Determine whether infinite METR time horizons imply human-level general intelligence

Determine whether an AI system whose METR time horizon tends to infinity would thereby become as intelligent as a human across the full range of cognitive abilities.

Background

The METR time horizon measures how long an AI system can reliably complete certain routine professional computer-use tasks. The paper notes that this measure resembles mean time between failures: divergence to infinity can correspond to perfect reliability on a limited class of tasks rather than unbounded general intelligence.

The unresolved question is whether an infinite score on this benchmark would generalize to broader cognitive abilities, including continual learning, sample efficiency, creativity, and performance on substantially more challenging tasks.