Refitting all of METR's data yields ~1.6x AI capability acceleration, above Toby Ord's 1.4x
austinc3301 · x · 2026-11-10
Refitting METR's full 16 frontier-snapshot AECI series: Toby Ord's 1.4x estimate only holds if you stop at Mythos 5, and that fit underpredicts all 5 later models. A jump-plus-slope-change fit on all data gives 1.6x (95% CI 1.2–2.4x). Comparison chart: METR's slope-only fit gives 1.9x; slope change + small jump fits slightly best at 1.7x; jump-only fits worst at 1.0x.
More from AGI Musings
- Frogdie exec: compiling code once needed a datacenter — few saw desktop power coming — draginol · 2026-11-10
- Remote work dying the same year everyone got an agent is 'the dumbest timing' — StewartalsopIII · 2026-11-10
- Duvenaud: fearing 'loss of meaning' from AI job loss is really fearing ostracism — zetalyrae · 2026-11-10
- Redditor argues frontier labs should prove AI power by curing cancer — ProxyLumina · 2026-11-10
- Kai-Fu Lee tells Columbia Business School top AI agents now do a workday in one coffee break — kaifulee · 2026-11-10
- AI safety researcher: window to pace AI may be shorter than expected, urges government action — austinc3301 · 2026-11-10