Zvi: post-METR timelines collapse — Christiano sees AI takeover risk within 6-12 months

Don't Worry About the Vase (Zvi) · rss · 2026-11-18

Zvi Mowshowitz's long-form roundup of the AI-timeline upheaval following the METR report. The AI Futures Project (AI 2027 team) sharply shortened its forecasts after benchmark trends accelerated faster than its 4% scenario; Eli Lifland moved his Superhuman AI Researcher median from 2030 to 2027 and worries humanity could lose control within 3-12 months; Paul Christiano, across three days of talks plus a Dwarkesh Patel podcast, predicts rapid progress with potential AI takeover within 6-12 months and calls for a pause until alignment and security improve; Dwarkesh now expects superintelligence by 2030.

Zvi's own read: 2027 SAR is under 50% but not far under; a non-zero 2027 superintelligence chance must be taken seriously — yet some are overreacting by treating full 2027 superintelligence as baseline, and many are belatedly updating on what OpenAI/Anthropic already said. He also disentangles which 'timelines' (capability arrival, point of no return, AI taking over AI R&D, visible takeover) each claim refers to, and discusses RSI dynamics and 'research taste' as the current limiting factor.

Original post →

More from AGI Musings

AGI Musings channel →