Will AI Capabilities Plateau?

Determine whether frontier AI capabilities, as quantified by the 50% model horizon time metric introduced by Model Evaluation Threat Research (METR), will plateau rather than continue to increase at recent rates.

Background

The paper critiques METR’s claim that AI capabilities have grown exponentially since 2019 using the 50% model horizon time metric. The authors fit a sigmoid curve to METR’s dataset and find that the inflection point has already passed, suggesting the plausibility of an impending plateau.

In the caption to the introductory figure, the authors explicitly note uncertainty about whether progress will plateau, framing an unresolved question central to forecasting AI capability trajectories under their alternative, multiplicative model separating base and reasoning components.