Redacted METR report cross-checked with public data: Anthropic's AI-driven 2x acceleration is probably real
ben_j_todd · x · 2026-11-07
Karthik Tadepalli triangulates the heavily redacted METR report from inside Anthropic with public data. Key points:
Findings
- Between Oct 2025 and Oct 2026, Anthropic model capabilities grew 2x faster than in the prior 18 months (internal AECI: 15 points/year to 29 points/year, a 1.9x jump)
- 80% of the acceleration is attributed to AI itself, 20% to non-AI factors — the strongest evidence yet of the recursive self-improvement feedback loop
Three questions answered
- Is the acceleration real? Yes
- Is it because of AI? Probably
- Will it continue? Nobody knows
Ben Todd adds that wherever data exists, less-verifiable tasks are improving just as fast as verifiable ones (1.64x vs 1.67x), weakening the verifiability-bottleneck objection.
More from AGI Musings
- AI central planning revisited: markets still produce demand signals models can't replace — NathanpmYoung · 2026-11-08
- AI safety circle memes shared 'terror' over METR report via Hubinger pfp — danfaggella · 2026-11-07
- Proposal: freeze frontier AI capability, redistribute compute to cheap inference — austinc3301 · 2026-11-07
- Kundaje: Even After Validation, Delivering AAV/CRISPR to Causal Cell Types Remains Unsolved — anshulkundaje · 2026-11-07
- Stanford Prof: Once LLMs Can Act on Noncoding Variants, Validation Becomes the Bottleneck — anshulkundaje · 2026-11-07
- Demis Hassabis: one AI-savvy person can outperform a startup team; 1/10 headcount in 5 years — smtabatabaie · 2026-11-07