奥特曼向员工谈SI三种终局：国际监察、实验室合并，或一家独大「熔掉对手的GPU」
据科技媒体 The Information 报道，OpenAI 首席执行官奥特曼（Sam Altman）上周在公司内部员工会议上谈到先进超级智能可能的三种「终局」，称这是业界讨论过的三种安全开发路径。据两名知情人士透露，奥特曼在会上逐一介绍了这三种设想，并对第一种表现出支持。
第一种是设立一个类似国际原子能机构（IAEA）的国际监察组织，为最先进的 SI 系统制定安全标准、执行检查与审计。第二种是把处于领先位置的各家 SI 实验室合并为一个机构，在同一框架下安全地开发新系统，这一设想与 Google DeepMind 负责人 Demis Hassabis 提出过的方案相近。第三种被称作「Eliezer 方案」，以 SI 安全理论家 Eliezer Yudkowsky 的名字命名：一家公司率先实现能自我改进的 SI、取得压倒性的经济与算力优势，进而能「熔掉其他所有人的 GPU」、令对手出局。据报道，奥特曼表示第三种只是为完整起见才列入，并称其「糟糕」。
这番内部讲话正值特朗普政府开始为 OpenAI 这类「超级智能」开发商建立首套监管安排之际。国际监察的构想并非奥特曼首次提出：2023年，他与 Greg Brockman、Ilya Sutskever 在 OpenAI 发表的《超级智能的治理》一文中就提出，未来的先进 SI 系统需要类似 IAEA 的机构来检查系统、要求审计并核查安全合规。据 The Information 报道，奥特曼还表示，业界至今没有找到为 SI 进展强制「限速」的办法，并希望 OpenAI 在一年内理清自身的安全政策。（GSM 据 The Information 报道）
Altman Tells Staff of Three SI Endgames: An International Inspector, a Lab Merger, or One Firm That Could 'Melt Everyone Else's GPUs'
OpenAI chief executive Sam Altman last week outlined three possible long-term outcomes for advanced super intelligence at an internal staff meeting, describing them as three approaches the industry has discussed for developing SI safely, The Information reported, citing two people with knowledge of the comments.
The first scenario is an international inspector modeled on the International Atomic Energy Agency (IAEA) that would set safety standards for the most capable SI systems and carry out inspections and audits, an approach Altman appeared to support. The second is a merger of the leading SI labs into a single body to develop new systems safely, an idea similar to one associated with Google DeepMind chief Demis Hassabis. The third, dubbed the 'Eliezer plan' after SI safety theorist Eliezer Yudkowsky, envisions one company achieving self-improving SI and gaining such overwhelming economic and computing power that it could 'melt everyone else's GPUs' and knock rivals offline. Altman reportedly said he included that scenario for completeness and called it 'terrible.'
The remarks come as the Trump administration sets up its first system for overseeing developers of 'superintelligence' such as OpenAI. The inspector idea is not new for Altman: in OpenAI's 2023 essay 'Governance of superintelligence,' he, Greg Brockman and Ilya Sutskever proposed an IAEA-like body to inspect systems, require audits and check safety compliance for the most capable models. Altman also said the industry has not found a way to enforce 'speed limits' on SI progress and pushed for OpenAI to settle its safety policies within a year, according to The Information. (GSM, based on reporting by The Information)
发布：GSM·美国（发布链接）
图片来源：图：奥特曼（Sam Altman）2025年在 TED 大会发言（资料图），Steve Jurvetson 摄，CC BY 2.0，经 Wikimedia Commons。Photo: Sam Altman speaking at TED in 2025 (file photo), by Steve Jurvetson, CC BY 2.0, via Wikimedia Commons.
GSM Global · 本篇英文版 English edition of this story
