Are AI Capabilities Increasing Exponentially? A Competing Hypothesis

Exponential growth claims lack empirical support: The paper challenges the widely-cited METR report's assertion that AI capabilities have grown exponentially since 2019, arguing that the data does not actually support this conclusion—a significant finding given how exponential growth narratives shape policy and safety discussions.

AI capability improvements may be slowing, not accelerating: By fitting sigmoid (S-curve) models to existing data, the authors find that the inflection point—where growth transitions from accelerating to decelerating—has likely already passed, suggesting AI capability gains may plateau sooner than exponential growth models predict.

Current AI forecasts are fragile and model-dependent: The paper demonstrates that different mathematical models applied to the same data produce dramatically different conclusions about future AI development, highlighting that predictions about AI's trajectory should be treated with caution rather than as settled facts.

AI capabilities are not monolithic: The authors propose decomposing AI capabilities into separate "base" and "reasoning" components with different improvement rates, suggesting that blanket statements about "AI capabilities" may oversimplify how different types of AI systems actually develop.

Policy and safety planning should account for forecast uncertainty: Since major decisions about AI regulation, labor policy, and safety investment often rely on capability forecasts, the fragility of these predictions underscores the importance of building flexibility into governance approaches rather than planning based on a single growth trajectory.