Analysis Suggests AI Inference Scaling Lags Behind Human Time Scaling

The analogy that giving an AI more tokens is equivalent to giving a human more time may be flawed. Analysis of recent benchmarks suggests that AI models scale sub-linearly with increased inference compute compared to how humans scale with additional time. While humans gain performance linearly or super-linearly with time, current AI systems require significantly more resources to achieve equivalent improvements.

Scaling Efficiency Gap

Data from the UK AISI and METR reveals that AI models exhibit diminishing marginal returns as token budgets increase. On a log-log plot, the relationship between token count and the 80% time horizon (the task duration an AI can complete 80% of the time) has a slope γ less than 1. This indicates sub-linear scaling.

• Early Models: Required approximately 10,000x more tokens to achieve a 10x improvement in time horizon.
• Recent Models (up to May 2026): The requirement has dropped to roughly 20x more tokens for the same 10x improvement.
• Trend: The parameter γ has been increasing over time, suggesting AI systems are becoming more efficient at utilizing additional compute, potentially approaching or exceeding human scaling efficiency in the future.

Implications for Recursive Self-Improvement

This scaling discrepancy has critical implications for theories of recursive self-improvement (RSI). Researchers model RSI using a parameter r, which depends on how well research scales with parallel labor (λ) and diminishing returns to research (β).

• If AI systems scale worse with time than humans (γ < 1), the effective feedback loop for intelligence explosion is weaker than previously assumed.
• Adjusting the RSI model to account for this yields r = λγβ. When γ < 1, this makes an intelligence explosion harder to trigger.
• Conversely, if future models achieve γ > 1, the risk of runaway self-improvement could increase significantly.

Limitations and Future Tracking

The analysis relies on limited datasets, primarily from AISI’s capture-the-flag hacking tasks and METR’s earlier benchmarks. Key limitations include:

• Task Specificity: Results may not generalize across all domains.
• Context Length: Data points involve very long chains of thought requiring multiple compaction rounds, which may affect scaling slopes.
• Measurement Uncertainty: Human-equivalent time horizons are complex constructs; direct human performance data would provide clearer comparisons.

Despite these caveats, the evidence suggests that tracking γ is essential for accurately projecting AI capabilities and safety risks.