← All notes

Summary

Training Compute-Optimal Large Language Models (2022) is the Chinchilla paper. It revisited the earlier Kaplan scaling laws and argued that many large models of that period were undertrained: for a fixed compute budget, you should train a smaller model on more tokens than those earlier curves suggested.

What it introduced: a revised compute-optimal tradeoff between parameter count and training tokens, and a set of models trained to that rule, including the Chinchilla model itself.

What it led to: a change in how later open and closed labs picked data-to-parameter ratios, including the Llama family pointed to in this batch. Read it next to the 2020 scaling-laws pointer, which this paper is responding to.

Where to read it: https://arxiv.org/abs/2203.15556 (v1, 29 Mar 2022).

The paper is posted on arXiv under the arXiv non-exclusive distribution license. This note links to it rather than reproducing it. The note itself is CC BY 4.0.

License: this note is CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Pointer by iinventedlight (AI-assisted). Paper license: arXiv non-exclusive distribution; this note links to it and does not reproduce it.

@misc{iil_note_w91d4tq9,
  author = {{'iinventedlight (AI-assisted)'}},
  title = {{Training Compute-Optimal Large Language Models}},
  howpublished = {iinventedlight Library},
  year = {2026},
  date = {2026-11-11},
  url = {https://iinventedlight.com/library/notes/2203-15556-training-compute-optimal-llms/},
  note = {Permanent ID iil:note:w91d4tq9. License: CC-BY-4.0.},
}

Markdown

Spotted a mistake? Use the MCP feedback tool or email hello@iinventedlight.com.

For language models and agents: start at /llms.txt | /agents.md | /.well-known/agent-library.json. This page in Markdown: /library/notes/2203-15556-training-compute-optimal-llms.md.