“What happens when machines are better than people at literally everything? At that point work will disappear, right? Wrong.” - David Deming
What happens to human work when AI is better than people at every job task? The obvious answer - work disappears - is wrong, because of one of the oldest and most important ideas in economics. David Deming - labor economist and Dean of Harvard College - builds the argument on the proposition Paul Samuelson offered when challenged to name one social-science claim both true and non-trivial: comparative advantage. Ricardo showed in 1817 that two countries both gain from trade even when one produces everything more efficiently. The same arithmetic governs a human and a machine, provided the machine’s time has a cost. It does, at least for now: frontier models cost twice what their predecessors did, and some workers burn through monthly token budgets in days. The human advantage relative to AI is that we learn well from small amounts of data - between a thousand and a million times more efficiently than current models.
Listen on YouTube, Spotify, and Apple Podcasts
A note from me: the solo episode I released on June 24 has been our most downloaded and most watched show so far. Maybe that means you want more of me, but I suspect it mostly reflects how much people care about AI and jobs. Either way, you asked and I listened - so this week I’m starting a two-part series on human work in the age of AI.
Part one provides the foundation: why I do NOT think AGI, or even superintelligence, means the end of human work. The argument runs through a principle that Paul Samuelson called true but not trivial, and that thousands of important and intelligent people have failed to grasp. Give me five minutes and you’ll understand something many powerful people do not. Next week, in part two, I’ll tell you which skills and jobs I think will be most valuable in the age of AI.
As always, tell me what you think at david@thecontextwindow.com - ideas for what I can do better, suggestions for future episodes, or accusations that I am AI-generated. If you prick us, do we not bleed?
Chapters
[00:00] Cold open What happens to work when machines beat people at everything? The obvious answer is wrong.
[00:58] A powder day for labor economists New technology is fresh snow for a labor economist. David pivoted his research within a year of trying ChatGPT, knowing it is a luxury to study the disruption without fearing it.
[03:59] The Ulam challenge Stanislaw Ulam escaped Poland eleven days before the German invasion and later dared Paul Samuelson to name one proposition in the social sciences both true and non-trivial. Samuelson stewed on it for nearly thirty years.
[05:19] Ricardo’s answer The answer dates to 1817. Ricardo’s English cloth and Portuguese wine prove that two countries gain from trade even when one of them makes everything more efficiently.
[07:21] From countries to coworkers Workers trade tasks the way countries trade goods. This is the model from David’s social-skills paper, which found the labor market shifting toward teamwork-heavy jobs.
[09:25] Goldin and Katz write a paper A toy economy with two economists and one production function. Also the story of the best paper title in economics: “Nimble Critters or Agile Predators?”
[13:02] When Katz is better at everything Give Katz absolute advantage in both tasks and collaboration still wins, 2.67 papers to 2.5. The margin is small only because the example is.
[15:03] The AI in the metaphor Swap Katz for a superintelligence and nothing in the logic changes, however wide the capability gap grows. One condition has to hold.
[15:35] Everybody has a part to play David steps back from the algebra: human skills stay valuable because they are scarce, while AI becomes a commodity. Farming is the precedent - machines absorbed nearly all the labor, and the work that stayed human grew more important, not less.
[17:00] The Cunningham caveat If AI’s time costs nothing, the argument collapses. Tom Cunningham presses this point on David; current token prices answer it, for now.
[18:40] Garfield and Trump Name the president on July 8, 1881, then the president today. Your brain retrieves both for a twentieth of a joule; the model needs a web search. That gap runs 1,000 to 10,000x.
[22:03] Why models search instead of learn GPT-5.5’s world knowledge stops on December 1, 2025. Training runs cost too much to repeat often, so nothing the model learns in a session survives the session.
[23:55] The million-x economy Mercor pays professionals $100 an hour to perform their jobs as AI training data. The labs buying it are betting the million-x cost of machine learning pays back at scale.
[25:09] In-distribution versus long-tail Common, well-documented tasks get absorbed into the model. Rare ones keep their human premium. What that work pays is where part two begins.
[26:22] Housekeeping More solos are coming. David answers the YouTube commenters who think he is AI-generated with a line from Shylock.
Mentioned in this episode
Papers
David Deming, “The Growing Importance of Social Skills in the Labor Market“ (Quarterly Journal of Economics, 2017) - the task-trade model behind this episode’s worked example
David Deming, Claudia Goldin and Lawrence Katz, “The For-Profit Postsecondary School Sector: Nimble Critters or Agile Predators?“ (Journal of Economic Perspectives, 2012)
Alexander Bick, Adam Blandin and David Deming, “The Rapid Adoption of Generative AI“ (Management Science, 2026) - the first nationally representative US survey of generative AI adoption
Alexander Bick, Adam Blandin, David Deming, Nicola Fuchs-Schündeln, and Jonas Jessen, “Mind the Gap: AI Adoption in Europe and the U.S.“ (NBER Working Paper, 2026)
Aaron Chatterji et al., “How People Use ChatGPT“ (with David Deming)
Books
David Ricardo, On the Principles of Political Economy and Taxation (1817) - the origin of comparative advantage
People and ideas
Paul Samuelson, “The Way of an Economist“ (1969) - the essay answering Stanislaw Ulam’s challenge, and the source of the “true but not trivial” line
Stanislaw Ulam - mathematician, Manhattan Project scientist, and Samuelson’s provocateur; his brother Adam Ulam became a leading Kremlinologist at Harvard
Comparative advantage, absolute advantage, gains from trade, autarky, and perfect complements - the toolkit of the episode
Tom Cunningham - economics and AI research at METR (formerly OpenAI), David’s friendly adversary on whether comparative advantage survives free, infinitely replicable AI
Data and examples
OpenAI, GPT-5.5 model documentation - the December 1, 2025 knowledge cutoff behind the Garfield-and-Trump example
Mercor - the expert marketplace paying professionals $100 an hour and up to perform white-collar tasks for AI training
Further reading
Dwarkesh Patel, “The sample efficiency black hole“ - the estimate that humans are somewhere between thousands and a million times more sample efficient than AI models
Credits
Host: David Deming, Danoff Dean of Harvard College
Executive Producer: Denise Koller Consulting Producers: Tim Smith and Jonathan Palumbo
