Hasabis' „Bunker-Plan“: Das neue Buch zeigt, wie die Sicherheitsideale von DeepMind scheiterten

据 1M AI News 监测，《大西洋月刊》刊发科技史学家、金融史学家 Sebastian Mallaby 新书《The Infinity Machine》节选，基于 2023 至 2026 年间的持续采访，还原了谷歌 DeepMind 联合创始人 Demis Hassabis 自 2014 年谷歌收购至今的 AI 安全理念演变。
According to monitoring by 1M AI News, The Atlantic published an excerpt from the new book “The Infinity Machine” by technology historian and financial historian Sebastian Mallaby, which, based on ongoing interviews from 2023 to 2026, reconstructs the evolution of AI safety concepts of Google DeepMind co-founder Demis Hassabis since Google’s acquisition in 2014.

Hassabis 早年构想由单一科学团队代表全人类安全地建造超级智能，甚至计划在关键时刻带顶级研究员撤入一个秘密地堡，类似曼哈顿计划。一位加入不久的研究员回忆，Hassabis 在面试尾声警告他做好「随时飞往摩洛哥某个秘密地点」的心理准备。
In his early years, Hassabis envisioned a single scientific team building superintelligence safely on behalf of all humanity, even planning to take top researchers into a secret bunker at critical moments, similar to the Manhattan Project. A researcher who joined not long after recalled that Hassabis warned him at the end of the interview to be mentally prepared to “fly to a secret location in Morocco at any time.”

2014 年谷歌收购 DeepMind 时，Hassabis 开出罕见条件：设立独立外部监督委员会，严禁军事应用，谷歌不得完全控制技术部署。谷歌接受了。次年他在马斯克位于加州霍桑的总部召集了一场秘密会议，试图将潜在竞争者纳入统一阵线，结果适得其反：马斯克随后与 Sam Altman 联手创办了 OpenAI。
When Google acquired DeepMind in 2014, Hassabis set rare conditions: establish an independent external oversight committee, prohibit military applications, and Google could not have full control over technology deployment. Google accepted. The following year, he convened a secret meeting at Musk’s headquarters in Hawthorne, California, attempting to bring potential competitors into a united front, but it backfired: Musk subsequently teamed up with Sam Altman to found OpenAI.

Hassabis 之后发起代号「Project Mario」的秘密行动，试图从谷歌手中夺回治理自主权，组建法律团队，从 LinkedIn 创始人 Reid Hoffman 处获得 10 亿美元资金承诺，并考虑将 DeepMind 从谷歌拆分独立。这场斗争持续三年，最终失败。联合创始人 Mustafa Suleyman 也在 2019 年被迫离开公司。
Hassabis then initiated a secret operation codenamed “Project Mario,” attempting to regain governance autonomy from Google, forming a legal team, securing a $1 billion funding commitment from LinkedIn founder Reid Hoffman, and considering splitting DeepMind from Google to become independent. This struggle lasted three years and ultimately failed. Co-founder Mustafa Suleyman was also forced to leave the company in 2019.

2022 年底 ChatGPT 发布后，Hassabis 彻底放弃了只做有益科学的高姿态路线。他对 Mallaby 说「这是战时」，随后全力投入 Gemini 与 ChatGPT 的竞争。数千亿美元涌入 AI 军备竞赛，而无论国家监管还是公司治理结构都无法阻止竞争。Hassabis 的安全哲学也随之根本转变：「安全不在于治理架构，就算有一个治理委员会，关键时刻它大概率也不会做出正确的事。」他的新策略是确保自己「在决策桌上有一席之地，当安全问题出现时，能参与决定解决方案」。与此同时，谷歌已在积极向美国国防系统推销 AI，与 Hassabis 当年为收购设下的军事禁令形成鲜明反转。
At the end of 2022, after the release of ChatGPT, Hassabis completely abandoned the high-minded approach of only doing beneficial science. He told Mallaby, “This is wartime,” and then fully committed to competing with Gemini and ChatGPT. Hundreds of billions of dollars flooded into the AI arms race, and neither national regulation nor corporate governance structures could stop the competition. Hassabis’s philosophy of safety also fundamentally shifted: “Safety doesn’t lie in governance structures; even if there is a governance committee, it is highly unlikely to make the right decision at critical moments.” His new strategy is to ensure that he “has a seat at the decision-making table so that when safety issues arise, he can participate in deciding the solutions.” Meanwhile, Google has been actively marketing AI to the U.S. defense system, marking a stark reversal from the military ban set by Hassabis at the time of the acquisition.