California AI safety report proposes compute-disclosure rules and RSI caps
sethlazar · x · 2026-11-19
California's expert panel on AI safety released a new report that Seth Lazar and Sneha Revanur call the world's best near-term policy opportunity to manage RSI (recursive self-improvement) risk.
Sharp ideas in the report:
- Requiring companies to disclose what % of compute goes to frontier AI R&D
- Outcome-based standards complementing safety cases
- Bringing companies up to >SL3 model weight security
The Governor's press release mentions RSI near the top. Per the announcement, next steps include emergency shutdown authority, baseline standards for all AI developers, mandatory ongoing third-party oversight, and authority to cap recursive self-improvement.
Lazar argues parts could become law in months if CA moves fast — the biggest determinant being how loudly lab employees push.
More from AGI Musings
- Physicist urges researchers to ride the AI dragon, deep learning theorists agree — jasondeanlee · 2026-11-19
- Coinbase CEO Brian Armstrong: 99.999% of GDP transactions will be agent-to-agent in 10 years — davidsenra · 2026-11-19
- Work as Ritual Play-Acting in the AI Era? We Undervalue Play Itself — generativist · 2026-11-19
- Superintelligence is not an adequate theory of scientific progress — StewartalsopIII · 2026-11-19
- AI didn't break trust online — it just removed the excuse that a photo counts as proof — Yamapama · 2026-11-19
- DeepMind model: AI is competent but poorly coordinated, firms get leaner before jobs vanish — thealexbanks · 2026-11-19