All podcasts / Dwarkesh Podcast / Summary

Alex Imas and Phil Trammell - What remains scarce after AGI?

2026-06-04 - 76 min - source - Read full transcript

Dwarkesh Patel (host)Alex ImasPhil Trammell

Key insights

Labor's share of national income has held near 60% for roughly two centuries despite repeated waves of automation, and much of the apparent recent decline disappears once accounting-definition changes are held constant.

Imas calls this a Kaldor fact and notes some economists think the recent decline is largely an artifact of changed national-accounting conventions over the last 30-40 years, citing a paper by Atkinson showing labor share hasn't fallen when the accounting is held fixed.

labor-share

Whether AI shrinks labor's share of income hinges on an empirical quantity nobody currently measures: whether the variety of new capital-intensive goods AI creates keeps expanding faster than people satiate on old ones.

Trammell frames this with a Mongolian-economist-in-1400 thought experiment: someone extrapolating from fixed categories (horses, yurts, yogurt) would have wrongly predicted humans would end up spending all their money on singers, because they couldn't foresee the huge variety of new capital-produced goods that would appear instead.

labor-share

A 'relational sector' of goods and services - therapy, performance, caregiving, anything where a human being in the loop is part of the value - should stay scarce even under full automation, because the preference for human involvement appears intrinsic, not just scarcity-driven.

Imas cites an experiment on willingness-to-pay for art prints: a human-made print loses much of its price premium once buyers learn 500 copies exist, but an AI-made print shows no such effect, suggesting people value the artist's presence itself, not just the print's rarity.

relational-sector

Both AI-run firms and human individuals who never satiate on capital accumulation face a selection advantage, meaning a small number of 'greedy' agents could end up controlling a disproportionate and growing share of future output without any population-wide preference shift.

Trammell notes that historically these accumulators have been diluted by 'dissipation shocks' - heirs who squander wealth or foundations that spend it down - but if some agents (human or AI) can sustain indefinite reinvestment (e.g. by living forever or being run purely by algorithm), that dissipation stops applying.

relational-sector

The 'messy middle' scenario, where AI destroys jobs faster than it generates redistributable wealth, is judged unlikely because an AI capable enough to eliminate entire job categories is almost certainly also making the overall economic pie dramatically bigger, not just replacing labor at a marginally lower cost.

Imas argues the scenario requires AI to be exactly threshold-capable: good enough to fully replace software engineers, say, but only a hair cheaper than paying them, so there's no abundance effect to redistribute. Historically automation has expanded the technological frontier substantially rather than just marginally undercutting existing wages.

ai-wealth-redistribution

Universal basic capital - giving citizens broad equity stakes instead of cash transfers - is philosophically attractive because it makes people shareholders rather than dependents on a discretionary government check, but it runs into a hard targeting problem.

Imas notes indexing 'the AGI economy' is much harder than indexing the S&P 500: if the wrong company ends up capturing AI's gains (e.g. a robotics firm displacing today's leading labs), a portfolio built on today's winners could miss the payoff entirely.

ai-wealth-redistribution

Popular AI-recession narratives, such as the viral Citrini essay, require an implausible economic condition: capital owners hitting a hard consumption ceiling and refusing to reinvest their gains, a pattern with no real historical precedent even amid extreme wealth concentration.

Imas wrote a direct rebuttal working backward from 'assume negative GDP growth' to find the conditions that would produce it, concluding they require demand among the wealthy to hit a hard bound rather than merely diminish, which hasn't happened even during past periods of severe inequality.

ai-wealth-redistribution

Countries outside the AI supply chain (India, Nigeria, and similar) are better served buying equity exposure to AI now, via sovereign wealth funds or subsidized citizen ownership, than betting primarily on retraining programs.

Trammell and Imas argue retraining assumes a strong education system, which is often exactly what a poor country lacks, whereas 'buying the index' converts a modest amount of current savings into large future consumption if AI drives high growth and falling capital-good prices.

developing-world-ai-access

Whether ordinary people and developing countries can capture AI's gains just by holding a broad market index depends on whether AI ends up structured like electricity (rents flow mostly to downstream users) or like social media (rents concentrate at the platform).

Imas notes ConEd, an electricity monopoly, has never accumulated outsized political or social power because the benefits of cheap power flowed to users; social media, by contrast, is used by everyone but the profits accrued mainly to the platforms - the guests are uncertain which pattern AI will follow.

developing-world-ai-access

Current labor-market data shows no economy-wide 'white-collar AI apocalypse' yet - only a modest, below-trend slowdown in hiring for junior software engineers, alongside rising demand for senior engineers, not a broad level shift in employment.

Imas cites analysis from the Budget Lab at Yale and cautions that firms may be laying people off partly for narrative reasons - to signal they are 'keeping up' on AI adoption - rather than because of measured productivity gains, which could make current layoff anecdotes a poor guide to the underlying economics.

labor-share

Individual economists' forecasts about automation have a poor historical track record, so the guests argue for building prediction markets and explicit scenario models rather than trusting any single expert's prediction, including their own.

Imas cites David Ricardo's 1820s prediction of mass unemployment from industrial automation: he correctly identified which jobs would be automated but missed that displaced spending would flow into new services, leaving prime-age employment near an all-time high two centuries later - the lump-of-labor fallacy Ricardo didn't anticipate.

ai-forecasting-methodology

Media referenced

What Will Be Scarce - article - Alex Imas's Substack essay proposing the 'relational sector' - goods and services whose value depends on a human being in the loop - as what stays scarce once AI automates everything else.
Blog post on economists' labor-market forecasts (Fradkin, Jabarian, Koh) - article - Cited by Imas to show economists' individual forecasts about AI's effect on jobs disagree wildly, motivating his call for prediction markets over single expert predictions.
Paper on labor share under constant accounting (Atkinson) - paper - Cited to argue labor share has not actually fallen once accounting-definition changes over the last few decades are held constant.
Essay on the 'Messy Middle' scenario - article - Molly Kinder's essay on a scenario where AI automates many jobs without generating enough surplus to cleanly compensate the people it displaces; used as the episode's jumping-off question.
Blog post on the politics of AGI - article - Andy Hall's essay, cited for the observation that even a 2% rise in unemployment radically changes the political environment around automation.
QJE paper on telephone operator automation, 1920-1940 - paper - Cited as a historical precedent for the 'messy middle': operators were reabsorbed into the economy over 20 years but at lower wages, rather than being cleanly laid off.
Research on the falling economic share spent on computing - paper - Chad Jones's finding that even as transistor supply has grown by orders of magnitude, the share of the economy spent on computing has fallen, cited to explain the 'pessimistic' framing of Moore's Law.
Citrini scenario-planning essay on an AI-driven recession - article - A viral essay predicting recession from white-collar automation and lost wages; Imas wrote a rebuttal arguing the required conditions (a hard consumption ceiling for capital owners) are historically implausible.
O-ring automation model (Gans and Goldfarb) - paper - A model showing that if AI can only automate part of a job to a lower quality standard than a human, firms may not automate that part at all - explaining slower-than-expected automation of white-collar work.
Dwarkesh Podcast episode with David Reich - podcast - Referenced for Reich's point that humans remain under strong ongoing natural selection, used to argue a preference for human connection could get reinforced even under initial indifference.
Astronomical waste argument (Bostrom) - paper - Referenced when discussing why a von-Neumann-probe-like optimizer would place near-zero value on human-provided (relational) goods relative to colonizing more solar systems.

Companies

Google DeepMind - Alex Imas's employer; he is Director of AGI Economics there.
Epoch - Phil Trammell's employer; he is Head of Economics there.
Anthropic - Used repeatedly as the example frontier AI lab - discussed re: universal basic capital targeting, the Defense Production Act threat, and going public.
Meta - Mark Zuckerberg's wealth (mostly Meta stock, reinvested rather than converted to consumption) used as an example of a human who behaves like a 'greedy optimizer' for capital.
The Budget Lab at Yale - Cited for current labor-market data showing no economy-wide white-collar AI job losses yet, only a modest below-trend slowdown for junior software engineers.
OpenAI - Discussed alongside Anthropic as a frontier lab whose eventual IPO would make its gains easier for ordinary investors to index.

Techniques and frameworks

Relational sector - Imas's framework for goods and services where a human being in the loop is intrinsically part of the value, not just an efficiency input - the category argued to remain scarce after full automation.
O-ring theory of production - The idea that a single unreliable component (or task) can sink an entire good or service, which the guests use to explain why full automation of complex jobs lags behind AI capability.
Kaldor facts - Stylized long-run regularities of economic growth, including labor's roughly 60% share of income holding constant for two centuries despite repeated waves of automation.
Universal basic capital - Proposal to redistribute AI-driven wealth by giving citizens broad-based equity ownership rather than cash transfers, discussed alongside its targeting problem.
Investment-specific technical change - Macro concept where the price of capital goods falls relative to consumption goods over time, used to explain how interest rates and labor share could behave under explosive AI-driven growth.

Summary

Dwarkesh Patel interviews two economists working directly on AI's economic implications: Alex Imas, Director of AGI Economics at Google DeepMind and a University of Chicago professor, and Phil Trammell, Head of Economics at Epoch and a Stanford research scholar. The conversation works through what happens to wages, labor's share of income, and wealth distribution as AI automation advances, organized around a central question: what will remain scarce, and therefore valuable, once AI can do almost everything?

Imas's central answer is the "relational sector" - goods and services where the fact that a human is in the loop is intrinsically part of the value, not just an artifact of scarcity. He points to an experiment showing buyers discount a human-made art print once they learn 500 copies exist, but show no such discount for an AI-made print, suggesting the preference for human involvement is not merely about rarity. Trammell complicates the picture with an "increasing variety" argument: even if we fully satiate on any fixed set of automated goods, AI could keep inventing genuinely new categories of capital-intensive products fast enough to keep capital's share of spending growing, illustrated by a thought experiment about a 1400s Mongolian economist who could never have predicted the vast expansion of goods beyond horses and yurts. Both guests stress that labor's income share has held remarkably stable near 60% for roughly two centuries (a "Kaldor fact") despite continuous automation, and argue much of the recent apparent decline is an accounting artifact rather than a real structural shift.

A substantial stretch of the episode interrogates the "messy middle" scenario - AI displacing many jobs without generating enough surplus wealth to cleanly compensate the people displaced. Both guests find this scenario narrow: an AI capable of eliminating whole job categories is almost certainly making the economic pie much bigger, not merely undercutting wages by a small margin, so there should be ample surplus to redistribute even if the political mechanics of doing so remain messy. They apply similar reasoning to debunk viral "AI causes recession" narratives (the Citrini essay), which require wealthy capital owners to hit a hard consumption ceiling and simply stop reinvesting - a pattern with no real historical precedent.

On redistribution mechanisms, the guests compare negative income tax, UBI, and "universal basic capital" (broad citizen equity ownership). Imas favors capital ownership on political-economy grounds - it makes people shareholders rather than dependents on a discretionary government check - but flags a hard targeting problem: indexing "the AGI economy" is much harder than indexing the S&P 500, since the company that ultimately captures AI's gains might not be today's obvious winner. This targeting problem is especially acute for developing countries, which the guests argue should prioritize buying equity exposure now (via sovereign wealth funds or subsidized ownership) rather than betting on retraining programs, since retraining assumes an education system many poor countries lack. Whether this strategy works depends on whether AI ends up structured like electricity, where downstream users capture most of the benefit, or like social media, where rents concentrate at the platform - the guests are genuinely uncertain which pattern will hold.

The conversation closes on longer-horizon speculation: selection pressure could favor both AI-run firms and human individuals who never satiate on capital accumulation, meaning a small number of "greedy" agents (potentially including something like a self-replicating von Neumann probe) could end up controlling a disproportionate and growing share of future output, historically checked mainly by "dissipation shocks" like squandering heirs, which may not apply if some agents can sustain accumulation indefinitely. Throughout, both guests are candid about the state of the evidence: Imas repeatedly notes that individual economists' forecasts (including Ricardo's famously wrong 1820s prediction) have a poor track record, and advocates building prediction markets and explicit scenario models rather than trusting any single expert's guess about how this plays out.

Notable Quotes

"It's incredibly surprising that it's over 60% after the Industrial Revolution and all of the automation we've ever seen." - Alex Imas

"I like the pessimistic framing of Moore's law: every 18 months, the value of computation halves. We're running out of uses for computation so fast that it's sustaining Moore's law." - Phil Trammell

"It's more difficult to imagine a good thing that doesn't exist than losing something that exists. It's much easier for somebody to go on a podcast and say, 'These jobs that you like, they're going away,' than for somebody to spin up a utopia which doesn't exist yet." - Alex Imas

"Is AI going to be like electricity or social media? ... With electricity, a lot of the downstream benefits actually came to the users of the electricity rather than the actual entity producing it. On the other hand, with social media ... the rents went to the platform." - Alex Imas

"For abundance to generate negative economic growth, that's really hard to get." - Alex Imas