Models and training

Scaling laws

working definition

Definition

Scaling laws are empirical relationships that estimate how a model's loss or measured performance changes as resources such as training compute, data, parameter count, or inference-time compute increase. They are fitted regularities over particular model families, data regimes, objectives, and measurement ranges.

Distinguish it from nearby terms

A scaling law is not a universal law of intelligence and does not guarantee that every downstream capability improves smoothly. Aggregate loss can scale predictably while specific tasks remain noisy, saturate, regress, or change abruptly under a different evaluation.

Check your understanding

Before extrapolating, identify the measured outcome, resource axis, fitted range, architecture and data assumptions, uncertainty, and whether the target deployment resembles the observed regime.