Palantir CEO Warns AI Misuse Risks Regulation, Nationalization

Palantir首席执行官警告：滥用AI恐引发监管与国有化风险 Alex Karp指出，企业的“tokenmaxxing”（盲目消耗Token）制造了无用的副产品，而未解决的政治风险更使人工智能行业面临被政府接管的危险。 Palantir首席执行官Alex Karp做客TBPN，探讨了为何品味、领域专业知识以及类似Palantir Ontology的结构化软件原语，要优于盲目的AI算力和原始的LLM输出。他告诫企业警惕“tokenmaxxing”——即将AI模型用于缺乏明确战略影响力的低价值任务，并将浮于表面的炒作与深度的企业级落地执行进行了对比。此外，Karp还探讨了国防技术的竞争护城河、AI在提升一线员工技能而非引发大规模裁员方面的作用，以及AI面临即将到来的监管与国有化的关键风险。 Palantir首席执行官Alex Karp保持着顶级的体能训练习惯，并分享了一套每周精准执行的训练方案，用于最大化握力耐力并创下悬吊5分30秒的个人纪录。 (0:00) 企业在普及AI应用时正面临阻力，原因在于“tokenmaxxing”现象，即员工在浮于表面、低价值的任务中过度使用LLM，耗费资金却未能解决核心业务问题。 (2:57) Palantir在内部开发了一项被戏称为“反自嗨工具”的能力，旨在遏制无目的的AI使用，促使员工专注于富有成效地解决问题，而非沉迷于生成式炒作。 (2:57) 在AI时代，企业的终极竞争优势在于“品味与资金的结合”——即能够识别关键问题并围绕这些问题组织数据结构的运营洞察力。 (5:09) 国防、金融和物流等领域的一线企业级问题，需要持久的知识库和结构化原语，这是独立的LLM所无法复现的。 (5:09)

Palantir CEO Alex Karp contends that enterprise artificial intelligence has devolved into low-value productivity theater driven by unguided token consumption. He argues that lasting competitive advantage belongs to organizations that combine capital with organizational taste and structured data architectures rather than raw model generation. Enterprise AI Disillusionment and Tokenmaxxing (2:57) Karp observes that enterprise AI adoption has moved past initial skepticism into a phase of quiet frustration. While financial markets have rallied around frontier models, corporate executives increasingly recognize that large language models (LLMs) fail to resolve core business challenges on their own. Instead, organizations face widespread 'tokenmaxxing'—employees consuming expensive compute on low-value tasks like classifying emails or checking the weather, yielding superficial 'tool-shaped objects' that obscure true productivity. To counter this aimless usage internally, Palantir engineered a capability informally dubbed the 'de-masturbatory tool.' Karp emphasizes that throwing capital and compute at operational bottlenecks cannot substitute for clear strategy, noting that unguided model deployment generates capital waste rather than durable enterprise value. Architectural Primitives Versus Probabilistic Code (5:09) Critical enterprise and defense workflows—such as financial underwriting, energy exploration, and military logistics—require persistent knowledge stores and deterministic logic rather than probabilistic text generation. Karp outlines a three-tier hierarchy of enterprise code: foundational primitive infrastructure developed over millions of engineering hours, forward-deployed engineer code that integrates operations into a scalable product framework, and lightweight, probabilistic AI-generated code. While probabilistic code excels at ad-hoc dashboards, Karp argues it cannot serve as a reliable knowledge store. Furthermore, Karp notes that frontier models can expose security vulnerabilities at 10x to 100x speed without offering native, on-premises remediation. Because commercial enterprises and national security entities refuse to transmit sensitive formulas or classified intelligence to public clouds, Karp views LLMs not as standalone enterprise solutions, but as catalysts driving clients toward structured governance platforms like Palantir's Ontology and Apollo. Organizational Taste and Structural Moats (14:14) Karp identifies 'taste'—the discernment required to separate genuine operational insights from hollow industry buzzwords—as an enterprise's decisive asset. In practical terms, taste dictates architectural governance: determining which datasets must remain strictly on-premises, which workloads can safely leverage the public cloud, and how technical teams are deployed. Addressing competition, Karp argues that the influx of approximately 50 defense tech entrants ultimately validates the sector and expands total defense tech procurement budgets. He maintains that surface-level competitors cannot quickly replicate Palantir's structural moat, estimating it would take a rival at least three years to recreate the company's deep enterprise integrations and operational Ontology. Regulatory Exposure and Frontline Upskilling (18:10) Karp warns that the technology industry risks heavy-handed government intervention or outright nationalization. He criticizes Silicon Valley executives for assuming immunity from regulation, cautioning that failing to proactively engage in policy and risk discussions will invite restrictive rules from politicians across the political spectrum who distrust the industry. Simultaneously, Karp cautions Fortune 500 leadership against attributing mass layoffs to artificial intelligence, warning that using AI to mask underlying operational bloat invites intense public backlash. Instead of viewing AI as an instrument for headcount reduction, he advocates leveraging software to upskill frontline workers—such as soldiers, truck drivers, and technicians—allowing non-specialists to execute sophisticated, data-driven decisions on the ground. While Karp positions frontier language models as accelerators for Palantir's infrastructure, his thesis rests on the presumption that unanchored AI usage will trigger corporate fatigue and regulatory pushback. The long-term efficacy of this strategy depends on whether enterprises prioritize deep, structural integration over cheaper, commoditized AI tools.

Palantir首席执行官Alex Karp指出，当前企业级人工智能正陷入缺乏实际价值的“Tokenmaxxing”内耗，企业真正的核心壁垒在于将资本与战略“品味”结合以构建深层本体架构，而非盲目调用通用大模型；同时他警告称，科技行业正面临严重的监管干预风险与不负责任的裁员叙事反噬。 企业AI的Tokenmaxxing陷阱与算力虚耗 (2:57) Karp指出，企业界对人工智能的认知正从早期的“怀疑其真实性”演变为“虽认可其能力却无法解决核心业务问题”。这一脱节催生了普遍的“Tokenmaxxing”现象，即员工将大量算力消耗在给电子邮件打标签或查询天气等琐碎低效的任务上。Karp将这类产出称为“工具状物体”，认为它们表面看似具备生产力，实质上是在浪费企业资本并掩盖了真正的商业价值，犹如在下沉的客轮上整理甲板躺椅。为此，Palantir内部甚至专门部署了限制机制，以遏制员工漫无目的的大模型依赖。 Karp进一步强调，复杂的商业挑战绝不可能仅靠堆砌资金或算力来解决。前沿基础大模型虽然将暴露系统安全漏洞的速度提高了10倍到100倍，却缺乏在企业本地（On-premises）实施修复与防御的能力。对于国防情报、能源钻探合规以及特色农业供应链等涉及核心机密的领域，企业与政府机构绝不会允许数据直接流向公共云端大模型，缺乏底层工作流集成的生成式技术正引发企业管理层的普遍疲惫。 本体架构与“品味”构筑的长期护城河 (5:09) 在Karp看来，AI时代的决定性竞争优势是“品味与资金的结合”（Taste combined with money）。基础大语言模型擅长批量生成可同质化扩展的内容，但无法替代解决现实复杂问题所需的精准决断。在军事作战或高难度商业运营中，核心价值在于辨识哪些问题真正至关重要，这高度依赖持久的“知识库”与结构化软件原语。Karp将企业代码划分为三类：凝聚数百万工程小时沉淀的基础架构原语代码、由前线部署工程师（FDE）维护的结构化应用代码，以及由AI生成的临时性、轻量级概率代码，并指出后者无法承载长期组织资产。 针对市场竞争，Karp认为模仿者的出现不仅没有削弱先锋企业，反而帮助开拓了市场认知。例如在国防科技领域，约50家竞争对手的加入向采购方证实了该板块的战略价值，进而扩大了整体市场预算。但他强调，竞争对手要完全重现Palantir的本体架构（Ontology）及深层组织集成至少需要3年时间，届时技术格局早已更迭。因此，大模型热潮实质上充当了Palantir业务的加速剂，迫使寻求AI落地的机构反向依赖其底层软件与Apollo基础设施。 监管介入风险与科技界的公共责任 (18:10) Karp在访谈中向硅谷同僚发出预警，称其在过去半年内持续提醒科技界：AI行业正面临严厉监管乃至被国家化的实质威胁。许多科技企业创始人天真地认为这种干预在美国绝不可能发生，误以为自身产品广受欢迎且天然创造价值。Karp反驳指出，公众与政治舆论正快速向强化政府监管倾斜，科技领袖若不主动坦诚面对安全风险与伦理挑战，对技术缺乏实际理解的外部立法者势必会强行介入并制定粗糙规则。 针对外界关于他曾呼吁“重启征兵制”的批评，Karp明确予以澄清，表示其真实提议是设立某种广泛参与的国家公民义务（例如每周参与公共公园维护等社区服务），以期在技术剧变期修复社会撕裂。他主张，让更广泛的普通公众融入国家重大战略决策本质上是一种反战举措，因为历史上的政策失误往往发生在工薪阶层被完全排斥在决策层之外的时期。他敦促持理性中间立场的科技从业者摆脱松懈心态，积极参与严肃的公共辩论。 一线劳动力赋能与对裁员叙事的警示 (21:07) 针对部分《财富》500强企业首席执行官对AI与用工规模的规划，Karp对那些炫耀“因部署AI而裁撤三分之二员工”的言论提出了严词批评。他指出，许多高管实际上是在将自身经营不善或组织冗余导致的裁员甩锅给AI，这种将技术作为大规模裁员借口的炒作行为极度危险，必然会引发强烈的社会反弹与公众敌意。 与追求替代人力相反，Karp主张将人工智能视为提升基层人员专业能力的杠杆。在国防等实际应用中，Palantir的系统使受过基础职业培训的士兵、卡车司机和电池技术人员能够胜任原本复杂的数据驱动型操作。Karp预测，未来胜出的企业并非那些盲目裁员的机构，而是能够将高层战略眼光与一线操作人员的领域专业“品味”紧密结合、并借助AI赋能全体员工的组织形态。 Karp的论点凸显了当前企业AI在投资炒作与落地现实之间的巨大落差。通用大模型在缺乏结构化本体与专业品味的前提下，容易退化为产出虚假生产力的算力黑洞；而科技界若不能在深耕核心业务架构的同时妥善处理劳动力转型与监管风险，将面临严峻的社会与政治挑战。

## Background 该视频由频道 TBPN 于 2026 年 6 月 23 日发布在 YouTube 平台上，内容为在 AIPCon 10 现场对 Palantir 首席执行官 Alex Karp 进行的深度专访。根据页面介绍，对话围绕 AI 时代的“品味”（Taste）竞争优势、企业在将 AI 炒作转化为实际商业价值时的挣扎、“代币暴食”（Tokenmaxxing）带来的虚假繁荣，以及 AI 行业面临的强监管与国有化风险展开。 ## TL;DR Alex Karp 在对话中尖锐剖析了当前科技圈沉迷于“Tokenmaxxing”这一疯狂消耗算力却缺乏实际产出的乱象，指出单纯依靠通用大语言模型无法解决企业真正的核心难题。他提出“品味加资本”（Taste plus money）才是终极壁垒，强调企业必须依托本体架构（Ontology）与专有业务流程来构建核心竞争力，大模型只是增强器而非替代品。此外，Karp 还对硅谷精英盲目自满的现状发出严厉预警，指出行业若不直面自身外部性与国家现实，恐将面临被严苛监管乃至收归国有的政治灾难；同时他驳斥了将裁员归咎于 AI 的遮羞布做法，主张通过技术赋能并提升一线劳动者的价值。 ## Atmosphere & style 整场对话在轻松幽默与极度严肃的政治商业洞察之间自由切换，开场谈及单杠悬垂健身时充满调侃与欢笑，而一旦切入行业格局则瞬间转入犀利、直率的硬核批判。Alex Karp 展现出标志性的反建制与哲学家式气场，语锋尖锐且毫无企业公关辞令的掩饰，既有对硅谷同温层狂妄自大的冷嘲热讽，也有对地缘政治与监管悬崖的深刻危机感。现场互动张力十足，听众既能感受到其天马行空、离经叛道的智识魅力，又能体会到那种对现实工业底座与严谨软件工程的绝对务实敬畏。 ## Bulletin - [03:03](https://www.youtube.com/watch?v=S9V-5VQ_Epg&t=183s) Alex Karp 批判当前企业界陷入盲目消耗算力的“Tokenmaxxing”陷阱，指出单纯堆叠 Token 无法解决核心商业问题，底层模型终将商品化，唯有界定高价值难题的“品味（Taste）”无法被算法替代。 - [05:09](https://www.youtube.com/watch?v=S9V-5VQ_Epg&t=309s) 企业级 AI 落地必须依托深厚本体架构（Ontology）与本地化（On-Prem）部署，高壁垒工业与国防部门绝不可能将核心绝密数据或知识产权交付公有云模型处理。 - [18:22](https://www.youtube.com/watch?v=S9V-5VQ_Epg&t=1102s) Karp 警告硅谷科技精英正盲目自满，若不直面外部性与地缘竞争，AI 行业未来极可能面临严苛监管甚至被“国家化（Nationalized）”的系统性危机。 - [21:05](https://www.youtube.com/watch?v=S9V-5VQ_Epg&t=1265s) 针对 500 强企业裁员潮，Karp 揭露许多高管将业务不顺借 AI 为名裁员是在推卸责任与玩火自焚；AI 的真正价值在于赋能并技能升级一线员工，而非盲目缩减人力。 ## Key takeaways - [03:03](https://www.youtube.com/watch?v=S9V-5VQ_Epg&t=183s) 企业界对 AI 的认知经历剧烈波动，从最初的观望到明知在业务中未生效却生怕显得愚蠢而不敢承认，催生出大量脱离实际业务产出的“Tokenmaxxing”泡沫。 - [03:03](https://www.youtube.com/watch?v=S9V-5VQ_Epg&t=183s) 许多公司员工沉迷于用大模型查天气、打标签等“工具形状物体”，看似繁忙实则如同在泰坦尼克号沉没前重新摆放甲板躺椅，纯属自欺欺人的工作伪装。 - [05:09](https://www.youtube.com/watch?v=S9V-5VQ_Epg&t=309s) AI 时代的核心竞争公式是“品味加资本”（Taste plus money），模型能力终将彻底商品化，能否敏锐界定出真正具备商业价值的难题才是终极分水岭。 - [05:09](https://www.youtube.com/watch?v=S9V-5VQ_Epg&t=309s) 特殊油气钻探、国防供应链与高端制造等高壁垒领域依赖专有知识库与严密流程，大语言模型仅能充当增强工具，绝无法直接取代底层业务流程。 - [05:09](https://www.youtube.com/watch?v=S9V-5VQ_Epg&t=309s) 明星大模型公司往往在风投圈光鲜亮丽，但在实际企业复杂场景落地中屡屡碰壁；Palantir 甚至通过鼓励客户先去试用大模型踩坑来转化为自身的销售优势。 - [08:24](https://www.youtube.com/watch?v=S9V-5VQ_Epg&t=504s) 公众与硅谷的狂热存在严重脱节，普通人对科技圈花着大众资金自嗨的刷量行为感到厌烦，科技企业自我循环印钞的假象掩盖了公众对其实际价值的质疑。 - [09:48](https://www.youtube.com/watch?v=S9V-5VQ_Epg&t=588s) 企业软件代码存在三个层次：由数百万工时硬编码凝结的确定性“原语（Primitives）”底座、FDE 编写的受管产品扩展代码，以及容许容错但极易让人上瘾的 LLM 自由代码。 - [11:48](https://www.youtube.com/watch?v=S9V-5VQ_Epg&t=708s) 大语言模型的爆发没有直接解决大企业核心痛点，但却像类固醇般催化了 Palantir 的业务，让全行业痛定思痛后意识到必须基于本体架构（Ontology）进行底层平台重构。 - [13:19](https://www.youtube.com/watch?v=S9V-5VQ_Epg&t=799s) 大量模仿者不仅帮 Palantir 完成了市场教育并做大国防与工业采购预算蛋糕，还在客户面前充当了天然的对照组与“表外销售资源”。 - [16:01](https://www.youtube.com/watch?v=S9V-5VQ_Epg&t=961s) 大模型生成的伪代码与表面架构无法解决实战问题，深度植入大型组织肌理的本体论架构需要数年积累，构成了追随者难以逾越的技术护城河。 - [17:11](https://www.youtube.com/watch?v=S9V-5VQ_Epg&t=1031s) 在组织管理与系统架构中，包括数据云端接入还是保留本地、哪些核心知识必须封锁等所有底层决策，本质上全依赖决策者具备公信力的“品味”仲裁。 - [18:22](https://www.youtube.com/watch?v=S9V-5VQ_Epg&t=1102s) 科技巨头普遍盲目深信美国绝不可能将企业收归国有，但政治动量正站在国家化一边，仅靠游说团体根本无法抵御缺乏技术认知的监管封锁与接管风险。 - [21:05](https://www.youtube.com/watch?v=S9V-5VQ_Epg&t=1265s) 将企业失败归咎于 AI 裁员是在向激进反资本主义阵营递刀子；技术的真正威力在于让受教育程度有限的一线人员完成技能重塑，实现单兵作战价值的指数级跃升。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=S9V-5VQ_Epg&t=0s) 录制开场与 Alex Karp 的单杠悬垂（Dead Hang）健身哲学 节目在轻松热烈的寒暄中开场。TBPN 主持人对 Alex Karp 再次做客表示欢迎，并询问这是第十届 AIPCon（AIPCon 10），与以往相比有什么新变化。Karp 笑着调侃主持人们现在更加成功、荷包满满（"some tendies in your pocket"），身形似乎也变得更加强壮。主持人则连忙致谢，坦言 Karp 上一次做客节目彻底带火了 TBPN，他是节目有史以来最大牌的嘉宾。随后话题自然转到了健身细节上，主持人追问 Karp 目前最专注的训练项目——单杠悬垂（Dead Hang）。Karp 上次透露自己的悬垂成绩接近 5 分钟，而如今在过去几个月里已经进入平台期，稳定在惊人的 5 分 30 秒。 主持人感叹，听众可能会觉得“双手挂在杠上”听起来很简单，但真正去尝试就会发现难度极高；主持人们自己也开始练习，但在 1 分 30 秒左右就会感觉肌腱快要撕裂，成绩始终卡在 2 分钟以内。Karp 点评道，1 分 30 秒已经非常体面了，而 2 分钟则属于“超一流（super elite）”水平。针对普通人训练悬垂时常犯的错误，Karp 毫无保留地分享了自己的恢复与周期训练方案：绝大多数人最大的误区就是每天都去硬吊，这会导致过度疲劳，而肌肉和肌腱最需要的是恢复。如果想要取得实质性进展，正确做法是一周只挑一天冲击极限（Maximum），但也不必拼得过于鲁莽；随后的几天以极低负荷活动为主（比如练两次 1 分钟并保持较长休息），两天前仅做 4 组 15 秒的热身微调，冲极限的前一天彻底休息。 谈笑间，主持人调侃 Karp 应该去开一门健身课作为 Palantir 的全新收入来源。Karp 则笑称自己的部分身体指标可能带有某种“外星人级别（alien territory）”的天赋基因。当主持人问及水下憋气能力时，Karp 坦言自己从小游泳，但在水下憋气方面并不擅长，甚至自认比不上主持人。身为潜水长（dive master）的主持人表示自己能憋气 3 分钟，Karp 幽默地反唇相讥，形容主持人静止时消耗极少氧气，“简直就像一头浮在水面下等待换气的鲸鱼”，引得全场大笑。 - **核心健身数据**：Alex Karp 的单杠静态悬垂目前平台期成绩为 5 分 30 秒；他认为普通人能坚持 1 分 30 秒即算合格，坚持 2 分钟便已达极高水准（Super Elite）。 - **训练建议**：每周仅安排一天冲击最大悬挂时长，严禁每日盲目死撑；其余训练日大幅递减负荷（如 2 次 1 分钟、4 次 15 秒），并在极限日前一天彻底停训以供神经与肌腱恢复。 --- ### [03:03](https://www.youtube.com/watch?v=S9V-5VQ_Epg&t=183s) AI 行业的炒作周期与“Tokenmaxxing”（代币暴食）困局 随着话题切入商业现实，Karp 梳理了过去一段时间企业界对人工智能认知的剧烈动荡。在最早期阶段，市场情绪普遍持观望态度，认为“AI 可能是真的”；随后大约两周前，整个行业陷入了一种诡异的集体心理：大家虽然惊呼“天哪，AI 确实是真实的”，但同时又心知肚明它在企业实际业务中根本没有发挥作用，只是迫于行业压力不敢公开承认，生怕自己显得愚蠢无知（"cuz we'll look stupid"）。与此同时，投资界却在一旁狂欢印钞，制造了大量的估值泡沫。如今，公众与企业终于开始认清现实，看穿了所谓的“Tokenmaxxing”（指盲目追求模型 Token 消耗量、以为 Token 烧得越多产出就越大的畸形现象）。 主持人顺势追问 Palantir 对于 Token 消耗以及企业真实落地 AI 的哲学。Karp 毫不避讳地揭露了企业内部普遍存在的 AI 滥用乱象。他透露，Palantir 内部甚至戏称某款产品模块为“戒手淫工具（de-masturbatory thing）”。在许多企业里，大量员工整天把玩 AI 工具的状态犹如沉迷于黄色内容，陷入了对“工具形状物体（tool-shaped objects）”的虚假沉迷：每天忙着用大语言模型查天气、给每封电子邮件打标签、反复微调毫无意义的格式。主持人形象地补充道，这就像在“泰坦尼克号沉没前不断重新摆放甲板上的躺椅”，纯属一种自欺欺人的工作伪装。 ```mermaid flowchart TD accTitle: AI 企业应用路径对比 accDescr: 展示盲目 Token 消耗与以解决高价值业务问题为导向的两种模式差异 subgraph Tokenmaxxing["Tokenmaxxing 虚幻模式"] A["采购前沿 LLM"] --> B["全员盲目调用 Token"] B --> C["沉迷工具表象 / 打标签 / 查天气"] C --> D["无业务产出 / 算力资本浪费"] end subgraph ValueCreation["价值驱动落地模式"] E["业务品味 Taste"] --> F["锁定核心高价值难题"] F --> G["融入行业专有本体知识"] G --> H["本地化防御部署与持续微调"] end ``` Karp 指出，当前还存在一种危险的政治叙事，尤其是欧洲那些完全不懂基础经济学的政客反而掌握了话语权，让整场关于 AI 生产力与监管的辩论脱离了商业逻辑。企业必须明白，商业问题几乎从不可能仅仅依靠砸钱、消耗更多计算 Token 就迎刃而解；资本固然可以加速进程，但前提是必须找准正确的目标与落地路径。 --- ### [05:09](https://www.youtube.com/watch?v=S9V-5VQ_Epg&t=309s) 业务问题的本质：“品味加资本”（Taste plus Money）与大模型局限性 Karp 提出了他在 AI 时代最核心的竞争哲学——“品味加资本”（Taste plus money）。他强调，包括底层基础模型在内的大多数 AI 技术组件，未来都会以极高价值但彻底商品化（commoditized）的方式快速扩展；唯独“品味”是无法被规模化复制的。所谓品味，即精准甄别“究竟哪一个业务问题才是真正值得且必须去解决的核心难题”。不论是乌克兰军队前线作战、以色列安全行动，还是大型商业实体的运作，最关键的永远是那个能敏锐判断出“这个问题极具价值，而那个问题一文不值”的决策大脑。 通用的低门槛问题（例如“写一份关于中国 GDP 增长的分析报告”）大模型固然信手拈来，但真正决定企业生死的难题都需要专属的“知识库与本体（knowledge store）”。Karp 列举了一系列硬核产业场景：符合法律合规、道德标准且能切实降低生产成本的特殊油气勘探开采技术；涉及军工、机箱制造或汽车工业的复杂供应链变革；以及金融保险业独特的承保逻辑。这些场景背后是严谨、精密的持续业务流程，大语言模型可以作为绝佳的“增强器（enhanced by LLMs）”，但绝不可能直接“取代（replaced by LLMs）”这些流程。 此外，Karp 重点强调了在模型能力提升后的安全与落地现实。大模型确实能将代码或系统漏洞的识别效率提高 10 倍乃至 100 倍，但真正的核心在于“由谁来进行修补”以及“如何在本地（on-premises）环境中打好补丁”。任何正规企业、军工部门或情报机构，绝不可能把自己的绝密数据或核心知识产权（哪怕只是大豆精准种植的专有诀窍）上传到公共云端。最后，Karp 揭示了行业内的一个巨大反差：前沿大模型公司在投资者眼中具有无与伦比的魅力（charisma），但在实际的企业客户面前却毫无吸引力。Karp 自豪地分享了 Palantir 的“独家销售秘诀”：当企业客户找上门时，他常常会说：“别急着找我们谈，你们先去跟那些明星大模型公司待上两天。要是你们运气好，等你们被折磨完了，我再考虑让你们进 Palantir 的大门。”而事实证明，在真实踩坑之后，这些企业往往会争先恐后地重回 Palantir 的怀抱。 - **核心商业公式**：品味 + 资本（Taste + Money）。通用的计算与模型能力终将商品化，唯有界定高价值问题的“品味”无法被算法扩展。 - **专有场景属性**：高壁垒产业（特殊油气钻探、国防供应链、汽车制造、特殊核保等）依赖闭环流程与专有知识库，LLM 仅能作为流程的增强组件而非替代品。 - **安全与本地化（On-Prem）诉求**：虽然模型能以 10-100 倍速度识别系统漏洞，但高密级企业与政府机构绝不会将核心数据上传公有云，修补漏洞与留存专有知识必须在本地完成。 - **企业端“反差魅力”**：明星前沿模型公司深得风险投资人青睐，但在落地企业真实业务时往往铩羽而归；Palantir 借此将“先让客户体验前沿大模型挫败感”转化为极具统治力的销售壁垒。 ### [08:24](https://www.youtube.com/watch?v=S9V-5VQ_Epg&t=504s) 脱离现实的“Tokenmaxxing”与两极分化的企业评价 Alex Karp 尖锐地指出，普通民众与硅谷当前的 AI 狂热之间存在着巨大的脱节。如果你走上街头，去和一名海军陆战队员、一位公交车司机或是公交公司的老板交谈，会发现他们不仅毫无欣喜，反而充满反感。大众已经对科技圈无休止的“tokenmaxxing”（疯狂堆砌与消耗 Token 的刷量自嗨）感到厌烦，这种行为在普通人眼里就像是在花着大众的钱进行昂贵的“自慰”。在谈及公众好感度时，Karp 提到了一个在美国商业圈常被避而不谈的词汇——“受喜爱度”（likeability）。他坦言，Palantir 在全球拥有数千万甚至上亿的支持者，但每天早晨同样有大约五百万人醒来直接把他咒骂为“撒旦”。 与许多科技巨头不同，Palantir 具有鲜明的两极分化属性（polarizing），既有狂热的信徒，也有坚定的反对者。而当下的许多 AI 追捧者和社交媒体公司则陷入了另一种困境：所有人都在使用它们，但没有任何人真正喜欢它们。不过，这些 AI 企业生活在一个自我循环的吸金闭环中，“当你照镜子时发现自己刚刚印了一大笔钱，你看起来自然容光焕发”，这种账面上的巨额收益掩盖了公众对其实际价值的质疑。 ### [09:48](https://www.youtube.com/watch?v=S9V-5VQ_Epg&t=588s) 代码的三种层次：大模型的魔力、代码成瘾与技术宗教 针对主持人的提问——前沿大语言模型（LLM）是否因其某种“魔法般”的特质，使得背后的模型厂商即便在其他业务上一塌糊涂也能保持增长，Karp 展开了深入的技术解构。他认为 LLM 确实具有魔力，但这种能力具有特定的适用边界。以 Palantir 为参照，软件代码在企业中实际上分为三个不同层次： ```mermaid flowchart TD accTitle: 代码的三种层次架构 accDescr: 展示 Alex Karp 提出的三层代码分类及其特征与适用场景 subgraph Core["底层与工程"] A["基础设施与原语<br/>(Primitives)"] -->|"历经数百万工时<br/>硬编码物理世界规律"| A1["钢铁大梁般的确定性底座<br/>(乌克兰军援/国防部系统)"] B["FDE 托管代码<br/>(Managed Code)"] -->|"围绕成熟产品库构建<br/>受控扩展"| B1["企业级敏捷交付"] end subgraph Generative["生成层"] C["自由生成代码<br/>(Free Code / LLM)"] -->|"概率驱动，快速生成<br/>允许大致准确"| C1["仪表板 / 财务分析 / 临时流转"] end ``` 第一类是基础设施代码与“原语”（primitives），例如乌克兰军方、美国国防部（Karp 戏称为“战争部”）以及大型核心企业所依赖的底座。这些代码如同建筑中的“钢铁大梁”，凝聚了数百万小时的技术积累和对复杂业务现实的深刻理解，必须具备高度的确定性，绝不可能由生成模型轻易替代。第二类是由前线部署工程师（FDE）编写的代码，其成功的秘诀在于这些代码是直接向受管理的成熟产品代码库提交并扩展产品功能，而非散兵游勇式的随意拼凑。第三类则是大模型生成的“自由代码”（free code）。这类代码极具魔力，能够以极快的速度编写出仪表板、临时性财务分析以及容许容错的概率性业务流，虽然只是“大致正确”，却足以让人惊艳。 Karp 形象地将大模型带来的快感比作某种令人上瘾的事物：人们明知过量使用有害、医生也再三警告可能带来系统性损害，却依然忍不住心想“再试一次也没什么大碍”，深陷其中不可自拔。而在特定的技术研究圈子中，AI 更是演化成了一种充满魅力的“新兴宗教”。对于许多从未有过传统宗教信仰（如犹太教、基督教或伊斯兰教）的技术人员而言，LLM 完美填补了他们内心对终极解释的空虚与渴望，甚至让人觉得所有的世间答案皆在其中。 ### [11:48](https://www.youtube.com/watch?v=S9V-5VQ_Epg&t=708s) LLM 成为催化剂：为什么企业重构必须扎根于本体（Ontology） 尽管大模型风靡一时，但 Karp 一针见血地指出，LLM 本身并没有真正解决大企业面临的根本业务痛点。然而，这并不意味着大模型毫无商业价值，相反，大模型的爆发直接给 Palantir 的业务“注射了类固醇”（putting our business on steroids）。 如果没有大语言模型的普及，外界就不会真正意识到 Palantir 底层本体架构（Ontology）的无可替代性，也不会关注到利用 Apollo 平台安全管理实战漏洞和全自动化部署的威力。当整个行业都在高喊要将业务全面代码化部署时，企业才痛苦地发现，缺乏工程底蕴的模型代码根本无法支撑严肃的生产系统。唯一的出路是基于 Palantir 进行全方位的平台重构（re-platform on Palantir）。Karp 讽刺道，那些缺乏工程品味、从未在企业软件领域摸爬滚打、纯靠臆想去构建企业级方案的外行，根本无法理解现实业务运行的精密与严酷。 - 对话中途还出现了一段关于佩戴耳麦的轻松插曲，主持人在调整设备时发现正反戴反，Karp 与主持人相视大笑，调侃着“颠倒过来就对了，这就是在追逐梦想（Living the dream）”。 ### [13:19](https://www.youtube.com/watch?v=S9V-5VQ_Epg&t=799s) 市场扩张的副产物：把模仿者变成“表外销售”与国防科技的验证 主持人提到，Palantir 过去二十年一直在深耕企业业务最底层的脏活累活，如今市场上涌现了投入数百亿美元的追随者，试图仅靠兜售“智能”概念就取代整个底层基础设施。Karp 对此显得非常坦然甚至觉得颇具戏剧性。他回顾道，当年 Palantir 默默无闻在地下室苦熬时，整个硅谷都把他们当成异类和“畸形秀”（freak show），而如今市场上至少有一半的抄袭者甚至根本没意识到自己是在亦步亦趋地模仿。 Karp 笑着吐槽一些播客节目：许多嘉宾在节目里洋洋得意地输出高见，主持人大赞“太有远见了”，但那些话不过是自己在二十或二十五年前早就反复讲过的原话。然而，从商业战略角度来看，这种模仿虽然在初期给市场带来了噪音和混乱，但最终却成为巨大的利好： - **做大市场预算蛋糕**：以国防科技（Defense Tech）为例，若全行业只有 Palantir 一家孤军奋战，政府绝不敢把过高比例的预算押注在单一供应商身上；正是因为涌现了 50 家宣称能做类似业务的初创公司，才彻底撬动并扩大了整个国防软件市场的采购预算。 - **充当“表外销售资源”（Off-balance sheet sales resources）**：这些蜂拥而至的竞争对手无形中帮 Palantir 完成了市场教育，并在客户面前构成了天然的对照组。当客户试过一圈发现其他方案漏洞百出后，最终依然会选择真正可靠的 Palantir。 - **重塑行业标准**：模仿者的失败反而拉高了企业对严谨软件工程的认知标准。Karp 还特别提到，自己的读写障碍（dyslexia）在应对剧烈市场变化时反倒成了巨大优势，因为他从来不受现成剧本（playbook）的思维束缚，能够极其敏锐地调整公司姿态以应对变革。 ### [16:01](https://www.youtube.com/watch?v=S9V-5VQ_Epg&t=961s) 无法复制的壁垒：深度本体与作为终极竞争优势的“品味”（Taste） 在剖析为何追随者难以逾越 Palantir 的护城河时，Karp 强调，许多东西表面上可以仿造，但内核完全不同——大模型生成的伪代码表面上看起来像 Palantir 的交付代码，但根本无法实战部署；其他厂商即便理论上试图抄袭 Ontology（本体结构），但这套深植于大型组织核心业务肌理的结构至少需要耗费三年才能勉强搭建起来，而三年之后，Palantir 早就迭代到了一个全新的技术世界。 在这一切技术与商业博弈的最顶端，Karp 提到了一个最决定性的无形壁垒——“品味”（Taste）。在 AI 时代，智商、勤勉与执行力固然是标配，但真正的核心竞争力在于能否拥有敏锐的品味去辨别真伪：在两个看似都在发表奇谈怪论的商业伙伴中，有品味的人能够瞬间识别出谁吐露的是离经叛道却直击本质的真知灼见，而谁只是在东施效颦般地背诵时髦词汇。任何卓越的企业内部都必须存在一位无可替代的“品味仲裁者”（taste arbiter）。Karp 强调，在 Palantir，品味渗透到了每一个微观细节之中——从每一个软件产品的研发方向、每一套现场部署的执行架构，到每一次将关键人才安排在关键战位上的排兵布阵。 ### [17:11](https://www.youtube.com/watch?v=S9V-5VQ_Epg&t=1031s) 本体论、组织管理与“品味”的壁垒 在技术架构层面，Palantir 的本体论（Ontology）能够负责将散乱的数据进行结构化组织，但在具体运作中，如何带着“品味”（taste）来管理整个组织却是一个更深层的人性与管理难题。企业必须做出无数复杂的权衡取舍：究竟由谁来统领项目？哪些数据集允许接入？哪些内容应该推送到公共云端，哪些必须保留在本地（On-Prem）部署？抛开法律与战争伦理不谈，企业究竟想要保护什么、应该保护什么，又有哪些数据因为需要投放到外部以换取更多数据反馈而选择不予封锁？卡普强调，所有这些决定在底层全由“品味”在裁决，而且决策者必须拥有具备品味的公信力。 对于大量硅谷科技公司而言，这构成了一个极其严峻的现实障碍。许多科技领袖在同温层的朋友圈里备受欢迎，却完全意识不到自己在传统企业客户群体中有多么不受待见。卡普幽默地借自身经历打趣：这就像他以前总以为自己和哥伦比亚大学的学者们有些小分歧，直到后来才发现“事情比想象中严重得多，他们简直觉得我就是撒旦”。他曾以为大家成长于同一个知识分子社群，还能坐下来“聊聊海德格尔”，结果对方根本不想同他探讨哲学。科技圈眼下同样沉浸在自我感觉良好的幻想中，无法看清外界对他们的真实态度。 ```mermaid flowchart TD accTitle: 基于品味的企业 AI 决策模型 accDescr: 展示本体论技术底座如何与品味驱动的管理决策相结合，最终落地于企业各应用场景。 subgraph TechBase["技术基础"] A["数据孤岛与底层架构"] --> B["Palantir 本体论 (Ontology)"] end subgraph TasteDecisions["品味裁决与管理边界"] B --> C{"由品味驱动的决策"} C --> D["云端部署 vs 本地化部署 (On-prem)"] C --> E["数据资产保护 vs 开放换取新数据"] C --> F["组织权责划分与数据源准入"] end subgraph Impact["现实落地"] D & E & F --> G["具备品味公信力的现代企业"] end ``` ### [18:22](https://www.youtube.com/watch?v=S9V-5VQ_Epg&t=1102s) 警惕国家化与监管灾难：技术精英的盲目自满 卡普向所有抱着“过度松弛、高枕无忧”（chillaxing）态度而不采取行动的行业人士发出严厉警报。在过去六个月里，他每隔几天就会给科技界的一众巨头大佬通电话，反复向他们敲响警钟：“我们会被国家化的（We're going to be nationalized）。”然而，那些科技巨头往往只把他当成语出惊人的娱乐型角色，完全不以为意地反驳：“凭什么会有人想把我们国家化？美国历史上从未发生过这种事。我们这么受人喜爱（We're so likable），创造了如此巨大的价值。”卡普对此回应，自己很清楚自己不受待见，也不会去拆穿对方有多讨人嫌，但政治动量显然站在主张国家化的一方。如果科技界不整肃自身，坦诚直面 AI 带来的冲击并正视国家面临的战略对手，整个行业首先会被一帮对技术一窍不通的人用严苛法规锁死，随后便会直接步入国家化。指望私下找游说团体去华盛顿疏通根本无济于事。 卡普进一步澄清了外界对自己此前言论的误读。有人懒得深入阅读便断章取义，指责他鼓吹在全美恢复“征兵制”并给他扣上好战的帽子。他澄清自己绝不支持征兵制，他的初衷是在剧烈变革的动荡时代，美国需要某种共同的公民纽带让大家记起彼此同为美国人——哪怕只是所有人抽出一周时间共同打理公园。他直言，许多战争之所以爆发，恰恰是因为决策层中没有任何工人阶级的参与；如果能让社会各个阶层的人都参与进来，国家参与的战争反而会少得多，这本质上是坚定的反战立场。目前无论极左翼还是极右翼，充斥着完全不懂技术、满腔仇恨的声音，而中间温和理智的人群却在安于现状、梦游般地走向国家化的悬崖。 - 卡普在过去半年内频繁致电多位科技巨头，警告 AI 行业面临被监管封杀甚至收归国有的系统性风险，但对方普遍盲目自满，深信“美国不可能发生国家化”。 - 游说团体无法抵御来自缺乏技术认知的监管者的干预，科技公司必须直面自身外部性与国家地缘竞争的现实。 - 卡普主张建立类似社区服务的全民纽带，认为让劳工阶层深入参与国家决策是避免草率发动战争的最佳反战路径。 ### [21:05](https://www.youtube.com/watch?v=S9V-5VQ_Epg&t=1265s) 裁员借口、员工赋能与企业真正的未来 针对过去一年《财富》500强企业的大规模裁员潮，主持人提问企业究竟是因 AI 带来了显著效率提升才缩减人手，还是借 AI 为掩护来掩盖业务停滞或组织臃肿的真相。卡普表示，他长期与跨国巨头管理层、工会、一线士兵和消防员保持深入对话，现实规律非常明确：如果能让员工掌握新技术并完成技能升级（upskill），他们只会变得更有价值。无论是动力电池产线工人、卡车司机还是前线士兵，技术的注入都在大幅放大一线人员的生产力。 卡普警告企业高管在对外叙事上必须保持节制与诚实。如果企业一边被竞争对手打得落花流水，一边却宣称“因为部署了 AI，所以裁掉了三分之二的员工”，这种推卸责任的做法无异于直接向激进反资本主义阵营递刀子，堪称是“去给伯尼·桑德斯的宣言背书”。普通民众已经敏锐察觉到了 AI 带来的潜在破坏力，大公司如果把这种公众恐惧当成免责掩护，就是在玩火自焚。卡普以国防战场为例：技术的普及并未淘汰底层步兵，反而让只有高中职高学历的一线战士能够熟练操作尖端产品，使其作战价值呈指数级提升。未来具备竞争力的现代企业绝非通过 AI 盲目砍人，而是由高瞻远瞩的领导层挂帅，同时在组织架构的每一个层级都配备具备创造力、审美与品味的高素质人才。最后，由于下一位重磅企业嘉宾即将登场，卡普在轻松的气氛中结束了本次对话。 ## Notable quotes > [00:00](https://www.youtube.com/watch?v=S9V-5VQ_Epg&t=0s) 绝大多数人最大的误区就是每天都去硬吊，这会导致过度疲劳，而肌肉和肌腱最需要的是恢复。 — Alex Karp > [03:03](https://www.youtube.com/watch?v=S9V-5VQ_Epg&t=183s) 商业问题几乎从不可能仅仅依靠砸钱、消耗更多计算 Token 就迎刃而解；资本固然可以加速进程，但前提是必须找准正确的目标与落地路径。 — Alex Karp > [05:09](https://www.youtube.com/watch?v=S9V-5VQ_Epg&t=309s) 别急着找我们谈，你们先去跟那些明星大模型公司待上两天。要是你们运气好，等你们被折磨完了，我再考虑让你们进 Palantir 的大门。 — Alex Karp > [08:24](https://www.youtube.com/watch?v=S9V-5VQ_Epg&t=504s) 当你照镜子时发现自己刚刚印了一大笔钱，你看起来自然容光焕发。 — Alex Karp > [11:48](https://www.youtube.com/watch?v=S9V-5VQ_Epg&t=708s) 大模型的爆发直接给 Palantir 的业务注射了类固醇（putting our business on steroids）。 — Alex Karp > [11:48](https://www.youtube.com/watch?v=S9V-5VQ_Epg&t=708s) 颠倒过来就对了，这就是在追逐梦想（Living the dream）。 — Alex Karp > [13:19](https://www.youtube.com/watch?v=S9V-5VQ_Epg&t=799s) 正是因为涌现了 50 家宣称能做类似业务的初创公司，才彻底撬动并扩大了整个国防软件市场的采购预算。 — Alex Karp > [17:11](https://www.youtube.com/watch?v=S9V-5VQ_Epg&t=1031s) 事情比想象中严重得多，他们简直觉得我就是撒旦。 — Alex Karp > [18:22](https://www.youtube.com/watch?v=S9V-5VQ_Epg&t=1102s) 我们会被国家化的（We're going to be nationalized）。 — Alex Karp > [18:22](https://www.youtube.com/watch?v=S9V-5VQ_Epg&t=1102s) 凭什么会有人想把我们国家化？美国历史上从未发生过这种事。我们这么受人喜爱（We're so likable），创造了如此巨大的价值。 — 科技巨头高管 > [18:22](https://www.youtube.com/watch?v=S9V-5VQ_Epg&t=1102s) 如果科技界不整肃自身，坦诚直面 AI 带来的冲击并正视国家面临的战略对手，整个行业首先会被一帮对技术一窍不通的人用严苛法规锁死，随后便会直接步入国家化。 — Alex Karp > [21:05](https://www.youtube.com/watch?v=S9V-5VQ_Epg&t=1265s) 因为部署了 AI，所以裁掉了三分之二的员工，这种推卸责任的做法无异于直接向激进反资本主义阵营递刀子，堪称是去给伯尼·桑德斯的宣言背书。 — Alex Karp ## Mentioned Alex Karp, Palantir, TBPN, AIPCon 10, Palantir Apollo, Ontology (本体论), 乌克兰武装部队, 美国国防部, 哥伦比亚大学, 马丁·海德格尔, 伯尼·桑德斯 ## Video description This is our full conversation with Alex Karp, recorded live at AIPCon 10. We discuss why he believes taste is the most important competitive advantage in the age of AI, how enterprises are struggling to turn AI hype into real business value, why "tokenmaxxing" and AI overuse can distract companies from solving meaningful problems, and the growing risk of AI regulation and nationalization. Sign up for TBPN’s daily newsletter at TBPN.com Follow TBPN: https://TBPN.com https://x.com/tbpn https://open.spotify.com/show/2L6WMqY3GUPCGBD0dX6p00?si=674252d53acf4231 https://podcasts.apple.com/us/podcast/technology-brothers/id1772360235 https://www.youtube.com/@TBPNLive [Source](https://www.youtube.com/watch?v=S9V-5VQ_Epg)

[0:00] Alex Karp: 绝大多数人最大的误区就是每天都去硬吊，这会导致过度疲劳，而肌肉和肌腱最需要的是恢复。 [3:03] Alex Karp: 商业问题几乎从不可能仅仅依靠砸钱、消耗更多计算 Token 就迎刃而解；资本固然可以加速进程，但前提是必须找准正确的目标与落地路径。 [5:09] Alex Karp: 别急着找我们谈，你们先去跟那些明星大模型公司待上两天。要是你们运气好，等你们被折磨完了，我再考虑让你们进 Palantir 的大门。 [8:24] Alex Karp: 当你照镜子时发现自己刚刚印了一大笔钱，你看起来自然容光焕发。 [11:48] Alex Karp: 大模型的爆发直接给 Palantir 的业务注射了类固醇（putting our business on steroids）。 [11:48] Alex Karp: 颠倒过来就对了，这就是在追逐梦想（Living the dream）。 [13:19] Alex Karp: 正是因为涌现了 50 家宣称能做类似业务的初创公司，才彻底撬动并扩大了整个国防软件市场的采购预算。 [17:11] Alex Karp: 事情比想象中严重得多，他们简直觉得我就是撒旦。 [18:22] Alex Karp: 我们会被国家化的（We're going to be nationalized）。 [18:22] 科技巨头高管: 凭什么会有人想把我们国家化？美国历史上从未发生过这种事。我们这么受人喜爱（We're so likable），创造了如此巨大的价值。 [18:22] Alex Karp: 如果科技界不整肃自身，坦诚直面 AI 带来的冲击并正视国家面临的战略对手，整个行业首先会被一帮对技术一窍不通的人用严苛法规锁死，随后便会直接步入国家化。 [21:05] Alex Karp: 因为部署了 AI，所以裁掉了三分之二的员工，这种推卸责任的做法无异于直接向激进反资本主义阵营递刀子，堪称是去给伯尼·桑德斯的宣言背书。

[0:00] tendons are going to rip [0:00] Alex Karp: alien territory [0:00] Alex Karp: lung capacity of a whale [2:57] Alex Karp: printing tendies [2:57] Alex Karp: de-masturbatory tool [2:57] Alex Karp: rearranging deck chairs on their personal Titanic [2:57] Alex Karp: tool-shaped objects [5:09] Alex Karp: taste combined with money [7:27] Alex Karp: clamoring [7:27] Alex Karp: literally calling me Satan [7:27] Alex Karp: steroids [14:14] Alex Karp: parrot [14:14] Alex Karp: huge advantage [18:10] it's never happened in America [18:10] Alex Karp: chillaxing

[0:00] Yeah, yeah, get in here class. This is [0:01] good. Okay, [0:03] how is it going? How is ARP con this [0:05] time around? What's changed? [0:07] >> Um, well, we're in a phase. All each one [0:10] of these things like marks a time. First [0:12] of all, you guys are even more ball or [0:14] more successful. [0:15] >> Thank you. [0:16] >> Some tendies in your pocket. [0:17] >> I think it might have been part two you [0:19] got to say thank you. You blew us up. [0:21] You were the biggest guest. [0:22] >> You're looking bigger and stronger [0:24] somehow. [laughter] [0:25] Hey, are you more attractive in your [0:26] personal life now randomly? [0:27] >> Well, here's what we're actually focused [0:29] on dead hangs. [0:30] >> Yeah, so you came on last time you said [0:32] your dead hangs around like five. [0:33] >> Oh, no. It's it's well, it's plateaued [0:36] in the last couple months at 5:30. [0:37] >> 5:30. Okay, so we [0:40] the thing is like people are going to [0:41] hear that. They're going to think [0:42] hanging on a bar how long could it be? [0:44] You got to go and do it. The audience [0:46] has to go try to do it. We've started [0:47] doing it. We're still in the [0:49] >> Under two minutes. [0:50] >> in one between yeah, between somewhere [0:52] around a minute 30 you feel like your [0:55] tendons are going to rip. [0:56] >> 30 [0:57] 1:30 dead hang is respectable. [0:59] >> Okay. [1:00] >> Two minutes is super elite. [1:01] >> It doesn't feel respectable when you [1:03] have the five minute number you're [1:04] looking at the time. [1:05] >> Strength strength strength matters. Now, [1:08] the the the thing is I don't want to go [1:10] into rabbit hole in training. The single [1:11] biggest mistake people make is they try [1:14] to hang every day. You need recovery [1:15] second thing else. So, if you want to [1:18] mimic and get progress, you just do what [1:20] I do which is once a week you hang as [1:22] long as you can. Doesn't have to be [1:23] super macho. And then that's your day. [1:25] So, let's just say you can do 1:30. [1:27] >> sets? [1:28] >> No, no, you know, one day a week you do [1:30] your maximum. [1:30] >> Maximum. Wow. [1:31] >> So, like let's say you could do two [1:32] minutes. You try to do at least 1:30. [1:35] You fight to get to 1:30, but you don't [1:36] fight to get to two minutes. [1:37] >> Got it. [1:37] >> That's your dead hang day. And then you [1:39] can basically [ __ ] around the next day [1:40] and do whatever you want. Don't overdo [1:42] it, but you could do two times one [1:44] minute with a long break. And then can't [1:47] you just screw around and do less and [1:48] less and less. Two days before if you [1:50] two minute dead hang you do like four [1:52] times 15 seconds. The day before before [1:55] off. [1:56] And you do that. just keep doing that, [1:58] and your day What The mistake people [2:00] make is they hear my ball or time. [2:02] >> [laughter] [2:03] >> [ __ ] that guy. I'm going [2:05] >> I mean, the mistake you're making is not [2:07] doing a course. This could be a whole [2:08] new revenue [2:10] Wait, you guys I mean, you guys have a [2:12] war for you guys. You could be like, [2:13] "Hey, call in." [2:15] >> So, um yeah, I mean, the dead hang is [2:18] that it's like And also, some of it's [2:19] just genetic. Like, my other metrics are [2:22] elite, but that this is somehow alien [2:24] territory. [2:25] God-given gift. [2:26] >> What about What about breath hold under [2:28] water? [2:28] >> I don't do that, and I'm not sure I have [2:30] a Like, I grew up swimming. I I think [2:32] I'm weaker at that. Like, I bet you I [2:33] would be in your guys' [2:35] >> [laughter] [2:36] >> I I I'm a I'm a dive master. I can hold [2:38] my breath for 3 minutes. [2:39] >> I'm just say I'm just say So, I think [2:41] you'd be like [2:42] I think you'd be crushing me on that, [2:43] honestly. But, you have like the lung [2:44] capacity of a whale. [2:45] >> That's true. That's true. [2:46] >> I mean, like [2:47] >> Yeah. [2:47] >> You got like you got [2:47] >> Yeah, if I'm not moving, I'm not using [2:49] any oxygen. [2:50] >> It's like you're like you're like you're [2:51] like a like a like a whale floating out [2:54] there [laughter] under the [2:55] waiting to surface. [2:56] >> That's true. [2:57] >> So, you know, I'll say I'll tell you the [2:59] difference. Uh it is just I got they're [3:01] always minding me out there. But, like [3:03] okay, when we first met, it was like, [3:05] "AI maybe real." Okay. Then, I would say [3:09] somehow until about 2 weeks ago, there [3:11] was like a holy [ __ ] this is real, but [3:14] somehow it's not working, but we're not [3:16] allowed to say it publicly cuz we'll [3:17] look stupid. And then there's a lot of [3:19] investor hype. [3:20] >> Yep. [3:20] >> Um there still is like investors [3:22] printing tendies. So, you have the [3:24] investors on one side. I think people [3:26] realize it's real, but [3:28] you know, it's like you know, you have [3:29] the whole token maxing. [3:31] >> Yep. [3:31] >> And people are on onto that. And then [3:34] so, there's a whole value lecture there. [3:36] Um uh you have a political situation [3:40] um where, you know, people who do not [3:43] understand basic economics [3:44] are winning the political argument. So, [3:46] you can you could we could talk about [3:48] where AI [3:48] >> Yeah, there's a lot here. Let's break it [3:50] Let's break it up. Let's start with the [3:52] token maxing thing. Let's start with [3:53] what's what's real. [3:55] How are you actually thinking about [3:57] deploying AI [3:59] >> First of all, what what is your what is [4:00] what is Palantir's philosophy around [4:03] token consumption? [4:04] >> Well, like we okay, we have a product [4:06] that will allow you to be able I mean, [4:07] internally it's called [4:09] something, but externally, but really we [4:12] call it the de-masturbatory [4:14] like get off masturbation thing [4:15] internally. [4:16] >> It's like [4:17] >> People are just like print like sitting [4:19] there all day kind of like a porn [4:21] addiction, and enterprises are like, [4:23] "Okay, we knew this we believe this will [4:26] create value, but we cannot have people [4:28] just like some people [4:29] >> checking the weather with it, just like [4:31] and just re-re-rearranging deck chairs [4:33] on their personal Titanic. [4:35] >> literally like porn. [4:36] >> Like people are like full on [4:38] >> It feels it feels so [4:40] >> objects, right? [4:41] >> Tool-shaped objects you're looking at [4:43] more than you want. You hope no one [4:44] notices. You're kind of before dinner [4:46] after dinner. [4:47] >> Yeah, and every email classified with [4:49] tags. [4:50] >> It comes down to like business problems [4:52] can never be I mean, sometimes they can [4:55] be solved purely with money and just [4:57] spending more, but very often [4:58] >> Actually, I think it's the opposite. So, [5:00] just to give you a weird analogy [5:01] >> No, and I was going to say very very [5:02] often it's more the opposite where it's [5:04] about figuring out the right way to do [5:05] something, and then you can use capital [5:08] to fuel that process, but [5:09] >> Let me give you a thing that's sui [5:11] generis for you guys. Okay. It it it's [5:15] taste plus money. [5:16] >> Okay. [5:17] >> And there is no the like AI Like if you [5:20] look at like to pick any issue we want [5:21] to talk about. Token maxing is a [5:23] maxing uh what's going on with deploy [5:26] codes. Are other people going to build [5:28] ontologies? Why are why why does the our [5:32] political class not understand AI, [5:33] especially in Europe? It's like, "Yes, [5:36] because all of these things can be [5:37] scaled in a very valuable, but largely [5:40] going to commoditized way. But you can't [5:42] scale the taste of like, "What is the [5:45] business problem you want to have to [5:46] solve and need to solve? At the end of [5:48] the day, whether it's in the four [5:50] whether it's the Ukrainians fighting, [5:51] the Israelis, commercial entities, it's [5:54] there's somebody sitting there who's [5:56] like, "Okay, but this problem is [5:57] valuable. This problem isn't." And once [6:00] that value that problem is always has [6:02] that problem [6:03] alway almost always but not always has [6:05] attributes. So, there are some problems [6:06] you could solve with this like, "I want [6:08] to write a report on GDP growth in [6:10] China, right?" Okay, but if it's a [6:12] problem that requires a knowledge store, [6:14] like I want to understand the [6:15] specialized way I underwrite. We're [6:17] going to have a guest here. I want to [6:19] understand the specialized way I drill [6:21] for oil and gas that's both legal, [6:23] ethical, and reduces the cost of [6:24] production. I want to change the the the [6:28] supply chain of my industry, whether [6:29] that's military or whether that's [6:31] building boxes, or whether that's cars. [6:33] These things require actual precise, [6:36] ongoing processes. [6:38] They are enhanced by large language [6:40] models. [6:41] They are not replaced by large language [6:42] models. And then you get to security [6:44] issues like But for us like the whole [6:46] myth of things is it's just a boon [6:48] because like yeah, we can take any [6:50] model, their model, [6:52] OpenAI model, open model, we can [6:55] identify we can now identify [6:57] vulnerabilities at like 10 100x. Yeah, [7:00] but then who patches them? Yeah. How do [7:02] you patch them on prem? How do you patch [7:04] them on prem so that your specialized [7:06] knowledge stays on prem? Like if you're [7:08] any business or intel service, there's a [7:10] lot of these things are very similar. [7:11] Like you're not putting your classified [7:13] data in a public cloud. Yeah. Same thing [7:16] if you're like, you have a special way [7:18] of farming soybeans. Well, you're not [7:20] So, it's like how do you have how do you [7:22] get So, all these problems are exposed, [7:25] identified, and then [7:27] you always have a thing of where's the [7:28] charisma, which people really [7:30] underestimate. And it's it's not global. [7:32] There's no global charisma now. Like So, [7:34] right now the large language models are [7:36] very [7:37] Frontier companies are super charismatic [7:39] with investors. [7:40] >> Mhm. [7:40] >> I'll give you some news. They're super [7:43] not un-charismatic with enterprises and [7:45] the people. [7:46] >> Yeah. [7:47] >> In a way even with enterprise. [7:48] >> No, no. I mean it's [7:49] >> Cuz I understand with the people. [7:50] >> no. The enterprise people I have a [7:52] secret I have a secret like every [7:53] company has a secret way of selling. You [7:55] know what my secret way of selling is? [7:57] Don't even call it don't don't come talk [7:59] to us. There's a frontier company. Go [8:01] spend 2 days with them. And if you're [8:03] lucky after you're done, I'll let you in [8:05] my door. They're like [laughter] [8:06] clamoring. [8:07] >> Yeah. [8:07] >> They're like hey, I'll take your bad [8:09] brand. [8:10] >> Yeah. [8:10] >> We have a great brand in enterprise. But [8:11] like it it's like it's like secret [8:13] knowledge cuz the investors love this. [8:15] They're like hey, my stocks are all up [8:16] everything's up. I mean Palantir's done [8:18] very well. We're like it but it's like [8:20] and you know you guys are doing very [8:21] well I imagine, right? And it's okay. [8:23] We're we're you know we can but I'll [8:24] tell you what you go down the street you [8:26] talk to a marine, you talk to a bus [8:28] driver, you talk to the person who owns [8:30] the bus driving company. [8:32] >> Mhm. [8:33] >> They are not happy. They do not like [8:35] these people. They're tired of people [8:37] token maxing. [8:38] >> Yeah. [8:38] >> It it looks like masturbation at their [8:42] that's cost them money. They they're [8:44] like and honestly then you have [8:45] something we're not allowed to talk [8:46] about in this country likeability. [8:48] >> Mhm. [8:48] >> Like Palantir, we have I think we have [8:50] like 50 100 million global bands. We [8:52] have like 5 million people that wake up [8:54] in the morning [8:55] literally calling me Satan. [8:56] >> Mhm. [8:57] >> I didn't know I had that kind of [8:58] >> Mhm. [8:59] >> warm hand. But uh you know it's like [9:01] that's what they believe. [9:02] >> Yeah. [9:02] >> Yeah. [9:03] >> And like and they really believe it. [9:04] Okay. [9:06] What people are not allowed to really [9:07] address is like we have fans and [9:09] enemies. [9:10] >> Yeah. [9:10] >> Yeah. [9:10] >> These people are polarizing. [9:11] >> Yeah. [9:12] >> We're polarizing which means both sides. [9:13] >> Yep. [9:13] >> These people have one side. [9:14] >> Yep. [9:15] >> They're just it is so it's like you and [9:18] then it's like it's a really big [9:19] >> Social media companies too have the same [9:21] problem. [9:21] >> Yeah, no. [9:21] >> Everyone uses them but no one likes [9:23] them. [9:23] >> Yeah, but but then they also live in a [9:25] circle and that circle's printing money. [9:27] >> Yep. [9:27] >> So it's like you know when you look in [9:29] the mirror and you just printed a lot of [9:30] money [9:31] you look pretty fresh. [9:32] >> Is part of it is is part of it that that [9:35] some element of the technology, let's [9:37] just say LLM's, [9:39] is so magical that [9:42] the companies involved, that the [9:43] companies that are making and selling [9:45] frontier intelligence, can be bad at a [9:47] bunch of other things and still grow. [9:48] >> Well, no. No, no. They are magical at a [9:51] certain kind of thing, allowing you to [9:54] write, for example, code. Now, that code [9:56] doesn't can't be used as a knowledge [9:57] store. So, if you look at code in like [9:59] three different ways, like just using [10:00] pound as a model. We have code that's [10:02] basically infrastructure. So, what what [10:03] are the Ukrainians using? What is the [10:05] Department of War using? What do a lot [10:07] of our enterprises that that we call [10:08] that primitives? So, they're basically [10:09] hard-coded things that that understand [10:11] the world. Well, way do you do? It would [10:13] take millions of technical hours and an [10:16] understanding of all these enterprises [10:17] to do it. So, it's it's much more like [10:19] how do you build a steel beam? Then you [10:21] have like code that is written by FDEs. [10:25] Okay, so that's kind of managed. It's [10:28] The reason why FDEs work, the secret is [10:30] it's actually managed in something that [10:31] we as a product. So, you're writing to a [10:33] code base for managing that. We're [10:34] increasing our product. It's not just [10:36] random people writing. Then you have, [10:38] let's call it free code. That free code [10:40] is that's magical. Like you can do it [10:42] very quickly. It's almost right. It [10:44] doesn't have to be exact. [10:45] >> Dashboards. [10:46] >> Dashboards, financial stuff, [10:48] >> Little little flow stuff. [10:49] >> probabilistic stuff where you you just [10:50] have to get it [10:51] >> One-off analysis. [10:51] >> it was it magical. [10:53] >> Yep. [10:53] >> By the way, it's magical and not only [10:55] creates it and it's magical in a way I [10:57] don't I don't know people like the porn [10:59] thing, but it's also addicting. It's [11:00] like you know it's not good for you, but [11:02] you know, it may lead to damage. [11:04] >> One more time. It can't hurt that much. [11:07] I know my doctor says it I shouldn't do [11:08] it. But it's like it's like that, right? [11:11] And you just keep going and like and if [11:13] you're involved in the that thing, [11:14] you're also making money. [11:15] >> Yeah. [11:16] >> And then, last not least, in certain [11:18] circles, like if you have you want to be [11:21] a research researcher or you believe, [11:23] essentially it's a religion. So, like, [11:25] you know, and like one of the things is [11:27] very charismatic, especially people [11:28] who've never had a religion, cuz all of [11:30] a sudden that hole in your heart that [11:31] was yearning for I don't know. I would [11:33] say, you know, a a established religion, [11:36] Judaism, Christianity, Islam is like [11:39] being filled. And and all the answers [11:41] are there. But it is very very [11:44] successful at doing things that a [11:46] company has to do. [11:48] But it is not actually solving the [11:51] problem that enterprises are. It is Now, [11:53] it can solve them. That's the trick. [11:55] It's not It's not binary. It's not like [11:57] You can't say they're not valuable. [11:58] They're totally putting our business on [12:00] steroids. Like without LLMs, nobody [12:03] would be talking about our ontology, [12:06] about Apollo managing secure exploits, [12:09] about our ability to manage an [12:10] enterprise essentially, uh turning all [12:12] these companies into FDEs. These deploy [12:14] codes, we love them cuz now every [12:16] company wants to deploy code. You know [12:17] how you do that? [12:18] You re-platform on Palantir. And like [12:21] and it actually works. It's not somebody [12:22] with no taste who's never done [12:23] enterprise, who has no earthly clue how [12:25] these things work, who's done something [12:26] else and is like just imagining they [12:28] know how to do it, right? [12:29] >> Yeah, is part is part of this moment [12:32] quite entertaining for you because [12:35] you guys have been working on [12:37] understanding businesses at a deep [12:40] fundamental level, creating [12:43] You guys have effectively been doing the [12:45] work that it that [12:47] people are promising AI could do for 20 [12:49] years now, but actually doing it, [12:51] finding all the really rough [12:54] finding all the really rough edges and [12:57] uh [12:57] and not and and being at a point where [12:59] you don't have to oversell the [13:00] technology. You can sell both things. [13:02] But now there's maybe [13:05] Here we go. We got it together. Um [13:07] now there's maybe [13:09] >> Oh, it's the wrong side. That's why. [13:11] Flip it around. [13:12] >> There you go. [13:15] There you go. There you go. Living the [13:16] dream. Living the dream. [13:19] >> I I trailed off, but uh but is part of [13:22] it entertaining to you that it feels [13:24] like [13:26] you You Palantir has always been in some [13:29] ways um [13:30] not had competitors because there's [13:32] nobody with Alex Karp running a company [13:35] that is does what Palantir does besides [13:37] Palantir. But at the same time, there's [13:39] been [13:40] tens of billions of dollars deployed now [13:42] to effectively do what Palantir does, [13:45] but [13:46] just selling the intelligence part, not [13:49] selling all the underlying kind of [13:51] infrastructure that you guys [13:52] >> Well, they're doing two things. They're [13:53] selling They're trying to sell the [13:54] intelligence part and they're trying to [13:56] pretend if you just hire a bunch of [13:58] people and let them run around their [13:59] FTEs. Now, the the very cool thing is [14:03] when you've been in your [14:04] basement doing your thing and everyone [14:07] kind of use it as the freak show, it [14:10] it's really interesting and and and [14:12] great to have adoption. The pretty uh [14:14] ironic thing is half the people adopting [14:17] now don't even know they're copying. But [14:19] now, the copying thing it helps and [14:22] hurts. [14:23] Where it hurts is in the beginning it [14:25] puts clutter in the market. [14:27] >> Yeah. [14:27] >> Um and there's there's no doubt about [14:29] it. Where it helps and then we saw this [14:31] with defense tech, honestly. So like in [14:33] defense tech, we were the only people [14:35] we're the first people despite what it's [14:37] I I love these honestly other [14:40] podcasters. They're interviewing people [14:41] who are parroting things I said 20 years [14:43] ago they don't know it. And it's like, [14:44] "Oh, that's so insightful." It's like, [14:46] "Yeah, of course it's insightful. Karp [14:47] said it 25 years ago." [14:48] >> [laughter] [14:49] >> Uh and like but it's it's but so that [14:51] kind of that part is super weird. But uh [14:54] and like but um but it's but what really [14:57] happens we see is like it expands the [14:59] market. So like in defense tech, we [15:01] would not be doing this well [15:03] in just purely in government unless [15:06] there weren't 50 companies that were [15:07] doing similar things because then the [15:10] people are like, "Okay, first of all, [15:11] >> You view it as like off-balance sheet [15:13] sales resources where people are [15:15] basically doing it themselves. [15:17] >> now that that's the largely it's They do [15:19] two things. They increase the size of [15:22] the market because de facto nobody wants [15:24] to find it under in a market where [15:26] there's only one person. Sure. So like [15:28] if you're the one person, the percentage [15:30] of the defense budget you can get is [15:31] much smaller. And two, they set up a [15:34] comparator. It's like, you know, you may [15:36] not like the freak show. [15:38] Okay. If like but have you noticed the [15:40] people who are serious buy it? And then [15:42] then it it change and then three it [15:44] changes the standard. Now, what you're [15:45] seeing now is like that times a 100x. [15:49] >> Yeah. [15:49] >> And it does change like recruiting, [15:51] retention, and like how you build a [15:53] company and we're always think you have [15:55] to think about how to being dyslexic [15:56] huge advantage there cuz like you don't [15:58] have a playbook and now that we need you [15:59] need things to shift and we're doing [16:01] that. [16:01] >> Um [16:02] >> The the the central thing though that [16:05] it's just cannot be developed it even if [16:09] you understood the playbook a lot of [16:10] these things are like appear like it's [16:11] like, you know, LM code appears like [16:14] Palantir code but isn't for a deployed [16:16] thing appears like Palantir it isn't. [16:17] Ontology you could a theoretically copy [16:20] parts of it but they're essentially [16:22] structures that are built deep into [16:23] organizations that we own. And by the [16:25] way, take you 3 years. And then 3 years [16:28] we're in a completely different world. [16:29] But [16:30] there is this magical thing called [16:32] taste. Like it in the end of the day the [16:35] reason why you guys have done so well. [16:36] It's [16:37] of course there's aptitude and diligence [16:39] and showing up and all those things. [16:41] Yeah, but you have to be able to [16:42] differentiate between two people who are [16:45] in business one of whom is saying [16:46] something that sounds weird that is [16:48] insightful. One of whom is parroting [16:50] something that sounds weird and that's [16:52] all they're doing. And a lot of people [16:54] very few people can do that. And it's [16:56] the same thing like the enterprises that [16:59] succeed there is a taste arbiter. And at [17:02] Palantir we have taste we have taste in [17:05] every product, taste in every [17:07] deployment, taste in every casting. Who [17:09] puts the people there? How do you put [17:11] them there? How do you organize the [17:12] thing? Our ontology then does that [17:14] technically. How do you manage the whole [17:16] org with taste? Who should be in charge? [17:18] What data sets should come in? What What [17:20] are What are the ways in which you [17:21] protect What is What should you push [17:23] into the public crowd? What should be on [17:25] prem? What What should I mean, leaving [17:27] aside the law and like wars war ethics, [17:30] what do you want to protect? What should [17:32] you protect? What should you not protect [17:33] because quite frankly, you want that to [17:35] be out there so you can get more data. [17:37] All those things are arbitrated by taste [17:39] and then you have to have the [17:41] credibility of having taste. [17:43] That's a real problem for a lot of these [17:45] places cuz they don't have They They're [17:46] popular with their friends. They don't [17:47] They really don't understand how [17:49] unpopular they are in enterprise. Like [17:50] [laughter] they think it's like, "Oh, [17:51] yeah, it's like the way I think I have a [17:53] problem with like professors at [17:54] Columbia." It's like, "No, it's a real [17:57] problem." Like they think I'm Satan. And [18:00] you know, it's like I I I I think, you [18:02] know, we grew up in the same community. [18:03] Let's talk about Heidegger. They're like [18:04] they don't want to talk about Heidegger. [18:06] So, it's like it's like Yeah, and so [18:08] that's just a It's a weird thing. It's [18:10] going to be a super The one thing I [18:12] would say for anyone listening, [18:15] if you're listening to this and you're [18:17] chillaxing and not active I'm not saying [18:19] you have to agree with me politically or [18:21] anything. [18:22] There the like partly because of this [18:24] dynamic and it very self-inflicted [18:26] because [18:27] I I I tell you I can't name names. I [18:29] called [18:31] many of the titans of this world and and [18:33] like started this 6 months ago like [18:34] every couple days. [18:36] We're going to be nationalized. [18:37] >> them every couple days? [18:38] >> Uh like some of them I'm like, "Yeah, [18:40] we're going to be I mean, you know, it's [18:41] like honestly they're like the batch [18:42] They They They find me very [18:43] entertaining." [clears throat] Like I'm [18:44] not sure Like so they call cuz yeah, [18:46] it's like it's like Yeah, this is going [18:47] to be entertaining. [18:48] >> to pick up. [18:49] >> Yeah, so in any case, so I've been [18:51] telling them for 6 months 6 We're going [18:53] to be nationalized. [18:54] >> Yeah. Yeah, you talked about [18:55] >> We're going to be nationalized. And [18:57] they're like, "Why would anyone [18:58] nationalize? Never happened in America. [19:00] It's never Why would anyone nationalize [19:02] us? We're so likable. We're creating so [19:04] much value." Like, "Okay, I'm not going [19:06] to debate that. I know how likable I am. [19:09] I'm not going to tell you how likable [19:10] you are. But I am telling you and you [19:12] know the momentum on this is on the side [19:15] of people in national and we don't get [19:16] our act together and figure out ways we [19:18] can say, "Hey, look, there are problems [19:20] here. We're going to deal with these [19:21] things that are not going to Yes, they [19:23] are going to create opportunities. You [19:24] have to talk openly about how these [19:26] things are valuable because we have [19:27] adversaries. You can't just say these [19:29] are All that stuff." So, the primary [19:31] risk, honestly, to Palantir and to a lot [19:33] of these other countries is And then, [19:35] it's going to be nationalized before [19:37] national is going to be regulated by [19:39] people who don't understand understand [19:40] this. And now, they'll tell you in [19:41] private, "I'm working on this. I'm da da [19:43] da and this and this lobbyist." It's [19:44] like, "Not going to work." So, like, [19:46] that's something like if you're [19:47] listening to this and you're like, [19:49] "Look, you know, you don't have to agree [19:50] with me on all my proclamations. I got a [19:52] lot of By the way, there's some people [19:54] who think I'm saying we should have a [19:55] draft." Too lazy to read. I'm just [19:57] saying we should Like, in a world where [19:59] everything is changing, everything is [20:01] changing. [20:02] Don't we have to find some communal [20:04] structure to remember we're American? [20:06] You don't like my You don't like my idea [20:07] of like we all do a week in the park. [20:09] Great. Come up with some other idea. [20:12] But, we can't have no idea. You know, [20:14] and like and then they're like, "Well, [20:15] I'm saying I I do not want to draft." [20:17] Just to be explicit. They're like, "Oh, [20:18] that's pro-war." Now, honestly, you know [20:20] what most of our wars are fought because [20:21] no working-class person is making a [20:23] decision. You start making sure everyone [20:25] is involved in everything. I'll see you [20:26] how few wars we fight. It's actually the [20:28] anti-war position. But, any case, [20:30] disagree with everything. If you're We [20:33] have on the right and on the left people [20:35] people who have no earthly clue what [20:38] they're talking about, right and left. [20:39] All they're talking about is how much [20:41] they hate us. And those of us who are [20:43] sensible in the middle, [20:44] you know, too many of us are chillaxing. [20:46] Like, it Like, nationalization, it can't [20:49] happen. It's America would never do [20:50] that. [20:51] Sleepwalking into And you guys have [20:53] tendies to protect now. You [laughter] [20:55] guys should be on the front line of [20:56] this. Like, you got full Oh, I'm sorry. [20:58] I have a I have a full-on very [21:00] impressive corporate leader coming on. [21:03] >> Yes. [21:03] >> So, I got to I got to turn it down. [21:05] >> Last Last question if we have time. [21:07] How are your conversations going with [21:10] Fortune 500 CEOs around headcount [21:14] planning? There's been so many layoffs [21:15] this last year that people were saying, [21:17] "Hey, we're getting so much out of AI. [21:19] We're able to, you know, cut back here [21:21] or there." [21:22] People inside tech often know like these [21:27] maybe there's just reduction because [21:30] there needs to be a reduction. The org [21:31] got bloated. Maybe they do need to fund [21:34] some AI initiatives. [21:35] >> business model like they're getting out [21:36] competed by someone else. Yeah, yeah. [21:37] Yeah, the business just doesn't have [21:38] momentum. But how are those [21:39] conversations going? [21:40] >> What does it look like? [21:41] >> Like the like I talk By the way, I talk [21:42] to Fortune 500 companies. I talk to [21:44] unions. I talk to soldiers. I talk to [21:46] fire [21:47] If you upskill somebody, [21:50] they're more valuable. And like all [21:52] these whether it's people working on [21:54] batteries, people driving trucks, [21:56] people, corporate leaders, and again, [21:58] this is where I think we have to be very [22:00] careful to be more disciplined on the [22:01] corporate side. Like if you run around [22:03] saying AI allowed you to fire 2/3 of [22:06] your workforce and you did it because [22:07] maybe your competitor's kicking your [22:09] ass. [22:10] That could that is a really like you [22:12] might as well just go sign up for Bernie [22:14] Bernie Sanders manifest. And part of the [22:16] thing is they really believe that can't [22:18] happen. So they're free riding on the [22:20] fact that it could. Like we have And [22:22] then it just cannot work anymore. These [22:24] things are very, very explosive. The [22:25] American people sense that there is [22:28] something dangerous here. And when [22:29] people are playing with that fire, it's [22:32] like it's a they assume the fire won't [22:33] burn their hands. Well, that's not the [22:35] world we're in. That fire is going to [22:36] consume us. And what we see again, the [22:39] war-fighting example is just the most [22:41] neutral, not for everybody. But like the [22:43] soldiers at the bottom have gotten much [22:46] more valuable. Well, and and I don't [22:47] even just mean the special operators, [22:49] which obviously they're in a different [22:50] league. But like every the people doing [22:53] a lot of the operations now are doing [22:54] our product their high school [22:56] vocationally trained. You see this [22:58] everywhere. The the modern enterprise is [23:01] going to have like we have a [23:03] a true like very very very smart [23:07] person coming on and it's like you're [23:09] going to have a very smart executive. [23:10] He's much better at hiding it than I [23:12] would be if I were him, but that's you [23:14] can talk to him about that. But um [23:17] uh um and [23:19] uh and then very talented creative [23:22] people with taste all up and down the [23:24] stack. In any case, I think this is time [23:26] for me to [23:27] >> this is time. [23:27] >> We uh [23:28] >> Thank you so much. [23:29] >> Great to be here. [23:30] >> Yeah, [23:30] always fun.