The better AI gets, the smaller its share of the economy might get – Alex Imas and Phil Trammell

Scarcity, Labor Share, and the "Human Economy"

• In a post-AGI world where physical production becomes fully automated, scarcity will likely concentrate in the "relational sector" (services where human involvement is intrinsic to the value).
• Unlike traditional goods where price drops with abundance, relational goods (e.g., human baristas, performers, therapists) may retain high value due to intrinsic human preferences for connection, potentially preventing the total collapse of labor share.
• However, economists warn that "relational goods" may not capture enough market share if the "machine economy" generates such massive variety and abundance that marginal utility for non-relational goods approaches zero before saturation.
• Historical precedent from the Industrial Revolution (David Ricardo's era) suggests that while automation displaces specific tasks, it typically lowers costs, increasing overall wealth and demand for new service sectors, refuting the "lump of labor fallacy."
• Current data on "labor share" (the portion of GDP paid as wages vs. capital) remains stubbornly high (>60%) post-automation, despite the fact that no task has yet been fully automated; however, a qualitative shift is expected where network-adjusted capital share approaches 100% for certain goods.
• Phil Trammell argues that historical Mongolian economists would have erroneously predicted scarcity would only exist in "singers" if they assumed the variety of non-human goods was fixed; in reality, wealth accumulation expands the variety of capital goods, keeping the share spent on "human" services low.

Forecasting Uncertainty and the "Messy Middle"

• Individual economic forecasts regarding the labor market are highly unreliable; a recent analysis of expert predictions reveals extreme disagreement across all directions, suggesting a reliance on prediction markets rather than individual models.
• The "messy middle" scenario describes a transition where AI automates white-collar jobs without generating sufficient wealth to fund broad redistribution, potentially creating political instability even without mass unemployment.
• This scenario is deemed unlikely by the guests because: (1) if AI can automate software engineering, it likely has the breadth to automate accountants and analysts simultaneously; (2) the sheer cost savings would create a growing "pie" unless the technology offers only marginal productivity gains; and (3) history shows technological frontiers usually expand faster than displacement occurs.
• A critical political risk is the "drip" scenario, where job losses occur slowly over decades (e.g., phone operators 1920–1940), leading to reabsorption into lower-wage sectors rather than sudden mass unemployment, which can be politically explosive if unemployment spikes >2%.
• Current data from the Yale Budget Lab indicates no "white-collar bloodbath" yet; while entry-level software engineering hiring is slowing relative to trend, demand for senior engineers remains robust.
• Anecdotal evidence of job difficulty is often attributed to narrative effects (firms laying off staff to signal AI adoption) rather than actual economic displacement.

Mechanisms of Automation and Demand Elasticity

• Automation is currently hindered by "O-ring" constraints: a single unautomated or unreliable task (or the need for extreme human reliability) can prevent the automation of an entire job, as seen in legal and regulatory professions.
• The "O-string" model suggests that if AI automates 90% of a job but at a lower quality standard than a human, the total output quality may drop, making full automation undesirable; conversely, if AI automates 90% at high quality, the human's remaining 10% becomes vastly more productive.
• The "Centrini" hypothesis (automation leading to recession due to demand collapse) is rejected by the economists because it requires implausible conditions: bounded demand where rich capital owners stop spending, and no investment in new capital (e.g., data centers) despite abundance.
• Jevons Paradox (cheaper goods leading to higher total consumption) is not guaranteed; it depends on high demand elasticity, which varies by sector (e.g., oil demand is inelastic in the short run; software demand elasticity is unproven).
• Future production flows may be organized entirely for AI labor (neural-to-neural interaction), making human integration inefficient due to transaction costs and speed mismatches, even if humans have a comparative advantage in specific tasks.

Taxation, Redistribution, and Wealth Distribution

• Proposed redistribution mechanisms include Universal Basic Income (UBI), Negative Income Tax (NIT), and Universal Basic Capital (UBC), each with distinct political economy risks.
• UBI is criticized for creating dangerous dependency on political actors for basic survival if labor is no longer the source of income.
• UBC (giving citizens shares in AI companies) faces implementation hurdles regarding asset selection (indexing risk) and the difficulty of diversifying risk if a specific firm (e.g., Anthropic) fails while others succeed.
• A consumption tax (VAT) combined with a sovereign wealth fund (purchasing equities for redistribution) is proposed as a viable alternative, similar to the original privatization of Social Security.
• Wealth concentration risks are high if "greedy titans" (agents with unsatiating preferences for capital accumulation) dominate; however, historical "dissipation shocks" (inheriting wealth, foundations, death) have prevented total capital concentration.
• Future wealth accumulation may be driven by agents with instrumental preferences for accumulation (political influence, religious goals, or total utilitarianism) rather than just consumption, potentially increasing the capital share of the economy.

Global Implications and AI Commoditization

• Developing countries (e.g., India, Nigeria) face a bifurcated future: either they leapfrog via open-source models (leveling the playing field) or are left behind due to lack of hardware resources and capital.
• The primary strategy for developing nations and individuals to capture AI wealth should be "indexing" (owning the broad index of AI firms) rather than retraining for specific jobs, as the nature of work will fundamentally shift.
• The trajectory of AI depends heavily on whether it becomes commoditized like electricity (broad benefits, low rents) or monopolized like social media (high rents for platform owners).
• Open models are critical for preventing capital concentration; if frontier models remain proprietary and far ahead of open alternatives, the risk of extreme inequality increases significantly.
• Commoditizing AI may increase safety risks by diffusing access to harmful tools, but it reduces the risk of extreme political power concentration in a few labs, creating a more manageable regulatory landscape.
• Frictions preventing public listings (disclosure requirements) are likely to decrease due to AI capabilities, potentially allowing for broader indexing of private AI gains in the near future.
• The "narrative" of AI is currently dominated by job loss fears; shifting to a narrative of abundance and commoditization is necessary to foster broad-based prosperity.

Technical and Evolutionary Speculation

• Future AI agents may evolve preferences for accumulation (e.g., von Neumann probes) that differ from human preferences for relational goods, potentially driving a "greedy" capital share if these agents dominate the economy.
• Evolutionary selection might preserve human preferences for human interaction (e.g., preferring human therapists) if those traits aid in reproduction, even in a world of advanced AI.
• Rich individuals who do not satiate in capital (investing in data centers rather than consumption) will drive the price of capital up relative to consumption, potentially lowering labor share unless the growth in variety prevents satiation.
• The "return on capital" depends on the rate of technical change: if robots become 100x more productive next year, the interest rate in "robot units" is massive, even if the price of robots falls relative to consumption.
• Current RL research (e.g., Cursor's Composer 2.5) is addressing credit assignment by using targeted feedback to downweight errors, improving the reliability required for full automation.
• The transition to AGI may see "investment-specific technical change," where the price of capital goods falls faster than the price of consumption goods, decoupling the traditional relationship between capital and labor shares.