Ryan McKenna

Scaling laws for differentially private language model training accurately model compute-privacy-utility tradeoffs and identify optimal training configurations for privacy-preserving LLM development.

“Author of "Tool-space interference in the MCP era: Designing for agent compatibility at sca" on arXiv”

Google released VaultGemma, the largest open-weight language model (1B parameters) trained from scratch with differential privacy, alongside new scaling laws quantifying compute-privacy-utility trade-offs.

“Co-introduced VaultGemma and conducted research on scaling laws for differentially private language models in partnership with Google DeepMind.”