Will METR state by 31 March 2027 that an AI model's 50% time horizon is at least 24 hours?

Критерии

YES if between 22 Oct 2026 and 31 Mar 2027 an official METR publication (blog post, report, paper or model evaluation on metr.org or arXiv) states that some AI model's measured 50% task-completion time horizon is at least 24 hours: a point estimate of 24 h or more, or a stated lower bound of 24 h or more. Predicted or extrapolated future horizons do not count; third-party recomputations do not count. Otherwise NO.

Вероятность во времени

Метод

base: ~6 METR time-horizon publications in 19 months -> P(at least one in 160 d) = 0.81, times 0.5 that one states a frontier number at or above the ceiling. signals: LR 0.7 the horizon page is 'no longer actively updated' (08.09) and TH 1.1 is unreliable above 16 h; LR 1.3 the Frontier Risk Report already gave a 16-20 h point estimate despite the ceiling. trend: 12 h (Feb-Mar 2026) doubling every 129 d gives 100 h by the deadline, so capability is not the constraint; 0.75 that a suite exists to say so. Ensemble = mean in log-odds. second: independent probability from a second model (Claude Fable 5.1) given the evidence and criteria, never my number.

Модели объединяются как среднее в логарифмах шансов; см. Как мы прогнозируем.

Все обновления