Toby Ord revisits METR capability trends: full-data fit yields 1.6x acceleration, not 1.4x
tobyordoxford · x · 2026-11-10
Toby Ord and @austinc3301 debate curve-fitting methodology in METR's model capability trends chart.
- Ord notes METR fits curves only on the white-area data, holding out the last 5 models (grey area) for validation — which is why the jump-only fit underpredicts those points
- austinc3301 replicates the 1.4x result but shows it underpredicts all 5 later models, suggesting the slope is too low; using all data, the jump + slope-change hypothesis gives 1.6x (95% CI: 1.2–2.4x)
- Ord defers to the 1.6x estimate from all 8 data points, while noting substantial uncertainty remains about where the data points should even sit
Related event: Debate over METR capability acceleration: full-data fit yields 1.6x(3 posts)→
More from Models
- GDPx leaderboard shows Fable 5.2's RSI plateauing at Sonnet 5.5 level, matching Anthropic's training caps — morgymcg · 2026-11-10
- Anthropic capped Fable 5.2's AI-research capabilities in training, system card reveals — EntropyMagnets · 2026-11-10
- Running open-weight models locally: Kimi K3.1 needs 594GB, ~$38k for 15 tok/s — ravikantagrawal · 2026-11-10
- Consultant: banks larping 'sovereignty' with on-prem OSS models; bullish on Chinese providers — teodorio · 2026-11-10
- The most complete list of LLM writing tics yet: negation-first reveals, aphoristic closers — gleech · 2026-11-10
- Amazon swaps Sonnet 4.6 for a decision model in ticket labeling, cutting cost by 98% — victormustar · 2026-11-10