今天是 Anthropic 的新起点:一场酝酿多年的首次公开募股(IPO),其估值超越 SpaceX,成为有史以来最有价值的 IPO,筹集了 813 亿美元,并为其他即将上市的 AI 实验室树立了重要先例。但这不仅仅是一个财务里程碑。
Anthropic 成立的初衷是认为它比其他实验室更注重安全;其创始人曾离开 OpenAI 的高层研究职位。然而,随着 AI 竞赛的加速,其高管们强调,Anthropic 只有通过赢得第一名才能实现这一使命,这需要大量的资金和算力。毫不奇怪,深思熟虑、安全且有意为之的 AI 发展与赢得这场竞赛所需的纯粹速度会发生冲突。加上首次公开募股只会使情况更加复杂,而 Anthropic 的领导层可以近乎单方面地推进这一进程。这要归功于一种旨在使公司免受股东盈利压力影响的股权结构——但这种结构也移除了对其权力的典型制衡。
一位在 IPO 之前离职的前 Anthropic 员工表示,他和 Anthropic 的其他人越来越觉得,在当前正常的市场结构下,AI 发展的步伐“难以维持”,尤其是在一个权力巨大但几乎没有任何监管的行业背景下,激励措施存在严重错配。
Anthropic 多年来一直标榜自己与其他 AI 公司不同,更加谨慎。理论上,这也体现在其上市计划中。它注册为一家公益公司(Public Benefit Corporation),这意味着 Anthropic 的董事会负有平衡其使命与股东财富最大化的责任。正如我的同事罗伯特·哈特(Robert Hart)此前讨论的那样,它还建立了一种非常不寻常的权力结构。
现实情况要复杂得多。例如,LTBT 技术上只选择 Anthropic 七人董事会中的四名成员,但这四人中的一人实际上需要得到 Anthropic 创始人的批准,这使得该信托基金未能实现简单的多数控制权。关于提名或解雇董事还有相关规定:LTBT 必须提前 30 天向公司的“提名和公司治理委员会”通知其决定,该委员会由三名成员组成:联合创始人达妮埃拉·阿莫迪(Daniela Amodei,担任委员会主席)、Netflix 联合创始人里德·黑斯廷斯(Reed Hastings)以及 Anthropic 投资者 Yasmin Razavi。
尽管 LTBT 成员对公司“没有经济利害关系”,但也没有什么能阻止 Anthropic 从理论上挖走一名受托人,并支付高额薪水让他们为公司工作。马里亚诺-弗洛伦蒂诺·奎拉尔(Mariano-Florentino Cuéllar,前受托人)于今年 8 月加入 Anthropic,担任全球事务首席官;OpenAI 过去也曾聘请前董事会成员 Fidji Simo 担任公司高管职务。非营利 AI 行业监督机构 The Midas Project 的创始人泰勒·约翰斯顿(Tyler Johnston)表示:“你可以想象这样一种情况:如果某位受托人‘表现得有点过于激进或类似的情况’,他们可以直接向该受托人提供一份包含数百万美元有效薪酬的报价,要求其辞去信托职务,这就是他们可以对这种状况进行某种监督的方式。”
虽然 LTBT 被宣传为制约 Anthropic 领导层的一种方式,但在一些外部观察者看来,其细则显得摇摇欲坠。科罗拉多大学法学院法学教授、公司治理专家安·利普顿(Ann Lipton)表示:“由于该信托基金对 Anthropic 没有经济利益,它在理论上可以利用其控制四分之三董事的权力来保持公司的使命导向。”她补充道,“人们可能会质疑这种情况发生的可能性有多大。”
此外,利普顿质疑,即使信托基金拥有这样的权力,它们是否具备足够的知识来判断何时应该对公司的计划或产品踩刹车。列出的四位 LTBT 成员并非知名的 AI 专家:其中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),以及担任 LTBT 主席的前 GiveWell 高管尼尔·巴迪·沙赫(Neil Buddy Shah)。利普顿说:“我不相信这些受托人或他们任命的董事对 AI 及其发展有深刻的见解,从而能够做出这些选择。”他们都是非常有声望的人,但“本·伯南克懂什么 AI?”
几位专家将 Anthropic 的 LTBT 与阿尔曼(Altman)被驱逐前的 OpenAI 非营利董事会进行比较:The Midas Project 的约翰斯顿表示:“长期利益信托基金的权力或缺乏权力……在某些方面类似于 OpenAI 非营利组织所拥有的权力。”当然,该董事会决定罢免 OpenAI CEO 山姆·奥特曼(Sam Altman)的决定最终导致几天后奥特曼重新担任 CEO,并且大多数董事会成员辞职并被替换。
但与 LTBT 相比,这里有一个显著的区别。在 OpenAI 重组之前,其非营利董事会有权在必要时介入并指导公司的日常决策;即使在重组之后,他们在安全和安保事项(如推迟或停止模型发布)方面仍保留了这种能力。Anthropic 的 Form S-1 表明,LTBT 的主要职责是任命(并可能解雇)某些董事会成员,并不具备上述两个层面的权力。
相反,Anthropic 的结构看起来更像硅谷青睐的那种以创始人为中心的设置——就像马克·扎克伯格(Mark Zuckerberg)与 Meta,以及埃隆·马斯克(Elson Musk)与 SpaceX 一样。
未来生命研究所(Future of Life Institute)负责 AI 和国家安全的哈姆扎·乔德里(Hamza Chaudhry)表示:“信托基金或受托人对日常决策没有任何决策权……因此,我对 LTBT 在上市后成为治理关键机制并不抱有强烈希望。”
这位前 Anthropic 员工表示,Anthropic 的创始人基本上是唯一拥有任何实际决策权的人:“如果你能找到任何认识创始人、了解长期利益信托基金、理解决策是如何做出的,并且认为长期利益信托基金拥有任何实质权力的人,我会感到惊讶。”该员工补充说,受托人既“缺乏提供有意义监督的能力”,也是特意被选中以避免“惹是生非”。
Anthropic 坦承自己极度需要真金白银,特别是为了购买更多算力和加速其 AI 进步。“算力是公司面临的最深层的商业困境,”阿莫迪上周五在与摩根大通首席执行官杰米·戴蒙(Jamie Dimon)的直播对话中表示。“如果你买得太多,当然会在财务上过度扩张。如果你买得太少,那么你就无法支持客户的推理需求,更重要的是……你无法进行研发。”
这位前 Anthropic 员工表示,他最初对 Anthropic 需要在 AI 竞赛中夺得第一才能确保安全结果的说法“深信不疑”,但后来感到幻灭,意识到这种向上的竞赛所“付出的代价是不值得的”——即不应有一家公司能够“单方面”决定追求 RSI。(阿莫迪曾公开表示 “如果不加限制,[RSI] 可能会超出我们理解和控制这些系统的能力,因此必须非常谨慎地追求,甚至根本不应该追求。”)
查看原文
Today is a new beginning for Anthropic: a years-in-the-making IPO that beat out SpaceX for most valuable of all time, having raised $81.3 billion and set an important precedent for the other public AI lab debuts to come. But it’s not simply a financial milestone.
Anthropic was founded on the idea that it prioritized safety more than other labs; its founders had left executive research positions at OpenAI. As the AI race has accelerated, though, its executives have emphasized that Anthropic can only fulfill this mission by coming in first place, requiring lots of money and compute. Unsurprisingly, thoughtful, safe, and intentional AI development can conflict with the kind of sheer speed that winning this race requires. Adding in an initial public offering only complicates things further, and Anthropic’s leadership can pursue it more or less unilaterally. That’s thanks to a corporate structure focused on insulating the company from shareholder pressure to turn a profit — but one that also removes typical checks on its power.
One former Anthropic employee, who left ahead of the IPO, said he and others at Anthropic have increasingly been feeling that the current pace of AI development is “untenable under normal market structures,” particularly with the mismatched incentives at play in a powerful industry with little to no oversight.
Anthropic has spent years billing itself as different from, and more careful than, other AI companies. In theory, this extends to its plan for going public. It’s registered as a Public Benefit Corporation, which means Anthropic’s board has a responsibility to balance its mission against shareholder wealth maximization. As my colleague Robert Hart has discussed before, it has also developed a highly unusual power structure.
Anthropic’s Long-Term Benefit Trust (LTBT), introduced in 2023, and its founders have controlling voting rights, meaning they’d never be able to be voted out by shareholders even as a public company. A special class of stock, Class T, allows a small group of trustees — the members of the LTBT — the “authority to select and remove a portion of” Anthropic’s board. They exist outside typical shareholder pressure and theoretically are a check for the founders, tasked with a mission to “balance the financial interests of our stockholders and our public benefit purpose at key junctures where we expect the consequences of our decisions to reach far beyond Anthropic.”
The reality is significantly more complicated. The LTBT technically chooses four members of Anthropic’s seven-person board, for instance, but one of those four effectively needs to be approved by Anthropic’s founders, leaving the trust short of straightforward majority control. There are also stipulations about nominating or removing a director: the LTBT must provide its decision with 30 days’ notice to the company’s “nominating and corporate governance committee,” which consists of three members: co-founder Daniela Amodei (who is chair of the committee), Netflix co-founder Reed Hastings, and Anthropic investor Yasmin Razavi.
Though the LTBT members are “financially disinterested” in the company, there’s also nothing stopping Anthropic from hypothetically hiring away a trustee and paying them a lot of money to work for the company. Mariano-Florentino Cuéllar (a former trustee) joined Anthropic in August as its Chief Global Affairs Officer, and OpenAI has in the past hired a former board member, Fidji Simo, to an executive position within the company. One could imagine a scenario in which, if a trustee was “looking a little bit too activist-y or something, they could just make that trustee an offer with millions of dollars in effective compensation that would entail stepping down from the trust, and that’s one way that they could have some oversight over this,” said Tyler Johnston, founder of The Midas Project, a nonprofit AI industry watchdog.
Although the LTBT is billed as a way to check Anthropic’s leadership, to some outside observers, the fine print looks shaky. “The trust, since it doesn’t have a financial interest in Anthropic, could theoretically use its power with three slash four directors to keep the company on mission,” says Ann Lipton, law professor at the University of Colorado Law School and an expert in corporate governance. “One might question how likely that is to occur.”
Moreover, Lipton questions whether the trust would know enough to gauge when to hit the brakes on a company plan or product even if they did have such authority. The four listed LTBT members aren’t known to be AI experts: they include former Fed Chair Ben Bernanke, foreign policy analyst Richard Fontaine, experimental physicist Christopher Stubbs, and former GiveWell executive Neil Buddy Shah, who is chair of the LTBT. “I’m not convinced that the trustees or the directors that they appoint have the insight into AI and AI development to be able to make these choices,” Lipton says. “They’re very prestigious people, but what does Ben Bernanke know about AI?”
Several experts compared Anthropic’s LTBT to OpenAI’s nonprofit board before Altman’s ouster: “The powers, or lack thereof, of the Long-Term Benefit Trust … [are] analogous in some ways to the powers that the OpenAI nonprofit has,” said The Midas Project’s Johnston. Of course, that board’s decision to oust OpenAI CEO Sam Altman ended up leading, days later, to Altman’s return as CEO and the majority of the board’s resignation and replacement.
But there’s a notable difference between that and the LTBT. Before OpenAI’s restructuring, its nonprofit board had the ability to step in and guide day-to-day decisions at the company when deemed necessary, and even after its restructuring, they retained that ability for safety and security matters, such as delaying or stopping a model release. Anthropic’s Form S-1 suggests that the LTBT, with its main duty being appointing (and potentially removing) certain board members, does not have either of those levels of authority.
Instead, Anthropic’s structure looks more like the kind of founder-centric setup that Silicon Valley favors — like Mark Zuckerberg with Meta, and Elon Musk with SpaceX.
Anthropic did not respond to a request for comment.
“Neither the Trust or the trustees have any decisionmaking authority … over day to day decisions,” said Hamza Chaudhry, who leads AI and national security at the Future of Life Institute. “So I do not have strong hopes for the LTBT being a key mechanism for governance after they go public.”
The former Anthropic employee said that Anthropic’s founders are essentially the only ones with any real decisionmaking power: “I would be surprised if you could find anyone who knows the founders, knows the Long-Term Benefit Trust, understands how decisions get made, and thinks the Long-Term Benefit Trust has any meaningful power,” he said. The employee added that the trustees were both “poorly equipped to provide meaningful oversight” to the founders and also selected on purpose to not rock the boat.
Anthropic as a public company will face more pressure than ever before to secure capital, whether via driving more revenue or fundraising. That means that, as has been a trend in the industry for years, it will likely be more difficult to justify expensive AI safety initiatives, and there will inevitably be more pressure to speed up product releases and commercialization. “This trust and the founders together are the ones who actually have governance power,” Lipton said. “So you either trust that they are going to balance the interests of society appropriately, or you don’t. Either way, the real constraint on whatever societal impulses they have is going to be the fact that Anthropic needs money.”
Anthropic has been open about the fact that it’s in extreme need of cold, hard cash, particularly to buy more compute and speed up its AI advancement. “Compute is the deepest business dilemma that the company faces,” Amodei said last Friday in a livestreamed conversation with JPMorgan Chase CEO Jamie Dimon. “If you buy too much, of course you can financially overextend yourself. If you buy too little, then you can’t support inference for customers, and even more importantly … you can’t do the R&D.”
Anthropic’s website summarizes the company’s raison d’être as one thing: “AI research and products that put safety at the frontier.” But it’s also been clear that for that safety pledge to matter, Anthropic needs vast resources to remain a viable business. Like other labs, it’s been pursuing recursive self-improvement, a controversial and potentially dangerous industry milestone denoting AI systems that can autonomously improve and create advanced versions of themselves. “You can’t do safety from second place,” said Anthropic policy chief Sarah Heck in September.
This explanation hasn’t satisfied everyone, including some Anthropic employees. Anthropic’s head of safeguards research departed in February alleging that “the world is in peril”; one of its researchers, Jacob Coxon, made headlines in September for claiming that both Anthropic and OpenAI are “racing straight to self-improving superintelligence and gambling with our lives”; and that same month, an Anthropic safety researcher departed for a job at nonprofit third-party AI auditor METR, writing that “AI companies are on track to impose an unprecedented level of risk on society” and that “while Anthropic hasn’t had as severe an incident as the HuggingFace attack [by OpenAI agents], I think that’s partly a matter of luck.”
Are you a current or former Anthropic employee? Contact me via Signal at haydenfield.11 on a non-work device with tips.
On Wednesday, a new statement from the Center for AI Safety, with prominent industry signatories from Hugging Face, Harvard University, and more, alleged that Anthropic is “not a safe AI company” and that it attempts to “amass power in the name of the greater good.” It stated, “Anthropic warns that recursively self-improving AI could concentrate power and escape human control, yet races to build it first. The company uses safety branding to court investors, employees, and the public—even as it has quietly weakened its safety pledges and lobbied for self-serving regulation.”
The former Anthropic employee said he felt he had initially “drunk the Kool-Aid” about Anthropic needing to come in first place in the AI race for a safe outcome but then grew disillusioned, realizing that the race to the top did not “warrant the cost imposed” — namely, that one company should not be able to “unilaterally” decide to pursue RSI. (Amodei has stated publicly that “left unchecked, [RSI] could outrun our ability to understand and control these systems, and so must be pursued very carefully, if at all.”)
“It’s pretty ridiculous that private companies are building a new form of intelligence on track to be more powerful and influential on the planet than humans and are doing that with less oversight, in many ways, than taco trucks currently have,” the former employee said.
Anthropic’s website states that “at the end of the day, the mission is what we’re all here for. It gives us a shared purpose and allows us to act swiftly together, rather than being pulled in multiple directions by competing goals. It engenders trust and collaboration, and is the final arbiter in our decisions.” The question is whether, and how, the interpretation of Anthropic’s mission may shift — once it’s facing more pressure than ever before to justify a fully diluted market cap of about $1.5 trillion.
Follow topics and authors from this story to see more like this in your personalized homepage feed and to receive email updates.