It argues that available evidence points to sublinear growth in AI task time horizon as token budgets increase, while human performance appears to scale more strongly with time.
Ord discusses a METR comparison with early models and a UK AI Security Institute chart based on capture-the-flag hacking tasks. In the latter, the estimated token increase needed for a tenfold improvement in time horizon fell from roughly 10,000× for earlier models to about 20× for models through May 2026. The essay treats this as evidence of improvement, not proof of parity: the datasets are limited, comparisons depend on task and evaluation design, and the trend may plateau.