What it means
Scaling laws are the empirical finding that model loss falls in a smooth, predictable relationship with compute, dataset size, and parameter count. They are not laws of nature — they are fitted curves that have held across many orders of magnitude.
Their practical importance is enormous. They let a lab run small, cheap experiments and forecast how a much larger training run will perform, which turns a nine-figure training decision from a gamble into a projection. The entire investment thesis behind large-scale AI rests on that predictability.
Refinements have mattered too: earlier models were substantially under-trained on data relative to their size, and correcting that ratio produced better models at smaller sizes. More recently, attention has shifted to scaling inference compute — reasoning models — as an additional axis.
Why it matters
Scaling laws are why enormous capital keeps flowing into compute: the returns are forecastable rather than speculative. The open question — whether the curves eventually flatten, and where — is effectively the industry's central uncertainty.
In practice
Mostly context for reading the news. It explains why labs announce compute commitments as if they were capability commitments, and why "more compute" is a coherent strategy rather than a lazy one.