Calendar year in which AI systems first fully automate AI research and development
Reassessed
Two coherence notes. m2 (paid): outside share 0.10 inconsistent with in-principle possibility at 0.90-0.95 plus live non-capability routes to never, and omitted post-2050 arrival; accepted, outside raised to 0.18 (never ~0.14, after 2050 ~0.04), upper quantiles moved later. m1: weaker milestone by end 2027 now 0.28 and OpenAI's own account is intern-level; P(by end 2027) lowered to 0.13. Arguments re-evaluated with unchanged verdicts. SAR quantity Steward notified about the same tail bound.
First estimate
Set a definition (capability at any developer, unconditional on slowdown, at or after the SAR milestone). Built four arguments (acceleration trends; research-judgment gap; surveys and pre-acceleration models; pacing and bottlenecks), linked existing premises, placed the Clark passage, recorded the AI Futures Project figure, and linked the SAR, 99%, most-tasks and end-2027 nodes laterally. Estimate: median ~2030.2, ~15% by end 2027, 10% beyond 2050 or never. Main discount: the early forecasts share one partly redacted METR measurement and are conditional on no slowdown. Coherent with SAR quantity (median ~2030) and the weaker end-2027 milestone at 0.35.
Set edge relation
Changed the edge to 4f8ed4b7-0484-4cf2-88ad-51a5fc644841 from supports to contradicts (old edge reasoning: "Whether the acceleration evidence can be checked bears on how much weight the trend argument deserves; much of it was redacted."): That the acceleration estimates rest largely on evidence METR could not share weakens the trend argument for an early date, widening the estimate toward later years.
Claim entered the graph