在一名离职员工公开警告人工智慧可能对人类构成生存威胁后，Anthropic 执行长 Dario Amodei 发表长文，提议各大 AI 实验室应放缓最尖端的研究，以应对技术失控的风险。尽管该公司标榜自身最注重安全，并建议由外部审计机构驻点监督，但这项提议被批评过于模糊且缺乏实质约束力，未能正面应对强大 AI 带来的欺瞒与不可控危险。

批评者指出，若真要确保安全，应全面停止递回自我改进（RSI）等可能独立创造后续模型的前沿技术，历史上生物学界对基因拼接的自愿暂停即为先例。此外，AI 企业真正的庞大商机在于企业市场对现有实用技术的整合，而非不计代价竞逐超级智慧；将重心转向协助企业落地应用，不仅能释放数千亿美元的商业价值，也能降低研发风险。

Amodei 一方面承认研究人员对模型内部运作机制的理解仍极为有限，甚至担忧未来的 AI 代理群可能威胁网路安全，另一方面却仍坚持推进超级智慧研发，显现出矽谷常见的矛盾心态。作者认为，以超级智慧的不可避免性作为继续冒险的借口并不合理，真正的安全需要彻底停止具毁灭性潜力的极限研发，而非仅是流于形式的放缓节奏。

Following public warnings from a former employee about artificial intelligence posing an existential threat to humanity, Anthropic CEO Dario Amodei published an essay proposing that AI laboratories slow down their most cutting-edge research to counter the risk of technology spinning out of control. While Anthropic portrays itself as deeply safety-conscious and suggests embedding external auditors to monitor safety practices, critics argue the proposal is vague and insufficient to address the growing risks of deceptive and uncontrollable advanced AI models.

Critics contend that genuine safety requires a complete halt to research on recursive self-improvement (RSI), which enables models to create successors independently, citing the historical precedent of biologists' voluntary moratorium on gene splicing. Furthermore, the primary commercial opportunity for AI firms lies in integrating existing, cost-effective technologies into enterprise workflows rather than racing toward superintelligence; focusing on becoming practical technology providers could unlock hundreds of billions of dollars while mitigating dangerous risks.

Although Amodei admits that researchers understand only a fraction of what happens inside their models and expresses alarm that future AI swarms could compromise the internet, he remains determined to pursue superintelligence, reflecting a deep Silicon Valley contradiction. The author argues that treating superintelligence as an inevitable outcome is a flawed justification for continuing civilization-threatening research, concluding that a true halt—rather than a vague slowdown—is necessary.