Chapter 9 · How an LLM Is Actually Built

Spend a Compute Budget

Split a fixed number of FLOPs between model size and training tokens with a published scaling law, place real models on the map, and see how serving costs change the answer.