Five of the most valuable software companies in the world — Microsoft, Google, Meta, Anthropic, and OpenAI, commonly grouped together as the Frontier Labs — have, within the past 18 months, independently re-oriented themselves around a common strategy that not a single one of them originated. It is a strategy that seeks to mimic the core organizational structure, technological primitives, and the philosophy of progress embodied by a single, much smaller software company.全球市值最高的五家软件公司——微软、谷歌、Meta、Anthropic和OpenAI（统称为“前沿实验室”）——在过去18个月里，不约而同地围绕着一种并非由它们中任何一家开创的共同战略进行了重组。这一战略旨在模仿一家规模小得多的软件公司所体现的核心组织结构、技术原语以及进步哲学。

It’s a company that does not share the Frontier Labs’ foundational focus on developing frontier AI models: systems that they, at least publicly, believe will soon be capable of automating all economically valuable work.这家公司并不像前沿实验室那样，将核心重心放在开发前沿AI模型上：即那些它们（至少在公开场合）认为很快就能实现所有具有经济价值的工作自动化的系统。

Its inner workings were, until very recently, more known and understood by executives at companies like General Mills and bureaucrats at three-letter agencies in Washington than by the average Silicon Valley AI researcher, engineer, or investor.直到最近，相比于硅谷普通的AI研究人员、工程师或投资者，通用磨坊（General Mills）等公司的高管以及华盛顿那些三字母机构的官僚们，反而对其内部运作更为熟悉和了解。

It is a company which professional investors on Wall Street still think bears a stronger resemblance to GameStop, AMC, and Nikola, than Google, SpaceX, or Amazon. Many engineers in Silicon Valley think it bears a stronger resemblance to Big Brother than to its antithesis.在华尔街的专业投资者眼中，这家公司依然更像GameStop、AMC和Nikola，而非谷歌、SpaceX或亚马逊。而硅谷的许多工程师则认为，它更像“老大哥”，而非其对立面。

That company is Palantir — founded 23 years ago and long dismissed as a company that provides data integration services, a category so unglamorous it’s functionally an insult.这家公司就是Palantir——成立于23年前，长期以来被视为一家提供数据集成服务的公司，这一类别平庸到几乎是一种侮辱。

But the defining technologies of the modern era are themselves clever flavors of data integration: the internet has integrated the world’s public knowledge, GitHub has integrated the world’s code, and the frontier models are, in a real sense, a compression of both. Each time a previously illegible corpus of knowledge and knowhow has become programmable: crawled, indexed, linkable, searchable, and versioned, discontinuities in progress seem to follow. Palantir’s founding insight, in some sense, was that the largest, most valuable corpus of all — the operational knowledge and knowhow of the institutions that run the world — remains API-incomplete: scattered across software systems that can’t talk to each other or, in many cases, in mediums that are undigitized altogether. It’s trapped in email inboxes, groupchats, ERPs, CRMs, Excel files, the airwaves of phone calls and also in the heads of (and in the conversations between) doctors, line engineers, and other front-line workers who embody the valuable knowhow they’ve earned through experience.然而，现代时代的决定性技术本身就是巧妙的数据集成形式：互联网整合了全球的公共知识，GitHub整合了全球的代码，而前沿模型在某种意义上正是两者的压缩。每当一种此前难以解读的知识和技术语料库变得可编程（即：被抓取、索引、链接、搜索和版本化）时，进步的断层似乎随之而来。Palantir的创始洞察在某种意义上是：最庞大、最有价值的语料库——即运营世界的机构所拥有的操作知识和诀窍——仍然处于API不完整的状态：它们分散在无法互通的软件系统中，或者在许多情况下，存在于完全未数字化的媒介中。它们被困在电子邮件收件箱、群聊、ERP、CRM、Excel文件、电话语音中，也存在于医生、一线工程师和其他通过经验积累了宝贵知识的一线工作者的头脑中（以及他们的对话里）。

Crucially, Palantir recognized that this fragmented corpus would not reveal, version, and programmatically assemble itself. There would be no GitHub repo or version control system that encodes a real-time understanding of how SpaceX puts 100 metric tons into orbit at less than $200 per kg or how Airbus assembles four million discrete parts into an A380, unless someone tirelessly coaxed and captured it into existence. In the case of SpaceX, Elon had it handled, but Airbus and many others like them required a new kind of software foundry.至关重要的是，Palantir认识到，这些碎片化的语料库不会自动显现、版本化并以编程方式整合。除非有人不懈地引导并将其捕捉成型，否则不会有GitHub仓库或版本控制系统来编码实时理解，以说明SpaceX如何以每公斤不到200美元的成本将100公吨载荷送入轨道，或者空中客车如何将四百万个独立零件组装成一架A380。在SpaceX的案例中，埃隆解决了这个问题，但空中客车及许多类似企业需要一种新型的软件铸造厂。

Palantir bet that the next generation-defining company would natively align itself with institutions to make their data legible, integrate it into computable ground truth, and in the process enables these institutions to solve problems and serve society more effectively. The dream was that they could finally connect the rapid progress the economy has been experiencing in the world of bits for the past 50 years to the relatively stagnant world of atoms, and, in this process of extending increasing-returns-to-scale characteristics to more of the economy, become fabulously rich themselves.Palantir押注，下一家定义时代的公司将与机构原生对齐，使它们的数据变得可读，将其整合为可计算的基准事实，并在此过程中使这些机构能够更有效地解决问题和服务社会。其愿景是，他们最终能够将过去50年来经济在比特世界中所经历的飞速进步，与相对停滞的原子世界连接起来；并在此过程中，将规模收益递增的特性扩展到更多的经济领域，从而使自己变得极其富有。

This 23 year bet has (rightfully) made Palantir the envy of techno-capitalism: since the November 2022 launch of ChatGPT, Palantir has seen its market capitalization rise more than 20 fold, outperforming every other publicly listed company on US major exchanges, including the most lauded beneficiaries of the AI buildout like Nvidia (~13x), Broadcom (~8x), and Micron (~17x). This stock-price performance is downstream of mindboggling business fundamentals — just look at the second derivative! Palantir’s revenue growth has accelerated for 12 quarters straight, from 13% year-over-year in Q2 2023 (the quarter GPT-4 was released) to 93% in Q2 2026, and profitability (operating margin) has expanded from 2% to 47%.这场为期23年的赌注（理所当然地）使Palantir成为了技术资本主义羡慕的对象：自2022年11月ChatGPT发布以来，Palantir的市值增长了20多倍，超过了美国主要证券交易所的所有其他上市公司，包括英伟达（约13倍）、博通（约8倍）和美光（约17倍）等AI建设浪潮中最受赞誉的受益者。这种股价表现源于令人瞠目结舌的商业基本面——看看二阶导数就知道了！Palantir的营收增长已连续12个季度加速，从2023年第二季度（GPT-4发布的季度）的同比增长13%提升至2026年第二季度的93%，而盈利能力（营业利润率）则从2%扩大到了47%。

That sort of market vindication is hard to watch without jealousy, particularly if your business model has been tied up in a different bet on where the value accrues as AI progress accelerates.这种市场验证很难让人不心生嫉妒，尤其是如果你的商业模式与“随着AI进步加速，价值将在何处积累”这一问题的不同赌注挂钩时。

The frontier labs’ theory has been that scale conquers all: acquire or create more publicly legible data, spend more compute training on it, and out will pop surprisingly powerful general-purpose models that first will be able to improve themselves, and eventually will be capable of automating economically valuable work writ large. Given the pace at which code is getting shipped, mathematical conjectures are falling, and frontier lab revenue is growing, that bet looks like it’s working.前沿实验室的理论是规模征服一切：获取或创造更多可公开读取的数据，投入更多的计算资源进行训练，随之而来的将是令人惊讶的强大通用模型，它们首先能够自我改进，最终将能够实现大规模的经济价值工作自动化。鉴于代码交付的速度、数学猜想的不断攻克以及前沿实验室营收的增长，这个赌注看起来是有效的。

But it’s not the only one you can make. You can also bet that as models become more generally capable, contextually-appropriate judgment, embodied privately within institutions, and the institutional right to exercise that judgement, is what remains scarce. In other words, general intelligence is much more useful in permissioned contexts, and cheap tokens make choosing the right prompt (and privately owning the upside of the output) more valuable.但这并非你唯一的选择。你也可以押注：随着模型变得更加通用，私下嵌入机构内部的、情境恰当的判断，以及行使该判断的机构权利，才是稀缺资源。换句话说，通用智能在许可环境下更有用，而廉价的Token使得选择正确的提示词（并私下拥有输出带来的收益）变得更有价值。

Palantir popularized a very specific role around this idea, the Forward-Deployed Engineer: someone who can understand a customer’s business problem, propose a technical solution, and, yes, write mountains of code that connect database X and third-party product Y to internal dashboard Z. This requires a mix of technical skills, business sense, and the discipline to give realistic answers instead of overselling. One of the core benefits of this structure is that it solves a classic incentive-alignment problem in sales: the person making specific promises is also the one who has to make those promises come true. It’s a clever way to organize information and skills within a company.Palantir围绕这一理念普及了一个非常具体的角色——前沿部署工程师（Forward-Deployed Engineer）：他们能够理解客户的业务问题，提出技术解决方案，并且，没错，编写海量代码将数据库X和第三方产品Y连接到内部仪表板Z。这需要技术技能、商业敏锐度以及给出现实答案而非过度推销的纪律的结合。这种结构的核心优势之一在于，它解决了销售中经典的激励对齐问题：做出具体承诺的人，同时也必须是实现这些承诺的人。这是一种在公司内部组织信息和技能的巧妙方式。

The labs are increasingly realizing this. Within the last four months, OpenAI, Anthropic, Google, Meta and Microsoft have deployed $30B of capital, acquired or funded four ex-Palantir founded or staffed companies (Tomoro, Fractional, Northslope, and Thrive Holdings), launched partnerships with the largest PE firms (e.g, Blackstone, TPG, Bain Capital, Thoma Bravo), and have set out on a hiring tear for 9,000+ FDEs of their own — equivalent headcount to two Palantirs — to gain native access to and deploy AI within institutions. Microsoft in particular has taken this mimicry to almost comedic lengths: naming its enterprise AI data-integration and deployment platform Foundry, roughly a decade after Palantir launched its flagship enterprise offering of the same name. All of this serves the same basic goal: get trusted access to, and make programmable, the scarce institutional context that Palantir has long known is necessary to turn machine tokens into economic value.实验室正在日益意识到这一点。在过去四个月里，OpenAI、Anthropic、谷歌、Meta和微软投入了300亿美元资本，收购或资助了四家由前Palantir员工创立或拥有前Palantir员工的公司（Tomoro、Fractional、Northslope和Thrive Holdings），与顶级私募股权公司（如Blackstone、TPG、Bain Capital、Thoma Bravo）建立了合作伙伴关系，并开始大规模招聘9000多名他们自己的FDE——相当于两个Palantir的人员规模——以获得对机构的原生访问权限并在其中部署AI。微软尤其将这种模仿发挥到了近乎滑稽的地步：将其企业级AI数据集成与部署平台命名为“Foundry”，而这距离Palantir推出其同名的旗舰企业产品已过去近十年。所有这些都服务于同一个基本目标：获得对稀缺机构背景的信任访问权限，并使其可编程化，而Palantir早已知道这是将机器Token转化为经济价值的必要条件。

So, we must ask: why are the labs trying to copy Palantir? Are they copying the right thing? Can you even copy it — or is a company’s model of delivering value more like a constitution: something that evolves out of a specific context, and that can’t be copy-pasted into some new context and expected to deliver the same results? It might turn out that it’s relatively easy to distill frontier models, but that it’s almost impossible to distill institutions. And, in the big picture, why does this matter beyond the narrow world of big enterprise software contracts?因此，我们必须问：为什么实验室要试图模仿Palantir？它们模仿对了吗？你甚至能模仿它吗——还是说一家公司的价值交付模式更像是一部宪法：它是从特定环境中演化出来的东西，不能简单地复制粘贴到新的环境中并期望获得同样的结果？事实证明，提炼前沿模型可能相对容易，但提炼机构几乎是不可能的。而且，从宏观角度来看，除了大型企业软件合同这个狭窄的世界之外，这为什么重要？

To motivate things, we’ll answer the last question first. If valuations (which are well into the trillions) are any signal, the frontier labs represent the capital markets’ best judgment on what kind of architectural knowledge — organizational, technological, and philosophical DNA — is necessary to enable and benefit from the cognified economy: a future that promises to scale economic growth and human prosperity with abundant machine intelligence.为了激发思考，我们先回答最后一个问题。如果估值（已达数万亿美元）是一个信号的话，那么前沿实验室代表了资本市场对实现并受益于“认知经济”所需的建筑性知识——组织、技术和哲学DNA——的最佳判断：这是一个承诺通过丰富的机器智能来扩大经济增长和人类繁荣的未来。

But the fact that all of these firms have converged, recently and independently, on a new common strategy that they didn’t originate — focused on FDEs, deployment, and institutional knowledge — illuminates a potential shift in their inside view. A shift in what our best-capitalized technological institutions actually believe is required to achieve their shared goal of beneficial artificial general intelligence, and their intuition about what they are missing. The labs independently converging on Palantir’s strategy and founding insights, and aggressively deploying capital and labor to mimic it, tells us they believe that Palantir is what’s missing. We’re going to argue they’re right: the development of beneficial AGI is in fact contingent on deeply understanding and internalizing Palantir’s lessons, but that it’s far from clear whether the labs actually have.但所有这些公司最近独立地汇聚于一种它们并未开创的共同战略——专注于FDE、部署和机构知识——这一事实揭示了它们内在观点的潜在转变。这表明我们资本最雄厚的技术机构实际上认为，要实现它们共同的“有益通用人工智能”目标，需要什么，以及它们对自己所缺失东西的直觉。实验室独立地向Palantir的战略和创始洞察靠拢，并积极投入资本和劳动力进行模仿，这告诉我们它们相信Palantir正是它们所缺失的。我们将论证它们是对的：有益AGI的发展实际上取决于对Palantir经验的深刻理解和内化，但实验室是否真正做到了这一点还远未可知。

The labs seem to have a general sense that Palantir has solved AI deployment. But we think Palantir solved a deeper problem, one that the labs don’t seem to grok: the problem of connecting powerful computation to local institutional knowledge without destroying an institution’s sovereignty or its incentive to reveal that knowledge.实验室似乎有一种普遍的共识，认为Palantir解决了AI部署问题。但我们认为Palantir解决的是一个更深层次的问题，一个实验室似乎无法领悟的问题：如何在不破坏机构主权或其揭示该知识的激励机制的情况下，将强大的计算能力与本地机构知识连接起来。

Envy is usually a sign that you strongly desire someone’s capabilities and traits, or at least the rewards that come with possessing them. The unhealthy response to envy is to signal or superficially copy those capabilities, which makes you averse to actually understanding their nature, why they matter to your goals, and how to truly develop them yourself. (Whenever someone plaintively complains that they did everything right, and didn’t get a fair outcome, pay attention to how much of what they did right involves credentials rather than skills, or is plainly unfalsifiable.) The healthy response to envy is to acknowledge the feeling without judgment, then use it as a map to learn why the capabilities the other person has are important to achieving your goals, and use that as motivation to truly understand those capabilities and develop them. Done well, envy becomes inspiration.嫉妒通常表明你强烈渴望某人的能力和特质，或者至少是拥有这些能力所带来的回报。对嫉妒的不健康反应是去标榜或肤浅地模仿这些能力，这会使你排斥真正去理解它们的本质、它们为何对你的目标重要，以及如何真正亲自去发展它们。（每当有人哀怨地抱怨他们做对了一切却没得到公平的结果时，请注意他们所做的“对”的事情中，有多少是涉及资历而非技能，或者明显是不可证伪的。）对嫉妒的健康反应是不带偏见地承认这种感觉，然后将其作为一张地图，去学习对方拥有的能力为何对实现你的目标至关重要，并以此作为动力去真正理解这些能力并发展它们。如果处理得当，嫉妒就会成为灵感。

The labs are currently responding to envy in an unhealthy way, leading them to create Palantir Cargo Cults that achieve the opposite outcome the labs’ desire, the outcome that makes Palantir enviable in the first place. Palantir’s chairman, Peter Thiel, likes to say that competition is for losers. Or as Rene Girard, one of Thiel’s strongest intellectual influences, made a related point: imitation starts with external forms but ultimately reflects borrowed desire.实验室目前正以一种不健康的方式应对嫉妒，导致它们创造了“Palantir货物崇拜”，这反而导致了与实验室愿望相反的结果——即Palantir之所以令人羡慕的那个结果。Palantir董事长彼得·蒂尔（Peter Thiel）喜欢说竞争是给失败者准备的。或者正如蒂尔最强大的思想影响者之一勒内·吉拉尔（Rene Girard）所指出的相关观点：模仿始于外部形式，但最终反映的是借来的欲望。

The charitable framing is that the labs are Cargo Culting Palantir because they don’t actually understand it. The uncharitable framing is that the labs are Cargo Culting Palantir because Palantir threatens their credibility and long-standing strategy, and the best plan they have for addressing that is to have a Palantir of their own.仁慈的解读是，实验室之所以对Palantir进行货物崇拜，是因为它们并不真正理解它。苛刻的解读是，实验室之所以对Palantir进行货物崇拜，是因为Palantir威胁到了它们的信誉和长期战略，而它们应对这一威胁的最佳计划就是拥有一个属于自己的Palantir。

The Essence of PalantirPalantir的本质

What’s the essence of an institution that gets more name recognition for the assassination of Osama Bin Laden than the Navy SEAL who shot him? That routinely convinces MIT Course 6 graduates to turn down offers from the best prop trading firms to spend their Friday nights optimizing processes at General Mills? That makes those same graduates feel like the highest-status people in their peer group, a peer group whose modal view also happens to be that Palantir is a company whose sole purpose is making ICE more effective at deporting parents in front of their children?一个因刺杀本·拉登而获得的知名度比执行任务的那个海豹突击队员还要高的机构，其本质是什么？一个经常说服麻省理工学院6号课程的毕业生拒绝顶级自营交易公司的录用通知，转而利用周五晚上在通用磨坊优化流程的机构，其本质是什么？一个让这些毕业生在同龄人群体中感到自己地位最高，而该群体的主流观点恰好认为Palantir是一家唯一目的是让ICE更有效地在孩子面前驱逐父母的公司的机构，其本质是什么？

Palantir’s approach to building software is, in some sense, a bet on the power of the intellect; a bet that the organization with the clearest real-time view of their internal and external world will be best suited to make positive expected value decisions within it, whether that’s the CIA preventing a terrorist attack, Tampa General Hospital detecting sepsis in a patient early enough to cure it, or Airbus being alerted that it has a defective part before continuing an expensive assembly run. Palantir helps customers natively connect their operations to software to prevent very costly, often existential, errors of commission and omission. The institutional world models Palantir builds for customers (and that are ultimately owned by customers) help them model the real-time state of their institution, predict the impact of potential decisions, execute optimally based on that understanding, and learn from the outcomes to more accurately understand the current state and make better predictions in the future.Palantir构建软件的方法在某种意义上是对智力力量的押注；它押注于拥有对其内部和外部世界最清晰实时视图的组织，将最适合在其中做出正期望值的决策，无论是CIA预防恐怖袭击，坦帕总医院在败血症患者病发早期检测到并治愈，还是空中客车在昂贵的组装运行继续之前收到有缺陷零件的警报。Palantir帮助客户将他们的运营与软件原生连接，以防止代价高昂、往往是生存层面的作为和不作为错误。Palantir为客户构建（并最终由客户拥有）的机构世界模型，帮助他们模拟机构的实时状态，预测潜在决策的影响，基于这种理解进行最优执行，并从结果中学习，从而更准确地理解当前状态并在未来做出更好的预测。

These world models are constructed by Palantir’s Forward Deployed Engineers, FDEs, who CTO Shyam Sankar has described as “pointillist painters” who “metabolize pain and excrete product,” turning an institution’s dark matter, its intangibles, into a computationally legible structure. Palantir calls this digital representation the Ontology. They go to work crawling and indexing a company’s assets, capabilities, knowledge and knowhow, and, importantly, the relationships between them to form this Ontology. When it comes to looking at a firm or institution as the unit of analysis, consilience is everything: understanding how all the disparate streams of information fit together to reveal some important truth about the nature of your business, market, or the world, is where all the alpha lies.这些世界模型由Palantir的前沿部署工程师（FDE）构建，CTO Shyam Sankar将他们描述为“点彩派画家”，他们“代谢痛苦并排泄产品”，将机构的暗物质——即无形资产——转化为计算上可读的结构。Palantir将这种数字表征称为“本体”（Ontology）。他们致力于抓取和索引公司的资产、能力、知识和诀窍，重要的是，还有它们之间的关系，以形成这一本体。当把公司或机构作为分析单位时，融会贯通（consilience）就是一切：理解所有分散的信息流如何结合在一起，以揭示关于你的业务、市场或世界本质的某些重要真相，这就是所有Alpha（超额收益）所在的地方。

Google made a bet that by virtue of powerful computers existing and being linked together, the world’s important information and secrets would continuously reveal themselves for Google to organize and make useful. Traditional enterprise software companies like Microsoft made the bet that by giving everyone in an organization access to rows, columns, and formulas to manipulate them, that all of an enterprise’s information would automatically be represented in exactly the way that’s best suited to solving problems and avoiding catastrophes.谷歌押注于：凭借强大计算机的存在和互联，世界上重要的信息和秘密将不断自我显现，供谷歌组织并使其变得有用。微软等传统企业软件公司则押注于：通过让组织中的每个人都能访问行、列和公式来操作它们，企业的所有信息都会自动以最适合解决问题和避免灾难的方式呈现出来。

Palantir understood that consilience is rare and that the disparate streams of information that inform it don’t reveal themselves: FDEs must tirelessly kick down factory doors, scour hospital floors, and travel to clandestine military bases to uncover these secrets and piece them together into beautiful pointillist paintings. Importantly, they understand that these paintings make very little sense until they are near completion, and that while the map (Ontology) never fully matches the territory, the goal of the FDE is to ensure that it’s always getting less wrong.Palantir理解到，融会贯通是罕见的，而提供这一信息的各种分散信息流并不会自动显现：FDE必须不懈地踢开工厂大门，搜寻医院楼层，并前往秘密军事基地，以揭示这些秘密并将它们拼凑成美丽的“点彩画”。重要的是，他们理解这些画作在接近完成之前几乎没有意义，并且虽然地图（本体）永远无法完全匹配领土，但FDE的目标是确保它变得越来越不偏离。

“The enterprise software industry’s focus on custom software tools and applications is misplaced. Those approaches often only work briefly, if at all. The problems and needs of an organization often change before the software can even be deployed. Our partners require something more. They need generalizable platforms for modeling the world and making decisions. And that is what we have built.”“企业软件行业对定制软件工具和应用程序的关注是错误的。这些方法通常只能短暂奏效，甚至根本无效。组织的各种问题和需求往往在软件部署之前就已经发生了变化。我们的合作伙伴需要更多。他们需要用于模拟世界和做出决策的通用化平台。而这正是我们所构建的。”

- Alex Karp, Palantir S-1, 2020——Alex Karp，Palantir S-1招股书，2020年

Ted Mabrey, Head of Palantir Commercial, responsible for their roughly $3B enterprise business, growing 157% year-over-year, tells Palantir FDEs to “act as if they are the CEO, but with zero authority.” We’d take it even further, and say that the best Palantir FDEs act as if they are Elon, who many would consider the best CEO in the world, but with zero authority. Like Elon, the prototypical engineer-CEO, the best FDEs are deeply curious and ruthlessly pragmatic. They seek to possess an almost superhuman understanding of the real-time state of capabilities and opportunities within their businesses, identify the key blockers (problems) to executing on those opportunities, and go to ground truth as a way to access and instrument the context required to devise systematic solutions.负责Palantir约30亿美元企业业务（同比增长157%）的商业主管Ted Mabrey告诉Palantir的FDE们：“表现得好像你是CEO，但没有任何权威。”我们甚至会更进一步，说最好的Palantir FDE表现得就像埃隆（许多人会认为他是世界上最好的CEO），但没有任何权威。像埃隆这位典型的工程师兼CEO一样，最好的FDE充满好奇心且极其务实。他们寻求对业务内部的能力和机会拥有近乎超人的实时状态理解，识别执行这些机会的关键阻碍（问题），并深入到基准事实，以此作为获取和仪表化所需背景的方式，从而设计出系统性解决方案。

In repeatedly doing this, they not only develop an exceptional big picture understanding of the businesses they work within, but also translate that understanding into software that embodies that representation with more and more accuracy over time, and allows everyone in an organization to use that better global understanding to better solve their local problems.在反复这样做时，他们不仅发展出了对所工作业务的卓越宏观理解，还将这种理解转化为软件，随着时间的推移，这种表征变得越来越准确，并允许组织中的每个人利用这种更好的全局理解来更好地解决他们的局部问题。

If you can master the meta-skill of figuring out what problems in arbitrary domains are computationally tractable, you have the opportunity to be a kind of “meta-genius.” You might not start out knowing the answer to anything about CPG supply chains or hospital triage, or even how to find it, but you have ever-better heuristics for the exact right questions to ask and the right temperament to make meaningful progress along the margin. This is the core process knowledge and IP Palantir has refined with their FDE model for the past two decades. And now, with their AI FDE product, they are turning that human FDE “meta-genius” into a computable system that inherently improves as frontier AI capabilities scale.如果你能掌握弄清楚任意领域中哪些问题在计算上是可处理的这一元技能，你就有机会成为一种“元天才”。你可能一开始并不知道关于CPG供应链或医院分诊的任何答案，甚至不知道如何找到它，但你拥有越来越好的启发式方法来提出正确的问题，并拥有在边缘取得有意义进展的性情。这是Palantir过去二十年来通过其FDE模型提炼出的核心过程知识和IP。现在，随着他们的AI FDE产品，他们正在将这种人类FDE的“元天才”转化为一种随着前沿AI能力扩展而内在改进的计算系统。

For example, one reason Toyota’s cars have long been more reliable than Ford’s is the “andon cord.” At Toyota’s factories, anyone on the assembly line can pull a cord and stop the entire process if they see a defect, and the workers with the most local knowledge of the problem are expected and empowered to fix it on the spot. The economist Masahiko Aoki called companies like Toyota J-firms, where production is organized horizontally and where whoever has the most accurate local picture has the ultimate authority to act. This is in contrast to what Aoki calls the H-firm (H for hierarchy), where a worker who spots a defect must first report it to a line manager, who then escalates it up the chain until it reaches someone with the authority to act but without the local context or local actuation to act well or quickly enough for it to matter. The J-firm model only works because its workers carry a deep model of the whole plant and possess very strong judgment, earned through standard Japanese business practices like job rotation, broad training, and lifetime employment within the same firm. These practices are largely unheard of in the US, which explains why, when American car companies installed the andon cord in their factories, all they got was a much slower production process and almost no discernible increase in production quality to show for it. Palantir’s Ontology breaks this tradeoff by making the real-time institutional knowledge, knowhow, and judgment — previously only embodied in people who had spent thirty years absorbing it — accessible to every local decision maker at the exact moment of decision. Palantir’s software takes the andon cord, increases its upload and download bandwidth by a few orders of magnitude, and hands it to everyone. By giving you the ability to see reality clearly and act in accordance with that reality, Ontology allows you, like the mast did for Odysseus, to hear the siren song, experience the power of accelerating digital computation, without crashing into the cliffs.例如，丰田汽车长期以来比福特更可靠的原因之一是“安灯拉绳”（andon cord）。在丰田工厂，装配线上的任何人如果发现缺陷都可以拉动拉绳并停止整个流程，而对问题拥有最多本地知识的工人被期望并授权当场解决它。经济学家青木昌彦（Masahiko Aoki）将丰田这样的公司称为J型企业，其生产是水平组织的，拥有最准确本地图景的人拥有最终的行动权。这与青木所称的H型企业（H代表层级）形成对比，在H型企业中，发现缺陷的工人必须首先向生产线经理报告，然后逐级上报，直到到达拥有行动权限但缺乏本地背景或本地执行力的人手中，以至于无法及时或有效地解决问题。J型企业模式之所以有效，仅仅是因为其员工对整个工厂有着深刻的模型，并拥有非常强的判断力，这是通过工作轮换、广泛培训和同一公司内的终身雇佣等标准的日本商业实践获得的。这些实践在美国几乎闻所未闻，这解释了为什么当美国汽车公司在工厂安装安灯拉绳时，他们得到的只是慢得多的生产过程，而生产质量几乎没有任何明显的提高。Palantir的“本体”通过使实时的机构知识、诀窍和判断——此前仅体现于那些花了三十年时间吸收它们的人身上——在决策的瞬间对每个本地决策者可用，从而打破了这一权衡。Palantir的软件采用了安灯拉绳，将其上传和下载带宽提高了几个数量级，并将其交给了每个人。通过赋予你清晰地看到现实并根据该现实采取行动的能力，本体允许你像桅杆对奥德修斯那样，听到海妖的歌声，体验数字计算加速的力量，而不会撞上悬崖。

A Philosophy Department with a P&L拥有损益表的哲学系

But to really understand Palantir, you have to take a step back and realize that it is, first and foremost, a deeply intellectual institution. It’s basically the only available medium for the world’s most elaborate cross-disciplinary PhD thesis. (Granted, Alex Karp already has a PhD. But he’s achieved the impossible: being a genius dropout-coded person who can correctly be addressed as “Doctor.”) It’s an organization that’s architected for a never-ending quest to understand the nature and future of institutional fitness in a world dominated by ever more capable machines, and crucially, the logic that facilitates and even necessitates the connection between them.但要真正理解Palantir，你必须退后一步，意识到它首先是一个深刻的知识分子机构。它基本上是世界上最详尽的跨学科博士论文唯一可用的载体。（当然，Alex Karp已经拥有博士学位。但他实现了不可能的事情：成为一个天才辍学气质的人，却可以被正确地称为“博士”。）这是一个为永无止境的探索而设计的组织，旨在理解在一个由能力越来越强的机器主导的世界中，机构适应性的本质和未来，以及至关重要的是，促进甚至促成两者之间联系的逻辑。

Palantir understands that the progression and proliferation of machines, on their own, confer no inherent moral quality — technological progress does not automatically guarantee that humanity understands reality more clearly — or uses that understanding to increase human flourishing. It understands, too, that the coupling between machine progress and institutional fitness — each reinforcing each other — is not automatic, either. And that this symbiotic coupling is a fundamental requirement for technological progress to turn into civilizational progress, and what lets both continue.Palantir理解到，机器的进步和扩散本身并不赋予任何内在的道德品质——技术进步并不自动保证人类更清晰地理解现实，或利用这种理解来增加人类的福祉。它也理解到，机器进步与机构适应性之间的耦合——两者相互加强——也不是自动的。而且，这种共生耦合是技术进步转化为文明进步的基本要求，也是让两者都能持续下去的原因。

Palantir’s implicit belief is that Silicon Valley has lost sight of this fundamental truth, believing instead either that technology will do so automatically (think the effective accelerationists), or that its own superior understanding of said technology anoints it to decide how institutions and society ought to be organized (think effective altruists). Palantir is a deeply paradoxical company in that what it shares with these movements is a grand vision, and what it adds is grounding this mission in deep respect for the parts of that status quo that work well. They take Chesterton’s fence, particularly in the context of institutions, seriously: you should never destroy or reinvent institutional norms without first deeply understanding why they exist.Palantir的隐含信念是，硅谷已经失去了对这一基本真理的关注，转而相信技术会自动实现这一点（想想有效加速主义者），或者相信其自身对上述技术的优越理解赋予了它决定机构和社会应如何组织的权力（想想有效利他主义者）。Palantir是一家极其矛盾的公司，它与这些运动的共同点是宏伟的愿景，而它所增加的是将这一使命建立在对现状中有效部分深切尊重的基础上。他们认真对待切斯特顿的栅栏（Chesterton’s fence），特别是在机构背景下：在没有深刻理解它们为何存在之前，绝不应销毁或重塑机构规范。

Palantir is a “cult of skeptical people who care deeply and want to argue about where the world is going and how software fits into it.” – Nabeel Qureshi, Reflections on Palantir, 2024“Palantir是一个由持怀疑态度的人组成的邪教，他们深切关心并想要争论世界正走向何方以及软件如何融入其中。”——Nabeel Qureshi，《Palantir反思》，2024年

Palantir is a cult that continuously asks: How can machines, particularly powerful and ubiquitous digital computers, continuously improve critical institutions’ ability to effectively serve the needs of their markets and constituencies, without trading off personal or institutional sovereignty and liberty? But, because they’re working with specific customers, that’s not a general question that leads to a ‘90s Wired cover story about how This Changes Everything, but a specific, concrete way to use particular tools to improve one thing at a time.Palantir是一个不断追问的邪教：机器，特别是强大且无处不在的数字计算机，如何才能在不牺牲个人或机构主权与自由的情况下，不断提高关键机构有效服务其市场和选民需求的能力？但是，因为他们与特定客户合作，这并不是一个会导致90年代《连线》杂志封面故事关于“这如何改变一切”的一般性问题，而是一种使用特定工具一次改进一件事的具体、具体的方法。

Palantir understands that high-g, generally intelligent agents capable of high-fidelity tool use have roamed the earth for hundreds of thousands of years, but it’s only been in the last few hundred that those agents have been somewhat optimally “deployed” to continuously improve our collective understanding of the world and translate that understanding into solutions that benefit humanity. Even in the last generation, talent-spotting has gotten far more rigorous across many domains — exceptionally tall and athletic kids will hear about the NBA earlier even if they don’t care about sports, kids with lightning-quick mental math skills recognize the initials “IMO,” and writers can find fame online even if the New York Times would have instantly discarded their résumé.Palantir理解到，具有高保真工具使用能力的高重力、通用智能代理已经在地球上游荡了数十万年，但只是在过去几百年里，这些代理才在某种程度上被最优地“部署”，以不断提高我们对世界的集体理解，并将这种理解转化为造福人类的解决方案。即使在上一代，人才发掘在许多领域也变得严谨得多——身材高大、运动能力强的孩子即使不关心体育也会更早听说NBA，拥有闪电般心算能力的孩子能认出“IMO”（国际数学奥林匹克竞赛）的首字母，而写作者即使《纽约时报》会立即丢弃他们的简历，也能在网上成名。

Why? Primarily because of institutions. Institutions are generators of economic and civilizational reward functions. The reward functions that drive progress are the ones that reinforce actions that improve our collective understanding of the universe — giving us dopamine from productive curiosity, for improving our ideas — and that translate that understanding into useful solutions to human problems. Reward functions that reinforce essentially anything else (particularly sensory novelty at the expense of epistemic novelty, or certainly anything that creates fear), will probably lead to stagnation. And stagnation almost always reinforces zero-sum behaviors that tend toward civilizational collapse.为什么？主要是因为机构。机构是经济和文明奖励函数的生成器。驱动进步的奖励函数是那些加强改善我们对宇宙集体理解的行为——通过生产性的好奇心给予我们多巴胺，用于改进我们的想法——并将这种理解转化为解决人类问题的有用方案。强化基本上任何其他事物的奖励函数（特别是以牺牲认知新颖性为代价的感官新颖性，或者肯定是任何制造恐惧的事物），很可能会导致停滞。而停滞几乎总是会加强趋向文明崩溃的零和行为。

Joel Mokyr, Philippe Aghion, and Peter Howitt won the Nobel Prize last year for formalizing this insight: progress accelerates during periods of time when the feedback loop between improving our collective understanding of the universe (science) and “making things happen” with that understanding (technology) are tight. Institutions that tighten the feedback loop — that reward epistemic novelty, and the application of epistemic progress to solving real problems — facilitate progress. Institutions that do this with the highest fidelity and speed create upward discontinuities in progress. Those that do almost anything else create downward ones.Joel Mokyr、Philippe Aghion和Peter Howitt去年因将这一洞察形式化而获得诺贝尔奖：在改善我们对宇宙的集体理解（科学）与利用这种理解“使事情发生”（技术）之间的反馈循环紧密时期，进步会加速。紧缩反馈循环的机构——奖励认知新颖性，并将认知进步应用于解决实际问题的机构——促进了进步。以最高保真度和速度做到这一点的机构创造了进步的向上断层。而做其他任何事情的机构则创造了向下的断层。

Capitalism, free markets, and Western institutions created the preconditions that allow us to continuously generate reward functions aligning our actions to the objective improvement of our collective understanding of nature and ourselves. Technology is a physical instantiation of that understanding, and when it’s applied against these reward functions, it helps us solve important problems. And the ultimate goal of technology is to ensure that our institutional reward functions increase our collective consciousness.资本主义、自由市场和西方机构创造了前提条件，使我们能够不断生成将我们的行动与对自然和我们自身的集体理解的客观改善相一致的奖励函数。技术是这种理解的物理实例化，当它应用于这些奖励函数时，它帮助我们解决重要问题。技术的最终目标是确保我们的机构奖励函数增加我们的集体意识。

Free-market capitalists would say that this reward function is as close as we have ever gotten (and are likely to get) to one that leads to never-ending progress and human flourishing: People who invent world-changing technology get heavily compensated for it. They might be panglossian about this from time to time — sports gambling may be a triumph of capitalism for particular capitalists, but it isn’t Western civilization’s greatest achievement. But overall, Silicon Valley would probably say that its reward function, its understanding of capitalism, truly reinforces this ideal today. They’d say that it’s a place where social validation and sensory novelty happen to correlate with actual progress. Bad environments give you dopamine for signaling; good ones give you dopamine for building, and Silicon Valley, they’d argue, is a good one. It’s good because the people who built it and live in it have high IQ scores so their approval tends to track genuine insight rather than performativity.自由市场资本主义者会说，这种奖励函数是我们有史以来（且很可能得到）最接近导致永无止境的进步和人类繁荣的函数：发明改变世界技术的人会为此获得丰厚的补偿。他们有时可能对此过于乐观——体育博彩可能是特定资本家眼中资本主义的胜利，但它并不是西方文明最伟大的成就。但总的来说，硅谷可能会说，它的奖励函数，它对资本主义的理解，在今天确实加强了这一理想。他们会说，这是一个社会认同和感官新颖性恰好与实际进步相关联的地方。糟糕的环境会因为信号传递而给你多巴胺；良好的环境会因为建设而给你多巴胺，而他们会争辩说，硅谷就是一个良好的环境。它之所以好，是因为建造它并生活在其中的人拥有高智商分数，所以他们的认可往往追踪真正的洞察力而非表演性。

One of Palantir’s core insights (mostly that of Palantir cofounder and chairman Peter Thiel) is that the Church of Silicon Valley has been largely usurped by the Pharisees (the signallers). That the reward function of capitalism, Silicon Valley’s or not, has become increasingly decoupled from progress (we wanted flying cars, but we got 140 characters); but also that this can be corrected. Palantir’s counter-position is that institutions are the ultimate meta-technology, and that natively connecting them to material technology is fundamental to re-accelerating progress. In other words, we must harness material technology to improve our institutions’ ability to generate the right reward functions. Material technology lets us mutate the world in countless ways, but it is our institutions that select the useful among these mutations. Human brains actually weren’t all that useful before institutions connected them to these reward functions and coordinated them at scale toward productive ends!Palantir的核心洞察之一（主要是联合创始人兼董事长彼得·蒂尔的洞察）是，硅谷教会已在很大程度上被法利赛人（信号传递者）所篡夺。资本主义的奖励函数，无论是否是硅谷的，已经越来越与进步脱钩（我们想要飞行汽车，但我们得到了140个字符）；但也认为这是可以纠正的。Palantir的对立立场是，机构是终极的元技术，将它们原生连接到物质技术对于重新加速进步至关重要。换句话说，我们必须利用物质技术来提高我们机构生成正确奖励函数的能力。物质技术让我们以无数种方式改变世界，但正是我们的机构在这些突变中选择了有用的那些。在机构将人类大脑与这些奖励函数连接起来并大规模协调它们以实现生产性目标之前，人类大脑实际上并没有那么有用！

Palantir doesn’t believe that superior understanding of computers confers some supernatural ability to create more effective institutions, to dictate how society ought to be organized, or to posit the optimal set of moral and philosophical norms that should govern it. Being a billionaire, or an AI researcher, or an enterprise software founder is orthogonal to your ability to litigate the big questions facing humanity. Instead, Palantir’s answer, in the context of software, is to go to ground truth. That means embedding would-be Silicon Valley engineers into institutions that have trusted access to rich, local realities. These are environments where domain experts already understand their problems on a fundamental level but would benefit enormously from connecting that localized knowledge to the accelerating power of digital computation and making it programmable.Palantir并不认为对计算机的优越理解赋予了某种超自然的能力来创造更有效的机构，来决定社会应如何组织，或来提出应治理它的最优道德和哲学规范集。成为亿万富翁、AI研究人员或企业软件创始人与你处理人类面临的重大问题的能力是正交的。相反，在软件背景下，Palantir的答案是回归基准事实。这意味着将未来的硅谷工程师嵌入到那些对丰富、本地现实拥有信任访问权限的机构中。这些环境中的领域专家已经在根本层面上理解他们的问题，但如果能将这种本地化知识与数字计算的加速力量连接起来并使其可编程，将受益匪浅。

To scale the fitness of pro-human institutions with the increasing power of digital computation, Palantir had to fundamentally be in the business of breaking tradeoffs. From the beginning, particularly in the post-9/11 era, it rejected the pervasive notion that leveraging computation to improve state security capabilities inherently required sacrificing privacy and sovereignty, a false dichotomy back in vogue as AI capabilities accelerate. The core innovation of Palantir’s original Gotham platform was an architecture that gave intelligence analysts the exact information necessary to achieve their mission at the right time, but enforced strict, purpose-based access and full accountability. It didn’t allow for a sweeping ability to find out anything about anyone.为了随着数字计算力量的增长来扩展亲人类机构的适应性，Palantir必须从根本上致力于打破权衡。从一开始，特别是在9/11后的时代，它就拒绝了一种普遍的观念，即利用计算来提高国家安全能力必然要求牺牲隐私和主权，这是一种随着AI能力加速而再次流行的错误二分法。Palantir最初的Gotham平台的核心创新是一种架构，它在正确的时间为情报分析师提供了实现其任务所需的精确信息，但强制执行了严格的、基于目的的访问和完全的问责制。它不允许以扫荡式的方式查找关于任何人的任何信息。

It’s fruitful to contrast this symbiotic model to the big consumer tech companies built on ads like Google and Meta. These companies are also trying to understand the world, and their profits measure how well they’re doing it. But ad companies have a double incentive to hoard their knowledge. First, their users and their regulators care about privacy: Google might target an ad for a product that helps with an embarrassing medical condition, or a substance abuse problem, but they don’t want to be in the business of selling advertisers a bundle of personal information about everyone whom they’ve inferred has these problems. Second, an ad system that’s an inscrutable black box to advertisers, where they set a budget and a goal and have no idea how this is accomplished, is great for pricing power. Advertisers know if they’re getting an acceptable return on their investment, but they don’t know how much of the return the platform itself is capturing. Over time, that commercial imperative is to cede more of the customer relationship, business knowledge, and their own agency to a few big platforms.将这种共生模型与谷歌和Meta等基于广告构建的大型消费科技公司进行对比是很有意义的。这些公司也在试图理解世界，它们的利润衡量了它们做得有多好。但广告公司有双重动机来囤积知识。首先，他们的用户和监管机构关心隐私：谷歌可能会针对有助于解决尴尬医疗状况或药物滥用问题的产品投放广告，但他们不想从事向广告商出售关于他们推断出患有这些问题的所有人的个人信息包的业务。其次，一个对广告商来说是不可知的黑盒的广告系统，在这种系统中，他们设定预算和目标，却不知道这是如何完成的，这对定价权非常有利。广告商知道他们是否获得了可接受的投资回报，但他们不知道平台本身捕获了多少回报。随着时间的推移，这种商业命令是让客户关系、业务知识和他们自己的代理权更多地让渡给少数几个大平台。

Palantir funds this quest, and commercializes its understanding, by building software that, by design, helps improve the fitness of institutions that empower humanity, and hurts those that don’t. Crucially, beyond a clear commitment to the US’ elected government (Karp has said that their goal is to power the West to its obvious innate superiority), it doesn’t claim final authority over which institutions are moral and which are not. Palantir instead delegates the question of “how society should be organized or what justice requires” to democratic institutions and the free market. Palantir’s job is to help them execute that vision better. Their technology, like the Company’s namesake (the palantiri, seeing-stones in The Lord of the Rings), strives to help pro-human institutions more effectively discover the Thielian secrets that make them better at serving their markets and the public, and more antifragile against adversaries.Palantir通过构建软件来资助这一追求并商业化其理解，这些软件在设计上旨在帮助提高那些赋能人类的机构的适应性，并伤害那些不赋能人类的机构。至关重要的是，除了对美国民选政府的明确承诺（Karp曾表示他们的目标是推动西方发挥其显而易见的内在优越性）之外，它并不声称对哪些机构是道德的、哪些不是拥有最终权威。Palantir反而将“社会应如何组织或正义需要什么”的问题委托给民主机构和自由市场。Palantir的工作是帮助他们更好地执行这一愿景。他们的技术，就像公司名字的来源（《指环王》中的真知晶球，palantiri），努力帮助亲人类机构更有效地发现那些使它们更善于服务市场和公众，并对对手更具反脆弱性的蒂尔式秘密。

“Effective Software” and Economic Computing“有效软件”与经济计算

Palantir calls this technology “effective software,” and is wholly oriented toward its successful creation. To Palantir, effective software is that which enables critical institutions to turn information about the world into the decisions — useful states — that benefit their constituents in the good times, and ensure survival in the bad. It lets pro-human institutions scale with the increasing power of digital computation, rather than be left behind, obviated, or destroyed by it.Palantir将这种技术称为“有效软件”，并完全致力于其成功创造。对Palantir而言，有效软件是指那些使关键机构能够在好时光中将关于世界的信息转化为造福其选民的决策（有用的状态），并在坏时光中确保生存的软件。它让亲人类机构能够随着数字计算力量的增长而扩展，而不是被其抛在后面、淘汰或摧毁。

“Our culture and means of organizing ourselves are preconditions for the creation of effective software.”“我们的文化和组织方式是创造有效软件的前提条件。”

- Alex Karp, Palantir S-1, 2019——Alex Karp，Palantir S-1招股书，2019年

All this might sound like it has very little to do with software. To many, it probably bears a closer resemblance to how we traditionally think about the ideal form of Firms, Economies, or Governments themselves: systems that, when functioning properly, continuously improve their ability to turn information into economically or societally valuable solutions.所有这些听起来可能与软件关系不大。对许多人来说，它可能与我们传统上思考企业、经济或政府本身理想形式的方式更为相似：当功能正常时，这些系统不断提高其将信息转化为具有经济或社会价值的解决方案的能力。

One way to conceptualize Palantir is that it is not a software company at all, but a meta-institution architected to turn institutional fitness into an engineering and information-management problem. A system that continuously searches over complex institutional environments with a reward function tied exclusively to improvements in institutional fitness, and with Ontology-armed FDEs as the actuators that close the learning loop.将Palantir概念化的一种方式是，它根本不是一家软件公司，而是一个旨在将机构适应性转化为工程和信息管理问题的元机构。一个不断在复杂的机构环境中搜索，且奖励函数完全与机构适应性改进挂钩，并以装备了本体的FDE作为关闭学习循环的执行机构的系统。

Palantir is spiritually and organizationally closer to TSMC than it is to OpenAI or Anthropic, in the sense that they embody extremely complex process knowledge, born from an organizational structure and execution philosophy that has, from day one, prioritized trusted access to scarce institutional context — and over and over again, honored that trust by increasing their customers’ ability to wield computation in an increasingly sovereign way. TSMC may help you manufacture your chips, but they will never, ever compete with your design, or use your IP to help others compete with you. It’s organized in every way to maximize your odds of success. Similarly, Palantir will help you build the computational substrate that makes programmable your process of turning local information about human needs into economically and societally valuable solutions, but without ever using that information to compete with you and in a way that you can own.Palantir在精神和组织上更接近台积电（TSMC），而非OpenAI或Anthropic，因为它们体现了极其复杂的流程知识，这些知识源于一种组织结构和执行哲学，从第一天起就优先考虑对稀缺机构背景的信任访问——并且一次又一次地通过提高客户以日益主权的方式运用计算的能力来尊重这种信任。台积电可能会帮助你制造芯片，但它们绝不会与你的设计竞争，也不会利用你的IP帮助他人与你竞争。它在各方面都旨在最大化你的成功几率。同样，Palantir将帮助你构建计算基底，使你将关于人类需求的本地信息转化为具有经济和社会价值的解决方案的过程变得可编程，但绝不会利用这些信息与你竞争，且是以一种你可以拥有的方式。

Progress in this economic computing revolution that Palantir is powering is not measured by Moore’s Law, FLOPs per watt, “intelligence per watt,” tokens per dollar, or the like but by something much more ambitious: the density of programmable economic complexity (economically valuable knowledge and knowhow) accessible by and sovereign to pro-human people and institutions, per unit energy.Palantir正在推动的这场经济计算革命的进步，不是由摩尔定律、每瓦FLOPs、每瓦“智能”、每美元Token等来衡量，而是由更雄心勃勃的东西来衡量：每单位能量下，亲人类的人和机构可访问且拥有主权的可编程经济复杂性（具有经济价值的知识和诀窍）的密度。

For example, Sepsis Hub, built on Palantir’s Foundry and owned by Tampa General Hospital, is a system that encodes clinicians’ tacit knowledge of what early sepsis looks like, monitors around 1,000 inpatients at a time against that embodied knowledge, and integrates vitals, labs, and other indicators to accelerate antibiotic treatment. Since launch it’s credited with halving overall sepsis mortality, a 68% reduction in 48-hour mortality, and saving roughly 900 lives as of June 2026. It’s also led to a 30% reduction in sepsis length of stay. Other systems that Palantir helped build for Tampa General have led to hospital wide efficiencies, including an 83% reduction in patient placement time.例如，构建在Palantir Foundry之上并由坦帕总医院拥有的“败血症中心”（Sepsis Hub），是一个编码了临床医生关于早期败血症表现的默会知识的系统，它一次监控约1000名住院患者，对照这些体现的知识，并整合生命体征、实验室数据和其他指标以加速抗生素治疗。自推出以来，据信它将整体败血症死亡率减半，48小时死亡率降低了68%，截至2026年6月挽救了约900人的生命。它还使败血症住院时间缩短了30%。Palantir帮助坦帕总医院构建的其他系统也带来了全院范围的效率提升，包括患者安置时间缩短了83%。

“We are paid, and have always aspired to be paid, as a derivative of value creation. Our results represent a subset of the economic value that our software has created for our customers.” – Alex Karp, Shareholder Letter, Q2 2026“我们获得报酬，并且一直渴望获得报酬，作为价值创造的衍生品。我们的结果代表了我们的软件为客户创造的经济价值的一个子集。”——Alex Karp，股东信，2026年第二季度

Palantir makes institutional memory, an institution’s search and learning process, computable. But these computers (which Palantir helps design and manufacture) are wholly owned by the institutions they power. Palantir understands that you cannot separate this economic computer from the institution it powers: doing so is like removing memory from a GPU, stripping a business of its core means of production. There may be different bundles for different use cases — a personal computer is useless without a monitor, but giving a data-center computer a monitor is useless when its behavior is being orchestrated at scale. You also cannot aggregate the memory of many different economic computers to create one that performs equally well for all institutions: the state of the computer, and its ability to create new useful states, is entirely a function of the singular context from which it’s born. In other words, “geniuses in a data center” is an incoherent concept. Aggregating genius misunderstands what genius is: singular intellectual synesthesia born of selective adaptation to a specific pocket of reality.Palantir 将机构记忆——即机构的搜索与学习过程——变得可计算。但这些由 Palantir 协助设计和制造的计算机，完全归其所赋能的机构所有。Palantir 深知，你无法将这种经济计算机与它所赋能的机构剥离：这样做就像从 GPU 中移除内存，剥夺了企业核心的生产资料。针对不同的使用场景，可能会有不同的配置方案——个人电脑如果没有显示器就毫无用处，但当数据中心计算机的行为在大规模环境下被编排时，为其配备显示器则是多此一举。你也不能将许多不同经济计算机的记忆进行汇总，从而创造出一台对所有机构表现同样出色的计算机：计算机的状态及其创造新有效状态的能力，完全取决于它所诞生的独特环境。换句话说，“数据中心里的天才”是一个逻辑不通的概念。汇总天才误解了天才的本质：天才是一种源于对特定现实领域进行选择性适应而产生的独特智力联觉。

“We have always declined, and will continue to decline, entering into a parasitic relationship with our partners. It is true that our approach has diverged sharply from the market in this regard. And yet we are one of the few companies in the technology industry that is fully aligned with its customers.”“我们一直拒绝，并将继续拒绝与我们的合作伙伴建立寄生关系。诚然，我们在这一点上与市场的做法截然不同。然而，我们是科技行业中为数不多的与客户完全保持一致的公司之一。”

- Alex Karp, Q2 2026 Shareholder Letter—— 亚历克斯·卡普（Alex Karp），2026 年第二季度股东信

Palantir vertically integrated to understand the logic of economic computing, to tease out its value networks, and to build a map for optimal institutional design in a world of accelerating computation. Most fundamentally, they believe that durably selected knowledge, embodied in Western institutions and free markets, should be made more effective, and preserved, by technological progress. It is not something that a few engineers, by virtue of understanding computers, have the mandate of heaven to rewrite in their own conception.Palantir 进行了垂直整合，旨在理解经济计算的逻辑，梳理其价值网络，并为在计算加速的世界中实现最优机构设计构建蓝图。从根本上说，他们认为，体现于西方机构和自由市场中的持久选择性知识，应当通过技术进步变得更加高效并得以保存。这并非少数工程师凭借对计算机的理解，就拥有“天命”可以按自己的构想去重写的东西。

So, in a way the labs are right to try copying Palantir, but they don’t realize they are actually copying the wrong thing, and missing why Palantir’s customers actually trust them. The labs’ implicit goal with the vertical integration of deployment seems to be increasing their models’ ability to do everything in the economy, by seeing everything in the economy. But being aligned with your customers, partners, and the broader ecosystem of economic actors is logically incoherent with that goal. So, the labs end up competing directly with the enterprises building on top of them — from encroaching on Cursor’s territory with native coding agents (think Claude Code and OpenAI Codex) to threatening Harvey with specialized offerings in the legal domain.因此，在某种程度上，这些实验室试图模仿 Palantir 是对的，但他们没有意识到自己模仿错了对象，也错失了 Palantir 的客户真正信任他们的原因。这些实验室在部署方面进行垂直整合的隐含目标，似乎是通过洞察经济中的一切，来增强其模型在经济中处理一切事务的能力。然而，与客户、合作伙伴以及更广泛的经济参与者生态系统保持一致，在逻辑上与该目标是相悖的。因此，这些实验室最终直接与其生态系统中的企业形成了竞争——从通过原生编码代理（如 Claude Code 和 OpenAI Codex）侵占 Cursor 的领地，到通过专门的法律领域产品威胁 Harvey。

And lately, this has also meant competing with the small businesses that form the character of our economy. For example, Thrive Holdings (a private equity firm partially owned by, and deeply intertwined with, OpenAI) has acquired 70 account and IT services firms, and essentially turned them into reinforcement learning environments for OpenAI’s frontier models to hill climb. It’s easy enough to understand why any private equity firm would be interested in such businesses and making them run better with technology; but here, the interest is tied to the business practices, clickstreams, and localized context that OpenAI’s researchers and engineers need to build models capable of automating work in those domains. In this paradigm, vertical integration is purportedly about empowering the customer, but in the fullness of time is about capturing their means of production.最近，这意味着还要与构成我们经济特征的小型企业竞争。例如，Thrive Holdings（一家由 OpenAI 部分持有并与其深度交织的私募股权公司）已经收购了 70 家会计和 IT 服务公司，并实质上将它们变成了 OpenAI 前沿模型进行爬坡训练的强化学习环境。不难理解为什么任何私募股权公司都会对这些业务感兴趣，并希望利用技术让它们运行得更好；但在这里，这种兴趣与业务实践、点击流以及本地化背景紧密相关，而这些正是 OpenAI 的研究人员和工程师构建能够实现这些领域工作自动化模型所需的内容。在这种范式下，垂直整合表面上是为了赋能客户，但从长远来看，却是为了攫取他们的生产资料。

“The businesses we acquire represent the right reward systems for this evolution, bringing together industry expertise and real-world data that can help improve models on specific tasks and capabilities.” – Joshua Kushner, OpenAI Partnership Announcement“我们收购的企业为这种演进提供了正确的奖励机制，汇集了行业专业知识和真实世界数据，有助于改进模型在特定任务和能力上的表现。”—— 约书亚·库什纳（Joshua Kushner），OpenAI 合作伙伴关系公告

“By training the most advanced models for specific tasks within our businesses, guided by both company-specific data and expert feedback, we believe we can continuously improve model capabilities…” – Thrive Holdings, OpenAI Partnership Announcement“通过在我们的业务范围内针对特定任务训练最先进的模型，并在公司特定数据和专家反馈的指导下，我们相信能够持续提升模型能力……”—— Thrive Holdings，OpenAI 合作伙伴关系公告

“There are Marxist overtones and undertones to our business. Others, including many of those building large language models, intend, knowingly or otherwise, to capture the means of production of their purported partners.” – Alex Karp, Q2 2026 Shareholder Letter“我们的业务中存在马克思主义的弦外之音和潜在含义。其他人，包括许多构建大语言模型的人，无论是有意还是无意，都打算攫取其所谓合作伙伴的生产资料。”—— 亚历克斯·卡普（Alex Karp），2026 年第二季度股东信

Palantir, on the other hand, vertically integrated deployment to build the capabilities and process knowledge (Forward Deployed Engineering, Ontology, AIP) required to make building institution-specific economic computers as easy to build as possible. If Palantir builds a sovereign economic computer for the CIA, the UK’s NHS becomes more amenable to the idea, and so do market leaders like Airbus and Kirkland & Ellis. Soon, the most critical institutions in society are the ones most invested in pushing forward the economic frontier: they reap some of the rewards, while gaining more control over their own destinies. But if FDE and vertical integration of AI deployment simply mean using trusted access to institutional context as a way to improve your ability to rent-seek, or worse, to build your customers’ or partners’ replacement (in the case of Thrive Holdings’ partnership with OpenAI), you jeopardize the whole promise of economic computing: making progress programmable in a way everyone can contribute to and own.相比之下，Palantir 进行部署的垂直整合，是为了构建所需的能力和流程知识（前沿部署工程、本体论、AIP），从而使构建机构特定的经济计算机尽可能简单。如果 Palantir 为中央情报局（CIA）构建了一台主权经济计算机，英国的国家医疗服务体系（NHS）就会更容易接受这一理念，空中客车（Airbus）和凯易（Kirkland & Ellis）等市场领导者也是如此。很快，社会中最关键的机构将成为最致力于推动经济前沿发展的机构：它们在获得部分回报的同时，也获得了对自己命运的更多掌控权。但如果 FDE 和 AI 部署的垂直整合仅仅意味着利用对机构背景的信任访问权来增强寻租能力，或者更糟的是，构建替代客户或合作伙伴的产品（如 Thrive Holdings 与 OpenAI 的合作案例），那么你就危及了经济计算的整个承诺：使进步变得可编程，让每个人都能参与并拥有。

The lesson from Palantir’s 23 years is to vertically integrate for innovation. Do it if it’s the only way to get permissioned access to the ground truth knowledge needed to help your customers and society uncover new, valuable secrets. But don’t vertically integrate for rent or to concentrate power. Not only does rent-seeking actually reduce your own institutional fitness, but it also creates a broader negative selection pressure in society — one born from the false idea that there are no secrets left to discover.Palantir 23 年历程带来的教训是：为了创新而进行垂直整合。如果这是获得许可访问基础事实知识以帮助客户和社会发现新的、有价值秘密的唯一途径，那就去做。但不要为了寻租或权力集中而进行垂直整合。寻租不仅会降低你自身的机构适应性，还会给社会带来更广泛的负面选择压力——这种压力源于一种错误的观念，即认为已无秘密可寻。