Tyler Cowen on AI Adoption, Human Agency, and the Search for Beauty

Tyler Cowen imagines that, 30 years from now, people will look back on having lived in two different worlds: before and after AI. Joel Hellermark begins by asking why, despite the models' intelligence, the effects are still hard to see in macroeconomic data.

Cowen says people already use AI informally at work: a memo that once took four hours may take ten minutes plus editing, with the time saved becoming leisure rather than measured output. He thinks the aggregate gain in leisure is already meaningful, even if conventional statistics miss it.

The larger advances, he argues, will come when AI makes science more effective and speeds the testing of hypotheses. Some fields could take ten years or more. Coding is already changing, but profitability and human bottlenecks mean that he expects AI's takeoff to be slower than many anticipate.

Hellermark asks what would make Cowen raise his forecast of a roughly half-percentage-point growth boost to more than two points. Cowen says he would update immediately if inefficient institutions enthusiastically integrated AI, but he regards that as highly unlikely.

Cowen points to government, nonprofits, higher education and inefficient parts of health care as large, slow-moving parts of the economy. In many cases, he thinks startups will need to displace incumbent institutions, a process that can take a generation.

Coding is an obvious area where the revolution has arrived, Cowen says. He compares it with the US information-technology and inventory-management improvements of the late 1990s, when productivity rose by roughly half a percentage point. Even if AI is more capable, he sees human adoption as the constraint.

Hellermark challenges the assumption of a persistent human bottleneck, asking whether autonomous agents could automate call centers. Cowen expects much more complete automation there within ten years, but says the sector is a small part of most economies and change will proceed one sector at a time.

Retraining is also an organizational challenge. Telling an experienced worker that their entire job has changed and must be relearned immediately can be a difficult sell, even without a layoff. Cowen expects that friction to slow implementation.

Asked about India's exposure to better coding models and firms such as Infosys, Cowen predicts that Indian companies will be adept at applying AI. He sees an advantage in the ability to apply ideas cheaply and quickly, and expects startups to play a particularly important role.

Cowen thinks many American organizations outside technology have not yet fully adjusted. He singles out higher education, where AI is often treated as a cheating problem rather than a reason to rethink education. The country has talent and resources, he says, but institutional change remains difficult.

Hellermark asks what becomes valuable if intelligence approaches zero cost. Cowen answers that, without universal humanoid robots, people are still needed to persuade, carry out physical work, assume legal responsibility, engage in politics and motivate others.

When Hellermark asks why large numbers of people would still be needed, Cowen invokes comparative advantage and new projects. Solar installations, care for older people, energy infrastructure and biological experiments all require physical work or a human touch, even if AI and robots help.

As a historical analogy, Cowen tentatively offers electricity. Its adoption took decades, partly because of human and institutional obstacles. He expects AI to spread faster, but finds the earlier delay sobering; electrification also changed life profoundly and had disruptive effects.

Hellermark notes that AI has spread quickly among knowledge workers and asks why the macroeconomic effects are not clearer. Cowen says historical growth statistics are imperfect and that genuinely new capabilities are difficult to express as growth rates: a previously unavailable medical cure, for example, does not fit neatly into a price comparison.

Cowen suggests judging transformation by what life feels like over decades. Each individual year might seem fairly normal even as the world becomes unrecognizable over 30 years. He hopes AI will help him live long enough to experience the contrast between pre- and post-AI eras.

Hellermark asks whether abundant intelligence could produce a return of the polymath. Cowen expects initiative, intuition, charm, charisma and the ability to combine knowledge to matter more than simply having attended a good school or knowing many facts.

He expects inexpensive intelligence and services to improve life for poorer people, while some highly paid professionals who principally supply knowledge could face competition from AI. His example of a lower-paid lawyer illustrates a shift in earnings and location, not what he calls a humanitarian disaster.

Cowen says some lawyers will build companies using AI agents and may become wealthier, while firms may need fewer associates. Human lawyers could concentrate on attracting and keeping clients, and legal rules will affect what AI is allowed to do. He also expects humans to write laws governing AI even if machines could do it better.

Asked about diminishing returns to intelligence, Cowen says more intelligence does not eliminate every bottleneck. A unified theory of physics may require scarce experimental evidence, not just smarter reasoning. He similarly sees limits in some art and music, even if far more work reaches a high standard.

Hellermark asks about mathematics and science. Cowen thinks some difficult mathematical problems may be solved, but translating knowledge into improved living standards still requires physical production, regulation, testing and trust. He expects social resistance to grow as AI becomes more powerful.

On adoption inside companies, Cowen's near-term picture is less cheerful than simply making every worker more capable. He notes that AI companies are hiring teams to help firms redesign workflows because few people know how. Education faces the same challenge: ideas for an AI-centered curriculum remain largely untested.

He expects trial and error, frustration and a period when some things get worse. At the same time, current models can already draft memos, check facts or proofread without needing to become dramatically smarter. He does not expect most of the economy to remove people from the loop in the next 20 or 30 years.

Beyond coding, Cowen expects major advances in bioscience. Converting discoveries into medicines and devices is slower, but he speculates that over roughly 40 years many deadly diseases could be cured. He recognizes health care as a heavily bottlenecked sector, which could limit measured growth even if the benefits are enormous.

Cowen says he has already shifted about two-thirds of his effort away from writing and toward podcasts, public talks and mentoring. He could have relied on tenure and established income, but wants to remain actively involved as AI changes the competitive landscape.

He believes AI writing quality will soon rival or exceed his own for many purposes, although readers may still value his particular perspective. Hellermark suggests that Cowen's distinctive references matter; Cowen replies that live talks and podcasts make those references vivid in a way he doubts AI can replicate.

Cowen says early access to GPT-4 was the decisive moment. Within minutes he concluded that, given scaling and lessons from computer chess, he needed to change his plans. He had anticipated an AI revolution in his 2013 book Average Is Over, but says it arrived several years sooner than expected.

He uses several models to keep up with their changing strengths, relying most on GPT Pro at the time of the interview. For travel he asks about restaurants, hotels and performances; he says the model's recommendations can be well tailored because his books, blog and podcasts give it extensive material about his tastes.

Cowen also reads with language models open, asking about unfamiliar references. While studying eighteenth-century Persia, he uses them to clarify names and facts. He does not claim the model has offered him new historical theories; what he values is the help learning the material.

Hellermark asks whether repeated stories about evil AI could influence models and how they should be trained to act. Cowen says he worries about such influences and favors teaching cooperation and gains from trade. His broader advice is to articulate constructive solutions and a positive vision rather than dwell only on fear.

The discussion turns to national attitudes. Cowen speculates that American apocalyptic traditions and different expectations of state support help explain contrasts with Sweden. He also describes China as comparatively practical and pro-AI, while noting that unusual American ideas have helped make the US a center of AI development.

Hellermark asks how Cowen spots talent early. Cowen looks for energy, determination, attention to detail and a strong sense that the person's work matters. He wants intelligence too, but cautions against weighting IQ so heavily that other qualities are missed.

Cowen no longer relies on his old favorite interview questions because people have learned to game them. Instead, he tries to get someone talking about a subject they care about without a prepared pitch: a place, a meal or a film. Detailed observations can reveal attention and genuine interest.

To update his own information diet, Cowen pays attention to where talented young people are going. He says recent conversations with teenagers interested in quantum computing made him take the field more seriously, even though he cannot independently judge the underlying science.

Asked what great groups share, Cowen names ambition, work ethic, belief in the importance of the work and the willingness to set other things aside. He also stresses rich sources: the surroundings of Renaissance Florence, or the British and American musical influences available to the Beatles.

For the Beatles, Cowen highlights the rivalry and cooperation between John Lennon and Paul McCartney. They competed over leadership and songs but also helped each other. In his account, their creative friction and changing contributions eventually contributed to the group's breakup, after producing extraordinary work.

Cowen thinks there is still great music today but fewer enduring groups. He contrasts the longevity and determination of artists from the 1960s and 1970s with a contemporary industry that he sees as more centered on streaming hits and tours. This is his judgment about the culture, not a claim that good music has disappeared.

Hellermark asks about Cowen and Patrick Collison's call for a new aesthetic. Cowen says they are struck by the contrast between beautiful older architecture and much postwar construction. Their starting point is a call for help rather than a prescription that all buildings follow a single style.

Cowen acknowledges that many interiors are beautiful and comfortable today, even when exteriors disappoint him. He hopes AI will create enough wealth and information for people to make public beauty a source of pride and status again. He is encouraged by the response to the call, but says he cannot predict whether it will succeed.

He doubts that a diverse, globalized world will settle on one unified aesthetic movement. Still, he points to attractive newer areas of Copenhagen and examples in Latin America as signs that people can try to make places beautiful and succeed.

Asked whether AI can discover beauty, Cowen recalls uploading Vermeer's The Art of Painting to GPT and receiving an explanation he considered superb. Explaining beauty is not the same as creating it, he says, but he expects AI to produce visual art and music that many honest observers find beautiful within ten years, perhaps sooner.

In rapid-fire questions, Cowen recommends Joseph Conrad's The Secret Agent, which he reread while thinking about political assassination. He also says economics has made substantial empirical progress but, in his view, little recent theoretical progress.

A pair of books that changed his thinking are Olaf Stapledon's Last and First Men and Star Maker. Cowen recommends their speculative exploration of intelligence and radically changing worlds, drawing from them the lesson not to underestimate what is possible.

One economic surprise for Cowen is that countries including the US and Japan have sustained high public debt without the crises that many observers, including him, would once have expected. Borrowing costs have risen, he says, but the economies still function; he remains surprised and uncertain.

Cowen says he has not resolved how a future of rapidly changing jobs will affect career planning, or how AI might change religious life and authority. He expects energy to constrain extremely rapid economic expansion and regards speculative proposals such as a Dyson sphere as distant.

Asked for the most contrarian place, Cowen suggests parts of Sicily or rural Mexico where people expect relatively little change and traditional relationships persist. He thinks they may be wrong, but says that such expectations are more contrarian now than the technology-oriented views of Silicon Valley.