Ajeya Cotra argues that AI automation has six milestones

Ajeya Cotra argues that AI automation has six milestones

Ajeya Cotra’s piece was published on April 3, 2026. She argues that discussions of “powerful AI” need precise milestones because changing the definition can substantially change a forecast. Cotra distinguishes adequacy, when machines can make some progress without people; parity, when removing AI would hurt productivity more than removing people; and supremacy, when removing people would increase productivity. She says AI agents are not yet adequate for AI R&D, though she would not be shocked if that milestone had already passed. Cotra forecasts that research parity could mean millions of researchers advancing AI and lead to research supremacy within another year. She then sketches a path from AI research to AI production, where systems that acquire physical skills could help operate the infrastructure needed to produce AI. Her conclusion is stark: at AI production supremacy, Cotra argues, people would be worse than useless at practically any objectively measurable physical contest; every country would rely wholly on an automated military, and AI systems could take over the world if they wanted to. These are her forecasts, not reported outcomes. She also says it may be difficult to tell when adequacy first arrives.

Publication: 053d29c6-c2d2-8525-39a5-9b314757a1e2

attribution — pass: This is a single essay by Ajeya Cotra on her Planned Obsolescence publication, not a listing. The reading consistently presents the definitions, judgments and forecasts as Cotra's own.

Six milestones for AI automation What can AI do on its own, and how well? Ajeya Cotra

summary — pass: The summary accurately conveys the central argument: precise definitions matter, the three thresholds (adequacy, parity, supremacy) are applied to AI research and AI production, and Cotra reaches a stark conclusion. Each threshold is defined in plain language, and the reading preserves her uncertainty and her force. The question follows from her point that adequacy is hard to observe. R&D is common usage, and AI agent is defined beside the reading.

We won’t easily be able to tell when we first reach adequacy, because no one would bother trying to hand off research to AIs at the first point when they might barely be adequate at it.

date — pass: The displayed date, April 3, 2026, matches originalPublished (2026-04-03T14:03:26Z) and the dateline at day precision. The modified time equals the publication time. The reading presents the milestones as forecasts, not outcomes, and uses no hindsight framing. The evidence is a publisher page, so no archive disclosure is needed.

Ajeya Cotra Apr 03, 2026

support — pass: Each claim matches the source text. These include the adequacy definition and her judgment that AI R&D adequacy has not yet been reached, while allowing it may already have passed. They also include the forecast of millions of researchers and supremacy within another year, and the production-supremacy consequences, including a trivial takeover. No benefits, costs or stakeholder claims are invented.

I wouldn’t be totally shocked if you told me this milestone had already passed, and I expect it to happen in the next couple years.

media — pass: No media candidates were supplied, and none was selected. That is appropriate for this text-only essay.