S9 Ep18: Will AI transform economic growth?

Episode Summary

Executive Summary: Anton Korinek argues AI may trigger explosive economic growth sooner than many economists expect, especially if AI automates enough of AI research itself to create feedback loops in hardware, software, and general innovation. He explains why standard models understate this possibility, why bottlenecks and measurement issues still matter, and why policy must focus on shared prosperity, alignment, and better tracking of the AI economy.

Main Topics: AI’s potential to drive explosive growth (Priority: 5/5): Korinek argues that AI could move beyond modest productivity gains and, if it automates enough research, produce feedback loops that sharply accelerate economic growth. Why standard AI productivity estimates may be too low (Priority: 4/5): He criticizes conservative modeling assumptions like Daron Acemoglu’s 0.07% annual growth estimate, saying they low-ball the share of jobs, tasks, and gains affected by AI. Recursive self-improvement and the intelligence explosion (Priority: 5/5): The discussion centers on the ambition of leading AI labs to reach recursive self-improvement, where AI improves AI and innovation compounds rapidly. Limits and bottlenecks in the economy (Priority: 4/5): Korinek notes that even very capable AI faces constraints from physical labor, robotics lags, and human preference for human-delivered services. Modeling explosive growth with semi-endogenous growth theory (Priority: 5/5): He describes how standard growth models can generate finite-time divergence when feedback from a larger economy outweighs diminishing returns to research. Measuring the AI economy and missing GDP (Priority: 4/5): Korinek says current statistics undercount AI because they miss AI work on future AI and consumer surplus from digital tools that exceed what users pay. Policy, distribution, and transformative AI research (Priority: 5/5): He frames the key issue as who benefits from AI, emphasizing shared prosperity, alignment, and the Economics of Transformative AI initiative at UVA.

Key Arguments: AI’s short-run measured effect may be small, but current capabilities likely capture only a fraction of the eventual economic impact. Conservative estimates in AI growth models can understate effects because they multiply several small assumptions about affected jobs, tasks, and productivity gains. Leading AI labs are pursuing recursive self-improvement, which could create accelerating feedback loops in innovation. Explosive growth becomes possible in semi-endogenous models if the gains from a larger economy and more research resources outweigh diminishing returns. Hardware, software, and general scientific progress mutually reinforce each other, so automating any major part of the innovation stack can speed up the whole economy. Physical and social bottlenecks mean AI cannot instantly transform every sector; some tasks still require labor or human-to-human interaction. National accounts and GDP will miss much of AI’s value unless they account for AI investment in future AI and unpriced consumer surplus. The central policy question is distribution: AI could create broad prosperity, but without governance it could also concentrate benefits among machines and their owners. Economists should be more willing to ask speculative, future-facing questions about transformative AI rather than relying only on historical growth patterns. AI alignment matters because powerful technologies can be used for good or evil, and economics has tools for shaping technology toward social welfare.

Data Points: Estimated annual growth effect of AI: 0.07% per year - Korinek cites Daron Acemoglu’s estimate as a reasonable figure for AI as of 2024, but likely too low for future capabilities. Expected timeline for human-level AI capabilities: within 2 to 5 years - Korinek says it is plausible AI could reach the level needed for recursive self-improvement within this timeframe, maybe sooner or slightly later. Chip capability doubling period: every 2 years - He uses Moore’s Law as an example of sustained technological progress in hardware. Workforce growth needed for chip progress: 8% per year - Korinek says maintaining chip progress has required about 8% annual growth in the scientists working on chips. Publication date of discussed paper: January 2026 - The paper 'When Does Automating AI Research Produce Explosive Growth?' is cited as published in January 2026. Publication date of related paper: March 2020 - The paper 'Scenarios for the Transition to AGI' is cited as published in March 2020.

Pivotal Quotes: "I think at this point, we may have seen just about that much growth impact of AI at the global level, maybe like a tiny bit more, but not much more." — Anton Korinek: On why the recent measurable impact of AI has been small so far, despite rapid progress. "If the second force is larger than the first force, then you can obtain conditions under which growth becomes explosive." — Anton Korinek: Explaining the semi-endogenous growth mechanism that could lead to finite-time divergence. "AI will be everywhere except in the productivity statistics." — Tim Phillips: A summary line highlighting the measurement problem discussed in the interview.

Implications: If Korinek is right, AI could reshape growth theory, statistics, and policy within years, not decades. Governments and firms should prepare for distributional conflict, measurement gaps, and the need to steer AI toward broad prosperity and safe alignment.