Anthropic, OpenAI's proposed AI risk evaluators may not have enough power to prevent disasters | Blerbz