The better AI gets, the smaller its share of the economy might get – Alex Imas and Phil Trammell
In a Nutshell
As AI automates non-relational production, labor share could collapse toward zero if demand for new capital varieties outpaces saturation in human-intrinsic services like empathy and performance. Historical stability of the ~60% labor share may not persist once entire supply chains become fully automated, and redistribution via negative income taxes or broad capital indexing faces targeting, political, and timing challenges. Developing countries benefit more from early indexing into AI-driven assets than from domestic retraining, while gradual displacement creates a politically fragile "messy middle" even if aggregate wealth eventually expands.
These notes were generated by AI and may contain inaccuracies.
Alex Imas is Director of AGI Economics at Google DeepMind and Professor of Economics at the University of Chicago. Phil Trammell is Head of Economics at Epoch and research scholar at Stanford. The discussion focuses on what economics predicts about wages, labor share, taxation and redistribution of AGI-generated wealth, and what will remain scarce in a highly automated world.
Scarcity determines where value accrues. The relational sector includes services and goods where human involvement is intrinsically part of the value. Even if automation eliminates scarcity elsewhere, scarcity persists in activities requiring human participation. In a world where AI and robotics handle all physical production, humans would have no reason to participate in the machine economy. However, humans may still value human involvement in services like performances or cafe experiences. This creates a human economy where humans provide services to each other, though some wealth flows out to purchase automated goods from the machine economy.
The human economy is not a closed loop. Machines do not demand human services, so the human-only economy's share may shrink. Individual economic forecasts are unreliable. A blog post by Andrey Fradkin, Brian Jabarian, and Andrew Koh found substantial disagreement among economists' labor market predictions. Prediction markets may better aggregate forecasts and capture crowd wisdom. Economic forecasting has historically been poor.
This debate dates back 200 years to David Ricardo during the Industrial Revolution. Ricardo initially predicted automation would lower prices and benefit everyone, then reversed course, predicting mass unemployment and political unrest. Ricardo correctly identified that jobs existing in his time would be automated, but he missed the structural change effects. Automated tasks became cheap, freeing income for new spending on services. This illustrates the lump-of-labor fallacy—the incorrect assumption that automation reduces total work available.
Rather than making point predictions, models should examine boundary conditions. Models can explore scenarios where labor share reaches zero or remains constant. Current data limitations are severe: consumer demand elasticities are unknown, job creation and destruction tracking is inadequate, and the O*NET database is outdated and low quality. A "Manhattan Project for data" is needed. The useful approach is mapping potential scenarios by identifying which dimensions of scarcity generate each outcome.
Labor share is the portion of total goods and services paid as wages. Capital share includes rents on buildings, shareholder returns, and payments to owners of machines and land. For centuries, wages have accounted for approximately 60% of the economy, with 30-40% going to capital owners. A Kaldor fact is that labor share remained remarkably stable near 60% despite centuries of automation. Some debate whether recent declines are real or accounting artifacts. Atkinson has shown that holding accounting methods constant eliminates apparent labor share decline.
When labor and capital are complements, both must be paid. Even automated processes retain labor contributions throughout the supply chain. Computer and electronic products show a stable network-adjusted capital share around 50%. A qualitative shift occurs when entire supply chains become fully automated with no intrinsic human component, pushing network-adjusted capital share to 1 for some goods.
Two sectors can be distinguished: the human-intrinsic sector and the automated sector. Full automation of non-relational supply chains could drive quantities to infinity while marginal utility falls faster, potentially affecting overall capital share. The ballerina example is limited. Most jobs involve multiple tasks, and consumers may pay premiums only for specific human tasks within otherwise automated processes.
The task-based model views jobs as bundles of tasks. A doctor performs insurance paperwork, pharmaceutical coordination, and patient interaction. If consumers pay more when a human delivers diagnosis and support, that component qualifies as relational. Data on willingness-to-pay differences between fully automated and partially human services is lacking. Conjoint analysis could measure elasticities for human involvement.
Fully automated goods that do not yet exist cannot be studied for demand. New varieties of capital goods may prevent satiation in the automated sector. If variety expands rapidly in capital while remaining static in the human sector, labor share can fall to zero even if relational goods are available. Phil Trammell draws an analogy to a Mongolian economist in 1400 predicting that automation would eliminate spending on transportation and food, leaving only spending on singers. Instead, new varieties expanded non-singer consumption while singer spending remained negligible.
The number of transistors has increased by trillions or quadrillions, yet the share of the economy spent on computing has declined. Chad Jones documented this result. Moore's law can be framed pessimistically: every 18 months, the value of computation halves because uses are exhausted faster than supply grows. AI may break this pattern. An H100 now costs more to rent than three years ago despite technological improvement, because smarter models raise the opportunity cost of compute. New varieties of compute demand can reverse share declines.
Economic models often treat demand as exogenous. Research on the relational sector reveals intrinsic preferences for empathy and human interaction, not merely scarcity effects. An art print experiment found people pay significantly more for human-made prints when scarcity signals connection to an artist. When 500 copies exist, the human premium disappears. AI prints show no scarcity premium. Unlike horses, which are pure inputs, humans may add value through the output itself when consumers prefer human involvement.
Jane Street pairs new hires with senior mentors in an apprenticeship model and runs classroom lectures and bootcamps. Courses cover reverse engineering with strace and gdb, profiling to the cache hierarchy, and applying neural networks to trading with specific workarounds. Training is treated as core work with dedicated classroom space in every office.
Molly Kinder described a "Messy Middle" where automation displaces workers without generating sufficient wealth for Pareto-improving redistribution. A trivial accounting view holds that saved labor costs remain in the economy. However, allocative inefficiencies, government information problems, and political constraints on redistribution may prevent smooth adjustment. A Meta worker earning $200,000 may not receive equivalent ongoing payments when many earn far less.
A 2% unemployment increase can shift political dynamics substantially. The "drip" scenario—gradual displacement into lower-paying sectors—is politically problematic. Phone operators were automated over 20 years despite available technology, becoming underemployed at lower wages. COVID demonstrated rapid fiscal response to sudden unemployment spikes. Gradual displacement may not trigger equivalent support. If automation merely reallocates existing resources without expanding the technological frontier, redistribution capacity remains limited.
Multiple conditions must hold simultaneously for a low-wealth-growth displacement scenario: automation must proceed piecemeal across white-collar jobs rather than broadly; intelligence sufficient to replace software engineers must not generalize; and cost savings must be marginal rather than substantial. Historical patterns suggest technological frontiers expand with automation. Intelligence breadth makes narrow, job-specific automation without broader capability unlikely.
Redistribution tools differ in implementation complexity and speed of effect. Universal basic capital generates no immediate returns for near-term displacement. A negative income tax provides immediate insurance with an earnings floor that phases out gradually. UBI raises political economy concerns: dependence on elected officials for basic needs creates power imbalances. Universal basic capital provides property rights in capital rather than ongoing transfers.
Universal basic capital faces a fundamental targeting problem: determining what assets to put into people's portfolios. If indexing is difficult, universal basic capital becomes difficult because it's unclear what to target. For example, if Anthropic goes to zero but some random robotics company takes over, this illustrates the core risk of universal basic capital.
A negative income tax shares the same vulnerabilities as UBI: a new administration could eliminate the program, removing the floor for people who can no longer work.
There is no politically sustainable equilibrium at a 0.5% wealth tax. The income tax began low for war funding and gradually escalated to marginal rates around 40% in the US, with certain states exceeding 50%. A capital tax raises concerns about distorting investment—people might ask why they would invest in Anthropic or Intel if the government will take larger shares and dilute ownership.
Revenue raising, taxation, and distribution should be separated. One approach is for the government to use broad-based taxes to purchase shares of companies like Anthropic and distribute them widely, rather than expropriating particular well-known companies through populist measures.
Optimal taxes could target externalities or land. A consumption tax like the European value-added tax would allow government to purchase stocks and distribute them broadly. David Autor has discussed this approach.
Privatizing Social Security was essentially giving everyone a basket of stocks. This has functioned but raises questions about long-term sustainability.
The Budget Lab at Yale released a report showing little evidence of mass automation or unemployment from AI. Even examining the most exposed sectors like software engineering, there is minimal observable impact. There may be a slight signal that junior developers are getting jobs at lower rates than before, but this reflects slower growth rather than a level shift. Demand for senior software engineers has actually increased.
Anecdotal reports of graduating CS students struggling to find jobs and AI-related layoffs may reflect normal labor market frictions being reframed through an AI narrative rather than genuine structural change.
Public coordination devices create concerning dynamics: if firms that don't lay people off are perceived as failing to adapt to AI, this can trigger cascade effects where companies lay off workers to maintain the appearance of technological adaptation, even when the layoffs may leave the firm worse off.
If AI automates nine out of ten tasks in a job, the remaining task becomes more valuable. If demand is elastic enough that cheaper products lead to increased purchases, this can result in more hiring rather than less. The elasticity of demand is critical to understanding automation outcomes.
Jevons paradox occurs when something becomes cheaper and demand is sufficiently elastic that total spending increases. This happened with coal in Britain approximately 200 years ago. However, this only occurs when demand is highly elastic. Many goods lack super-elastic demand—cheap oil doesn't automatically create dramatically more cars, and cheap insulin doesn't generate proportionally more demand.
Long-run elasticity exceeds short-run elasticity. Agriculture demonstrates that even with massive productivity gains, food demand has natural limits.
Software may be distinctive because as it becomes cheaper, demand increases substantially rather than saturating.
Citrini's scenario planning predicts recession from white-collar automation: automated workers lose salaries, causing demand collapse. The unemployment component may be plausible if automation speed is rapid. However, negative economic growth requires specific, improbable conditions.
Negative growth would require capital holders to have bounded demand—they must eventually refuse to spend more, and that money must not enter as investment. Even in a singularity scenario where people don't want more consumption goods, investment in data centers and fabs to run AGI would continue driving growth.
The Citrini essay usefully illustrates that demand collapse could shrink the economy, but abundance generating negative growth is extremely difficult when the technological frontier is expanding.
Google's Gemini Omni demonstrates advanced video editing: users can upload video and instruct changes to backgrounds, lighting, or elements while maintaining consistency. The model accepts text, audio, or video inputs and is architecturally capable of outputting images or text.
This reflects the multimodal data transfer hypothesis: models improve at predicting one data type by training on others. Omni renders text in video accurately without specific targeting of that capability. Video models require deep understanding of physics and spatial dynamics to predict subsequent frames.
Good video models could help close the Sim2Real gap in robotics, where simulation data is easier to collect than real-world data.
O-ring theory, referencing the Challenger shuttle failure from a single malfunctioning component, suggests goods production requires extreme reliability. Even if AI can perform tasks probabilistically, current reliability levels prevent full job automation because a single failure destroys the finished good.
However, advanced AI could reverse this dynamic. Production flows organized entirely around AI labor—operating in neuralese at thousands of times human speed—would make integrating humans costly due to transaction costs and reliability concerns.
If AI automates nine-tenths of a job but at lower quality than humans, firms may decline to automate even those nine-tenths. Symmetrically, humans may become unemployable for the remaining tenth because they cannot match AI quality or speed on other components, pulling down overall output quality.
This explains why lawyers, accountants, and software engineers resist automation. Clients pay lawyers not just for probable correctness but for assurance that companies won't fail, plus regulatory compliance and ownership structures. Licensing requirements and regulatory layers maintain human involvement independent of capability.
Political decisions—legislation, judging, jury service, professional licensing—currently trust only humans. These expectations have changed throughout history from hunter-gatherer bands to empires. More efficient AI-run political systems would likely outcompete alternatives.
Evolution selected human preferences because they promoted survival; these preferences now shape a hundred-trillion-dollar economy. Similar dynamics may apply to AIs. Even without catastrophic misalignment, evolution among firms using AIs will favor agents that grow. Selection pressures will favor firms or agents that accumulate resources, likely preferring compute over human-intrinsic goods.
An AI with autonomous welfare has no inherent reason to prefer human interaction. However, human preferences for interacting with other humans may persist through evolutionary mechanisms. People with moral emotions against offloading social interactions to AI may reproduce more successfully than those indifferent to human interaction.
Wealth of the richest people exists primarily as equity rather than consumption. Mark Zuckerberg could convert Meta stock to dividends for personal consumption but instead prefers compounding through data center construction. This suggests selection for agents with unsatisfiable demand for capital accumulation.
If one group of people satiates in capital while another continues accumulating due to exploration ambitions, the latter group will have higher savings rates and eventually hold most wealth. Their capital share preferences will dominate overall production.
Elon Musk's investments in mass drivers demonstrate that wealthy individuals may not distinguish between human and AI researchers when allocating capital.
If returns to data centers remain high, investment incentives persist. If capital satiation occurs, returns fall and consumption rises. High labor share scenarios require lower returns to capital, which requires lower growth rates.
Capital stock could grow rapidly while prices of capital goods fall faster than stock grows. High growth environments offer high-return investments or compelling consumption options. Human-intrinsic goods may command higher relative value if robot marginal utility falls.
If every robot can turn into 100 robots next year, the interest rate in units of robots could be extremely high, but prices adjust. The key dynamic is investment-specific technical change where the price of capital falls relative to the price of consumption. Standard macro models treat output as a single chimera that can be allocated one-for-one to capital or consumption, but that will not hold when capital goods and consumption goods diverge sharply. One robot now turns into many robots next year, yet the number of ballerinas remains fixed.
If the extra robots represent new varieties and demand for them does not saturate, the marginal value of capital stays high. On the investment side, a single greedy titan of industry who keeps demanding more robots could sustain high marginal value of capital and therefore reduce labor share. Historically, wealthy individuals have often built libraries or foundations rather than endlessly accumulating, partly because they die.
New types of agents could emerge whose sole goal is wealth accumulation. Historically, wealth accumulation has been embedded in social status and peer recognition. Many titans accumulate capital and then spend it; near the end of life they often pass wealth to children who are poorer stewards or place it in foundations that spend it down. These dissipation mechanisms have prevented any single lineage from dominating the economy. If people live longer or create trusts aligned with perpetual accumulation, a few agents with this preference could dominate because the capital they control grows faster than the rest of the economy.
An intrinsic preference for accumulation would validate the scenario in which capital owners capture most gains. Most preferences, however, involve hedonics followed by social status motives, as noted by Rousseau and St. Augustine. Even without intrinsic accumulation preferences, instrumental motives exist: desire for political, philosophical, or religious influence, and total utilitarian philanthropy aimed at creating new happy beings. Bostrom's astronomical waste argument illustrates how vast wealth could be used to simulate enormous numbers of happy minds via Dyson spheres and similar projects.
A pure von Neumann probe represents an extremely greedy optimizer that values additional solar systems because each one produces more probes. Under current GDP accounting, only final consumption and investment goods are counted; self-replicating probes optimizing between more probes and other goods would not register in conventional labor share statistics. How AI beings and autonomous replicators are treated in national accounts will determine measured labor share.
A major current challenge in reinforcement learning is credit assignment over extremely long rollouts. Cursor researcher Sasha Rush described targeted RL with textual feedback for training Composer 2.5: the system generates a token sequence, a model identifies problematic turns, hint tokens are injected via text manipulation, a second forward pass is run, and the original model is trained to match the adjusted probabilities that downweight the identified errors.
Countries not producing AI models, advanced chips, HBM (Korea), leading-edge fabs (Taiwan), or EUV lithography equipment (Netherlands) face two divergent paths. In one, AI capabilities diffuse and level the playing field. In the other, automation allows rich countries to produce commodities domestically, eliminating export markets for developing nations and leaving them behind. The economics profession has allocated too little attention to middle-income developing countries in the AI era.
If interest rates are very high because capital is extremely productive, even modest savings in developing countries can generate large future consumption. The messy middle becomes narrower because the price of capital-produced goods falls rapidly. Developing countries can benefit from savings placed in developed-world assets without needing direct redistribution. However, they start from a lower base and have less existing indexing to the global economy, making early action important.
Before index funds, it was extremely difficult to own a broad slice of economic growth; missing the few companies that later created most value left wealth stagnant. The brief period when broad indexing was straightforward may be ending as returns concentrate in private companies. Most household wealth, especially housing, is poorly positioned to capture gains from AI or robotics. A Georgist land tax would therefore raise limited revenue relative to future needs.
AI could resemble electricity, where downstream benefits accrue widely to users rather than concentrating in the producing firms, or social media, where rents concentrate at the platform level. If AI transforms every company the way electricity did, owning the broad index again captures the gains. If returns remain concentrated in a few private frontier labs, indexing becomes much harder. Nigeria, for example, does not own significant stakes in SK Hynix or Anthropic, and the S&P 500 alone may not suffice.
Whether open models remain only months behind the frontier determines how quickly developing countries gain access. This technical question connects directly to runaway gains, recursive self-improvement, continual learning, and online learning. Faster diffusion via open models makes indexing a viable strategy for countries such as Nigeria or Uganda.
For countries outside the AI production chain, indexing the AGI economy is likely cleaner and faster than retraining programs or building domestic data centers. In long-timeline or messy-middle scenarios, however, retraining to use the latest computing tools retains value. Countries with weak education systems may find it difficult to become world leaders in AI-related retraining. Leapfrogging remains possible, as seen with mobile banking in Nigeria.
Well under 20 % of the total market cap of non-tiny U.S. companies is currently private. Even leading labs are expected to go public eventually. AI itself may reduce disclosure frictions that currently keep firms private. The long-term trend toward easier indexing is therefore likely to resume despite the recent increase in privatization.
Widespread access to frontier models increases the chance that AI produces broad prosperity similar to electrification. However, commoditization also diffuses capabilities that could be misused. Concentrated labs create clearer political targets, as illustrated by the Defense Production Act threat against Anthropic. A world with a modest capability gap but widely distributed public ownership may reduce both safety and distributional trade-offs. Narratives matter: negative stories about job loss spread more easily than positive stories about nonexistent future abundance.
Keep Dwarkesh Patel in your library
Save the videos and channels worth coming back to, and find them again in one place.
