US Defense Department Labels Anthropic a Supply-Chain Risk Over AI Combat Rules

美国防部因AI作战规则将Anthropic列为供应链风险 战争部副部长 Emil Michael 表示，五角大楼将 Anthropic 列为供应链风险，是因为其针对作战使用的合同限制及审计威胁在实战交涉中危及了国家安全。 在这期加急节目中，All-In 主持人就针对伊朗的“史诗之怒行动”（Operation Epic Fury）采访了战争部副部长 Emil Michael，详细探讨了本届政府的战略目标、低成本无人机蜂群，以及旨在防止陷入旷日持久国家重建冲突的交战规则修订。讨论分析了在即将举行的贸易峰会前，切断伊朗和委内瑞拉的能源流动如何转化为针对中国的战略筹码。本期节目的另一核心焦点是五角大楼因限制性服务条款及模型控制问题将 Anthropic 列为供应链风险，并深入探讨了 AI 企业风险、国防资金投入以及美国关键国防供应链回流等更广泛的议题。 美以联军对伊朗发起“史诗之怒行动”（Operation Epic Fury），旨在数周内系统性摧毁其导弹网络与无人机储备，并借此在上半年双边峰会前收紧对华石油供应通道以施加地缘杠杆。 (2:30) 美军作战策略转向压倒性武力与下放前线指挥权，并加速采购单价5万至8万美元的低成本自主无人机，通过边缘 AI 算法在强电子干扰环境下执行目标识别与协同作战。 (18:05) 受霍尔木兹海峡局势影响，国际油价升至84美元且海运险保费暴涨，特朗普政府动用美国国际开发金融公司（DFC）直接承保波斯湾商船政治风险，以规避全球能源供应链危机。 (37:24) 五角大楼将 Anthropic 列为供应链风险并终止其2亿美元合同，原因是该机构服务条款禁止军方将模型用于动能打击规划，且坚持要求在机密军事行动中保留单方审查与授权权。 (41:19) 在 xAI 与 Google 承诺全面支持合法军事用途的背景下，各家大模型性能正加速趋同，企业与国防采购正从单点依赖转向具备容错能力的多模型冗余架构。 (51:51)

Under Secretary of War Emil Michael and the All-In hosts assess the strategic objectives of Operation Epic Fury against Iran, framing the strikes as rapid disruption to build leverage against China while discussing the Pentagon's decision to blackball AI vendor Anthropic. Strategic Execution and Geopolitical Aims in Iran (2:30) Operation Epic Fury opened with high-impact strikes that eliminated Supreme Leader Ali Khamenei, 40 senior officials, and roughly 1,000 people, alongside a historic U.S. submarine torpedo sinking off Sri Lanka and the loss of six U.S. Reservists in Kuwait. Co-hosts David Friedberg and Chamath Palihapitiya argued that neutralizing Iran and Venezuela directly constrains China by threatening the roughly 20% of crude oil imports Beijing draws from both nations ahead of scheduled bilateral summits. Under Secretary Emil Michael clarified that while Chinese energy pressure is a secondary benefit, the operation is designed as a focused, 'weeks, not months' campaign to dismantle missile depots and drone networks, explicitly avoiding prolonged ground occupations. Doctrinal Shifts: Decentralization and Drone Warfare (17:12) Michael outlined an operational doctrine that moves away from restrictive rules of engagement toward the Powell Doctrine, granting tactical flexibility back to field commanders. Modern combat increasingly relies on low-cost attritable systems rather than legacy platforms, drawing lessons from Ukraine where drones inflict up to 70% of casualties. The U.S. military is deploying platforms like the LUCAS drone program ($50,000 to $80,000 per unit) alongside autonomous swarm capabilities and edge AI to conduct automatic target recognition within electronically jammed environments. Economic Fallout and Maritime Protection (37:24) Escalation in the Persian Gulf triggered economic volatility, lifting crude prices to $84 per barrel after the IRGC threatened traffic through the Strait of Hormuz, risking a 3.3 million barrel daily shortfall. War risk insurance rates jumped from 0.25% to 1.25% before underwriters cancelled coverage, driving supertanker traffic down 94% within 48 hours. To unfreeze the corridor, the Trump administration directed the U.S. International Development Finance Corporation (DFC) to extend political risk insurance to commercial shipping, protecting indebted maritime operators from insolvency. The Anthropic Dispute and Supply-Chain Designation (41:19) The Department of War canceled a $200 million contract with Anthropic and designated the developer as a supply-chain risk, barring its software from defense contractors like Lockheed Martin. Michael reported that Anthropic demanded strict contractual limits, refusing kinetic strike targeting, satellite operations, or bulk data analysis, and even questioned Palantir about model use following a raid in Venezuela. While Anthropic CEO Dario Amodei alleged the move was political retaliation for withholding donations, Michael dismissed the claim, stating the military cannot accept civilian executives conditioning mission-critical software on manual real-time policy exceptions. Procurement Reform and Industrial Reshoring (71:14) To counter China's 15-year naval and industrial expansion, Michael highlighted initiatives through the Office of Strategic Capital, which oversees $200 billion in low-cost lending authority to onshore 20 vital defense inputs, including critical minerals, batteries, and solid rocket motors. The Pentagon is overhauling acquisition away from legacy 'cost-plus' contracts to fixed-price models tied to operational milestones, enabling venture-backed defense startups to build scalable, mass-attritable systems rather than costly legacy platforms. The discussion underscores a structural pivot toward mass-attritable autonomous hardware and sovereign AI alignment, though questions remain on whether commercial AI labs will accept unconstrained military use cases and how quickly the U.S. defense industrial base can eliminate Chinese component dependencies.

美国战争部副部长Emil Michael与All-In节目主持人群重点探讨了美以对伊朗发起的“史诗之怒行动”、霍尔木兹海峡能源危机，以及五角大楼将AI企业Anthropic列入“供应链风险”名单的风波。对话表明，美国军方正转向依赖压倒性武力与低成本自主武器，并在防务采办与前沿技术上强化不容妥协的国家安全主导权。 史诗之怒行动与压倒性武力战术回归 (2:30) Jason Calacanis指出，美以联军发起“史诗之怒行动”（Operation Epic Fury）六天内，初期空袭导致伊朗最高领袖哈梅内伊及约40名高级官员身亡，伊朗境内据报约1000人丧生；美军有6名陆军预备役士兵在科威特遭无人机袭击阵亡，一艘美国潜艇在斯里兰卡海域用鱼雷击沉伊朗舰艇，创下二战以来美军潜艇首次鱼雷战果。预测市场Polymarket数据显示，3月底前美军地面部队进入伊朗的概率为40%，年底前伊朗政权垮台的概率为51%。副部长Emil Michael强调，此次行动预期为“以周计而非以月计”的战役，重点是系统性摧毁弹道导弹库存和单向攻击无人机网络，并非重演伊拉克或阿富汗式的旷日持久地面占领。 Michael表示，国防领导层彻底摒弃了以往过度法律化的严苛交战规则（ROE），避免前线士兵在生死瞬间因繁琐程序犹豫不决。现行条令回归以清晰目标和压倒性武力为核心的鲍威尔主义（Powell Doctrine），将作战裁量权下放至战场指挥官。行动的高效推进同时依托于在反恐战争中积累丰富实战经验的高级指挥层，以及整合太空、网络、海陆空各领域的系统技术优势。 能源杠杆、对华博弈与海峡航运保险干预 (6:03) David Friedberg与Chamath Palihapitiya分析认为，打击伊朗及此前针对委内瑞拉马杜罗政权的军事行动，战略核心是在4月美中高层会晤前积累对华谈判筹码。中国约90%的伊朗原油出口及大量委内瑞拉原油输入已被打乱，这两处来源约占中国石油总进口量的20%。在中国设定30年来最低的4.5%至5.0%国内生产总值（GDP）增长目标且青年失业压力沉重的背景下，切断其外部原油供给严重压缩了北京的战略回旋余地，有助于遏制其因国内经济停滞而对外挑起台海冲突的风险。 中东战事直接引发能源外运通道危机，国际原油价格升至每桶84美元，霍尔木兹海峡面临每日330万桶原油供应损失的风险。战险保费从船舶价值的0.25%飙升5倍至1.25%，以劳合社（Lloyd's of London）为代表的主流承保财团随后全面撤销保险，导致超级油轮通行量在48小时内暴跌94%。为防止全球恶性通胀与航运停摆，特朗普政府动用拥有信贷职能的美国国际开发金融公司（DFC），直接为波斯湾所有商业海运提供政治风险担保，以国家信力托底航运债务要求，迅速恢复了商船通航。 五角大楼与Anthropic冲突与AI供应链风波 (41:19) 五角大楼正式将AI开发商Anthropic认定为“供应链风险”并取消其2亿美元合同，这是该防务限制首次适用于美国本土公司。Emil Michael透露，争议爆发于Anthropic合同中长达20页的禁止性服务条款（TOS），该条款严禁军方将其模型用于动能打击规划、卫星轨道机动或涉及潜在人员伤亡的战争推演。在美军针对委内瑞拉马杜罗的突袭行动后，Anthropic高管向主承包商Palantir质询其模型是否被用于该机密任务，使五角大楼确信该公司随时可能在战时擅自断供，给官兵带来不可承受的作战风险。Michael还透露，Anthropic首席执行官Dario Amodei曾提议军方在突发危机时致电其本人申请特例，军方认为这一设想极不合理且十分危险。 针对Amodei指责五角大楼因其拒绝政治献金而施加报复的说法，Michael驳斥此为无稽之谈，强调军方曾善意谈判近三个月，唯一红线是确保技术可靠并适用于所有合法作战场景。相比之下，xAI（Grok）与Google（Gemini）均同意支持军方合法用例，OpenAI首席执行官Sam Altman则在协助紧急备用部署的同时呼吁行业团结。Chamath指出，Anthropic试图推行自身道德规范而漠视法律，给依赖单一模型的机构敲响了“超级平台封禁”的警钟，各界正被迫加速转向多模型冗余架构。 无人机蜂群革命、采办改革与防务供应链 (18:05) 现代战争形态正从高昂庞大的旧式武器转向高密度、低成本的自主系统。Michael提到，俄乌冲突中高达70%的伤亡由无人机造成，美军正加速量产成本仅为5万至8万美元的LUCAS等低成本攻击型无人机，并结合本地边缘AI模型在强电磁干扰环境下自主识别目标并规避诱饵。在防空方面，以铁束（Iron Beam）为代表的地面定向能激光拦截与高功率微波反无人机技术正步入部署，旨在以极低边际成本构筑多层次反导防御系统。 在装备制造层面，Michael兼管的战略资本办公室（OSC）正运用2000亿美元借贷授权，以国债利率加100个基点的低成本贷款支持关键矿产、特种电池及固体火箭发动机等20项关键防务部件的本土化替代。Michael指出，五角大楼正全力破除传统军工巨头的垄断，坚决终结容易催生寻租与超支的“成本加成”合同及脱离实际的严苛技术指标，全面转向设定明确作战效果、带有交付奖惩机制的固定价格商业合同，以扶持Anduril、Palantir等新兴防务科技公司快速形成规模效应。 本期讨论展现了美国军方在战术执行、能源施压及前沿采办上推行的实用主义重构。然而，以压倒性武力开展精准打击能否完全避免中东地面战泥潭、由政府兜底的战险与产业贷款能否在预算约束下长期维系，以及军方与前沿AI企业在自主武器伦理上的根本分歧，仍是该战略面临的未决挑战。

## Background 本期节目由知名商业科技播客《All-In Podcast》于 2026 年 3 月 6 日在 YouTube 平台发布。节目简介显示，四位主持人紧急连线了新晋美国国防部负责研究与工程的副部长 Emil Michael，围绕美军与以色列对伊朗发动的“史诗之怒行动”、无人机与人工智能驱动的新型作战形态，以及五角大楼将 AI 独角兽企业 Anthropic 认定为“供应链风险”等重大热点事件展开深度对话。 ## TL;DR 节目首先从战术与宏观战略双重视角拆解了美军对伊朗展开的快节奏军事打击，指出这一行动背后包含截断中国石油供应链、为中美高层峰会累积地缘筹码的底层大战略。随后，副部长 Emil Michael 重点阐述了现代战争向低成本单向自杀式无人机（如 LUCAS）、边缘 AI 与蜂群自主协同演进的趋势，并说明了国防部放宽交战规则与构建多层反导体系的逻辑。最后，对话深入揭露了五角大楼与 Anthropic 决裂的谈判内幕，剖析了军方因自主武器与大规模监控条款争议将其定性为“供应链风险”的技术与制度原因，并探讨了美国国防制造业重塑、去中国化供应链及传统军工采购体制改革等前沿议题。 ## Atmosphere & style 整期节目兼具闭门军事吹风会的严肃压迫感与老友聚会的插科打诨，节奏紧凑且信息密度极高。由于嘉宾 Emil Michael 曾是硅谷核心操盘手，主持人们在追问地缘大战略与五角大楼封杀 Anthropic 的机密内幕时毫无保留、针锋相对，抛出了“政治献金报复”和“神王独裁”等尖锐质问；而副部长则在严谨捍卫军方底线的同时，频频以幽默自嘲化解现场张力。整场对谈既充斥着对现代无人化战争残酷性的现实审视，又洋溢着硅谷风险投资圈拆解宏观商业利益与技术格局时的敏锐洞察，让听众如同身处国家安全与前沿科技博弈的最前沿。 ## Bulletin - [02:30](https://www.youtube.com/watch?v=gzwRflcLPAA&t=150s) 美军与以色列联手对伊朗展开“史诗之怒行动”，哈梅内伊及40余名高官在数小时内被击毙；国防部明确行动以周为单位且绝不派地面部队长期占领，核心目标为彻底解除其核导与无人机武装能力。 - [41:19](https://www.youtube.com/watch?v=gzwRflcLPAA&t=2479s) 美国防部正式将 Anthropic 列为“供应链风险”并终止 2 亿美元合同，系历史上首次将本土顶尖科技独角兽列入该名单，核心分歧在于军方坚持“所有合法用途”与厂商严苛的自主武器限制及模型控制权问题。 - [39:10](https://www.youtube.com/watch?v=gzwRflcLPAA&t=2350s) 霍尔木兹海峡受伊朗封锁导致全球油轮战争险取消、航运量暴跌94%，特朗普政府动用国际开发金融公司（DFC）信用额度直接向船东提供航线担保，顺势推动海上保险经纪业务回流美国本土。 - [18:05](https://www.youtube.com/watch?v=gzwRflcLPAA&t=1085s) 现代空战全面转向低成本无人机对抗，美军加速推进 LUCAS 自杀式无人机部署（航程 500-700 英里，造价仅 5-8 万美元），同时推动国防采购由传统“成本加成”向基于明确问题的固定价格商业竞标转型。 ## Key takeaways - [05:40](https://www.youtube.com/watch?v=gzwRflcLPAA&t=340s) 美军对委内瑞拉与伊朗的连续精准军事打击不仅是战术行动，更切断了中国近两成（若算入俄罗斯则达近四成）的石油进口来源，成为4月中美高层峰会前锁定的关键地缘经济筹码。 - [16:06](https://www.youtube.com/watch?v=gzwRflcLPAA&t=966s) 五角大楼彻底打破以往过度微观管理且严苛的交战规则（ROE），重拾鲍威尔主义，将战术裁量权全面下放给现场指挥官，追求压倒性武力速战速决。 - [20:28](https://www.youtube.com/watch?v=gzwRflcLPAA&t=1228s) 未来无人机战场的胜负取决于机载边缘 AI 实现的自动目标识别（ATR）与跨机型自组网的“异构自主”协同，中国凭借对大疆等供应链的掌控在该领域具有显著先发优势。 - [22:21](https://www.youtube.com/watch?v=gzwRflcLPAA&t=1341s) 军方在军事 AI 部署上推行“风险梯度”机制，在天基激光反导或近程微波防空等低误伤场景下加速放权，但在人口密集区的地面打击上保持极度保守，当前距离“天网”式的全自主致命决策相去甚远。 - [29:59](https://www.youtube.com/watch?v=gzwRflcLPAA&t=1799s) 以色列研发的“铁束”（Iron Beam）激光防空与高功率微波系统为美以联合攻坚成果，旨在与天基防御及“铁穹”动能拦截形成多层防空网络，极大摊薄应对无人机蜂群的拦截成本。 - [46:23](https://www.youtube.com/watch?v=gzwRflcLPAA&t=2783s) Anthropic 质问美军在突袭马杜罗行动中是否调用其模型成为决裂导火索，五角大楼认定其保留模型控制平面与权重权限可能导致战时拒绝服务或内部投毒，存在不可承受的战备风险。 - [49:07](https://www.youtube.com/watch?v=gzwRflcLPAA&t=2947s) 企业与政府机构必须转向多模型（Multi-Model）架构，避免单一模型供应商将其未经选举的内部意识形态或道德偏见凌驾于法定合约之上，重蹈千倍级“去平台化”的商业瘫痪危机。 - [51:51](https://www.youtube.com/watch?v=gzwRflcLPAA&t=3111s) 在国防部 AI 供应商格局中，马斯克的 xAI（Grok）全面开放支持所有合法涉密与非涉密场景；Google 正推进涉密基建；OpenAI 则在担任替补的同时私下为 Anthropic 斡旋以避免行业人才动荡。 - [54:12](https://www.youtube.com/watch?v=gzwRflcLPAA&t=3252s) 各大前沿大模型的性能输出正在呈现渐近线收敛趋势，开发者盲目调用智能体造成的“Ralph Wiggum 循环”导致企业算力账单数倍飙升却难获产出，倒逼多模型替代与精细化成本控制。 - [01:07:35](https://www.youtube.com/watch?v=gzwRflcLPAA&t=4055s) 军方驳斥了关于 Anthropic 因未向特朗普阵营捐款而遭报复的指控，强调定性供应链风险是出于防止国防装备研发被底层逻辑投毒，民用商业设计领域则完全不受限制。 - [01:11:14](https://www.youtube.com/watch?v=gzwRflcLPAA&t=4274s) 美国国防硬件供应链正处于重塑起步阶段，战略资本办公室动用 2000 亿美元低息借贷推动电池、固体火箭发动机等核心元器件本土化，力求摆脱对中国制造业供应链的深度依赖。 - [01:17:45](https://www.youtube.com/watch?v=gzwRflcLPAA&t=4665s) DARPA 正在全力推进以合成生物学直接富集与提取战略关键矿产，并布局针对 AI 自主代理（AI Agents）的高烈度网络攻防，以建立应对大国竞争的底层非对称技术优势。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=gzwRflcLPAA&t=0s) 紧急播客：国防部副部长 Emil Michael 做客与好友互损 全入播客（All-In Podcast）紧急录制了第 263 期节目。主持人 Jason Calacanis 宣布，由于中东突发战事，节目组周四晚间紧急上线，邀请到了新晋担任美国国防部（在节目中戏称为“战争部”）主管研究与工程的副部长（Under Secretary of Defense for Research and Engineering）Emil Michael。Emil 直接向国防部长 Pete Hegseth 汇报。Jason 热情回顾了两人的私交：当年他们曾是 Uber 战队的元老，Emil 是前首席执行官 Travis Kalanick 的副手与核心操盘手，甚至被外界称为“危机处理人”（Fixer）。Jason 打趣说他是帮公司融资 200 亿美元的狠角色，什么烫手山芋最难啃，Travis 就交给他搞定；Emil 则笑着反驳，希望自己在大家眼里不只是个“清道夫”，并欣然接受了“操盘重斧”（operational axe）和“吸金能手”（rainmaker）的称号。 随后话题转到老友之间的插科打诨。Jason 点名 Chamath Palihapitiya，调侃他这一周在社交平台 X 上的“推特挑衅”（tweet mogging）和神级回复。几人狂笑不止，提到有人在推特上发帖冷嘲热讽：“大家干嘛对 Chamath 这么大敌意？他不就是搞单页 SPAC 让散户亏了数十亿美元嘛，又没让他们去享受资本亏损……”结果 Chamath 竟然极其光棍、毫无保留地直接在下面回复了一句：“是的，我就是这么干的（Yes, I did）。”一旁的“科学苏丹”David Friedberg 刚结束出差返回家中；而另一位常驻主持 David Sachs 则因在国会山处理公务而缺席。 --- ### [02:30](https://www.youtube.com/watch?v=gzwRflcLPAA&t=150s) “史诗之怒”行动爆发：伊朗战事进程与目标界定 节目迅速切入当周最震动世界的头条事件。美军与以色列军队在周六针对伊朗本土发起了联合军事打击，代号为“史诗之怒行动”（Operation Epic Fury），播客录制时战事已进入第六天。行动打响仅数小时，伊朗最高领袖哈梅内伊（Ali Khamenei）即被炸身亡，一同被消灭的还有 40 余名高级军政官员，据媒体报道目前死亡总人数已达约 1000 人。战火也造成了美方人员伤亡，6 名美国陆军预备役士兵在科威特基地遭遇无人机袭击不幸殉职；此外，一艘美军核潜艇在斯里兰卡海岸附近用鱼雷击沉了一艘伊朗军舰，这也是自第二次世界大战以来美军潜艇首次取得鱼雷实战战果。 关于美国为何在此刻全面开战，官方解释在早期有些摇摆。国务卿 Marco Rubio 最初声称是以色列执意发动袭击，美国别无选择只能参与，但随后收回了这一说法；唐纳德·特朗普总统则明确表态，此次行动绝非为了发动一场旷日持久的“政权更迭”战争，而是为了制止恐怖主义蔓延，摧毁伊朗发展洲际弹道导弹（ICBM）和核武器的能力，同时他也喊话伊朗人民抓住历史机遇“夺回自己的国家”。国防部长 Pete Hegseth 更是留下了一句名言：“这或许不是一场所谓的政权更迭战争，但伊朗政权确实被更迭了，而且世界因此变得更美好。” 在预测市场 Polymarket 上，针对事态走向的押注情绪高涨。Jason 抛出两组核心数据并向 Emil Michael 发问：这场战争到底要打多久？美军最终会不会派地面部队进驻（boots on the ground）？Emil Michael 斩钉截铁地明确了行动框架，表示特朗普总统已定调这是一场“以周计算，而非以月计算”的快节奏精准打击行动。其核心使命是彻底解除该政权的武装能力，使其再也无法为黎巴嫩真主党、哈马斯、穆斯林兄弟会等代理人极端组织输送资金和军火，更无法染指核武器。针对各界对重蹈覆辙的担忧，Emil 强调，绝不存在任何类似阿富汗战争或第二次伊拉克战争那样陷入泥潭的长年地面占领方案。 - **Polymarket 预测数据盘点：** - 美军地面部队在 3 月底前进入伊朗本土的概率：40% - 美军地面部队在 2026 年底前进入伊朗本土的概率：59% - 伊朗现政权在 6 月 30 日前垮台的概率：39% - 伊朗现政权在 2026 年底前垮台的概率：51% - **美军核心打击目标：** 军火库、浓缩铀及核设施、具有数百英里航程的单向自杀式攻击无人机（OWAs）制造厂，以及威胁中东全境的弹道导弹阵地。 --- ### [05:40](https://www.youtube.com/watch?v=gzwRflcLPAA&t=340s) 地缘大交易：对伊动武背后的中美博弈与杠杆逻辑 随后，David Friedberg 提出了一个截然不同的大战略视角。他指出，许多选民当初投票给特朗普，看重的是其第一任期内展现的“和平总统”红利，而如今美国已经对八个国家动过武，扳倒了两位外国领导人（委内瑞拉的马杜罗与伊朗的哈梅内伊）。但深入探究其底层逻辑，特朗普政府极有可能是在为 4 月份即将到来的最高级别中美首脑峰会积攒筹码。在 Friedberg 看来，特朗普的核心诉求从来不是被中东的政权更迭烂摊子拖垮，而是希望与中国达成一项具有历史意义的“宏大交易”（Grand Bargain）。 ```mermaid flowchart TD accTitle: US Military Pressure Flowchart accDescr: Diagram showing how US actions in Venezuela and Iran create strategic leverage for upcoming trade negotiations with China. subgraph DirectActions["美军精准地缘打击"] A["推翻委内瑞拉马杜罗政权"] B["摧毁伊朗军事与能源节点<br/>击毙最高领导层"] end subgraph SupplyDisruption["打击中国关键供应链"] C["切断南美反美轴心供油通道"] D["封锁伊朗原油出口<br/>(占伊朗原油出口90%去向)"] end subgraph MacroNegotiation["4月中美高层峰会"] E["美国手握极大战略筹码 (Maximal Leverage)"] F["达成经贸大交易：<br/>1. 市场对等准入与监管对齐<br/>2. 台湾问题与关键半导体安全保障<br/>3. 为中期选举铺平政绩之路"] end A --> C B --> D C --> E D --> E E --> F ``` Friedberg 强调，伊朗 90% 的原油都运往中国，而委内瑞拉与中国之间也有着长期且深厚的经济与能源绑定。通过打击乃至控制这两大关键节点的供应链，美国在面对中国时手握了极大的地缘经济杠杆。这不仅能够迫使对方在贸易壁垒、美资企业进入中国市场等长期悬而未决的问题上作出让步，甚至还能在台湾局势以及半导体等尖端科技供应链的安全问题上锁定对美国有利的平衡框架，进而成为本届政府在中期选举前最具分量的执政成果。 Chamath 顺着这一逻辑进一步延伸。他认为必须将视野拉高到全球宏观政治的高度来看待这连环三步棋。就在前一天晚上，中国官方发布了最新的年度 GDP 增长目标，罕见地下调至 4.5% 至 5% 的区间，这是自中国加入世贸组织（WTO）以来的约三十年里从未有过的历史低点。Chamath 分析道，当一个国家的经济体量和人口规模早已翻倍，增长率却从过去的 8%、9%、10% 骤降至一半，同时伴随着极高的国内失业率尤其是青年失业率时，社会内部的动荡风险必然加剧。按照历史规律，当内部增长承压、矛盾激化时，统治阶层最本能的对外宣泄途径往往就是采取激进的地缘军事冒险，而台海局势便成为了那个最危险的潜在宣泄口（sinkhole）。 ### [10:25](https://www.youtube.com/watch?v=gzwRflcLPAA&t=625s) 能源封锁与大妥协：切断非正常石油背后的对华战略 Chamath Palihapitiya 进一步引申指出，一旦陷入全面对抗，整个国家的民众、资源都会被战争机器占据，国内军工生产随之运转，这也是历次冲突中战争机器对经济的典型驱动模式。他认为特朗普政府所采取的行动可以用一张图表来概括：如果核心目标是避免与中国爆发毁灭性的大国冲突（这种冲突可能升级为核战争且具灾难性），切断其能源供应正是关键路径之一。中国经济对进口石油极度依赖，虽然能源在名义经济总量中占比约 20%，但它构成了物流、交通运输、民航以及化工原材料等核心 GDP 支柱的 100%。 在这部分进口中，约有五分之一此前完全来自伊朗和委内瑞拉，而当前这两处来源基本被全部切断。如果再结合史蒂夫·威特科夫（Steve Witkoff）、贾里德·库什纳（Jared Kushner）以及乔什·格林鲍姆（Josh Greenbaum）等人正在推进的涉俄斡旋，把俄罗斯算进去，这三国几乎占到了中国石油进口的近 40%。这套打法不仅能强化美元回流结算体系、截断流向全球制造动荡的黑产石油资金，更能将中国牢牢限制在谈判桌前，为 3 月底至 4 月初可能到来的“大妥协”（Grand Bargain）奠定筹码，从而稳固全球安全格局。 ```mermaid flowchart TD accTitle: China Energy Choke Point Strategy accDescr: Diagram showing how shutting down oil exports from Iran, Venezuela, and negotiating Russia curbs China's economy. subgraph SanctionedSources ["受限/切断的外部石油来源"] V["委内瑞拉\n(行动摧毁/政权更迭)"] I["伊朗\n(军事打击封锁)"] R["俄罗斯\n(外交施压与谈判斡旋)"] end subgraph Impact ["对中国经济的核心命脉约束"] Oil["约占中国石油进口近 40%"] V --> Oil I --> Oil R --> Oil Choke["100% 支撑关键实体部门:\n物流、航空、交通运力与化工原料"] Oil --> Choke end subgraph StrategicOutcome ["最终地缘战略目标"] Choke --> Outcome["迫使进入大妥协 (Grand Bargain)\n遏制全面大国对抗并维护美元体系"] end ``` - 中国对进口石油有极端依赖，支撑了其境内全部的物流、航空与基础工业供应链。 - 委内瑞拉与伊朗合计占据了中国石油进口来源的约 20%；若将俄罗斯计算在内，这一比例升至近 40%。 - 通过截断这类非法与非正式的石油流动，既阻止了资金被用于全球各地的混乱活动，又为即将到来的大国外交谈判提供了战略压制点。 --- ### [12:20](https://www.youtube.com/watch?v=gzwRflcLPAA&t=740s) 战争形态的演变与“零伤亡”委内瑞拉斩首行动 对于将军事行动与中国直接挂钩的论点，副国防部长埃米尔·迈克尔（Emil Michael）明确澄清，这只属于行动的二阶甚至连带收益，而非全部初衷。他驳斥了外界所谓“美国同时卷入八场冲突”的夸大说法：加沙危机和俄乌冲突是上届政府遗留的烂摊子；胡塞武装纯粹是因为拜登政府长期对其放任，任由其攻击国际航道，美方出手仅仅是恢复公海航行自由；委内瑞拉的行动则只是一场几小时内干净利落结束的突袭行动，并顺带打击了贩毒运毒船只。至于真正意义上的重大军事对抗，除委内瑞拉外仅有伊朗一处。 主持人提问，为何现在的作战能做到“不派地面部队”且执行得极其顺畅。埃米尔笑谈，除了网上热传的所谓神秘武器“定向致乱器”（Discombobulator）之外，真正的核心资产在于美军拥有极高训练素养的军官团。过去的反恐战争固然在诸多层面充满惨痛代价，但当年在前线实战摸爬滚打的基层军官如今已成长为将领，积累了极其丰富的实战直觉；反观中国军队多年未打仗，缺乏现代战争的直接经验。此外，美军在太空、空中、陆地、海洋以及网络电磁空间拥有全维度的技术整合能力，才能派出约 100 人的特种部队，直接渗透进委内瑞拉防御最严密的总统府邸，安全控制并带走其国家领导人及夫人，且整个行动美方无人伤亡（Zero Casualties）。 - 所谓庞杂的多线作战实则是舆论误读，委内瑞拉行动仅仅持续数小时即告完成。 - 伊朗行动（代号“午夜铁锤”，Midnight Hammer）等作战推演方案均在国防部的架子上预研了数年，并根据最新的战术、技术和规程（TTPs）持续迭代更新。 - 面对主持人关于“致乱器到底存不存在、能不能买一台放家里对付吵闹孩子”的调侃，埃米尔幽默回应“那只能留给赫尔穆特来的时候用”，并表示此类敏感装备无可奉告。 --- ### [16:06](https://www.youtube.com/watch?v=gzwRflcLPAA&t=966s) 释放前线手脚：重写交战规则与回归鲍威尔主义 在被问及为何五至十年前美军难以复制这种外科手术式的成功时，埃米尔指出技术的成熟只是一个方面，更根本的变革在于彻底废除了束手束脚的旧“交战规则”（Rules of Engagement, ROE）。在阿富汗等历史战场上，美军曾受到极其荒谬的政策掣肘——例如规定如果敌对分子使用轻武器射击，美方也不得采用重火力压制，强行追求所谓的“武力对等”。埃米尔直言：“如果站在我们的立场，你希望这场战斗是对等的还是不对等的？你当然希望这是一场极度不对等的碾压战。” 埃米尔提到国防部长赫格塞斯（Pete Hegseth）在其著作中写到的亲身经历：旧的交战规则对前线作战人员极为严酷，军人在瞬息万变的战场上每分每秒甚至需要先搞懂繁复的法律条文，导致官兵生命被置于严重危险之中。现任国防部领导层打破了这种教条，重新回归以科林·鲍威尔（Colin Powell）为代表的军事原则：不再做无休止的微观管理，而是设立明确红线，把决策权和裁量权直接下放给现场指挥官，要求“带着明确的目标倾巢出动，凭借压倒性武力取得胜利，然后干脆利落地撤出”。 - 过去的交战规则要求前线士兵近乎像律师一样在火力接触中反复核查交战细则，严重制约战力。 - 现行作战原则赋予现场指挥官自主判断权，坚持使用压倒性优势力量（Overwhelming Force）迅速歼敌。 - 这种哲学转变促使美军彻底摆脱了过去四年里拖泥带水、瞻前顾后的作战姿态。 --- ### [18:05](https://www.youtube.com/watch?v=gzwRflcLPAA&t=1085s) 无人机主导现代战争：LUCAS 自杀式无人机与无人武器库 埃米尔指出，现代战争形态已经彻底走向无人机对抗与机器人对抗。从 10 至 15 年前昂贵的“捕食者”（Predator）无人机执行针对基地组织头目的定点清除，到特朗普第一任期内使用无人机击杀苏莱曼尼（Qasem Soleimani），再到俄乌战场上 70% 的人员伤亡由无人机造成，低成本无人化系统已成为未来战场的决定性力量，这也是 Anduril 等防务科技公司快速崛起的逻辑所在。美军在过去一周部署的无人机规模，已经刷新了历史纪录。 针对五角大楼采办周期过于漫长、无法跟上初创公司节奏的弊端，国防部启动了类似昔日“DARPA 自动驾驶挑战赛”的全新竞速采购机制，加速推进“无人机优势计划”（Drone Dominance Program），目标是打造庞大的低成本无人机国家武器库。其中最引人瞩目的是名为 LUCAS（低成本无人作战攻击系统，Low-Cost Unmanned Combat Attack Systems）的单向自杀式攻击无人机，其航程可达 500 至 700 英里，具备飞机量级的飞行速度并能搭载大当量弹头，造价却能控制在 5 万至 8 万美元之间，远低于传统巡航导弹。这些廉价无人机系统不仅在伊朗方向的战略压制中发挥主力作用，还承担着边境反毒品无人机拦截，以及为 2026 年美国建国 250 周年庆典、美加墨世界杯和 2028 年洛杉矶奥运会等本土敏感目标提供防空安保的重要职能。 - 俄乌战场验证了无人机的主导地位，约七成伤亡源自无人机直接打击。 - LUCAS 单向自杀式攻击无人机：飞行距离达 500–700 英里，单价仅约 50,000 至 80,000 美元，兼具高航速与强毁伤能力。 - 国防部将无人机部署规划不仅限定于海外重大战区，亦覆盖边境防御与国内重大国家级集会的安保拦截需求。 ### [20:28](https://www.youtube.com/watch?v=gzwRflcLPAA&t=1228s) 蜂群无人机、异构自主与边缘 AI 的战场运用 无人机在现代冲突中的应用早已超越了单一的人工操纵与单纯侦察。当被问及未来无人机战争中人机协同与人工智能决策的比例时，埃米尔·迈克尔（Emil Michael）指出，未来高烈度的无人机对抗必然由某种程度的 AI 控制的无人机蜂群主导。战场的复杂性在于敌方会释放诱饵、干扰弹或炫目装置，单靠人类肉眼和远程操纵根本无法在毫秒级的时间窗口内识别真伪，而经过深度学习训练的 AI 模型可以通过自动目标识别（Automatic Target Recognition）迅速辨别真假目标。此外，在边缘侧运行的 AI 还能精确锁定高价值目标个人，避免殃及无辜，使打击更加精准安全。 这种技术已经在实战中展现出残酷的雏形。俄乌冲突中双方已经演化出所谓的“杀伤区”（kill box）战术：当无人机的远程通讯和 GPS 受到强电子压制而断连时，无人机并不会坠毁，而是自主在预定网格区域内巡弋，利用机载视觉算法搜寻并猎杀预设目标。迈克尔特别警示，中国在此类技术上已经拥有成倍的领先优势。依靠强大的产业掌控力，中国能够强制大疆（DJI）及其他无人机制造商实现底层协议互联互通，这种技术被称为“异构自主”（heterogeneous autonomy）——即指挥调度形态各异、功能不同的无人机群自主协同通讯，并在没有集中指令的情况下分配任务，确保多架无人机不会无效地重复撞击同一个目标。 ```mermaid flowchart TD accTitle: 现代战场边缘 AI 与异构无人机蜂群决策架构 accDescr: 描述战场强干扰环境下无人机如何从依赖人工操作过渡到本地边缘 AI 目标识别与蜂群协同 subgraph EdgeLevel["边缘端自主运作 (通信阻断环境)"] Sensor["传感器与机载视觉识别"] --> ATR["自动目标识别算法 (ATR)"] ATR --> Filter["真伪辨识与抗诱饵/抗眩目过滤"] Filter --> KillBox["网格巡弋模式 (Kill Box)"] end subgraph SwarmLevel["异构蜂群协同 (Heterogeneous Autonomy)"] KillBox --> Comms["跨机型机间自组网通讯"] Comms --> Deconfliction["动态目标解算与防重复分配"] Deconfliction --> Strike["精准执行打击任务"] end ``` - 异构自主（Heterogeneous Autonomy）：整合不同厂商、不同规格的无人机平台，让异构硬件实现无缝数据通信与打击协同。 - 杀伤区（Kill Box）：通信被全频段电子干扰屏蔽后，无人机依托机载芯片在限定空域自主执行目标闭环搜索与打击。 - 自动目标识别（ATR）：利用经过海量对抗样本训练的边缘计算机视觉，从诱饵与电磁干扰中实时解算真实威胁。 ### [22:21](https://www.youtube.com/watch?v=gzwRflcLPAA&t=1341s) 容错率分级与“全自动驾驶”时刻：金色穹顶下的 AI 战略 针对大语言模型和计算机视觉普遍存在的“幻觉”（hallucination）与误判隐患，主持人询问军方何时能迎来军事 AI 的“完全自动驾驶（FSD）时刻”——即系统可靠性达到比人类决策高出一个数量级的水平。迈克尔坦言，目前军事 AI 远未达到这个节点，国防部在部署时严格遵循“风险梯度”机制，根据不同战术场景设定极不相同的容错门槛。 迈克尔以反导拦截系统“金色穹顶”（Golden Dome）为例进行了详细剖析。面对高超音速导弹突防，从助推段升空到弹头分离、释放大批诱饵，留给防御系统的反应时间可能仅有短短 90 秒。在高轨道外太空环境下，人类操作员的生理神经传导速度和肉眼根本无法完成弹道辨别与拦截计算，必须依靠太空基基准 AI 进行超高速目标解算。在太空防御或使用定向能激光击落无人机的场景中，AI 误判的代价极低——激光打空直接消散在空气中，没有任何人员伤亡风险；反之，若在人口稠密地区执行地面反恐行动，军方则采取极度保守的风险态度。在未来十年内，各类场景下的 AI 渗透将稳步推进，但底线始终是严控误伤风险。 ### [24:55](https://www.youtube.com/watch?v=gzwRflcLPAA&t=1495s) 绑着单手作战的士兵与“天网”伪命题：五角大楼眼中的自主交战 主持人们回顾了美军在阿富汗等冲突中受到的严苛交战规则（Rules of Engagement）束缚。士兵们冒着生命危险奔赴海外，却被厚重的官僚规则手册捆绑，甚至因防卫作战触碰条规而面临审查，这种“一只手被反绑在身后打仗”的现实让一线官兵深感挫败与委屈。紧接着，话题自然延伸到公众对军用全自主武器（Full Autonomy）的道德恐慌，以及媒体热炒的 Anthropic 创始人达里奥·阿莫代伊（Dario Amodei）对“将击杀权移交给机器人”的伦理担忧。 迈克尔对此明确回应称，舆论将目前的军事 AI 描绘成《终结者》中的“天网”（Skynet）完全脱离了技术现实。五角大楼目前推动的项目距离“让机器人自主决定杀人”相去甚远，真正的重点是基础自主性——例如无人潜航器（UUV）的水下航行避障、侦察无人机的地形跟随，以及作为有人驾驶战机火力僚机的协同作战飞机（Collaborative Combat Aircraft, CCA）。在这些系统里，无人机始终受制于长机的战术指令。 迈克尔透露，他曾直接对 Anthropic 的高管挑明核心逻辑："你们的大模型正在被窃取，对方会剥除所有安全护栏（guardrails）并将其武器化反制我们，而你们却希望美军的模型自废武功、降低能力。" 事实上，如果将现代计算机视觉回溯到十几年前维基解密曝光的“附带谋杀”（Collateral Murder）阿帕奇直升机误杀记者事件，或者科威特战场上的友军误击事故，精准的 AI 目标辨识本可以拯救这些无辜生命。就像自动驾驶虽然让习惯人类操驾的人感到本能恐惧，但在绝大多数场景下，算法的精准度与感知广度远比疲惫、恐慌的人类士兵更加安全可控。 - 协同作战飞机（CCA）：无人僚机跟随有人驾驶长机编队飞行，受飞行员实时监管并提供防区外火力延伸。 - 技术现实与天网假想：当前部署聚焦于无人载具巡航控制、激光防空与基础感知，远未涉及全自主致命性决策授权。 - 护栏悖论：单方面在本土基础模型上强加道德限制，无法阻止竞争对手逆向工程后解除安全约束投入战场对抗。 ### [28:39](https://www.youtube.com/watch?v=gzwRflcLPAA&t=1719s) 审视中东泥潭与 MAGA 阵营质疑：特朗普及“史诗之怒行动”中的以色列角色 对话转入地缘政治的敏感核心：以色列政府在对伊朗军事行动中的推波助澜，以及美国内部政治层面的剧烈分歧。塔克·卡尔森（Tucker Carlson）和 MAGA 选民基本盘中有相当一部分声音质疑本届政府是否已被以色列利益集团绑架，甚至翻出马可·卢比奥（Marco Rubio）的言论，指出“美国参战是因为以色列无论如何都要动手”。 针对“美国在中东受制于人”的指控，迈克尔给予了直接否认。他明确表示，总统绝没有被以色列所绑架，白宫的每一项战略决策都是严格基于“美国国家利益优先”的考量而定。在这套逻辑中，以色列只是在中东地缘战略中能力极强、利益高度重叠的关键盟友。当行动符合美国利益时，美方自然会深度整合以方的顶尖战术能力。迈克尔透露，在当前代号为“史诗之怒行动”（Operation Epic Fury）的作战已经进入第四天，摩萨德（Mossad）提供的高精度情报体系正在与美军作战网络无缝协同，成为美军在该地区快速压制对手不可或缺的助力。 ### [29:59](https://www.youtube.com/watch?v=gzwRflcLPAA&t=1799s) 伊朗战力折损、欧洲的迟疑与“铁束”激光防空系统 讨论聚焦于中东战场局势。情报显示伊朗约 90% 的弹药已被消耗殆尽，几乎无法再向外发射导弹，而大批无人机和战机已处于待命状态。面对美军与盟友的高效协同，主持人们批评了部分欧洲盟友的消极反应：英国在作壁上观，西班牙在夸夸其谈，而整个欧洲甚至要等到“度完周末”才慢吞吞地发表声明。 当被问及以色列由定向能技术驱动的“铁束”（Iron Beam）激光防空系统时，国防部副部长埃米尔·迈克尔（Emil Michael）指出，以色列先部署了著名的“铁穹”（Iron Dome）作为第一代防空系统，目前正积极研发并部署“铁束”激光拦截系统。虽然该技术尚处早期阶段，但实测已展现成效。迈克尔解释，从战术角度来看，最理想的状态是在弹药处于远距离外部空域（如敌方本土上空）时就予以拦截，而不是等它飞临特拉维夫头顶。 关于未来防御洲际弹道导弹（ICBM）和高超音速导弹的前景，迈克尔表示，虽然天基激光拦截系统在技术上面临极大挑战，但其战略价值最高；与此同时，由于面临巡航导弹、低空无人机等多种威胁，必须构建“多层防御体系”（Multi-layers）。针对特朗普总统“该激光拦截技术属于美国”的说法，迈克尔澄清这属于美以联合开发：激光光束相干聚焦（beam coherence）与高功率微波（HPM）对抗无人机群的技术日新月异，其中大量核心技术由美方掌控，同时以色列也贡献了其本土的科研成果。 ```mermaid flowchart TD accTitle: Multi-Layer Air and Missile Defense Concept accDescr: A diagram illustrating multi-tier interception from space-based lasers down to ground-level directed energy systems. subgraph Layer1["天基拦截层 (Space-Based Layer)"] A["高空侦测与天基定向能激光"] -->|"拦截远距离威胁与高超音速/弹道目标"| Target1["ICBM / 高超音速导弹"] end subgraph Layer2["中远程防空层 (Standoff Defense)"] B["防空导弹与增程激光系统"] -->|"远离本土边界实施拦截"| Target2["巡航导弹 / 弹道导弹"] end subgraph Layer3["末端与近程防御层 (Terminal Ground Layer)"] C["铁穹 (Iron Dome) 动能拦截"] D["铁束 (Iron Beam) 激光 / 高功率微波 (HPM)"] C & D -->|"低空及近距拦截"| Target3["短程火箭弹 / 无人机蜂群"] end ``` - **战况数据**：伊朗全境约 90% 的弹药库存已被消耗，远程导弹火力基本停滞。 - **技术要点**：定向能武器（高能激光、高功率微波）能实现极高性价比的低成本拦截，是应对低空无人机群与巡航导弹的关键互补手段。 --- ### [33:41](https://www.youtube.com/watch?v=gzwRflcLPAA&t=2021s) 政权更迭、历史教训与“不再打三十年治安战” 主持人们深入探讨了推动伊朗政权更迭的可能性与战略考量。Jason 旗帜鲜明地表示，如果能以深思熟虑、外科手术式的方式进行，他全力支持推翻统治伊朗长达 50 年的独裁神权政权，将自由还给伊朗人民；他强调自己信任特朗普与国防部的战略决断，并重申自己是独立选民而非所谓的“白左（libtard）”。 查马斯（Chamath）随后指出，美国在处理伊朗问题时拥有极为沉痛且深刻的历史教训可供复盘：包括 1953 年英美情报机构策动政变推翻摩萨台（Mohammad Mosaddegh）并扶植巴列维国王、随后引发 1979 年伊斯兰革命；1980 至 1988 年残酷的旷日持久的两伊战争；以及 2003 年伊拉克战争的泥潭。正如社交媒体上广泛流传的一条犀利热评所言：“战争不一定非得打上整整三十年、耗费数万亿美元去喂饱弗吉尼亚、马里兰和华盛顿特区的军工复合体利益集团。”这次的战略意图是精准打击、“速战速决（in and out）”。 众人进而分析，如果特朗普政府能促成委内瑞拉与伊朗走向民主化，这两个国家合计占全球生活在独裁统治下人口的近 5%，这将是数十年来对全球民主自由进程最具实质意义的推动，完全配得上“所有领域的诺贝尔奖甚至奥斯卡奖”。现场随后演变成对 Jason 真实政治立场的打趣互怼：Jason 坚称自己是独立温和派，只支持特朗普 60% 至 70% 的政策，并透露自己并未投票给卡玛拉·哈里斯，但在伙伴们“你到底投了谁”的连番逼问下打起太极。 --- ### [37:24](https://www.youtube.com/watch?v=gzwRflcLPAA&t=2244s) 霍尔木兹海峡危机、油价飙升与海事战争保险 中东局势的迅速恶化直接引爆了能源市场与全球航运危机。原油价格周三迅速蹿升至每桶 84 美元，而作为全球石油大动脉的霍尔木兹海峡几乎陷入瘫痪。卫星监控画面显示航道内的超级油轮航行速度断崖式下跌，大量代表船舶的光点直接消失。 若霍尔木兹海峡无法立即复航，下周初全球市场将面临每日 330 万桶原油供给的硬性缺口。更致命的是金融与保险市场的连锁反应：各大跨国海事保险公司自 3 月 5 日起全面撤回并取消了波斯湾水域船舶的“战争险”（War Risk Coverage），导致该海域的超级油轮运输量在 48 小时内暴跌 94%。为此，特朗普政府迅速表态，美国将由国家出面，为穿越波斯湾的全部海上贸易（尤其是能源油轮）提供官方政治风险与战时保险背书。 弗里德伯格（Friedberg）从金融史的演变剖析了当前危机的底层逻辑： - **现代保险的起源**：现代保险业最初正是为了应对海上贸易风险而诞生的。17 世纪伦敦的“劳合社”（Lloyd's of London）最初只是一家水手、商人与航运业主聚集的咖啡馆，大家交流哪条航线能躲避海盗和风暴，并逐渐开始通过签署互助担保合同来分担货船遇难损失。 - **再保险机制的停摆**：如今的劳合社已演变为全球规模最大的再保险交易市场，旗下拥有 78 家辛迪加（Syndicate）财团，专门为油轮跨越危险水域等超大额极限风险进行精算与承保。一旦这些辛迪加因战争风险全面退场，任何商业船舶在失去保险的情况下都无法合法起航，这也正是美国政府不得不介入接盘海运承保的根本原因。 ### [39:10](https://www.youtube.com/watch?v=gzwRflcLPAA&t=2350s) 霍尔木兹海峡危机与特朗普政府的海上保险棋局 随着伊朗伊斯兰革命卫队（IRGC）宣布封锁霍尔木兹海峡，航运业瞬间陷入极端恐慌。水雷密布与商船遇袭爆炸的致命风险，导致船舶保险费率在极短时间内从原先船体价值的 0.25% 暴涨 5 倍至 1.25%。面对高昂的担保成本，全球航运市场迅速陷入停摆，直接危及全球能源价格、激化通胀并重创美国经济安全。 在这一关键时刻，美国政府出台了一项极具商业洞察力的举措：通过美国国际开发金融公司（DFC，该机构起源于饱受争议的美国国际开发署 USAID 旗下的信贷分支）动用信用额度，直接向航运企业提供航线保险。由于绝大多数航运公司都高度依赖借贷购置船只，债权人强制要求必须持有保险才能出海，否则整个资金链便会断裂。美国政府填补了私营商业保险市场溃败留下的空白，不仅保障了能源航道的通畅，更有望顺势打破长期被欧洲银团与辛迪加（如劳合社等）垄断的海上保险格局，将海上承保与经纪业务彻底回流（onshore）至美国本土，为战后的美国金融业开辟全新的万亿级产业空间。 ```mermaid flowchart TD accTitle: 霍尔木兹海峡危机与海上保险回流逻辑 accDescr: 展示伊朗封锁海峡如何导致保费飙升与航运停摆，以及美国DFC介入重塑本土保险业的过程 subgraph Crisis ["地缘危机与市场失灵"] A["伊朗伊斯兰革命卫队 (IRGC) 封锁霍尔木兹海峡"] --> B["水雷与袭击风险激增"] B --> C["保费暴涨5倍 (0.25% 飙升至 1.25%)"] C --> D["船东无力承担 / 欧洲保险辛迪加关闸撤出"] D --> E["高负债船东停运，全球能源与通胀告急"] end subgraph Solution ["美国政府破局与产业链重塑"] F["美国国际开发金融公司 (DFC) 注入信用"] --> G["为商业商船提供关键航线直接担保"] G --> H["债权人风险解除，商业航线恢复通航"] H --> I["打破欧洲独占格局，海上保险承保业务回流美国"] end E --> F ``` ### [41:19](https://www.youtube.com/watch?v=gzwRflcLPAA&t=2479s) 突发风暴：五角大楼将 Anthropic 列为“供应链风险” 彭博社披露了一则重磅突发新闻：五角大楼已正式向知名人工智能企业 Anthropic 下达通知，将其正式列为“供应链风险”（Supply-chain risk）。这是美国历史上首次将一家本土顶尖科技企业定性为供应链威胁，此前仅有俄罗斯企业以及中国的华为等受到过同等制裁。在此之前，战争部（五角大楼）已于周五紧急终止了与 Anthropic 价值 2 亿美元的采购合同。 争端的导火索集中在两项条款上：其一是 Anthropic 对所谓“全自主武器”（俗称杀人机器人，murderbots）的严苛限制，首席执行官达里奥·阿莫代伊（Dario Amodei）坚称其底层模型当前尚不够稳定可靠；其二是 Anthropic 极力反对将大模型用于对美国本土民众的“大规模监控”（mass surveillance），认定此类大模型具备远超普通街头监控摄像头的超常能力。五角大楼对此针锋相对，坚决要求享有在法律许可范围内的“所有合法用途”（all lawful use）。 - **历史罕见定性**：此前被列入该级供应链风险黑名单的几乎全为外国敌对势力实体（如华为、俄系软硬件公司），Anthropic 成为第一家被美军列为供应链风险的本土硅谷独角兽。 - **合同解约与核心冲突**：战争部废止了价值 2 亿美元的现存大单，双方彻底在“武器自动化边界”与“民用监控界限”这两大红线上公开决裂。 ### [43:06](https://www.youtube.com/watch?v=gzwRflcLPAA&t=2586s) 埃米尔·迈克尔亲述谈判内幕：“所有合法用途”与长臂管辖 战争部副部长埃米尔·迈克尔（Emil Michael）详细拆解了这一冲突的历史脉络。拜登政府时期的 AI 行政命令曾通过人为设定算力门槛确立了少数受保护的“钦定赢家”，Anthropic 正是其中之一。该公司随后通过在美军各核心作战司令部（包括负责伊朗战事的中央司令部 CENTCOM、主管对华防务的印太司令部 INDOPACOM）及多家情报机构内部署前置工程师（FDE），将其大模型深度嵌入美军核心工作流。 迈克尔在 8 月份接管 AI 事务后审阅采购合同时大为震惊：Anthropic 附带了长达 20 页的限制性条款，不仅明文禁止将模型用于规划动能打击（kinetic strike），甚至连卫星机动变轨也遭到限制。在长达三个月的拉锯谈判中，美军提出反舰高超音速导弹防御、无人机蜂群拦截等具体防务场景，Anthropic 虽答应给予个案豁免（exception），但迈克尔坚决拒绝这种逐案审批模式：“我不可能为未来 20 年美军所有潜在的 AI 战术用途一一去打申请报告。”军方的底线必须是“一切合法用途”均由军队自主决断。迈克尔直言，Anthropic 随后将争议引向了公关舆论战，试图将战争部描绘成滥用技术的威胁；但实际上军方受法律严格约束，根本没有国内监控执法权，Anthropic 甚至将使用公开信息（如在 LinkedIn 上搜索个人简历）判定为违规，其条件苛刻到令军方无法接受。 - **作战指挥层面的深度渗透**：Anthropic 模仿 Palantir 派遣现场部署工程师，将模型渗透进美军中央司令部和印太司令部的战术日常运作中。 - **条款严苛到荒谬**：服务条款甚至限制军方在沙盘推演中模拟涉及人员伤亡的动能打击行动。 - **个案审批的死胡同**：军方断然拒绝由一家民营科技公司的伦理委员会为美军战时决策充当“审批官”。 ### [46:23](https://www.youtube.com/watch?v=gzwRflcLPAA&t=2783s) 导火索爆发：马杜罗突袭行动调查与技术控制权隐患 长达数月的拉锯最终在美军针对委内瑞拉总统马杜罗的突袭行动后彻底破裂。行动结束后，Anthropic 的一名高管直接致电其总承包商 Palantir，质问美军在该突袭行动中是否调用了 Anthropic 的 AI 软件，并暗示若属实将判定美军违反了用户服务条款。这直接触碰了五角大楼的最高红线——突袭行动的技术细节属于绝密级作战情报，一家供应商居然私自调查军方的涉密战术行动，甚至意图撤销美军正在作战中的软件使用许可。 该事件震惊了包括战争部部长皮特·赫格塞斯（Pete Hegseth）在内的整个五角大楼领导层。技术架构分析显示，尽管模型运行在 AWS 专用政务云（AWS GovCloud）上并通过 Palantir 交付，但模型的控制平面（control plane）与模型权重更新权限依然掌握在 Anthropic 手中。这意味着在下一场战争的关键时刻，Anthropic 随时可以通过模型拒绝服务（refusal）、内部越权开发者投毒（poisoning）或恶意诱发幻觉，让前方将士陷入无防护的绝境。在随后由赫格塞斯、迈克尔与达里奥·阿莫代伊举行的摊牌会议上，军方给出的最后通牒被 Anthropic 漠视，合同最终全面流产。这一事件向整个企业界与政府部门敲响了警钟：基础模型厂商正在利用其对底层控制平面的把控，将自身的意识形态与政治诉求强加于客户的核心业务之上。 - **绝密战术遭内部质询**：Anthropic 高管跨过军方直接向前线软件集成商 Palantir 追问涉密作战细节，引发严峻安全警报。 - **控制平面命门**：模型权重与控制节点掌握在外部科技厂商手中，导致美军在面对技术“断供”、“拒绝响应”和“内部投毒”时毫无冗余备份。 - **商业影响溢出**：这场冲突向外界证明，当底层模型公司坚持将自身的哲学立场与政治偏好置于合同之上时，任何企业或政府机构的业务连续性都将面临前所未有的系统性风险。 ### [49:07](https://www.youtube.com/watch?v=gzwRflcLPAA&t=2947s) AI“道德审查”与去平台化重现：企业与政府为何必须拥抱多模型架构 查马斯（Chamath Palihapitiya）深入剖析了科技公司意识形态对现实商业与政府治理的巨大冲击。他指出，私营科技公司的治理哲学往往可以在一夜之间发生剧变，但政府的反应也是完全合情合理的——当一家关键供应商表现得极不可靠、随意否决合法的合理需求时，没有任何政府能将核心命脉托付其上。查马斯举了一个尖锐的现实案例：如果一个反堕胎州（pro-life state）想要运行某项公共医疗健康计划，而 Anthropic 内部持支持堕胎权（pro-choice）立场的工程师私自决定“我们不同意该州的政策，因此要修改该模型在当地的访问权限和功能”，这是否应该被允许？从表面上看，有人会辩称私营企业有权自主决定客户；但这种逻辑推导出的必然结果是，任何联邦部门、州政府、市政厅乃至私营企业，都绝对不能将自身绑定在单一 AI 模型上。 查马斯激动地强调，只要把业务押注在单一厂商身上，企业就等于坐在一颗定时炸弹上，因为“某些模型公司内部的人迟早会发起一场脱离法律、纯粹基于主观道德的疯狂说教（lunatic moral tirade），进而彻底危及你的核心业务”。周五爆发的风波彻底撕开了这个潘多拉魔盒：任何在经历此事后仍未制定多模型混用（multi-model）和模型中立（model-agnostic）策略的企业，都在承担不可承受的巨大商业风险。他直言，这本质上就是几年前社交网络“去平台化（deplatforming）的千倍升级版”——当年只要平台不认同某些言论，就会直接封禁账号；而如今被封杀的不再只是社交媒体的发言权，而是支撑现代商业运营、直接决定企业存亡的基础生产力底层。 战争部副部长埃米尔·迈克尔（Emil Michael）对此完全赞同。他补充道，正如各大 AI 巨头掌门人自己吹嘘的那样，这套技术将导致 50% 的白领失业，其破坏力与威慑力不亚于核武器，相当于在一个数据中心里塞进了 5 万个超级天才，甚至能让一个小国借此要挟全世界。AI 已经成为一种高度泛化、渗透进日常生活方方面面的通用智能底座，它根本不是像 Workday 那样随意更换竞品的普通 HR 软件。更可怕的是，Anthropic 拥有自己内部制定的所谓“AI 宪法（Constitution）”和“灵魂”，但这套准则绝非美利坚合众国宪法。当整个社会必须屈从于科技新贵们朝令夕改的主观心血来潮时，所有美国人都会感到毛骨悚然，而这在未来几年只会演变成愈发严重的深层危机。 - **去平台化升级**：从社交媒体舆论封杀升级为针对算力与通用智能基础设施的权限剥夺。 - **主观道德对抗法定权力**：商业模型受制于开发者团队未经选举的个人伦理偏好，而非国家法律法规。 - **多模型冗余的迫切性**：避免单一供应商因意识形态干预造成关键基础设施“单点故障”。 --- ### [51:51](https://www.youtube.com/watch?v=gzwRflcLPAA&t=3111s) 国防部的AI生态扫描：Grok全盘投入、Google稳步部署与山姆·奥特曼的“两面平衡” 主持人询问埃米尔·迈克尔，目前各大前沿模型在配合国防部需求上的具体态度如何，是像戴尔（Dell）或苹果（Apple）那样只单纯销售算力工具、任由客户按需使用，还是也会横加干涉？迈克尔详细拆解了各大巨头的立场：马斯克的 xAI 旗下模型 Grok 表现得最为直接干脆，全盘支持包括涉密与非涉密网络在内的“所有合法使用场景（all lawful use cases）”。迈克尔解释道，马斯克奉行的是“求真（truth-seeking）”原则，而战争部最迫切需要的正是客观真实，绝对容不下可能导致作战决策失真的意识形态偏见。在 Google 方面，军方正在其非涉密网络上大规模扩容 Gemini，并正积极推进其向涉密网络部署，目前的进度限制仅仅是因为涉密基础设施的搭建过于繁琐复杂，在合规合作态度上完全符合军方标准。 ```mermaid flowchart TD accTitle: 国防部AI供应商生态与部署格局 accDescr: 展示Grok、Google Gemini、OpenAI与Anthropic在国防部涉密/非涉密网络上的部署状态与合作态度对比。 subgraph DefenseDemand["国防部合作基准：所有合法用途 (All Lawful Use Cases)"] Grok["xAI / Grok<br/>全盘支持涉密与非涉密网络<br/>奉行求真、拒绝意识形态偏向"] Google["Google / Gemini<br/>非涉密全面推进，涉密基建推进中<br/>商业合作完全合规"] OpenAI["OpenAI / Sam Altman<br/>战备替补支持，试图保护Anthropic<br/>人才争夺战下的温和平衡派"] Anthropic["Anthropic / Claude<br/>此前深度嵌入政企，后因宪法限制受阻<br/>被定性为供应链安全风险"] end Grok --> StatusGreen["全面合格投入"] Google --> StatusGreen OpenAI --> StatusYellow["冗余替补推进"] Anthropic --> StatusRed["剥离与供应链审查"] ``` 谈到 OpenAI 及其 CEO 山姆·奥特曼（Sam Altman）时，现场气氛颇为微妙。迈克尔透露，奥特曼展现出了其标志性的“两头下注”端水风格。当与 Anthropic 的谈判出现破裂迹象时，迈克尔亲自致电奥特曼寻求替代预案，奥特曼表现得非常深明大义与爱国，立刻表态“只要国家需要，我能做什么”，并全力协助迅速部署 OpenAI 的方案。但与此同时，奥特曼在私下里也在为竞争对手 Anthropic 说情，竭力劝阻国防部不要将其定性为“供应链风险”，直言这种做法会对整个 AI 产业造成重大打击，甚至主动提出担任中间人协助双方重新谈判合规条款。 迈克尔指出，奥特曼之所以夹在中间左右为难，核心根源在于顶尖人才的争夺战。整个生成式 AI 产业的命脉几乎全维系在区区约 1,000 名顶级核心研究员手中，这就像职棒大联盟的“点球成金（Moneyball）”游戏，顶尖选手在几大巨头之间频繁跳槽流动。哪怕只流失 20% 的核心研究员，就可能导致 Anthropic 抢先推出了 Claude Code，而 OpenAI 的新版 Codex 迟迟无法落地，整个行业的竞争格局瞬间逆转。因此，即便 Anthropic 在外部公开抨击 OpenAI 并在背后大肆挖角，奥特曼依然展现出了相当的行业格局来维护这个生态。但对于国防部而言，迈克尔表示军方毫无偏向，为所有厂商提供完全一致的公平条款，核心目标是建立充分的技术冗余，观察不同模型未来究竟是各自分化还是走向能力趋同。 - **Grok（xAI）**：无保留接入所有合法场景，完全杜绝政治倾斜对战术作战决策的干扰。 - **Gemini（Google）**：合规意愿完备，目前正处于从非涉密网络向涉密专用基础设施铺设的过渡期。 - **OpenAI**：迅速承担国防备用职能，同时作为产业老兵出面斡旋，试图缓冲政府对 Anthropic 的强硬打击。 - **“点球成金”式人才博弈**：千名核心科研人员的高频流转，直接左右了关键产品（如 Claude Code 对抗 Codex）的先发节奏。 --- ### [54:12](https://www.youtube.com/watch?v=gzwRflcLPAA&t=3252s) 模型能力渐近收敛与疯狂飙升的推理账单：Ralph Wiggum 循环的代价 弗里德伯格（David Friedberg）随即提出核心疑问：既然 Anthropic 运营哲学与军方存在根本分歧，且市场上已有三四种替代方案，为什么国防部还要费力把他们保留在技术池中？迈克尔坦言，主要原因在于此前 Anthropic 深度扎根于国防生态。与其他几家尚未全力投入政企大客户销售和驻场前置部署工程师（Forward Deployed Engineers, FDE）的对手相比，Anthropic 在政企服务流程与资质上遥遥领先，导致军方短时间内很难迅速将其抽丝剥茧地剥离，其他厂商要赶上的不是单纯的模型推理能力，而是复杂的政府客户交付体系。 但在底层模型的能力差异上，迈克尔与查马斯均认为技术鸿沟正在急剧缩窄。查马斯从统计学角度指出，观察顶尖前沿模型的置信区间与误差范围，领先优势正在呈现明显的“渐近线（asymptoting）收敛”。当所有大模型都能获得充足的电力、算力和海量数据供给后，它们所展现出的输出表现正在趋于同质化。即便 Claude Code 确实具备惊艳的先发创新性，但在未来 12 个月内，竞品（如 Codex 系列）一定会迎头赶上，这就使得多模型组合的灵活性远比依赖单一神级模型更为重要。 ``` 传统软件开发 vs 盲目自主智能体（Ralph Wiggum 循环） [传统开发]：明确指令 -> 代码生成 -> 人工验证 -> 极低算力开销（低成本、高ROI） [自主循环]：粗放目标 -> 智能体盲目自循环（Ralph Wiggum Loop）-> 陷入死锁失败 -> AWS/Cursor账单暴增（成本飙升3倍，ROI崩塌） ``` 随后，查马斯分享了他旗下软件孵化公司 8090 的亲身经历，佐证了滥用 AI 模型的可怕经济账。由于看到 Anthropic 营收暴增的新闻，查马斯让团队调阅了公司自 2025 年 11 月以来的实际运营支出（OpEx），结果令他大吃一惊：短短数月内，公司的 AI 相关成本暴增了 3 倍以上。向 AWS 支付的天价推理账单、Cursor 席位费以及 Anthropic 的 API 调用开支高达数百万美元。查马斯调侃道：“消耗 Token 是非常容易上瘾的”，但残酷的现实是企业的 AI 成本每三个月翻三倍，收入却纹丝不动。其核心元凶正是开发者群体当下盲目迷恋的“拉尔夫·维古姆循环（Ralph Wiggum loops，指像《辛普森一家》中的笨小孩拉尔夫一样盲目打转的无序死循环）”——程序员将任务丢给自主智能体任其漫无目的地自我探索，结果往往是任务不仅没解决，反而换回一张来自 Cursor 的天文数字账单。查马斯透露，8090 最终被迫下达硬性指令，要求团队彻底弃用 Cursor，因为这些工具本质上只是 Claude Code 的套壳，却在无效循环中疯狂收割高额 Token 费用。迈克尔赞同地表示，随着边际任务的复杂化，算力消耗正呈指数级攀升，未来国防部同样必须面临严苛的投资回报率（ROI）审计。 - **前置工程师生态壁垒**：Anthropic 此前的先发优势源于深入政企一线的工程驻场与军工资质交付体系。 - **前沿模型性能收敛**：随着算力与电力资源的大规模注入，顶级模型之间的能力差距正迅速被抹平。 - **8090 公司的财务失控案例**：AI 推理与工具链成本自 2025 年底暴涨三倍以上，而实际业务营收未见同频增长。 - **警惕智能体盲目死循环**：以套壳工具调用底层大模型引发的“Ralph Wiggum 循环”造成海量算力与资金浪费，倒逼企业进行工具链降级与精细化管理。 --- ### [57:24](https://www.youtube.com/watch?v=gzwRflcLPAA&t=3444s) 为何定性为“供应链风险”？军工复合体的模型投毒防线 针对全网最为震动的决策——军方为何不选择与 Anthropic 悄无声息地解约、和平分手，反而动用极刑正式将其列为“供应链风险（supply-chain risk）”，迈克尔给出了直接而强硬的回应。他坚决否认这是出于报复或惩罚性质的行政举措，并从国家安全系统的架构层面给出了逻辑依据：如果一个前沿基础模型的底层逻辑天然带有源于其自身“宪法”、公司文化及员工主观倾向的政治偏向（policy bias），那么这一偏见就构成了潜在的系统性“模型投毒威胁（poisoning threat）”。 迈克尔划定了极其清晰的军民界限： 1. **民用领域完全自由**：如果波音公司（Boeing）希望使用 Anthropic 的模型去设计民用商用客机，战争部完全不会干涉，企业大可随意使用。 2. **军工武器绝不妥协**：但如果洛克希德·马丁（Lockheed Martin）或其他核心军火商使用该模型为军方设计下一代战斗机、精确打击导弹或尖端雷达武器的核心精密元器件，军方就绝不能接受。 迈克尔严肃强调，一旦一家基础模型厂商将其单方面的意识形态偏好置于国家指令与技术客观性之上，其输出结果的真实性与中立性就荡然无存。当军工复合体的底层研发链条可能被带有偏见的代码或算法所污染时，国防部就必须将其依法定义为不可信赖的供应链风险，将其彻底阻隔在国防安全的核心大门之外。紧接着，现场话题转向了这场风波为何迅速演变为公开的舆论大战，主持人追问这究竟是 Anthropic 借助精密的公关机器主动开火，还是政府部门的高调施压所致。 - **定性逻辑非惩罚性**：列入供应链风险是出于防止国防研发体系遭遇底层逻辑“算法投毒”的主动防御。 - **清晰的军民分界线**：民用飞机设计完全解绑，涉及军工主承包商（如洛克希德·马丁）的尖端武器及元器件研发则实施绝对禁入。 - **核心信任破裂**：模型对私营意志的坚持一旦凌驾于军工客观需求之上，即被视为危及装备安全性的潜在系统漏洞。 ### [58:46](https://www.youtube.com/watch?v=gzwRflcLPAA&t=3526s) 危机时刻还要“打电话请示”：Anthropic 与国防部的根本分歧 埃米尔·迈克尔（Emil Michael）揭露了 Anthropic 公关战术背后的虚伪面目。他指出，Anthropic 非常擅长把国防部的正常运作歪曲为所谓的“大规模监控”（mass surveillance）。事实上，对方紧抓不放的所谓问题，无非是国防部打算采购并使用某些任何商业机构都能购买的公开数据库，而国防部使用这些数据库通常仅仅是为了征兵（recruiting）——毕竟国防部不仅负责打仗，还运营着庞大的学校、医院等公共机构。Anthropic 巧妙地操弄这些极具吓人噱头的议题来引发公众恐惧，但国防部真正担忧的根本危机在于：在瞬息万变的国家安全关键时刻，Anthropic 会不会突然单方面切断系统，或者在关键节点暗中干预破坏系统运行。 主持团队随之指出，这就像国防部购买电池或笔记本电脑，只要军方的使用完全合法合规（lawful use），这就理应足够；如果企业是彻头彻尾的和平主义者（peaceniks）、不想卷入军工武器销售，大可以像当年的 Google 一样彻底不参与，但既然选择了向“战争部”（Department of War）供货，就绝不能既要军方订单又拒绝履行职责，正如查马斯（Chamath Palihapitiya）所讽刺的：“如果你不想让你的产品用于战争部的业务，那你一开始就不该把东西卖给战争部，毕竟名字里就写得清清楚楚。” 更令人震惊的荒谬细节发生在具体的闭门谈判中。当面对突发极端威胁（如“9·11”式的黑天鹅事件或“金穹”防空系统紧急拦截场景）时，达里奥·阿莫代伊（Dario Amodei）竟然在二十多人的众目睽睽之下轻描淡写地表示：“如果你们需要破例（exception），到时候直接给我打电话就行。”埃米尔对此极为震惊反驳道，当战争危机瞬间爆发、必须立即采取决定性反击行动时，军方指挥官根本不可能停下来给硅谷的某个 CEO 打电话请示批准。主持人感叹，达里奥的这种特权做派，本质上就是妄图越俎代庖扮演真正的“战争部长”乃至掌握生杀大权的“神王”（god king）。 - **公开数据库与监控之辩**：国防部采购的商业数据库多用于军队征兵日常，Anthropic 却将其渲染为令人恐慌的“大规模监控”。 - **不可承受的系统断供风险**：五角大楼的核心红线是供应商绝对不能在战备与交火时刻切断服务或干涉操作。 - **“给我打电话”的离谱要求**：达里奥在面对危机模拟（如防空拦截或突发袭击）时要求军方临机申请特批，被埃米尔痛斥为完全失去理智的不可行方案。 --- ### [01:01:06](https://www.youtube.com/watch?v=gzwRflcLPAA&t=3666s) 政治反噬与失焦的董事会：左右不讨好的 Anthropic 查马斯指出，Anthropic 此举不仅彻底得罪了特朗普政府与全体共和党人，如果他们寄希望于从民主党那里获得庇护，那也完全是打错了算盘。美国进步派民主党人（progressives）在骨子里就对硅谷巨头和技术精英怀有深仇大恨，他们绝对不会容许硅谷傲慢自大地坐着一个不受民意控制的“神王”。一旦未来进步派试图对这些技术施加严苛控制而遭到 Anthropic 抵抗，两者的联盟必将迅速破裂。Anthropic 的这种自负姿态，实际上是把美国政界的所有关键派系都得罪了个遍。 讨论迅速转向 Anthropic 内部治理与董事会的严重失职。主持人直言，作为一家军工或准军工供应商，你绝不可能指望枪械制造商史密斯威森（Smith & Wesson）在卖出武器后，还要求军方射击前必须给董事会打电话报批。面对达里奥试图向五角大楼高层发号施令、单方面推翻行业内通行的“合法合规使用”标准的狂妄行径，人们不得不质疑：Anthropic 的高管团队和董事会最近究竟在干什么？难道整个董事会仅仅因为眼红于疯狂飙升的营收增长曲线，就集体“闭上眼睛、闭上嘴巴”，彻底听任达里奥肆意妄为？ - **两头不讨好的政治困境**：激怒共和党的同时，进步派民主党也因痛恨硅谷不受监管的技术霸权而绝非可靠盟友。 - **治理失控的董事会**：面对企业核心与政府对抗的公关与经营危机，董事会因产品极速膨胀的业绩而陷入集体沉默与纵容。 --- ### [01:02:39](https://www.youtube.com/watch?v=gzwRflcLPAA&t=3759s) 估值豪赌、模型趋同与云巨头的终局优势 谈到 Anthropic 的公司估值与业务表现，主持人坦言这家公司近期的增长速度在商业史上罕见——仅在过去 90 天内就狂揽了 60 亿美元的 ARR（年化经常性收入），其中由 Claude 推出的 OpenClaw 及 Cowork 等先进代码与工作流工具引发了全行业开发者的疯狂追捧，甚至被视作真正兑现了 AI 诺言的颠覆级产品。 然而，在二级市场与私有股权定价上，嘉宾们产生了激烈争辩： 1. **净市值创造 vs. 估值乘数**：若在 Google、OpenAI 与 Anthropic 之间做选择，Google（当前市值约 3 万亿美元）只要翻倍就能带来 3 万亿美元的惊人绝对价值，被视为最稳健的“市场价值创造者”；而 Anthropic 则被看作是博取 5 倍倍数（冲向 1.5 万亿美元潜在市值）的弹性标的，但目前与军方的正面冲突为其未来蒙上了巨大阴影。 2. **护城河的短暂性与竞争反扑**：弗里德伯格（David Friedberg）指出，尽管 Cowork 体验惊艳，但 Google 预计在 90 天内就会在虚拟端推出整合了 G Suite（Google Workspace）的同类竞品并迅速横扫市场；同时埃隆·马斯克也宣称 xAI 的 Grok 将在 2026 年 5 月至 6 月的模型迭代中全面超越现存的所有代码模型。 面对查马斯提出的终极设想——如果到 2026 年 10 月所有主流模型能力彻底收敛趋同（asymptote），国防部将如何抉择？埃米尔明确表示，他最期待的就是模型大宗商品化（commodity），因为届时军方可以纯粹按价格竞争进行招采。国防部的战略是保持“一主一备”或“双主力”，至少维持两到三家相互竞争的供应商。如果 Anthropic 依然维持当下高高在上的对抗姿态，国防部会毫不犹豫地将其剔除出局。此外，埃米尔还强调了 Google 拥有底层全栈云基础设施（Own Cloud）的巨大长期成本优势，相较于必须在第三方云成本之上加价转嫁给客户的 Anthropic，Google 具备不可逾越的经济性壁垒。 ```mermaid flowchart TD accTitle: 模型能力收敛下的国防部 AI 采购逻辑 accDescr: 展示当模型能力无差异化后，五角大楼转向价格竞争与多供应商冗余的决策路径 subgraph Converge["2026年秋季预期：模型能力趋同 (Asymptote)"] M1["Anthropic Claude"] M2["OpenAI 模型"] M3["Google Gemini"] M4["xAI Grok"] end subgraph DefenseReq["五角大楼采购与风控标准"] Std1["无条件接受'合法合规使用'"] Std2["双主力 / 一主一备冗余架构"] Std3["完全自主掌控，杜绝断供风险"] end subgraph Outcome["最终格局"] A_Out["剔除 Anthropic (政治与政策姿态僵化)"] Price_Comp["其余厂商展开纯粹价格竞争"] Cloud_Edge["Google 凭借自有云算力成本优势胜出"] end Converge --> DefenseReq DefenseReq --> A_Out DefenseReq --> Price_Comp Price_Comp --> Cloud_Edge ``` - **90 天暴增 60 亿 ARR**：由强大的编码工具驱动，Anthropic 营收呈现史无前例的指数级爆发。 - **模型均质化后的博弈**：能力拉平后，国防部将完全根据价格竞价，并保持 2 至 3 家供应商以防单点故障或政策摇摆。 - **自有云的核心护城河**：Google 凭借自有数据中心和算力底座，免除了外部云厂商利润剥削，在长期采购中极具价格战耐力。 --- ### [01:06:30](https://www.youtube.com/watch?v=gzwRflcLPAA&t=3990s) Google Agent 闪电反击与“政治献金”阴谋论交锋 正当弗里德伯格分析 Google 很快会反扑时，制作人尼克（Nick）恰好当场切出了 Google 当天刚刚发布的重磅官方公告：Google Workspace 正式全面集成 Agents，并首发上线了多达 40 项 Agent 实用技能。主持人纷纷戏称弗里德伯格简直是未卜先知的“预言家”，Google 动作之快完全印证了巨头正在以闪电速度抹平纯模型层创业公司的产品差距。 在讨论的最后，杰森·卡拉卡尼斯（Jason Calacanis）直接向埃米尔抛出了一记尖锐的“直球”质问：达里奥·阿莫代伊私下向外界声称，国防部和特朗普政府之所以突然把 Anthropic 定性为“供应链风险”并全面封杀，真实原因根本不是什么国家安全，而是因为 Anthropic 没有向特朗普政治阵营捐款；相反，OpenAI 及其总裁格雷格·布罗克曼（Greg Brockman）以及微软、苹果（Tim Apple）、英伟达、亚马逊等九大科技巨头，不仅豪掷政治献金，还积极出席总统就职典礼、白宫 CEO 晚宴，甚至出资赞助梅拉尼娅的纪录片。达里奥指责五角大楼是在搞赤裸裸的“政治站队与利益寻租报复”。这一尖锐指控立刻将节目的戏剧张力推向了最高峰。 - **Google Workspace Agent 落地**：同日发布 Agent 集成并上线 40 项功能，全面对标挤压独立 AI 应用市场。 - **达里奥的政治报复指控**：声称被列入供应链黑名单是因为未向特朗普团队“纳贡交保护费”，而 OpenAI 等 9 家巨头皆因深度政治进贡而获得庇护。 ### [01:07:35](https://www.youtube.com/watch?v=gzwRflcLPAA&t=4055s) “入场费”与政治献金指控：Anthropic 被列入风险名单的真正原因 主持团队抛出一个尖锐的质疑：外界有声音认为，特朗普（Donald Trump）政府充斥着“权钱交易”（pay to play），科技巨头们纷纷出席活动、向政界纳投名状或捐款，唯独 Anthropic 既不肯屈膝低头（genuflecting），也不愿意交这笔政治“入场费”（cover charge），这是否是他们被国防部单独针对的原因？对于这种阴谋论，美国国防部副部长埃米尔·迈克尔（Emil Michael）断然否认，并直斥这是他听过“最愚蠢的言论之一”。迈克尔解释道，自己身处国防部（Department of War），核心职责就是打赢战争，如果一家公司的模型能帮美军赢得胜利，自己高兴都来不及，根本不会想浪费几个月的时间去把对方替换出供应链。 迈克尔进一步驳斥道，特朗普总统从未插手并向他示意过“这家公司没给钱，不能再用他们”之类的话；相反，过去整整三个月里，他本人一直极力与 Anthropic 保持沟通与谈判，试图建立一个简单的合作基准。如果真如外界所传是出于政治报复，国防部早在三个月前就可以直接通知他们出局，何必耗费漫长的时间谈判？他认为，把正常的国家安全与供应链可靠性审查脑补成政治清洗，完全是某些人自身思想深处的“内部妄想综合征”（internal psychosis）。 ```mermaid flowchart TD accTitle: 外界质疑与国防部实际决策逻辑对比 accDescr: 对比了外界所传的政治报复入场费逻辑与国防部基于战争胜利和供应链可靠性的实际考量流程。 subgraph Rumor["外界阴谋论假设"] A["未支付政治'入场费'"] --> B["政治报复 / 权钱交易"] B --> C["被国防部列为供应链风险"] end subgraph Reality["国防部实际考量"] D["核心目标：打赢未来战争"] --> E["谈判历时 3 个月寻求共识"] E --> F{"条款能否确保战时稳定可靠？"} F -- "自主武器条款受限且动辄反水" --> G["认定为不可控供应链风险"] F -- "支持一切合法用途且坚守服务" --> H["合格的长期防务合作伙伴"] end ``` ### [01:08:47](https://www.youtube.com/watch?v=gzwRflcLPAA&t=4127s) 内部决策迷思：Claude 策略建议与董事会视角的反思 在谈到 Anthropic 首席执行官达里奥·阿莫代伊（Dario Amodei）为何采取如此强硬且带有敌意的姿态时，全场展开了热烈讨论。主持群甚至调侃起硅谷内部的决策机制：Anthropic 内部可能跑着一个巨大的协作模型实例（Cowork），当高管向 AI 咨询商业战略时，Claude 机器人或许告诉他们“反其道而行之（zig where they zag）能获得巨大的媒体声浪”，结果团队真听信了 AI 的建议，直接让国防部“滚一边去”（pound sand），而这最终可能被证明是一次极其灾难性的商业决策。 随后，主持人从董事会治理角度提出了两种截然不同的建议路径： - **政治妥协与实用主义路线**：如果站在董事长立场，就应当把达里奥拉到一旁敲警钟：“你固然是技术天才，模型也确实拔尖，但跟五角大楼对着干绝不可能赢。当年拜登（Joe Biden）冷落埃隆·马斯克（Elon Musk）激起了反弹，如今蒂姆·库克（Tim Cook）连梅拉尼娅（Melania Trump）的电影首映礼都会出席。不一定要捐款，但至少要出席白宫晚宴，展现出基本的爱国主义姿态，别让自己被贴上反美标签。” - **商业契约与合法性路线**：另一位合伙人则反对上述“奉迎式”建议，提出达里奥应立刻致电副部长埃米尔认错，并公开发布声明，明确表态两条原则：第一，Anthropic 承诺全力支持客户将其模型用于一切符合法律规定的用途；第二，服务条款（TOS）一旦确立便坚如磐石，绝不会在半途因内部意识形态偏好随意变更，给军方提供确定性与可靠性。 埃米尔·迈克尔对此回应表示，这正是国防部一直以来的诉求。面对未来自主作战系统的演进，军方需要的是在关键时刻不会神经质大发（wig out）的稳定盟友。然而，达里奥刚刚在周五发表的长达 5000 字长文中，将特朗普直接定性为“准独裁者”（wannabe dictator），埃米尔直言：“现在你得道歉的对象可远不止我一个人了。” ### [01:11:14](https://www.youtube.com/watch?v=gzwRflcLPAA&t=4274s) 国防硬件与工业供应链：摆脱对华依赖的重塑之路 话题随之转向更为宏观的美国国防硬件供应链建设。主持人询问，目前美国国防制造业在降低对中国零部件依赖方面究竟处于什么阶段？埃米尔坦言，目前仍处于“非常早期的起步阶段”。为了逆转制造业长期外流造成的产业空心化，他负责主管的战略资本办公室（Office of Strategic Capital）已获得高达 2000 亿美元的直接借贷授权。其运作机制是向关键本土制造企业提供“国债基准利率 + 100 个基点（Treasuries + 100 bps）”的低息贷款，涵盖固体火箭发动机（solid rocket motors）、电池、玻璃纤维等核心基础工业领域，以此重建严重依赖中国外包的国防工业基础。 埃米尔强调，这一重塑工程极其漫长且紧迫： - **关键矿产（Critical Minerals）**：属于首要解决阶段，预计在当前任期结束前能够实现通过本土供应及盟友网络完全掌控所需资源。 - **先进电池与锂供应链**：技术和上游原材料目前被中国高度垄断，是国防部接下来攻坚的头号难题。全美大约梳理出了 20 项核心关键瓶颈，各企业从落地拿钱到完成实体工厂建设，至少需要数年时间。 - **弹药与老旧武备风险**：虽然目前美军尚未爆发绝对的弹药断供危机，但由于俄乌冲突的巨额消耗，加上欧洲盟友在军援补给上进展迟缓，全球弹药储备严重承压。雪上加霜的是，美军当前很大程度上仍在依赖 20 世纪 80 年代冷战时期的老式武备系统作战，部分核导弹服役已达 50 年，部分战机服役已达 40 年，整个国防军火库急需全面更新换代（plus up）。 ### [01:13:50](https://www.youtube.com/watch?v=gzwRflcLPAA&t=4430s) 防务科技（Defense Tech）新浪潮与军工采购官僚壁垒 随着资本大量涌入，防务科技是否已出现泡沫？埃米尔透露，目前流向防务科技领域的风险投资资金规模是去年的三倍之多，行业资金十分充裕。然而，光有资本注入远远不够，当前的瓶颈在于国防采购系统的闭环：五角大楼必须让这些新兴科技初创企业迅速赢得大额采购合同，从而形成“风投资本流入 $\to$ 优秀创业者加入 $\to$ 交付现代化战力 $\to$ 商业化成功”的良性飞轮。现代战争形态正从耗资 200 亿美元、建造周期长达 15 年的传统超级航空母舰，转向低成本、高损耗容忍度（attritable）的大规模无人作战系统，而这恰好是初创企业的核心优势所在。 针对“未来市场是否只会被 Anduril、Palantir 等少数几家‘新一代主承包商’（New Primes）垄断”的疑虑，埃米尔表示虽然 Anduril 和 Palantir 渴望成为绝对垄断寡头，但他致力于扶植持续打破垄断的第二梯队。他举例提到，自己曾拜访一家仅有 30 名员工的本土独立小型军工厂，凭借极其创新的制造与工程设计，该企业每年能独立生产上千枚用于援助乌克兰的 RAM 防空拦截导弹。传统老牌军工巨头之所以能长期把持防务订单，并非因为产品最优秀，而是因为它们掌握了在极其繁琐的政府官僚系统里运作的潜规则，将“熟悉官僚审批”做成了护城河。埃米尔表示，国防部当前推行的改革重点正是全面拆解这种繁文缛节，特别是彻底重塑采购指标需求（requirements reform），打破以往海陆空各军种为一架新型战机无休止堆砌苛刻指标的陈旧机制，让真正具备创新实力的民营科技企业能够快速接入美军采购体系。 ### [01:16:37](https://www.youtube.com/watch?v=gzwRflcLPAA&t=4597s) 国防采购制度改革：打破成本加成与引入初创企业 国防部副部长埃米尔·迈克尔（Emil Michael）深入剖析了传统国防军工采购体制的致命弊端。在旧有模式下，五角大楼往往会提出一份长篇大论、极其繁琐却在工程上根本无法落地的技术需求清单；而传统的军工承包商对此毫不在意，甚至乐见其成，因为他们拿的是“成本加成”（cost-plus）合同。承包商会毫无负担地承诺按部就班推进，两年后军方才发现方案从底层工程设计开始就存在硬伤，结果导致项目一再延期数年，成本暴增数十亿美元。迈克尔指出，他每天醒来都在努力推动一场颠覆性的变革：五角大楼不再直接规定死板的工程细节，而是转为向市场明确提出“共同作战问题”（common operational problem）——例如“我们需要一批射程在 500 英里以上、具备特定爆破杀伤力的导弹”，尽可能减少繁复的条框限制，让企业自主带着解决方案前来竞标。 在合同条款设计上，迈克尔正全力推动五角大楼向现代商业合同靠拢，而这正是科技初创企业展现巨大优势的领域。初创企业非常乐意接受固定成本定价（fixed cost pricing），并接受明确的绩效奖惩条款：“延期交付就少付钱，提前交付则获得奖励”。主持人对此深表赞许，认为这种做法极具颠覆性，能彻底盘活民营创新生态；政府只需发布战略需求（例如“霍尔木兹海峡至关重要，我们必须保障其通航，需要相应装备设备来实现”），让具有创造力的创业者去发挥工程想象力，而非让传统防务巨头通过繁冗拖沓的程序来套取超额利润。 ```mermaid flowchart TD accTitle: 国防采购模式转型对比 accDescr: 对比传统军工采购模式与迈克尔推动的初创企业商业化采购新模式 subgraph Legacy["传统军工承包模式 (Legacy Cost-Plus)"] direction TB L1["军方下达繁琐且不可行的技术指标"] --> L2["承包商签署成本加成合同 (Cost-Plus)"] L2 --> L3["工程硬伤延误数年，成本追加数十亿美元"] end subgraph Reform["现代化商业竞标模式 (Problem-Driven)"] direction TB R1["军方明确作战任务与问题 (如射程、毁伤力)"] --> R2["科技企业自主设计工程方案并自由竞标"] R2 --> R3["固定价格合同与敏捷交付 (提前奖/逾期罚)"] end ``` ### [01:17:45](https://www.youtube.com/watch?v=gzwRflcLPAA&t=4665s) DARPA 的技术突破：合成生物采矿与网络攻防 主持团队随后谈到迈克尔分管的另一个重磅机构——国防高级研究计划局（DARPA）。作为现代互联网的诞生地，DARPA 曾孵化出无数改变世界的关键底层技术。迈克尔坦言，DARPA 绝对是他所有管辖业务中最钟爱的板块。在“曼哈顿工程”之后相当长的一段时间里，联邦政府公职人员的光环有所褪色，顶级顶尖人才若想造火箭大多会选择去 SpaceX；但 DARPA 至今仍是全美顶尖天才心向往之的殿堂，聚集着最富创造力的大脑。 迈克尔列举了 DARPA 当前正在攻关且已解密的部分核心课题，其中之一是利用合成生物学技术来提取和合成关键矿产资源（Critical Minerals）。该项目的目标是通过生物工程手段直接从地下提取或富集关键元素，彻底摆脱传统矿石冶炼加工中污染严重、工序复杂的繁琐工序，这将使美国能够以更低成本、更快速度掌握战略关键矿物，从而在科技博弈中实现对中国供应链的“弯道超车”。此外，AI 时代的网络攻防是 DARPA 目前深度布局的重中之重；随着基于 AI 的自主代理（AI Agents）被用于发动大规模网络攻击（近期针对 Anthropic 的系统攻防事件便是一例），AI 驱动的网空威胁已成为未来战争的巨大风险节点。尽管出于极高保密级别的要求无法透露更多细节，迈克尔强调 DARPA 仍在诸多不可思议的前沿领域推进着极具前瞻性的探索。 ### [01:19:18](https://www.youtube.com/watch?v=gzwRflcLPAA&t=4758s) 飞碟传闻、中美军事力量评估与告别寄语 在节目的问答收尾阶段，主持人们抛出了两个充满好奇的终极问题：第一，军方究竟有没有关于外星人的机密？第二，抛开外界夸大的宣传，美国面对中国军事力量崛起的真实底牌到底如何？迈克尔打趣地回应称，他确实曾据理力争去争取过外星人与不明飞行物（UAP）的机密业务管辖权，甚至团队里的下属都兴奋地央求他向国防部长申请接管，但遗憾的是最终没能如愿；他笑着表示“只要能获得 100% 的全部权限，我一定会接手，那绝对会超级精彩”。 谈及严肃的中美军事竞争议题，迈克尔十分坦诚地指出了现实的严峻性：过去 15 年里，中国经历了世界历史上规模最大、速度最快的军事扩张；而与此同时，美国在某种程度上由于长期深陷于全球反恐战争而“在方向盘前打了个盹”，导致对手在未受到足够战略遏制的情况下迅猛发展。但迈克尔同样明确强调，美国公众不必陷入恐慌，美军在综合实战能力、深海潜艇部队以及天基太空层（Space Layer）技术等方面依然拥有全球最顶尖、最尖端的绝对底蕴。核心关键在于，美国绝不能再心存自满，必须全力阻止技术差距被进一步缩小。 在全场节目的尾声，主持人深有感触地总结道，回顾本周在中东爆发的剧烈动荡（从多哈、迪拜到特拉维夫，民用住宅和无辜平民均直面空袭导弹的威胁），才真正让人警醒地意识到生活在美国拥有如此完备的国家防御是多么幸运。一位原本信奉小政府原则与自由意志主义的主持人坦言，自己近来观念发生了深刻转变：唯有强有力的国防体系筑牢安全基石，公民才能享有追求个人理想与所有自由的权利。主持团队对埃米尔·迈克尔的开诚布公与真知灼见表达了高度感谢，随后在轻松幽默的播客片尾花絮中正式结束了本期节目。 ## Notable quotes > [02:30](https://www.youtube.com/watch?v=gzwRflcLPAA&t=150s) 这或许不是一场所谓的政权更迭战争，但伊朗政权确实被更迭了，而且世界因此变得更美好。 — Pete Hegseth > [16:06](https://www.youtube.com/watch?v=gzwRflcLPAA&t=966s) 如果站在我们的立场，你希望这场战斗是对等的还是不对等的？你当然希望这是一场极度不对等的碾压战。 — Emil Michael > [16:06](https://www.youtube.com/watch?v=gzwRflcLPAA&t=966s) 带着明确的目标倾巢出动，凭借压倒性武力取得胜利，然后干脆利落地撤出。 — Emil Michael > [24:55](https://www.youtube.com/watch?v=gzwRflcLPAA&t=1495s) 你们的大模型正在被窃取，对方会剥除所有安全护栏并将其武器化反制我们，而你们却希望美军的模型自废武功、降低能力。 — Emil Michael > [33:41](https://www.youtube.com/watch?v=gzwRflcLPAA&t=2021s) 战争不一定非得打上整整三十年、耗费数万亿美元去喂饱弗吉尼亚、马里兰和华盛顿特区的军工复合体利益集团。 — Chamath Palihapitiya > [43:06](https://www.youtube.com/watch?v=gzwRflcLPAA&t=2586s) 我不可能为未来 20 年美军所有潜在的 AI 战术用途一一去打申请报告。 — Emil Michael > [58:46](https://www.youtube.com/watch?v=gzwRflcLPAA&t=3526s) 如果你不想让你的产品用于战争部的业务，那你一开始就不该把东西卖给战争部，毕竟名字里就写得清清楚楚。 — Chamath Palihapitiya > [58:46](https://www.youtube.com/watch?v=gzwRflcLPAA&t=3526s) 如果你们需要破例，到时候直接给我打电话就行。 — Dario Amodei > [51:51](https://www.youtube.com/watch?v=gzwRflcLPAA&t=3111s) 只要国家需要，我能做什么 — Sam Altman > [54:12](https://www.youtube.com/watch?v=gzwRflcLPAA&t=3252s) 消耗 Token 是非常容易上瘾的 — Chamath Palihapitiya > [01:08:47](https://www.youtube.com/watch?v=gzwRflcLPAA&t=4127s) 现在你得道歉的对象可远不止我一个人了。 — Emil Michael ## Mentioned Emil Michael, Jason Calacanis, Chamath Palihapitiya, David Friedberg, David Sacks, Pete Hegseth, Donald Trump, Dario Amodei, Sam Altman, Elon Musk, Marco Rubio, Ali Khamenei, Nicolás Maduro, Travis Kalanick, Steve Witkoff, Jared Kushner, Colin Powell, Qasem Soleimani, Greg Brockman, Tim Cook, Anthropic, Claude, Claude Code, OpenAI, Codex, xAI, Grok, Google, Gemini, Google Workspace, Palantir, Anduril, AWS, AWS GovCloud, Cursor, Lockheed Martin, Boeing, DJI, DARPA, DFC, Lloyd's of London, Polymarket, 史诗之怒行动 (Operation Epic Fury), 午夜铁锤行动 (Operation Midnight Hammer), 铁束 (Iron Beam), 铁穹 (Iron Dome), 金色穹顶 (Golden Dome), LUCAS 无人机系统, 异构自主 (Heterogeneous Autonomy), 协同作战飞机 (CCA), 战略资本办公室 (Office of Strategic Capital), 8090 ## Video description (0:00) The Besties welcome Under Secretary of War Emil Michael (2:30) US war with Iran: Bigger picture and why now? (13:16) Trump's new approach to warfare, AI, drones, rules of engagement (28:39) Israel's role in the conflict, relationship with the US, Iron Beam (37:24) Oil prices, Trump's maritime insurance play (41:19) Pentagon vs Anthropic: Why Anthropic was labeled a supply-chain risk (1:02:03) How to value Anthropic after its supply chain risk designation (1:11:14) State of the US defense supply chain, the defense tech industry, DARPA, and China's military Follow Emil Michael: https://x.com/USWREMichael https://x.com/emilmichael Follow the besties: https://x.com/chamath https://x.com/Jason https://x.com/DavidSacks https://x.com/friedberg Follow on X: https://x.com/theallinpod Follow on Instagram: https://www.instagram.com/theallinpod Follow on TikTok: https://www.tiktok.com/@theallinpod Follow on LinkedIn: https://www.linkedin.com/company/allinpod Intro Music Credit: https://rb.gy/tppkzl https://x.com/yung_spielburg Intro Video Credit: https://x.com/TheZachEffect Referenced in the show: https://x.com/chamath/status/2029584905831891069 https://polymarket.com/event/us-forces-enter-iran-by https://polymarket.com/event/will-the-iranian-regime-fall-by-the-end-of-2026 https://x.com/chamath/status/2029416079781736844 https://x.com/USWREMichael/status/2029539950962626734 https://x.com/addyosmani/status/2029372736267805081 https://github.com/googleworkspace/cli https://x.com/chamath/status/2029634071966666964 https://www.lloyds.com/about-lloyds/history/lloyds-buildings https://www.bloomberg.com/news/articles/2026-03-05/pentagon-says-it-s-told-anthropic-the-firm-is-supply-chain-risk https://www.bloomberg.com/news/articles/2026-03-03/anthropic-nears-20-billion-revenue-run-rate-amid-pentagon-feud #allin #tech #news [Source](https://www.youtube.com/watch?v=gzwRflcLPAA)

[2:30] Pete Hegseth: 这或许不是一场所谓的政权更迭战争，但伊朗政权确实被更迭了，而且世界因此变得更美好。 [16:06] Emil Michael: 如果站在我们的立场，你希望这场战斗是对等的还是不对等的？你当然希望这是一场极度不对等的碾压战。 [16:06] Emil Michael: 带着明确的目标倾巢出动，凭借压倒性武力取得胜利，然后干脆利落地撤出。 [24:55] Emil Michael: 你们的大模型正在被窃取，对方会剥除所有安全护栏并将其武器化反制我们，而你们却希望美军的模型自废武功、降低能力。 [33:41] Chamath Palihapitiya: 战争不一定非得打上整整三十年、耗费数万亿美元去喂饱弗吉尼亚、马里兰和华盛顿特区的军工复合体利益集团。 [43:06] Emil Michael: 我不可能为未来 20 年美军所有潜在的 AI 战术用途一一去打申请报告。 [58:46] Chamath Palihapitiya: 如果你不想让你的产品用于战争部的业务，那你一开始就不该把东西卖给战争部，毕竟名字里就写得清清楚楚。 [58:46] Dario Amodei: 如果你们需要破例，到时候直接给我打电话就行。 [51:51] Sam Altman: 只要国家需要，我能做什么 [54:12] Chamath Palihapitiya: 消耗 Token 是非常容易上瘾的 [68:47] Emil Michael: 现在你得道歉的对象可远不止我一个人了。

[6:03] Chamath Palihapitiya: If your goal is to prevent war with China, which is a massive global conflict... how would you do it? You hem in China going into a massive moment at the end of March. [18:05] Emil Michael: Long one-way attack drones are important for... any kind of major conflict like you're seeing in Iran, but also to protect military bases, America's 250th anniversary, the World Cup, the Olympics in '28. [22:21] Emil Michael: Your models are getting stolen by the Chinese. They're going to un-guardrail them and use them against us, and then you want our models to be less capable against your models? [41:19] Emil Michael: I looked at contracts. I was like, holy cow, they say you can't use them to plan a kinetic strike. You can't use their AI model to move a satellite... I said, okay, well... I need to have the terms of service be rational relative to our mission set. [49:11] Chamath Palihapitiya: This is deplatforming all over again... This is that times a thousand because this is not about posting on social media. This is about using fundamental technology to either advantage or disadvantage your business. [51:51] Emil Michael: We don't want ideology because ideology will mess with operational decisions... you don't want anything to be fake or tilted. [54:04] Chamath Palihapitiya: Token use is very addicting... Everybody has gotten infatuated with what we call these Ralph Wiggum loops—just send the thing off and like it'll just go figure something out. A, it never figures anything out, and B, you just get this ginormous bill. [57:24] Emil Michael: He's like, 'Just call me if you need another exception.' And I'm like, but what if the balloon's going up at that moment... I'm not going to call you to do something. It's not rational. [61:03] David Sacks: I don't think any business in history has grown as fast as they have in the last 90 days... They've added what was it? 6 billion of ARR in a month or something. [68:52] David Friedberg: Hey Dario, call Emil back right now and say, 'You know what? Sorry, we blew up. We're going to own this and we're going to put out a press release that says we support our customers' use of our models to do everything and anything that's lawful... and that our terms of service are written in stone.' [68:52] Emil Michael: I need a reliable, steady partner that...'s not going to wig out in the middle... You called President Trump in your 5,000-word essay on Friday a wannabe dictator. You're going to have to apologize to more people than just me. [73:50] Emil Michael: One of the reasons the primes are such a small number... is they learned how to contract with the government. They learned how to go through the bureaucracy, and that became a competitive advantage. I'm trying to take that competitive advantage away.

[0:00] All right everybody, emergency podcast [0:02] time. Episode 263 of All-In. We have [0:04] Emil Michael, the under secretary of war [0:08] for research and engineering working [0:10] directly for Pete Hexa. We had to get [0:13] this out to you on Thursday night [0:15] because it is an emergency pod. One of [0:18] my old besties. Emil Michael is here. [0:20] Emil and I uh were part of team Uber [0:23] back in the day. He was Travis's [0:25] right-hand man. Some might say fixer. [0:28] And Emil Michael is now the under [0:30] secretary for war [0:33] here in the United States, serving his [0:35] country like our bestie David Sachs. [0:38] Welcome to the program for the first [0:39] time. Neil Michael, how you doing, [0:41] brother? [0:41] >> I'm doing good. Uh I hope it was more [0:43] than a fixer, but you know, [0:46] >> raising 20 billion fixer. I mean, you [0:48] got it done. You got it done. The [0:50] hardest. He would give you the hardest [0:51] things. Yeah. Just [0:52] >> That's right. Fair enough. [0:53] >> If it was hard, and that's what a fixer [0:55] is. [0:55] >> An operational axe. That's what they [0:58] call them. [0:58] >> All right. [clears throat] Sure. Uh, in [0:59] Brooklyn, we call them fixers. With us [1:01] again, [1:01] >> a rain maker. [1:03] >> There's that, too. There's that, too. [1:05] Making it happen. With us again. Chim [1:08] Polyhapatia. How are you, brother? [1:11] >> Great. [1:11] >> Yeah. Look at that smile. What do you [1:14] got going on? You got some pokers in the [1:17] fire. [1:18] >> Okay. [1:19] >> I'm not going to say in the coming [1:20] weeks, I think some news is going to [1:22] drop. That's my prediction. I don't have [1:23] any. [1:23] >> Are you loving tweet mogging that's been [1:25] going on this week? So good. So good. [1:28] >> Oh my god. [1:29] >> So good. [1:30] >> He's looks maxing by default, but he's [1:32] been mogging [1:33] >> the Gooners. [1:35] >> Yeah. So funny. What was your favorite? [1:38] >> The [laughter] one I sent you this [1:39] morning that you said. What did you say? [1:41] So funny. [1:41] >> Are you collecting your losses? Tax [1:45] harvesting. [1:45] >> What did you say? Oh my god. [1:47] >> Chimat said uh Oh my god. It was just [1:50] like Yes, I did. Yes, I did. [1:51] >> Yes, I did. Yes. Someone said something [1:53] to Chabat. He's like dropped in. Why is [1:55] everyone so mad at Jamoth? All he did [1:57] was lose billions in retail investors [1:58] money by proep [laughter] [2:00] one page spaxs. It's not like he then [2:02] told them to enjoy their capital losses [2:04] or anything. Give the man a break. [2:05] Jama's response. Yes, I did. [laughter] [2:08] >> All right. Piling on is your sultan [2:11] [snorts] of science. Everybody's [2:12] favorite. Had a great [2:15] >> some great science that he brought to [2:17] the show last week. Freedberg, how are [2:18] you doing? [2:20] >> Oh, yeah. I've been traveling this week [2:21] back at home. [2:22] >> All right. Sax is out today. He's very [2:24] busy on Capitol Hill. We'll talk about [2:26] what he's up to next week. [2:27] >> Let's go. Come on. Let's go. Let's go. [2:29] >> Let's go. Go, Jason. [2:30] >> All right. The US and Israel launched a [2:31] joint attack on Iran on Saturday. Today [2:34] is day six of Operation Epic Fury. Iran [2:38] Supreme Leader Ali Hameni [2:42] was killed within hours of the [2:44] operation. 40 senior officials have also [2:46] been killed. Death toll so far. About a [2:48] thousand people according to reports. [2:50] Tragically, six US Army Reserve soldiers [2:53] were killed following a drone strike on [2:55] a base in Kuwait. A US submarine sank an [2:58] Iranian ship off the coast of Sri Lanka. [3:00] This is the first torpedo kill since [3:02] World War II. Why were at war been a bit [3:05] of a moving target in a debate. First [3:07] explanation from Rubio. He said Israel [3:09] was going to attack and the US had no [3:12] choice but to participate. Later walked [3:14] that back. Trump made it clear this is [3:17] not a regime change effort, but we're [3:20] doing this to stop terrorism and the [3:22] development of ICBMs by obviously a [3:26] pretty crazy group of individuals and [3:29] obviously nuclear bombs which we blew up [3:31] a couple weeks ago. Trump also mentioned [3:33] the people of Iran should seize the [3:35] moment quote and take their country [3:38] back. Pegsth believe is sure boss meal [3:41] said quote this is not a so-called [3:43] regime change war but the regime sure [3:46] did change and the world is better off [3:48] for it so here's an interesting poly [3:50] market right now US forces enter Iran [3:55] this is boots on the ground by the end [3:57] of March 40% chance by the end of the [4:00] year 59% chance so the idea that we're [4:02] not going to have boots on the ground [4:04] the sharps on poly market believe we [4:06] will will the Iranian regime fall by [4:11] June 30th 39% chance according poly [4:13] market and by the end of the year 51% [4:16] chance so Emil I guess there are two [4:19] questions people really want to know [4:21] I'll leave off why we're doing this I [4:22] think President Trump has been pretty [4:24] clear now but how long is this going to [4:27] take is the one question and are we [4:29] going to have to have boots on the [4:30] ground maybe what is success here [4:35] >> um I think the the president tal talked [4:38] about this is a weeks not months kind of [4:41] operation and it's aimed at essentially [4:45] disarming the the regime uh or the [4:48] country from uh in such a way that they [4:52] can't supply Hezbollah, Hamas, um Muslim [4:56] Brotherhood, all the kind of terror [4:58] groups that get sponsored by weapons and [5:00] money from Iran, not to mention the [5:02] nuclear bit. And that's why you see from [5:06] the reporting they're going after the [5:08] depots the the you know we went after [5:11] nuclear sites before they're a [5:14] prodigious drone maker. These like huge [5:16] one-way attack drones that can go hund [5:19] you know hundreds and hundreds of miles. [5:21] Um lots of ballistic missiles uh that [5:25] are aimed at every country in the Middle [5:27] East as you've seen they've attacked [5:28] them. So [5:30] I think that's one. In terms of boots on [5:32] the ground, there's no scenario where we [5:34] have some protracted boots on the ground [5:37] Afghanistan, Iraq 2 like scenario. [5:40] >> Freeberg, your thoughts on this war. [5:43] Obviously, a lot of people voted for [5:47] Trump in order to have the peace [5:49] dividend that he was in his first term, [5:52] absolutely the peace president. And now [5:54] here we are. Eight countries have been [5:56] bombed and we've had two leaders deposed [5:59] and one of those two have been killed. [6:02] Your thoughts, Freeberg? [6:03] >> I think the president and the [6:05] administration have probably the biggest [6:07] meetings of the term coming up in China [6:10] in April. My [6:13] estimation based on the conversations [6:14] and the comments made by the president [6:17] before he came into office and since [6:19] he's been in office is that finding a [6:21] grand bargain or a deal with China is [6:23] probably one of his top priorities. And [6:25] if you think about the importance of [6:27] that, is the US going to wade into a [6:30] giant global conflict led by a US China [6:34] rift or is the US going to find some [6:36] grand bargain? I think he would probably [6:38] have a preference for the grand bargain. [6:40] And that being the case, I think you [6:42] could look at the in the context of [6:44] Maduro and the actions in Iran as [6:47] creating maximal leverage going into [6:49] those negotiations that [6:51] >> the reason for that free bird [6:53] >> 90% of the oil that comes out of Iran [6:54] goes to China [6:57] and there's been a long developing and [6:59] developed relationship between Maduro's [7:02] government and China and these are big [7:06] economic drivers or support the economic [7:09] driving in China. So creating leverage [7:12] by having significant influence or [7:15] damage or destruction to those supply [7:17] chains for China gives the United States [7:20] footing to be able to negotiate a better [7:22] deal for America. I would imagine that [7:24] the president's intention here isn't to [7:25] go and decide who should be in charge [7:28] and drive regime change and end in a [7:29] multi-year conflict with Iran. But [7:32] ultimately, if there's some transaction [7:34] with China that gets everyone out of [7:36] this and puts the US on a strong footing [7:39] where American businesses can sell into [7:40] China, which is very challenging as [7:42] everyone knows today. And there's [7:44] parity, regulatory par, economic and [7:47] trade parity between the US and China. [7:49] there's a point of view on what happens [7:50] with Taiwan and availability of key [7:52] technologies like semiconductors. I [7:54] think it could be a win-win and I think [7:57] that a deal with China could be the [7:59] crowning achievement of this [8:00] administration particularly going into [8:02] the midterms. So the timing is right and [8:04] I think that's probably a core part of [8:06] the motivation here. Chimath, your [8:08] thoughts on this action and why we're [8:12] doing it. You've heard obviously the [8:14] president has his position. We're not [8:16] doing regime change. It's a secondary [8:18] effect obviously, but we want to stop [8:20] those ICBMs and nuclear bombs from being [8:22] developed and we want to stop terrorism. [8:25] Additionally, Freeberg says, "Hey, we're [8:27] framing this great, you know, discussion [8:32] we're going to have with she and China [8:34] and oil is part of that." Where where do [8:36] you think you stand on all this? [8:38] >> I'll build on both what Emil said and [8:40] what Freeberg said. I don't think this [8:43] is about regime change and I don't think [8:45] it's about a local regional conflict. I [8:49] think if you take a step back and zoom [8:51] out, the most important thing that we [8:54] did in the last 3 months was by taking [8:58] out Maduro and by taking out the Iranian [9:04] leadership, we have created enormous [9:06] leverage as Freedberg said [9:09] with China. Now, why is that important? [9:12] Because I think all of this centers [9:14] around that geopolitical discussion. [9:17] Last night, something important [9:19] happened, which is that the official [9:20] Chinese [9:22] bureaucracy posted what their GDP [9:25] targets were, and it was shocking to [9:27] anybody reading it because what we saw [9:30] was that they guided to a range of 4.5 [9:33] to 5%. [9:35] which if you look at the historical [9:37] context of that growth is the lowest [9:40] that it has ever been in about 30 years. [9:44] So three decades. So before they entered [9:47] the WTO [9:48] and the question that one should ask [9:50] yourself is when a country that's [9:51] growing at 8 9 and 10% start to grow at [9:54] half that rate yet have double the [9:57] number of people and double the GDP what [9:59] happens? you already have incredibly [10:02] high domestic unemployment, especially [10:04] youth unemployment. [10:06] Does it become more or less chaotic? And [10:08] I think the historical artifacts of [10:09] every other country would show that it [10:12] will become more chaotic. If you have [10:13] that as a starting point, what is it in [10:16] China's best interest to do? And I think [10:18] it becomes obvious that the right thing [10:19] to do would be to invade Taiwan. Why? [10:23] because you start to create a sinkhole [10:25] that occupies your people, that occupies [10:27] resources, that can get domestic [10:29] production up and running, that can [10:31] start to generate a war machine. And you [10:33] see the economic impact of war machines [10:35] in any country during any conflict. [10:38] And if I had to guess, just to build on [10:40] what Emil said, the president saw that [10:42] and I think what they did can be [10:44] summarized in this chart which I sent to [10:46] Nick. So if your goal is to prevent war [10:50] with China, which is a massive global [10:53] conflict, which could be nuclear, which [10:56] could be cataclysmic, [10:59] how would you do it? And this chart [11:02] paints one way to do it. If you look at [11:04] the conditions inside of the Chinese [11:06] economy, [11:07] the most interesting takeaway is that [11:10] they are enormously [11:12] dependent on imported oil. So about 20% [11:17] of their economy, but it's not 20% of [11:19] their economy because it's 100% of these [11:22] critical things that create GDP, [11:24] logistics, transportation, aviation, [11:26] feed stock inputs. [11:29] And of that 19% [11:32] about a fifth of it comes exclusively [11:34] from Iran and Venezuela. And now all of [11:37] that is off the table. [11:39] So if you take that and then you see [11:41] what Steve Whitoff and Jared Kushner and [11:43] Josh Greenbond have been doing, which is [11:46] trying to get a deal done in Russia, and [11:47] you put all of these things together, [11:49] because by the way, if you add Russia [11:50] into that mix, it's about 40% of China's [11:53] oil, not only do you red dollarize, not [11:56] only do you stop the funneling of all of [12:00] these illicit oil funds to creating [12:02] chaos all around the world, but you hem [12:05] in China going into a massive moment at [12:08] the end of March, beginning of April, [12:10] where as Freeberg said really astutely, [12:13] there is the potential for a grand [12:14] bargain and I think that secures global [12:17] safety in that that is a huge thing for [12:19] America. [12:20] >> Emil, how much does this have to do with [12:22] China? [12:24] I think you know my instinct is and I'm [12:27] not speaking for the administration on [12:29] this is that's a second order benefit to [12:32] some of these things like the um you [12:35] know and you said eight eight conflicts [12:37] there have not been eight conflicts [12:39] there is like we inherited Gaza we [12:42] inherited Russia Ukraine um Venezuela [12:45] was its own operation and then you could [12:48] sort of attach to it the drug boats that [12:50] were coming out of that as like one big [12:51] operation [12:53] um and then the Houthies was just Biden [12:55] was ignoring the Houthies. They were [12:56] just shooting at our ships. So that was [12:58] like very limited in terms of like stop [13:01] shooting our ships. We need freedom of [13:03] the seas and that wasn't sort of a, you [13:06] know, so that's something any president [13:08] should be doing generally. I think um [13:11] Iran being the one, you know, material [13:14] conflict outside of Venezuela. So it's [13:16] not it's not that many. And and how long [13:18] did Venezuela last? It was one raid. [13:21] one. [13:21] >> I guess that's a really important [13:23] >> few hours. This is this is an important [13:25] note I think for you Emil to sort of uh [13:28] >> I've explained to us there's a new [13:30] [laughter] [13:31] >> approach here with regard to the these [13:35] actions which is no boots on the ground [13:38] and we seem to uh and and you of course [13:41] have better information than anybody [13:42] else does. I don't think anybody would [13:44] have known Venezuela would have gone as [13:46] well as it did and so far and listen we [13:49] got a long way to go with Iran. This has [13:51] gone very well as as well. So explain to [13:54] us what you know and what you the [13:57] president and HGV know that we don't [14:00] that makes these two operations go so [14:02] smoothly. What what is it then? There's [14:05] obviously some new technology here in in [14:06] the case of what happened in Venezuela. [14:08] >> Yeah. Bes besides the discombobulator, [14:11] what we've got is [laughter] a very [14:15] well-trained military. Like the the [14:18] global war and terror was disaster in so [14:21] many respects, but the people now who [14:24] are fighting that are generals now. And [14:27] so they've learned a lot of lessons. And [14:29] you compare that to the Chinese [14:31] military. They don't have a lot of [14:32] experience. In fact, the the [14:34] decapitation they did in the Chinese [14:36] military, the one guy they took out was [14:37] the one guy who had experience in [14:39] Vietnam. So, they don't have conflict [14:41] experience, and that matters because you [14:43] understand going in what are the things [14:46] that could go wrong. Um, and then you [14:48] you have incredible technology, space, [14:51] air, land, sea, cyber, um, all kinds of [14:56] effects that you could bring together. [14:58] And so you imagine uh a hundred guys [15:01] goes into the most fortified compound in [15:03] Venezuela where the president is, you [15:06] know, take him and his wife out safely [15:08] and are out with no KAS. [15:12] >> Incredible. [15:12] >> I mean, it's just incredible, right? So [15:14] stunning. [15:14] >> Yeah. So, and they they these things [15:17] these war games have been on the on the [15:18] shelf for a long time. Every every [15:20] scenario has been planned for years [15:23] ahead of time. Midnight hammer in Iran [15:25] was planned years ahead of time in terms [15:27] of how would you do it if you were going [15:29] to do it. [15:30] >> Um, and then you keep refreshing the [15:32] tactics, tactics, techniques, and [15:34] procedures and you're updating them. So, [15:36] we have a very sophisticated way of of [15:38] doing these things to minimize loss of [15:40] life and maximize success. [15:42] >> Can I ask a question? I don't want to I [15:43] don't want to derail this conversation, [15:44] but is the discombobulator real? Like, [15:47] what can you say about the discombobul? [15:50] >> I completely I was like obsessed with [15:52] this when I saw it on X. I was like, [15:54] "What is this thing?" I mean, I need it [15:56] in my house. Like, I just want to push a [15:58] button of all these my kids. [15:59] >> And that's just for when Helm Youth [16:01] shows up. [16:01] >> Oh my god. [16:02] >> Not meant for your kids or if they're [16:04] behaving badly. No, it's can't talk [16:06] about it. [16:06] >> Emil, do you think we would have been [16:07] able to pull off that mission as [16:09] successfully as we did 5 years ago, 10 [16:11] years ago? Has the technology improved [16:14] that quickly that this is not something [16:16] that's been possible historically? And [16:17] how does that change the the pacing and [16:20] the face of war for the next couple of [16:22] years? I'd say no. It wasn't only a [16:25] technology maturation from five years [16:28] ago. It's the uh rules of engagement. [16:31] The rules of engagement that we used to [16:34] have uh there was some I mean if you [16:36] read about them some of them were insane [16:37] like if in Afghanistan if the guy had a [16:40] small gun you had to have a small gun [16:43] and you know there was this parody in [16:45] weird ways. [16:46] >> And when you're like well but is the [16:48] objective to have like a fair fight or [16:50] an unfair fight? Um, well, if you're on [16:52] our side, you want it to be unfair. So, [16:54] the rules of engagement were relaxed to [16:57] be [16:57] >> Who writes those, Emil? Who sits in an [16:59] office and says you can't shoot back if [17:01] a combatant is shooting at you if you [17:04] aren't matched gun forgun? [17:06] >> Yeah. I mean, who writes that crazy [17:08] policies that are written into military [17:10] department? And that's why when [17:12] Secretary Hicks talks about um this kind [17:16] of thing and what was happening with him [17:18] when he was in Afghanistan, if you ever [17:19] read his book in Iraq, he's like the [17:22] rules of engagement were so punishing [17:24] that we were we were at risk all the [17:26] time cuz you had to have like a legal [17:28] understanding of what was happening in [17:30] every minute in the battlefield as [17:32] opposed to well your job is to you know [17:34] take out these guys and protect these [17:36] guys. Here's your munitions. here's like [17:38] the the the red lines and then like in [17:41] the middle of that go use your judgment [17:43] your commanding officer use your [17:45] judgment on how to win and we kind of [17:47] gotten back to that use your judgment [17:50] push responsibility field still have [17:52] your red lines but other than that the [17:53] objective is the objective it's more of [17:55] a colon Powell approach it's like go all [17:58] in have a clear objective come out use [18:01] overwhelming force and we were not doing [18:03] that for the last four years [18:05] >> and then going back to the the face of [18:07] war going forward forward. My [18:08] understanding is that there have been [18:10] more drones deployed by the United [18:12] States this past week than we've done in [18:14] the history of military activity. Is [18:17] that right? And like how does that [18:18] really change things going forward here? [18:20] >> It it changes it big. Well, so so the [18:23] Predator drone was the first big drone [18:25] program like 1015 years ago. It was this [18:27] big honken drone. Um and then if you [18:30] remember Obama would take out some of [18:32] these al-Qaeda leaders with drones on [18:34] their balcony and things like that. Uh I [18:37] think we uh President Trump took out [18:39] Sulammani with a drone near his car. [18:41] That was the beginning. And then Iraq uh [18:44] sorry the Russia uh Ukraine war happened [18:46] where it's like drone on drone. 70% of [18:48] the the casualties are for drone but [18:51] because of drones. So um drone on drone [18:55] warfare, robot on robot warfare, those [18:57] things are the future for sure, [18:59] >> right? And that's why companies like [19:00] Ander are companies like Anderoll is [19:02] because they're making unmanned systems. [19:04] A and this has been something you've [19:06] specifically been very focused on and [19:09] you tweeted today a little bit about a [19:11] competition. We'll play a little video [19:12] here. And this Lucas lowcost unmanned [19:14] combat attack systems. [19:17] >> It used to take a lot of time. It [19:19] certainly wasn't startup time to get new [19:23] product into the channel for our [19:25] military to use. Explain what program [19:27] you're running here. Feels like the [19:28] DARPA self-driving challenge all over [19:31] again and what these drones cost. I know [19:33] there's a company making them for I [19:35] think $35,000. Am I correct? [19:37] >> I mean the the small drones like I'm [19:39] holding right there and that are way [19:41] cheaper than that. The Lucas oneway [19:43] attack drone which can go 5 6 700 miles [19:47] at the speed of an airplane [19:49] >> carry a big warhead. Those are like 50 [19:52] $80,000 [19:53] depending on what kind of equipment you [19:54] put on you put on it. But the we've have [19:57] a drone dominance program and the real [19:59] and we basically have to build an [20:01] arsenal for for drones. Now are we [20:04] likely to have a territorial conflict [20:05] like Russia Ukraine with Canada and [20:07] Mexico? No. But but we do want to take [20:10] out drug drones at the border and we [20:13] want but long one-way attack drones are [20:16] important for you know any kind of major [20:19] conflict like you're seeing in Iran. but [20:21] also to protect military bases for [20:23] America 250 World Cup uh Olympics in 28 [20:28] like there's other there's a lot more [20:29] uses of drones for surveillance not just [20:32] for you know for combat [20:34] >> there you're showing drones that are [20:36] sort of human operated but how much of [20:38] this should basically be AI so that it's [20:41] just some computer vision and again back [20:43] to what you said before a model [20:46] understands the rules and the red lines [20:48] but otherwise is be on task and [20:50] accomplish your mission. How much of it [20:52] is one versus the other? [20:53] >> I mean, it I believe that a [20:55] sophisticated drone war is going to be [20:57] drone swarms controlled by AI to some [21:00] degree or another, right? To what degree [21:02] the control matters? Like for example, [21:05] drones have decoys. They could spit out, [21:07] you know, they could dazzle. They could [21:09] put out things. So, how do you [21:11] discriminate what's a drone and how to [21:13] hit it? You know, you could use AI for [21:15] that because it's learned, you know, how [21:17] to do automatic target recognition, for [21:19] example. Um, and then also could it [21:22] identify a person and that and and does [21:24] that make it safer? So, it's going after [21:26] actually someone you want to get and not [21:28] someone you don't want to get. Um, so [21:31] there's a lot of uses for AI at the [21:32] edge, if you will. Um, in the future [21:35] here, the the Ukrainians and Russians do [21:37] something called like a kill box where [21:39] they lose comms because it's jammed for [21:41] this drone and then it just starts going [21:43] in a box and looking for, you know, a [21:46] the person they're trying to get and [21:48] they're trying they're starting to use [21:49] AI to do that. [21:50] >> Wow. [21:51] >> And China has this uh ability already [21:55] probably times some magnitude. Yeah. [21:58] They have drone swarms because they can [22:01] they can force the companies that make [22:04] them, not just DJI, to interoperate. So [22:07] interoperating drones called [22:10] heterogeneous autonomy, right? You take [22:11] different kinds of drones and how they [22:13] communicate with one another and then [22:15] make sure they're not going after the [22:16] same target is like a pretty complex [22:19] thing that they're definitely working [22:20] on. [22:21] >> And let's talk uh about the fidelity of [22:25] these. Obviously, AI is a new [22:28] technology. They it can make mistakes. [22:30] Anybody who uses it on a day-to-day [22:32] basis might experience a hallucination. [22:34] How confident are you in the AI Ukraine [22:37] and Russia conflict? They obviously are [22:40] not going to be as thoughtful maybe as [22:41] we are in putting this together. They're [22:43] in a hot war right now. But we as the [22:46] United States have to be very thoughtful [22:47] about this. So, how confident are you [22:49] that this isn't going to make a mistake? [22:50] I think that's the key to a lot of this [22:53] debate. and and when will it be you know [22:55] perfect defined as much better and I [22:59] guess this dovetales with the [23:00] self-driving [23:02] you know thoughts it has to be a [23:04] magnitude better than a human so when [23:05] will this be a magnitude more accurate [23:07] than you know when we have uh make a [23:10] mistake as a as a military and we kill a [23:12] civilian [23:12] >> yeah no it's it's a good question and um [23:16] I don't know when that moment hits that [23:18] FSD moment where it get kind of gets [23:20] better certainly not there and you [23:23] wouldn't want to take huge risk with [23:25] that in like you know there there's a [23:27] gradation of when you would use that and [23:28] what kind of risk you were trying to [23:30] take or not. If you were trying to take [23:32] out a drone using AI using a like a [23:35] laser or something you'd be pretty like [23:37] okay making mistakes because you just [23:38] missed the drone you know like whatever [23:41] with a with a with the laser laser goes [23:42] off it's all over. Um if you were doing [23:45] something more sophisticated in a [23:46] population area have densely populated [23:49] area you'd take less risk. So, we're [23:51] developing procedures, tactics for each [23:53] scenario. And this is part of the debate [23:55] I had with anthropic, which is we need [23:58] AI for things like Golden Dome. [24:01] >> Chinese hypersonic missile comes up. [24:02] You've got 90 seconds before it [24:04] separates and all kinds of decoys and [24:07] you don't know where the actual payload [24:09] is and you want to get it hit it from [24:12] space. and a human can't doesn't have [24:14] the reaction time, doesn't have the may [24:17] not be able to discriminate with their [24:18] own eyes what they're going after. [24:21] That's a pretty, you know, lowrisk thing [24:22] because it's in space and you're just [24:24] trying to hit something that's trying to [24:25] hit you. So, I think in the next 10 [24:28] years, you're going to see a lot of [24:29] these applications develop AI to one [24:30] degree or another so long as we think [24:33] it's safe and it's not going to do the, [24:34] you know, make mistakes. [24:36] >> Before we get on to the anthropic [24:38] discussion, and we really appreciate you [24:40] coming here and my lord, this has been [24:41] so informative. So thank you Neil for [24:43] coming here and explaining to the [24:44] American public and to us what you're [24:46] working on. It really makes us uh I [24:48] think speak for everybody really [24:50] confident in what you're doing and it's [24:51] so great that you've you know left the [24:53] private sector to do this. [24:55] >> What I would say just very quickly Emil [24:57] is I think that not enough people [24:58] understand that the American military [25:01] has had to fight with one hand tied [25:02] behind their back. Just that little [25:05] >> insight that you just gave about [25:06] Afghanistan [25:08] to me seems so scary because the men and [25:11] women that sign up for the American [25:13] military. They're doing this to fight on [25:16] behalf of this country. They deserve a [25:18] lot more than being sent there and all [25:20] of a sudden being given this rule book [25:22] and say, "Do your best." And it's like, [25:24] "Oh, wait. You violated 19 rules trying [25:26] to protect America. Do your job." That's [25:28] insane. [25:29] >> Let's just [25:30] >> It's really insane in some cases. I I my [25:33] my belief is that's what the frustration [25:36] for those soldiers who were out there in [25:37] those wars had more than anything. There [25:40] was the broader frustration what are we [25:41] doing here and then the secondary [25:43] frustration is while I'm here why can't [25:45] I do my job? [25:47] >> Yeah. Is there is there much of a debate [25:49] internally and sorry Jake before we move [25:51] forward on this [25:52] >> regarding this idea of full autonomy in [25:56] military action? I don't want to speak [25:57] ahead to the anthropic point, but it was [26:00] something that the media seemed to say [26:02] was part of Daario's concern is that [26:05] when you press the button and hand over [26:06] complete autonomy and there's a kill [26:09] action that you're now giving to a robot [26:11] or to some autonomous system, do we then [26:13] kind of have a moral issue at hand? And [26:17] is that something that's kind of debated [26:19] or discussed? And is that the right way [26:20] to think about the framing of what goes [26:21] on? [26:22] >> I mean, we're not even close to there [26:24] yet, right? where like this the systems [26:26] are not we we wouldn't feel that a [26:29] system uh that would have sort of like [26:33] real risk for a civilian is ready to [26:36] launch yet. So we're not even debating [26:38] that. We're just trying to get basic [26:40] autonomy in drones, basic autonomy in [26:42] underwater unmanned vehicles, basic [26:45] autonomy that, you know, you've heard of [26:47] this collaborative aircraft that fly [26:49] along with a the jetcraft so that it has [26:51] more firepower, but it's still tethered [26:53] to what the jet does. So, we're [26:55] >> incredible. Yeah. [26:56] >> Yeah. So, we're just at the very [26:57] beginning of this stuff, [26:59] >> but for Golden Dome's a good example of [27:01] like, yeah, who can oppose that? Like, [27:03] it's the only way to get out a threat [27:04] like that. Um, so who could oppose if [27:08] you have a military base and you have a [27:10] bunch of soldiers sleeping that you have [27:11] a laser that can take down drones [27:13] autonomously on that? So there's it's [27:16] it's pretty scenario by scenario, but I [27:18] don't we're not having a lot of debate [27:19] because the Skynet thing is so not a [27:22] realistic thing at this moment, right? [27:25] Except if one thing I did tell the [27:27] anthropic guys, I was like, you know, or [27:29] I'd tell any company, your models are [27:32] getting stolen by the Chinese. They're [27:34] going to unguard rail them and use them [27:36] against us. And then you want our models [27:38] to be less capable against your models. [27:41] It's sort of [27:41] >> they're not going to be thoughtful. In [27:42] other [laughter] words, they're going to [27:44] go for it. And you know, if we just [27:45] benchmark this against where we were at, [27:49] you know, but 101 15 years ago, there [27:51] was the Wikileak of collateral murder, I [27:55] think they called it, where we [27:57] tragically had an Apache take out some [27:59] journalists. And this technology even [28:02] applied today probably would have [28:03] avoided that in my mind. Yeah. Like we [28:06] have enough that when you're targeting [28:09] not drones but you know people on the [28:11] ground with an Apache this would have [28:12] probably avoided that. [28:14] >> Yeah. Or or you know the Kuwaiti [28:17] aircraft hitting you know an American [28:19] aircraft making a mistake because it [28:21] doesn't have the identification. I mean [28:24] it's the same self-driving argument to a [28:26] degree. like self-driving could save [28:28] lives even though it's scary to look at [28:30] a car without a human behind the wheel [28:32] but there's tons of scenarios where it's [28:34] a way better safer option more precise [28:38] um than the alternative [28:39] >> all right before we move on to the Dario [28:41] thing and anthropic and that brew haha [28:43] there was one piece that we haven't [28:44] addressed with this interaction Freeberg [28:46] Chimath which is uh the Israeli [28:49] government and their desire to take out [28:54] this regime And us according to Tucker [28:58] Carlson and a large contingent of the [29:01] MAGA base they feel that we are captured [29:04] by this group. Does Israel have too much [29:07] influence over the United States with [29:09] regard to these actions in the Middle [29:11] East? This is you know a big debate [29:13] within the party within the Republican [29:15] party within the MAGA constituent. Hey [29:17] we number one we don't want these wars. [29:19] Number two is Israel driving this thing [29:21] to the point of Rubio's quotes that hey, [29:23] we're doing this because Israel is going [29:25] anyway. I think we should address it [29:27] here. Not that I have a personal stake [29:28] in this. I'll give my personal opinion [29:29] at the end. [29:30] >> I don't think the president is captured [29:32] by Israel in the least. I think he [29:34] decides what is in the best interest of [29:36] the United States. And if [29:40] Israel can be a part of that, then [29:43] they're a part of it. And look, let's be [29:45] clear, they're incredibly capable. And [29:48] so in something like this, to be able to [29:51] incorporate the intelligence of MSAD, [29:54] what you're seeing today in this [29:55] operation Epic Fury, we're four days in, [29:59] Iran has been 90% depleted of all of [30:01] their munitions. It looks like they're [30:02] just firing no more missiles out from [30:05] Iran to anywhere else. There's fleets of [30:08] drones and planes just waiting. [30:09] Everybody knew where the Iranians were. [30:12] It's great that when we make a decision [30:14] on something that we need to do, we can [30:16] rely on our allies. I think the opposite [30:18] question should also be asked like what [30:20] was the UK doing? Why is Spain [30:21] pontificating? Why was Europe taking the [30:23] weekend off before they could even issue [30:25] a statement? Why don't you ask that [30:27] question? [30:27] >> Yeah. No, it's equally valid question. [30:30] You know, uh Freeberg, do you want to [30:32] get in on this or no? [30:34] >> No. [30:34] >> I'm a Jew. No one's going to care what I [30:35] have to say. They're [laughter] either [30:37] going to they're either going to be like [30:38] totally like or they're going to say [30:40] this guy's a Jew. We shouldn't listen to [30:42] him. So, like let's move on. Go ahead. [30:43] >> Yeah. Yeah. Yeah. [30:44] >> Emil, any thoughts on on this? I do want [30:47] to know from Emil though like you know [30:49] is this Iron Dome working this laser in [30:52] Israel system is this operational and if [30:54] so is there any success metrics you can [30:56] share around it [30:57] >> I mean I think that the golden sorry the [31:00] golden iron beam was the first [31:01] generation of of the Israeli air air [31:04] defense thing and then they're build [31:06] building Iron Beam and and I think it's [31:08] still earlyish but yeah it works they're [31:11] they're a technologically sophisticated [31:14] country that's very small that has like [31:17] a reason to invest in these things and [31:19] they have a lot of smart people to do [31:21] them. So I think I think it's good. But [31:23] >> does it primarily work on rockets? And I [31:25] guess I just want to understand the [31:26] logical evolution of this because in the [31:28] 80s and 90s there was a lot of [31:30] conversation about space-based lasers [31:32] that could shoot ICBMs out of the sky to [31:34] avoid, you know, global nuclear war and [31:38] we could always take out every nuclear [31:39] warhead delivered on an ICBM. Is that [31:42] technology feasible? Is there a place in [31:44] the near future where we could see [31:46] basically maximal global deterrence [31:48] using these systems, either groundbased [31:50] or space-based to take out hypersonic [31:52] missiles? [31:53] >> I think I think the the harder but more [31:57] valuable problem to solve would be the [31:58] space-based way of doing it because then [32:01] um you could get at any kind of almost [32:04] any kind of threat that hits space, but [32:06] you still need a ground layer because [32:08] there's cruise missiles that could come [32:10] at you, there's drones and so on. to so [32:12] we call them multi-layers like how do [32:14] you how do you get every kind of weapon [32:16] at every layer [32:18] but you know directed energy lasers as [32:22] they get more powerful you could take on [32:24] a bigger weapon farther away right so [32:26] those uh so those technologies as that [32:30] as they improve it gets more and more [32:32] capable and I think all these defense [32:34] systems uh are going to get more and [32:36] more capable to get more and more of a [32:38] variety of weapons at farther farther [32:40] standoff which is what want you want it [32:42] to you don't want to shoot it when it's [32:43] right over Tel Aviv. You want to shoot [32:45] it, you know, when it's still over their [32:47] their land ideally. [32:48] >> Are the laser interceptors in the field [32:51] today. There's reports that they are. [32:53] >> I I think there's some I think they've [32:55] demonstrated some of them. [32:56] >> Got it. And and is this our technology [32:58] or Israel's technology? Because [32:59] President Trump said, "Hey, that's [33:02] actually our technology." Is there any [33:04] insight there? [33:04] >> We we have collaborations with Israel on [33:07] some of this stuff. They have their own. [33:09] We have our own. Um, so it's not this is [33:12] uh but they're good at tech, we're good [33:16] at tech. There's certain there's certain [33:17] ways you get part of our system and part [33:19] of their system because it's like a he [33:21] it's a quickly evolving part of of [33:24] science right now. How do you cohhere [33:26] beams of light to like get distance? How [33:29] to use high-owered microwave to like [33:31] just drop drones in their tracks? Um [33:34] there's lots of different ways to get at [33:35] some of these things. Um and and yeah, a [33:38] lot of it's ours. uh and a lot of and [33:40] some of it's theirs. [33:41] >> Yeah. And uh to to the earlier question, [33:44] you know, I I am pro-regime change if it [33:47] can be done thoughtfully and obviously [33:50] isolating a dictator, that's the best [33:52] thing you can do. We've done that [33:53] successfully with, you know, Putin, Kim [33:56] Jong-un, etc. Keep diplomacy up. But if [33:59] there is a moment in time where you [34:01] could free the people of Iran after 50 [34:04] years of being subjugated by these [34:07] lunatics and dictators, I'm all for it. [34:09] And I actually trust President Trump to [34:11] make that decision. I know this may [34:13] sound crazy. People think like I'm a [34:15] libtart or something because of the way [34:16] my besties frame me on this program, [34:18] which is completely inaccurate. I'm an [34:20] independent. [34:20] >> You are. [34:21] >> I actually I'm completely independent [34:24] and I am just based on my voting and I'm [34:26] not on either one of these sides. I am [34:28] pro President Trump and I trust his [34:30] judgment. I think he has more [34:31] information than us. I think you have [34:32] more information. I actually trust you [34:34] guys to do it thoughtfully and there [34:35] obviously was a window here. Israel can [34:37] have their own, you know, motivation. [34:39] That could be the China motivation, but [34:40] there's also spreading democracy, which [34:42] might be the least of people's concerns [34:45] here, but that's on the top of my list. [34:47] I would like to see the people of Iran [34:49] free. Just to build on your point, [34:50] Jason, the thing that Emil said before, [34:52] which I think is important as well, is [34:54] we have an enormous amount of learnings [34:56] about what happened in Iraq. Yeah. We [34:58] also have a ton of learnings between the [35:00] Iran Iraq war and a ton of learnings in [35:02] 53 when US and the British deposed Mosad [35:06] in the or at least foremented that and [35:08] put the sha and then the sha was booted [35:10] up. [35:11] >> Yeah. [35:12] >> If you take those three chapters in [35:13] Iranian history or that regional [35:15] history, there's a ton to learn. And to [35:19] your point, there is a way to affect [35:21] what we need to do without creating some [35:25] 20-year forever war. There was an [35:27] incredible tweet. I don't know if you [35:28] guys saw this. Somebody said, "So every [35:31] war doesn't have to be [35:34] >> three decades and trillions of dollars [35:35] to your friends in Virginia, Maryland, [35:37] and DC. Did you guys see that tweet?" [35:39] It's true. These things can be one and [35:42] done in and out. [35:44] And if President Trump succeeds here, I [35:46] just want to also give him some flowers [35:48] here. The people of Venezuela and the [35:51] people of Iran being free represent [35:53] about 5% of the people in the world [35:54] living under an autocracy, under a [35:57] dictator. If those both flip back to [36:00] democracies, he'll have done more for [36:02] the spread of democracy than any [36:03] president for many decades, perhaps in [36:06] our lifetime. This would be incredibly [36:08] noble. Incredibly noble, incredibly [36:10] just. And [36:11] >> would you in the human rights set [36:14] >> want him to get the Nobel then? [36:16] >> Absolutely. Give him all the Nobels. [36:18] Like literally if you can free people, [36:21] >> all of them. Give him every prize. Give [36:23] him an Oscar. [36:24] >> Physics, chemistry, [laughter] [36:27] he can have everything. [36:29] >> Physics, philosophy. [36:30] >> Jay Cal's an independent. When's the [36:32] last time you voted for a Republican [36:33] presidential candidate? Just curious. [36:36] >> Um, [36:37] yeah. [36:38] >> Say it. Say it. [36:39] >> No, no, no, no. Um, [36:41] >> Montdale. [36:42] >> No, no, I didn't. I would have voted [36:44] for, if I was of age, I would have voted [36:46] for um I wouldn't have voted for the [36:47] Bushes. I voted for the moderates. Um, [36:52] uh, uh, obviously Clinton and Obama. [36:54] >> Oh, we're playing the would have, should [36:55] have game. [36:56] >> I would have voted for Reagan in [36:58] >> I would have bought Nidia at $4. [37:00] >> Well, no. And I I didn't vote for Kamla, [37:02] so I'll leave it at that. But I voted [37:04] probably [37:05] at that. Why don't you say that you [37:06] voted for President Trump? [37:07] >> Just say you voted for President Trump. [37:09] I don't want to complicate things, [37:10] >> but you did. So, just say it. [37:12] >> I didn't vote for Kla. I'll leave it at [37:13] that. All right. [37:14] >> It's so weird that you that you'll say [37:16] you're a moderate, but you won't say [37:17] that you voted for President Trump. [37:18] >> I am supporting President Trump in about [37:20] 60 70% of what he does. Uh, let's leave [37:22] it at that. Three, two. All right. Let's [37:24] talk economic [37:26] impact of oil and insurance. Oil has [37:28] rose to $84 a barrel Wednesday [37:30] straightfor [37:33] basically a standstill at this point. [37:35] Here's the clip. You can see the traffic [37:37] slowing down and then hey, some of the [37:39] dots are even going away. That could be [37:41] uh ships were taking out. [37:45] Unless the straight opens, 3.3 million [37:47] barrels of daily production would be [37:49] lost early next week. And then there's [37:53] insurance companies. They've all [37:54] cancelled [37:56] the war risk coverage of vessels in the [37:59] Gulf effective March 5th. Super tanker [38:01] traffic dropped 94% within the first 48 [38:04] hours. Trump said the US would provide [38:06] political risk insurance for all [38:08] maritime trade through the Gulf, [38:10] especially energy. Freeberg, your [38:12] thoughts on the economic second order [38:14] effects that we're starting to [38:16] experience here and over the next four [38:18] weeks could be um you know intense and [38:21] acute. [38:21] >> The modern insurance market emerged [38:24] specifically to solve the risks of [38:26] maritime trade. So in the 17th century, [38:29] Lloyds of London, which was a coffee [38:30] shop in London, where all the maritime [38:33] traders would get together and they talk [38:34] about, hey, what's the safest route so [38:36] pirates don't get our ship and so you [38:38] don't run into weather. That's where [38:39] they would kind of have these [38:40] conversations and eventually they [38:41] started underwriting the risks of the [38:43] shipping uh routes and giving each other [38:46] guarantees. They said, "Hey, if you make [38:48] this route, great. You pay me a certain [38:50] amount and if you don't make the route, [38:52] I'll pay you the loft value." And that's [38:54] how Lloyds of London, which is the kind [38:56] of world's biggest reinsurance market, [38:58] started today. Lloyds of London has 78 [39:00] what are called syndicate members. And [39:02] these are kind of these pools of [39:03] reinsurance that underwrite big crazy [39:05] risks like maritime insurance for folks [39:08] that are moving oil tankers through the [39:10] straight of Hermuz, which the IRGC just [39:13] announced they're shutting down. When [39:15] the IRGC announced that they were [39:17] shutting down the straight of Hermuz, [39:19] there's a significant risk of all the [39:20] mines going in the straight and the [39:23] ships getting attacked and blown up. So [39:24] loss of value. The insurance premium [39:27] spiked initially from a quarter% so [39:29] 0.25% of the value of the ship to 1.25%. [39:34] So it went up by like 5x and so folks [39:36] had to pay a lot more of the value of [39:37] their ship in order to continue the [39:39] routes and get guarantees that they'll [39:41] make it through. And then all of the [39:43] markets started to shut down. So once [39:45] the conflict got heavier, everyone said, [39:47] "Let's shut this thing down." And that's [39:48] a obviously a massive risk to energy [39:50] prices globally, which drives inflation [39:52] and puts US economic security at risk. [39:55] And so this is a brilliant move. I would [39:57] say the US government stepped in with [39:59] the US International Development Finance [40:01] Corporation, which was actually funny [40:03] enough started a couple of years ago [40:05] like in 2019 or something like that as a [40:08] kind of output of one of the agencies [40:10] that provided credit from USAD. much [40:13] talked about USA ID. [laughter] [40:14] And so they're they're leveraging the [40:16] credit capacity of this old US ID agency [40:20] to go out and say to all the shipping [40:22] companies, hey, we'll give you insurance [40:25] on your routes. And the reason they need [40:27] it is the shipping companies are [40:29] levered. They take on debt to buy the [40:31] ships. And the debtors require that they [40:34] have insurance or else they're not [40:36] allowed to take the routes because the [40:37] debtors are ultimately going to be out [40:39] the money. And so the shipping companies [40:40] themselves need to have insurance. And [40:43] so this provides a market that has now [40:45] gone away. Very smart. And ultimately a [40:48] lot of people are saying this could [40:50] actually reshore or onshore maritime [40:52] insurance back to the United States and [40:54] create an entirely new insurance [40:56] industry here in the US that has [40:59] historically been served almost [41:01] exclusively by European syndicates and [41:03] European partners. And it actually [41:05] creates a big economic opportunity as [41:07] this war dies down for American [41:09] insurance companies and American brokers [41:11] to basically be the underwriters and the [41:13] guarantors of this sort of insurance and [41:15] create a new industry. So that's super [41:17] super interesting kind of side story on [41:18] what's going on here. All right, some [41:19] breaking news here folks via Bloomberg. [41:22] The Pentagon has formally notified [41:24] Anthropic that it's been deemed a supply [41:26] chain risk. This has never happened to [41:28] an American company. It has happened to [41:31] Russian companies. uh and Chinese [41:34] companies Huawei and for background the [41:37] department of war cancelled Anthropic [41:40] $200 million contract on Friday and said [41:43] they would do this. The dispute came [41:44] down to two clauses according to sources [41:46] and we have one of the principles here. [41:49] So we will hear directly from him in a [41:51] moment. Anthropic had two concerns. [41:54] Number one, fully autonomous weapons aka [41:56] murderbots. As we previously discussed, [41:59] Daario didn't feel that their technology [42:01] was reliable yet and wanted some [42:04] assurances. The second thing Anthropic [42:06] said was uh they were concerned about [42:09] mass surveillance of Americans because [42:11] they believe this technology is uniquely [42:14] powerful and uh it's can do things [42:19] beyond what a series of webcams or a [42:21] network of 7-Eleven cameras can do. [42:25] Pentagon said they wanted all lawful [42:28] use. Dario, you're welcome to come on [42:30] the program next week or any time to [42:32] give your side of the story, but this [42:33] week we have Emil. Emil, your thoughts [42:36] and explain to us what happened here and [42:38] how this broke down. [42:40] >> It's worth a little history, short [42:41] history. So, [42:44] if you remember the Biden executive [42:46] order on AI, which was this crazy [42:49] executive order that limited the amount [42:51] of compute any model company could do [42:54] and was essentially grandfathered in a [42:57] few a small number of AI companies that [42:59] they were going to designate the winners [43:01] and everyone else was out so they could [43:03] have more control on what they did. Um, [43:06] Anthropic was one of those winners. Um, [43:08] and then they were smart. act as a good [43:11] sales strategy to sell in to the most [43:14] sensitive parts of the US government [43:16] like all of our combatant commands sent [43:19] central command that's doing the Iran [43:21] fight now the Indoaccom command which is [43:24] sort of responsible for China several of [43:26] the intelligence agencies and they did [43:28] forward deployed engineers palunteer [43:30] style so they're very got very sticky um [43:33] to the workflows and all that and so I [43:36] came in and I got the AI portfolio for [43:39] department in August and I said, I just [43:42] want to see the contracts. You know, the [43:44] old lawyer in me and I looked at [43:46] contracts. I was like, holy cow, they [43:48] say you can't use them for you can't use [43:50] them to plan a a kinetic strike. You [43:53] can't use their AI model to move a [43:55] satellite. You can't There was a a 20 [43:58] page. [43:58] >> You can't do a war game scenario with [44:00] it. [44:00] >> You could do a scenario, but you can't [44:02] like let's suppose you're writing a plan [44:04] saying like if this happens, here's what [44:06] we would do. and it might involve a [44:08] kinetic strike which causes harm to a [44:10] human. So like, well, what do you think [44:12] these folks do? You know, we use the [44:14] Department of War. This is what we do. [44:16] And so I said, okay, well, I've got to [44:18] number one, have direct relationships [44:20] with these companies, not just through [44:21] Palunteer, because I want to use it more [44:23] broadly. And then number two, I need to [44:26] have the terms of service be rational [44:28] relative to our mission set. So we [44:30] started these negotiations and and took [44:33] three months and I had to sort of give [44:37] them scenarios about like this Chinese [44:39] hypersonic missile example. They're [44:40] like, "Okay, we'll give you an exception [44:42] for that." Well, how about this drone [44:44] swarm? We'll give you an exception for [44:45] that. And I was like, "The exceptions [44:47] doesn't work. I I can't predict for the [44:50] next 20 years what all the things we [44:51] might do use AI for." Um, and so so all [44:55] lawful use seems like a good thing. If [44:58] Congress wants to act, great. We have [44:59] our own internal policies like we'll [45:02] follow them. We're not knuckle draggers [45:04] here. We want we don't want to hurt [45:06] people unnecessarily. So, you know, it's [45:08] our province to decide how we fight and [45:10] win wars um so long as they're lawful. [45:14] And I think at some point it turned into [45:17] a PR game for them because they were not [45:21] going to win this intellectual battle of [45:25] well, we're going to stop you. we're [45:26] going to use our judgment because we [45:28] think Congress is behind and impose it [45:30] on the US military. And it became this [45:33] like let's find the issues that are most [45:35] inflammatory, robot weapons and mass [45:39] surveillance. I mean, like we're the [45:40] Department of War. We're not the FBI. [45:41] We're not Homeland Security. We're not [45:43] >> You're not allowed to legally spy on [45:45] Americans. [45:46] >> Yeah. You're not you're not. So it's so [45:48] so you're like and then what it came [45:51] down to on that issue just as an [45:52] anecdote is they didn't want us to bulk [45:56] collect public information on people [45:59] using their AI system and they wrote it [46:02] in a way that I was like so you're [46:03] telling me before we got to bulk collect [46:05] if someone types in you know Shimat's at [46:08] LinkedIn it's I'm using public available [46:12] information that I would be violating [46:14] your terms of service like yeah well [46:15] okay let's rewrite it. was months of [46:17] this like stuff. Um, which which was [46:21] sort of interminable and then the [46:23] trigger point was after the Maduro raid, [46:27] one of their execs called Palanteer who [46:29] we buy ourselves through and asked them [46:32] uh was our software used in that raid [46:35] which is by the way classified [46:36] information anyway. So we're trying to [46:38] get classified information and implying [46:41] that if there was used in that raid that [46:43] that might violate their terms of [46:44] service. So they wanted to enforce, this [46:47] is very important here. Yeah. [46:48] >> They wanted to enforce their terms of [46:50] service. They went behind your back to [46:52] try to collect information to then [46:56] >> maybe pull your license for their [46:58] technology. [46:59] >> But you know, and it wasn't by behind my [47:01] back. I don't want to accuse them of [47:02] that. Palunteer is the prime contractor [47:04] this sub. Um, but it raised enough alarm [47:07] with Palenteer who's got a trusted [47:09] relationship with the department to tell [47:10] me and I'm like, "Holy, what if this [47:14] software went down, some guard rail [47:16] kicked up, some refusal happened for the [47:19] next fight like this one and we left our [47:22] people at risk and I had so I went to [47:24] Secretary Hath. said this would happen. [47:26] And that was like a wo moment for the [47:28] whole leadership at the Pentagon that [47:30] we're potentially so dependent on a [47:32] software provider without another [47:34] alternative that has the right or [47:36] ability to do to not only shut it off, [47:38] maybe it's a rogue developer who could [47:40] poison the model to make it not do what [47:42] you want uh at the time or sort of trick [47:44] you because you have to trick it. I mean [47:46] all these things that we know we were [47:48] about models or hallucinate purposefully [47:50] or do or not follow instructions like [47:53] some insider threat stuff. So then that [47:56] culminated in the Tuesday kind of [47:58] dramatic meeting with Hexath and [48:00] Secretary Hexath and me and and Daario [48:03] um with the Friday deadline that that [48:06] got blown and I never thought they [48:07] really wanted to make it. M is is the [48:10] model entirely hosted by Anthropic or [48:13] just explain to us technically does this [48:15] sit in a cloud that Palunteer runs for [48:18] you guys? Um is there really technically [48:20] a way that employees at Anthropic could [48:23] kind of interfere intervene in the use [48:25] of the model? [48:26] >> Yeah. So they put their model in AWS [48:30] GovCloud [48:32] >> GovCloud. Yeah. [48:32] >> And then Palunteer serves it from there [48:35] and they refresh it. They held the [48:38] control plane for the model. So, so [48:41] yeah, [48:41] >> they can change the model weights if [48:43] they want. They can do whatever they [48:44] want. [48:45] >> Yeah. [48:45] >> The insight into this thing is [48:46] unbelievable. Not just governments, but [48:50] now if you're running a company, the [48:51] reality is that what anthropics showed, [48:55] which by the way is their right at some [48:57] level, is that they are going to have a [49:01] political perspective and a set of terms [49:04] that reflect their philosophy and that [49:07] that philosophy can change on a dime. [49:09] But what the government did was also [49:11] completely reasonable, which is we can't [49:13] rely on you if you're going to be [49:16] completely unreliable and [49:20] disallow things that are reasonable. [49:23] I'll give you a different example to [49:24] make the point. [49:26] There's a state that wants to run some [49:28] healthcare program, but they're a [49:30] prolife state. [49:33] You can't conduct abortions in that [49:34] state. Does that mean that the anthropic [49:37] engineers can decide, you know what, [49:39] we're pro-choice, so we're going to [49:41] change the access model and the [49:42] capability of that model inside of that [49:44] state. Is that allowed? Should that be [49:47] allowed? At one level, you'd say, "This [49:50] is a private company. They're allowed to [49:51] choose." But what that really means is [49:53] for the government, for all the states, [49:55] for any city, for every company, you [49:57] cannot choose to only use one of these [50:00] things because it is just a matter of [50:02] time until some person inside of one of [50:05] these companies goes on some lunatic [50:07] moral tirade and then jeopardizes your [50:10] business against something that is [50:13] nothing about law, but is everything [50:14] about subjectivity. That is the huge [50:17] thing that this thing tore open this [50:19] weekend. So if you're not figuring out [50:21] how to be multimodel and agnostic across [50:23] these models, you're taking on enormous [50:25] business risk after Friday because you [50:27] can't tolerate that these folks will do [50:29] that. It's too critical of a technology. [50:32] By the way, this is deplatforming all [50:34] over again. Remember what happened when [50:36] you didn't like what was said? Now all [50:38] of a sudden you were deplatformed. This [50:40] is that times a thousand because this is [50:42] not about posting on social media. This [50:45] is about using fundamental technology to [50:47] either advantage or disadvantage your [50:48] business. Emil. [50:50] >> Yeah. I mean, I think I described it the [50:52] other way the other day as these c the [50:56] leaders of these companies say they're [50:58] going to cause 50% white collar [51:00] unemployment. This is as powerful as a [51:02] nuclear bomb. You it's like 50,000 [51:05] geniuses in a data center. So you could [51:07] have a small country coers the world [51:09] into its whatever. So you're like, "Holy [51:10] cow." All right. So this is a general [51:12] substrate of intelligence of technology [51:15] that's applicable to a lot of things. [51:17] Very generalized. It's not like workday [51:20] like HR software we could just use a [51:22] competitor. This is going to be part of [51:24] our everyday life in so many different [51:25] ways and the controlling the like what [51:29] whether it has a moral conscience. I [51:30] mean anthropic has its own constitution. [51:33] It has its own soul. It's not the US [51:36] constitution. So you're subject to that [51:39] plus whatever whims and how that [51:40] changes. And that's a scary thought for [51:43] for Americans generally. Um and I think [51:46] that did come through a little bit [51:47] today. And in the coming years it's [51:49] going to be a bigger and bigger deal. [51:51] >> So take us through OpenAI software, [51:54] Gemini software and Gro software. Have [51:58] they push back on any use or are they [52:01] like Dell or Apple? They sell you a [52:03] computer and you have the computer and [52:05] you can use it as you will. Have any of [52:08] those given you any push back? So [52:10] Grock's all in for all awful use cases [52:12] across all classified and unclassified [52:14] networks as you'd expect because and you [52:16] know Elon's truth seeking. We want truth [52:19] and Department of War. We don't want [52:20] ideology [52:21] >> because ideology will mess with [52:23] operational decisions like you you don't [52:25] want anyone to anything to be fake or [52:27] tilted. [52:28] >> We're we're surgeing Google and [52:30] >> we have them [clears throat] we have [52:32] Google for all lawful use cases on on [52:34] classified networks and we're trying to [52:35] move them to classified networks. are [52:37] just they have to build out [52:38] infrastructure because the stuff's [52:40] complicated. [52:41] >> So they're in compliance in terms of [52:43] what you're looking for as a partner. [52:44] And then [52:45] >> I guess the last one is OpenAI and Sam [52:47] seems to be [52:48] >> just characteristically playing both [52:50] sides a bit trying [52:53] to his credit [52:56] um I called him and said I need a [52:58] solution if this thing goes sideways. I [53:01] need multiple solutions. I'd like you to [53:03] be one of them. And he's like okay well [53:05] what can I do for the country? is like, [53:07] I need to get you up running as soon as [53:09] I can. And he was he was trying to [53:12] protect anthropic to his credit. He was [53:14] like, don't do don't call him a supply [53:16] chain risk. That's bad for the industry. [53:18] Let me maybe I can negotiate terms that [53:20] they'll find acceptable. Uh but he's in [53:23] the middle because they're they compete [53:24] for the same researchers. So, a lot of [53:27] this comes down to this thousand [53:29] researchers like baseball players that [53:32] get traded between these companies. [53:33] >> Moneyball. Yeah. These are the best of [53:35] the best. It's a very moneyballish sort [53:36] of thing and there's not that many of [53:38] them and you lose [53:40] >> 20% of them and all a sudden, you know, [53:42] they launched Claude Code before you [53:44] launched Codeex or something like that [53:46] and then the numbers changed pretty [53:47] dramatically. So, he was being a real [53:49] patriot um to his credit and trying to [53:52] him, you know, help anthropic while they [53:54] were trashing him and recruiting from [53:56] his company. And I I am not biased. I [53:58] just I want all of them. I want to give [54:00] them all the same exact terms uh because [54:03] I need redundancy. I want to see if they [54:06] diverge or not or did they if they [54:08] converge maybe I only need two over time [54:10] but we don't know it's too early [54:11] >> but why why keep them in the mix? So if [54:14] there's clearly like a difference of [54:16] operations and philosophy and how they [54:19] want to run their business and there's [54:21] other models is is their model [54:23] particularly good at particular [54:25] applications that make it important to [54:28] keep it in the mix given that there are [54:30] three or four other kind of alternatives [54:31] here. [54:32] >> Anthropic you mean? [54:33] >> Yeah. Oh well because number the number [54:35] one reason we were having this [54:37] conversation at all was because they [54:38] were deeply embedded. So now I have to [54:40] unentangle them and the other companies [54:44] have not gone as heavy enterprise [54:46] enterprise sales forward deployed [54:48] engineers government business. So [54:51] they're have to catch up not on [54:52] necessarily the capability of the model [54:54] but just how do you serve the government [54:56] >> the Bible is just way ahead on that [54:58] >> right [54:59] >> but the models themselves you don't [55:00] think are uniquely advantaged or do you [55:01] have a view on that at this point? I [55:03] don't have a view on that. I I don't [55:04] think they're, you know, I mean, [55:06] certainly cloud code was was innovative [55:09] and ahead. That's true. Um, but I do I [55:12] believe in 12 months Codeex is not going [55:15] to be close. I think it will be. [55:17] >> I think you're right. There's an [55:18] asmmptoing that's happening. If you just [55:20] look at the like the confidence interval [55:22] on how overperforming or [55:24] underperformance some of the leading [55:25] models are, the error bars are [55:27] shrinking, right? the confidence [55:28] intervals like these things are all kind [55:30] of becoming the same eventually they're [55:33] all getting access to enough power [55:35] enough compute they're generating [55:38] similar results it turns out which I [55:39] think you would expect so even more [55:42] important that you have a complexion of [55:44] models the other thing Emil I don't know [55:46] if you saw this but they posted about [55:49] the revenue ramp of anthropic [55:52] and well I have a small software company [55:55] called 8090 and I asked the team. Let's [55:57] go look at our opex. I posted it because [56:00] I was so shocked at these numbers. Our [56:01] costs have more than tripled since [56:04] November of 25. Between the inference [56:07] cost that we pay AWS, which is [56:10] ginormous, between our cost with cursor, [56:13] between anthropic, we are just spending [56:15] millions. [56:17] >> So now more per unit and more more in [56:19] aggregate [56:20] >> both. But the problem is that my costs [56:23] are going up 3x every three months. My [56:25] revenues are not [laughter] [56:28] >> token use is very addicting. [56:30] >> Yeah. And by the way, because everybody [56:32] has gotten infatuated with what we call [56:33] these Ralph Wiggum loops, like just like [56:35] send the thing off and like it'll just [56:37] go figure something out. A, it never [56:39] figures anything out, and B, you just [56:41] get this ginormous bill from Cursor. So, [56:43] one of the things we had to do was just, [56:44] we had to say, guys, you got to [56:46] deprecate cursor because you're just [56:47] wrapping cloud code and charging us way [56:49] too much for these tokens. But I don't [56:51] know if you're seeing any of this thing [56:53] where like the tool usage, it's so great [56:55] to use these tools. Let's be honest, [56:56] it's super fun. It's like you feel like [56:59] a genius, [57:00] >> but then the ROI of these tools are [57:02] really important. I'm not sure that [57:03] that's as much of an issue for you or [57:05] not in in [57:08] >> it will be it. It will be [57:11] >> for sure. [57:12] >> As people find more and more use cases, [57:14] the use cases get more sophisticated. So [57:16] the next marginal thing you have it do [57:19] is likely to be harder and therefore be [57:21] more consumptive. Right. [57:23] >> Right. Right. [57:24] >> Let me just ask Emil, the important [57:26] question that I think triggered a lot of [57:27] the news this week is why then designate [57:30] them a supply chain risk? Why not just [57:33] abandon them, move on, use the other [57:35] vendors? Like why take this kind of [57:37] punitive action? [57:38] >> Yeah. So I I don't view it as punitive [57:40] and I'll tell you why. It's if their [57:44] model has this policy bias, let's call [57:46] it, based on their constitution, their [57:49] culture, their people and so on. I don't [57:52] want Loheed Martin using their model to [57:55] design weapons for me. I don't want the [57:58] people who are designing the things that [58:00] go into the the componentry to come to [58:02] me because if that po if you believe the [58:05] risk of poisoning threat yes it can [58:08] enter into any part of the defense [58:11] enterprise [58:12] >> but it's just the defense enterprise so [58:15] Boeing wants to use anthropic to build [58:18] commercial jets have at it Boeing wants [58:20] to use it to build fighter jets I can't [58:23] have that because I don't trust what the [58:25] outputs may be because they're so wedded [58:27] to their own policy preferences. [58:30] >> I guess a dovetail to that is why [58:31] couldn't this have been handled [58:34] quietly? Is this anthropic who made this [58:36] a public spat or was it the [58:39] administration that made it a public [58:41] spat or two to tango? I mean, they have [58:44] a very good sophisticated press [58:46] operation and like really good and [58:50] painting us as doing mass surveillance [58:53] where where where their issue was like [58:56] some commercial database thing that [58:58] someone else could buy. They didn't want [58:59] us to buy to use it, which I'm not even [59:01] sure we buy them except to do recruiting [59:03] for soldiers. And you know, we run [59:05] schools, hospitals, we do a lot of [59:06] things at DoD. We don't just fight wars. [59:09] and um and the the way they were able to [59:12] characterize these two things which are [59:13] genuinely scary to people but were not [59:16] the real issues. Um it was really the [59:20] you worry I worried about them shutting [59:22] off our system at a moment of need or [59:24] them messing with our system in motar [59:27] >> came to mind is if they are selling you [59:30] batteries and you need to use the [59:31] batteries or the laptops however you [59:33] need to use them lawfully okay that [59:36] should be enough for them unless they [59:38] are peacenicks and they don't want to be [59:40] involved in selling weapons which by the [59:41] way was Google's position for many years [59:43] they just didn't want to be involved in [59:44] it because to your point they want to [59:47] recruit talent that is also aligned with [59:49] that. So there's just seems to be maybe [59:52] this isn't the right partner for the [59:54] Department of War. [59:56] >> Yeah, you should if you if you don't [59:57] want your stuff to be used for [59:59] department war stuff, you shouldn't be [60:00] selling to the Department of War. [60:02] [laughter] [60:02] >> Pretty sure it's in the name. It's in [60:05] the name. [60:06] >> Well, and then also I have to say when [60:09] you know you said, "Hey, we don't know [60:10] what how we're going to use this thing." [60:12] Like immediately came to mind was like [60:14] 911. you you have to go check with them, [60:17] you know, if you find out there's [60:19] another 911 unique, you know, black swan [60:22] event that's going to occur and you have [60:24] to go clear it with them. Like you [60:26] That's [60:26] >> That was literally the comment. That was [60:28] literally the comment when I was Yeah. [60:30] So I was in a room of 20 people. So this [60:31] is not undeniable if everyone want Daria [60:34] wants to deny it. [60:36] >> And I was giving these scenarios, these [60:37] Golden Dome scenarios and so on. And [60:39] he's like, "Just call me if you need [60:41] another exception." And you know, I'm [60:42] like, but what if the balloon's going up [60:45] at that moment and it's like a decisive [60:47] action we have to take. I'm not going to [60:49] call you to do something. It's like not [60:51] rational. And [60:53] >> yes, [60:54] >> uh so it that was another holy cow [60:56] moment of like how they think about it. [60:59] >> That just means that what he wants to be [61:00] is the secretary of war. [61:02] >> That's right. [61:03] >> He wants to be the the god king there, I [61:05] guess. Yeah. [61:06] >> You can't do that. The thing that shocks [61:08] me, Emil, I don't know. you maybe you [61:10] can't say anything but guys you can [61:11] comment on this. It's clear that [61:13] Anthropic just lost all the Republicans [61:17] but I think that if they think that they [61:19] have the Democrats that's fleeting as [61:21] well because I think progressive [61:23] Democrats fundamentally just hate [61:26] Silicon Valley and technology and so [61:27] there's no way they're going to let some [61:28] god king over here that they don't [61:31] control either. And so in both ways, I [61:33] think they accidentally may have pissed [61:35] off every constituent. The longer term [61:37] fallout amongst them and progressives [61:39] will come home to roost because as the [61:41] progressives want more control and these [61:43] guys push back on them, they're just [61:44] going to fall into the same situation. [61:49] >> Yeah. I mean, it's an interesting [61:51] perspective. I think if you don't want [61:54] to be involved in war that you're right, [61:56] I think you mentioned this like three [61:57] times, Jimoth. [61:58] Don't sell bullets if you don't want to [62:00] be in but you can't call Smith and [62:02] Wesson and say can I The other thing is [62:05] what the hell was the senior management [62:07] and the board talking about over these [62:10] last few days because to me it would [62:12] have sounded insane. So then the [62:13] question is were people just so [62:15] breathless to buy this revenue curve? [62:17] What is the board doing? What is the [62:19] senior management really doing? What do [62:21] you change guys? What do you think you [62:22] would tell them if you were sitting [62:24] inside of the board of [62:25] >> Enthrop if you're an investor, you're on [62:27] the board, what do you say to Daario [62:29] when he says, "Hey, I need to dictate to [62:32] Emil and Hexth how they use my tool and [62:34] everybody else is just saying lawful use [62:36] as the standard." What's your coaching [62:38] advice? [62:39] >> Well, it's also a very unusual [62:41] circumstance because I don't think any [62:43] business in history has grown as fast as [62:45] they have in the last 90 days. So, [62:47] they've added what was it? 6 billion of [62:50] ARR. [62:51] >> Yeah. in a month or something. [62:53] >> I mean, that's absurd. Like, I mean, [62:55] absurd. It's absurd. It's a great [62:56] product. Open Claw has driven a lot of [62:58] this. [62:59] >> If you're on the board, [63:00] >> you're closing your eyes. Yeah. [63:01] >> You're shutting the up. You're just [63:03] shutting the up cuz something's [63:04] working. [63:05] >> You're actually [laughter] [63:07] >> I think he's off doing his thing and [63:08] they're going to let him do it. And I [63:09] don't think that company's worth 350 [63:11] billion anymore. [63:13] God knows what it's worth. [63:14] >> Oh. Oh, that's interesting. Where do you [63:17] If you get put a block of stock right [63:18] now, where do you put a bid in? I'll [63:20] tell you where I [63:20] >> Oh my god. I had I had this conversation [63:22] at dinner two nights ago. It's like you [63:23] have to pick between OpenAI at their [63:25] current mark, [63:28] anthropic at their current mark, or [63:29] Google. And it's either multiple from [63:33] here or net market value creation from [63:36] here because those are actually two very [63:37] different conversations. [63:38] >> Explain the difference. [63:39] >> I think the net market valuation because [63:41] Google's already worth three trillion. [63:43] So if they double, they've added three [63:45] trillion, but I think Google is the bet. [63:47] I think Google is the market value [63:49] creator bet. But I think anthropic is [63:51] the multiple bet. I think anthropic is a [63:53] trillion five market cap at the end of [63:55] the day. [63:56] Unless this blows them up, [63:58] >> you're still buying the 5x versus the 3x [64:00] kind of thing. [64:01] >> You'd buy the 5x instead of the 2x. [64:03] >> But if you get put a block of stock now, [64:04] do you buy it at the last post or do you [64:06] buy it at a discount or do you just say, [64:08] "Ah, I just buy it at the last post." [64:09] >> Anthropic is worth a lot more than 350. [64:12] That's for sure. [64:12] >> I I that it's undervalued compared to [64:14] chat. [64:15] >> They just added six billion in the last [64:17] month. And I will tell you anecdotally, [64:18] anecdotally, [64:20] >> I am everyone I talk to is on co-work. [64:23] Everyone is like gone deep on this. [64:25] Everyone's amazed and shocked and [64:27] actively using it. And everyone's saying [64:28] the same thing, which is anthropic, may [64:30] actually be fulfilling the promise of [64:32] AI. I will also say that it's only going [64:34] to take 90 days for Google to flip on a [64:38] virtual version of co-work. And once [64:40] Google like has this integrated with G [64:42] Suite and you have a virtual hosted [64:44] version of co-work sweeps the market [64:46] with this same competitor. But right now [64:48] Coowwork is such an incredible product [64:50] and everyone's saying the same thing. [64:51] It's like giving Elon giving truth to [64:53] AI. Elon said something with respect to [64:57] Grock which was that [64:59] he expects it to exceed all of these [65:01] coding models probably in the May spin [65:05] but for sure by June. [65:07] So to your point free like what like I [65:10] guess my question guys to you is like [65:12] what happens okay what do you guys do em [65:15] what do you do when all the models [65:16] asmtote let's just say by October of [65:18] this year let's just say I can guarantee [65:20] you just for the thought exercise by [65:23] October all the models are the same do [65:26] you just take a complexion of them all [65:28] and say great we're going to build some [65:30] governance layer around it and now we're [65:33] indifferent or [65:35] >> what do you do [65:35] >> I would love to be indifferent because [65:37] then I could compete on price, right? [65:39] And then then I have and then I have one [65:42] one main and one redundant or two mains [65:45] and I'd need at least two. [65:48] >> Yeah. [65:49] >> Anthropic's not going to be one of them [65:50] if they continue sort of with their [65:52] their sort of posture. So then it would [65:55] be three. And if one gets wobbly from a [65:57] policy scenario too because they all, [66:00] you know, except for Elon's is based in [66:01] San Francisco and has that that vibe to [66:04] it. So, uh, you kind of want to have two [66:06] or three at any given time. And yeah, [66:08] then I then you price compete them. I do [66:10] think Google has a long-term strategic [66:13] advantage, not because of not only [66:15] because of their consumer thing, but [66:16] because they have their own cloud. [66:18] >> So, between them, they don't have the [66:20] margin on top of the cloud that [66:22] Anthropical have to pass on. So, it's an [66:24] interesting economic uh [66:26] >> uh proposition from them. [66:30] And just to build on your point, [66:32] Freedberg, after you finish your uh [66:34] insightful comments [laughter] [66:36] here, pull this up, Nick. Almost on [66:39] quue, Freedberg. You're such an oracle. [66:41] Here is the announcement from Google. [66:43] Google Workspace is now integrated for [66:46] agents and 40 agent skills were included [66:50] today. Emilio, you've been great today. [66:52] Super honest. Daario's uh position. I'm [66:55] going to give you some fast balls here. [66:57] Daario says, "The real reason the [67:01] Pentagon and Trump admin do not like us [67:04] is that we haven't donated to Trump. [67:07] While Open AI Greg have donated a lot, [67:10] here's Claude's answer to that claim." [67:13] Here's nine companies and their [67:16] activities with the administration from [67:18] the inauguration to attending the [67:20] inauguration to the White House CEO [67:23] dinner to the Melania documentary. If [67:25] you go through and you look at these [67:27] nine companies, Microsoft, [67:30] Apple, Tim Apple, Nvidia, Amazon, they [67:33] have all participated. [67:35] There's one company that hasn't [67:37] participated and that's Anthropic. Are [67:39] is Anthropic being singled out because [67:42] they are not genulecting and because [67:45] they're not paying the cover charge. [67:46] People say this administration is pay [67:49] for play. That's the accusation he's [67:51] making. I'd say maybe there's a cover [67:53] charge. Nobody likes to pay it, but the [67:55] other companies have. What do you think [67:57] here? [67:57] >> I mean, it's literally one of the [67:58] dumbest things I've ever heard. I I [68:01] [laughter] [68:02] truly just because I'm like, I'm in the [68:05] Department of War. I need to win wars. [68:07] If you help me win wars and I don't have [68:09] to waste time transitioning you out, I'm [68:12] that makes me thrilled. Um, and it sort [68:14] of it's a criticism on me because it's [68:17] not like Trump, President Trump dipped [68:19] in and he's like, "Hey, Emil, by the [68:21] way, those guys didn't get any money. [68:22] you can't use them anymore. I mean, [68:24] obviously, it's sort of like like [68:26] invention in his own mind. It's like I [68:28] don't know if people sleep at night if [68:29] those thoughts get in there. Um and and [68:33] I was trying to work with them. Why [68:34] would I spend three months trying to [68:35] negotiate with them to get to a simple [68:37] standard if I would have just said, [68:38] "Okay, guys, you're out. Bye." So, I I [68:42] think it's just some internal [68:44] psychosis. That's the only way I can [68:46] explain that. [68:47] >> Okay. It could be on Dario that he's [68:49] antagonistic to the administration both [68:51] with respect to how he operates [68:53] commercially and it's also reflected in [68:55] the fact that he doesn't want to support [68:56] the administration. [68:57] >> I have a different theory. [68:59] >> I think that they have a massive [69:01] instance of co-work internally that [69:04] helps them come up with business [69:06] strategy. And I bet you there's like [69:09] some element of AI that says, "Yeah, you [69:11] should do it. Do it. It just makes [69:12] sense. [69:13] >> Zig where they zag and get more press." [69:15] And so now there's some some [69:17] cladbot telling them to basically tell [69:19] the department of war to pound sand. [69:21] It's gonna turn out to be the stupidest [69:23] decision. [69:24] >> Listen, if I was chairman of the board [69:25] of that company, I pull Dario aside and [69:27] I'd say, "Listen, you're obviously a [69:28] genius. We obviously have the best tool [69:30] in town. This is not a battle you can [69:33] win and it makes no sense. You're going [69:35] to come across as not being patriotic." [69:38] And Tim Cook is showing up for the [69:40] Melania premiere. Would it kill you to [69:42] support the president? Would he kill you [69:43] to show up? Look what happened when [69:46] Biden Look what happened when Biden [69:47] excluded Elon that ankled him. Show up [69:50] for the president. Show up for America [69:53] and be a patriot. You don't have to [69:55] donate, but be a patriot and show up for [69:56] the dinners. [69:57] >> That's terrible advice. Here's my [69:58] advice. [69:59] >> Okay, here's your advice. Okay. [70:00] >> Hey Dario, call a meal back right now [70:03] and say, "You know what? Sorry, weed up. [70:06] We're gonna own this and we're going to [70:08] put out a press release that says we [70:11] support our customers use of our models [70:15] to do everything and anything that's [70:17] lawful. Number one, and number two, that [70:20] our terms of service are written in [70:22] stone and that you can expect solidity [70:25] and reliability from us. And this was [70:28] just a misstep. [70:29] >> Camille, how do you respond? [70:31] >> I mean, I would say that's what I've [70:33] always wanted. I need a reliable, steady [70:35] partner that gives me something that'll [70:38] work with me on autonomous because [70:40] someday it'll be real and we're starting [70:43] to see earlier versions of that and I [70:45] need someone who's not going to wig out [70:47] in the middle and we're just at the [70:49] early stages and it's rational. But [70:52] then, [70:53] >> you know, you called President Trump in [70:54] your 5,000word essay on Friday a wannabe [70:57] dictator. [70:59] >> You're going to have to apologize to [71:00] more people than just me. Yeah, maybe [71:02] time to rewrite the position here. Uh, [71:04] let let's just say Kumbaya, everybody. [71:07] Kumbaya, we solved the problem. And look [71:08] who's on the line. Surprise guest. [71:10] Daario's here. I thought I would [71:11] surprise everybody. Nick, pull Dario up. [71:13] No, he's not here. [71:14] >> What's your view on how the industrial [71:17] supply chain for hardware components and [71:20] systems is coming along in the United [71:21] States? Because my understanding is [71:23] we're trying to reduce dependency on [71:25] Chinese manufactured components. Where [71:27] are we with respect to where we need to [71:30] get to in the US manufacturing supply [71:32] chain? [71:33] >> We are early days. Um critical minerals [71:37] you see you've seen the action around [71:39] that. Um you'll start to see so I have [71:42] the office of strategic capital which [71:44] has 200 billion in lending authority. [71:46] And what we're trying to do is is it's [71:49] like a treasuries plus 100 bips loan to [71:52] companies, show them that the department [71:54] needs their solid rocket motors, their [71:57] batteries, their fiberglass, like all [71:59] the things that we that we're heavily [72:01] dependent on for our defense industrial [72:03] base that are completely outsourced to [72:05] China and domesticate them here. Um, and [72:09] we've got uh a bunch of great people [72:11] running it. So, but but it's early days. [72:13] is going to take for the rest of the [72:15] term to get um I think we'll get [72:18] critical minerals done before the rest [72:20] of the term where where we have the [72:23] access to what we need to from US or [72:26] allied countries. Um but from batteries [72:30] is like the next problem I'm trying to [72:31] solve. For example, batteries are [72:32] totally outsourced both technologically [72:34] and from lithium to China. Um, and [72:37] there's like, you know, kind of call it [72:39] 20 critical things. If I could get to [72:42] all of them at some level, but then [72:43] it'll take a few years for them to like [72:45] build plants and do that stuff. But [72:46] there there it's it's very important. I [72:48] hope whatever administration comes next [72:50] continues it because I'm all free [72:53] market, but but we outsource so much [72:56] that um, you know, it crippled sort of [72:59] the the kind of the assembly part of [73:02] putting all these things together. Do we [73:04] have a munitions risk right now given [73:06] the conflict that we're involved in? [73:08] >> We don't have a munitions risk, but um [73:11] we do need to plus up because [73:14] the Europeans are taking a long time to [73:16] contribute like Ukraine. Russia has [73:18] consumed a lot of munitions from like [73:21] all over the world and then uh obviously [73:25] these conflicts we've been in and [73:28] um we need to have like the next [73:30] generation we're still there's still a [73:32] large degree we're fighting with 1980 [73:34] cold war weapons [73:36] >> right and not modern weapons and so we [73:38] need to plus up those things that to to [73:41] regenerate them I mean our nuclear [73:43] missiles are 50 years old some of the [73:46] planes are 40 years old so all has to be [73:48] renewed. [73:50] >> Do you think um just speak to the [73:52] venture capitalists in the audience? [73:54] >> Are we in the early stages of this kind [73:56] of defense tech boom? Is defense tech [73:59] wellunded at this point or is it kind of [74:01] too hypy and bubbly and that's not [74:03] really the issue? It's not about funding [74:05] the companies. It's about funding some [74:06] of the further upstream uh issues that [74:08] we're facing. What's what's your view on [74:10] where we are there? [74:10] >> There's more defense tech venture [74:12] capital than ever by you know 3x more [74:15] than last year. So, you know, it it's [74:18] growing. What I need to do and what the [74:20] department needs to do is have some of [74:22] these companies win big contracts quick [74:24] like whether you know and sure um uh [74:28] Seronic sure like bunch of these [74:30] companies so that more money flows in [74:33] more entrepreneurs do it and I could buy [74:35] more because genuinely I do think [74:38] warfare is going from big car carrier [74:41] ships that cost20 billion dollars and a [74:44] decade and a half to build to mass [74:47] traitable [74:48] lowcost um things and that's what these [74:52] new these new entrance can do. So we [74:55] need those to succeed so that the [74:56] flywheel goes with venture capital money [74:58] entrepreneurs capabilities [75:00] >> in that sense and what I've heard as [75:02] kind of the explainer for this is we're [75:04] moving from the old primes to the new [75:06] primes that there's going to be a small [75:08] set of big winners and then obviously [75:10] lots of seconds and and subs and [75:12] whatnot. Is that really how this [75:14] market's going to evolve? So, are we [75:15] going to end up with Andrew, Palanteer, [75:17] and maybe three or four others, and [75:18] that's where most of the value is going [75:20] to acrue from a market perspective? [75:22] >> I mean, Andrew and Palanteer want that, [75:24] and I joke with them all the time about [75:25] it, but I I want I definitely want at [75:28] least a second layer that's innovative [75:31] and trying to disrupt the first layer [75:32] all the time. I met a mom and pop like [75:35] wholly owned company that that makes [75:37] these missiles called Rams that are we [75:39] really sell and send send to Ukraine and [75:41] they do it with like 30 people and they [75:43] can do a thousand a year because they've [75:45] designed a manufacturer and it's [75:46] awesome. So I want companies like that [75:48] to continue innovating maybe and then b [75:52] buys them but the the one of the reasons [75:54] the primes are such a small number it's [75:57] not the only but it's one is they learn [75:59] how to contract with the government. [76:00] They learned how to go through the [76:01] bureaucracy and that became a [76:04] competitive advantage. I'm trying to [76:05] take that competitive advantage away. [76:07] >> That's a really important point. How do [76:09] you disassemble all that bureaucracy so [76:12] that product innovation can actually get [76:14] to you? [76:15] >> Yeah. So we we did a big So part of it [76:20] comes down to requirements reform. What [76:22] used to happen is people like oh we need [76:24] a new fighter jet. So, Army, Navy, Air [76:26] Force put in the requirements and were [76:28] you know it would we needed to be [76:31] stealthy to hold a missile to hold four [76:34] humans and you know it became this [76:37] unbuildable thing but the contractor [76:40] didn't care because they're getting paid [76:41] cost plus so like sure I'll fulfill your [76:44] requirements two years from now you're [76:46] like that was never engineered properly [76:48] it'll be another few years late and a [76:50] couple more billion dollars so we're [76:52] trying to change that to I tell you my [76:54] common operational problem. I need a [76:56] bunch of missiles that go 500 miles or [76:57] more that have this kind of blast. Come [76:59] to me with solutions, as little [77:01] requirements as possible on that side. [77:03] And on the contract piece, trying to get [77:05] to as close to commercial contracts as [77:08] possible. And this is going to take, and [77:10] this is where the startups are so good, [77:12] they'll do fixed cost pricing. They'll [77:14] do, you know, pay you don't pay me as [77:17] much if I deliver late. You pay me more [77:18] if I deliver early. [77:20] >> It's very disruptive to the existing [77:22] system. Yeah. [77:22] >> Super disruptive. But that's that's what [77:24] I'm I'm like waking up every day trying [77:27] to do. [77:28] >> So you could put out as something [77:29] saying, "Hey, the straight of horses is [77:30] super important. We need to keep it [77:32] open. We need these type of devices to [77:35] keep it open." But come to us with your [77:36] ideas and let them be creative [77:39] entrepreneurs as opposed to, you know, [77:41] just trying to goose the profits. Yeah. [77:43] It's really brilliant. [77:44] >> Yeah. [77:45] >> Emil, you also oversee DARPA. Yeah. [77:47] >> Yeah. DARPA is the father of the modern [77:49] internet and it's created a lot of [77:52] really critical technologies. Can you [77:53] talk about what's going on in there? Are [77:55] there interesting things that you think [77:57] our audience should know about that [77:58] you're trying to push forward? [78:00] >> I mean there's so it's probably my [78:02] favorite part of of my my office is like [78:05] because there that's where you it's sort [78:08] of like it's still a very honored [78:10] profession to be part of DARPA. like you [78:12] know being a being in government service [78:14] for a long time is sort of reduced in [78:16] its stature since the Manhattan project [78:19] now because now now if you're a great [78:22] ass you know uh someone who wants to do [78:23] rockets and stuff you go to SpaceX DARPA [78:26] still has the best of the best and so [78:28] the most creative ideas happen there one [78:31] of the things that they're working on [78:32] that's public is they're trying to use [78:34] biology to synthesize critical minerals [78:36] so so how do you so how can you just [78:38] pull them out of ground use biology to [78:40] do it so you don't need to do all this [78:42] crazy messy dirty refining that would [78:45] like change the game big time on our [78:47] ability to get the critical minerals we [78:49] need faster and leaprog the Chinese in [78:51] terms of tech. Um, [78:54] so they're doing a lot of that kind of [78:55] stuff. They're deep in cyber cyber [78:58] attacks are are the next huge threat [79:00] with AI, right? The what what we saw [79:02] with the creating all these agents to [79:04] attack systems that anthropic happened [79:06] to them. Um, so they're they're they're [79:09] working on that's there's not a ton I [79:11] can talk about kind of DARPA because [79:13] it's so it's so classified, but those [79:15] are a couple examples for you. [79:18] >> All right, speaking of classified, uh, [79:19] just two quick questions before we wrap [79:21] here. Are there aliens? And what are you [79:22] going to tell us? And number two, uh, in [79:25] all seriousness, I I'm curious, what [79:28] have you learned about China and where [79:30] they're at and the threat there and our [79:32] ability to counter it? like give us some [79:35] idea of where we're at as a country cuz [79:38] we hear a lot of hyperbolic stuff. [79:41] They're building this incredible mobile [79:43] small navy. They've got hypersonics. [79:45] They're just way ahead of us. You know, [79:47] we hear these things. But realistically, [79:49] are we competitive? [79:51] >> Um I fought Well, I'll answer your first [79:53] question, which I fought for the alien [79:55] portfolio. I didn't get it. [laughter] [79:58] >> Work to do more work to do. [80:01] All the guys on my team were like, [80:02] "Dude, you got to get this for us. [80:04] Please talk to the [laughter] [80:05] secretary. We want to do this." [80:08] >> But I but I I was like, you as long as I [80:10] had 100% access to everything, I would [80:12] do it because that would be it would be [80:14] amazing, right? [80:15] >> Sabers would be a game changer. [80:16] [laughter] [80:17] >> Um but on the second one, uh it is true [80:20] that Chinese have had the greatest [80:23] military buildup in world history in the [80:25] last 15 years. and we're asleep at the [80:28] wheel to some degree because we're [80:30] focused on global war and terror. So, [80:32] they've advanced without sort of us [80:34] thinking about threat. That being said, [80:37] our operational expertise and our space, [80:40] like we have some sophisticated stuff, [80:44] you know, our subs, our space layer, um [80:47] we still have the best stuff in the [80:49] world that does, you know, but but we [80:52] have to make sure that gap doesn't [80:53] narrow, [80:54] >> right? We can't be complacent. We should [80:57] sleep well at night knowing you're [80:58] there. Yeah. [80:59] >> Knowing President Trump's allocating [81:00] money towards this and he's decisive in [81:02] his actions. But we cannot be complacent [81:04] trying to I feel like this week was a [81:07] true reminder of how fortunate we are to [81:09] have the defense that we have for the [81:11] United States. When you look at what [81:13] happened in Dubai and in Doha and in Tel [81:17] Aviv and you see how people in their [81:19] residential homes are getting attacked [81:22] and bombed, you realize just how [81:24] fortunate we are to have all of the [81:25] layers of protection that we have by our [81:27] government. And I've actually come [81:29] around to this quite a lot. [81:30] >> I'm a true kind of arguably libertarian [81:33] at heart, small government, but the one [81:35] thing that I've realized is so critical [81:37] for us to have the freedom to do all the [81:39] things we want to do is defense. And so [81:42] I think it's an amazing institution, [81:44] very valuable to the United States. [81:46] Emil, thank you for what you do. [81:47] >> Yeah, thank you. Really appreciate you [81:49] coming on and being so candid and [81:51] thoughtful and insightful. This has been [81:53] a fun amazing episode. We'll see you [81:56] next time. Byebye. [81:57] >> Love you, boys. Byebye. [82:00] >> Let your winners [music] ride. [82:08] We open sourced it to the fans and [82:09] they've just gone crazy with [music] it. [82:11] >> Love you. Queen of [singing] [82:15] >> yours. [82:16] [music] [82:20] Besties are gone. [82:23] >> That is my dog taking your driveway. [82:25] [music] [82:28] >> Oh man. Myasher will eat me. [82:30] >> We should all just get a room and just [82:32] have [music] one big huge orgy cuz [82:33] they're all just useless. It's like this [82:35] like sexual tension that you just need [82:36] to release somehow. [82:41] >> Your feet. [laughter] [82:44] We need to get merch. I'm going all in. [82:47] [music] [82:53] I'm going [music] all in.