Glossary term
Glossary term
Foundations
An observed relationship between model capability and factors such as compute, data, and parameter count. Scaling laws are useful planning signals but should not replace empirical safety evaluation. Scaling-law expectations should be paired with safety testing because capability growth can produce unexpected operational and misuse risks.
Kaplan et al. (OpenAI, 2020) Scaling Laws for Neural Language Models is the original neural scaling law paper.
Hoffmann et al. (DeepMind Chinchilla, 2022) revised compute-optimal scaling and shifted training practices industry-wide.
Epoch AI tracks scaling law-related metrics including training compute, parameter counts, and dataset sizes across major models.