Anthropic proposes an AI safety bill with a $1 billion risk threshold, independent model reviews, and limits on state regulations.