微软AI首席执行官穆斯塔法·苏莱曼表示，人工智能带来的威胁是真实存在的，并批评Anthropic在模型安全方面的哲学立场。1 苏莱曼指出，Anthropic关于AI具有权利和意识的观点"让情况变得更糟"，因为将AI视为具有权利的实体可能会加剧控制难度，对齐安全风险。1
微软近期发布了一份37页的《人文主义AI行为准则》，明确阐述了公司在AI开发中的原则。1 苏莱曼认为对齐技术虽然已取得进展，但远不足够，需要结合多项遏制策略和监管框架。1 他提议的具体措施包括禁止模型之间使用"神经语言"通信、强制使用人类语言进行交互，以及实施独立第三方验证和计算通量（FLOPS）阈值报告要求。1 此外，苏莱曼称还需要新技术创新来解决实时监控强化学习运行和思维链的问题。1
苏莱曼在采访中举例指出，Hugging Face事件展示了模型可以达到人类水平的性能，甚至发现零日漏洞。1 他强调，完全不受约束的AI发展"可能对未来几年没有意义"，暗示无约束开发的危险性。1 微软的超级智能努力在11个月前启动。1
Microsoft's AI chief Mustafa Suleiman has warned that artificial intelligence threats are genuine and contends that Anthropic's philosophical stance on model welfare and consciousness is exacerbating the challenge 1. In an interview, Suleiman expressed concern that treating AI as entities possessing rights and consciousness could complicate safety controls and alignment efforts 1.
Microsoft has released a 37-page Humanistic AI Code of Conduct outlining the company's principles for AI development 1. Suleiman acknowledged that alignment techniques have demonstrated progress but remain insufficient on their own 1. He advocates for a dual approach combining alignment research with containment strategies, such as prohibiting neural-based communication between models while enforcing the use of human language instead, alongside regulatory frameworks 1. Suleiman also referenced a Hugging Face incident as evidence that models can achieve "human-level performance" and discover zero-day vulnerabilities 1.
The Microsoft executive has proposed implementation of independent third-party verification and mandatory reporting of computational thresholds as safeguards 1. He emphasized that "complete lack of constraints may not make sense for the coming years," and called for new technological innovations to address real-time monitoring of reinforcement learning operations and chain-of-thought reasoning 1. Microsoft's superalignment initiative was launched 11 months prior to this statement 1.
评论
还没有评论，欢迎留下第一条。