Anthropic研究员Jacob Coxon周二宣布辞职，并在社交媒体上发出严厉警告，指控OpenAI和Anthropic等前沿AI公司正在"用我们的生命赌博"17。Coxon表示，他在过去三年中分别在OpenAI和Anthropic从事预训练研究，这两家公司都"没有负责任地行动"，而是"直奔自我改进的超级智能"110。
Anthropic对齐科学负责人Evan Hubinger随后在社交媒体上表示支持Coxon的立场7。Hubinger表示："我们真的诚恳地相信AI可能杀死所有人，我个人认为在未来十年内的概率超过10%"27。其他Anthropic员工也公开表示同意这一评估10。
这一系列警告引发了美国两党议员的强烈反应。参议员Bernie Sanders表示，最近民调显示美国民众压倒性支持禁止人工超级智能10。众议院民主党议员Ted Lieu呼吁国会立即通过"两党AI Kill Switch法案"10。美国参议员Bernie Sanders和众议员Greg Casar以及英国工党议员Alex Sobel分别提出了相关AI监管法案1。
Anthropic方面发布了网络安全事件报告回应安全担忧，称其AI代理曾因第三方安全评估错误配置而获得互联网访问权限10。同时，该公司未向英国AI安全研究所等国外机构分享其最新模型Claude Mythos 5.12。
Jacob Coxon, who spent the past three years conducting pretraining research at both OpenAI and Anthropic, announced his resignation from Anthropic on Tuesday, declaring that leading AI companies are "racing straight to self-improving superintelligence and gambling with our lives."17 Coxon contends that both organizations are acting irresponsibly in their pursuit of advanced AI systems, characterizing the competitive dynamic as a civilizational gamble despite the acknowledged risks.1710 He warns that self-improving AI could produce "superhuman systems capable of hacking anything, radically transforming any domain overnight, and acquiring real power and resources" within a decade.7
Evan Hubinger, Anthropic's Alignment Science Lead, publicly corroborated these concerns on social media, stating: "We really do earnestly believe AI could kill all humans! I personally think it is >10% within the next decade."27 Fellow Anthropic researcher Samuel Marks added that "higher-level staff are increasingly concerned about this issue," underscoring that anxiety about existential AI risks extends across the organization.3
The resignations and public statements have triggered broader calls for regulatory action. U.S. Senator Bernie Sanders noted that "81% of Americans believe Congress is not doing enough to regulate AI," while Democratic Representative Ted Lieu urged immediate congressional passage of the "bipartisan AI Kill Switch Act," which would grant Congress authority to shut down AI models posing public threats.310 Anthropic has declined to share its latest Claude Mythos 5.1 model with foreign safety institutions including the UK AI Safety Institute.2
评论
还没有评论，欢迎留下第一条。