CourseTransformers: From Attention to Trained, Served and Interpreted Models · Module 9: Scale and Distributed Training · part 64 of 111
Part 64 · Module 9: Scale and Distributed Training

Lesson 1: Scaling laws, and what a small grid can say

28 min read·13 Oct 2026
Module 9 · Lesson 1 · About 45 minutes · Needs: Module 6 (the base recipe, the seed noise floor); Module 8, Lesson 4 (6ND and tokens per parameter)

The situation

Sol 426 at Aurora Station. Hana Aoki, the Compute Officer, holds the budget for GPU time hired on Earth, capped at $1,500 a run. She asks: "What would a model 1,000 times bigger than base buy us?"

You cannot train that model here, but you can train small ones, fit a curve and read it at the size you cannot afford. Kofi Mensah adds his rule: the answer comes with error bars, or it does not go to Hana. The running example is Hana's question and the ten-run grid that tries to answer it.

The rest of this course is yours to keep

This course is bought on its own, once, and stays readable afterwards, including the parts added to it later.