Toby Ord sanity-checks METR time horizons: OpenAI's number may run 14 months behind expectations
tobyordoxford · x · 2026-10-22
Oxford philosopher Toby Ord sanity-checks METR's time horizon graphs. He notes that if real-world AI R&D were as hard as METR's task suite, OpenAI's reported number would suggest we are about 14 months behind where one might expect.
He acknowledges a mundane explanation: real AI R&D has a harder task mix than METR's suite. METR's chart of time horizons across suites shows OpenAI's number roughly in line with the red 'RLBench' trend line, so it may not imply overstated capability. He posted to invite community verification.
More from AGI Musings
- Founder laments replying to bots all day as AI agents flood the internet — bindureddy · 2026-10-22
- Gauss's Seal Read 'Few, but Ripe' — a Motto Worth Remembering When Quantity Gets Cheap — littmath · 2026-10-22
- Follow-up: Google UK's non-disparagement clause lacks protected-disclosure exemptions — BlackHC · 2026-10-22
- Bentham's Newsletter argues Nature's case against AI consciousness fails — dioscuri · 2026-10-22
- Martin Casado: Cyber Had Become Checklists and Compliance, AI Brings Real Combat Back — Kyrannio · 2026-10-22
- Interactivity, not autonomy, will define the most iconic agentic products — _AustinCalvert_ · 2026-10-22