Scaling law

A simple formula, usually a power law, that predicts how a model's test loss falls as its size, its training data or its training compute grows. It lets developers estimate the result of a large training run from smaller ones.