“The ‘compute’ required to successfully navigate a 5-minute unstructured conversation with another person is massive, much bigger than figuring out the optimal move in the middle of a chess game.” - David Deming
In 1961, TIME magazine warned that automation might permanently outrun the economy’s ability to create new jobs. Sixty-five years later the machines are smarter and the worry is exactly the same. In part two of his series on work in the age of superintelligent AI, David Deming - labor economist and Dean of Harvard College - argues that technology doesn’t end work, it reprices skills, and that the next skill to be repriced upward is one we barely think of as a skill at all. Along the way: what a benchwarmer taught the NBA about hidden value, what a parking lot full of food trucks reveals about the limits of machine social reasoning, why AI can name the emotion on your face but can’t tell you what your coworker’s bad mood means, and why a relationship is not a transcript.
Listen on YouTube, Spotify, and Apple Podcasts
Last week I made the case that superintelligent AI won’t end human work. This week I want to answer the obvious follow-up: fine, but what kind of work will we do when machines are smarter than us? My answer runs through a 1961 TIME magazine article, a basketball player who never showed up in the box score, and a research finding that being a team player matters about as much as IQ.
One thing you can do while you listen: my colleagues Ben Weidmann and Yixian Xu built a test of emotion perception called PAGE, and you can take it yourself at skillslab.dev to see how you stack up. Fair warning - the AI models are pretty good at it. That’s not where our edge lives, which is sort of the whole point of the episode.
Let me know what you think at david@thecontextwindow.com. Next week I’ll get more precise about how social skills actually create economic value, and whether the jobs of the future will pay decent wages.
Chapters
[00:00] Cold open Social skills don’t mean cocktail parties. Why the deep context of personal relationships must be built one at a time - and what that means for AI.
[00:46] “Do morons know they are morons?” Everyone from Musk to Gates to Yang predicts the end of work. David thinks they’re all wrong, and interrogates his own confidence on the way in.
[02:05] A warning from 1961 A quote about automation killing jobs that could run in tomorrow’s paper, and the LBJ commission that recommended a guaranteed minimum income - sixty years before Andrew Yang.
[02:59] The great repricing Mechanized farming made muscle cheap and literacy valuable. What happens when intelligence itself becomes abundant.
[04:00] The 1% that remains American farmers lost 99% of their tasks to machines and ended up with a median household wealth of $1.6 million. Your great-grandparents would not believe what you do for a living.
[05:39] Speed runners and completionists David’s childhood video game habit - searching every room of the dungeon - as a model of how AI learns, and why exhaustive learning only pays if the floor plan never changes.
[06:14] Low-context, high-context Code either compiles or it doesn’t. A conversation has an infinite action space and a murky scoreboard. Where each kind of mind wins.
[08:51] The no-stats all-star Shane Battier, plus-minus, and the identification problem: many people on a team, one outcome. Who gets the credit?
[10:58] Team players The Econometrica experiment that randomly assigned people to teams, over and over. Being a team player matters about as much as IQ - and is completely unrelated to it.
[12:36] Reading the mind in the eyes The RMET, its successor PAGE, and a sample item you can try on screen. Take the full test at skillslab.dev.
[14:01] Ted Bundy would ace the test AI already beats humans at reading emotions in faces. Why emotion recognition is necessary for teamwork but nowhere near sufficient.
[15:10] Predictably irrational Models can forecast the behavior of homo economicus. Most of us are not him. What thousands of real risk choices reveal about learning who a person actually is.
[17:17] The food truck problem Inverse planning in a parking lot: what a person’s route reveals about their preferences, and where GPT-4 falls apart.
[18:59] One step the wrong way Inverse-inverse planning - moving so that someone else can read your mind. Humans pick the subtle signal; the AI almost never does.
[20:47] “Did you mean the sales report?” Conversational grounding, and the Microsoft finding that LLMs ask for clarification 3x less and make follow-up requests 16x less than people do.
[22:49] Hardwired Ten-month-old infants doing inverse planning, thin-slice judgments, and Joe Henrich’s argument that culture is a second inheritance system.
[25:29] 35 trillion ways to know 10 people The combinatorial explosion of social networks, and how people infer office politics from six short interactions.
[27:22] A relationship is not a transcript Why memory features don’t solve the problem: storing facts about a person is not building a mental model of them, and shared histories can’t be brute-forced.
[28:26] Next week How social skills create economic value, and whether the jobs of the future will pay decent wages.
Mentioned in this episode
Papers
Ben Weidmann and David J. Deming, “Team Players: How Social Skills Improve Team Performance” (Econometrica, 2021) — the repeated random-assignment experiment behind the “team player” result.
Simon Baron-Cohen et al., “The ‘Reading the Mind in the Eyes’ Test Revised Version: A Study with Normal Adults, and Adults with Asperger Syndrome or High-Functioning Autism” (Journal of Child Psychology and Psychiatry, 2001) — the emotion-recognition measure later used to predict who makes teams better.
Junqi Wang et al., “Evaluating and Modeling Social Intelligence: A Comparative Study of Human and AI Capabilities” (Proceedings of the 46th Annual Conference of the Cognitive Science Society, 2024) — the food-truck and restaurant-grid experiments on inverse reasoning and inverse-inverse planning.
Ryan Liu et al., “Large Language Models Assume People Are More Rational Than We Really Are” (arXiv preprint, 2024; revised 2025) — the study showing that LLMs model people as more rational and predictable than they really are.
Omar Shaikh et al., “Navigating Rifts in Human-LLM Grounding: Study and Benchmark” (ACL, 2025) — the chat-log study behind the finding that LLMs were three times less likely to initiate clarification and sixteen times less likely to make follow-up requests.
Shari Liu et al., “Ten-Month-Old Infants Infer the Value of Goals from the Costs of Actions” (Science, 2017) — the gate-jumping study.
Nalini Ambady and Robert Rosenthal, “Thin Slices of Expressive Behavior as Predictors of Interpersonal Consequences: A Meta-Analysis” (Psychological Bulletin, 1992) — the research on judgments from brief observations that was later popularized by Blink.
Isaac Davis, Julian Jara-Ettinger, and Yarrow Dunham, “Inferring the Internal Structure of Groups Through the Integration of Statistical Learning and Causal Reasoning” (Nature Communications, 2026) — the office-politics experiment.
Articles
Michael Lewis, “The No-Stats All-Star” (The New York Times Magazine, 2009) — the Shane Battier profile.
“The Automation Jobless,” TIME, February 24, 1961 — the source of the automation warning quoted in the episode.
Books
Joseph Henrich, The Secret of Our Success: How Culture Is Driving Human Evolution, Domesticating Our Species, and Making Us Smarter (Princeton University Press, 2016) — culture as a second inheritance system.
Angela Duckworth fans will remember that Blink came up in a different spirit: Malcolm Gladwell, Blink: The Power of Thinking Without Thinking (Little, Brown and Company, 2005).
Try it yourself
The PAGE test (Perceiving AI-Generated Emotions), built by Ben Weidmann and Yixian Xu at the Skills Lab.
Further reading and listening
Part one of this series: “Superintelligent AI Won’t End Human Work. Here’s the Math.”
Related solo episode: “Why New College Grads Can’t Find Jobs (It’s Not AI)”
Credits
Host: David Deming, Danoff Dean of Harvard College
Executive Producer: Denise Koller Consulting Producers: Tim Smith and Jonathan Palumbo
