AI ChronicleAI 纪年 Full timeline →

Research breakthrough · curated entry

Chinchilla scaling laws correct the parameters-to-data ratio

Training Compute-Optimal Large Language Models

Date
March 29, 2022
Category
Research breakthrough
Organization / Model
Google DeepMind
Evidence
1 source · 1 evidence family

DeepMind's Chinchilla showed that under a fixed compute budget, model size and training data should grow in proportion. It corrected the earlier tendency to scale parameters alone.

Curator's note: Cross-checked against the arXiv paper.

Curated sources · 1

  1. Paper preprints · arXiv2022-03-29

    Training Compute-Optimal Large Language Models

    arxiv.org

Cite this page

This page is a curated static snapshot and can be cited directly. Preferred format GB/T 7714—2025 (electronic resource [EB/OL]); the other 5 styles open below, together with the machine-readable files.

GB/T 7714—2025 (electronic resource [EB/OL])

AI 纪年. Chinchilla 缩放定律修正参数与数据配比[EB/OL]. (2022-03-29)[2026-11-03]. https://aihist.net/e/chinchilla-2022.html.

APA 7 · American Psychological Association

AI 纪年. (2022, March 29). Chinchilla 缩放定律修正参数与数据配比 [Web page]. https://aihist.net/e/chinchilla-2022.html

MLA 9 · Modern Language Association

"Chinchilla 缩放定律修正参数与数据配比." AI 纪年, 29 March 2022, aihist.net/e/chinchilla-2022.html. Accessed 3 Nov. 2026.

Chicago 17 · notes-bibliography

AI 纪年. "Chinchilla 缩放定律修正参数与数据配比." Published March 29, 2022. Accessed Nov. 3, 2026. https://aihist.net/e/chinchilla-2022.html.

IEEE · citation-reference list

"Chinchilla 缩放定律修正参数与数据配比," AI 纪年, 29 Mar. 2022. [Online]. Available: https://aihist.net/e/chinchilla-2022.html. [Accessed: 3 Nov. 2026].

ACM · author-date variant

AI 纪年. March 29, 2022. Chinchilla 缩放定律修正参数与数据配比. Retrieved November 3, 2026 from https://aihist.net/e/chinchilla-2022.html.

All six styles come from one template and can be compared word for word. [引用日期], Accessed, urldate and Y2 all carry this page's build date, 2026-11-03 — a static page is an immutable snapshot, so substitute the date you actually consulted it. The address in every style is the archive's canonical page: this English page translates the same snapshot, and the work keeps its name, so one record cites both. For one-click copying use the same entry in the timeline, which takes the day you opened it. Expand the other styles before printing.