A former Google whistleblower argues that AI safety should rest on verifiable rules, not corporate promises. Tracking the computing power behind advanced models could help—but it is no simple fix.