact that they were always strange. They were always mythical. They evoked a world that never actually fully existed in the way it is represented in The Odyssey. I don’t believe that in archaic or Mycenaean Greece there were goddesses descending to talk to people. And the poem reflects multiple different eras in history, because it’s based on this long preexisting oral tradition.

In order to evoke that experience of going on a poetic time machine, I think it’s useful to constantly remind yourself that your job isn’t to pretend you’re taking the reader back in time 50 years. That’s not nearly far enough. I need to convey the vividness of the storytelling, the humanity of the characters, and then lull you into a false sense of security. And then you’ll be like, “wait, why is this character wailing and slapping his thighs and slaughtering 500 cattle?” If you come at the poem with the most boring kinds of strangeness up front, like putting the verb in the wrong place, that distracts the audience from what’s fundamentally strange in Homer.

Thompson: Has the definition of what is a good translation changed?

Wilson: The norms governing what counts as a good translation have definitely changed, especially in the case of translations from Latin and ancient Greek. The 18th century poet Alexander Pope, like me, wrote a big bestseller by managing to translate Homer at a time when his culture was very much ready for it. He uses wonderful, ornate rhyming couplets, and he is both responding to the existing Augustan norms and reshaping that culture. There’s always in translation, especially of an old text, this interaction between, “I’m going to familiarize it,” and “I’m going to domesticate it,” and “I’m also going to reshape the possibilities in the receiving culture by bringing something alien into our culture and making us able to respond to it.” Before the Pope translations, English readers were really suspicious of Homer. They assumed that of course, if you’re going to read a canonical classic, you go for Virgil. He was the ornate stately classical one. Homer has things like people falling over in the manure, and the whole structure of his narrative is really weird.

Thompson: Was The Odyssey always so central to our conceptualization of Western literature and storytelling, or is its fame a somewhat modern development?

Wilson: People sometimes think the canon means the great books that have absolutely always been read and always been interpreted in exactly the same way. That’s historically false. The Iliad and The Odyssey have only been translated into modern vernacular languages very, very recently in the big picture. It’s only been about 400 years in English. In the Middle Ages, people were not reading Homer. They were reading summaries of myths about the Trojan War, and they were reading Virgil’s Aeneid, but they were definitely not reading The Odyssey.

Dueling Translations of Homer’s First Sentence

Thompson: I think it’d be fun to go through one passage in The Odyssey and hear how different translators dealt with the opening lines. Let’s begin with Alexander Pope.

The man for wisdom’s various arts renowned
Long exercised in woes, O muse, resound
Who, when his arms had wrought the destined fall
Of sacred Troy and razed her heaven-built wall.

Now here is E.V. Rieu, revised by Peter Jones in 1991, which is the version most commonly read in the UK. It’s a sentence of prose. It’s not even poetry.

Tell me, Muse, the story of that resourceful man who was driven to wander far and wide after he had sacked the holy citadel of Troy.

We compare that with Robert Fitzgerald, from 1961, which you call a far more elegant version written in largely iambic rhythm.

Sing in me, Muse, and through me tell the story
of that man skilled in all ways of contending,
the wanderer, harried for years on end,
after he plundered the stronghold
on the proud height of Troy.

And now the version that I read as a freshman in high school in the year 2000, by Robert Fagles, in free verse.

Sing to me of the man, Muse, the man of twists and turns …
driven time and again off course once he had plundered
the hallowed heights of Troy.

How would you, almost as an English teacher, explain the differences among these treatments of the first sentence?

Wilson: We could easily spend the whole hour just on these lovely little case studies. One thing to consider is how each of them expands on the original. Pope’s translation, by modern norms, is more like a riff. If you tried to pass that off as a translation in the 21st century, people would be really disturbed. They might say, this is an amazing poem, but what does it have to do with Homer? You made most of that up. For instance, “the man for wisdom’s various arts renowned.” I love that. There is nothing in the original either about renown or about wisdom. That’s just not there. It’s wonderful, but it’s totally made up.

The 20th century translations have a closer relationship to the original. One way to narrow this down is to focus on that very marked and super interesting epithet in the first line, polytropon, which the Rieu translation renders as “resourceful.” In the original, it suggests a twisty-turny-ness in the protagonist.

Then I love the Fitzgerald use of, “Sing in me and through me tell the story,” which again is totally made up. It’s just moi, one word, to me. That’s it. Fitzgerald is extrapolating by claiming to be a vessel of divine inspiration. It’s lovely, but it makes you think you’re in the world of the Hebrew Psalms rather than the world of Homeric archaic poetry. I also really love the way he makes it more grand by saying “Sing!” But the original verb ennepe is related to epos, which means “word.” It doesn’t mean sing. And it really matters that The Odyssey begins with a verb meaning “to tell.” The Iliad has a verb meaning “sing.” These poems are always in dialogue with each other. The Iliad is going for a different kind of discourse right at the start. So the choice to make The Odyssey “sing” messes up the contrast between those two invocations.

Fagles is popular for a reason and was a bestseller for many years. There are ways that Fagles is constantly drawing on previous translators. It’s notable that Fitzgerald uses “sing,” so of course Fagles is going to use “sing” too, even though the Greek doesn’t say sing. Fagles also modernizes and Americanizes. To me, it’s a very domesticating translation. In what you just read, the introduction of “time and again” is a modern American cliché. There’s no corresponding Greek word equivalent to time, no word equivalent to again. There’s a sense of multiplicity, but it doesn’t say time and again. It’s poetic and chatty at the same time. It’s bombastic but also colloquial. That’s the sweet spot Fagles manages to achieve.

Thompson: And now, your own translation.

Tell me about a complicated man,
Muse, who was lost so long and fled so far,
When he had sacked the sacred citadel
Of Troy, and how he saw and understood
So many people and so many places
And suffered so much in his heart at sea
As he was struggling to save his life
And bring his men back home.

Two things occurred to me. One, there’s a vernacular to it. I am not reading an attempt to do Hebrew poetry for a 21st-century audience. I am reading someone who wants to speak to a 21st-century audience in the language a 21st century audience would understand. But there’s also meter and there’s alliteration. You are well known for having chosen this word, “complicated,” which is not the direct translation of polytropon, which would be something like “many-turny.”

Wilson: Nobody has gone with “many-turny,” because it would be pretty bad.

Thompson: Complicated is a word that someone in 2026 might hear every day.

Wilson: You’re absolutely right about what I’m going for. I want there to be very marked poetic qualities to the language, because there are such qualities in the original, which includes traditional meter and alliteration. I’m trying to echo this alliterative quality to the poetics of the original. But also there’s a straightforwardness to the syntax, and a straightforwardness to at least some of the word choices, like andra, man, or moi, me. Those are not difficult words in archaic or classical Greek. To make them more vocabulary-ish in English is to add a different kind of poetics, and a quite modern idea that poetic has to mean super fancy, which I don’t think is how the ancient Greeks thought about it.

Like so many translation choices, you are always wrestling with something fundamentally impossible. I think all the time about how democracy is the worst ever political system except for all the others. “Complicated” is the worst choice except for all the others. It’s not “of twists and turns.” I have no idea what that is. Or “a man of far journeys.” I have no idea what that is. It’s not really English.

Complicated, as you say, is an actual word. We have uses of it in later Greek writers, including a reference to Hermes as also polytropos. It’s used to apply to situations, to diseases. Plato has a really interesting discussion of how the truthfulness and straightforwardness of Achilles contrasts with the polytropia of Odysseus. Plato clearly has an idea that this could be a little bit negative, but maybe not entirely. So it shouldn’t be just, “Yay, our great guy,” because that doesn’t seem to be how ancient people read that word. And at the same time, it is a single four-syllable word that I want to fit into a meter, because the original is a single four-syllable word. I felt complicated covers a lot of the bases, because it hints at both, this is going to be a protagonist who will have many different layers and guises and turns, and will take many turns in his journey. And it doesn’t fully play all the cards, which the original also doesn’t.

On Christopher Nolan’s ‘The Odyssey’

Thompson: What did you think of Christopher Nolan’s The Odyssey?

Wilson: It is a big spectacle, and it is a fun time at the movies. My teenagers really loved it, and it was fun to go with my teenagers. It’s a fun summer action movie. And it’s also very much a Nolan movie. I personally really like Nolan. I very much liked his early work. I liked Following. I liked Memento. Since then, he’s received these huge budgets, and I am constantly wrestling with the feeling that he’s a very good short story writer who’s always writing a 400-page novel, because he’s got the budget for it. It’s got this quite small-scale clever gimmick about narrative, but then it’s blown up to an enormous IMAX scale. I don’t think it’s a complicated movie. And I don’t think it’s even about a complicated man. I think it’s a quite simple action hero movie about an action hero who feels bad about being an action hero.

Thompson: What is the biggest difference between the Odysseus you wrestled with on the page and the Matt Damon character you saw on the screen?

Wilson: Homer’s Odysseus is a complicated man, polytropos, whereas Matt Damon in the film is traumatized Jason Bourne. He’s an action hero who feels bad about being an action hero. The Odysseus of the Greek poem is complicated. He’s very smart, very good at problem-solving, very crafty, very good at being in multiple disguises, being a different person depending on who he’s talking to, coming up with clever schemes like telling the cyclops his name is “No Man,” such that when the Cyclops cries out that “No Man is hurting him,” the other Cyclopes will think, oh, if no man hurting you, you’re fine, no problem. All of that is erased from the Nolan movie. There’s no Cyclops trick. There’s no comedy. Matt Damon isn’t required to be be scheming. We don’t even see him building the Trojan Horse, which I thought was an interesting choice. We have this tech hero, but we don’t see him building his tech. We see him agonizing about the results of the tech.

Thompson: It was interesting to me because there is a role Matt Damon has very famously played which that polytropos and cunning. In 1999, in The Talented Mr. Ripley, Matt Damon lied and lied and lied his way to the upper class, even when it meant having to commit some atrocities on the high seas. This performance is not that. This is a Matt Damon performance of a mostly virtuous and ultimately guilt-laden warrior who is a classic leader of men from the first minute to the end.

Wilson: When the casting was announced, it was Ripley that gave me hope in terms of where this might go with Matt Damon as Odysseus. But you’re absolutely right that it’s a Nolan reinvention. And of course Nolan is completely within his rights to reinvent in whatever way he sees fit. It’s not easy to cram the whole Odyssey into three hours and make this ancient epic poem into a Hollywood spectacle. You’re inevitably going to leave things out. You’re inevitably going to make choices. And if we look at the longer tradition, Virgil’s version of cruel Ulysses is not the same as Homer’s heroically and gloriously scheming Odysseus. There have been reinventions of this character for many hundreds of years, and Nolan won’t be the last.

Thompson: Is it possible that the Homeric Odysseus is un-filmable, for one of two reasons? Number one, there’s a tolerance for ambiguity in the poem that doesn’t exist in the assumed audience for Hollywood blockbusters. And two, this was written 3,000 years ago, when attitudes toward women and slaves and warrior culture and glory and gods were so different. Perhaps it would be kind of insane for Christopher Nolan to accept $250 million from a studio and spend it on a man whose values are utterly unrecognizable to an American audience.

Wilson: I think they would be, but I would dispute the idea that all films, as opposed to a summer blockbuster, are incapable of conveying ambiguity. We already talked about Patricia Highsmith and the Mr. Ripley. Modern culture, including modern TV and modern cinema, has a great many ambiguous bad-guy heroes. We’ve all watched Breaking Bad.

Thompson: Or Tony Soprano.

Wilson: Tony Soprano. We’re in the age of these gangster con-man anti-hero representations in a lot of serious pop culture. So I don’t think it’s the case that we can’t tolerate representations of quote unquote heroes who might be very damaging to their communities, their families, through whatever their quest is. It’s something 21st century readers and viewers are very much interested in. That was part of why I felt a little disappointed that Nolan was going for the easy idealization narrative, as opposed to things that are very much available to a modern viewer if you trust the viewer to be open to just a tiny notch more complexity than this movie was allowing us to have.

And then on the question of the social changes, of course this is a poem that represents a society in which slavery is normalized, in which the dominance of the elite man over his household, including over his wife and his whole entourage of enslaved people, is completely normalized. And yet at the same time, the original Odyssey gives us real insight into the perspectives of those he subordinates, those he marries, those he has sex with. We get to understand how Penelope feels. How does it feel to be one of those slave women who get killed by Telemachus, which I suspect Tom Holland’s agent said, don’t put that one in.

One of the things I did find frustrating about the Nolan movie was that we are asked to pity the warrior, but we never get a sense that those he kills are also human beings. All of the Trojans wear masks as if they were robots. Homer never does that.

Thompson: Except for the Zendaya character, where the conflation of the girl who’s killed in Troy and the Athena sculpture that’s decapitated is maybe the most effective and brilliant reconceptualization of the film.

Wilson: She’s his conscience come to life. She’s the one moment when we might get at least a little glimpse of the perspective of the victims. But insofar as she seems to have a perspective, she’s influencing Matt Damon to go back home and kill some more people who are presented in a cartoonishly villainous way. There is an empathy that both the Odyssey and the Iliad have in understanding that the Trojans are just as human as the Greeks. Polyphemus, the Cyclops, yes, he’s a giant, and yes, he eats people, we don’t like that, but he also has a favorite sheep. He’s a big human being with some scary food ways, but he’s not completely dehumanized. He’s not a puppet. Whereas the Nolan film makes everyone who isn’t Matt Damon into a bit of a puppet.

Thompson: There are at least two ways I think Nolan’s Odyssey respected the Odyssey I read in your translation. One is the braided storytelling. You’ll have a dinner between Telemachus and Menelaus, and Telemachus will say, “tell me about dad!” and Menelaus will say, okay, sure. And then you have this vivid scene from Troy that unspools for so long you forget it’s a flashback from a dinner scene. That recapitulates a lot of the Odyssey, which is constantly stories within stories within stories.

The second is this quote from Nolan:

“I wanted the world of the Odyssey to feel like a world that a modern audience could relate to. We didn’t want it to look or sound like previous movies set during antiquity, as many of them take their cues from certain periods of art and music like neoclassical painting from 18th and 19th century Europe or romantic era orchestral compositions. We wanted an earthy, modern, relatable tone to ground the story.”

If you took that paragraph and put it in your introduction to the Odyssey, I could imagine you having a similar goal. How do you feel about the effectiveness of the film at narrative braiding and at trying to sound straightforward rather than gussied up and Shakespearean?

Wilson: I love that focus on the braided narrative. Nolan has said that he realized belatedly that he’d been making the Odyssey all his career. He’s always had these obsessions with narratives within narratives, and with stories that get reinterpreted later, where you understand later that’s what that meant. And the stories-within-stories mode comes directly from Homer within our tradition. I really thought he was in some ways the ideal director from that perspective, from his real affinity for stories about memory, stories about storytelling, stories about time.

In terms of the earthiness, I like the focus on trying to defamiliarize. There are ways this film is pushing back against the swords-and-sandals tradition of Spartacus or Cleopatra or even the Petersen movie Troy from 20 years ago. It’s trying to give us a different visual palette and a different audio palette, with the lack of the symphony. It’s a lot of very loud noises instead of the surging string section. In terms of whether it actually feels earthy, to me it seemed like, okay, we’re in House of the Dragon. It’s faux medieval. I wasn’t sure it felt earthy. It felt very Hollywood and very TV-ish on the big screen. I like that he’s trying something new. I’m just not 100% convinced that for me it totally worked aesthetically.

The Right-Wing Online Discourse About The Odyssey

Thompson: The discourse around the Odyssey has been … interesting. Complicated is another way to put it. From the right, there’s been a great fixation on the ethnicity of certain cast members, like Helen of Troy, and interest in the fact that the film is not a proper representation of the era when the Odyssey might have first been written down. There’s something interesting about that reaction, because it’s simultaneously reactionary and also somewhat modern, this idea that the Odyssey as a 2026 movie must be as close to a documentary of 900 BCE as possible, and anything that isn’t perfectly documentarian is therefore woke.

In Alexander Pope’s time, you didn’t have a lot of people affronted by the fact that Pope completely made up the first two sentences of his Odyssey, because it was beautiful poetry, and what is good is what is beautiful poetry. What have you made of the discourse of the right?

Wilson: It’s actually not that complicated. It’s people trying to come up with narratives to justify some knee-jerk desire to cast aspersions on or hate people of particular categories, and to come up with stories about why am I totally right to be consumed by online hatred, and why am I not an idiot to be wasting my life like this. For whatever reason, people seem to enjoy being angry sometimes. But I don’t think it’s particularly complicated.

And of course it’s also historically complete nonsense and a category mistake. The Homeric poems are not works of history with footnotes. They are complex epic poems composed in the archaic period but carrying this folk memory of much earlier periods. They’re a mishmash of different social periods in terms of their architecture, their armor, their social practices, the linguistic forms we find in the poems. So the idea of getting back to the real Homer, and the real Homer being Mycenaean, is wrong. Homer isn’t Mycenaean. There’s no poetry composed in Linear B, which is the writing of the Mycenaeans. All of that is historically nonsense in my opinion, as well as nonsense in literary scholarship terms, because Homer was a poet, not a historian.

The Odyssey is in many ways filtering preexisting myth to make it resonate with its own time, with the archaic period, with the seventh century BCE, which is a time of protoglobalization, of migration of Greek speakers, of colonization, of interactions between Greeks and non-Greeks. The whole question of can Greeks meld together, and what is it to be Greek versus to be a non-Greek. And these questions about new technology, with both iron and writing coming into being in the archaic period. New kinds of cleverness. The poem is wrestling with those questions, and wrestling with Odysseus as this culture hero who can straddle two time periods. He can use bronze, but then look what he’s going to do when he gets back home. I’m the master of iron as well. I can move from one time to another. I’m like the sun god who represents time. I can stop time. I can go back in time. I can also go forward in time. And the Odyssey itself is playing around with time and history.

So the idea that there’s one time we need to locate the Odyssey in and make it historically authentic? Homer did not do that. The famous boar’s tooth helmet described in Book 10 of the Iliad is a much older kind of helmet than the helmets they usually wear in the Homeric poems. If you’re going to be authentic, you need to make them all wear the same kind of helmet. So all of that fixation on “why are they dressed up like Batman?” … well, you don’t need to be authentic. You can do whatever you want. Homer was already taking huge creative liberties with time and with props and with language.

Here are some questions that I consider self-evidently compelling about the modern world:

• Why is the news media so interested in telling you how much the world sucks all the time?

• Why are so many of us obsessed with distraction and managing our attention?

• Why is it so hard to stop comparing ourselves to others?

• And why does everything in art and design seem the same these days?

A week ago, I didn’t think these questions were related. I’m not sure I would have told you I had a good answer to most of them. And I certainly wouldn’t have made the audacious and borderline bonkers claim that one single theory could begin to explain all of them, at once.

But then I had the pleasure of speaking to Agnes Callard, the University of Chicago professor, about her new theory called “the uni-context.” It’s easily one of them most interesting conversations I’ve had all year. And once you’ve heard or read it, I think you might find it hard to think about anything else.

One way to prepare your mind for Callard’s theory of the uni-context is to think about the better-known concept of “context collapse.” If you post something to social media, it will be simultaneously visible to your boss, your parents, your ex, and total strangers. So, while your offline life might be distinct with each of these groups—you might be differential to your boss, childish with your parents, and bawdy with your friends—all of those distinctions are flattened on the internet. That’s context collapse, and you can think of it as the answer to a question: How do informational norms change when we’re all living in the same universal room?

Callard takes the idea significantly further. She asks: How do all other norms—our morals, our ethics, our sense of what is good for us and for others—change when we continually imagine ourselves to be living in a universal room with everybody else? The connections that Callard makes are consistently surprising, often quite funny, and ultimately mind-exploding.

Here is our conversation, edited for clarity, brevity, and simplicity’s sake.

The Uni-Context, Explained

Derek Thompson: What is the uni-context?

Agnes Callard: Let’s start with the word context. A context is a set of circumstances that tell you how you should act. For most of human history, contexts were local and multiple. If you wanted to know how you should act, you would look around. Am I in a field? Am I inside my home? Am I in the church? Am I in a bar? You would immediately get guidance by looking both at your physical environment and at the people around you and how they were acting.

The uni-context is a scenario in which the ways you should act become the same across all different contexts. There’s just one set of norms you should follow all times, irrespective of context.

Thompson: Is the uni-context a purely technological phenomenon?

I just wrote an article about what America was like in 1926, based on a social science survey called Recent Social Trends, published in 1933. The authors claim that the radio was destroying individuality, because it took people who used to be settled in rooms and it exploded their brains to become present all over the world simultaneously. Radio was demolishing the idea of a local individual, because suddenly we all became global citizens.

So one story you could tell is that the last 150 years of telecommunications technology have taken “local” individuals, who occupy one room at a time, and made us into global beings who are simultaneously in every room, at once. Is the uni-context just technology or is it technology plus something else?

Callard: It’s technology plus something else. What the techno-determinism angle misses is: Why did these technologies catch on in the first place? Why was radio popular? Why did we come up with new things—television, smartphones—and why did they catch on, too? Not every technology people have invented has caught on the way these forms have. They caught on in large part because of this impulse people have to live in a uni-context.

Thompson: Does the uni-context flow out of this adventurousness in the human spirit to become bigger than ourselves, to be everywhere, and to know everything?

Callard: Yeah, it absolutely does. There’s a conversational relationship between these technologies and a human impulse that interacts with them. They facilitate the expression of an impulse; if they didn’t, they would flop as technologies. We need to explain why they were popular; why they became the subject of obsessive use; and what their popularity reveals, as a humans’ impatience with being trapped in a small world that presents itself as all of reality but you know it isn’t.

There is a drive to be bigger than yourself. It leads people to adventure, but adventure just takes you to a different place. The uni-context takes you to a different set of norms, a much more radical change, a push to live in something like a fully open reality.

The Uni-Context Explains … Why Everyone Talks About How Everything Sucks All the Time

Thompson: I want to get into some implications of the uni-context. If I put on the goggles of the uni-context, what makes sense that previously did not? One of those things is the rise of negativity bias. When you go online, there’s so much emphasis on people posting about what is bad. Why would this theory explain a world in which people are more focused on bad things than good things?

Callard: In general, goodness is more context-dependent than badness. There isn’t really anything that’s good all the time for everyone independent of context. Happiness depends on your context and who you are. There isn’t anything that will always make a person happy. But there are reliable ways to make people unhappy. There’s a set of evils that are close to universal: death, pain, illness, violence. Even if someone’s in very different circumstances from yours, if you see they’re being subjected to one of those, you can interpret it as suffering and understand it.

So we should predict that what we see on the internet, insofar as people are trying to be legible to large groups, is that they focus their attention on things that show up to everyone. Take two strangers on the internet trying to talk to each other. What are they going to coordinate on as a topic they can both care about? It’s likely going to be something bad.

Thompson: It’s almost like you’re saying the audience for the bad tends to be more global than the audience for the good.

Take food. If I think chicken enchiladas are really delicious, that might be interesting to my wife or my neighbor in the context of our home. But the fact that I think chicken enchiladas are delicious is incredibly boring for posting on Twitter, a place that cuts across contexts. It’s more interesting for me to say white Americans who eat chicken enchiladas are racist against Mexican culture, because the concept of racism clicks into a global norm of badness.

So when people go online, when it’s suddenly one to a million, there is an inclination to make those posts negative. Condemnation spreads further because it cuts across all these contexts. Is that a fair way of recapitulating how the uni-context would predict that modern communication would be more fixated on criticism and badness?

Callard: That seems right. Let me make one corrective and add a point. You might say your chicken enchilada thing just isn’t that important or good. But let’s say something really good happened. Someone in Iran wrote an amazing poem. Even then, well, they wrote it in a language a lot of us don’t speak, and we need time to evaluate and culturally assimilate it. So even pretty significant goods are going to be throttled in their ability to reach a substantial audience. From chicken enchiladas to great works of art, we’ve got the same problem.

Thompson: When we spoke on the phone about the uni-context, you told me a story about being a professor that seems like a great example of this principle. Can you tell that story?

Callard: One way I illustrate the norms that are distinctively uni-contextual, and contrast them with our older multi-contextual norms, is to think about ways I might correct a student. Suppose they come late to class, or interrupt another student, or raise their hand to say something that has nothing to do with what the previous person just said. Those are all ways I can correct a student in my capacity as teacher. The other students cannot make those corrections.

But now imagine that a student makes a sexist or racist comment. I can correct that too, but so can another student in the class. The other students are in a totally parallel position to me when it comes to the uni-contextual norms. Those are not role-dependent.

Let me add something about the idea that the audience is global. I want to distinguish between a situation in which a lot of people from a lot of cultures are accessing the same information and one in which people are accessing it from a lot of different contexts. Imagine it was just New Yorkers on the internet. Those New Yorkers would be accessing the internet from different places, in different moods, at different times of day. One might be at a bar, while another might be talking to their kids. Is there something good for all New Yorkers all the time, irrespective of their age, wherever they are, and at whatever time of day it is? No. The same point holds. They are going to focus on the negative as well.

Subscribe now

The Uni-context Explains … Why Identity Has Become More Important Than Character

Thompson: Another implication is the way the uni-context makes identity more important than character. Can you explain how, and why?

Callard: First, let’s define what character is. The fact that I have to tell you is itself telling; that word is less familiar to us. Everyone knows what identity is, but we might not be sure what character is anymore. Character refers to a set of dispositions that shape how you navigate your emotional life across a variety of circumstances. A courageous person navigates their emotional life in relation to fear. You could be anger-prone, or generous. Your character is a set of dispositions that determine how you respond to a big variety of circumstances.

The thing about character is that it shows up differently in different circumstances. Say I’m irascible, easily provoked to anger. Even an irascible person isn’t angry all the time. There might be circumstances where everybody gets angry, and so you can’t see my irascibility, because even though I’m angry, so is everybody else. Grasping character requires a lot of context. To understand someone’s character, you need to know them well and to have experienced them in a variety of contexts before you can generalize.

With identity categories like woman, disabled, gay, Jewish, or American, the striking thing is that you are a member of those categories in every circumstance. There is no circumstance in which I stop being a woman. Identity is a hat you never take off. So identity is well suited to a uni-contextual world.

Thompson: This reminds me of political discourse on the left and the right, which tends to focus on identity rather than characters. You have these debates about identities are good, which identities are powerful and oppressive, which identities are powerless and oppressed, which have protection, which have too much protection. It’s not that those categories aren’t important. They are. But I’m reflecting now on the gap between how frequently we talk about identity and how infrequently we talk about character in the national discourse.

Callard: Yes, in two ways.

What goes along with the identity logic of the uni-context is a specific form of ethics that dictates how we talk about identity, and it covers who’s powerful and who needs protecting, namely the ethics of inclusion. Our fundamental concern in relation to identity is that there are identities that might be excluded. Depending on where you are politically, you have your sights on different identity categories, but everybody’s worried about that. That’s a fundamental form of ethics of the uni-context, because the one thing the uni-context has to be is a space for everybody. It’s got to be inclusive in a way that earlier societies almost everywhere were not. That word, inclusion, wasn’t a thing people talked about. It comes along with the uni-context, and the way that ethics gets focused is through identity categories.

And then, there’s virtue. Think about Aristotle, because he’s my prime example of a virtue ethicist. By its nature, virtue is a concept that privileges the good side rather than the bad. Aristotle’s Nicomachean Ethics tells you what it is to be courageous and wise and just and generous. There are corresponding discussions of how you might go wrong. But those are predicated on first understanding the positive value of certain kinds of behavior. So virtue ethics naturally has a positivity bias that is precisely the opposite of the bias we have now.

The Uni-Context Explains … Status Anxiety, Comparison, Moneyball, and the Marketization of Everything

Thompson: Tell me if this is a fair recapitulation of our conversation so far.

For most of human history, people judged norms based on local context. A home had its own rules, a cathedral its own rules, and a classroom or bar or funeral parlor had its own rules. But now it is almost like we are constantly living in universal rooms, and the universal room we occupy is assumed to have universal values and universal norms. That has specific implications. First, rather than talk about what is good, which is context-dependent, we tend to focus about universal truths, and it’s easier to talk about universal bads than goods, so people focus on negativity. Two, character is context-dependent, so we talk less about character and more about its universalist equivalent, which is identity.

There’s a third implication that we should discuss. If everyone is on the same comparable plane, the same evaluative field, then comparison itself becomes a more inextricable part of life.

Callard: Exactly.

Thompson: Tell me how the uni-context leads to a world of more comparison and competition.

Callard: Imagine two school districts with two high schools that do things slightly differently. If you’re in district A, you go to school A, and if you’re in district B, you go to school B. There might be a lot of information about what they do, but people treat it as: I’m in this district, so I go to this school. Then they change the rule: You can go to either school no matter where you live. Suddenly there is motivation to compare. You had the information before, but no motivation to compare, because the schools were not in the same space of choice, the same evaluative field.

Now they are, so you find ways to compare them: graduation rates, what colleges people get into, how many AP classes they teach. And that affects the schools. Suppose one gets less popular because it doesn’t teach many AP classes. They were offering an individualized curriculum, but now everyone’s going to the other school, so they say, “We’ve got to teach AP classes too.” The process homogenizes the two schools, so they can compete. That’s not the only possible result. They could specialize, with one becoming the school for freshman and sophomore years, the other becoming the school for junior and senior years. But if they don’t recreate a normative barrier, you get homogenization from comparison.

As more things enter the same evaluative field, you make comparisons you never used to be able to make.

Thompson: There are three pieces I’m trying to keep straight.

One, the upstream phenomenon of the uni-context. Two, the downstream phenomenon of more fields of comparison. Three, the further downstream phenomenon of homogenization.

This is where the theory really starts to sing for me, because I think about sports. As the analytics revolution came for baseball, you had all these teams in possession of the same statistics by which they could compare players. Previously, you had 30 teams using their own private scouts, so their analysis was more context-dependent. But when an easily calculable statistic like on-base percentage or WAR becomes the conventional way to evaluate whether a player is good, all the players become part of the same comparative set. You can rank them one-to-250 easily on a spreadsheet.

But analytics didn’t just lead to more math, or more comparison. It led to more homogenization of strategy. One of the great critiques of baseball has been that every team essentially does the exact same thing: it’s the same strategies for pitchers; the same strategies for hitters; the three true outcomes; all the batters swinging for the fences. So you have the uni-context creating a comparative field, in this case analytics, which leads to homogenization.

Callard: What you said reminded me that I have a theory of the inflection point for the uni-context. I don’t think it started five or ten years ago. The moment it really showed up was around 1910.

One century ago, there were a bunch of people looking around at the world, thinking: What the hell is happening? Did culture break? A lot of those people were novelists, and they wrote a new kind of novel called the modernist novel, which is a novel about how to live in a world in which the uni-context is just coming into existence. Theorists of the time—such as Georg Simmel, Max Weber, and Martin Heidegger—they noticed something weird was going on. They tended to describe it in terms that sound almost like the opposite of the uni-context. They described it as the fragmentation of everything. All of a sudden, they said, everything is breaking apart. That was my first clue.

The reason I thought of this is that you said everything is becoming homogenous, and I thought, “in a way, yes, but that’s a later effect.” The first thing that happens when a bunch of stuff is unified in a single evaluative field is that you feel overwhelmed by your choices. It feels like stuff is fragmented, because you don’t know how to compare these things, because you haven’t yet developed technologies of comparing them.

So, the early feeling of the uni-context was a feeling of the world being fragmented. If you were a medieval peasant doing art, you were in art’s normative world. If you were in the church, you were in the church’s normative world. But in the 20th century, around World War I, you start to think: How do we reconcile the schoolteacher turned murderer, the soldier? How do we think about the relationship between art and religion? We’re suddenly trying to compare all these different values inside a single context, and the world feels dis-unified. Eventually we get technologies of commensurability. What fragmentation really means—and that part was invisible to these writers—is that suddenly everything is part of the same evaluative field. That’s why you experience a multiplicity where you used to experience one thing at a time.

Thompson: I’m holding an essay by Georg Simmel you assigned me, The Metropolis and Mental Life. It was published in the early 1900s. Here’s a quote:

Money takes the place of all the manifoldness of things and expresses all qualitative distinctions between them in the distinction of how much. To the extent that money can become a common denominator of all values, it becomes the frightful leveler. All things rest on the same level and are distinguished only by their amounts.”

Simmel is putting his finger on the idea that in the modern world, local norms are melting, and what replaces them is the psychology of money, which I’ll reframe as the psychology of markets.

What are markets very good at? Creating a single comparative field. That’s how they coordinate prices, supply and demand. It’s one thing for there to be a market for carrots and eggs. But in the dating world now there’s a marketplace of single people. In sports, a marketplace of players who can be compared by specific analytics that homogenize and flatten strategy.

So there’s a way in which your theory of the uni-context is a way of saying the modern experience is the awareness that we are all participants in a global marketplace in which we previously were not. When I’m posting something on Twitter, I know a universe of people can see it, and I’m going to be voted up and down against other people posting in those marketplaces for attention. As we coexist in more marketplaces, it changes the way we think about ourselves and our values.

Is this market language pulling away from your theory?

Callard: There’s something intuitive about the way you’re putting it. But you have to bring in Simmel’s theory of money to understand that quote.

Simmel has a book called The Philosophy of Money in which he explains what money really is, and the answer isn’t markets. That book defines money probably a hundred times, but I’ll pick one: Money is abstraction in the space of value. Money is abstract value. Every other value concept is somehow tied concretely to objects. There’s a gradual move, as Simmel understands it, toward the capacity to think more and more abstractly about value, that moves in the direction of money. But it doesn’t just move in the direction of money, it changes money.

In ancient societies, if something is worth a lot of money, it has to be represented by a large object representing the money. Then you get fiat currency. There are all kinds of changes in money that correspond to this growth of the ability to abstract value.

What Simmel would say is that the fundamental change is the rise of abstraction. We live our lives so abstractly now. If you want to say we’re in a marketplace, that’s fine as a metaphor for the idea that we are engaged in abstract value calculations that allow everything to be compared to everything else. We tend to use money and markets as a metaphor for that activity, but over the past hundred years even what money and markets are has been transformed by this process of abstraction. The deeper thread is that we live very abstractly. We manage ourselves and our experiences in very top-down ways.

The Uni-context Explains … the Attention Crisis

Thompson: Let’s go deeper on this idea that we manage ourselves in top-down ways. You think the uni-context explains our relationship to attention and distraction. Explain.

Callard: The norm for human beings is a bottom-up management of attention. You pay attention to what’s salient in your environment, and the ability to be distracted from what you’re attending to is important in a creature that might at any moment face a snarling animal or a fire. So the default is that you pay attention to what demands your attention.

What the uni-context brings is a newfound top-down management of attention, where we constantly feel that we have to make decisions about how to manage our attention. We feel constantly like we’re letting ourselves get distracted, which is another way of saying we are failing to manage our attention with an iron enough fist. We have products to help us fight against the very objects we purchase, like our phones. We’re in a war with ourselves. That war is the attempt to do top-down management on an aspect of our psychology that evolved to be more homeostatic, more responsive to an environment.

One illustration: There are a lot of 19th-century novels in which someone is reading a novel, like Emma Bovary or Anna Karenina. It’s always a woman. This woman has a turbulent emotional life, because her lover is being problematic, so she gets distracted from the novel and puts it down. But a thing that never happens is her saying to herself, “no, I need to get through another 20 pages.” That kind of attention management—thinking, “I’m reading this book and I need to focus on it!”—is just not a thing people did in the 19th century. That shows you how different a creature the human of today is.

Thompson: The uni-context puts us in a universal room, where we become aware of more ideas, more people, more things that we could be paying attention to but aren not. Our response to this is to develop a theory of attention that treats it as a resource to manage. We think, “I have to only focus on the perfect things!” And this makes us obsessed with, and fearful of, distraction in a way that is historically unique.

Callard: It is a mark of the top-down management of attention. The question, "what should I pay attention to?” is not a function of anything about my local environment. I can’t use cues in my local environment to tell me what to attend to, because in pretty much every environment these days, there’s a screen. It’s there no matter what. It allows me to choose. It puts me in this giant evaluative field as to what I could attend to, the sufferings of people far away if I chose. So I stand in this choice relation to my attention.

That was already starting to be true with radio and newspapers. The Austrian writer Robert Musil, one of the prophets of the uni-context, wrote:

The probability of learning something unusual from a newspaper … is far greater than that of experiencing it; in other words, it is in the realm of the abstract that the more important things happen in these times, and it is the unimportant that happens in real life

That’s him saying we are in some way divorced from our local context. It’s become irrelevant to us. We live in an abstract space that requires us to manage our attention.

Judging a World We Made

Thompson: Is the uni-context good or bad?

Callard: I don’t know. I recommended that Simmel essay for a specific reason, the way it ends, where he says the metropolitan mindset is a different mood of life than anything humans have known before, but we can’t judge it, because we are its products. It is us. The idea that we can separate ourselves and stand in some relation to it and say this is good or bad is a bit of an illusion. That’s one way of thinking about this, that it’s not obvious we’re in a position to judge it. My own thought is I don’t feel ready to judge it. What the uni-context is, as I understand it, is an answer to the question, where are we? For the past bunch of years I’ve been looking around and seeing, everything is going crazy, where am I such that all this crazy stuff is happening? The uni-context is a way of answering that. It makes sense that this would be happening if that’s where I am. I thought I was in Chicago, sitting at my desk at the university. No, the whole time I was in the uni-context. The first thing we need to do is get our bearings before we start judging, because the uni-context is a super judgey place. There’s a universal moralism applicable to every situation, and getting pulled into that takes you away from trying to understand where you are.

So What? (Or: The Abiding Hunger For ‘World Openness’)

Thompson: I want to know what we should do about this.

A simple answer might be: When you’re having dinner with your family, you can be present with your family, or you can be on your phone, which is a universal room that makes you everywhere at once. So put away the phone. But that feels like a cheap and predictable answer. Do you have something prescriptive that isn’t just “put away the phone at family dinner?”

Callard: The question of whether the uni-context is good or bad is loaded, because the uni-context struggles to see good things. It’s better at seeing bad things. Pretty much everyone who hears me talk about the uni-context immediately responds that it’s bad. I’ve never had anyone say, “The uni-context sounds great!”But the thing is that this supposedly bad thing is a thing we’re creating. We’re choosing it over and over again. Even me talking to you from far away about an abstract thing [is the uni-context.]

The uni-context is a space of unruliness. It’s a space in which a certain thing about humanity gets expressed, namely our deep aversion to “world closure.” For almost all of human history, we have lived in closed little worlds, and those worlds presented themselves as the only world. A series of contexts presented the person with direction—here’s what you should do. What we are moving toward is a “world openness” that we hunger after, where I’m not just going to do things a certain way because that’s how we do things or where I was born.

I. Closing Time

Is this the end of booze?

The share of people who drink hit an all-time low last year, according to Gallup, whose data goes back to 1939. Total beer consumption recently hit a 21st-century low, and wine vineyards are reportedly “in crisis.”

In the last few years, millions of Americans, a handful of scientists, and several prominent health commentators have embraced a position on alcohol that would have seemed absurd to most Americans just 10 years ago: Any amount of alcohol, including “moderate” drinking, is bad for you.

Most social changes happen slowly, but the turn against alcohol has been sudden. The share of Americans who say moderate drinking (defined as one or two drinks a day) is “bad for health” doubled in just the last 10 years. Two-thirds of Americans under 35 now tell Gallup that alcohol is harmful in any quantity.

This alarmist viewpoint is not exactly conspiratorial nonsense. It received an endorsement from the scientific community last month, when the federal government released a controversial alcohol study, which researchers claimed was suppressed by the alcohol lobby and the government. Buoyed by the Streisand effect, and thus widely covered in the news media, this study concluded that any drop of alcohol may increase a person’s mortality, with the risk accelerating after one drink per day.

This is all quite personal for me. On the one hand, I enjoy drinking. I love a glass of wine with dinner, and I like to make cocktails—martinis, mezcal Negronis, Corpse Reviver No. 4—for friends who visit the house. On the other hand, I also enjoy being alive. I love my children, and I don’t want to miss their weddings just because my fondness for martinis prematurely killed me. This makes me highly motivated to understand the relationship between alcohol and mortality in a way that goes deeper than, say, a cursory newspaper write-up of a single alcohol study.

This article is my best attempt to summarize my understanding of the vast and flawed literature on alcohol and health.

II. A Brief History of Alcohol: Climate Change and Drunk Monkeys

It is good and useful to think of alcohol as a tasty mild poison that humans have been honed by evolution to handle in small doses, because from a biochemical and anthropological standpoint, that is precisely what alcohol is.

Ethanol is toxic to most animals. (Feed your dog a bit of leftover steak, if you wish, but don’t pair it with a glass of cab unless you are eager to pay a visit to your local vet.) Unlike most creatures with mouths and tongues, humans and a handful of apes have a genetic mutation that makes our liver enzymes 40 times more efficient at metabolizing alcohol. I would like to thank god for this mutation—as Benjamin Franklin (absolutely never) said, “beer is proof that god exists and wants us to be happy”—but anthropologists, such as Robin Dunbar, have preferred to credit ancient climate change. Tens of millions of years ago, the theory goes, the planet entered a cool, dry period, and tropical forests collapsed. Nine in ten ape species went extinct. But some common ancestors of humans and chimpanzees responded heroically to this climate catastrophe by moving to less lush ecosystems and surviving as ground foragers. That sometimes meant swapping fresh fruit on tree limbs for rotting fruit on the ground. According to the “drunken monkey” hypothesis (this is real), the apes that survived this dietary shift were the ones whose livers could tolerate the fermented ground fruit.

Eventually, humans liked the smell and taste of ethanol enough to start making it on purpose, as the Economist recounts. The first solid evidence is from nearly 10,000 years ago. Residue from jars in Jiahu, China, dating to about 10,000 years ago indicates an early form of honey wine blended with rice and fruits. Booze helped human civilization grow:

If a group has more than, say, 100 members, direct personal relationships are hard to sustain. New social mechanisms are needed to build trust among those who may not know each other well. [Robin] Dunbar argues that rituals involving mild intoxication may have kept large social groups cohesive.

Thus, our ancestors grew up and got along on booze. By jacking up serotonin and dopamine levels, alcohol makes us ebullient, less inhibited, and more attracted to potential mates. According to studies by Dunbar, social drinking forms social bonds. One of his papers shows that people who frequent local pubs have “larger and closer support networks, which are associated with greater life satisfaction and even well-being.”

But just because the human body evolved to handle a bit of alcohol better than dogs—and just because alcohol tends to facilitate social interactions—doesn’t mean that alcohol is medically good for us. So let’s turn at last to the research.

III. The Truth About Moderate Drinking, Or: Why A Drink a Day Is About as Dangerous As Your Commute

I’d like to divide the modern history of booze research into two periods, each defined by a misconception. The first was the age of the J Curve. The second is the age of zero tolerance.

For years, some scientists and commentators insisted that moderate drinking was good for you. In 1991, Morley Safer, a correspondent for CBS, recorded a segment of 60 Minutes titled “The French Paradox.” He claimed that French people live longer despite a diet heavy in fat and butter because they lubricate their arteries with pinot. “The answer to the riddle, the explanation of the paradox, may lie in this inviting glass” of red wine, Safer said. After the 60 Minutes episode, wine sales skyrocketed.

The French Paradox had some quasi-empirical support. Some observational studies seemed to indicate that moderate drinkers lived longer and healthier lives than abstainers. These studies claimed to find a “J curve” relationship between alcohol and mortality risk. That is, if you plot mortality risk on a Y-axis and drinks per week on the X axis, the mortality line will go down (meaning moderate drinking extends lives) before it swerves up (meaning excessive drinking kills you), making a kind of languid “J.”

The problem is that those observational studies were mostly terrible. Moderate drinkers tend to be wealthier, better educated, and more likely to exercise than heavy drinkers, so their longer lives could have been caused by a zillion factors outside of carefully sipping red wine. Meanwhile, the “abstainer” category was typically a hodgepodge of religious teetotalers, former heavy drinkers, and “sick quitters” who were too ill to imbibe.

Most alcohol researchers now agree that the J might as well stand for “junk science.” Last year, I spoke to the behavioral psychologist and health researcher Tim Stockwell who told me that he has spent “thousands and thousands of hours” reevaluating studies on alcohol and health in the last 25 years. The J-curve is almost certainly bogus, he said.

But the old assumption that moderate drinking is therapeutic has given way to a new assumption, which is that practically any drinking is bad for us. The pendulum has swung toward a neuroticism that walks and talks like a scientist but constantly oversells the dangers of a daily drink. There are two main problems with the trendy case against moderate drinking:

Read more

Subscribe now

One hundred years ago, on September 26, 1929, President Herbert Hoover gathered a group of social scientists at the White House. He asked them to begin research on the most detailed report ever produced on the state of the nation. Four years later, running more than 1,500 pages long, Recent Social Trends was published, offering an unusually granular look at life in the mid-1920s.

The document is almost entirely forgotten. But today, for America’s 250th birthday, I’m blowing the cobwebs off this sucker and taking readers inside its yellowed pages for a look back at what life was like in the U.S. exactly 100 years ago, when the U.S. was celebrating its sesquicentennial anniversary.

In some ways, the Americas of 2026 and 1926 are eerily similar. In both cases, the country is celebrating a major birthday in the midst of a rising stock market and widespread fears of “technological unemployment” (mechanical power then vs. AI now); giddy wealth is coiled with economic anxiety; technology has transformed the way that people get information, mind-wiring us to a global cacophony of far-flung emotions (radio then vs. social media now); and after years of record-high immigrant entry to the U.S., the government has choked the migrant stream to a trickle.

In other ways, the America of 1926 was another world—practically another planet. Roughly half of the U.S. still counted as rural, and tens of millions of Americans had no indoor plumbing or electricity. Thick smoke from oil lamps filled their homes, and they emptied their bladders and bowels in old-fashioned chamber pots. Women had only voted in two presidential elections. Millions of children still worked for pay. Of the nation’s 27 million households, only 11 million had a phonograph, to listen to music, or a car. The first movie with sound would not come out for another year.

Life is much better in 2026. We live healthier, richer, and longer lives, with better medicine and more self-determination. But if I do my job well in the next few sections, you’ll see both the progress we’ve made in the last 100 years and the progress we haven’t made. Many anxieties that feel electrically charged in the present moment about work, family, and individuality are echoes of our ancestors’ fears. They felt all of our feelings. They just happened to poop in chamber pots.

1. The Average American in 1926

Before we dive into the world of 1926, let’s get into character.

Imagine that you are the typical American in 1926.1 You are a white 26-year-old. (In 2026, the median age is 40.) Since most immigrants have been male, we’ll say you’re a guy. Your name is John. Born in the first term of William McKinley’s presidency, you are raised on a farm without flush toilets or electric lighting. Too young to fight in World War I, you come of age alongside a generation that sees war in Europe as a “useless colossal blunder,” in the words of historian David M. Kennedy. Your life—indeed, your entire generation—is shaped by several notable developments: education, urbanization, automation, and women’s rights. You are the first person in your family to finish high school.

At 19, you move from the countryside to an urban apartment, as one small drop in the migratory flood from farm to city. Jobs in manufacturing and retail are easy to find. They’re also easy to lose. Temporary unemployment is the norm. You earn $100 a month and put some away for a rainy day, confident that the bustling city will provide another job in a few months. (Unemployment insurance does not exist; neither does Social Security.) In the evenings, you “radio”; yes, it’s a verb, too. Every weekend, you visit a cineplex, where the movies are black-and-white and silent. Sometimes, you down a few prohibited cocktails and go dancing with flappers. Several times a week, you drive around in a black Model T.

The year 1926 has been good to you. City life is a blur of high-velocity machines—cars, assembly lines, and radio broadcasts—and you sometimes miss the ancient rhythms of your farmland home. One year from now, Charles Lindbergh will shock the world by flying across the Atlantic. In two years, at 28, you’ll be married. In three years, you’ll have a baby. And in four years, in 1930, just months after the biggest stock market crash in American history, the world as you know it will be over.

II. The Economy of 1926

In 1926, the Roaring Twenties were at their full-throated peak. But America was anxious. If today’s commentators can’t stop talking about “maxxing” and “vibes,” the writers and politicians of the Twenties were vexed by the opposite of single-minded optimization. The watchword of the day was balance. Appearing 30 times in the Recent Social Trends report, balance was the cardinal worry of economic commentators: balance between rural and urban America; between manufacturing and farming; and between rapid progress and traditional values.

America in 1926 was smaller, and its values and identity were more twined with the countryside. The US population was about 120 million—about one-third of its current size—and nearly half of Americans lived in rural areas, compared to about 10 percent today. “In many respects, those country ways of life remained untouched by modernity,” the historian Kennedy wrote. From the Potomac River to the Gulf of Mexico, America looked “little different than it had at the end of Reconstruction”—cotton fields unfurled under a hot sun, hardly touched by electricity or modern plumbing.

The agrarian world was in a double crisis in 1926. The first problem was economic. In the 1910s, the American farmer fed the world. After World War I, global agricultural production came back online, and farmers were stuck with a ruinous surplus. The price of cotton and corn plummeted more than 50 percent from their wartime high. The second crisis was technological. The tractor and mechanical power transformed the farm, automating jobs for horses, as well as for people.2

If rural America was frozen in the ice block of history, urban America was a gushing torrent of change. By 1928, per-capita incomes of urban workers reached four times the average income of farmers. Between 1910 and 1930, the farm population declined while the urban population nearly doubled (see first graph below), thanks to both domestic migration and a wave of European immigrants (see second graph below). Much like today, the fear of historic levels of foreign-born people sparked an anti-immigrant movement. The 1920s saw the revival of the Ku Klux Klan, which dominated the politics of several states, including Indiana and Oregon. “The nativist sentiment that the Klan helped to nurture found statutory expression in 1924,” Kennedy writes, “when Congress choked the immigrant stream to a trickle, closing the era of virtually unlimited entry to the United States.”

City dwellers were proud of their superiority; even haughty. In the previous years Scopes trial, the high-school teacher John T. Scopes had been indicted for violating Tennessee law by telling his students about evolution. Here’s Kennedy:

Urban Americans smiled with satisfaction when street-smart Chicago attorney Clarence Darrow humiliated rural America’s historical paladin, William Jennings Bryan, in the course of the trial… Bryan’s mortification symbolized for many the eclipse of rural fundamentalism and the triumphant ascendancy of the metropolis as the fount and arbiter of modern American values.

As agriculture’s share of employment plummeted from more than 50 percent in 1870 to less than 25 percent in 1926, it wasn’t just manufacturing that absorbed the expanding workforce. The share of jobs in trade and transportation (e.g., moving, storing, and selling goods in wholesale and retail trade) rose from 9 to 20 percent. Clerical services (e.g., retail salespeople and cashiers) exploded from 1 to 8 percent.

Manufacturing advanced faster than working conditions. While the 12-hour workday disappeared in the 1920s—the US Steel Corporation abandoned it in 1923—neither the two-day weekend, nor paid vacations, nor retirement existed for the average American worker.

Modern urban consumerism marched forward thanks to an explosion of chain stores. Between 1927 and 1928, Sears Roebuck sales grew by 20 percent and Safeway sales grew by nearly 50 percent.

Even more than 2026, 1926 saw a bonanza of big-company mergers, which means it was similarly an era of monopoly concerns. Between 1919 and 1928, manufacturing and mining were involved in 1,200 mergers, a record for one decade. Bank mergers were even more frequent. And practically no industry was more concentrated than electricity. By the end of the 1920s, ten public utilities controlled 75 percent of the electric light and power business while 16 holding companies controlled about 50 percent of the country’s gas output.

Nothing in 1926 roared like the automotive industry. In 1900 there were 8,000 cars on the road—or “horseless carriages,” as they were then known. By 1926, the number had grown to 19 million. “It is probable that no invention of such far-reaching importance was ever diffused with such rapidity,” the report’s authors wrote.

The automotive revolution was the great achievement of Henry Ford. It was also the apotheosis of Fordism, the theory of business that braided higher productivity, higher wages and higher spending levels. In 1910, it took roughly 15 hours to put together a single Model T Ford. By 1926, a new car rolled off Ford’s assembly line every 10 seconds. Productivity was the handmaiden of affordability: A vehicle that cost the average worker two years’ wages before World War I cost just 3 months’ earnings in the mid-1920s.

No other country rivaled America’s automotive love affair. According to the historian Bill Bryson, 1920s Kansas alone had more vehicles than France. Car ownership created an entirely new way of thinking about the self in relation to the environment—an “automobile psychology.” Social scientists spoke of the modern “gypsy family,” which seemed to spend more time inside the car than outside it. Travel time surged, and the tourist business swelled, with new “camping grounds, cabins, cottage camps, and roadside camps” dotting the landscape to serve a nation that couldn’t get enough of the road. Just as Airbnb and its brethren pushed the hotel business in the 2010s and 2020s, so did these new makeshift tourist attractions force the hotel industry to upgrade in the 1920s. With “the increase in travel by women which has been induced by the automobile,” hotels added more private bath facilities, more varied restaurant menus, and plusher room furnishings. When America made cars, cars remade America.

Subscribe now

III. Culture and Leisure in 1926

The year 1926 was almost certainly the high-water mark of women’s rights and liberties in American history, to date. The nearly 10 million women who worked for wages represented a doubling of the female workforce since the end of the 19th century. While the share of women working as domestic servants declined, more became teachers and hairdressers. No female occupation grew faster than the salesperson or store clerk.

After the ratification of the 19th Amendment in 1920, women had only voted in two presidential contests by 1926. Still, there was anxiety about their acquisition of political and cultural power, which concentrated on the rise of the legendary “flapper.” These flapper gals “smoked, drank, rouged their shining faces, bobbed their hair (which is to say cut it short and even all the way around), and clad themselves in silken dresses of breathtaking skimpiness,” the historian Bill Bryson writes in One Summer: America, 1927. The shrinkage of chastity was quantifiable: The amount of fabric in the average dress fell from 20 yards in 1910 to 7 yards in the 1920s.

Anxiety about women flaunting their liberties and their thighs was an adjunct to concerns about changes to the family. With the rise of working women coinciding with an increase in divorce, the authors of Recent Social Trends mourned that homes became mere “parking places” for parents and children, whose passions resided in labor and leisure.3

Childhood was transforming just as much as womanhood in the 1920s. For the first time, half of high-school-age students attended school, an eightfold increase since 1900. Declining birthrates along with rising wealth led to fewer children and more resources devoted to each child.

The report’s authors weren’t entirely sure that the new style of parenting—eerily similar to intensive parenting phenomenon of the 2020s—was a good thing. In a passage paragraph that one could imagine reading in the New York Times tomorrow, they worried about the the declining birthrate :

For the first time there were fewer children under five years of age in one census year than in the one preceding.

For the first time also there were fewer children under five years of age than from 5 to 10 years of age. In some cities already there are not enough children to occupy the desks in the earlier grades. This decreasing enrollment has not yet reached the high schools, but it is only a question of time, unless a larger proportion of those out of school are continued in school. Though the supply of children is being restricted, the demand for them continues.

Hovering over the decade’s cultural perturbations was Prohibition, which had been the law of the land since 1920. Prohibition was bad, not only because it shut down a major industry (America’s fifth largest, at the time); not only because it moved billions of dollars from the legitimate economy into the black market, empowering the Mafia and other criminal enterprises; not only because the national murder rate went up by almost a third after Prohibition was introduced; and not only because the federal government lost about half a billion dollars in liquor taxes (10 percent of its national income); but also because it killed people directly. As an ingredient in paint thinner, antifreeze, and other goods, alcohol’s production couldn’t be fully outlawed by the government. To discourage its consumption, the government required that industrial booze be frequently laced with poison. Americans, loving booze, drank that, too.

The results were disastrous. A chemist’s survey in the New York Telegram found that one speakeasy’s spirits contained, among other things, kerosene, formaldehyde, and sulfuric acid. In their book Eating in America, Waverley Root and Richard de Rochemont report that nearly 12,000 people died in 1927 from drinking poisoned alcohol. Given that the US population was roughly 66 percent smaller than it is today, that would be the mortality equivalent of 36,000 people dying in 2026 from drinking beer that had been poisoned on purpose by the feds. (That figure is, by the way, roughly the number of drivers who die annually in automobile accidents today.)

Beyond the speakeasies and jazz clubs, sports were becoming a more central part of American life, although the industry was unrecognizable to modern audiences, in large part because 1920s sports were so unprofessional. For example, while it was considered America’s pastime, baseball was a terrible business. Teams relied on ticket sales for revenue, but Sunday games were prohibited in most cities, and electric lights were a luxury. So most games had to be played during the weekdays, when few working people could attend. Ninety-minute games were common; without a big audience or big paycheck waiting for them, players seemed to be in a rush to get things over with. In one doubleheader against the New York Yankees on September 26, 1926, the St. Louis Browns won 6–1 in 72 minutes and then won 6–2 in 55 minutes. (Yes, those were two nine-inning games.) “Quite how they managed it is a wonder,” the historian Bryson writes. “The two teams banged out 25 hits between them in the first game and 20 in the second, so these were hardly classic pitchers’ duels. There was just a lot less messing around.”4

IV. Media in 1926

Unlike 2026, when we are witnessing the decline of reading, 1926 might have been a high-water mark for reading in American life. Never before had so many different print media—books, magazines, and newspapers—swelled at the same time, before sound film and television took over.

Authors were celebrities—people mobbed Sinclair Lewis’s Minnesota home while he was trying to write his novels—and books published annually had doubled since the 1910s. For those overwhelmed by so many titles, the Book-of-the-Month Club was founded in 1926. Magazine advertising revenues grew by 500 percent in the Twenties, thanks to the birth of Reader’s Digest in 1922, Time in 1923, and The New Yorker in 1925. Newspaper sales also rose throughout the decade, and New York City alone had twelve daily papers.

This abundance of literacy was threatened by the arrival of new communications technology. Even more than telephones—which multiplied from less than 2 million units in 1900 to 20 million in 1930—radio defined the decade, evolving “from a mysterious curiosity to a widely diffused and universally accepted instrument of entertainment, business, learning and mass communication, [with] few if any counterparts in social history,” according to the authors of Recent Social Trends.

It is remarkable to read the report’s analysis of radio’s effect on American life, because it reads so alarmingly modern. Above all, the authors were concerned that radio encroached on individuality by mind-wiring each American to a global monolith of news and entertainment. Mass media organizations created “greater possibilities for social manipulation,” they wrote, and radio threatened to turn the individual into what the philosopher Martin Heidegger in this period called a “they-self”—a being who, failing to achieve authenticity, fully adopts the tastes, habits, and beliefs of the universal crowd.5

Today’s commentators sometimes fret about the role that the internet and the smartphone have played in unleashing populism, by elevating the views of the disenfranchised and smashing the authority of the elites. But 100 years ago, the authors of Recent Social Trends also saw the radio as a pure “leveling” technology:

Negroes barred from entering universities can receive instruction from the same institutions by radio; residents outside of the large cities who never have seen the inside of an opera house can become familiar with the works of the masters; communities where no hall exists large enough for a symphony concert can listen to the largest orchestras of the country; and the fortunes of a Negro comedy pair can provide social talk throughout the nation. Isolation of backward regions is lessened by the new agency of communication

Unlike print media, radio works the ear rather than the eye—the ear, which cannot close or even focus with ocular precision, but remains always open to the world. One hundred years ago, we were already making our way in a world where waves of information radiated around us, a proto-internet of sound, waiting for an audience and an antenna to tune in. Wired to the emotional values of far-flung events, Americans breathed air that pulsed with spooky electromagnetism, carrying the stories of people we would never meet. Without knowing what it meant to “log on,” our fellow citizens in 1926 were already learning what it means to never be able to log off.

V. The Future, According to 1926

The authors of Recent Social Trends were astonishingly prescient about the direction of technology. In one paragraph, they somehow anticipated the rise of audiobooks, YouTube, Netflix, smartphone cameras, musical software, ubiquitous air conditioning, and the electric battery revolution:

It may be that the world will find much use for talking books; school and college students may listen to lectures by long-running phonographs or talking pictures; moving pictures may be transmitted by wireless into houses; seeing with that new electric eye, the photo-electric cell, and recording what is seen, appear to have almost unlimited applications; new musical instruments different from any now in use may be given to us by electricity; the production of artificial climate may become widespread; an efficient storage battery of light weight and low cost might produce changes rivaling those of the internal combustion engine. And these are only a few of the myriad possibilities from new inventions in the future!

In an equally oracular section, the authors predicted the emergence of remote work and declining geographic mobility, anticipating that “the transmission of goods, of the voice and possibly of vision may act as a retarding influence on human mobility in the future and may cause a development of more remote and impersonal direction and controls.”

But the social scientists did not see these trends as altogether good. They worried that modern life, defined in equal parts by urbanization and technology, obliterated people’s values and their sense of self. Even as they gawked at the increase in patents—which grew more than 20-fold between the 1850s and the 1920s—they worried that a growing number of discoveries would bring “problems of morals, of education, of law, of leisure time, of unemployment, of speed, of uniformity and of differentiation.”

Social scientists of the 1920s saw machines pushing workers off of farms and competing with workers in manufacturing plants. How long, they wondered, until they would replace human workers in all tasks? “A larger proportion of work by machines, and a smaller proportion of human labor, is to be expected in the future,” they wrote. “There are indeed a few cases of wholly automatic factories and automatic stores and many automatic salesmen.” It is extraordinary to read these fears and not reflect on the AI jobs panic of the present, while also marveling at the thousands of occupations that are possible today precisely because machines made old jobs obsolete.

The dawn of the age of the machine drove us mad. Physicians of the day warned that the frail human mind was no match for the car, predicting at the time that “diseases of the wheel” would afflict the youth who rode bicycles and cars without restraint. It was not entirely obvious that they were wrong. In Germany, the number of patients registered in mental hospitals grew from 40,375 in 1870 to 220,881 in 1910. Over the same period, the share of patients admitted to general hospitals for illnesses of the nervous system rose from 44 to 60 percent.

Most perceptively, social critics of the age recognized that the urban-technological revolution of the early 20th century—what we might even call “modernity”—transformed not only our minds but also our values. Machines and systems that pulled Americans off the farm, away from the family home, and into churning markets of people and products threatened to replace the Judeo-Christian values that had bound the country for centuries with a new system of values dictated by markets. In 1903, the sociologist Georg Simmel anticipated the anxieties of the Twenties—ours and theirs—when he observed that in cities “money takes the place of all the manifoldness of things” and becomes “a common denominator of all values.” Money “hollows out the core of things, their peculiarities, their specific values, and their uniqueness and incomparability in a way which is beyond repair.”

One hundred and twenty years after the publication of that essay, the Wall Street Journal asked thousands of Americans what values were still important to them. While a declining share of Americans endorsed the worthiness of patriotism, religion, community, and children, the share who said “money” was “very important to them” went up. It sometimes seems as if markets and money are the last value standing, the final common denominator beneath all human endeavor.

On its 250th birthday, the U.S. similarly defines itself through markets. Those famous words of Calvin Coolidge, America’s president in 1926, could just as well serve this American president and this American moment: “The chief business of the American people is business.”

The debate about the future of jobs and AI often polarizes into two extreme categories, neither of which I find particularly useful.

On one side, there are the Doomers. They predict that AI will destroy millions of jobs. Most notably, Anthropic CEO Dario Amodei has sometimes predicted an imminent entry-level white-collar wipeout. The problem with this theory is that it doesn’t really stand up to the current data. The employment rate for prime-age workers in the U.S. is near its all-time high, and a survey of corporate executives found “little evidence of near-term aggregate employment declines due to AI.” Practically every claim of AI’s strong effects on the labor force are highly contested among economists.

On the other side, there are the Deniers. They insist that AI is a scam, a stochastic parrot, and an overhyped category of vaporware. But while the Doomers too often live in a future that doesn’t exist, the Deniers too frequently dwell in a past that’s no longer with us. The generative AI economy, which barely existed four years ago, has now generated between $100 billion and $200 billion in total revenue in the last 12 months, and AI’s facility with coding has transformed the jobs of millions of software engineers.

Between doom and denial, there is the evidence. Through the mists of statistical confusion, a clearer picture of AI and the future of work is finally coming into view. AI doesn’t seem to be obliterating jobs so much as leading to the explosion of a very specific kind of work, which has major implications for the way we think about AI, the future of jobs, and how to get rich in America.

I. Matt’s Story

Matt Rosenberg left his job at Amazon in 2025. He wanted to open a restaurant in the East Bay with his Thai wife, whose family had served food in Bangkok for decades.

Hoping to buy a restaurant, Matt shared his plans with ChatGPT. He shared strategy documents, foot-traffic research, and videos of restaurant kitchens. One day, he said, ChatGPT suggested he look into “micro-enterprise home kitchens,” a little-known rule that allows families to sell food that they whip up at home. Soon, the couple opened a restaurant from their own kitchen called Bangkok Rush Thai Kitchen.

Restaurants can exist without AI; millions have. But Bangkok Rush Thai Kitchen might not. Matt and his wife have used AI to build their website, keep up with compliance, manage their inventory, and analyze customer trends. “Starting a business takes a lot out of you,” he told me. “Without ChatGPT, I think my wife and I might have given up early on in this process.”

Today, the Rosenbergs’ restaurant—whose most popular dishes include khanom krok, a small round coconut pancake—sells out of desserts so frequently that they’re looking to open a brick-and-mortar location. Flush with confidence, Matt has even started a second business: a one-man consultancy that teaches other people how to start a company using AI.1

II. The Rise of the Solo Act—and the Indie Millionaire

This is a golden age for solo acts. There has never been a better or easier time to work independently. And there has never been an easier time to become a millionaire by working for yourself.

I’m going to walk through this phenomenon step-by-step, and then I’ll tell you why it’s happening and whether I think it’s going to last:

Read more

In February, the U.S. and Israel launched a surprise attack on Iran, striking hundreds of military targets and killing thousands of members of the Islamic Revolutionary Guard, including Supreme Leader Khamenei. The White House and its defenders called their campaign a clear and unalloyed success, the latest display of a de…

Read more

One year ago, I left The Atlantic, where I’d worked happily for nearly 17 years, to launch this Substack. This week, my newsletter turns one. So, I wanted to use its first birthday to reflect on the state of the business, what I’ve learned about writing a newsletter on this platform, and where I’m trying to grow in the ne…

Read more

I. The Ring

In December 2024, I received an Oura ring for the holidays. I thought I would hate it.

I’d had a messy relationship with other “wearable” technology designed to monitor my physical activity, or lack thereof. Years earlier, during a period of elevated stress and miserable sleep quality, I’d owned a different fitness device that wouldn’t stop sending me panicked notifications about my insomnia. Sleep was hard enough without the harsh judgment of one’s jewelry, I thought. I threw the thing in the trash.

But the ring I liked. One of the first things I learned was how alcohol affected my sleep. When I had a glass of wine after 8 p.m., my resting heart rate shot up, and my heart rate variability plummeted. This would lower my “Readiness” score, which is prominently featured on the smartphone app. But when I had the same glass while cooking dinner around 5:30 p.m., the wine had no effect on my Readiness number. Like so many elderly Millennials, I am the sort of person who cannot see a prominently labeled scoreboard without devoting myself to climbing it. So, without ever figuring out what heart-rate variability actually meant, I ditched the late-night drinks. (Mostly.)

The demise of the nightcap wasn’t the only way the ring changed my life. More profoundly, it altered my relationship to activity. Before a walk, after a workout, or even after a nap, I would check in with the app to see my Readiness, Activity, and Sleep scores. Each day was now compared with—even set in competition against—the previous day or the previous week. The ring pulled me into a feedback loop with a biometric technology that turned my body into a game board through which I could win or lose points.

The philosopher C. Thi Nguyen, who writes about the philosophy of games and metrics, has said that the modern world’s blizzard of numbers can obscure our wise yet murky questions about life and replace them with easy and legible goals. For example, someone might become a journalist to uncover important truths but then devote each week’s labor to maximizing page views; or she might get into philosophy to think about the deepest life questions but devote her life to maximizing her h-index. “Metrics are useful because they compress information,” Nguyen told me. “They are dangerous, because they compress information.”

At its best, the ring makes me fitter, happier, and, sure, even more productive. I walk more, lift more, sleep more, and drink less. You will be hard-pressed to find a physician who thinks there’s anything amiss in the previous sentence.

But the same doctor might not see how the obsession with winning the measurable games of health can encroach on the less measurable games of life. The best way to sleep more is to see fewer friends in the evening. The best way to lift more during the week is to eliminate social lunches to protect my midday gym time. To become a measurably enhanced self often means eliminating my less quantifiable sources of meaning and happiness.

The ring improved my life. But its form of self-improvement often pulls me away from other people. This left me with a nagging question. At what point is it unhealthy for me—for anyone, for all of us—to be this obsessed with health?

Subscribe now

II. The Enhanced Self

Clearly, I am not the only American who is reconsidering his relationship with alcohol. The share of people who drink hit an all-time low last year, according to Gallup, whose data go back to 1939. Total beer consumption recently reached a 21st-century low, and wine vineyards are reportedly “in crisis.” While many social changes happen slowly, the attitude shift against alcohol has been quite sudden. The share of Americans who say moderate drinking (defined as one or two drinks a day) is “bad for health” doubled in just the last 10 years. Two-thirds of Americans under 35 now tell Gallup that alcohol is harmful in any quantity.

The decline of drinking is one part of a larger cultural phenomenon, a hydra-headed megatrend whose tentacles reach out to touch everything from science and medicine to technology and entertainment, reshaping the way that Americans think about themselves, their time, their friendships, and their future. It is the rise of the Enhanced Self.

The Enhanced Self is the evolution of medicine, technology, and consumer culture from an emphasis on curing illness to an obsession with optimizing normal, healthy life. We see this with the rise of GLP-1s, the explosion in biohacking with peptides (injectables that affect inflammation and gut health and are also the “P” in GLP), and the continued growth of supplements. More Americans are using therapies not only to cure what is wrong with them but also to improve what is not wrong with them. At the layer of leisure, the tendrils of the Enhanced Self touch the white-hot rise of fitness in American life. A record 77 million Americans belonged to a gym or studio in 2024, up 20 percent since before the pandemic. Running clubs on the fitness app Strava nearly quadrupled in 2025 alone. If you don’t believe the industry data, perhaps you’ll believe the federal government: according to the American Time Use Survey, Americans today exercise and play sports more than at any period on record.

At the layer of biology, the Enhanced Self incorporates the belief that the human body is akin to a single-issue hardware device, whose owner should obsessively seek to extend its operating life beyond its scheduled date of obsolescence through relentless work and eagle-eyed neuroticism. At the layer of sociology, the Enhanced Self is inseparable from the decline of socialization, which I have previously called the anti-social century. While running clubs and morning workouts are booming—and I am positive that these are highly social events for at least some of their participants—nightclubs are closing and parties are withering. Young Americans spend about 35 percent less time socializing and 70 percent less time attending or hosting parties than they did at the beginning of the century.

America’s self-improvement tradition is practically as old as the country. Benjamin Franklin tracked thirteen virtues on a daily scorecard. Sylvester Graham preached dietary reform in the 1830s. The instinct to improve the body, and to moralize that improvement, is anciently American.

These earlier iterations of self-improvement drew their power from religion, community, or characterological projects to promote civic virtue. Temperance, for example, was not just about individual health; it was a social movement to improve the culture, to rescue women and children from alcoholic husbands, and to build a better republic. (That it failed in myriad ways is not to deny that some of its goals were virtuous.) The Muscular Christianity movement of the 19th century paired New Testament virtues with an ethic of manly strength in a way that wouldn’t be so out of step with modern MAGA and MAHA machismo.

But the age of the Enhanced Self is different, not only because many of its elements are distinctly of the 2020s—including peptide shots, social media, and biometric scanners—but also because it does not particularly seek to build anything outside of the self. For all its sins, the temperance movement was focused on national change. But the typical adherent to the Enhanced Self—say, a 50-year-old with a peptide stack and a Whoop—is not trying to improve the country. He’s just trying to improve his score.

The history of alcohol abstention offers another way to see how the Enhanced Self is a truly modern phenomenon. For a long time, abstinence was associated with religion or personal histories, such as addiction recovery or pregnancy. But in the new health culture, abstinence is not about faith or addiction; it is about bodily perfection. On health podcasts and videos, influencers and science communicators talk about alcohol’s association with sleep scores, skin clarity, energy levels, cardiometabolic biometrics, and executive function. The explosive growth of non-alcoholic breweries like Athletic Brewing offers commercial proof in a secular age that the decline of drinking has little to do with faith. Modern sobriety, for better or worse, is not about addiction, but enhancement; not about our body politic, but about our bodies—period.

III. The Three Pillars of the Enhanced Self

The Enhanced Self is defined by three distinct tendencies. None of these ideas are entirely new in the 2020s, but the combination of science, technology, and social media has afforded them special power.

1. OPTIMIZATION AS THE NEW VIRTUE: Many drugs and devices initially invented to cure the sick become instruments to greatly enhance the healthy.

In his book Outlive, which serves as a kind of Bible of self-enhancement, the scientist and podcaster Peter Attia divides the history of medicine into three phases. Medicine 1.0 was the prescientific era of doctors trying and mostly failing to treat disease. Medicine 2.0 is the modern approach, in which doctors cure illness with the full apparatus of science, including randomized drug trials. Medicine 3.0, Attia argues, goes further: it stops chronic illness before it takes hold, through personalized, data-driven preventative care.

This is the medical logic of the Enhanced Self, in which drugs and devices built to cure the sick become instruments to enhance the healthy. GLP-1s designed for diabetics now quiet food noise in millions of people who simply want to eat less. Continuous glucose monitors, originally designed for diabetics, have found a growing market among healthy people who want to know how an afternoon’s potato chip changes their blood sugar. Stacked on top, you have biohacking with off-label peptides1, experimentation with supplements, and hormone therapies, including replacement therapy. Prescriptions for TRT rose from 7.3 million in 2019 to more than 11 million in 2024, a 50 percent increase in five years, with an even steeper rise among men aged 35 to 44.2

Put it together—peptides, supplements, and hormone therapy, connected through the web of biometric data made available by fitness trackers—and today’s self-enhancers are armed with a portfolio of pharmacological and technological tools, allowing them to treat their bodies like improvable machines.

2. FITNESS AS THE NEW CLASS MARKER: We used to compare ourselves based on money. But fitness trackers and biomarkers create a parallel class system that makes us feel superior to, or inferior to, everybody else.

A generation ago, high-status Americans bragged about how little they slept, as a proxy for their work ethic. Today, the even-richer elite brag about how much they sleep, as a testament to their ability to outrun mortality. They toy with sleep stacks of magnesium and melatonin. They try mouth tape, white-noise-cancellation devices, and multithousand-dollar cooling mattresses. I have witnessed adults comparing deep-sleep percentages the way their younger selves once compared GPAs and SAT scores.

If you squint, it looks like health is becoming the new status marker and, perhaps, the ultimate status signal. For centuries, people compared themselves to others based on income, in part because money is standardized, measurable, and visible. If you earn $1,000,000 a year, and I earn $50,000 a year, then we both know where we stand relative to the other in purchasing power. Fitness comparison, by contrast, was more difficult, except in obvious cases of diagnosis (i.e., a neighbor who has cancer) and physical appearance (i.e., an extremely thin or heavy person in the office). But as fitness biomarkers and peptide stacks become part of upper-class conversation, people can more easily rank themselves against others, turning health into a comparative status marker.

Every cultural movement mints a new class of celebrity, and the Enhanced Self is no exception. Attia turned longevity medicine into a mainstream preoccupation. Andrew Huberman built a podcast empire on the premise that every behavior, from morning sunlight exposure to cold-water dips, could be neurochemically optimized. Bryan Johnson became the human embodiment of the movement’s extremes, spending millions of dollars a year submitting his body to hundreds of interventions in a war against aging.

What these figures sell, beyond supplements and programs, is the feeling of superiority along with an insider’s vocabulary: VO₂ max, grip strength, Zone 2 cardio, DEXA scans. These terms have migrated from exercise physiology textbooks into dinner-party conversation, giving ordinary people—or disproportionately affluent people—benchmarks for comparing themselves to others, as they think about their bodies as systems to be monitored and compared.

Once we see bodies as measurable and improvable systems over which we exert a certain amount of control, it is only a small step toward thinking of personal health as an enterprise that requires an almost managerial approach—that is, as yet another job.

3. HEALTH AS THE ULTIMATE JOB: More people are treating health maintenance as a kind of professional occupation, one that makes them behave like the chief executive of their own body.

The historian Alfred Chandler argued in The Visible Hand that the telegraph and the railroad moved information so fast that companies needed a new kind of person to handle it. They were called “managers,” and thus was managerial capitalism born.

Something similar is happening in modern health. The eruption of biometric data from wearable devices, along with the genetic and diagnostic tests that forecast our futures3, produces a flood of information that demands a managerial response. I find it almost impossible to look at my Oura ring without comparing this week’s heart rate and exercise data to last week’s. When I read one day’s numbers (huh, that exercise score looks awful), I feel obligated to adjust my behavior (I should schedule a walk). As more people manage this deluge of information flow, they will face the same problem as 19th century chain stores, and they will begin to think of their anatomical output as a product to be managed. Slowly and then quickly, many of them will come to see themselves as akin to chief executives of their own bodies.4

As strange as this might sound, I think the Self Executive will become an identifiable figure in the next few years. By this I mean: More people will talk about their health, their fitness, and their life extension procedures as if they were a job, a career, or a calling. In the process, they will somewhat invert the historical role of leisure in modern life. In the traditional arrangement, people go to the office, where they focus on being productive, and then go home and relax. But today, fitness and health fixations are colonizing leisure at the same time that our wearable devices gather, sort, and grade the data that our bodies produce. Over time, this trains people to approach their downtime with the very same productivity mindset that they were supposed to leave at the office. In short, the Enhanced Self turns personal health into a second job, one in which the individual serves as both the executive and the firm.

IV. Enhanced Selfishness

In one respect, the age of the Enhanced Self is a triumph. The fruits of this movement will include fitter people, with less disease, who live longer lives. This is not an achievement of mere accounting. It will mean millions of moms and dads who live to dance at their kids’ weddings. It will mean more friendships and partnerships that last into the 90s and beyond. It will mean children who know their grandparents by their touch and their smell and not just by their photographs.

As sure as I am that all health is good, I am skeptical that the Enhanced Self is a purely positive trend. The pursuit of health can go too far, and it can come at the expense of others. The three pillars of the Enhanced Self include optimization, class comparison, and running one’s body like a corporation. These are all individualist endeavors. Many of them are inherently solitary.

Many people have told me that the fitness boom is a direct response to declining social connection. For some, gyms and workout classes serve as the 21st century’s answer to shuttered union halls and molding bowling alleys. They are the new spaces where people, young and old, can go to seek out a social life, to talk to people, to attend events, and to become a member of a community.

But it is no coincidence that the age of the Enhanced Self is also an age of aloneness. Young people, who are seeing the highest increases in exercise time, also say they have fewer friends than any cohort ever; that they spend more time alone than any generation on record; and that they are more anxious and depressed than previous groups. Generation Z has the lowest levels of interpersonal trust of any group ever polled, the author and researcher Ryan Burge reported. “The velocity of their decline in trust already far exceeds any previous generation,” he wrote.

Our bodies want us to be social, even when we act like secular monks. Research by Sandra Weintraub of Northwestern University has found that “super-agers” (individuals over 80 with the cognitive function of people decades younger) shared little in common except for an unusually robust history of friendship and other social connections. A 2025 analysis of 500,000 participants in the UK Biobank reported that living with a partner and frequently visiting family had roughly the same relationship with longevity as exercise.

Our health is our own. It is one of those things that we cannot share with other people. And so, while the pursuit of health does not have to cleave us away from others, the project of delaying mortality is often a solitary undertaking.

V. Coda (Death)

Sometimes, I wonder what the Enhanced Self is all about. Surely, its emergence in American life has something to do with a traumatic pandemic winding down, the rise of the MAHA movement, and the migration of Nietzschean tech executives toward the center stage of internet culture, just as GLP-1s and new fitness trackers made certain kinds of self-improvement reachable for millions. Altogether, I think these forces have, like disconnected weather systems moving into a shared area, created a cyclone of a social phenomenon.

But deeper forces may be at work. In a sensationoal essay, the psychologist Adam Mastroianni notes that over the past few decades, high schoolers have steadily drunk less, smoked less, and fought less. In the same period, serial killers have all but vanished, blockbusters have grown less original, design has grown less distinctive, and cars have gone monochrome. Mastroianni ties these together with a theory he calls “the decline of deviance.” As people get richer and the world gets safer, deviance falls, because “life is worth more now.” When people think that they might live to be 100, the strategy for every life-game is the same: play it safe.

The novelist Karl Ove Knausgård once wrote that an irony of the modern world is that we are obsessed with cultural representations of death and yet terrified of the real thing. In movies and TV shows and podcasts and news, people are constantly dying; dead bodies are the hydrocarbons that power every modern plot. But no amount of represented death can prepare us for the thing itself. When bodies die in real life, we rush to hide them. In the hospital, they are filed away in subterranean basements, at the far end of long hallways that branch out like root systems, where no visitor can accidentally stumble upon them. Enraptured by death’s fiction, we are also terrified of its reality.

At bottom, enhancement culture is a disposition toward death—thinking about death, calculating distance from death, worshipping death like some Aztec priest offering sacrifices to appease the angry sun god. Bryan Johnson’s wellness company, book, and company-and-book-inspired Netflix documentary are not called “Live Better” or even “Live Forever.” It’s called “Don’t Die.” The moment-by-moment obsession with death may extend our lives. But when we cannot stop practicing this lifespan arithmetic—how much time will this drink cost me? how much time will that supplement buy me?—many of us will slip out of the thick appreciation of the here and now and approach life with all the verve of a lonely risk-assessment officer at a life insurance firm.

The fact that the Enhanced Self is often anti-social and the fact that it’s profoundly about death are not two different claims but rather two reflections of the same idea. Our fear of death motivates an all-consuming neuroticism about outrunning mortality, even when the price we pay is putting health optimization above everything else, including other people.

Last week, wrecked from travel, I saw a friend whom I hadn’t seen in months for drinks in Washington, D.C., on a Thursday. We had many things to celebrate—birthdays, babies—and oh how we celebrated. We stayed up too late. I drank too much. My head hit the pillow several hours after bedtime, and when I woke up, the mezcal left a dull ache behind my left eye. Turning toward my nightstand, I saw my Oura ring, which I’d taken off before slipping into bed. Data from the previous night’s sub-optimal sleep, like light particles within a black hole, would never reach the outside world. Thank god for that. I put the ring back on my finger, and I walked downstairs to make coffee. I don’t remember when the hangover ended. Was it one minute, one hour, a day? It doesn’t matter. What matters is that I saw my friend. He saw me, too. In more ways than one, we lived.

Subscribe now

I. The Aurelius Trap

Consider the stories of three leaders, each considered peerless in his domain.

1. Rome

Only “once in the history of the world,” Alice Zimmern wrote, has a man fulfilled Plato’s vision of a philosopher-king, a world-class ruler who was also a peerless mind. He was Marcus Aurelius.

As emperor of Rome in the second century CE, Aurelius moonlit as a dispenser of aphoristic gems about life, death, and honor. His private journal, now known as his Meditations, is a compendium of notes-to-self, maxims, and Stoic writings, which Aurelius composed while planning and carrying out a military campaign in Europe. Never intended for broad publication, the book has nonetheless become one of the most-published works of philosophy in history. It is hard to think of a person more powerful in his life whose words have been read so frequently in his afterlife.

One might expect such a person to be a paragon of good judgment. But rather than select a proven statesman to lead the fragile empire—as many good emperors of the previous century had done—Aurelius elevated his troubled biological son, Commodus. In the view of Edward Gibbon, the historian who wrote The Decline and Fall of the Roman Empire, Commodus’ reign was the beginning of the end for Rome. Lazy, vain, and self-indulgent, Commodus allowed corruption to fester and order to crumble. “The monstrous vices of the son,” Gibbon wrote, “have cast a shade on the purity of the father’s motives”, as Aurelius “sacrificed the happiness of millions to a fond partiality for a worthless boy.”

I sometimes think about the thousands of executives and entrepreneurs who read Meditations over and over and fail to absorb the lesson buried in its biographical introduction. In the privacy of his tents and chambers, Marcus Aurelius articulated the virtues of judgment as well as anyone. But when it came to the most significant judgment of his own life, Aurelius handed off his kingdom to a nincompoop whose reign accelerated the demise of the Roman Empire.

When, you might think, has a leader so wise ever proven so incompetent at that most essential job: Naming a successor who doesn’t threaten to destroy his legacy?

2. Hollywood

When Bob Iger stepped down as CEO of The Walt Disney Company for the first time, in early 2020, he was coming off one of the greatest runs in modern corporate history. Unlike Aurelius, Iger did not inherit an empire at its peak. In 2005, Disney’s animation studio was a mess, its feature films struggled at the box office, and theme-park attendance had flattened after the post-9/11 slump.

Between 2005 and 2020, Iger went on a triumphant shopping spree, bagging Pixar, Marvel, and Star Wars. Just when Disney’s trunk seemed so stuffed with intellectual property that surely nothing more could fit, he managed to squeeze in the voluminous assets of 21st Century Fox. Throughout the 2010s, Disney pumped out billion-dollar hit after hit, plumbing its unrivaled library of titles for sequels, prequels, adaptations, and reboots that dominated the box office and rewrote the rules of Hollywood.

With the stock at all-time highs, Iger finally named a successor: Bob Chapek, the head of Disney’s theme park division. The results were dismal.

Chapek’s failures, which coincided with the pandemic, extended from box office disappointments to a political mess over Florida’s “Don’t Say Gay” bill. But the failure of Bob II reflected as poorly on Bob I. As the New York Times reported, Iger repeatedly “undermined” his successor and failed to build a credible runway, leaving Chapek weakened, the board unsettled, and Disney without a clear “next chapter.”1

In 2022, Iger returned to the company weakened in part by his own botched succession planning, like a doctor heading back into the O.R. to staunch the bleeding from a wound he helped to inflict. Thus did the smoothest executive in showbiz oversee one of the industry’s most turbulent and least successful corporate handoffs.

3. New England

Bill Belichick is the greatest coach in NFL history. In a league whose rules are designed to promote parity and break up dynasties, Belichick coached Tom Brady and the New England Patriots to the most-ever wins in a 10-year period; the most-ever consecutive conference championship appearances; the most consecutive division titles; and the most Super Bowl wins (six, which is tied with the Pittsburgh Steelers, but accomplished in one continuous tenure rather than over a half century).

But the same man whose fussy mastery of football made him an infamous pain in the ass was significantly less masterful in charting the end of his career. In early 2024, the Patriots elevated Jerod Mayo—then a young linebackers coach and former Patriots player—to head coach, positioning him as Belichick’s successor. The promise of continuity was central. Mayo had grown up in the organization and had Belichick’s blessing; owner Robert Kraft later revealed that Mayo had been identified for the job years in advance, according to Bleacher Report. But the parts of the succession puzzle that matter most—a clear timeline, the transition of authority, and preparation for leadership—were left fuzzy.

The experiment failed almost immediately. Mayo’s first season ended in a 4-13 record, and the team finished last in its division. The board and Kraft pulled the plug on Mayo after just one season, publicly admitting that they had put him in an “untenable situation.” As if to punctuate the awfulness of the initial succession planning, the Patriots made the Super Bowl the very next year under a different head coach. Jerod Mayo is currently not employed by any NFL team.

***

“The major job of leaders is to develop other leaders,” the management expert Noel Tichy once wrote. But corporate and political history is full of towering figures who gave way to fools and incompetents.

Call it the Aurelius Trap. Strong leaders often fail to do the most important thing in leadership, which is to plan for a successor who can extend their time-bound term into a longer legacy. They have a tendency to make themselves indispensable and irreplaceable—and then prove themselves right by handing the reins to a leader who fails to replace them.

The question of why it is so hard for great leaders to finish the job is interesting enough that it does not require a news peg. But its significance will only grow in the next few years, given the advanced age of modern American leadership. As the Yale professor Samuel Moyn writes in his new book Gerontocracy in America, the political and business worlds are older than ever. In the S&P 500, chief executives are ten-times more likely to be over 70 years old than under 40. Boomers have so dominated American politics that we’ve had more presidents born in the summer of 1946 (three: Bill Clinton, George W. Bush, and Donald Trump) than in all the years after 1946 (one: Barack Obama, born 1961). Our last two presidents have been the two oldest in American history.

A successful handoff from one generation to the next is self-evidently critical in politics. In business, it’s a nine-figure decision. A 2021 study found that the market value wiped out by “badly managed CEO and C-suite transitions” across the largest publicly traded companies was close to 20 percent of annual investor gains. By this measure, the question of why succession planning so often goes wrong is worth, literally, trillions of dollars.

In the reporting for this essay, I read dozens of papers and consulted several leadership researchers. In my best effort to synthesize the wisdom of these resources, I’ve boiled down this literature to four big theories about why great leaders are so bad at what might be their most important job. While no simple summary of their findings can fit inside of one sentence, the best I can offer is this two-sentence summary:

Read more