The math behind AI benchmarks may quietly determine who gets left behind as compute budgets diverge.
Researchers building on prior scaling work have found that two broad classes of performance metrics point toward opposite futures. Bounded metrics — those with a natural ceiling, like accuracy on a fixed test — tend to show the gap between small and large models shrinking over time. Unbounded metrics — measuring things like raw software engineering output or rhetorical persuasion — show frontier models pulling further ahead, indefinitely. The paper offers formal mathematical conditions for sorting any given metric into one camp or the other, and notes that many common benchmarks have near-twin versions in the opposing category, making the choice of metric less innocent than it looks.
The practical stakes are significant. If a capability like synthetic biology reasoning or persuasive writing is genuinely unbounded in the terms that matter, then frontier-level performance stays locked inside a small number of well-funded labs and the governments or corporations that can afford them. If the same capability turns out to be bounded, cheaper models catch up and access spreads. The same underlying technology, measured differently, implies opposite policy responses — from antitrust intervention to export controls to open-source mandates.
This is the kind of paper that should make policymakers uncomfortable, because it means the benchmarks funding agencies and regulators currently lean on may be encoding assumptions about concentration of power without anyone noticing.