Researcher pushes back on METR's 2x AI acceleration claim, citing 1.4x figure
HaydnBelfield · x · 2026-11-13
AI safety researcher Haydn Belfield argues people may be over-updating on AIFP/METR acceleration findings, questioning three assumptions: the 2x speedup (the chart shows 1.4x, +24/yr vs +17/yr, when other developers' public models are included), the 80% AI-attributable share (could be 50%), and the absence of bottlenecks (research taste may be far harder than coding).
More from AGI Musings
- Armin Ronacher: I started 141 projects in a year, but only maintain 10-15 — TheTuringPost · 2026-11-13
- Daniel Kokotajlo: pretraining loss is a checkable task proxying nearly all downstream skills — nitarshan · 2026-11-13
- Nina Schick: In the 21st century, power is bound to artificial intelligence — NinaDSchick · 2026-11-13
- UW Prof. Abhishek Gupta muses on academia's role in robotics RL research at CSIR workshop — chris_j_paxton · 2026-11-13
- Eli Lilly CEO pushes back on Dario Amodei's claim cancer will be cured within a decade — AlexTensor · 2026-11-13
- AI wealth's 'third wave' of philanthropy: ~$37B a year, best spent catalyzing markets — RishiBommasani · 2026-11-13