THE HOOK
California has found a new way to embarrass AI labs: make them say out loud how much compute they spend keeping models safe.
THE GIST
State Senator Josh Becker announced the Frontier AI Safety Act on 20 October. Major frontier developers would publish quarterly reports on the share of research compute devoted to safety, with independent verification. CalOES could set minimum standards, enforced by the Attorney General.
THE TAKE
This is the smartest AI bill in ages because it measures something that's hard to fake. Compute is money, and money shows priorities. The catch is definitions: every lab will suddenly discover that half its training runs were secretly safety research. Expect creative accounting.
OPINION
Becker's office points out that Anthropic once reported about 6% of its research compute going on safety over a seven-day window, and the press release notes other big labs haven't shared comparable figures. If it passes, the first quarterly league table will do more to shape lab behaviour than a dozen voluntary pledges.
- Story type: Breaking news
- Impact: Medium
- Confidence: High (official announcement)
- Audience: Business
- What changed: A California senator unveiled a bill requiring frontier AI labs to report and verify the share of R&D compute spent on safety.
- What to watch: The bill's formal introduction, how 'safety' compute gets defined, and whether OpenAI and Google lobby against it.
- Reading time: 1 minute
