We Won't Be Missed: Work and Growth in the Era of AGI [pdf] (conference.nber.org)

Pascual Restrepo’s chapter develops a task-based growth model to explore the long-run economic effects of Artificial General Intelligence (AGI). He distinguishes bottleneck tasks (essential for unbounded output growth) from accessory tasks (non‑essential). Under two premises—AGI makes it technically feasible to replicate any human task with some compute cost α(ω) that eventually falls, and aggregate computational resources Q(t) grow without bound—the model predicts that all bottleneck work will be automated, some accessory work may remain human-only, and output transitions from multiplicative human×compute complementarities to an additive form driven by compute. Crucially, long-run growth becomes pinned to the expansion rate of compute, wages converge to the opportunity cost of the compute needed to reproduce human work, and the labor share of GDP tends toward zero as income accrues to compute owners. Technically, Restrepo models output as F({X(ω)}) where work X(ω)=L(ω)+Q(ω)/α(ω), with human labor L and computable work Q constrained by a total compute budget Q(t). AGI is formalized as α(ω) reaching finite (and falling) values; abundant compute lets Q(t) grow large. Extending the framework to science, automating scientific bottlenecks likewise pins technological progress to compute growth—allowing sustained exponential growth without a singularity. The paper highlights distributional stakes: humans retain roles by saving scarce compute or doing inaccessible accessory tasks, but their relative contribution and wages are ultimately bounded by computing costs, raising policy challenges around ownership of compute and the meaning of work.