Nadella says stronger AI does not automatically solve alignment
Microsoft CEO Satya Nadella said he no longer believes stronger AI will naturally resolve alignment between model behavior and human values and goals. IT之家, citing Business Insider and a new episode of the 20VC podcast, reported that increased model size and capabilities have not resolved alignment, which Nadella now considers a more difficult and distinct problem. The report also said that in an October 10 post on X, Nadella urged enterprises to treat AI models as internal security threats and use deterministic system design, human oversight, operating procedures and industry standards to protect against behavior that is difficult to predict.
Nadella’s reported position links rising AI capability to unresolved alignment and uncontrollable behavior. It also identifies specific safeguards for enterprises deploying models whose behavior is difficult to predict.
微软 CEO 萨提亚·纳德拉对 AI"对齐"问题的判断,发生了公开反转。据《商业内幕》10 月 27 日报道,在新一期《20VC》播客里,主持人哈里·斯特宾斯问他过去一年什么想法变了,纳德拉坦言:自己过去总觉得,AI 越强大,让其行为贴合人类价值观与目标这件事,也会顺理成章地解决。
从"自然解决"到"更棘手"
现实却没按剧本走。纳德拉承认,模型规模一路膨胀、能力持续抬升,这部分确实如他所料;可"对齐"——也就是让 AI 行为守规矩——至今没被一并解决。他说:"我过去总以为对齐问题自然就会解决。我不再这么认为了。对齐可能是一个更棘手、性质也不同的问题。"
所谓对齐,关乎 AI 的行为究竟听不听人话、是否符合人类目标。当模型大到连设计者都难完全预判其举动,能力增长与可控性增长脱钩,成了纳德拉不得不直面的新现实。
把 AI 当"内鬼"来防
这种警惕早有苗头。10 月 10 日,纳德拉就在 X 平台发文警告 AI 风险,主张把外部安全管控拧紧。他的核心建议很刺耳:企业应当把 AI 模型视作"内部安全威胁",做防护时先假定模型已经被攻破。
他在帖子里写道,要给那些行为难以完全预测的模型套上"可靠、具有确定性的系统设计、人工管控和完善的操作规程"三道防线;标准还不够的地方,就赶紧立行业规矩。
纳德拉并非孤例。OpenAI 的奥尔特曼、Anthropic 的阿莫代伊都曾对 AI 的潜在威胁敲过警钟;今年 7 月,更有 1000 多名从业者联署公开信,提醒 AI 能力狂奔的速度,可能已经甩开了人类理解或控制它的能力。当顶尖操盘手集体调低对"强即安全"的信任,AI 治理恐怕要从口号走向更硬的工程与制度了。