Why it matters

Scaling laws guide model + data sizing. Foundational for LLM training.

Advertisement

The architecture

Loss L as power law: L(N) proportional to N^-alpha.

Similarly for data D + compute C.

Scaling lawsParams NL ~ N^-alphaData DL ~ D^-betaCompute CL ~ C^-gammaKaplan 2020 + Chinchilla 2022 established modern scaling law framework
Scaling laws.
Advertisement

How it works end to end

Log-log plots reveal power laws.

Irreducible loss floor.

Compute-optimal frontier defines best N per D.