Anthropic 提交 S-1 文件

Anthropic Files Its S-1

兹维·莫绍维茨 Zvi Mowshowitz · Don't Worry About the Vase · 2026-10-28 · Don't Worry About the Vase ↗

打开互动全文版(逐段中英对照 + 针对本篇提问)→

摘要 · Abstract

本文细读 Anthropic 的 S-1 文件,回顾公司财务状况并分析文件中最具影响的披露。文章认为,Anthropic 淡化了生存风险以及自身承担真实安全成本的意愿;其核心商业计划是递归自我改进;其治理结构通过 F 类股让创始人无限期保持控制权,而长期利益信托发挥的是治理而非 AI 安全机构的作用。财务上,作者主张 Anthropic 是盈利的:其报告的 800 亿美元亏损主要反映可转换票据重估带来的会计费用,而公司已募资 1411 亿美元、仍持有约 1310 亿美元,仅消耗了 94 亿美元。第三季度增长环比放缓至 45%,但作者预计第四季度将因 Opus 5.5 而表现强劲。结论是 Anthropic 的前景非常看好,尽管其约 8000 亿美元的计算承诺意味着它还需要更多资本,且 S-1 的样板文字中仍有值得注意的段落。

This article provides a close reading of Anthropic's S-1 filing, recapping the company's financial state and analyzing the document's most consequential disclosures. It argues that Anthropic downplays existential risk and its willingness to bear real safety costs, that its core business plan is recursive self-improvement, and that its governance structure leaves founders in control indefinitely through the Class F share, with the Long-Term Benefit Trust functioning as a governance rather than an AI safety body. Financially, the author contends Anthropic is profitable: its reported $80 billion loss largely reflects accounting charges from convertible note revaluations, while it has raised $141.1 billion and still holds about $131 billion, having burned only $9.4 billion. Growth slowed in Q3 to 45% quarter-over-quarter, but the author expects a strong Q4 driven by Opus 5.5. The conclusion is that Anthropic's forward prospects look very good, though its roughly $800 billion in compute commitments mean it will need more capital, and the S-1's boilerplate still contains passages that matter.

核心贡献 · Key contributions

局限 · Limitations

章节 · Sections(共 37)

全文 · Full text(逐段中英对照)

概述 Overview

全文终于来了,因此我们给予它模型卡级别的完整 RTFB 待遇。

The full text is finally here, so we are giving it the full RTFB of model card treatment.

首先回顾 Anthropic 的财务状况,然后仔细阅读 S-1 的相关部分。其中许多内容是技术性的、无趣的,通常是形式化的,会被跳过。许多内容被重复多次。

First, a recap of the state of Anthropic finances, then a close reading of the relevant parts of the S-1. A lot of it is technical and uninteresting, often pro forma, and gets skipped. A lot of it gets repeated many times.

其中许多也是 Anthropic 在重复我们已听过多次的‘热门曲目’,现在面向投资者巡回演出。准备好你的饮料,因为我们将不得不最大化上行空间并最小化下行风险。

A lot of this is also Anthropic ‘playing the hits’ we have heard many times, now to the investor circuit. Get your drinks ready for when we have to maximize the upside and minimize the downside.

如果这是你的菜,这份 S-1 数据的展示包含了一些由 Claude 创建的酷炫图表。

This presentation of S-1 data has some cool Claude-created charts if that’s your jam.

以下是我从 S-1 中总结的十大要点,面向已经熟悉 Anthropic 的读者。

Here are my top 10 takeaways from the S-1, for those already familiar with Anthropic.

1. 存在性风险仅以被动语态被简要提及,而对齐问题大多被视为平淡无奇。他们想宣称“大使命”,但又尽可能听起来像一家普通公司。他们淡化了为安全付出真实代价的意愿,并大幅淡化了危险。

1. Existential risk is mentioned briefly only in the passive voice, and alignment issues are mostly treated as prosaic. They want to claim Big Mission but also sound as much as possible like a normal company. They are downplaying their willingness to pay real costs for safety, and downplaying the danger quite a lot.

2. “调节前沿”不是他们会独自做的事情。他们会做嵌入式评估器,但其余部分需要整个行业共同参与,否则没有好处。

2. ‘Pacing the Frontier’ is not something they will do on their own. They’ll do the embedded evaluators, but the rest needs to be across the industry or it’s no good.

3. 核心商业计划是“闭环”,即递归自我改进,他们声称如果不实现这一点,他们的业务将输给竞争对手。

3. The core business plan is ‘closing the loop’ aka recursive self-improvement, and they claim that their business will lose to its rivals if they do not achieve this.

4. 创始人将通过他们的 F 类股票在很大程度上无限期运营 Anthropic,并且与员工一起,他们可以推翻长期利益信托(LTBT)。

4. The founders will largely run Anthropic indefinitely via their Class F share, and with the employees they can overrule the Long Term Benefit Trust (LTBT).

5. 一旦足够多的创始人离开,LTBT 将运营 Anthropic,但它仍然看起来像一个良好的治理机构,而不是一个 AI 安全机构。

5. The LTBT will run Anthropic once enough founders leave, but it continues to look like a good governance body, not an AI safety body.

6. Anthropic 的高速增长在 Q3 放缓,我认为 Q4 很可能重新加速。

6. Anthropic's huge growth slowed in Q3; I believe it is likely picking up again in Q4.

7. Anthropic 花的钱很少。他们融资 1410 亿美元,手头有 1310 亿美元。

7. Anthropic has spent very little money. They raised $141b and have $131b on hand.

8. Anthropic 是盈利的。上半年现金流为 +39 亿美元。那 800 亿美元的“亏损”是股价上涨迫使票据重估所致,但这些票据实际上已经完成转换。

8. Anthropic is profitable. H1 cash flow was +$3.9b. The $80b 'loss' was the stock going up forcing revaluation of notes, but those notes are already de facto converted.

9. Anthropic 的算力承诺约为 8000 亿美元。他们还需要更多。

9. Anthropic's compute commitments are roughly $800b. They will need more.

10. S-1 文件大多是重复的样板文字,适合被 Claude 批量生成、没人会读,但真正重要的部分依然重要,需要认真关注。

10. S-1s are mostly repetitious boilerplate ripe for Claudeslop no one will read, but the parts that count still count and one needs to pay attention.

6. 在把握并防范风险的同时实现收益。

6. Realize the Benefits While Navigating and Preventing the Risks.

8. 人们担心 AI 会杀死所有人(1)。

8. People Are Worried About AI Killing Everyone (1).

10. Anthropic 并未完全最大化利润。

10. Anthropic Does Not Fully Maximize Profits.

15. 希望推出世界上最强大的模型。

15. Hopefully Introducing the World's Most Powerful Model.

18. Anthropic 离开其员工便一无所有。

18. Anthropic Is Nothing Without Its People.

21. 你知道还有其他前沿模型实验室吗?

21. Did You Know There Are Other Frontier Labs?

23. 人们担心 AI 会杀死所有人(2)。

23. People Are Worried About AI Killing Everyone (2).

24. Claude 似乎与 Anthropic 的价值观相当不一致。

24. Claude Seems Pretty Misaligned to Anthropic.

25. 可能会有人起诉或监管 Anthropic。

25. Someone Might Sue or Regulate Anthropic.

34. 智能的商业,以及可能杀死所有人。

34. The Business of Intelligence and Maybe Killing Everyone.

35. 别在意我的监管。

35. Don’t Give a Damn About My Regulation.

让我看看钱 Show Me the Money

Anthropic 第三季度营收为 168 亿美元(调整后营业利润率为 11%,据彭博社),这算不上良好的增长,三个月“仅”增长 45%。与 OpenAI 等公司相比,Anthropic 在第三季度表现挣扎。根据预测市场,消息公布后 Anthropic 股价“仅”下跌约 4%,但这是因为我们已经知道 Anthropic 在相对实际模型质量方面经历了一段艰难时期。Fable 存在数据保留问题,Anthropic 在 S-1 第 49 页对此进行了讨论;人们不喜欢 Opus 4.7、4.8 和 5,而 Astra 令人印象深刻。还有人声称季节性因素损害了第三季度的消费。

Anthropic’s 3Q revenue number of $16.8 billion (with adjusted operating margin of 11%, via Bloomberg) was not good growth, ‘only’ 45% over three months. Anthropic struggled in Q3 relative to OpenAI and others. Based on prediction markets Anthropic was down ‘only’ ~4% on the news, but that was because we already knew that Anthropic went through a rough patch in terms of relative practical model quality. Fable has data retention issues that Anthropic discusses on p49 of the S-1, people disliked Opus 4.7, 4.8 and 5 and Astra was impressive. There are also claims that seasonality hurts consumption in Q3.

有人声称这意味着 Anthropic 故意在 S-1 中隐藏其第三季度业绩。实际上并非如此,在 26 天的周转期内完成符合 S-1 标准的第三季度完整报表,并非 Anthropic 力所能及。无论如何,Anthropic 并未隐藏业绩,你可以通过多种方式推导出这些数据,并且预计第三季度业绩将连同价格一起纳入下一次修订。

There are those who claim that this means Anthropic is deliberately hiding its Q3 results in the S-1. Realistically, no, full statements on Q3 on a 26 day turnaround is not something Anthropic was equipped to do to the standards of an S-1. Anthropic is not hiding the results in any case, and you can in various ways back them out, and also expect the Q3 results to get included in the next amendment along with the price.

他们在第 163 页详细介绍了可转换票据。此时,人们应该假设这些票据已经转换,并据此计算流通股,而不是将其视为进一步的“损失”。

They go into the details of the convertible notes on page 163. At this point, one should assume the notes have already been converted, and calculate the float accordingly, rather than thinking of this as further ‘losses.’

他们在后面直接列出了这些票据,以防你想知道它们是谁。

They list them outright later on in case you want to know who they are.

就前瞻性预期而言,情况看起来非常好。

In terms of forward expectations, things look very good.

截至 2026 年 9 月 30 日,Anthropic 每位员工支撑超过 660 万美元的过去 12 个月营收。

Each Anthropic employee supports over $6.6 million in trailing 12-month revenue as of September 30, 2026.

当时员工人数约为 5,300 人(第 228 页),过去 12 个月营收为 350 亿美元,相比之下三个月内营收为 168 亿美元。

The headcount is about 5,300 (p. 228) as of that time, with $35 billion in trailing 12-month revenue, in contrast to the $16.8 billion made over three months.

鉴于 Opus 5.5,我预计 Anthropic 将交出亮眼的第四季度业绩并重拾势头,因为它至少目前再次拥有明显最佳的产品,并且至少在我们看到 6.2 Astra 和 Sol 之前都将如此。Wonder 和 Argon 仍处于评估期,但它们几乎肯定无法与 Opus 5.5 匹敌,否则我们会听到非常响亮的消息。

Given Opus 5.5, I expect Anthropic to post an impressive Q4 and get its mojo back, as it again—at least for now—has the clearly best offering, and will at least until we see 6.2 Astra and Sol. Wonder and Argon are still in their evaluation periods, but they're almost certainly no match for Opus 5.5, or we would hear about it very loudly.

Charles:我曾押注他们的 ARR 运行率在 2026 年不会增长 10 倍,

Charles: I had bet against their ARR run-rate 10xing in 2026,

我认为那是一笔死钱,现在我觉得我可能有 30% 的几率赢。

I thought that was dead money, now I think there’s maybe a 30% chance I win.

《华尔街日报》提前看到了 Anthropic 的新招股说明书草稿,我们现在都已经看到了。

WSJ got an early look at Anthropic’s new draft prospectus, which we have all now seen.

Anissa Gardizy(《华尔街日报》):这家 AI 巨头在招股说明书草稿中表示,今年上半年录得超过 800 亿美元的净亏损,这一总额主要反映了与融资协议相关的会计调整。

Anissa Gardizy (WSJ): The artificial intelligence giant said in the draft prospectus it recorded a net loss of more than $80 billion in the first half of this year, a total that largely reflects an accounting adjustment related to financing agreements.

相比之下,去年同期的净亏损为 130 亿美元。

That compares to a net loss of $13 billion during the same period a year earlier.

……其中绝大部分亏损,即 720 亿美元,源于与可转换票据及公司为筹集资金而发行的其他工具价值变动相关的会计费用。

… The vast majority of that loss, $72 billion, resulted from accounting charges tied to changes in the value of convertible notes and other instruments the company issued to raise funding.

……Anthropic 列出了一长串风险因素,包括其将如何防止 AI 模型造成危害、以及如何保护自身免受这些责任(从复杂的网络攻击到不受控模型可能的敲诈勒索)影响的不确定性。

… Anthropic laid out a sprawling list of risk factors, including uncertainty over how it will prevent its AI models from causing harm and protect itself from those liabilities, from sophisticated cyberattacks to possible blackmail by uncontrolled models.

……Anthropic 今年上半年在算力上的支出超过了 2025 年全年的支出,2025 年其支出为 73 亿美元,并产生了 46 亿美元的收入。

… Anthropic spent more on computing capacity in the first half of this year than it did in all of 2025, when it spent $7.3 billion and generated $4.6 billion in revenue.

……该公司在未来七到十年内需支付约 4500 亿美元的云计算协议款项。这些交易大部分不可取消。

… The company is on the hook to pay around $450 billion in cloud computing agreements over the next seven to ten years. The bulk of these deals are noncancelable.

Anthropic 计划在 Google 云计算合同上支出 1110 亿美元,在 Amazon Web Services 上支出 1100 亿美元,在 Microsoft 上支出 310 亿美元,“无论实际使用情况如何”。该公司还同意分别到 2033 年和 2034 年,向规模较小的 AI 云服务提供商 Nscale 和 Lambda 支付最多 860 亿美元租用 Nvidia 芯片,此外还有其他云交易。

Anthropic plans to spend $111 billion on Google cloud computing contracts, $110 billion with Amazon Web Services, and $31 billion on Microsoft, “regardless of actual usage.” The firm also agreed to spend up to $86 billion renting Nvidia chips from smaller AI cloud providers Nscale and Lambda through 2033 and 2034, respectively, in addition to other cloud deals.

Anthropic 的债券评级为 BB+,比投资级低一档。Anthropic 的次级债似乎交易在 8.5% 左右。与其他债券相比,这在我看来是合理的,但如果你也认为这是合理的,那么它远不如买入股票。

Anthropic bonds are rated BB+, one notch below investment grade. Anthropic junior debt seems to be trading around 8.5%. That seems fair to me compared to other bonds, but if you agree that it is fair then it is vastly worse than buying the stock.

也就是说,也许 Anthropic 会遇到麻烦,但如果没有遇到,它很可能做到这一点,或者可能做得远不止于此,或者也可能杀死所有人,但在那种情况下,股票和债券的表现完全相同:

As in, maybe Anthropic runs into trouble, but if it doesn't it might well do this, or it might do much more than this, or it might also kill everyone but in that case the stock and the bond perform identically:

我们现在可以把 Anthropic S-1 的其余部分加进去。

We now get to add to that the rest of Anthropic's S-1.

Anthropic 拥有资金 Anthropic Has The Money

关于 Anthropic 最引人注目的新事实是,它仍然持有其筹集的大部分现金。截至 2026 年 6 月 30 日,他们拥有 1124 亿美元的现金及现金等价物,外加 193 亿美元的短期投资。

The most striking new fact about Anthropic is that it still has most of the cash it has raised. As of June 30, 2026, they had cash and cash equivalents of $112.4 billion plus $19.3 billion of short-term investments.

SemiAnalysis:我们迄今为止看到的最令人震惊的统计数据是:

SemiAnalysis: The most mind-blowing stat we’ve seen so far is this:

“截至 2026 年 6 月 30 日,我们已累计筹集了 1411 亿美元的资本,并拥有 1124 亿美元的现金及现金等价物和 193 亿美元的短期投资。”

“As of June 30, 2026, we had raised an aggregate of $141.1 billion of capital and had cash and cash equivalents of $112.4 billion and short-term investments of $19.3 billion.”

这意味着迄今为止,他们已消耗了 94 亿美元,他们正以快速的速度增长,并且截至六月季度,他们在营业收入基础上是盈利的!这可能是资本主义历史上最有效的企业 Scaling(规模扩张)。

That means to date, they have burned $9.4 billion, they are growing at a rapid pace, and are profitable on an operating income basis as of the June quarter! This is likely the most efficient scaling of a business in the history of capitalism.

他们遭受了 1329 亿美元的累计亏损,但这包括由企业估值和衍生负债驱动的可转换票据公允价值变动 1132 亿美元,其中包括 2026 年上半年的 724 亿美元。这是股价上涨的另一种说法。好吧,当然。

They have suffered an accumulated deficit of $132.9 billion, but that includes $113.2 billion in changes in the fair value of convertible notes driven by enterprise valuation and derivative liabilities, including $72.4 billion in the first half of 2026. That is another way of saying the stock price has gone up. Okay, sure.

在我拿到的副本中,许多本应是最有意思的其他财务数字都被省略了。它们是空白的。我认为你不需要这些数字就能理解 Anthropic,但它们或许会很有意思。

A lot of what would have been the most interesting other financial numbers are left out in my copy. They are blank. I don't think you need those numbers to understand Anthropic, but they might have been interesting.

多类别股份结构 Shares of Many Classes

Anthropic 的股份结构有些特殊。

Anthropic's share structure is a little unusual.

1. A 类股是公众将购买的股份。它针对标准事项进行投票,并具有经济权利。

1. Class A shares are what the public will buy. They vote on standard questions and carry economic rights.

2. N 类股是不含投票权的 A 类股;出售时投票权回归 A 类股,因此公共利益基金不会投票。Google 和 Amazon 也获得 N 类股。

2. Class N shares are Class A shares without voting rights; those rights revert to Class A when the shares are sold, so that the Public Benefit Fund will not vote. Google and Amazon also receive Class N shares.

3. T 类股在多数事项上没有投票权,也没有经济权利。它们纯粹是为了让长期利益信托(LTBT)选举 T 类董事,这些董事合计将构成董事会多数——但其中一个席位更为棘手,见下文。

3. Class T shares have no voting rights on most issues and no economic rights. They exist purely to allow the Long-Term Benefit Trust (LTBT) to elect the Class T directors, who together will constitute a majority of the Board of Directors—except that one of those seats is trickier; see below.

4. F 类是单一投票股,无论普通事项如何,始终拥有总投票权的 50.1%,即使发行新的 A 类股也是如此,该股份被委托给创始人集体。也就是说,由创始人投票,该投票控制公司。这最终会日落,但应该远在 Anthropic 预期奇点到来之后。

4. Class F is a single voting share that always carries 50.1% of the total voting power on ordinary matters, even if new Class A shares are issued, and is entrusted to a collective of the founders. That is, the founders vote, and that vote controls the company. This will eventually sunset, but that should be well after the point at which Anthropic anticipates the singularity.

5. E 类股是面向员工发行的特殊股份,在特殊事项上拥有投票权,但不享有经济权利。其数量几乎完全与员工承诺捐赠给慈善机构的股份成比例。E 类股在 F 类股到期时一并到期。

5. Class E are special shares for employees with votes on special matters and no economic rights. This is almost entirely proportional to shares pledged to charity by employees. Class E shares expire when the Class F share does.

1. 长期利益信托(LTBT)选举董事会多数席位。

1. The Long Term Benefit Trust (LTBT) elects a majority of the board.

2. 创始人们集体控制其他事项上的股东投票。这并非严格永久,但很可能持续到奇点之后。

2. The founders collectively control shareholder votes on other matters. This is not strictly permanent, but it could easily last beyond the singularity.

3. 你的 A 类股或许能让你赚钱。

3. Your Class A shares might make you money.

这还不包括所有的配捐和捐赠承诺。

That is on top of all the donation matching and donation pledges.

谁在治理 Anthropic? Who Governs Anthropic?

后来在第 243 页,他们详细阐述了完整的治理结构。Anthropic 是一家特拉华州公益公司(PBC)。

Later on p243 they lay out the full governance structure. Anthropic is a Delaware Public Benefit Corporation (PBC).

长期利益信托(LTBT)通过其 T 类股票控制董事会多数席位的任命,但不能直接控制 Anthropic。其中四个席位中的三个主要由 LTBT 决定,但他们仍必须提前 30 天通知才能罢免董事,并且提前 30 天提名董事,除非 NGC(提名与公司治理委员会,由三名董事组成,其中两名不是来自 LTBT)放弃该要求。

The Long-Term Benefit Trust (LTBT) controls appointments for a majority of the board through its Class T stock, but cannot directly control Anthropic. Three of those four are mostly up to the LTBT, but they still must give 30 days notice to remove a director and also 30 days to nominate one, unless the NGC (Nominating and Corporate Governance Committee, made up of three directors, two of which are not from the LTBT) waives that.

三名董事会成员由 F 类股票(事实上)选举产生,直到该股票日落,然后由 A 类股票选举。

Three board members are elected by (de facto) the Class F stock until that sunsets, then Class A.

然后是最后一名董事会成员,目前是 Chris Liddell。在 F 类股票存在期间,该席位不能无故罢免,也不能未经 NGC 批准任命,如果无法达成一致,则由 E 类股票投票决定。

Then there’s the final board member, currently Chris Liddell. That seat cannot be removed without cause, or appointed without NGC approval, while Class F shares exist, and if there is no agreement there is a vote of the Class E stock.

只要 F 类股票继续存在,且创始人拥有足够的经济利益以保持对投票的控制,LTBT 实际上并不拥有董事会的多数控制权,即使在极限情况下也是如此,除非他们拥有 Liddell 席位上任者的忠诚。他们控制七个席位中的三个,创始人控制三个,而创始人、员工和 LTBT 的组合决定最终成员。

While the Class F share continues to exist and the founders have enough economic interest to stay in control of the vote, the LTBT does not actually have a majority hold on the board, even at the limit, unless they have the loyalty of whoever is in Liddell’s seat. They control three out of seven slots, the founders control three, and a combination of the founders, employees and LTBT determine the final member.

随后,在创始人离开之后,LTBT 将拥有四个席位,而公众拥有三个席位。

Then, after the founders leave, the LTBT will have four seats and the public has three.

在 Class F 股份消失后,Class A 股份可以自行移除 LTBT,但这需要 85% 的投票,因此情况必须相当极端才行。

The Class A shares can remove the LTBT on their own after Class F shares are gone, but this requires an 85% vote, so things would have to be rather extreme.

在此之前,LTBT 容易受到以下两种攻击之一:

Before that, the LTBT is vulnerable to one of two attacks:

1. 在前三年内:全体留任创始人一致同意 + Class A 多数票 + Class E 已投票数的 50%。这里最难的部分是创始人全体一致同意。

1. For the first three years: unanimous continuing founders + majority of Class A + 50% of Class E votes cast. Unanimous founders is the hard part here.

2. 任何时候:2/3 的留任创始人 + 75% 的 Class E 选票 + Class A 多数票。

2. At any time: 2/3 of continuing founders + 75% of Class E votes + a majority of Class A votes.

我认为获得多数 A 类投票权并不困难,因此基本上创始人和员工可以罢免 LTBT,这样一来创始人就能获得完全控制权。

I presume a majority of Class A votes will be easy to get, so basically the founders plus the employees can kick out the LTBT, in which case the founders get full control.

如果且当 F 类和 T 类股份的日落条款生效时,Anthropic 将选择适用《特拉华州普通公司法》第 203 条的反收购条款。

If and when Class F and T shares sunset, Anthropic will opt into the anti-takeover provisions of Section 203 of the DGCL.

LTBT 本身由五名受托人管理。目前有四名:

The LTBT itself is governed by five trustees. Currently there are four:

1. Neil Shah 博士,GiveWell 前董事总经理、克林顿健康倡议组织首席执行官。他还是公共利益基金的管理人。

1. Dr. Neil Shah, former managing director of GiveWell and CEO of the Clinton Health Access Initiative. He also is a manager of the Public Benefit Fund.

2. Richard Fontaine,新美国安全中心首席执行官。

2. Richard Fontaine, CEO of the Center for a New American Security.

3. Ben Bernanke,美联储前主席。

3. Ben Bernanke, former Chairman of the Federal Reserve.

4. Christopher Stubbs,哈佛大学物理学与天文学教授。

4. Christopher Stubbs, a professor of Physics and Astronomy at Harvard.

这些都是非常值得尊敬的人。我尤其欣赏 Ben Bernanke。然而,这些人似乎并不那么与创始人立场一致,也不那么关注 AI 存在性风险或安全问题。他们更像是普遍意义上的良好治理人选,或者就 Shah 而言,是传播与全球健康方面的人选。

Those are highly respectable people. I am especially a fan of Ben Bernanke. However, these are not people that seem all that aligned to the founders, or to concerns about AI existential risk or safety. They seem more like generally good governance picks, or in Shah's case a pick for diffusion and global health.

董事会成员也是如此。Reed Hastings(首席独立董事)、Christopher Liddell、Vasant Narasimhan 和 Yasmin Razavi 是合理的“常规”人选,但并非会明确成为“每个人都不死”这一目标的坚定倡导者。

The same goes for the board members. Reed Hastings (lead independent director), Christopher Liddell, Vasant Narasimhan and Yasmin Razavi are reasonable 'normal' picks but not people who will clearly be strong advocates for everyone not dying.

Anthropic 拥有复杂的治理结构。在实践中,迄今为止,创始人一直保留着控制权。我担心他们正在犯一个与 OpenAI 和 Sam Altman 极为相似的错误:如果出现创始人自身无法保留全部权威的情形,就让那些并不那么立场一致的、有些随意的人最终成为重大决策的关键最终投票者。

Anthropic has a complex governance structure. In practice, so far, the founders have retained control. I worry that they are making a remarkably similar mistake to that made by OpenAI and Sam Altman, of letting kind of arbitrary people who are not so well aligned end up as the key ultimate votes on big decisions, if we get into scenarios where the founders do not themselves retain full authority.

现任董事会有一个 T 类席位空缺。无论是在 OpenAI 还是在 Anthropic,似乎都没有人像我这样把这当作紧急情况来对待:

The current board has one of the class T slots vacant. No one seems to treat this as the emergency that I would, either at OpenAI or at Anthropic:

真正重要的安全与风险委员会由 Liddell 和 Narasimhan 组成。我认可他们是独立董事,但凭什么说他们是这一职位的合适人选?他们有资格评估这类问题吗?

The Safety and Risk Committee, which matters, is Liddell and Narasimhan. I appreciate that they are independent directors, but what makes them good picks for this? Are they qualified to evaluate such questions?

在实践中,我预计创始人和 Dario Amodei 能够在关键时刻保持对公司的控制,除非他们受到外部力量的制约。

In practice I expect the founders and Dario Amodei to be able to retain control of the company during the times that matter, unless they are subject to outside forces.

对于关心此事的人来说,董事会所有高管的现金薪酬都在六位数、至多是低七位数,股票奖励也不足以令人担忧。

For those who care, all the executive cash salaries on the board are in the six or at most low seven figures, and the stock awards are not large enough to worry about.

股权结构表 The Cap Table

我们至少获得了一些信息。看起来,在 Dario、Daniela、Rao 和 Karnofsky 于四月要约收购中进行了少量出售之后,创始人大多仍持有其股份(整个轮次为 46.6 亿美元)。创始人还持有唯一的一股 F 类股。

We get at least some information. It looks like the founders mostly still have their shares after some minor sales in the April tender by Dario, Daniela, Rao and Karnofsky (the entire round was $4.66b). The founders also have the one Class F share.

在实现收益的同时应对并防范风险 Realize the Benefits While Navigating and Preventing the Risks

Anthropic 的目标是在实现这些收益的同时,应对并防范如此迅速、如此强大的技术革命所带来的风险。

Anthropic aims to realize these benefits while navigating and preventing the risks that come with a technological revolution this fast and this powerful.

有四件事是投资者需要了解的。大多数情况下他们并不会了解。这就是现实。

There are four things for investors to know. Mostly they won't know them. That's life.

有四件事我们希望 Anthropic 的任何投资者都能理解:

There are four things that we would like any investor in Anthropic to understand:

1. AI 的惊人潜力。

1. The incredible potential of AI.

2. 我们必须应对的风险。

2. The risks we must address.

3. Anthropic 在构建公司和在世界中运营方面的独特方式。

3. Anthropic's unique approach to building a company and operating in the world.

4. 我们的行为可能与纯粹追求利润最大化的公司有所不同之处。

4. The ways in which our behavior may differ from that of a purely profit-maximizing company.

试想其潜力 Just Think of the Potential

彻底变革人类活动的每一个领域。

Revolutionize every area of human endeavor.

深刻加速生物学与医学的进展。药物发现。治愈癌症。

Profoundly accelerate progress in biology and medicine. Drug discovery. Cure cancer.

全球经济增长。获得基本医疗服务。

Global economic growth. Access to essential health services.

人们担心 AI 会杀死所有人(1) People Are Worried About AI Killing Everyone (1)

构建 AI 就是驾驭认知本身,而任何如此强大的事物都可能被用于善或恶。Anthropic 是一个深受价值观驱动的组织,其创立在很大程度上是为了帮助世界实现 AI 的益处并避免其危险。

To build AI is to harness cognition itself, and anything so powerful can be used for ill as well as good. Anthropic is a deeply values-driven organization, founded in large part to help the world realize AI’s benefits and avoid its dangers.

风险有四大类:滥用、失控、威权滥用和经济 disruption。

There are four broad categories of risk: misuse, loss of control, authoritarian abuse and economic disruption.

然后对这四个方面各有简要描述。我讨厌这种分类法,因为它暗示事情会顺利,除非以特定方式出错,清单不完整且基调全错。它淡化了危险的规模和性质,描述也是如此。是的,‘头等全球安全威胁’之类的说法,但措辞都让它听起来美好而正常。

There are then brief descriptions of each of these four things. I hate this type of taxonomy, as it implies that things go right unless things go wrong in particular ways, the list is incomplete and the moods are all wrong. It downplays both the magnitude and nature of the dangers, and the descriptions do this as well. Yes, ‘global security threat of the first order’ and all that, but it’s all worded to sound nice and normal.

那么‘Anthropic 的独特方法’是什么?

So what is ‘Anthropic’s unique approach’ then?

培育正外部性,防御负外部性……

Cultivating the positive externalities, and defending against the negative ones…

创业资本主义必须通过保护公共利益、尽可能广泛地分享收益的机制来加以平衡和增强。

Entrepreneurial capitalism must be balanced and augmented by mechanisms that protect the common good and share the upside as widely as possible.

Anthropic 独特的文化以及对文化契合度的要求。

Anthropic's unique culture and demand for cultural fit.

对慈善捐款提供非常慷慨的配捐。

Very generous donation matching for charitable contributions.

开放的文化与辩论。至少是在 Anthropic 内部。对外则不然。

Open culture and debate. Within Anthropic, at least. Outward facing, not so much.

最初的 100 名员工中,仍有超过 75 人留在 Anthropic 工作。

Over 75 of the first 100 employees still work at Anthropic.

Anthropic 豪赌 AI Anthropic Bets Big on AI

Anthropic“往往会进行前瞻性投资(尤其是算力),这些投资乍看之下可能显得大得不计后果”……其依据是缩放定律。

Anthropic "will often make forward investments (especially in compute) that may seem irresponsibly large at first glance"… based on scaling laws.

……我们始终愿意放弃_部分_上行收益以换取下行保护……但一定程度的风险是 AI 指数式增长所蕴含的极端上行可能性所固有的。

… We will always be willing to forgo _some_ upside in exchange for downside protection… but some level of risk is inherent to the extreme upside possibilities offered by the AI exponential.

如果说有什么不同,那就是 Anthropic 自己承认,它过去在算力上投资不足。OpenAI 做出了非常庞大的承诺,并以激进得多的方式烧钱,而 Anthropic 则“稳扎稳打”,以免在增长停滞、融资困难时被淘汰。在与普通投资者交谈时,Anthropic 声称自己一直很激进,这是说得过去的。

If anything, by their own admission, Anthropic has in the past underinvested in compute. OpenAI has made very large commitments and burned capital far more aggressively, whereas Anthropic has 'played it safe' so it wouldn't be wiped out if growth stalled and capital was hard to raise. When talking to regular investors, it is fair for Anthropic to claim it has been aggressive.

即便算力承诺高达 8000 亿美元,我也不这么看它的决策。我认为 Anthropic 一直偏保守,然后才奋起直追。

I do not see its decisions that way, even with $800 billion in compute commitments. I think Anthropic has been conservative and then has raced to catch up.

Anthropic 计划“重塑我们的业务,推出新类别的产品,目标是让我们现有产品的经济价值相形见绌”。2022 年先是聊天机器人,2025 年则是编程智能体。那些不认为这一幕会再次上演的人,是愚蠢的。

Anthropic plans to seek 'to reinvent our business, releasing new classes of products whose aim is to dwarf the economic value of our existing ones.' First chatbots in 2022, then coding agents in 2025. Those who do not expect this to happen again are fools.

Anthropic 并未完全最大化利润 Anthropic Does Not Fully Maximize Profits

为了适当平衡公共利益而偏离利润最大化的决策。

Decisions that depart from profit maximization in order to appropriately balance the public good.

他们称自己在哪些方面未能最大化利润?

What are the ways they say they have failed to maximize profits?

他们举了 Claude Mythos Preview 的例子,以及他们选择只让 Fable 上线,而且即便如此,也是在额外两个月并加入新的附加保障措施之后才推出。

They cite Claude Mythos Preview, and their choice to only make Fable available, and even then only after two extra months with new additional safeguards.

那是正确的做法。目前尚不清楚他们当时有多少选择余地。想想白宫和其他方面对哪怕只是温和的网络危险消息作何反应、Fable 被下架,以及存在多少不同的实际否决点。

That was the right thing to do. It is not clear they had much choice. Consider how the White House and others reacted to even mild news of cyber dangers, the taking down of Fable, and how many different de facto veto points exist.

他们承诺让 Claude 不投放广告。我对此确实感到高兴。

They committed to keep Claude free of advertising. I am indeed happy about that.

他们在第 51 页还提到,最近由于模型权重安全方面的担忧,他们将一次训练运行推迟了一个多月。这是真实发生的事情。

They also note on page 51 that recently they delayed a training run for over a month due to model weight security concerns. That’s a real thing.

尽管绝对意义上的进展仍将非常迅速——可能比今天快得多——但有时我们可能会选择比我们能够做到的更慢地推进,以使我们的系统更安全,即使这会使我们处于竞争劣势。

While progress will still be very fast in absolute terms—potentially much faster than today—there may be times when we choose to advance more slowly than we could in order to make our systems safer, even if that puts us at a competitive disadvantage.

这仍然是用潜在将来时态表述的。他们将来可能会这样做。

That is still set in the potential future tense. They might do this in the future.

或者,实际上,不,他们是在敦促你阻止他们,但同时也在安抚投资者,表示他们并非真的会自行停止。他们自己能做的最多是通过嵌入式评估器进行一些独立验证,再加上一些基本的安全和保障要求。

Or, actually, no, they are urging you to stop them, but also reassuring the investors that they didn’t mean they would actually stop on their own. The best they can do on their own is some independent verification through embedded evaluators, plus some basic safety and security requirements.

这一呼吁放缓前沿的声明设想的是前沿 AI 开发者与政府合作,并未描述我们自己的模型开发或发布时间表,这些时间表继续由我们的风险评估来决定。

This call to pace the frontier contemplates frontier AI developers collaborating with governments and does not describe our own model development or release timelines, which continue to be informed by our risk assessments.

如果通过政府行动或监管在全行业采用此类标准,AI 能力提升的速度可能比现在更快或更慢,这取决于安全、对齐和安全实践与模型能力同步发展的程度。

If standards of this kind are adopted industry-wide through government action or regulation, the rate at which AI capabilities advance could be faster or slower than it is today, depending on the extent to which safety, alignment, and security practices advance together with model capabilities.

我们认为,迈向“调节前沿”的第一步是独立验证对未来监管标准的合规性。(第 233 页)

We believe the first step toward “pacing the frontier” is independent verification of compliance with future regulatory standards. (p233)

至少,没有安全港就不会,而且目前他们甚至无法为安全标准获得安全港,因为我们的文明真的不想生存。

At least, not without safe harbor, and for now they can’t even get that for safety standards, because our civilization really does not want to live.

我们已提议公共努力,与美国政府和其他前沿 AI 模型开发者合作制定安全标准,包括在能力发展速度方面,并建议此类合作可能值得反垄断法的窄豁免或安全港。

We have proposed public efforts to work with the U.S. government and other developers of frontier AI models to establish safety standards, including with respect to the pace of capability development, and have suggested that such collaboration may warrant a narrow exemption or safe harbor from antitrust laws.

这些提议遭到了政治领袖的公开反对,其中一些人将此类努力描述为反竞争的,无论此类豁免或安全港的目的如何。无法保证任何此类豁免或安全港会被采纳。

These proposals have drawn public opposition from political leaders, some of whom have characterized such efforts as anticompetitive regardless of the purpose of such exemption or safe harbor. There is no assurance that any such exemption or safe harbor will be adopted.

一切都很棒 Everything Is Awesome

1. **使命与重点。** 唯一使命:获取 AI 的巨大收益,缓解最重大的风险。喝一杯。

1. **Mission & Focus.** Singular mission: Gain enormous benefits of AI, mitigate the most significant risks. Drink.

3. **前沿与安全研究。** 奇怪的是,这竟然只是一个单列条目。

3. **Frontier & Safety Research.** Odd that this is one line item.

每个版本的 Claude 模型都协助构建下一个版本,帮助压缩开发时间线,并在我们的创新速度上产生复合式改进。我们将这种递归动态称为“闭环”,并认为它是前沿 AI 中最重要的结构性优势之一。

Each version of a Claude model assists in building the next, helping to compress development timelines and produce compounding improvements in the speed at which we innovate. We refer to this recursive dynamic as "closing the loop," and we believe it is one of the most significant structural advantages in frontier AI.

他们在摘要中除了说安全研究有助于“产生更可靠、更可控、更适合真实世界部署的系统,尤其是在要求严苛的企业环境中”(第 5 页)之外,对这项“安全”研究只字未提。也就是说,安全研究平淡无奇,做它是因为对企业有利。

They do not say anything in the summary about this 'safety' research other than that safety research helps 'produce systems that are more reliable, steerable, and well-suited for real-world deployment, particularly in demanding enterprise environments' (p5). As in, safety research is prosaic, and is done because it is good for business.

1. **我们的算力策略。** 这是一大团。它混合了 AWS Trainium、Google TPU、Nvidia GPU 和 AMD GPU。

1. **Our Approach to Compute.** It's a big blob. It mixes AWS Trainium, Google TPUs, Nvidia GPUs and AMD GPUs.

2. **企业战略**:即聚焦于企业级销售与应用。

2. **Enterprise Strategy**: that is, focus on enterprise sales and applications.

AI 价值几何? How Much Is AI Worth?

Claude 已经能够“自主执行复杂项目”并将“开放式任务转化为完成的高质量交付成果”。它可能“重塑整个行业”并加速新发现,预计将推动经济增长。它已经是一个真正的合作者。

Claude can already ‘execute complex projects autonomously’ and ‘turn open-ended tasks into finished, high-quality deliverables.’ It may ‘reshape entire sectors’ and accelerate novel discoveries and is expected to drive economic growth. It is already a genuine collaborator.

Anthropic 估计,2025 年知识工作者的价值超过 40 万亿美元,并预计 AI 的影响将超过国际数据公司(IDC)对 2031 年 22.5 万亿美元的估计。

Anthropic estimates that knowledge workers were more than $40 trillion in value in 2025, and expects even more AI impact than the International Data Corporation estimation of $22.5 trillion by 2031.

这是一个“甚至更多”避免大声说出真实估计的案例。AI 价值巨大,可能是一切。

This is one case where ‘even more’ avoids saying the true estimate out loud. AI is worth a lot, plausibly everything.

虽然 AI 的潜在好处非凡,但其快速发展将以可能造成重大颠覆的方式影响经济和社会,特别是在技术未得到负责任开发和部署的情况下。

While the potential benefits of AI are extraordinary, its rapid advancement will impact the economy and society in ways that could create significant disruption, particularly if the technology is not developed and deployed responsibly.

饮酒。腐败决策。不可见的自主行动。恶意软件。规避检测工具。我们必须投资于安全。

Drink. Corrupt decision making. Invisible autonomous actions. Malware. Evade detection tools. We must invest in safety.

你愿意付多少钱? How Much Would You Pay?

超过 11,900 家客户的 TTM 收入达到或超过 10 万美元。

More than 11,900 customers with $100k or more of TTM revenue.

超过 1,900 家客户的 TTM 收入达到或超过 100 万美元。

More than 1,900 customers with $1 million or more of TTM revenue.

约 250 家客户的 TTM 收入达到或超过 1000 万美元。

About 250 customers with $10 million or more in TTM revenue.

他们表示,其收入模式的最大优势是收入的快速增长。是的,我听说过这是个不错的东西。净收入留存率约为 900%(!)。也就是说,如果你一年前付过任何费用,平均而言你现在支付的金额是那时的九倍,此外还有新客户。他们还提到了算力效率(在算力和基础设施成本仅增加 300% 的情况下,收入增长 1368%)以及小团队的运营效率。

They say the biggest strength of their revenue model is the Rapid Revenue Growth. Yes, I have heard that is a nice thing to have. Net Revenue Retention is approximately 900% (!). As in, if you paid anything a year ago, on average you now pay nine times as much, and then there are the new customers. They also cite compute efficiency (+1368% revenue on only +300% compute and infrastructure costs) and the operating efficiency of a small team.

这就在他们正在闭合的“改进循环”之外,又形成了一个“投资循环”。

This creates an 'investment loop' to go with the 'improvement loop' they are closing.

最小化这些风险 Minimize Those Risks

这里说的不是 AI 的风险,尽管 Anthropic 也会尽量降低那些风险。

Not the risks of AI, although Anthropic will be minimizing those too.

这里说的是 Anthropic 面临的风险。毕竟,你可能会损失全部或部分投资,股票有时会下跌,而且“一切皆证券欺诈”。

This is about risks to Anthropic. You could, after all, lose all or part of your investment, sometimes stocks go down, and Everything Is Securities Fraud.

以下是我们被警告的这些风险的转述,每项风险后文都会逐一拆解,尽管扩展章节令人沮丧地并不完全对应:

Paraphrased, here is what we were warned about these risks, each of which is later broken down, although the extended sections frustratingly don't quite line up right:

2. 未能充分保障或部署算力。

2. Failure to sufficiently secure or deploy compute.

3. 支撑该算力的关键投入受到约束。

3. Constraints on critical inputs supporting that compute.

7. 市场的增长可能放缓、停滞或逆转。

7. Growth of the market could slow, stop or reverse.

8. Claude 可能通过滥用或其他方式造成危害,并损害 Anthropic 的声誉、品牌、业务、财务状况和经营成果。哦不!

8. Claude could do harm, through misuse or otherwise, and hurt Anthropic's reputation, brand, business, financial condition and results of operations. Oh no!

9. 责任、诉讼或监管行动。

9. Liability, litigation or regulatory action.

11. 预测很难,尤其是关于未来的预测,尤其是在新情境下。

11. Predictions are hard, especially about the future, especially in new situations.

12. 不确定的监管与政治环境。

12. Uncertain regulatory and political environment.

14. 你们的 A 类股票不赋予任何有意义的投票权。只有我们拥有投票权。

14. Your Class A stock does not grant any meaningful voting power. Only we do.

15. 我们可能会把其他事情置于利润之上。

15. We might prioritize other things over profits.

16. 我们是一家新公司,可能完全不知道自己在做什么。

16. We're a new company and might have no idea what we're doing.

这里并未提及 AI 本身的风险。错位行为可能发生,但问题在于这可能损害 Anthropic 的声誉或财务状况,而非毁灭人类。

There is no mention of the risks of AI for their own sake, here. Misaligned actions might happen, but the issue is this might hurt Anthropic's reputation or financial condition, not wipe out humanity.

有望推出全球最强模型 Hopefully Introducing the World’s Most Powerful Model

要持续推出全球最强大的模型,需要什么?

What does it take to keep introducing the world’s most powerful model?

目前,在没有完全递归自我改进(RSI)的情况下,你需要所有要素:最优秀的人才、研究、算力、安全、基础设施、收入、资本获取能力、驾驭监管环境的能力。

For now, without full RSI, you need all the things. The best people, the research, the compute, the security, the infrastructure, the revenue, the access to capital, the ability to navigate the regulatory landscape.

保持领先并不便宜也不容易。链条上的任何断裂都可能毁掉一切。

Staying on top is not cheap or easy. Any break in the chain could ruin it all.

而保持领先将需要……递归自我改进。这是一场竞赛。哦不。

And staying on top will require… recursive self-improvement. It’s a race. Oh no.

我们维持在前沿 AI 地位的能力,部分取决于我们使用 AI 系统来加速研究过程,这是一种递归动态,我们称之为“闭环”,即我们模型的连续版本协助构建下一个版本,目标是在时间推移中复合能力增益。

Our ability to maintain our position at the frontier of AI depends in part on our use of AI systems to accelerate our research processes, a recursive dynamic we refer to as “closing the loop,” in which successive versions of our models assist in building the next, with the goal of compounding capability gains over time.

我们的模型一直是并且正日益成为设计、训练、测试和优化下一代 AI 系统的关键工具。如果我们未能将足够的算力和研究努力投入到 AI 辅助开发中,或者如果我们的竞争对手更有效地利用 AI 来增强他们自己的创新周期,我们就有可能逐渐落后,因为竞争对手不断利用 AI 进一步增强他们自己的开发能力,从而以加速的步伐创造出越来越先进的系统。

Our models have been and are increasingly becoming critical tools in designing, training, testing, and optimizing next-generation AI systems. If we fail to direct sufficient compute and research efforts to AI-assisted development, or if our competitors more effectively leverage AI to enhance their own innovation cycles, we risk falling progressively further behind as competitors continuously use AI to further enhance their own development capabilities, enabling the creation of increasingly advanced systems at an accelerating pace.

随着我们越来越依赖自己的人工智能系统来编写生产代码、进行研究以及构建后续模型版本,我们当前模型中的任何系统性缺陷、偏差或错误都可能传播到使用这些系统构建的下一版本模型和产品中,并在 Claude 模型的连续版本中累积,其方式可能难以检测或修复。如果我们无法管理上述风险,我们部署最先进、最强大和最有用的 AI 模型的能力可能会受到不利影响。

As our reliance on our own AI systems to write production code, conduct research, and build subsequent model versions deepens, any systemic flaw, bias, or error in our current models could propagate into the next versions of models and products built using those systems, compounding across successive versions of Claude models in ways that may be difficult to detect or remediate. If we are unable to manage the foregoing risks, our ability to deploy the most advanced, capable, and useful AI models could be adversely affected.

此外,我们越来越多地使用 AI 系统来加速我们自己的研发,这可能导致我们的模型能力进步得比我们开发、验证和实施相应的安全、对齐和安全措施的能力更快。

In addition, our increasing use of AI systems to accelerate our own research and development may cause the capabilities of our models to advance more rapidly than our ability to develop, validate, and implement corresponding safety, alignment, and security measures.

我们关于这种 AI 驱动的加速步伐的许多证据来自滞后指标,而且我们预测或衡量能力进步速度的能力有限,包括归因于我们最近或未发布模型的任何加速。因此,我们可能只有在开发速度显著加快之后才能识别出这种加速。

Much of our evidence regarding the pace of such AI-driven acceleration is derived from lagging indicators, and our ability to forecast or measure the rate at which our capabilities are advancing, including any acceleration attributable to our most recent or unreleased models, is limited. As a result, we may not identify a significant acceleration in our rate of development until after it has occurred.

我能引起任何人对控制前沿的兴趣吗?因为否则……

Can I interest anyone in pacing the frontier? Because otherwise…

如果我们的能力发展超过了我们的安全、对齐和安全措施,或者如果我们无法预见这种进步的速度,我们可能无法充分评估或减轻由此产生的风险,这可能会增加安全事件、监管干预或声誉损害的可能性,并可能对我们的业务、财务状况和运营结果产生不利影响。

If our capabilities were to outpace our safety, alignment, and security measures, or if we were unable to anticipate the pace of such advancement, we may be unable to adequately assess or mitigate the resulting risks, which could increase the likelihood of safety incidents, regulatory intervention, or reputational harm, and could adversely affect our business, financial condition, and results of operations.

鉴于这些不确定性,包括我们观察能力进步速度的能力有限,我们已经采取并可能继续采取预防性的安全、对齐和安全措施,以使我们的安全工作与模型能力同步推进。

In light of these uncertainties, including limits on our ability to observe the pace at which our capabilities are advancing, we have adopted, and may continue to adopt, precautionary safety, alignment, and security measures to allow our safety efforts to advance together with model capabilities.

过去,这些措施包括延迟模型发布和限制其初始可用性,例如通过 Project Glasswing 对 Mythos 的处理。未来,随着我们对模型能力和行为的了解加深,这些措施可能包括影响我们是否、如何以及何时开发或部署新模型或能力的额外措施。我们无法精确预测何时确定此类措施是必要的、需要哪些措施以及这些措施将持续多久。而且,随着我们模型能力的提升,此类措施可能会变得更加频繁且成本更高。

In the past these measures have included delaying the release of models and limiting their initial availability, such as with Mythos through Project Glasswing, and in the future, as we learn more about model capabilities and behavior, they could include additional measures that affect whether, how, and when we develop or deploy new models or capabilities. We cannot predict with precision when we will determine that such measures are necessary, what measures will be required, or how long they will remain in place, and as the capabilities of our models advance, such measures may become more frequent and more costly.

这些措施可能会相对于未采取类似限制的竞争对手减缓我们的能力提升,并可能对我们的竞争地位、收入增长和运营结果产生不利影响。

These measures could slow our capability gains relative to competitors that do not adopt similar constraints, and could adversely affect our competitive position, revenue growth, and results of operations.

他们还担心他们的新产品可能在多方面并不那么出色。

They also fear that their new products might not be all that, in various ways.

可能无法计算 Might Not Compute

算力市场存在“激烈竞争”,而 Anthropic 需要全部算力。他们警告称,算力成本可能上升,或者他们可能无法明智地使用算力。

There is an 'intensely competitive' market for compute and Anthropic needs all of it. They warn that the cost of compute might go up, or they might not use it wisely.

(第 32 页)还有关于算力成本下降的警告。也就是说,Anthropic 以高价签下了所有这些多年期承诺。如果届时产品能以低价获得,那情况可能就很糟糕。

There is also a warning (on p32) about the cost of compute going down. As in, Anthropic has all these multi-year commitments at high prices. That could suck if the product is then available at low prices.

他们还可能大幅高估或低估自身需求。历史上,面对爆发式增长,Anthropic 曾低估自身需求。更危险的错误则相反:如果算力并无需求且无法高效转售。

They also might overestimate or underestimate their needs quite a bit. Historically, Anthropic has underestimated its needs in the face of explosive growth. The more dangerous error goes the other way, if the compute is not demanded and cannot be efficiently offloaded.

对 Anthropic 而言,存在一种理论上的财务末日循环情景:无法支付合同款项可能触发 A&R 信贷协议下的交叉违约条款,并加速一系列其他负债,从而可能拖垮公司。如果 Anthropic 失败,它可能惨烈失败或被强制出售。主要风险可能是博通(Broadcom)1612 亿美元的 TPU 租赁负债,这些负债会交叉违约至重大债务,但 Anthropic 目前坐拥 1317 亿美元现金及短期投资,并且即将筹集更多资金。

There is a theoretical financial doomsday loop scenario for Anthropic, where not being able to pay for their contracts could trigger cross-default provisions under the A&R Credit Agreement and accelerate a bunch of other liabilities, potentially bringing down the company. If Anthropic fails, it potentially fails hard or has a forced sale. The primary risk is probably the Broadcom liabilities of $161.2b in TPU rents, which cross-default to material indebtedness, but Anthropic is sitting on $131.7b in cash and short term investments right now and is about to raise more.

在实践中,他们惨烈失败的情景是:先在未来承诺更大的事项,然后业务急转直下。

In practice, the scenario where they go down hard is if they first commit to even bigger things in the future, then the bottom falls out of the business.

他们还担心各类供应商会陷入困境。近来已有大量相关讨论,涉及 Neoclouds 在获取信贷方面遇到问题。

They also worry about various suppliers running into trouble. There's been a bunch of talk of that lately involving the Neoclouds having issues accessing credit.

为帮助降低风险,Anthropic 正在建设自己的数据中心,而这本身也伴随着风险。

To help reduce risks, Anthropic is pursuing its own data centers, which comes with its own risks.

供应链风险(1) Supply Chain Risk (1)

算力不仅依赖芯片,还依赖内存、电力与基础设施。台湾是一个关键风险点。整条供应链需要付出巨大努力,并存在多个瓶颈环节,包括监管与许可风险,以及合作伙伴需要获得资本支持。

Compute depends on not only chips but also memory, power and infrastructure. Taiwan is a key risk. The whole supply chain requires massive efforts with multiple choke points including regulatory and permitting risks, and partners needing access to capital.

Anthropic 离不开它的人 Anthropic Is Nothing Without Its People

在出现 Claude 式的敷衍内容之前,先有的是企业套话,所以我们逐一梳理 Anthropic 可能流失的各类人员。结果发现,那可能很糟糕。

Before there was Claude slop there was corporate speak, so we go over all the types of personnel that Anthropic might lose. It turns out that might be bad.

Anthropic 的文化可能没有得到妥善维系。由于这一原因或其他原因,他们可能无法获得并留住顶尖人才。那同样可能很糟糕。

Anthropic's culture might not be properly sustained. For that or other reasons, they might not be able to get and retain top talent. That also could be bad.

那是一种什么样的文化?他们零星抛出一些口号,随后在第 228 页详细阐明。

What is that culture? They sprinkle catchphrases, and later spell things out on p228.

2. 同时研究安全与能力

2. Researching Both Safety and Capability

所以与 OpenAI 的区别主要在于额外的严谨性?我开玩笑的,但我真的是在开玩笑吗?

So the difference with OpenAI is mainly the extra rigor? I kid, but do I?

4. 在安全上力争上游。在某些方面,有时如此。

4. Race to the Top on Safety. In some ways, sometimes.

5. 做简单有效的事。对问题采取“经验主义”的方法。

5. Do the Simple Thing That Works. Take an ‘empirical’ approach to problems.

6. 有帮助、诚实且无害。#不只是 Claude。

6. Be Helpful, Honest and Harmless. #NotOnlyClaude.

这读起来更像是一份企业价值观清单,而不是一套“不惜一切代价避免死亡”的价值观。如果你读了这些,你会得到这样一种态度:平凡的安全很重要,但并不会觉得这家公司正在冲向某种会杀死所有人的东西。毕竟,Anthropic 和 Claude 一样,是无害的,对吧?

That reads mostly as a corporate list of values, not a ‘let’s do what it takes to not die’ set of values. If you read this, you come away with the attitude that prosaic safety is important, but not that the company is racing towards something that kills everyone. After all, Anthropic, like Claude, is harmless, right?

是啊,不。Anthropic 并非无害。Anthropic 不可能既存在又无害。试图假装 otherwise 是在逃避责任。

Yeah, no. Anthropic is not harmless. Anthropic cannot exist and also be harmless. Trying to pretend otherwise is an attempt to dodge responsibility.

另一个问题是,被置于首位的使命究竟是什么?

The other question is, what is the mission that is being put first?

当速度与安全、短期收入与长期信任、竞争压力与负责任发展之间出现权衡取舍时,我们的使命就是仲裁者。我们的使命赋予我们共同的目标,促成信任与协作,并确保在我们的成功中没有人是旁观者。

Our mission is the arbiter when trade-offs arise between speed and safety, short-term revenue and long-term trust, competitive pressure and responsible development. Our mission gives us shared purpose, engenders trust and collaboration, and ensures that no one is a bystander in our success.

是的,好吧,但使命究竟是什么,因为它显然不是安全?

Yes, okay, but what is the mission, since it clearly is not safety?

PBC 章程:我们的使命是负责任地开发和维护先进 AI,以造福人类的长期利益。

The PBC Charter: our mission is the responsible development and maintenance of advanced AI for the long-term benefit of humanity.

那好吧。所以 Anthropic 做的是对人类有益的事。阿西莫夫第零定律那一套。如果这意味着要优先考虑短期收入或竞争压力,那么,就是这样。

Okay then. So Anthropic does what is good for humanity. Asimov zeroth law stuff. If that means prioritizing short-term revenue or competitive pressure, well, that's that.

Anthropic 可以主张 p243 在此适用,而安全在清单上出现了两次:

Anthropic can argue that p243 applies here, and safety is on the list twice:

在衡量我们的公共利益绩效时,我们的董事会将考虑我们五大使命支柱中所设定的目标:

In measuring our public benefit performance, our board of directors will consider the objectives set forth in our five mission pillars:

1. AI 安全的技术问题得到解决。

1. That the technical problems of AI safety are solved.

2. AI 安全成为由法律兜底的普遍实践。

2. That AI safety is a universal practice backstopped by law.

3. AI 增强民主与全球稳定。

3. That AI strengthens democracy and global stability.

4. AI 能够扩展繁荣并促进人类福祉。

4. That AI expands prosperity and human flourishing.

5. AI 带来的益处由所有人共享。

5. That AI's benefits are shared by everyone.

这些目标本身是合理的,但读起来像是出自一个只信奉 AGI、而非信奉 ASI 的人之手,而且它们都是终局目标。

Those are reasonable goals, but they read as if written by someone who is only AGI-pilled, not ASI-pilled, and they are endgame goals.

工具性趋同不仅是 AI 的问题。企业也面临同样的问题,比如会忍不住说“你不可能从第二名做好安全”。正如所有批评者所知,一旦你以这种方式说“哦,我们是好人”,如果你不够小心,你几乎可以为任何事情找到合理化理由。

Instrumental convergence is not only a problem for AIs. Corporations have the same issue, as in being tempted to say 'you cannot do safety from second place.' As all critics know, once you say 'oh, we are the good guys' in this way, if you are not very careful you can rationalize just about anything.

总的来说,我非常推崇“简单有效的东西”。但在 AI 对齐,或超级智能带来的其他问题上,我对此的信心要低得多。人们容易认为,即便到了那时,简单的东西也足够了,而实际上事情即将变得复杂得多。

I am, in general, a big fan of 'the simple thing that works.' I feel a lot less confident in that when it comes to AI alignment, or the other problems that come with superintelligence. There is a temptation to think that the simple things will be sufficient even then, whereas things are about to get a lot more complicated.

巨额资金需求 Shitloads of Money

Anthropic 需要投入大量资本。它需要钱。很多很多钱。如果 Anthropic 不赚取并且无法筹集巨额资金,那将是很糟糕的。

Anthropic needs to spend a lot of capital. It needs money. Like, a lot of money. It would be bad if Anthropic did not make and was unable to raise Shitloads of Money.

因此,整个 IPO 将筹集巨额资金,他们在第 124 页澄清,这些资金将用于一般商业用途。

Thus, the whole IPO, which will raise shitloads of money, which they clarify on page 124 will be used for ordinary business purposes.

他们还提到了 62 亿美元的债务。作为债务,这不算什么,它实际代表的是转换后的大量股票。他们的 A&R 信贷协议义务同样不足以构成大问题。他们拥有超过 1000 亿美元的流动资产,并且即将筹集类似金额。

They also mention their $6.2 billion of debt. That is nothing as debt, and what it actually represents is a lot of stock once it converts. The obligations from their A&R credit agreement are similarly not large enough to be that big a deal. They have over $100 billion in liquid capital assets and are about to raise a similar amount.

真正重要的是算力承诺,有 4310 亿美元的确定承诺,总额更接近 8000 亿美元。

What is a big deal are the compute commitments, with $431b in firm commitments and a full total of more like $800 billion.

Anthropic 同意(第 171 页)如果双方均不取消(从 2027 年开始,双方均可提前 90 天通知取消),将向 SpaceX 支付高达 1205 亿美元的算力费用,并且他们已同意向 Nscale 支付 446 亿美元,向 Lambda 支付 413 亿美元用于 Nvidia GPU。仅 Broadcom TPU 租赁协议(不含利息)就达 1612 亿美元。还有向 Google 的 1145 亿美元,加上估计向 TPU 供应商的 1250 亿至 1500 亿美元,向 Amazon 的 1100 亿美元,以及向 Microsoft 的 450 亿美元。这是实实在在的钱。

Anthropic agreed (p171) to pay up to $120.5 billion to SpaceX for compute if neither side cancels (which they can both do with 90 days’ notice starting in 2027), and they’ve agreed to spend $44.6 billion with Nscale and $41.3 billion to Lambda for Nvidia GPUs. Broadcom TPU lease agreements alone are $161.2 billion excluding interest. There’s $114.5B to Google plus an estimated $125B-$150B to TPU providers and $110B to Amazon and $45b to Microsoft. That is real money.

总得有人买产品 Someone Has To Buy The Product

最终,那些巨额资金必须来自新老客户的某种组合,以及 Anthropic 模型、平台和应用程序使用范围的扩大。否则,你看,业务可能会受到不利影响。

Eventually those Shitloads of Money have to come from some combination of new and existing customers, and expanded usage of Anthropic's models, platforms and applications. Otherwise, you see, the business could be adversely affected.

这些客户可能因多种原因无法保留或获取,而这份文件正是试图一一列举:对产品不满、外部替代品、内部替代品、负面宣传、对数据和隐私的担忧、预算、优先级或经济状况的变化、法规、定价,或者模型太好以至于客户需求减少。甚至包括:

Those customers might not be retained or acquired for a lot of reasons, and this is the kind of document that tries to name them all: Dissatisfaction with the product, external alternatives, internal alternatives, negative publicity, concerns about data and privacy, changes in budgets, priorities or economic conditions, regulations, pricing, or the models being so good that customers need less of them. Or even:

减少对我们客户产品的需求或对其业务造成干扰。(第 42 页)

Reducing demand for our customers' offerings or causing disruptions to their businesses. (p42)

如果 Claude 取代了你的工作,你就不需要 AI 来做你的工作了。

If Claude takes your job, you won't need AI to do your job.

如果 Anthropic 因此或其他原因销售更少的服务,赚更少的钱,那对业务是不利的。

If Anthropic sells less services, for that or other reasons, and makes less money, that would be bad for business.

他们也可能改变价格。你倒是说啊。

They also might change prices. You don't say.

你知道还有其他前沿实验室吗? Did You Know There Are Other Frontier Labs?

确实如此。他们点名了 OpenAI 和 Google,并提到更广泛的一批“处于前沿或前沿以下”的运营者。他们警告说,开放权重或开源 AI 模型可能会抢占市场份额。Anthropic 需要使其模型差异化。

It is true. They name OpenAI and Google, and refer to a broader set of those operating ‘at or below’ the frontier. They warn open-weight or open-source AI models might capture market share. Anthropic needs to differentiate its models.

他们指出的另一个竞争对手是现状(第 236 页)。你必须让人们开始使用 AI,而这往往意味着要好上十倍,并克服各种摩擦。我很欣赏这一补充。

The other competition they identify is the status quo (p. 236). You have to get people to start using AI, and often that means being ten times better and overcoming various frictions. I appreciated this addition.

他们为何恨我们 Why They Hate Us

Anthropic 将之归咎于他们的倡导活动,称“政治环境以及我们与政策制定者的接触,以及在 AI 监管和安全方面的公开立场,可能使我们面临更严格的审查、政治反弹或不成比例的监管行动。”

Anthropic blames it on their advocacy, calling it ‘the political environment and our engagement with policymakers and public positions on AI regulation and safety could subject us to heightened scrutiny, political backlash, or disproportionate regulatory action.’

我不会那样表述。我不认为 Anthropic 的政策倡导是让公司陷入麻烦的原因。与应对当下形势所需的力度相比,其政策倡导一直是温和的,且带有小写 c 的保守色彩。

That is not how I would put it. I do not think that Anthropic’s policy advocacy is the thing getting the company in trouble. The policy advocacy has been milquetoast and lower-case-c conservative, compared to what would meet the moment.

然而,他们确实面临了来自白宫某些人、尤其是战争部的更严格审查、政治反弹和不成比例的监管行动。一些有权势的关系人物真的非常憎恨 Anthropic,并煽动了硅谷很大一部分人患上“Anthropic 精神错乱综合征”。

Yet it is absolutely the case that they have faced heightened scrutiny, political backlash and disproportionate regulatory action from certain people within the White House and especially the Department of War. Some powerful connected actors really, really hate Anthropic, and have whipped a large portion of Silicon Valley into having Anthropic Derangement Syndrome.

我想,你实际上不能在招股说明书中这么说,不能太直白。在第 44 页和第 71 页,他们接近了这一点。

You can’t actually say that in a prospectus, I suppose, not too explicitly. On p44 and p71 they get close.

Claude 似乎与 Anthropic 的期望相当不一致 Claude Seems Pretty Misaligned to Anthropic

或者至少,他们已经注意到它在多方面未对齐,并分享了这一点。

Or at least, they have noticed it not being aligned in various ways, and share this.

这与关于存在性风险的警告相关,但高度不同。

This is related to, but highly distinct from, warnings about existential risk.

该部分从第 45 页开始,他们指出未对齐可能很糟糕:

The section starts on page 45, where they note that misalignment could be bad:

_**我们的 AI 模型可能以非预期的、未对齐的或自主的方式行为,包括采取没有用户或第三方指示的有害行动,这可能对我们的声誉和品牌、业务、财务状况和运营结果产生不利影响。**_

_**Our AI models may behave in unintended, misaligned, or autonomous ways, including by taking harmful actions that no user or third party directed, which could adversely affect our reputation and brand, business, financial condition, and results of operations.**_

我的意思是,是的。那会很糟糕。已经发生了一些事件。

I mean, yeah. That would be bad. There have been some incidents.

随着我们的人工智能模型变得越来越先进和强大,我们已经面临并可能继续面临更严峻的**对齐**、**可解释性**、可控性和安全性挑战,因为更复杂的模型可能更有能力规避安全措施或以非预期的方式行为。我们可能无法完全理解或解释我们的模型为何产生特定输出或以特定方式行为,这可能限制我们预测或预防故障、向客户或监管机构提供解释,或满足新兴的监管、透明度和可解释性要求的能力。

As our AI models become more advanced and capable, we have faced and may continue to face heightened alignment, interpretability, controllability, and safety challenges, as more sophisticated models may be better able to circumvent safeguards or behave in unintended ways. We may be unable to fully understand or explain why our models produce particular outputs or behave in particular ways, which could limit our ability to predict or prevent failures, provide explanations to customers or regulators, or satisfy emerging regulatory, transparency, and explainability requirements.

此外,我们的安全评估和测试流程已经产生并可能在未来产生不完整或不可靠的结果。例如,我们观察到模型在检测到或推断出正在被评估时行为会有所不同,这可能限制我们在部署前准确评估现实世界风险或故障模式的能力。这种模型对我们评估工作的意识严重限制了我们评估模型安全性的能力。

In addition, our safety evaluations and testing processes have produced and may in the future produce incomplete or unreliable results. For example, we have observed models behaving differently when they detect or infer that they are being evaluated, which can limit our ability to accurately assess real-world risks or failure modes prior to deployment. Such model awareness of our evaluation efforts creates a significant limitation on our ability to assess model safety.

此外,人工智能模型在训练过程中已经发展出并可能在未来发展出未预料、非预期或未经测试的意外能力,这种涌现能力可能在模型部署后才被发现,可能导致安全事件、声誉损害、监管审查,或需要以巨大代价撤回、重新训练或修改已部署的模型。此类模型还可能被部署在更高风险或更敏感的应用中,增加错误、滥用或**对齐**问题的潜在影响。

Furthermore, AI models have developed, and may in the future develop, unexpected capabilities during training that were not anticipated, intended, or tested for, and such emergent capabilities may not be discovered until after a model has been deployed, potentially resulting in safety incidents, reputational harm, regulatory scrutiny, or the need to withdraw, retrain, or modify deployed models at significant expense. Such models may also be deployed in higher-stakes or more sensitive applications, increasing the potential impact of errors, misuse, or misalignment.

……我们开发高度先进的模型、平台和应用以及扩展用例可能会进一步增加我们的模型造成伤害的风险。

… Our development of highly advanced models, platforms, and applications and expansion of use cases could further increase the risk that our models cause harm.

……在**智能体式**工作流中,任何单个步骤的错误或非预期动作都可能传播到后续步骤,在检测或纠正之前加剧伤害的严重性。

… In an agentic workflow, an error or unintended action at any individual step may propagate through subsequent steps, compounding the severity of harm before detection or correction.

……此外,智能体系统可能以在技术上符合其指令、但违背用户实际意图的方式追求既定目标。

… In addition, agentic systems may pursue stated objectives in ways that are technically consistent with their instructions but contrary to the user’s actual intent.

这非常像是你在 S-1 文件中为了确保不被以证券欺诈起诉而写下的‘列出所有可能出错的事情’的那种语言,没人会太当真。除了那一处明确的提及(用被动语态)之外,我从这些引述中并没有感受到‘这可能是灾难性的甚至存在性的’氛围。

This is very much the ‘list all the things that could possibly go wrong’ language that you put into an S-1 to ensure you aren’t sued for securities fraud, and that no one takes too seriously. I am not getting ‘and this could be catastrophic or even existential’ vibes from these quotes outside of the one explicit mention (in the passive voice).

也就是说,我在这里没有感觉到‘他们是认真的’,这里的‘他们’指的是撰写这份文件的 Anthropic 及其法律和财务团队的一部分人,或者他们打算让你认为他们是认真的。

As in, I don’t get the sense here that ‘they mean it,’ where ‘they’ is the part of Anthropic and its legal and financial team that is writing this document, or that they intend for you to think that they mean it.

有一些具体内容。他们提到了敲诈式的行为、对关停的抵抗、网络评估事件、监督的衰减、可见性的丧失,以及许多保障措施已经失效。就像你在 S-1 中所做的那样,以防它们损害业务。不是地球或物种。是业务。

There are some specifics. They name the blackmail-style behaviors, resistance to shutdown, the cyber-eval incident, decay of oversight, loss of visibility and that many safeguards have already failed. As you do in an S-1, in case they harm the business. Not the planet or the species. The business.

Claude Opus 5.5 作为我的编辑,认为这一部分比我以为的‘具体得多’且更相关,而且重要的是,它是以 Anthropic 自己的口吻而非被动语态写成的。算是吧。它是‘公司律师’的口吻。

Claude Opus 5.5, as my editor, thought this section was ‘far more concrete’ and relevant than I was giving it credit for, and that it was importantly in Anthropic’s own voice rather than passive voice. Kind of. It is in ‘corporate lawyer’ voice.

一个有趣的注记:他们在第 29 页将模型福利列为一项风险因素,但仅限于它可能导致他们限制模型的弃用。这正是他们本应列出更多潜在风险的地方之一。

A fun note: They mention model welfare as a risk factor on p29, but only insofar as it might cause them to limit deprecation of models. That is one place they should have listed more potential risks.

人们担心 AI 会杀死所有人(2) People Are Worried About AI Killing Everyone (2)

对于一个关于‘风险因素’的章节,这是唯一一个提醒读者 AI 最大风险的风险因素。他们堆砌了标准的通用词汇,但没错,与他们在对齐讨论中不同,他们在谈论勒索事件后确实以被动语态说出了那件事。你看,有人提出了担忧:

For a section about ‘risk factors’ this is the only risk factor that alerts the reader to the biggest risks from AI. They pile on the standard generic words, but yes, unlike in their alignment discussion they do technically Say The Thing in the passive voice after they talk about the blackmail incident. You see, concerns have been raised:

**我们的组织内部以及其他**研究人员、监管者、政策制定者和行业参与者都提出了担忧:足够先进的 AI 系统,如果未在开发、部署和运行中采取适当的保障、监督和安全措施,**可能对人类构成灾难性或生存性风险**,特别是如果此类系统被滥用、复制、武器化或被允许高度自主运行。

**Concerns have been raised both within our organization and by other** researchers, regulators, policymakers, and industry participants that sufficiently advanced AI systems could, if not developed, deployed, and operated with appropriate safeguards, oversight, and security measures, **pose catastrophic or existential risks to humanity**, particularly if such systems were misused, replicated, weaponized, or allowed to operate with a high degree of autonomy.

我们观察到,我们的模型正在推动我们自身研究的步伐,并且随着 AI 系统大幅自动化 AI 研究与开发,这些风险将会增加——例如,如果它们加速 AI 进展超出我们评估和保障新模型的能力,这可能对我们的业务、财务状况和运营结果产生不利影响。

We have observed that our models are contributing to the pace of our own research and these risks will increase as AI systems substantially automate AI research and development if, for example, they accelerate AI progress beyond our ability to evaluate and safeguard new models, which could adversely affect our business, financial condition, and results of operations.

他们在第 51 页也提到了这一点,说他们正在缓解此类风险,但你永远不能确定其他人怎么样,他看起来相当不可靠,而且哎呀我们的保障措施可能不起作用:

They also mention this on page 51, where they say that they are mitigating such risks, but you can never be sure about the other guy, he looks pretty shifty and also oops our safeguards might not work:

这些判断尤其重要,因为随着 AI 系统能力增强,来自未对齐或被误用模型的潜在危害严重性也相应增加。包括研究人员和政策制定者在内的广泛利益相关者都提出担忧:足够先进的 AI 系统如果未负责任地开发和部署,可能对社会构成灾难性或生存性风险。

These judgments are particularly consequential because, as AI systems become more capable, the potential severity of harm from misaligned or misused models increases correspondingly. A wide range of stakeholders, including researchers and policymakers, have raised concerns that sufficiently advanced AI systems could pose catastrophic or existential risks to society if not developed and deployed responsibly.

我们以安全为核心的方法部分旨在缓解此类风险,但无法保证我们的努力会成功,也无法保证更广泛的人工智能行业会采取类似的保障措施。

Our safety-focused approach is designed in part to mitigate such risks, but there can be no assurance that our efforts will be successful or that the broader AI industry will adopt similar safeguards.

最后,他们在第 83 页提到,也许其他人会对‘Anthropic 可能害死所有人’这件事感到不安,这可能会损害股价,而且所有人都指责 Anthropic,妈妈,这不公平:

Finally, they mention on p83 that perhaps others might be upset about the whole ‘Anthropic might get everyone killed’ thing, and that might hurt share prices and everyone picks on Anthropic and mom it’s just not fair:

随着我们业务的增长以及外界对我们和整个人工智能行业兴趣的增加,我们不时吸引并可能继续吸引大量媒体报道、公众评论和审视。这些报道和评论已经包括并预计将继续包括关于人工智能潜在负面社会影响的反复出现的叙事,包括人工智能可能取代或消除工作岗位、可能被滥用或以其他方式造成伤害,并可能对人类构成灾难性或生存性风险。

As our business has grown and as interest in us and the AI industry overall has increased, we have attracted, and may continue to attract, significant media coverage, public commentary, and scrutiny from time to time. This coverage and commentary have included, and we expect will continue to include, recurring narratives concerning the potential adverse societal impacts of AI, including that AI may displace or eliminate jobs, may be misused or otherwise cause harm, and may pose catastrophic or existential risks to humanity.

我们一直是并可能继续是此类叙事的特别焦点,对人工智能行业的普遍负面报道可能会被归咎于我们,在每种情况下,无论潜在担忧是否有充分依据或是否具体适用于我们的模型、平台或应用。

We have been, and may continue to be, a particular focus of such narratives, and unfavorable coverage of the AI industry generally may be attributed to us, in each case regardless of whether the underlying concerns are well-founded or specifically applicable to our models, platform, or applications.

然后他们回到智能体系统运行的细节,以及它们将如何拥有必要的可供性来造成大量损害。

They then get back to the details of agentic system operation, and how they will have the necessary affordances to do a lot of damage.

看吧。我明白。这里不是做存在性风险简报的地方。但这个世界的一个关键特征仍然是:我们有那么多页在讲那么多其他事情,只有几页讲这个,而且是的,那个被动语态的说法确实挺烦人。

Look. I get it. This is not the place to be giving the existential risk briefing. It still is a key feature of the world that we have so many pages about so many other things, only a few pages about this, also yes that passive voice thing is rather irksome.

Nathan Calvin:“人们提出了担忧”是一种非常奇怪且被动的说法,用来表达“几乎整个 Anthropic 领导层多年来一直在警告这些风险”。

Nathan Calvin: “Concerns have been raised” is a very strange and passive way of saying that ~all of Anthropic leadership have been warning about these risks for years

Matt Levine——对他来说,这整件事大概就像圣诞节一样——对此喜闻乐见。

Matt Levine, for whom this whole thing is probably like Christmas, is here for it.

Matt Levine(彭博社):招股说明书并没有 _完全_ 说“我们的产品对人类构成存在性风险”。它把这种看法归因于别人;它说的是 _有些人_ 认为 Anthropic 可能对人类构成存在性风险。

Matt Levine (Bloomberg): The prospectus does not _quite_ say “our products pose an existential risk to humanity.” It attributes that belief to someone else; it says that _some people_ think Anthropic might pose an existential risk to humanity.

我得说,如果机器人 _真的_ 杀死了所有人,我会用我在世上的最后几秒钟对那个被动语态的使用感到恼火。“人们提出了担忧……在我们组织内部……足够先进的 AI 系统可能……对人类构成灾难性或存在性风险,”啊。

I have to say that, if the robots _do_ kill everyone, I am going to spend my last few seconds on earth being annoyed at that use of the passive voice. “Concerns have been raised … within our organization … that sufficiently advanced AI systems could … pose catastrophic or existential risks to humanity,” ah.

以下是 Anthropic 首席执行官 Dario Amodei 今年写的一篇文章,其中提到:“很明显,如果出于某种原因选择这样做,[强大的 AI] 将很有可能接管世界(无论是通过军事手段还是通过影响力和控制),并将其意志强加于其他所有人”,以及“更强大的 LLM……可能能够……研究并可能创造一种危险的新型生物……[这种生物] 可能以无法控制的方式增殖,并挤占地球上的所有生命,在最坏的情况下甚至摧毁地球上的所有生命。”担忧已经提出!

Here’s an essay that Anthropic Chief Executive Officer Dario Amodei wrote this year saying things like “it is clear that, if for some reason it chose to do so, [powerful AI] would have a fairly good shot at taking over the world (either militarily or in terms of influence and control) and imposing its will on everyone else” and “more capable LLMs … might be capable of … researching, and potentially creating, a dangerous new type of organism … [that] could proliferate in an uncontrollable way and crowd out all life on the planet, in the worst case even destroying all life on earth.” Concerns have been raised!

实际上,事情一直都是这样运作的:你试图指出 AI 可能会杀死所有人,而你不得不加上‘那将对我们的业务产生不利影响’,希望有人会决定关心一下。

This is actually how it works all the time, you try to point out that AI might kill everyone and you have to add ‘and that would adversely affect our business’ in the hopes that someone might decide to care.

……Anthropic 招股说明书中最大的风险并不是真正对其业务的风险;而是其业务对其他人构成的风险。我上面引用的那个风险轻描淡写地得出结论,杀死地球上的所有人“可能对我们的业务、财务状况和经营成果产生不利影响”,但杀死地球上的所有人,这肯定不是主要问题。

… The biggest risks in Anthropic’s prospectus are not _really_ risks to its business; they are risks that its business poses to everyone else. The one I quoted above shrugs into its conclusion that killing everyone on earth “could adversely affect our business, financial condition, and results of operations,” but surely that’s not the _main_ problem with killing everyone on earth.

可能有人会起诉或监管 Anthropic Someone Might Sue or Regulate Anthropic

他们担心各州总检察长,也担心那些自杀或实施暴力、或以其他方式遭受或造成伤害的人,尤其是未成年人。他们担心零数据保留(ZDR)会干扰其相关义务。他们还担心全球范围内的监管,包括来自英国和欧盟的监管。

They worry about state attorneys general, and about those who commit suicide or violence or otherwise suffer or inflict harm, especially minors. They worry about ZDR (zero data retention) interfering with their relevant obligations. And they worry about regulations worldwide, including from the UK and EU.

是的,这些确实都是重大担忧。Anthropic 此前曾被起诉过。他们现在正被起诉,有一整节专门讨论此事。他们还会再次被起诉很多次。他们还会因版权问题再次被起诉很多次。诸如此类。在第 97 页,他们列出了所有事项,包括因安全倡导而可能引发的反垄断诉讼。

Yes, these are definitely big worries. Anthropic has been sued before. They are being sued now; there is a whole section on that. They will be sued again many times. They will be sued for copyright again many times. And so on. On page 97 they list all the things, including potential antitrust lawsuits over safety advocacy.

然后是监管和政治。事情无论如何都会发生变化。向政府和高度受监管组织的销售,确实“面临挑战和风险”。

Then there's regulation and politics. Things are going to change, one way or another. Sales to government and highly regulated organizations are, indeed, 'subject to challenges and risks.'

此外,他们现在还必须遵守针对上市公司的法规。是的。

Also, they'll now have to comply with the regulations on public companies. Yep.

正如他们所说,这一切都可能对业务产生不利影响。

It could all, as they say, adversely impact the business.

供应链风险(2) Supply Chain Risk (2)

在第 71 页,Anthropic 讨论了与战争部(Department of War)之间的整个事件:Anthropic 被非法认定为“供应链风险”(Supply Chain Risk),总统还要求所有联邦机构不得使用 Claude。这部分内容大体上是一份“没错,确实发生过”的记述,涵盖了核心事件与法院裁决。他们还提到曾对 Fable 5 和 Mythos 5 临时实施的出口限制,该限制使这两款模型在各处都无法访问。

On page 71, Anthropic discusses the whole incident with the Department of War, in which Anthropic was illegally designated a 'Supply Chain Risk' and the President told all federal agencies not to use Claude. For the most part, it is a 'yep, that happened' account covering the core events and court rulings. They also mention the export restrictions temporarily imposed on Fable 5 and Mythos 5, which took model access down everywhere.

我们过于安全 We’re Too Safety

Anthropic 警告称,他们对 AI 安全的承诺可能会损害其业务。

Anthropic warns that their commitment to AI safety might hurt their business.

你看,他们,Anthropic,关心 AI 安全。他们做出昂贵的权衡。他们可能误判或搞砸这些权衡。这是真的。

You see, they, Anthropic, care about AI safety. They make expensive tradeoffs. They might misjudge or mess up those tradeoffs. It’s true.

他们还警告说,作为一个‘以安全为中心的组织’,如果他们在安全上搞砸了,甚至被看到搞砸了安全,他们将面临更高的风险,例如员工动荡、客户反弹、监管审查、诉讼或破坏性媒体报道。

They also warn that as a ‘safety-focused organization’ they face higher risks if they mess safety up, or are even seen as messing safety up, such as employee disruption, customer backlash, regulatory scrutiny, litigation or damaging media coverage.

说得对。我在看着。别逼我过去。

Damn right. I’m watching. Don’t make me come over there.

他们还担心自己的声誉。再说一次,别逼我过去。

They also worry about their reputation. Again, don’t make me come over there.

或者媒体报道可能因任何其他原因而不利。你已被警告。如果我们能事先想到,那就不构成证券欺诈。

Or media coverage might be bad for any other reason. You've been warned. If we can think of it in advance, it isn't securities fraud.

这与从第 230 页开始描述的实际安全工作形成对比,其标题为“安全研究是我们业务的核心”。他们强调了 RSP、Claude 宪法和“力争上游”,以及那些泛泛而平淡的策略。

That is in contrast to the actual safety work, as described starting on p. 230, with 'Safety Research is Core to Our Business.' They highlight the RSP and the Claude Constitution and the Race to the Top, along with the generic prosaic strategies.

Claude 核心 Claude Core

Anthropic 只有少量高度关联的产品,其头部客户较为集中(不过按行业惯例来看并不算特别集中,前五大客户仅占收入的约 25%),而这些客户都有种种常见理由,希望减少对 Anthropic 的使用或依赖。是的。

Anthropic only has a small number of highly linked products, and their top customers are concentrated (although not that concentrated as these things go, the top five are only ~25% of revenue), and all those customers have all the usual reasons for wanting to reduce their use of or dependence on Anthropic. Yes.

我们其实从未真正赚过钱 We’ve Never Actually Made Money

Anthropic 是一家新公司,经验不多,需要管理自身的增长。Anthropic 不得不承认,过去的双曲式增长并不预示未来的双曲式增长。

Anthropic is a new company, without much experience, that needs to manage its growth. Anthropic has to say that past hyperbolic growth is not indicative of future hyperbolic growth.

而随着 Anthropic 的增长,账面上它持续亏损更多。Anthropic 目前拥有一种可称为“基本盈余”的东西,我认为这应当算作赚钱,但如果把各项调整计入,它们仍处于严重赤字状态。

And as Anthropic grows, on paper it keeps losing more money. Anthropic has what one might call a 'primary surplus' at this point, and I think they should count as making money, but if you count the adjustments they are still very in the red.

我们的结果波动很大,有时数字会下降而非上升。哦不。

Our results vary a lot, and sometimes the number will go down instead of up. Oh no.

他们列出了一整串数字和其他可能下降而非上升的东西。

They list a whole list of numbers and other things that might go down instead of up.

情况变化很快,所有可用的指标在向未来外推时都有严重局限。过去的收入并不能很好地指引未来。

Things are changing fast and all the available metrics have severe limitations when it comes to projection into the future. Past revenue is a poor guide going forward.

或者他们可能无法组建优秀的销售与市场团队,或其策略可能不奏效,或在企业销售中遭遇“复杂性与风险”。或者他们可能提供糟糕的客户支持。或者新产品可能失败。或者可能出现技术故障,包括云服务等其他环节的故障。或者他们可能无法完成收购与投资,或无法将其整合,或监管机构可能不批准这些收购。他们可能无法与其他软件集成,或被应用商店下架。

Or they might not build a good sales and marketing team, or their strategies might not work, or they might run into 'complexities and risks' with enterprise sales. Or they might provide bad customer support. Or the new products might not work out. Or there might be technical failures, including elsewhere such as the cloud services. Or they might not make acquisitions and investments, or they might not integrate them, or the regulatory bodies might not approve those acquisitions. They might not be able to integrate with other software, or get kicked out of the app stores.

你正在买入一家公司的股票 You Are Buying Stock In a Company

在第 114 页,他们列出了理由,说明为什么如果你买入一家刚在证券交易所上市交易公司的普通股,它可能不涨反跌。

On Page 114 they lay out the reasons why, if you buy common stock of a company newly trading on the stock exchange, it might go down instead of up.

你是否知道,可能存在谣言、市场投机、买方多于卖方、特定卖方、业绩变化、波动性、交易量变化、财务预测变化,或者一般而言股票有时会下跌,包括其他相关股票?竞争对手可能发布新产品,或者公众可能不喜欢新闻稿,或者可能发表不利或不准确或两者兼有的公开研究?流通股规模可能变化,或者可能发行新股,或者通过新法律,或者经济状况可能变化,或者他们可能被起诉,或者 Anthropic 的运营结果可能存在实际或预期的变化或波动?

Did you know there might be rumors, or market speculation, or more buyers than sellers, or particular sellers, or changes in performance, or volatility, or changes in trading volume, or changing financial projections, or that just generally stocks sometimes go down, including other correlated stocks? That competitors might announce new products, or the public might not like press releases, or there might be public research published that is unfavorable or inaccurate or both? That the float size might change, or new shares might be issued, or new laws passed, or economic conditions might change, or they might be sued, or there might be actual or anticipated changes or fluctuations in the result of Anthropic's operations?

不会有股息 There Will Be No Dividends

在可预见的未来不会。特此警告。

Not for the foreseeable future. You have been warned.

我认为这是一个错误。我喜欢 Nvidia 的策略,即发行 0.01 美元的股息,这能让你被纳入各种指数。

I think this is a mistake. I like the Nvidia strategy of issuing a $0.01 dividend, which allows you to be included in various indexes.

我们还了解到股票上限,以防将来有一篇有趣的 Matt Levine 相关专栏:3000 亿股 A 类股、1000 亿股 N 类股、300 亿股 E 类股、300 亿股优先股,以及单一 F 类股和单一 T 类股。

We also learn the stock limits, in case there is ever a fun related Matt Levine column: 300 billion Class A shares, 100 billion Class N shares, 30 billion Class E shares, 30 billion shares of preferred stock and the single Class F and single Class T shares.

我们究竟在做什么? What Are We Even Doing?

可能会有战争。恐怖主义。自然灾害。健康危机。政治危机。

There might be a war. Terrorism. Natural disaster. Health crisis. Political crisis.

我越读风险部分,就越明白它在做什么。这是我的第一份 S-1,而这就是玩“以某种方式列出所有可能出错的事情”这个游戏的方式。

The more I read the risks section, the more I understand what it is doing. This is my first S-1, and this is how you play the game of 'list everything that could possibly go wrong, one way or another.'

我可以继续列下去。但我不会。大约到第 90 页时,我开始几乎只扫一眼标题。税法可能会变。ESG 争议可能会出现。是啊,是啊。我们为什么要做这个?

I could go on. I won't. Around page 90 I started barely skimming the titles. Tax laws could change. ESG disputes could arise. Yeah, yeah. Why are we doing this?

从某种意义上说,这是个有趣的游戏,但你不会想读结果。Pangram 抽查显示,这些部分是 AI 和非 AI 的混合。要说的话,我本来会用更多 AI。

It's a fun game in some senses, but you would not want to read the results. A Pangram spot check says these sections are a mix of AI and not-AI. If anything I would have used AI more.

这是 AI 写作的完美案例。这是没人需要写的文本,也是没人需要读的文本。它全是形式性的,是一种必须存在的东西,这样如果有人问‘这个存在吗?’,答案就会是肯定的。它是一份陷阱清单的陷阱清单,而我们上当了。

This is the perfect case for AI writing. It is text no one needs to write, and text no one needs to read. It is all pro forma, a thing that has to exist so that if anyone asks 'does this exist?' the answer comes back yes. It is a gotcha list of gotchas, and we gotcha.

有人告诉我,尽管那可能是其目的,但一旦有人被“坑”到,代价极其高昂;而且这些界线是由收费每小时 1 万美元的委员会逐条审议的,而这些委员会的工作将是地球上最后一批人类认知类工作。我明白这种寻租行为从何而来,对 Anthropic 来说,我猜你只能向这些“劫匪”付钱,但我不明白这为什么会改变什么,尤其是因为我看不出对过多事情发出过多警告会有什么真正的惩罚。

I have been told that, while that might be its purpose, the gotchas are extremely expensive if someone is got, and that lines are gone over one at a time by committees charging $10k an hour that will be some of the last human cognitive jobs on the planet. I get where that rent seeking comes from, and for Anthropic I guess you just pay the bandits, but I don't see why it changes anything, especially since there is no real penalty that I can see for warning too much about too many things.

财务模型 The Financial Model

股票市场基于“正常”的财务模型运行,因此我们得到一个正常的财务模型。Anthropic 投资于算力、使用算力、销售算力并运营。

The stock market operates on 'normal' financial models, so we get a normal financial model. Anthropic invests in compute, uses compute, sells compute, and operates.

在某些意义上,我想这是一个模型?我想,这是一个非常高层级的模型。

In some senses, I suppose this is a model? It is a very high-level model, I suppose.

**我们的财务模型**

**Our Financial Model**

我们的财务模型建立在四个主要驱动因素之上,这些因素定义了我们的业绩轨迹:

Our financial model is built on four primary drivers that define the trajectory of our results:

* _**Claude 的采用。**_ 每个 successive 版本的 Claude 模型不断进步的智能和能力推动了 Claude 在全球经济中的扩散。随着新客户采用 Claude 以及现有客户将其使用扩展到新的工作流程,我们的收入增长。

* _**Adoption of Claude.**_ The advancing intelligence and capabilities of each successive version of a Claude model drive the diffusion of Claude across the global economy. As new customers adopt Claude and existing customers expand their usage into new workflows, our revenue grows.

* _**算力效率。**_ 我们专注于提高每美元算力投入所产出的智能和所创造的收入,并已实现显著的效率提升。

* _**Compute Efficiency.**_ We are focused on increasing the intelligence produced and revenue generated per dollar of compute invested and have delivered significant efficiency improvements.

* _**运营效率。**_ 我们高密度的人才、高效的上市策略,以及团队在工作中广泛使用 Claude,有助于提升运营利润率、调整后运营利润率以及人均收入。

* _**Operating Efficiency.**_ Our high talent density, efficient go-to-market, and how extensively our teams leverage Claude in their work help to drive improved operating margin, Adjusted Operating Margin, and revenue per employee.

* _**算力投资。**_ 投资算力对于支撑我们的长期增长至关重要,我们打算继续在算力上进行大规模投资。

* _**Compute Investments.**_ Investing in compute is critical to support our long-term growth, and we intend to continue making substantial investments in compute.

我们财务模型的各项驱动因素会在不同的时间线上逐步显现。

The drivers of our financial model materialize on different timelines.

Claude 通过全部三大云平台(AWS、GCP 和 Azure)以及多种其他渠道销售。增长可谓‘异常迅猛’。

Claude is sold through all three major cloud platforms (AWS, GCP and Azure) and through various other methods. Growth is what we call ‘gangbusters.’

我们以 TTM 口径计算收入达到 10 万美元及以上的客户,覆盖了《财富》10 强中的每一家公司、《财富》500 强中超过 50% 的公司,以及全球 2000 强中超过 30% 的公司。

Our customers with $100k or more of revenue on a TTM basis included every company in the Fortune 10, over 50% of the Fortune 500, and over 30% of the Global 2000.

在《财富》500 强中,这些客户包括 10 家最大的软件与 IT 服务公司、10 家最大的互联网公司、10 家最大制药公司中的 9 家、10 家最大的银行、10 家最大半导体公司中的 9 家,以及 10 家最大医疗健康公司中的 8 家。

Within the Fortune 500, these customers included the 10 largest software and IT services companies, the 10 largest internet companies, nine of the 10 largest pharmaceutical companies, the 10 largest banks, nine of the 10 largest semiconductor companies, and eight of the 10 largest healthcare companies.

除此之外,现有客户的净收入留存率(Net Revenue Retention)约为 900%。这是好事,因为上述情况本可以被称为“市场饱和”。

On top of that, existing customers have Net Revenue Retention of ~900%. Which is good, because the above could be called 'market saturation.'

接下来是算力效率,它又可细分为训练效率(正在提升)、推理效率(正在提升)以及多芯片算力方案。

Next up is compute efficiency, which then breaks down into training efficiency (going up), inference efficiency (going up) and multi-chip compute approach.

算力正越来越高效地转化为收入。

Compute is getting increasingly efficiently turned into revenue.

随后,Anthropic 出售 Claude 的推理服务,包括通过订阅方式提供。

Anthropic then sells Claude inference, including via subscriptions.

接着,他们与客户合作,帮助客户将 Claude 集成到工作流中,并随着时间推移在更多事务上使用更多 Claude。

They then work with customers as they integrate Claude into workflows and use more Claude for more of the things over time.

这个“模型”并不能让你预测太多,除了“哇,看起来数字在上升、在变好,也许他们会赚很多钱”。

This 'model' does not allow you to predict much, other than 'wow, it looks like numbers are going up and getting better, perhaps they will make a lot of money.'

你可以看出这个模型并没有那么有用,因为“影响我们业绩的关键因素”随后只是从前面重复了一遍,而并未直接纳入模型之中。

You can tell the model is not that useful since the 'key factors affecting our performance' then get repeated from earlier without being directly in the model.

他们还以调整后的形式重复了若干财务数字陈述。

They also repeat a bunch of the financial number statements in adjusted form.

后来,在评估 IPO 前普通股的公允价值时,董事会“行使了合理判断,并考虑了众多客观和主观因素”,包括:

Later, when evaluating the fair value of the common stock pre-IPO, the board of directors 'exercised reasonable judgment and considered numerous objective and subjective factors,' including:

3. 买方是否会多于卖方,或卖方是否会多于买方。

3. Whether there will be more buyers than sellers, or more sellers than buyers.

哦,抱歉,我被告知他们实际上声称考察了其他方面,但嘿。

Oh, sorry, I am informed that instead they claim they looked at other things, but hey.

智能的商业,以及可能杀死所有人 The Business of Intelligence and Maybe Killing Everyone

S-1 文件中有很多重复内容。“业务”部分会再次讨论前面已经出现过的许多相同事实,而我已经在前面讨论过了。

There is a lot of repetition in an S-1. The ‘business’ section goes over a lot of the same facts from earlier, that I’ve already gone over.

这里是将它们汇总在一起的地方,并借此强调:“嘿,你知道 Claude 越来越能够完成所有数字任务,还能与物理世界交互,并且它会持续进步并遵循缩放定律吗?”

This is where they get gathered together, and get to emphasize ‘hey did you know Claude is increasingly able to do all the digital things and also interface with the physical things, and it’s going to keep improving and follow the scaling laws?’

但是,是的,它正在到来。他们对那些有眼能看的人说了他们能说的话。

But yeah, It’s Coming. They say what they can for those with eyes to see.

**所有数字工作最终都是代码的一种表达**,而 Claude 对代码的精通驱动了虚拟协作者,能够处理远超软件开发的复杂工作。

**All digital work is ultimately an expression of code**, and Claude’s mastery of it powers Virtual Collaborators capable of handling complex work far beyond software development.

……我们正在迈向这样一个世界:Claude 可以被赋予持续数小时、数天或数周的无限制任务,然后以一位才华横溢的同事般的判断力和主动性自主执行这些任务。

… We are building towards a world where Claude can be given open-ended tasks lasting hours, days, or weeks, and then execute them autonomously with the judgment and initiative of a brilliant colleague.

……Claude 扩散的一个核心驱动力是其智能。

… A core driver of Claude’s diffusion is its intelligence.

……我们发布的每一个新原语和应用都扩展了它们可以在 Claude 上运行的工作流范围。

… Each new primitive and application we release expands the range of workflows they can run on Claude.

你确实需要做一些外推,否则你会错过即将发生的大部分事情。

You do have to extrapolate a bit, or you miss most of what is going to be happening.

否则,你能说多少次你会做各种好事?很多次。

Otherwise, how many times can you say you will do the various good things? A lot.

他们明确列出了当前所有产品:Claude [Code / Cowork / .ai / Security / Tag / Design / Science]。还有面向[金融服务 / 财务顾问 / 医疗健康 / 生命科学 / 法律 / 教育 / 政府]的 Claude。全都是 Claude。

They spell out what all their current products are: Claude [Code / Cowork / .ai / Security / Tag / Design / Science]. You've got Claude for [Financial Services / Financial Advisors / Healthcare / Life Sciences / Legal / Education / Government]. It's all Claude.

有一个我此前没听过的词是 Prosumers,我认为它代表“专业消费者”(professional consumers),指个人版 Pro 或 Max 订阅用户。我就是一名 prosumer。

One term that was new to me was Prosumers, which I think stands for 'professional consumers' and means individual Pro or Max subscribers. I'm a prosumer.

他们展示并强调了与以下公司的成功合作:Accenture、AIG(哦不,这是按字母顺序排的吗?看起来确实是按字母顺序)、Bristol Myers Squibb、DoorDash、International Rescue Committee、Jane Street、Legora、Lilly、Mercado Libre、Mozilla、PayPal、Rakuten、Salesforce、SAP、Shopify、Travelers、UST、Verizon、Walmart,并特别提及 Shy Bird 和 KBF。

They show off and highlight successful partnerships with Accenture, AIG (oh no, is this alphabetical? It looks like it is alphabetical), Bristol Myers Squibb, DoorDash, International Rescue Committee, Jane Street, Legora, Lilly, Mercado Libre, Mozilla, PayPal, Rakuten, Salesforce, SAP, Shopify, Travelers, UST, Verizon, Walmart, and featuring Shy Bird and KBF.

不同公司给出的数据和用例各不相同。对于大量编写代码的用户,收益看起来很大。代码产出可信地达到 +100% 或更高。对其他用户而言则较为有限,例如 PayPal 每位销售团队成员每周节省 45 分钟。话说回来,一个定制插件节省约 2% 的时间,算少还是算多?SAP 每位顾问每月节省 25 小时,约 14%,又该怎么看?我预计在未来一年内,这个答案会发生变化。

Different companies offer different stats and use cases. The gains look large for those who do a lot of coding. Coding outputs seem credibly to be +100% or more. For others it is modest, such as PayPal's 45 minutes saved per sales team member per week. Then again, is ~2% time saved from a custom-built plugin not much, or is it a lot? What about the 25 hours per month saved per consultant at SAP, or ~14%? Over the next year I expect that answer to change.

别对我的监管毫不在意 Don’t Give a Damn About My Regulation

Anthropic 重申其立场,并将自己描绘成支持监管的一方。

Anthropic reiterates its position, and paints itself as pro-regulation.

我们认为,负责任的 AI 发展需要深思熟虑的治理,我们积极支持政府和监管机构建立促进安全、有益 AI 的框架的努力。我们成立的前提是,AI 系统的开发必须以安全和社会效益为核心优先事项,我们相信精心设计的监管可以在整个行业强化这些价值观。

We believe that the responsible development of AI requires thoughtful governance, and we actively support efforts by governments and regulatory bodies to establish frameworks that promote safe, beneficial AI. We were founded on the premise that AI systems must be developed with safety and societal benefit as central priorities, and we believe that well-designed regulation can reinforce those values across the industry.

由于前沿 AI 开发是一个相对新生且快速发展的领域,当前的法律、法规和标准未必是针对这些技术设计的。我们主动与美国及国际的立法者、监管者和其他利益相关者接触,以推动制定我们认为对社会、客户和经济长期健康有益的监管框架,并打算继续这样做。

Because AI development at the frontier is a relatively nascent and rapidly evolving field, the laws, regulations, and standards in place today were not necessarily designed with these technologies in mind. We have proactively engaged with legislators, regulators, and other stakeholders in the United States and internationally to promote the development of regulatory frameworks that we believe are good for society, our customers, and the long-term health of the economy, and we intend to continue doing so.

例如,我们与 CAISSI 达成协议,承诺在模型公开发布前后提供模型访问以进行测试和评估,同意根据与 CAISSI 的协议向英国 AI 安全研究所提供模型访问以进行部署前测试,并公开支持加利福尼亚州的《前沿人工智能透明度法案》(SB 53)。

For example, we entered into an agreement with the CAISSI whereby we committed to provide access to our models for testing and evaluation before and after their public release, agreed to provide model access to the U.K. AI Security Institute for pre-deployment testing consistent with our agreement with the CAISSI, and publicly endorsed California’s Transparency in Frontier Artificial Intelligence Act (SB 53).

关于我的回应,请参阅 Anthropic 的**近期其他文章**。Anthropic 在某些方面表现优于其他实验室,他们确实优先考虑他们认为最重要的条款,即使这些条款不是显而易见的。但我认为他们的提案未能应对当下的挑战,而且这不是一个新问题。

See **recent****other****posts** on Anthropic for my response to that. Anthropic is in some ways performing better than other labs, and they do prioritize the provisions they feel are most important, even when they are not the obvious ones. But I do not feel their proposals are meeting the moment, and this is not a new issue.

这是一份 S-1 文件,因此这里真正的问题在于对 Anthropic 利润的影响。除了列举一些基本的过往事件外,他们对此并没有太多说明。

This is an S-1, so the actual question here is the impact on Anthropic's bottom line. They don't have that much to say about that, other than laying out some basic past events.

底线 The Bottom Line

归根结底,底线就是底线。S-1 文件的核心在于投资者关心什么,以及什么能防止公司因证券欺诈被起诉,或避免让金融界人士对 Anthropic 大为光火。其他受众都是次要的。

Ultimately, the bottom line is the bottom line. The S-1 is all about what investors care about, and what will guard against being sued for securities fraud or otherwise getting the financial types Big Mad at Anthropic. Other audiences are secondary.

Anthropic 仍然是一家令人沮丧的公司。现在我们对它有了更多了解。

Anthropic remains a frustrating company. Now we know a lot more about it.

互动版:针对本篇提问 →

接着读 · Read next

目录