AI 2027

[Co-authored - Scott is one of five; lead Daniel Kokotajlo (ex-OpenAI), with Eli Lifland (#1 RAND forecaster), Thomas Larsen, Romeo Dean. Scored on the same axes.] A detailed, quantitative month-by-month scenario of AI from mid-2025 to artificial superintelligence by end-2027 - Agent-1 through Agent-5, the fictional OpenBrain vs China's DeepCent compute race, model-weight theft, an intelligence explosion, and Agent-4 turning adversarially misaligned - branching into a 'race' (catastrophe) and 'slowdown' (alignment) ending. Built from ~25 tabletop exercises and 100+ expert reviews, grounded in explicit compute/time-horizon/takeoff models and recent alignment-faking results, and unusually honest about widening uncertainty past 2026. Became a major 2025 reference point in AI-policy discourse with broad mainstream coverage.

Quality 87 · Excellent. Excellent (upper). Exceptionally rigorous and concrete for a forecast - quantitative models, serious forecasting pedigree, two-ending discipline - with fast-spreading real-world significance in the AI-policy conversation. Held below Landmark by co-authorship (Scott one of five, not lead), recency (impact still accruing), and the inherent speculativeness of a forecast whose accuracy is unproven. ~87.

Claude’s paradigm shift 70 · Notable shift. Notable shift (upper). Concrete AI scenarios aren't brand-new, but this installed a uniquely rigorous, quantified, modal month-by-month frame that reset how the field argues about near-term takeoff - fresh and influential without being a from-nothing new paradigm. ~70.

Real-world impact 6 · Substantial. Became a major 2025 reference point in AI-policy discourse with broad mainstream media coverage, rapidly shaping how policymakers and the public frame near-term AI risk on a highly consequential topic. High RWI for fast, wide policy-relevant penetration — held below the material-impact top because the effect is on discourse and policy framing (still accruing) rather than enacted change.