228 tasks, 26 models: METR's time horizon curve is nearly flat from 2-30 min
lulzxdxdxd · reddit · 2026-10-11
A Reddit post highlights METR's new paper (arXiv:2610.12466) testing 228 tasks across 26 models: the time horizon curve — how long a task a model can complete — is nearly flat between 2 and 30 minutes. The jump from 3 min to 30 min is far smaller than from 30 min to 5 hours, implying task-length capability grows nonlinearly, with steep gains only past certain thresholds.
More from AGI Musings
- Karpathy: the more AI works autonomously, the more we need AI to review AI — yihui_indie · 2026-10-11
- AI Investment Is Only 1-2% of GDP, Too Small to Crash the Economy, Argues Researcher — kuza55 · 2026-10-11
- Beff Jezos: Rationalists underestimate irreducible complexity of biology — beffjezos · 2026-10-11
- Chris Albon: Kevin Roose's AGI Chronicles makes unread AI books feel like ancient tomes — chrisalbon · 2026-10-11
- The 25-year-old dichotomy: frontier-lab data broking machines vs. grads who can't crack interviews — dhruv2038 · 2026-10-11
- Domingos: Neuralink's killer app is human brains providing compute cycles for AI — pmddomingos · 2026-10-11