Listen: https://serve.buildtall.systems/5f47f4bfb96569f62cfb1596515da7c585f273b217e2071cdaa440739251cee2.mp3
This episode of Marginal Revolution Read Out covers the post AI and complements, published November 15, 2026. The post examines a new working paper on how AI's economic promise and peril can be extremely asymmetric when production tasks are complements.
The first article is AI and complements. The author highlights a recent in-progress manuscript by Charles I. Jones and Christopher Tonetti, titled The Weak-Link Asymmetry: AI's Peril and Promise. The core idea is a weak links model of production, where tasks are complements. In such a model, output is constrained by the least productive tasks, so the gains from automation tend to arrive slowly. But the same complementarity makes attacks more harmful and easier to execute. The paper gives a striking example: with an elasticity of substitution across tasks of 0.2, automating tasks worth 1 percent of GDP with a tenfold increase in productivity raises output by less than one percent. Yet cutting their productivity tenfold lowers output by 68 percent on impact. The author's take is that when a larger AI sector simultaneously makes AI models more capable and attacks more likely, the positive productivity spillover is negligible compared with the potential damage, so the market overinvests in AI. In a weak-link economy, the costs and benefits from AI can be extremely asymmetric.
The second article is Lukas Freund. This is the source the author cites for the manuscript, a post on X by Lukas Freund. Freund describes the paper as rather topical and explains the weak-link logic in similar terms. He notes that if tasks are gross complements, sometimes called weak links or Kremer's O-ring, not only do the gains from automation arrive slowly, but the economy is also very vulnerable to shocks at the task level. He repeats the same numerical example: with an elasticity of substitution of 0.2, boosting productivity tenfold in tasks worth 1 percent of GDP raises output by less than 1 percent, whereas cutting their productivity tenfold drops output by 68 percent on impact. Freund applies this to AI, motivated by the risk of AI-enabled attacks on the financial system, electricity grid, and other critical infrastructure. He adds that the FAA ground stop is a pretty vivid illustration of the underlying fragility. The author's commentary simply notes that the manuscript came via Lukas Freund, but the post itself adds useful context by connecting the abstract model to concrete infrastructure risks and a recent real-world disruption.
Taken together, the two links point to a common theme: AI's upside may be slow and diffuse, while its downside could be sudden and severe. The weak-link framing suggests that standard productivity estimates may overstate the benefits of AI and understate the risks, especially as AI becomes more capable of enabling targeted attacks on critical systems.
