Artificial intelligence in research and development (www.nber.org)

Benjamin Jones’s NBER paper introduces a formal model for how AI can accelerate progress across research fields by isolating three decisive factors: the fraction of research tasks AI can perform, AI’s productivity on those tasks, and the severity of remaining bottlenecks that block progress. The model maps changes in AI capabilities to concrete research outcomes, quantifies the “marginal returns to intelligence” (the gain from incremental increases in AI ability), and shows how shifts in AI capability can reallocate returns to R&D investment. It also uses this framework to make abstract debates about superintelligence, Powerful AI, and Transformative AI more precise and empirically grounded. For the AI/ML community, the paper’s main technical implication is practical: impacts depend less on a single notion of “better AI” and more on where and how capability gains translate into task automation and whether persistent bottlenecks remain. That reframes priorities toward measuring task-level performance, estimating task shares across domains (e.g., cancer therapeutics vs. software design), and identifying bottlenecks that limit downstream gains. The proposed measurement agenda links benchmark development to field-specific opportunity maps, informing research focus, funding allocation, evaluation protocols, and policy discussions about the pace and distribution of scientific acceleration.