AI Saturation Shapes Future of Work

University of Pennsylvania computer scientist Konrad Kording and economist Ioana Marinescu released a working paper that reframes AI’s macroeconomic role by splitting production into two sectors: intelligence (AI and algorithmic capabilities) and physical (machines, manual labor). They embed this split into a constant elasticity of substitution (CES) production function to produce a parameter-driven mapping from AI scaling assumptions to wages, employment shares, and output.

Traditional macro models treat AI as a generic form of capital and therefore predict gradual change. Computer‑science scaling laws, in contrast, imply very rapid improvements in intelligence-capital. Kording and Marinescu’s contribution is to reconcile these views by modeling complementarity between intelligence and physical capabilities and explicitly allowing different price and technological trajectories across sectors.

Because intelligence and physical capital are complementary, the model produces diminishing marginal returns (saturation) to intelligence: even if AI capabilities keep improving rapidly, their incremental economic value falls once complementary physical capacity or labor becomes the bottleneck. The authors show that the price of AI capital is falling faster than physical capital, so intelligence tasks get automated earlier; displaced labor reallocates toward the physical sector. Critically, automation’s effect on wages is not monotonic: under many parameter settings wages can rise initially (productivity and output gains dominate) and later decline as reallocation dynamics and shrinking employment shares in the intelligence sector take over. The paper provides an interactive simulation interface to let users vary elasticities, substitution parameters, and automation rates to explore where the turning points occur.

The framework gives ML engineers, economists, and policy teams a concrete, testable way to reason about where automation pressure will land first (task type, sectoral exposure) and what complementary investments matter (physical capital, labor training). For practitioners designing deployment policies, business strategy, or labor-market interventions, the model clarifies that focusing only on software capability growth ignores binding constraints in physical capacity and worker reallocation. The non-monotonic wage result cautions against simple narratives that automation uniformly increases or decreases wages.

parameter estimates for cross-sector elasticities and empirical measures of substitution; evidence on relative price declines for AI vs. physical capital; sectoral employment shares over the next 1–5 years; and policy experiments that alter reallocation frictions (training, mobility, capital investment). These will determine whether the baseline rise-then-fall wage trajectory occurs in practice.