Andrei Fredkin, Brian DeBerrian, and Andrew Koh found in a recent blog post that economists' forecasts about the labor market show massive disagreement in every direction.
Alex Imas and Phil Trammell – What remains scarce after AGI?
For AGI-driven economic collapse to cause negative growth, the ultra-wealthy would have to stop investing entirely — economists call this nearly impossible, even in a singularity scenario.
Jun 4, 2026
1:16:08
Difficulty: Intermediate
Played
Dwarkesh Podcast
Alex Imas and Phil Trammell – What remains scarce after AGI?
For AGI-driven economic collapse to cause negative growth, the ultra-wealthy would have to stop investing entirely — economists call this nearly impossible, even in a singularity scenario.
Jun 4, 2026
1:16:08
Difficulty: Intermediate
Played
TL;DR
Economists Alex Imas and Phil Trammell join Dwarkesh Patel to map the economic landscape after AGI. They argue that demand collapse causing negative growth is nearly impossible [1] — Alex Imas "For AGI to cause negative economic growth, rich capital owners would have to stop investing entirely — not just consume less, but refuse to…" 35:10 , that the "messy middle" of automation without broad wealth creation is a narrow but real risk [2] — Phil Trammell "For AI to automate white-collar work without generating redistribution wealth, automation would have to be cheap enough to replace workers …" 19:35 , and that the relational sector — goods where human involvement is intrinsically valued — may survive automation but only if consumer data confirms genuine willingness to pay premiums [3] — Alex Imas "Scarcity after AGI will concentrate in goods where human involvement is intrinsically part of the value — not just because humans are capab…" 00:44 . The single clearest takeaway: developing countries should prioritize indexing into AGI's returns rather than retraining programs.
#AGI economics
#labor share
#relational sector
#automation policy
#wealth redistribution
#UBI vs. Universal Basic Capital
#developing country AI strategy
#demand elasticity
#Moore's Law reframing
#O-ring theory
#capital accumulation
#AI distribution models
#index investing
#lump of labor fallacy
#investment-specific technical change
#capital share
#automation
#redistribution
#UBI
#universal basic capital
#Moore's Law
#developing countries
#scarcity
#Kaldor facts
#wealth concentration
Alex Imas (Director of AGI Economics, Google DeepMind) and Phil Trammell (Head of Economics, Epoch AI) discuss what economics tells us about automation, labor share, optimal taxation, redistribution, and what will remain scarce after AGI.
Chapter list
-
Dwarkesh introduces Alex Imas and Phil Trammell. They explore labor share stability, the relational sector, the Mongolian economist thought experiment, and whether human intrinsic goods can survive automation. [1] — Alex Imas "Scarcity after AGI will concentrate in goods where human involvement is intrinsically part of the value — not just because humans are capab…" 00:44
-
Discussion of whether AI can automate enough jobs to cause political disruption without generating redistributable surplus. The 'drip' scenario of slow displacement leading to underemployment is explored. [1] — Phil Trammell "For AI to automate white-collar work without generating redistribution wealth, automation would have to be cheap enough to replace workers …" 19:35
-
Comparison of UBI, negative income tax, universal basic capital, wealth taxes, and consumption taxes as redistribution tools. Key debate: UBI creates political dependency; universal basic capital has targeting problems. [1] — Alex Imas "UBI makes basic needs contingent on who holds political power — a dangerous dependency. Universal Basic Capital gives people ownership stak…" 25:58
-
Analysis of the Citrini AGI recession scenario. Negative growth requires wealthy capital owners to stop investing entirely — an implausible condition. Yale Budget Lab data shows no white-collar bloodbath yet. [1] — Alex Imas "For AGI to cause negative economic growth, rich capital owners would have to stop investing entirely — not just consume less, but refuse to…" 35:10
-
O-ring theory explains current limits on AI automation. But once AIs are reliable enough, the same logic will make human integration into AI-native production flows practically impossible. [1] — Dwarkesh Patel "O-ring theory explains why automating 9 out of 10 job tasks sometimes lowers output quality enough to make automation counterproductive. Bu…" 39:55
-
Discussion of selection pressure favoring non-satiating capital accumulators — tech billionaires, AI agents, von Neumann probes. Even a few such agents could dominate economic preferences and drive labor share toward zero. [1] — Dwarkesh Patel "If even a small number of agents — whether AI firms, humans like Elon Musk, or von Neumann probes — have effectively unlimited demand for c…" 43:50
-
Policy recommendations for countries outside the AI supply chain. Primary advice: index into AGI returns via sovereign wealth funds rather than retraining programs. AI-as-electricity vs. AI-as-social-media distinction is key. [1] — Phil Trammell "The most robust strategy for developing countries facing AGI is to index into the returns — buy sovereign wealth exposure to the AI supply …" 1:02:20
- Labor share
- The fraction of total GDP paid out as wages and salaries to workers, as opposed to returns to capital; has hovered above 60% in most developed economies for centuries.
- Capital share
- The fraction of GDP paid to owners of capital — machines, land, buildings, and equity — rather than to workers; complements labor share to sum to roughly 100%.
- Kaldor fact
- A set of empirical regularities about economic growth observed by economist Nicholas Kaldor, including the remarkable long-run stability of labor's share of national income.
- Relational sector
- Goods and services where the human presence in the production process is itself part of the value — e.g., a human therapist or barista — not just the output.
- O-ring theory
- A model of production where the weakest link determines overall quality, inspired by the Challenger disaster; used here to explain why partial AI automation may not yet be viable.
- Lump of labor fallacy
- The mistaken belief that there is a fixed amount of work in an economy, so automation permanently destroys jobs rather than allowing new tasks and sectors to emerge.
- Conjoint analysis
- A survey method that asks respondents to trade off between product features to reveal their underlying willingness to pay for each attribute.
- Demand elasticity
- How sensitive the quantity demanded of a good is to changes in its price; highly elastic demand means a price drop leads to a more-than-proportional increase in quantity purchased.
- Jevons paradox
- The counterintuitive finding that making a resource cheaper can increase total consumption of it by so much that absolute usage rises, first observed for coal in 19th-century Britain.
- Investment-specific technical change
- A form of technological progress where the price of capital goods falls relative to consumption goods, so a given amount of investment buys more and more productive capacity over time.
- Universal Basic Capital
- A policy proposal where every citizen receives an ownership stake in productive capital assets — like a diversified stock portfolio — rather than cash transfers.
- Negative income tax
- A tax system where people below a certain income threshold receive supplemental pay from the government rather than paying taxes, effectively providing a guaranteed income floor.
- Wealth tax
- An annual tax levied on the total net worth of individuals, as opposed to income or consumption taxes; debated as a tool for redistributing AI-generated wealth.
- Network-adjusted factor share
- The share of value added by capital or labor in a good when you trace the entire supply chain, not just the final production step.
- Dissipation shock
- A term used in growth economics for events that disperse concentrated wealth — like an heir spending down a fortune or a billionaire donating to a foundation.
- Von Neumann probe
- A hypothetical self-replicating spacecraft that uses local resources to build copies of itself, used here as a metaphor for a maximally greedy self-replicating optimizer.
- Satiation
- The point at which additional consumption of a good yields no further utility; used here to describe whether rich people or AI agents will ever 'have enough' capital.
- Sovereign wealth fund
- A state-owned investment fund that holds financial assets like stocks and bonds on behalf of a nation, proposed here as a mechanism for developing countries to index into AGI returns.
- Neuralese
- Informal term used in the episode for the internal representations or communication protocols native to AI systems, which humans cannot natively read or participate in.
- Leapfrogging
- When developing economies skip intermediate stages of technological adoption and jump directly to the most advanced technology, as happened with mobile banking in parts of Africa.
Chapter 1 · 00:00
Will capital share increase?
Dwarkesh introduces Alex Imas and Phil Trammell. They explore labor share stability, the relational sector, the Mongolian economist thought experiment, and whether human intrinsic goods can survive automation. [1] — Alex Imas "Scarcity after AGI will concentrate in goods where human involvement is intrinsically part of the value — not just because humans are capab…" 00:44
Claims made here
Prime-age employment in 2026 is the second highest on record, with only the year 2000 being higher.
The O*NET database tracking job tasks has been rarely updated and is of very low quality.
Andy Atkinson's paper shows that if accounting methods are held constant over time, labor share has never actually fallen.
Chad Jones's research shows that the share of the economy devoted to paying for computing has been decreasing over time, despite massive expansion in the number of transistors.
Phil Trammell's pessimistic reframing of Moore's Law: every 18 months the value of a unit of computation halves because new uses for compute are exhausted so rapidly.
H100 GPU rental prices are higher today than they were 3 years ago, despite significantly improved technology and greater compute supply.
In Alex Imas's incentive-compatible experiment, people paid significantly more for human-made art prints than AI-made ones, and the human premium collapsed when 500 copies were produced while AI art showed no change.
The Relational Sector: What Stays Scarce After AGI
Scarcity after AGI will concentrate in goods where human involvement is intrinsically part of the value — not just because humans are capable, but because consumers actively prefer the human to be in the loop. The hypothesis only holds if willingness-to-pay survives replacement, and we currently lack the data to know.
David Ricardo Was Right About the Jobs and Wrong About Everything Else
Ricardo correctly predicted that industrial-era jobs would be automated. He completely failed to predict that new jobs would replace them, and that prime-age employment in 2026 would be near an all-time high. The lump-of-labor fallacy has fooled experts for two centuries — and may be fooling them again.
Prime-age employment at near-record high in 2026
Despite all historical automation, prime-age employment in 2026 is the second highest ever recorded, surpassed only by the 2000 peak — a fact David Ricardo would have found shocking.
O*NET database rarely updated, super low quality
The main U.S. database tracking job tasks has rarely been updated and is of very low quality, making it nearly impossible to track which jobs are being created or destroyed by automation.
Data point
60%
Despite every wave of automation since the Industrial Revolution, labor's share of GDP has remained above 60%, which economists call a Kaldor fact — almost suspiciously stable.
The Mongolian Economist Thought Experiment
A Mongolian economist in 1400 predicting scarcity would have assumed people satiate in horses and yogurt and concentrate spending on singers. They would have been wrong — because wealth expansion always generates new varieties. The same failure awaits anyone predicting AGI scarcity by holding varieties fixed.
The Pessimistic Reframing of Moore's Law
Moore's Law isn't just about supply — it's about demand evaporation. We create new uses for compute so fast that the value of each unit halves every 18 months. For the first time, AI may have broken this: H100 rental prices are higher now than three years ago despite massive supply expansion.
Pessimistic framing: Moore's Law = value of compute halves every 18 months
Phil Trammell reframed Moore's Law: rather than compute doubling every 18 months, you could say we run out of uses for computation so fast that the value of a unit of compute halves every 18 months.
H100 costs more to rent now than 3 years ago
Despite massive improvements in compute technology and supply, H100 GPU rental prices are higher than 3 years ago because smarter AI models raise the opportunity cost of compute.
The Art Print Experiment: Quantifying Relational Value
In an incentive-compatible experiment, people paid significantly more for human-made art prints than AI-made ones — but the human premium collapsed when 500 copies existed, while AI art was already treated as a commodity with no such drop. This is the kind of data needed to validate the entire relational sector hypothesis.
Art print experiment: human-made valued much higher than AI-made
In an incentive-compatible experiment, people paid significantly more for art prints made by a human versus AI, and this premium collapsed for humans but not AIs when 500 copies were produced.
Why the 'Messy Middle' Scenario Is a Narrow Risk
For AI to automate white-collar work without generating redistribution wealth, automation would have to be cheap enough to replace workers but not cheap enough to produce abundance. That's an implausibly narrow window — and history suggests technological frontiers always expand alongside job displacement.
Chapter 2 · 19:36
Messy Middle scenario
Discussion of whether AI can automate enough jobs to cause political disruption without generating redistributable surplus. The 'drip' scenario of slow displacement leading to underemployment is explored. [1] — Phil Trammell "For AI to automate white-collar work without generating redistribution wealth, automation would have to be cheap enough to replace workers …" 19:35
Claims made here
A 2% increase in unemployment is sufficient to completely shift political winds, according to Andy Hall's analysis of the politics of AGI.
Telephone operators were fully automated between 1920 and 1940 — over 20 years — and were reabsorbed into the economy at lower salaries and mostly underemployed.
Data point
2%
Political scientist Andy Hall found that even a 2% rise in unemployment is enough to dramatically shift political outcomes, making a slow 'drip' automation scenario particularly dangerous.
Phone operators automated over 20 years (1920–1940)
Even though the technology to automate telephone operators existed, the transition took 20 years, resulting in workers being reabsorbed into the economy but at lower salaries and mostly underemployed.
Chapter 3 · 25:57
How to tax and redistribute AI wealth
Comparison of UBI, negative income tax, universal basic capital, wealth taxes, and consumption taxes as redistribution tools. Key debate: UBI creates political dependency; universal basic capital has targeting problems. [1] — Alex Imas "UBI makes basic needs contingent on who holds political power — a dangerous dependency. Universal Basic Capital gives people ownership stak…" 25:58
Universal Basic Capital vs. UBI: The Political Risk Difference
UBI makes basic needs contingent on who holds political power — a dangerous dependency. Universal Basic Capital gives people ownership stakes with property rights, making them normal shareholders rather than welfare recipients. The tradeoff is targeting: which stocks do you put in people's portfolios?
Chapter 4 · 30:02
Why demand collapse is unlikely
Analysis of the Citrini AGI recession scenario. Negative growth requires wealthy capital owners to stop investing entirely — an implausible condition. Yale Budget Lab data shows no white-collar bloodbath yet. [1] — Alex Imas "For AGI to cause negative economic growth, rich capital owners would have to stop investing entirely — not just consume less, but refuse to…" 35:10
Claims made here
The Yale Budget Lab's analysis of the entire economy finds almost no signal of mass AI-driven unemployment, with only a slight below-trend decline in junior developer hiring.
No evidence of white-collar bloodbath in current data
The Yale Budget Lab's analysis finds almost no signal of mass AI-driven unemployment across the economy, with junior developer hiring only slightly below trend rather than sharply declining.
Why Demand Collapse Causing Negative Growth Is Nearly Impossible
For AGI to cause negative economic growth, rich capital owners would have to stop investing entirely — not just consume less, but refuse to build more data centers even as the technological frontier expands. Abundance causing recession requires conditions that have never existed in history.
Negative growth requires demand hard-bound to stop
For AI automation to cause negative economic growth, wealthy capital owners would need to completely stop investing — not just spend less — a condition economists consider nearly impossible.
Chapter 5 · 39:26
Human employees would be hard to integrate into the machine economy
O-ring theory explains current limits on AI automation. But once AIs are reliable enough, the same logic will make human integration into AI-native production flows practically impossible. [1] — Dwarkesh Patel "O-ring theory explains why automating 9 out of 10 job tasks sometimes lowers output quality enough to make automation counterproductive. Bu…" 39:55
Claims made here
Ganz and Goldfarb's O-ring automation model shows that if you can only automate 9 out of 10 parts of a job but at lower quality, you may not want to automate any of it.
O-Ring Theory and Why AI Can't Yet Automate Entire Jobs
O-ring theory explains why automating 9 out of 10 job tasks sometimes lowers output quality enough to make automation counterproductive. But this friction works symmetrically: once AIs are reliable enough, the same logic will make humans — who slow down AI production flows — impossible to integrate.
Chapter 6 · 43:08
What if some humans (or AIs) value wealth accumulation intrinsically?
Discussion of selection pressure favoring non-satiating capital accumulators — tech billionaires, AI agents, von Neumann probes. Even a few such agents could dominate economic preferences and drive labor share toward zero. [1] — Dwarkesh Patel "If even a small number of agents — whether AI firms, humans like Elon Musk, or von Neumann probes — have effectively unlimited demand for c…" 43:50
Greedy Optimizers and the Future of Capital Share
If even a small number of agents — whether AI firms, humans like Elon Musk, or von Neumann probes — have effectively unlimited demand for capital accumulation, they will compound faster than everyone else and gradually dominate the economy's preference structure. Labor share could go to zero through selection alone.
Investment-specific technical change: robots compound but ballerinas don't
Phil Trammell explained that in a world of investment-specific technical change, one robot today becomes many robots next year, but the number of ballerinas stays the same — a dynamic most macro models miss.
Chapter 7 · 1:01:28
What should developing countries do?
Policy recommendations for countries outside the AI supply chain. Primary advice: index into AGI returns via sovereign wealth funds rather than retraining programs. AI-as-electricity vs. AI-as-social-media distinction is key. [1] — Phil Trammell "The most robust strategy for developing countries facing AGI is to index into the returns — buy sovereign wealth exposure to the AI supply …" 1:02:20
Claims made here
Well under 20% of the total market cap of non-tiny U.S. companies is currently private, making broad index-based access to AI returns still feasible.
What Developing Countries Should Do Right Now
The most robust strategy for developing countries facing AGI is to index into the returns — buy sovereign wealth exposure to the AI supply chain — rather than rely on retraining programs. If AGI diffuses like electricity, index ownership captures the gains. If it concentrates, retraining won't save you anyway.
AI as Electricity vs. Social Media: The Stakes
If AGI diffuses like electricity, every company captures it and index ownership works. If it diffuses like social media, platform rents stay concentrated and ordinary people miss the gains. Which model prevails depends largely on whether open models stay competitive and whether frontier labs go public.
AI could look like electricity or social media — very different outcomes
If AI distributes like electricity, downstream users capture most gains; if it distributes like social media, rents accrue to platforms. This distinction determines whether wealth becomes highly concentrated.
Leapfrogging: mobile banking more common in Nigeria than Germany
Mobile banking is more prevalent in Nigeria than in Germany, illustrating how developing countries can leapfrog stages of technological development with transformative technology.
Every S&P 500 company of the future will have leveraged AI
If AGI becomes as fundamental as electricity, every major company will be built on it — making broad index ownership as effective a hedge as owning the economy itself.
Data point
20%
Despite concern about privatization of AI returns, well under 20% of the total market cap of non-tiny U.S. companies is currently private, meaning indexing remains feasible.
No indexed bits in this chapter.
Show stoppers
The Pessimistic Reframing of Moore's Law
Moore's Law isn't just about supply — it's about demand evaporation. We create new uses for compute so fast that the value of each unit halves every 18 months. For the first time, AI may have broken this: H100 rental prices are higher now than three years ago despite massive supply expansion.
The Mongolian Economist Thought Experiment
A Mongolian economist in 1400 predicting scarcity would have assumed people satiate in horses and yogurt and concentrate spending on singers. They would have been wrong — because wealth expansion always generates new varieties. The same failure awaits anyone predicting AGI scarcity by holding varieties fixed.
AI as Electricity vs. Social Media: The Stakes
If AGI diffuses like electricity, every company captures it and index ownership works. If it diffuses like social media, platform rents stay concentrated and ordinary people miss the gains. Which model prevails depends largely on whether open models stay competitive and whether frontier labs go public.
Snapshots ()
Key Quotes ()
This episode
Cast
-
Classical economist used as an example of expert forecasting failure — correctly predicted job automation but missed full employment recovery.
-
Cited as a real-world example of an ultra-wealthy person with effectively unsatiating demand for capital accumulation — e.g., mass drivers on the moon.
-
Political scientist cited for finding that a 2% rise in unemployment is sufficient to dramatically shift political outcomes — relevant to AGI transition risks.
-
Used as an example of a billionaire who prefers compounding wealth over personal consumption, suggesting a selection dynamic among the ultra-rich.
-
Referenced for the 'astronomical waste' argument — that vast wealth could be used to create happy simulated beings — as an example of non-satiating utility in capital.
-
Used as a key example of a private AI company whose concentrated ownership raises questions about broad wealth distribution from AGI.
-
Episode sponsor; cited for using targeted RL with textual feedback to train their Composer 2.5 model.
-
Episode sponsor; cited as a model for turning smart people into exceptional researchers through apprenticeship, lectures, and bootcamps.
-
Track
Dutch company cited as a key node in the AI hardware supply chain (EUV lithography machines), used to illustrate which countries are positioned in the AI production chain.
-
Alex Imas's employer; institutional home of the AGI Economics directorate he leads.
-
Track
Used as an example of how ultra-wealthy founders choose to reinvest wealth (building data centers) rather than take consumption income.
-
Cited as a private AI frontier lab whose concentrated private ownership raises wealth distribution concerns, but appears likely to go public.
-
South Korean HBM memory manufacturer cited as an example of a supply-chain-critical AI company that developing countries likely have no ownership stake in.
-
Research group cited for finding little evidence of mass AI-driven unemployment in current labor market data.
-
Used as a representative developing country asking what strategy to adopt given exclusion from the AI hardware and model supply chain.
-
Cited as a key node in the AI hardware supply chain through its semiconductor fabrication industry (TSMC implied), contrasting with countries excluded from AI production.
Stats
This episode
Claims & Sources
7 / 13 cited (54%)
Factual claims made this episode, and whether a source was named.
Prime-age employment in 2026 is the second highest on record, with only the year 2000 being higher.
Andy Atkinson's paper shows that if accounting methods are held constant over time, labor share has never actually fallen.
The O*NET database tracking job tasks has been rarely updated and is of very low quality.
H100 GPU rental prices are higher today than they were 3 years ago, despite significantly improved technology and greater compute supply.
Phil Trammell's pessimistic reframing of Moore's Law: every 18 months the value of a unit of computation halves because new uses for compute are exhausted so rapidly.
The Yale Budget Lab's analysis of the entire economy finds almost no signal of mass AI-driven unemployment, with only a slight below-trend decline in junior developer hiring.
A 2% increase in unemployment is sufficient to completely shift political winds, according to Andy Hall's analysis of the politics of AGI.
Telephone operators were fully automated between 1920 and 1940 — over 20 years — and were reabsorbed into the economy at lower salaries and mostly underemployed.
In Alex Imas's incentive-compatible experiment, people paid significantly more for human-made art prints than AI-made ones, and the human premium collapsed when 500 copies were produced while AI art showed no change.
Chad Jones's research shows that the share of the economy devoted to paying for computing has been decreasing over time, despite massive expansion in the number of transistors.
Well under 20% of the total market cap of non-tiny U.S. companies is currently private, making broad index-based access to AI returns still feasible.
Andrei Fredkin, Brian DeBerrian, and Andrew Koh found in a recent blog post that economists' forecasts about the labor market show massive disagreement in every direction.
Ganz and Goldfarb's O-ring automation model shows that if you can only automate 9 out of 10 parts of a job but at lower quality, you may not want to automate any of it.
