Work and Growth in the AGI World [pdf] (www.nber.org)

Pascual Restrepo’s chapter analyzes the long-run economic effects of Artificial General Intelligence (AGI) using a task-based model where every type of productive work ω ∈ Ω can be done either by human labor or by compute. Two core premises drive the results: AGI makes the computational cost to emulate human tasks, α_t(ω), finite (so all tasks are in principle automatable), and aggregate computational resources Q_t grow without bound. He distinguishes bottleneck work (tasks essential for growth) from supplementary work (non-essential). The main theoretical results: as compute expands, all bottleneck work becomes automated by AI, production shifts from a multiplicative compute×labor technology to an additive one, and long-run growth is determined by the growth rate of compute. Wages do not collapse to zero but converge to the opportunity cost of the compute required to reproduce human work; the labor share of GDP falls toward zero and most income accrues to compute owners. Technically, the model formalizes work completion X_t(ω)=L_t(ω)+Q_t(ω), with α_t(ω) measuring compute per human-equivalent task and Q_t the economy’s total flops. It extends to science: automating scientific bottlenecks makes technological progress compute-driven, permitting sustained exponential growth even with shrinking population but not an infinite singularity. Implications for the AI/ML community include predictions about where automation value concentrates (compute and algorithms), how compute scaling changes economic incentives, and why reducing α (algorithmic efficiency) or expanding Q remains central to shaping labor outcomes and long-run growth.