The empirical regularity that model performance improves predictably as compute, data, and parameters grow.
Broad technical consensus on what this term means. Disagreements that invoke it are about the world, not the word.
Scaling laws are measured curves, not theory: across many orders of magnitude, a model's loss falls as a smooth, predictable function of how much compute, data, and parameters go in. Their discovery turned training-run budgeting from a gamble into something like engineering, and underwrote the multi-billion-dollar bet that bigger would keep being better.
The definition is settled; the extrapolation is not. Whether the curves are bending, and whether new axes like inference-time compute reset them, is precisely the capability question the Argument Map tracks. A scaling law is an observed regularity, not a guarantee.