Why it matters

Kaplan started the scaling law era. Understanding evolution shapes context.

Advertisement

The architecture

Kaplan: GPT-3 sized for D that was too small.

Chinchilla: data starving found.

Scaling law evolutionKaplan 2020params-heavyChinchilla 2022balancedModern practiceover-train tokensKaplan underestimated data needs; Chinchilla + modern practice diverged
Kaplan vs Chinchilla.
Advertisement

How it works end to end

Kaplan: L ~ N^-alpha with early data-fitting.

Chinchilla: revised with matched compute frontiers.

Modern: over-train for inference.