Meta Research Challenges Chinchilla Scaling Laws on Data-Compute Interactions

burkov · x · 2026-08-20

A new paper from Meta questions the additive assumption behind the Chinchilla scaling laws, which assumes model size and data contribute independently to training error. Analyzing nearly 400 training runs, the authors found that the benefit of additional data increases as the model gets larger. This implies that scaling both together yields better returns than scaling either alone, suggesting current scaling formulas need to account for this interaction term.

Original post →

More from Research

Research channel →