Scaling Laws for Private LLMs

Google researchers derive empirical scaling laws for differentially private LLM training, showing performance depends on a 'noise‑batch ratio' and can be recovered by increasing compute or data. They validate this by releasing VaultGemma, a 1B‑parameter, open‑weight model trained with differential privacy that performs comparably to non‑private peers. — Quantifying privacy–compute–data tradeoffs gives developers and regulators a practical knob for legal‑compliant AI training that reduces memorization risks while maintaining utility.