Stanford and UC Berkeley hosted workshops to shape California’s first TFAIA report. TFAIA is a state law requiring transparency, incident reporting, and whistleblower protections for frontier AI. It defines frontier models by a 10^26 FLOPs compute threshold. Stakeholders argued compute alone is an imperfect risk proxy. They recommended expanding definitions to include distilled models and domain-specific risks. The report proposes broader catastrophic risk definitions to cover harms like mental health and loss-of-control incidents. It also suggests adding R&D spend as a metric for large frontier developers. The authors urge California to build capacity for robust AI governance. They see TFAIA as a model for other states. The report is not collectively endorsed but reflects workshop discussions.
What changed
California had no state-level regulatory framework for frontier AI before TFAIA.
What it unlocks
State agencies can now require transparency, incident reporting, and whistleblower protections from frontier AI developers.
- 10^26 FLOPs threshold for frontier models
- $500M annual gross revenue for large frontier developers
- $100M R&D spend threshold proposed
- 50+ deaths or $1B+ damage for catastrophic risk
What you need to act on it
- frontier model trained using >10^26 FLOPs
- developer meets revenue or R&D spend thresholds
Sources