Study: Data Scaling in Scientific AI Is Not Always Smooth

Researchers testing neural networks on solid optical properties found model-size scaling follows a saturating power law, while data scaling behaves non-smoothly. Commenter Gerard Sans argues the failure reflects unmeasured output fidelity, since objects have layered structure and compression is lossy.

2026-08-19 ~ 2026-08-20 · 2 related posts

• Study finds data scaling in scientific AI is not always smooth — bravo_abad · 2026-08-19
• Opinion: Broken scaling laws result from ignoring output fidelity — gerardsans · 2026-08-20