Debate: METR's 2x capability extrapolation may have missed the jump-plus-slope model
JacquesThibs · x · 2026-11-08
Jacques Thibs argues METR's AI capability extrapolation likely missed accounting for a one-time capability jump combined with a slope increase, treating the hypotheses as binary. Anthropic's August risk report and April system card both note Mythos preview departed from the prior AECI trend with an unusually large jump, supporting his case. bshlgrs counters that given moderate trust in the researchers, an argument that the evidence is weaker is itself an update that the redacted evidence was stronger. Thibs maintains his bottom line that METR's 2x estimate should shift if the AECI slope drops from 1.9x to 1.4x.
More from AGI Musings
- Gary Marcus calls out Anthropic CEO's AI job-loss claims: 'numbers out of a hat' — GaryMarcus · 2026-11-08
- Australia moves to restrict screens and AI in schools, drawing criticism it will leave kids behind — CtrlAltDwayne · 2026-11-08
- Valuation Guru Damodaran Dismisses 'AI Will Triple GDP' Hype: 40 Years of Data Show No Productivity Gain — GaryMarcus · 2026-11-08
- AI-Generated Resumes Are Flooding the Job Market, Shifting Hiring From Polish to Proof — mdancho84 · 2026-11-08
- Every's Thesis 2027 recap: livestream chat from 15 countries on AI and agency — danshipper · 2026-11-08
- Vitalik backs anti-data-center protests: algorithmic progress makes bigger compute unnecessary — paulnovosad · 2026-11-08