Why it matters

Scaling laws guide training investment. Compute is expensive; laws optimize spend.

Advertisement

The architecture

Loss ~ N^-a + D^-b + irreducible.

Chinchilla: balanced N and D optimal.

Scaling laws deep diveLoss decreaseswith N, D, CChinchillabalanced N, DOptimal spendcompute vs dataKaplan 2020, Chinchilla 2022; guide training budget allocation
Scaling laws deep dive.
Advertisement

How it works end to end

Kaplan 2020.

Chinchilla 2022.

Balanced N + D optimal.