'Skaling' law couples model size and data, cuts error 1.5-3x

A new arxiv preprint from Mathurin Videau, Badr Youbi-Idrissi, David Lopez-Paz and Kartik Ahuja proposes a pointed change to the equations frontier labs use to plan training runs. They call it the Skaling law, and the argument is that standard scaling laws break in a specific and expensive way: they "systematically under- and overestimate loss at data-scarce and overtraining extremes," which is exactly the territory a lot of current model development lives in.

The root cause, according to the abstract, is the underlying assumption that model size and training data impact the loss independently. Skaling drops that assumption and couples the two through a single interaction exponent. That one change is claimed to reduce Mean Absolute Percentage Error by 1.5 to 3x across both interpolation and extrapolation regimes.

The compute story is the part that matters for planning. Paired with a sparse grid strategy restricted to low-compute regimes, the paper says the method gets accurate full-grid extrapolation using roughly 10x less compute than uniform sweeps. If that holds, a small team can plan a large run from smaller experiments without absorbing the calibration bill that frontier labs currently eat.

A few things the abstract does not spell out are worth flagging before anyone rewrites a training plan around it. It does not report which model sizes or data ranges were used to fit the interaction exponent, does not say whether that exponent transfers across architectures or data mixes, and does not place its numbers side by side against the original Chinchilla and Kaplan fits on the same experimental grid — even though the paper's title ("Chinchilla's Exponents Meet Kaplan's Coupling") gestures directly at both. Those details will decide whether the 1.5-3x error reduction survives the move from paper to production.

If the result generalizes, the clearest immediate winners are academic groups and smaller labs, where a roughly tenfold cut in calibration compute is the difference between running a scaling study and not running one at all.