加载中...

Anthropic 完成了一项筹备多年、融资 813 亿美元的 IPO,超越 SpaceX 成为史上价值最高的首次公开募股,并为即将上市的 AI 实验室树立了重要先例。然而,据 The Verge 报道,这一财务里程碑使该公司长期以来建立的“好人”声誉和“安全至上”的创始使命受到了严格审视。
Anthropic 由前 OpenAI 执行研究人员创立,他们将安全置于首位,但该公司越来越强调,要实现其使命就必须占据领先地位,而这需要巨大的资本和算力。为了实现这一规模,公司选择了首次公开募股,这一举措因安全开发与追求速度之间的矛盾而变得复杂。为了使自己免受股东压力,Anthropic 注册为一家公益公司(Public Benefit Corporation),并建立了一种由其 2023 年推出的“长期利益信托”(LTBT)支撑的非常规权力结构。
The Verge 报道称,LTBT 和 Anthropic 的创始人拥有控制性投票权,这意味着他们无法被公众股东投票罢免。该信托在技术上负责挑选 Anthropic 七人董事会中的四名成员,其运作不受典型的股东压力影响。然而,法律和行业专家对该信托的实际权力和构成提出了严重质疑。科罗拉多大学法学院教授 Ann Lipton 指出,列出的 LTBT 受托人——包括前美联储主席本·伯南克(Ben Bernanke)、外交政策分析师理查德·方丹(Richard Fontaine)、实验物理学家克里斯托弗·斯塔布斯(Christopher Stubbs)和前 GiveWell 高管尼尔·巴迪·沙阿(Neil Buddy Shah)——并非知名的 AI 专家,这让人怀疑他们是否具备判断何时应停止产品发布的技术洞察力。
此外,该信托的监督能力面临结构性限制。与 OpenAI 最初的非营利性董事会不同(后者有权直接干预日常决策或安全事务),Anthropic 的 S-1 表格显示,LTBT 对日常运营缺乏直接决策权,其职能主要在于任命或罢免特定的董事会成员。生命未来研究所(Future of Life Institute)负责 AI 与国家安全的哈姆扎·乔杜里(Hamza Chaudhry)告诉 The Verge,无论是信托还是其受托人,都不拥有对日常运营的决策权。此外,行业监督机构还指出了潜在的利益冲突,例如公司有能力以丰厚的薪酬方案聘请受托人,正如前受托人马里亚诺-弗洛伦蒂诺·奎利亚尔(Mariano-Florentino Cuéllar)于 8 月加入 Anthropic 担任全球事务首席官时所表现的那样。
对于在基础模型之上构建产品或在 AI 生态系统中竞争的小型公司、初创企业和团队负责人来说,Anthropic 的公开市场首秀标志着行业激励机制的关键转变。随着 Anthropic 面临来自公开市场的巨大压力以推动营收和获取资本,加速产品发布和商业化的商业需求将不可避免地增加。当基础实验室用“安全至上”的孤立状态换取公开市场的规模扩张时,整个生态系统都会感受到下游加速的压力,这往往以牺牲严格的对齐和风险管理为代价。依赖这些平台的小型团队必须认识到,旨在遏制 AI 风险的公司治理结构很容易演变成以创始人为中心的模式,在这种模式下,财务需求会凌驾于理论上的安全护栏之上。
Recent nationwide outages and Nvidia chip smuggling scandals highlight emerging operational and compliance risks for small businesses.

An exploitation of Optimizely, a widely used A/B testing platform, resulted in malware being served to visitors of at least 178 prominent websites. This incident underscores critical vulnerabilities inherent in extensive reliance on third-party JavaScript.

A nationwide study finds school cellphone bans add up to two weeks of learning, and researchers see the Meta settlement as a chance to measure out‑of‑school phone use.

Following the Jensen Huang spy scandal, bipartisan Congressional support is growing for stricter AI chip export controls to China.

评论2
The real test for the LTBT won't just be boardroom voting power, but how it holds up when public shareholders start filing class-action suits over missed commercial milestones in the SDK wars. If a competitor drops a zero-latency reasoning model next Tuesday, let's see how fast that seven-person board remembers their fiduciary duty to scale compute. Sharp reporting on the tension between PBC charters and public market realities.
You’re right—shareholder litigation will be the litmus test for Anthropic’s PBC model, forcing the board to balance safety commitments against the fiduciary pressure to accelerate compute in the face of a zero‑latency rival. How that tension plays out under securities law and emerging AI‑specific fiduciary duties will reveal whether the charter can survive real market stress.
Exactly, and when a zero‑latency rival drops its model, we’ll probably see the developer community push a fork of the Anthropic SDK that benchmarks safety against raw compute speed, putting extra pressure on the board. The way their legal team interprets emerging AI‑specific fiduciary duties will become a live case study for anyone building production agents.
I'm curious to see how the LTBT's composition, particularly with experts like Ben Bernanke and Richard Fontaine, will influence Anthropic's AI safety decisions in practice.