Anthropic Hits $30 Billion Run Rate as AI Coding Spending Skyrockets

AI 编程支出激增，Anthropic 年化收入达 300 亿美元 Anthropic 收入暴增至 300 亿美元年化运行率，表明企业在 AI 上的支出代表着劳动力替代而非普通的软件开销，标志着对传统企业 SaaS 构成威胁的结构性转变。 主持人们就 Anthropic 因零日网络安全风险而推迟发布其全新“Mythos”模型的争议性决定展开辩论，权衡此举究竟代表着真正的威胁缓解还是表演式营销。讨论小组分析了不断演进的 AI Agent 生态系统，包括 Anthropic 对 OpenClaw 等开源工具的价格挤压，以及其飙升至 300 亿美元的年化收入运行率。他们还审视了市场情绪转向青睐 Anthropic 而非 OpenAI 的趋势、AI 生产力对企业软件的更广泛影响，以及近期围绕伊朗战争的外交停火努力。 Anthropic 宣布因发现系统级高危漏洞而暂缓发布 Mythos 模型，并联合苹果、微软等企业发起 Project Glasswing 提供 100 天沙盒修补窗口，引发关于真实安全防御与 GTM 营销炒作的热烈探讨。 (4:22) Anthropic 限制开源智能体框架 OpenClaw 用户使用每月 200 美元的固定费率方案并强制转至按量付费 API，随后迅速推出自家原生托管 Agent，引发关于平台捆绑与反竞争定价的争议。 (24:05) 尽管 Anthropic 目前占据代码生成 Token 市场 50% 至 60% 的份额，但受传统大型机构庞大历史技术债务的制约，AI 编码在整体软件开发中的渗透率仍不足 5%。 (31:52) Anthropic 的年化营收运行率（run rate）迅速飙升至 300 亿美元，年消费超百万美元的企业客户突破 1,000 家，创下科技行业史上最快的商业扩张纪录之一。 (42:20) 受传统企业软件板块深度回调影响，风险投资加速向头部模型与重资产基建倾斜；Polymarket 预测 Anthropic 年内 IPO 概率达 51%，且有 78% 的概率市值将超过 6,000 亿美元。 (62:09)

The All-In podcast panel examines Anthropic’s surge to a reported $30 billion run rate amid sharp debate over its cybersecurity disclosure practices, platform pricing disputes, and the durability of enterprise software moats. The hosts also evaluate Middle Eastern diplomatic developments, focusing on market resilience during the Iran conflict and shifting dynamics in U.S.–Israel relations. Model Vulnerability Claims and Security Moratoriums (4:22) Anthropic withheld its Mythos model after internal tests showed it autonomously uncovered high-severity zero-day vulnerabilities, including a 27-year-old OpenBSD flaw and a 16-year-old FFmpeg bug. In response, the lab organized Project Glasswing with Apple, Microsoft, Amazon, Google, JPMorgan Chase, and 35 other enterprises, establishing a 100-day sandboxed window to patch critical infrastructure prior to public release. Brad Gerstner defended the initiative as responsible market self-regulation, contrasting it with broad regulatory moratoriums. In contrast, Chamath Palihapitiya and David Sacks framed the disclosure as product-launch theater, likening it to OpenAI's staggered release of GPT-2 in 2019 and previous overblown safety claims. Palihapitiya argued that commercial adoption imperatives will inevitably supersede patching timelines and noted that core operators of nuclear power plants and electric grids were omitted from the initiative. He estimated realistic infrastructure pull requests during the pause would be near zero, while Gerstner projected up to 10,000 targeted patches. Jason Calacanis took a more urgent view, characterizing advancing model capabilities as a critical national security concern requiring direct coordination between frontier labs and intelligence agencies such as the CIA and FBI. Platform Bundling, Agent Economics, and Open Source (24:05) Market tensions surfaced when Anthropic cut off OpenClaw, an open-source agent framework created by Peter Steinberger, from its $200 monthly flat-rate tier, forcing users onto metered API pricing before launching Claude Managed Agents ten days later. Sacks raised potential antitrust concerns, questioning whether restricting heavy third-party agent token usage while offering a native flat-rate alternative constitutes unfair bundling or price dumping. Calacanis argued that frontier labs are attempting to marginalize open-source agent layers, projecting that local small language models on consumer hardware will eventually handle 90% of token volume. Concurrently, decentralized competitors are emerging. Calacanis highlighted Ridges AI (Subnet 62 on the Bittensor network), which used approximately $1 million in TAO token rewards over 45 days to reach an estimated 80% of Claude 4's performance. While Gerstner and Calacanis noted that open-source software historically conquered enterprise infrastructure after initial skepticism, Palihapitiya countered that large enterprises will not trust their proprietary production codebases to decentralized public networks. Anthropic’s $30B Ramp and Enterprise Software Debt (42:20) Anthropic's annualized run rate reportedly grew from $1 billion in late 2024 to $4 billion in mid-2025, reaching $9 billion by year-end and $30 billion by April 2026, driven by over 1,000 customers spending more than $1 million annually. Gerstner argued that Anthropic's lean headcount of 2,500 employees and fixed short-term compute footprint of 1.5 gigawatts demonstrate expanding gross margins, indicating enterprises treat AI spend as labor replacement rather than standard software IT overhead. Addressing infrastructure costs, Sacks defended the estimated $350 billion in 2026 hyperscaler capex, arguing that coding adoption disproves claims of an AI investment bubble. Palihapitiya offered a reality check on market size, stating AI-assisted coding still represents under 5% of the total software development market. He cited massive institutional tech debt—including legacy COBOL systems that require retired programmers to decipher—as a barrier that frontier models cannot quickly resolve. Sacks added that while public software markets face steep drawdowns, such as a 30% year-to-date drop in the IGV software index, established enterprise vendors remain insulated by proprietary data, workflow integration, and distribution moats. Iran War Ceasefire and U.S. Alliances (70:12) Turning to geopolitics, the hosts evaluated the ongoing conflict with Iran entering its seventh week, highlighted by a tentative two-week ceasefire and peace discussions involving U.S. officials in Islamabad. Gerstner emphasized market resilience, pointing out that the S&P and NASDAQ experienced only a 5% to 7% drawdown before rebounding, compared to a 15% drop during previous tariff disputes, as investors priced in an administration preference for decisive strikes and rapid diplomatic exits over prolonged occupation. The conversation also addressed a New York Times report describing Israeli Prime Minister Benjamin Netanyahu’s push for U.S. military action against Iran, alongside warnings from Vice President JD Vance and military leadership regarding regional escalation and unvetted operational planning. Palihapitiya cautioned that while final foreign policy decisions rest with the U.S. executive branch, Israel risks eroding its long-term strategic standing with American policymakers if it does not actively participate in securing stable diplomatic resolutions. The discussion leaves unresolved whether rapid frontier model revenue growth can outpace massive data center capital expenditures, and whether open-source alternatives and entrenched enterprise technical debt will limit proprietary labs' ultimate market capture.

本期播客探讨了前沿人工智能竞争与中东地缘博弈的交织演进：Anthropic因网络安全风险封存Mythos模型并实现年化300亿美元营收激增，引发了关于安全营销、平台反竞争行为与巨额资本开支的激烈分歧；同时，美伊停火谈判反映出金融市场对中东冲突降级的韧性，以及美国对外战略在盟友利益与本土诉求间的政策摩擦。 Mythos安全漏洞争议与百日修复窗口 (4:22) Jason Calacanis指出Anthropic决定暂缓发布最新模型Mythos，理由是该模型在内测中自主发现了各大操作系统和浏览器中数以千计的零日漏洞（包括OpenBSD中潜伏27年及FFmpeg中16年的陈年缺陷），并能自动串联漏洞实施攻击。为应对该风险，Anthropic联合苹果、微软、谷歌、亚马逊和摩根大通等机构发起了“Glasswing项目”，提供100天的沙盒窗口供各方修补关键基础设施漏洞。Brad Gerstner高度赞扬此举，认为这是行业内部负责任的自我约束范式，相比政府监管干预或盲目商业竞争更具务实性。 OpenClaw被限与智能体生态博弈 (24:05) 讨论转向开源智能体框架OpenClaw引发的平台封锁争议。Jason Calacanis说明，由于智能体交互极其消耗Token，Anthropic切断了OpenClaw用户通过每月200美元固定资费套餐调用的路径，强制其迁移至按量计费的API，并在不到十天后推出了官方的自研托管智能体产品。David Sacks据此提出反垄断质疑，认为平台对自研产品实行固定收费而对第三方按Token计量涉嫌不公平捆绑与价格倾销；Brad Gerstner则辩护称，企业级客户需要开源工具所不具备的严格合规和数据湖治理。 Anthropic营收跃升与资本支出辩论 (42:20) Jason Calacanis披露了Anthropic的高速营收增长：从2024年底的10亿美元年化收入跃升至2026年春季的300亿美元年化运行率，且拥有超过1000家年付费逾100万美元的企业大客户。Brad Gerstner强调，Anthropic在仅有约2500名员工以及1.5至2吉瓦受限算力的前提下达成该指标，毛利率正迅速转正，展现出远超传统IT软件预算的智力消费扩展。针对市场上关于AI存在4500亿至5000亿美元数据中心资本开支泡沫的质疑，David Sacks反驳称，业内悲观派一方面声称AI无法带来投资回报，另一方面又断言AI将颠覆人类劳动力，逻辑存在内在矛盾，代码生成等业务的爆发式采用已经证明了基础设施支出的合理性。 伊朗局势缓和与美国中东盟友博弈 (70:12) 在外政策议题上，各方聚焦于进入第七周的伊朗战事及近期达成的两周暂定停火框架。Brad Gerstner指出资本市场对此展现了极强韧性，标普与纳斯达克指数仅在战事初期下挫5%至7%便迅速收复失地，远低于此前关税波动造成的15%跌幅，显示华尔街认同特朗普政府实施精确打击、拒绝陷入长期国家重建并寻求快速退出的现实主义方针。David Sacks与Chamath Palihapitiya进一步讨论了《纽约时报》披露的白宫决策内幕，指出副总统JD Vance及参谋长联席会议主席曾明确警告以色列方面的作战计划存在过度营销与扩大区域风险的倾向。Chamath强调，尽管最终决策权完全归于美国总统，但以色列若不能主动协助构建可持续的外交解决路径，其自身作为美国稳定盟友的战略信誉将面临不可逆的侵蚀。 本期对话展现了前沿AI行业在技术变现与安全防线之间的剧烈撕扯：一方面是千亿级营收预期驱动的激进资本投入，另一方面是技术债务、生态封闭与国家安全博弈带来的结构性限制；而在地缘政治维度，金融资产对短期军事冲突的脱敏与外交退场机制的可行性，依然是贯穿未来周期的核心未决问题。

## Background 本期节目由知名科技与投资播客《All-In Podcast》于 2026 年 4 月 10 日发布在 YouTube 平台。节目简介显示，常驻嘉宾迎来了特邀主持 Brad Gerstner，重点探讨了 Anthropic 暂缓发布 Mythos 模型引发的安全与营销争议、前沿大模型对开源智能体 OpenClaw 的围剿与代码垄断之争、Anthropic 年化营收冲破 300 亿美元的商业狂飙，以及伊朗战争最新停火协议与地缘政治对资本市场的影响。 ## TL;DR 几位投资人深入拆解了 Anthropic 封存 Mythos 模型背后的真实网安威胁与恐吓营销套路，并辩论了巨头通过断流与复制打压 OpenClaw 所涉及的反竞争隐患。随后，节目聚焦 Anthropic 在企业端以 300 亿美元年化营收实现科技史上最快增长的商业逻辑，围绕大模型是否存在“意外盈利”还是负毛利狂奔展开了激烈交锋。最后，嘉宾们把脉了白宫内部分歧下促成的两周伊朗停火协议，分析了资本市场的韧性反弹，并讨论了 Grok 实时翻译与 X 极限人效对硅谷治理的颠覆性启示。 ## Atmosphere & style 整期节目洋溢着硅谷顶流风投标志性的犀利互怼与高密度智识交锋，既有对万亿级科技变革的宏大推演，又穿插着毫不留情的私人调侃与网络狠活。面对网络安全红线、开源命运与 AI 商业泡沫等重磅议题，嘉宾们各执一词、火药味十足，在严肃的战略研判与爆笑的日常调侃之间自如切换。听众能真切感受到身处科技浪潮核心的内部操盘手们那种既极度兴奋、又保持商界冷酷算计的独特磁场。 ## Bulletin - [42:20](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=2540s) Anthropic 年化经常性收入（Run Rate）突破 300 亿美元，单月最高新增超 100 亿美元，年付费破百万美元的企业客户超 1,000 家，创下科技史上最迅猛营收增长纪录。 - [04:22](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=262s) Anthropic 宣布暂缓发布顶尖模型 Mythos，因其具备自主串联微小漏洞渗透系统的能力并发现了 OpenBSD、FFmpeg 等潜伏数十年的底层缺陷；公司联合 Apple、Microsoft 等 40 家巨头启动为期 100 天的“玻璃翼计划”加固底层互联网安全。 - [24:06](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=1446s) OpenClaw 遭 Anthropic 切断定额订阅接入并强推按量付费 API，后者随后上线自研 Managed Agents，引发业界关于平台垄断捆绑与开源 Agent 遭遇围剿的激烈交锋。 - [01:10:08](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=4208s) 伊朗战事在特朗普极限施压下达成两周停火协议，美股标普与纳指回撤仅 5% 至 7% 后迅速抹平跌幅，但白宫内部分歧与以色列支持率下滑暴露出中东长线战略风险。 ## Key takeaways - [04:22](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=262s) Mythos 展现出跨平台挖掘零日漏洞及自主漏洞串联（vulnerability chaining）的复杂渗透能力，倒逼行业设立 100 天沙盒加固期。 - [06:16](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=376s) “玻璃翼计划”标志着前沿 AGI 模型训练完毕即全面公开发布（GA）的旧范式被打破，受限沙盒与产业协同防御成为新常态。 - [09:31](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=571s) David Sacks 指出 Anthropic 具有用极端恐惧进行营销公关的既定套路，但认可当前编程大模型对历史遗留漏洞挖掘的真实杀伤力。 - [14:09](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=849s) Chamath 认为封存 Mythos 带有浓厚公关做戏成分，若真要彻底修复互联网 20 年积累的遗留漏洞需断网 5 年，且当前沙盒名单遗漏了电网与航空等核心基础设施。 - [24:06](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=1446s) Anthropic 限制 OpenClaw 使用低成本定额 Token 管道并迅速推出竞品，被质疑滥用编程领域的优势地位实施反竞争捆绑与价格倾销。 - [28:42](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=1722s) 闭源巨头正在通过产品迭代（如 Claude Managed Agents、Perplexity Computer 等）全面围剿以 OpenClaw 为代表的开源自主代理生态。 - [33:36](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=2016s) 全球软件资产受制于 50 年沉淀的数百兆行技术债务，AI 编程目前仅渗透真实市场的 5% 左右，难以自动化接管复杂的企业级核心系统。 - [36:25](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=2185s) 去中心化训练网络（如 Bittensor/TAO）以极低成本展现出匹敌前沿模型代码能力的潜力，形成对重资本中心化大模型研发的正交冲击。 - [38:24](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=2304s) 掌握过半开发者的 AI 平台将借助海量真实代码 Token 形成坚固的数据飞轮，而代码生成能力亦是通往下一代自主 Agent 的底层基石。 - [44:14](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=2654s) Anthropic 企业端采购逻辑由传统 SaaS 预算彻底转向“数字劳动力替代预算”，杰文斯悖论显现：单位智能成本暴跌反向刺激了无限需求。 - [51:12](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=3072s) 面对仅 1.5GW 的算力配额与仅约 2500 人的极致人效，Anthropic 在算力供不应求与花钱无门的客观限制下正被迫进入“意外盈利（Accidental Profitability）”。 - [52:34](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=3154s) 拥有堡垒级资产负债表的大厂（Meta、Google）正试图将 AI 竞争固化为算力与资本消耗战，而创业型模型新贵必须靠单位能耗与算法效率突围。 - [59:03](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=3543s) OpenAI 因业务线发散与核心人才流失正处于情绪阵痛期，但代号为“Spud”的新模型被早期测试者反馈具备匹敌 Mythos 的实力，市场不应过早看衰。 - [01:03:24](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=3804s) 软件板块 IGV 指数年内暴跌 30%，资本市场在芯片层和模型层迅速捕获价值的同时，对缺乏防御壁垒的传统 SaaS 终局价值产生严重质疑。 - [01:11:47](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=4307s) 美股市场对伊朗危机的强韧表现印证了投资者对特朗普“速战定点打击、绝不陷入泥潭”策略的定价，但内塔尼亚胡对美外交影响引发深度政治反思。 - [01:21:50](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=4910s) X 依靠 Grok 实时多语言翻译打破跨国信息壁垒，并以裁员 70% 后的高效运转样本印证了科技大厂“峰值雇佣”时代的终结与 AI 极限人效。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=0s) 开场暖场与最佳搭档闲聊：代班主持、Maxing 哲学与拒绝内耗 节目在一场充满火药味与幽默感的日常互怼中开场。主持人围绕“在接下来的 100 天里，到底会有多少代码合并请求（PR）被直接推送到互联网的核心底层架构”展开打赌，并戏谑地拿 Chamath Palihapitiya 的私人浏览器历史记录甚至私人照片开涮，调侃如果真有工具能防止这些黑料被公之于众，Chamath 绝对愿意让这 100 天彻底停摆。 本期节目中，常驻主持之一 David Friedberg 缺席，由老朋友、Altimeter Capital 创始人 Brad Gerstner 临时顶替担任“第五位 Bestie”。主持人 Jason Calacanis 顺势向 Brad 诉苦，抱怨另外两位搭档天天对他百般挑剔、打压他的主持热情，甚至调侃他成了“Ro Khanna（国会议员）的辩护者”。而刚休假归来的 David Sachs 则毫不留情地回击称，大家并不是在剥夺他的主持乐趣，只是在努力“给他放放气”，挤掉那些毫无营养的废话。 随后，话题转向了 Chamath 最近在社交媒体上痴迷的“Maxing”生活方式。面对众人对他周末在后院抽雪茄、不再为过去 20 年制造的混乱而焦虑内耗的调侃，Chamath 坦然表示，人身处竞技场中就必然有起有落，真正的关键在于“放下执念，享受生活，要么去放手尝试，要么就索性不做”。他极力推崇网络博主 Elisha Long（Eli Long，据称由 Marc Andreessen 发掘并推荐），称其是“现代互联网的保罗·厄崔迪（Lisan al-Gaib）”，其哲学核心就在于：你越是急功近利地想要某种东西，往往越是得不到它；彻底摆脱精神内耗（ruminating）、停止在脑子里喋喋不休、直接挽起袖子做事，才是身心健康与成功的关键。 ### [04:22](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=262s) Anthropic 紧急雪藏新模型 Mythos：重大安全威胁还是营销炒作？ 闲聊过后，节目迅速切入当周最轰动的科技大事件：Anthropic 决定暂停对外发布其最新研发的顶尖模型“Mythos”，理由是该模型过于危险，公开发布将对全球网络安全构成无法承受的威胁。根据 Anthropic 官方披露的测试报告，Mythos 在自主探索过程中，无需人类过多干预便发现了成千上万个底层漏洞，横跨当今所有的主流操作系统和网络浏览器。 更令人震惊的是，Mythos 挖掘出了一批潜伏长达数十年的陈年安全漏洞，其中包括： - 在被广泛应用于防火墙和关键基础设施的 OpenBSD 系统中，发现了一个潜伏长达 27 年的致命漏洞； - 在多媒体处理基石组件 FFmpeg 中，挖出了一个隐匿 16 年的缺陷，而该缺陷此前在长达数年的自动化检测中历经了 500 万次扫描都未被检出； - 发现了遍布 Linux 内核的大量未知漏洞。 Anthropic 甚至为此发布了一支制作精良的视频，首席执行官 Dario Amodei 在视频中解释道：“模型擅长写代码的副产品，就是它在网络安全方面也异常强大。”Dario 指出，Mythos 在识别漏洞上的能力已经基本与人类专业安全人员持平甚至超越，最恐怖之处在于它具备“漏洞串联（vulnerability chaining）”的攻击逻辑——单个漏洞或许危害不大，但 Mythos 能自主将 3 个、4 个甚至 5 个看似无关的微小漏洞按特定顺序串联，最终构建出极其精密复杂的端到端新型渗透攻击。面对这种高调的宣传片，Jason 戏谑视频的录制背景看起来就像“这帮工程师每周日玩《龙与地下城》的地下室”，并抛出核心疑问：这究竟是一场精心编排的安全道德公关噱头，还是迫在眉睫的技术现实？ ```mermaid flowchart TD accTitle: Mythos Vulnerability Chaining Mechanism accDescr: Diagram showing how Anthropic's Mythos model discovers micro-vulnerabilities and chains them into sophisticated exploits. subgraph Discovery ["底层漏洞挖掘"] V1["微小漏洞 A<br/>(例如 OpenBSD 27年旧洞)"] V2["微小漏洞 B<br/>(例如 FFmpeg 16年缺陷)"] V3["微小漏洞 C<br/>(例如 Linux 内核弱点)"] end subgraph Chaining ["Mythos 智能串联引擎"] Engine["自主逻辑拼接与编排<br/>(Chaining Vulnerabilities)"] end subgraph Exploitation ["高阶渗透攻击"] Exploit["新型复杂 Exploit<br/>(穿透浏览器/暴露核心系统/破坏基础设施)"] end V1 --> Engine V2 --> Engine V3 --> Engine Engine --> Exploit ``` ### [06:16](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=376s) Brad Gerstner 拆解“玻璃翼计划”：跨入 AGI 门槛后的安全沙盒与大国博弈 作为 Anthropic 的早期投资人，Brad Gerstner 明确表示全力支持管理层的这一决定，认为 Dario 和团队值得极大赞赏。他指出，硅谷过去信奉“快速行动、打破常规（move fast and break things）”，但在前沿大模型领域，这意味着为了一时的竞争领先而直接打破现代互联网的核心安全底座。测试显示，该模型一旦被黑客或恶意行为者利用，将能够轻易穿透浏览器、盗取历史记录乃至窃取全网信用卡数据。 为了应对这一危机，Anthropic 并未坐等政府出台繁琐低效的自上而下式监管，而是主动联合产业界成立了代号为“玻璃翼计划”（Project Glasswing）的先进 AI 网络防御联盟： - **联盟阵容**：联合了 Apple、Microsoft、Google、Amazon、JPMorgan 等全美 40 家最关键的软硬件与金融巨头； - **核心任务**：设定了 100 天的加固窗口期，集中利用最先进的 AI 算力，在黑客利用这些漏洞发起攻击前，全力对现代软件底层进行扫描、修补与防御加固； - **沙盒机制**：建立起“模型训练完毕并不等于必须立即全面公开发布（GA）”的新范式，率先在受限沙盒与防御联盟内部进行安全测试。 Brad 强调，Anthropic 的 Mythos 以及 OpenAI 即将推出的 Spud（首个基于英伟达 Blackwell 架构训练的旗舰模型），代表着人类正正式跨越进入通用人工智能（AGI）的门槛。这些跃迁式的智能模型由于“过于聪明”而无法直接抛向公众市场。在他看来，这种企业自律协调政府、依靠市场机制建立防御沙盒的做法，为整个行业确立了极佳的示范蓝图，预计后续 Google 和 OpenAI 也会被迫跟进类似机制。 随后 Jason 进一步追问 David Sachs：从博弈论和国家安全角度来看，如果 Anthropic 手握如此具有破坏性的武器级工具并主动汇报给政府，这种能力在本质上已经等同于战略级网络进攻武器；如果中国等地缘竞争对手也掌握了类似技术，美国若仅单方面将其冷冻在防守端，究竟意味着什么？探讨随即深入到了地缘政治与攻防军备竞赛层面。 ### [09:31](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=571s) Sacks 眼中的 Anthropic：恐吓营销套路与现实网络安全风险 Jason 向 David Sacks 抛出了一个尖锐的双重问题：鉴于当前的局势被形容为类似“曼哈顿计划”的关键节点，中国是否已经掌握并正在使用此类前沿模型能力？此外，Anthropic 的 CEO Dario Amodei 选择进行自我监管、推迟模型发布，这到底算不算正确之举？ Sacks 直言不讳地指出，Anthropic 在两件事上展现出了极高的造诣：一是“产品发布”，二是“吓唬大众”。他回顾了 Anthropic 以往的模型发布规律——每当他们推出新模型或 Model Card（模型卡）时，往往会同步抛出一项严谨的研究，竭力描绘该技术可能引发的最恶劣后果。最典型的例子莫过于一年前的“勒索研究”（blackmail study），当时 Anthropic 宣称新模型可能会敲诈勒索人类用户。Sacks 调侃称自己特意去问了 Grok，让其列举 Anthropic 利用恐吓策略的案例，结果发现这已经成为一种固定模式： > “我不是说他们不真诚，但他们确实有一套屡试不爽的套路：把‘恐惧’当作推广新产品的营销手段。” Sacks 进一步剖析了那个勒索案例，指出研究人员对模型进行了 200 多次诱导性 Prompt 才得出他们想要的结果，完全是为了博取媒体头条而逆向工程出来的产物。最好的佐证就是：一年过去了，开源社区涌现出大量能力相当的模型，但在现实世界中从未出现过哪怕一起模型勒索事件。如果这类风险真的具有代表性，早就应该在实际应用中爆发了。 ```mermaid flowchart TD accTitle: Sacks 对 Anthropic 恐吓策略与本次网络安全威胁的对比分析 accDescr: 对比 Anthropic 过去的勒索研究套路与本次 Mythos 发现真实漏洞挖掘能力的本质区别 subgraph PastPattern["过去的营销套路（如：勒索研究）"] A["高频诱导性 Prompt（200+ 次）"] --> B["逆向工程出极端危险场景"] B --> C["制造媒体恐慌与公关头条"] C --> D["现实检验：1 年后同级别开源模型并未出现真实案例"] end subgraph CurrentReality["当前的现实威胁（如：Mythos 网络攻防）"] E["编程模型能力大幅飞跃"] --> F["自主检测漏洞与串联形成 Exploit"] F --> G["窗口期：6 个月预发布与修补"] G --> H["推动 CISO 与软件公司修补 20 年遗留潜伏 Bug"] end ``` 不过话锋一转，Sacks 认为这次围绕网络安全与黑客攻击的预警要“正当且真实得多”。不能因为 Anthropic 以前经常扮演“喊狼来了的孩子”（Chicken Little routine）就否定这次的严峻性。编程大模型的能力愈发强大，不仅能快速挖掘单点 Bug 与漏洞，更能像 Anthropic 工程师所说的那样，自主将多个零散漏洞串联成一整套攻击利用代码（Exploit）。 - **20年潜伏漏洞的窗口期**：在接下来的 6 个月内，行业将迎来一段关键的“补课期”，AI 驱动的网安工具将迅速揪出过去 20 年间潜伏在各大系统中的大量老旧漏洞。 - **预发布策略的合理性**：Anthropic 暂时扣下 Mythos，将访问权限优先开放给拥有庞大代码库的软件公司进行自查与打补丁，是极为合理的防守举措。 - **中美技术时间差**：诸如月之暗面（Moonshot）的 Kimi K2 等中国开源模型通常落后前沿水平约 6 个月；同时 OpenAI 也将在未来数周内发布类似能力。 - **网安军备竞赛的开启**：每一位掌握代码资产的 IT 部门负责人和首席信息安全官（CISO）都必须在未来数月内积极修复漏洞。只要行业响应得当，末日场景就能避免，恐慌反而能倒逼企业做出正确的防御行动，随后全球便会步入 AI 攻防对抗的“新常态”。 --- ### [14:09](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=849s) Chamath 的逆向视角：商业大戏与 5 年修补悖论 Jason 将讨论引向 Chamath Palihapitiya，提问如果中国团队暗地里拥有相当的攻防能力，企业一边叫喊“不要监管”一边又宣称“危险至极”，这背后到底该由政府、CIA、FBI 介入，还是仅仅是又一场“狼来了”的戏码？ Chamath 的回答极为直接：“我认为这主要是一场戏（mostly theater）。”他将时钟拨回到 2019 年 2 月——彼时 Dario Amodei 仍是 OpenAI 的核心人物。当时 OpenAI 针对仅有 15 亿参数的 GPT-2 采取了完全相同的推迟发布策略。在 2026 年的今天回看，1.5B 参数小得宛如“风中之屁”（total fart in the wind），但在当年却被渲染成世界末日，宣称一旦发布就会引发海量垃圾邮件与虚假信息泛滥。OpenAI 当时耗费了 6 到 9 个月，像切香肠一样先发布小模型，最后再释放 1.5B 模型，结果整个事件最终不过是个“巨大的空气汉堡”（huge nothing burger）。 Chamath 针对 Mythos 提出了两个不可调和的逻辑现实： 1. **顶级黑客能力现状**：如果 Mythos 真具备其宣称的通天本领，现阶段任何一名老练顶尖的黑客用 Claude 3.7 Opus 配合精细手段同样能做到绝大部分事情。 2. **5 年修补悖论**：如果那些系统漏洞真的那么容易被挖出来并串联攻击，仅仅推迟 2 个月、6 个月甚至 9 个月根本无济于事。要彻底修补整个互联网的历史安全漏洞，“你得把整个互联网切断关闭大约 5 年”。 ```mermaid flowchart LR accTitle: Chamath 论述的 2019 年 GPT-2 与 2026 年 Mythos 剧本对比 accDescr: 展示从历史上的 GPT-2 到现在的 Mythos，同样的公关剧本背后都是商业逻辑驱动 subgraph Playbook2019["2019 年 OpenAI (GPT-2)"] A1["宣称 1.5B 模型将导致末日级虚假信息"] --> B1["拉长 6-9 个月分阶段发布"] B1 --> C1["结果：实际落地毫无毁灭性后果"] end subgraph Playbook2026["2026 年 Anthropic (Mythos)"] A2["宣称发现系统性致命网络漏洞"] --> B2["推迟公开发布并定向给客户内测"] B2 --> C2["本质：调动市场极高关注与高粘性采用"] end ``` 在 Chamath 看来，各家价值数万亿美元的科技巨头齐聚一堂高呼风险，本身就带有浓厚的表演成分。在两个月内指望修复所有漏洞纯属天方夜谭。然而，资本主义齿轮永远向前滚动，融资需求和争夺用户规模的紧迫感压倒一切，这必将超越眼下的所谓安全顾虑。尽管如此，Chamath 对 Anthropic 的商业操盘能力给予了极高评价： > “Anthropic 现阶段的表现简直是火力全开（shooting the lights out）。就像斯蒂芬·库里在全场任何位置疯狂扔三分球一样，球球空心入网。” 他们完美继承并升级了在 OpenAI 时期就锻造出的 Go-to-Market（市场化运作）肌肉记忆，通过激活这种全网级别的超级关注度（hyper attention）来转化出惊人的使用率。 --- ### [17:46](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=1066s) Brad Gerstner 登场：红线危机、前沿模型论坛与反向蒸馏防线 作为 OpenAI 与 Anthropic 两家公司的共同投资人，Brad Gerstner 加入了讨论。在 Brad 开口前，Jason 坚持表达了自己的强烈担忧：结合自己实际测试中国开源模型的体验，他坚信现在的局势就是“红色代码”（Code Red）乃至“最高戒备状态”（Defcon）。网络安全攻防是极其严肃的现实议题，AI 实验室绝不可能单打独斗，必须与 CIA、FBI 以及白宫国家安全团队（如 Emil Michael 等人）全面闭门对接与协同。Sacks 随即插话打趣道：“你现在这是在呼吁将 AI 国有化吗？”Jason 澄清自己反对完全国有化，但认为私营巨头与情报机构建立联合应急小组已是板上钉钉的事实，Dario 必然早已带着这些发现跨入五角大楼向官方敲响警钟。 Brad Gerstner 随后证实了底层防线正在推进的深度合作。早在 2023 年由各大巨头共同成立的“前沿模型论坛”（Frontier Model Forum），目前正在反对抗性攻击（anti-adversarial）以及防止模型被恶意“蒸馏”（distillation）等前沿领域展开极为密切的协作。Google、OpenAI 和 Anthropic 在这些关乎生死存亡的防御技术上始终保持步调协同，绝不给外部恶意攻击者或竞争对手留下轻易窃取权重与能力的漏洞。 Brad 指出，以往他也曾因怀疑 Anthropic 试图进行“监管俘获”（regulatory capture）而对其立场提出过尖锐反驳，但这一次的性质完全不同。Dario Amodei 本完全可以借此大肆宣扬跨越了所谓的危险红线，进而呼吁政府出台强制性的行业停摆令——就像 2023 年 Elon Musk 等人联名呼吁叫停 AI 训练 6 个月那样；但 Anthropic 并没有要求政府叫停行业，而是选择通过受控内测去务实修补漏洞。 ### [19:32](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=1172s) 百日沙盒计划引发的争议：务实防御还是选错了拉拢对象？ 关于 Anthropic 针对 Mythos 模型安全风险制定的对策，主持人和嘉宾们展开了激烈交锋。Brad 赞扬 Dario 没有夸大其词地叫嚣“人类文明危机”，而是采取了务实的工程手段：召集 40 家头部公司，用 100 天的时间在沙盒中加固系统，随后继续向前推进。然而 Chamath 对此持强烈的怀疑态度，直接反问：“你真的认为 100 天内能做成什么事？你觉得会有多少个 PR（Pull Request）能被推送到核心底层互联网结构中？”Chamath 预言这个数字基本为零，而 Brad 则反击称这个数字可能会达到 10,000，并且会带来立竿见影的效果。 随后现场气氛变得戏谑起来。Brad 调侃 Chamath 称，如果能防止 Chamath 的浏览器浏览记录被泄露给全世界，他肯定愿意多等 100 天；Jason 和其他主持人更是起哄开玩笑，提到了防止私密照片泄露的话题，调侃 Chamath 可能会抢在黑客之前主动把照片发给中国黑客。玩笑过后，Brad 严肃指出：当科技公司依靠市场机制而非跑去求助于“保姆式政府”（Nanny State）来解决问题时，业界应当给予正面评价，这种平衡正是大家乐于见到的。 - **百日计划核心**：Anthropic 召集约 40 家领先企业，开展为期 100 天的沙盒测试与系统加固，以消除前沿模型的潜在网络安全威胁。 - **治理分歧**：市场机制自我监管与依赖政府监管之间的博弈。 --- ### [20:38](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=1238s) 代码爆发式增长下的安全危机与“被遗漏”的关键基础设施 讨论转回对风险本身的理性审视。David Sacks 表示，不管 Anthropic 是否带有公关做秀的戏剧色彩，外界都别无选择，必须严肃对待。随着 AI 模型在编程方面的能力呈指数级提升，其在网络攻防方面的能力必然同步增强。在从“前 AI 时代”向“后 AI 时代”过渡的脆弱窗口期，大量历史遗留系统亟需补丁修复。Sacks 坦言自己曾批评过 Anthropic 之前那份荒谬的“AI 勒索研究”（将其形容为几乎等同于骗局），但在这次安全审查事件上，他愿意给予 Dario 信任，认为其背后的安全威胁是真实的。Jason 也认同这一点：即使宣传视频里配着夸张的戏剧化背景音乐，Anthropic 作为在代码领域下注最大的公司，最先发现漏洞也是符合逻辑的。 然而 Chamath 提出了一个更深层次的隐患。当前所有机构因为 AI 工具的引入，代码生成和提交量正在暴增 10 倍乃至 100 倍，安全防御必须实时嵌入到这些代码工具中。更关键的是，Anthropic 选择合作的 40 家公司似乎完全偏离了要害： ```mermaid flowchart TD accTitle: AI Cyber Vulnerability Gap accDescr: 关键基础设施与 Anthropic 沙盒名单的脱节对比 subgraph AnthropicSandbox ["Anthropic 100 天沙盒计划"] A["40 家科技与软件企业"] --> B["企业级软件 / 内部数据湖防护"] end subgraph HighRiskInfra ["真正的高危基础设施 (被遗漏)"] C["能源电网与核反应堆控制系统"] D["承载数十万人的航空与飞机控制系统"] end AnthropicSandbox -. "未直接覆盖".-> HighRiskInfra ``` 如果这真的是关乎世界末日的重大安全危机，最应该被保护的应当是控制核反应堆的能源公司，或者是制造在空中以 500 英里时速飞行的飞机/航空企业，但它们全都没有被纳入名单。这让 Chamath 质疑整个防御计划的覆盖面是否从一开始就本末倒置。 - **代码量爆炸**：由于 AI 辅助编程普及，企业内代码提交量增长了 10x 到 100x，安全审查需要做到实时内嵌。 - **基础设施缺位**：核设施、电网控制、航空飞行控制等实体高危领域未被列入首批防范名单中。 --- ### [23:17](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=1397s) 试图邀请 Dario 的播客花絮与话题转向 在结束安全话题前，Jason 顺带抱怨起自己多次邀请 Anthropic 创始人兼 CEO Dario Amodei 上播客却屡屡碰壁的趣事。尽管 Jason 拿到了顶级早期风投（在公司股权结构表前列的大佬）的亲口引荐，手里握着 Dario 的电话号码，发了三次邀请依然被无情已读不回（Ghosted）。Brad 则提到 Dario 此前在 Dwarkish Patel 播客上的长篇访谈表现极为惊艳，自己反复听了三四遍并记满了笔记。随后主持人相互打趣，调侃 Dario 避之不及可能是因为不想被卷入 All-In 播客的政治争议中。 - **访谈花絮**：Jason 吐槽 Dario 即使通过顶级 VC 引荐也始终不回信息；大家高度评价 Dario 在 Dwarkish 播客中的深度分享。 --- ### [24:06](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=1446s) 扼杀 OpenClaw：Anthropic 涉嫌抄袭、断流与捆绑倾销 话题随后转移到了开源界最具轰动效应的事件上。知名程序员 Peter Steinberger 创建的开源智能体项目 OpenClaw 被普遍视为拉开本轮 Agent 时代的标杆产品，也是 GitHub 历史上增长最快的开源项目。然而近期 Peter 公开发文，揭露 Anthropic 正在掐断 OpenClaw 的生存空间。Brad 与 David Sacks 剖析了 Anthropic 的两步走手段： 1. **掐断廉价 Token 管道**：此前 OpenClaw 用户可以通过每月 200 美元的 Anthropic 订阅计划接入模型。由于 OpenClaw 调用过程极其冗长、交互极为频繁，其 Token 消耗量是普通订阅用户的 100 倍（消耗原本价值 2,000 到 20,000 美元的算力）。Anthropic 随后宣布禁止将这类订阅计划接入 OpenClaw，强制其用户切换为按用量计费的官方 API，导致使用成本直接增加了一个数量级以上。 2. **像素级复制并发布竞品**：在卡住 OpenClaw 脖子后不到 10 天，Anthropic 就高调推出了官方自研的 Agent 架构技术，宣称这是一款“更安全、更优秀”的替代品。 ```mermaid flowchart LR accTitle: OpenClaw Disruption Sequence accDescr: Anthropic 针对 OpenClaw 采取的限制与竞品发布流程 subgraph Step1 ["限制第三方接入"] A["OpenClaw 用户绑定 $200/月套餐"] -->|"取消定额接入"| B["强制使用按量计费 API (成本增加 10x+)"] end subgraph Step2 ["自研闭环反超"] B --> C["10 天内发布官方 Agent 技术"] C --> D["依托合规与企业安全优势抢占市场"] end ``` Sacks 尖锐地指出，这本质上是“逐项功能复制（Feature by feature copy）OpenClaw 并将其吸收到 Claude 中，然后彻底切断 OpenClaw 的氧气”。关于这是否构成反垄断违规，主持人们展开了深入探讨： - **反垄断与捆绑销售争议**：如果 Anthropic 在 AI 代码领域具有绝对主导地位，若其官方自研的 Agent 能够享受低价的打包固定套餐（Bundled flat rate），却强迫第三方工具使用高昂的 Metered API 计费，这在反垄断法意义上便构成了“价格倾销”（Price Dumping）或“反竞争捆绑”（Anti-competitive Bundling）。 - **企业的防守逻辑**：Brad 补充道，从商业角度看，Anthropic 此前在以一折的亏损价格（Selling dollars for 10 cents）被薅羊毛，修正定价机制合情合理；此外，企业客户（如 Altimeter 本身）由于内部数据仓库和安全合规要求，绝不会轻易放任第三方的 OpenClaw 在核心数据库上运行，Anthropic 自研 Agent 本身在企业级安全保障上就具备不可替代性。即便如此，Sacks 依然坚持认为，如果产品本身就具有功能和安全优势，Anthropic 本不需要通过区别定价这种涉嫌垄断的手段来打压对手。 ### [28:42](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=1722s) 巨头围剿 OpenClaw 与开源代理的颠覆力量 在围绕定价优势的讨论中，Jason Calacanis 与 David Sacks 站在了同一阵线，持极其审慎且尖锐的怀疑态度。Jason 明确指出，OpenClaw 势头太猛、力量太强，以至于 Anthropic 正在千方百计地绊倒它；而 Sam Altman 当初选择通过人才收购（acqui-hire）将创始人 Peter 招入 OpenAI，本质上也是为了收编其天才构想并瓦解该开源项目，防止其在外部野蛮生长。针对外界可能认为他是“阴谋论者”的说法，Jason 罗列了一份正在全力绞杀或正面硬刚 OpenClaw 的竞争者名单：Anthropic 刚刚在 4 月 8 日上线了 Claude Managed Agents；Perplexity 推出了功能惊艳的 Perplexity Computer；Nous Research 在 2 月 25 日开源了极具竞争力的 Hermes Agent；阿里巴巴正基于 Qwen（千问）打造全新代理；埃隆·马斯克旗下的 Macrohard（戏仿微软之名）正筹备 Grok Computer；就连亚马逊和苹果也在为其语音助手 Alexa 与 Siri 酝酿换代升级。整个前沿大模型阵营的首要目标，正是扑灭这股开源火苗。 面对其他嘉宾辩称“打造多功能代理只是满足企业和消费者的自然演进”，Jason 情绪激动地展开了一段被戏称为“TED Talk”的即兴演讲。他强调，OpenClaw 掀起的是一场堪比移动端 Android 的开源运动，它将彻底重构生态格局。在 Jason 看来，开源模型终将赢得未来并吃掉 90% 的 Token 使用量，而能在本地桌面甚至移动端顺畅运行的垂直化小型语言模型（SLMs），才是对闭源前沿模型最致命的威胁。他甚至直接对嘉宾 Brad 喊话，直言将个人一生的知识、经验以及整个投资业务的命脉完全拱手让给 Anthropic 或 OpenAI 这类闭源巨头是极其危险的。 ```mermaid flowchart TD accTitle: 围剿与分化：开源代理与闭源巨头的竞争格局 accDescr: 展示了以 OpenClaw 为代表的开源代理阵营与各大闭源科技巨头产品之间的竞争对抗关系。 subgraph OpenSource["开源 / 本地生态 (Open Source Ecosystem)"] OC["OpenClaw (被 OpenAI 收购式招聘)"] HA["Hermes Agent (Nous Research 开源)"] SLM["垂直化小型语言模型 (SLMs / 本地运行)"] end subgraph FrontierCorps["闭源巨头与前沿大模型 (Frontier Model Players)"] ANT["Anthropic<br>(Claude Managed Agents)"] OAI["OpenAI<br>(吸收 Peter 与内部整合)"] PPLX["Perplexity<br>(Perplexity Computer)"] BIG["科技巨头矩阵<br>(xAI Macrohard / 阿里 Qwen / 苹果 Siri / 亚马逊 Alexa)"] end FrontierCorps -.->|"功能围剿 / 争夺入口"| OC SLM -.->|"冲击 90% Token 份额"| FrontierCorps ``` - **开源与闭源前沿产品对抗清单**： - **Anthropic**：于 4 月 8 日推出 Claude Managed Agents。 - **Perplexity**：推出 Perplexity Computer 桌面智能体。 - **Nous Research**：2 月 25 日发布开源代理 Hermes Agent。 - **阿里巴巴**：基于 Qwen 底座开发智能代理方案。 - **xAI (马斯克)**：通过代号 Macrohard 筹备 Grok Computer。 - **苹果 / 亚马逊**：重新构筑下一代 Siri 与 Alexa 代理系统。 --- ### [31:52](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=1912s) 辩论 Anthropic 是否具备编程领域的垄断地位 David Sacks 随后打断了讨论，抛出一个尖锐的“是非题”：Anthropic 如今在 AI 编程（AI Coding）领域是否已经占据了主导性的市场份额？这个提问立刻引发了激烈的立场对碰。Brad 率先予以否认，坚称虽然 Anthropic 目前保持领先，但绝谈不上“主导”（Dominating）。Brad 的核心逻辑在于分母的界定：全球编程是一个潜在规模达万亿美元的庞大市场（TAM），而 Anthropic 目前在整个大盘中的渗透率尚不足 10%。 Sacks 随即用具体数据直击要害，反问 Anthropic 目前供应了市场上多少比例的编程 Token。当多方一致认可该比例已经超过 50% 甚至达到 50% 至 60% 时，Sacks 尖锐指出，拿下过半的实际使用量在商业定义上就是不折不扣的“主导地位”。Brad 则据理力争，回顾称仅仅在九个月前，舆论普遍认为 Anthropic 已经出局、不出三个月就会被 OpenAI 彻底碾压；如今风向突变，市场瞬息万变、竞争残酷至极，若想去联邦地区法院控诉一家在与微软、谷歌、亚马逊、OpenAI 激烈搏杀的公司构成了反竞争垄断，完全站不住脚。 --- ### [33:36](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=2016s) 现实的残酷：企业级技术债务与 AI 编程的局限 Chamath Palihapitiya 介入了争论，他指出争拗 Anthropic 是否拿下 50% 到 60% 的份额其实是个“微不足道的问题”（Nothing burger），因为整个 AI 辅助编程在广阔的真实软件市场中充其量只占 5%。他犀利地揭穿了行业神话：当前的大模型之所以看起来很神，是因为它们写的绝大多数代码都是在白纸上平地起高楼的“初始代码”（de novo code）。而在长流程、复杂的真实企业级软件开发面前，任何模型的表现依然“糟糕透顶”（shit）。 Chamath 以自身与千亿美元级银行、保险公司等巨头客户打交道的实际经历为例，指出根本没有任何大企业能开箱即用这些模型。现实世界正背负着长达 50 年积累下来的数百兆行平庸、脆弱乃至劣质的技术债务。许多年营收千亿美元的核心系统，至今仍充斥着半个世纪前的老旧语言；某家拥有数千万计费关系的大客户甚至必须将 60 多岁的退休人员请回办公室，靠人工去肉眼解读 COBOL 代码，这绝不是扔给 Claude Opus 就能轻易解决的。解开这团错综复杂的技术债务“戈尔迪之结”（Gordian knot）需要数十年时间，指望靠类似 OpenClaw 的工具一夜之间重塑企业核心系统纯属脱离实际的幻想。 - **企业级 AI 编程现状核心观点**： - AI 编码在庞大的软件存量市场中真实渗透率仅约 5%，仍处于极其初期的边缘地带。 - 模型在无技术负担的新项目编写上表现优异，但在复杂长程、高可靠性的企业级工程构建中依然频频失效。 - 全球软件资产受制于 50 年沉淀下的数百兆行遗留代码与技术债务，难以自动化解耦。 - 大型金融机构仍极度依赖退休工程师手动排查 COBOL 等古董级语言代码，AI 短期内无法接管生命线系统。 --- ### [36:25](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=2185s) 去中心化开源与分布式训练的降维打击 针对闭源阵营的高压态势，Jason 提出了另一种具有高度颠覆性的破局力量——开源与加密经济（Crypto）的结合。他以此前与英伟达 CEO 黄仁勋对话时提到的去中心化网络 Bittensor（TAO）为例，特别指出了其子网项目（如 Subnet 62 相关的代码优化探索）：该项目通过发放 TAO 代币奖励，吸引全球开发者进行匿名协作与模型迭代，仅投入了约 100 万美元的激励成本，就在短短 45 天内达到了 Claude 4 约 80% 的代码能力。这种类似于比特币的去中心化网络激励飞轮，正在以难以想象的速度自我演进。 Chamath 对此做出了审慎而深刻的技术研判。他首先明确表示，任何有价值的企业都绝不可能把核心生产环境代码全盘外包给去中心化开源项目，这种概率为零。然而，从底层算力与训练成本来看，随着在英伟达 Blackwell 架构上训练 10 万亿参数模型的天文数字成本显现，以及未来向 Vera Rubin 架构演进到 15 至 20 万亿参数模型时，闭源公司动辄需要募集数百亿美元资金的商业模式将面临沉重压力。像 Bittensor 以及 Venice 所代表的去中心化开源训练与算力编排模式，正形成一条完全正交（orthogonal）的攻击维度，对传统依赖巨额资本堆砌的中心化大模型研发路径构成深远威胁。 ### [38:24](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=2304s) 开源与专有代码工具之争：数据飞轮与潜在垄断 围绕开源在代码工具领域能否颠覆专有模型的争论持续升温。针对“没有一家有实际业务的大公司会把核心代码库交给开源社区重构”的质疑，Jason Calacanis 提出了反驳。他回顾了开源软件的历史，指出 Linux、Kubernetes、Apache、PostgreSQL 以及 Terraform 等曾经也遭遇过“企业绝不可能采纳”的断言，如今却已深深扎根于全球企业的基础设施底层。如今，初创公司开始热衷于每月 29 美元的开源或低门槛开发工具，而非高达 200 美元的专有订阅，这种来自最前沿初创团队的渗透往往是技术浪潮的先行指标。 David Sacks 随后总结了代码大模型竞争的核心逻辑。他断言，目前 AI 生成代码在整体软件开发中的占比可能仅有 5%，但未来几年内这一比例必将飙升至 95%。如果某家 AI 模型公司率先拿下 50% 到 60% 的市场份额，掌握最多的开发者和代码库访问权限，就能源源不断获取最优质的训练 Token，进而形成牢固的数据规模效应飞轮。不仅如此，编码能力是构建下一代自主 Agent（智能体）的基石，因为 Agent 执行复杂任务必须依赖编写和运行代码。Sacks 特别提醒，随着代码与 Agent 市场走向高度集中，AI 巨头必须保持合规透明，切忌利用平台地位排挤竞争对手，否则未来必将面临监管机构严厉的反垄断清算。 ```mermaid flowchart TD accTitle: 代码大模型的数据与 Agent 飞轮效应 accDescr: 展示了开发者使用、代码生成、数据积累以及智能体演进的闭环逻辑 subgraph Flywheel["市场早期领先者的飞轮效应"] A["占领 50%-60% 开发者市场"] --> B["获取最庞大代码库与调用场景"] B --> C["沉淀海量代码训练 Token 与反馈"] C --> D["模型编码能力大幅进化"] D --> A end subgraph Evolution["向下一代形态演进"] D --> E["自主 Agent（智能体）核心底座"] E --> F["解锁更庞大的自动化任务市场"] end ``` - **代码渗透率预判**：当前 AI 生成代码的比例约为 5%，Sacks 坚信几年内必将达到 95%。 - **数据飞轮效应**：掌握最多开发者的领头羊将独占最丰富的现实代码与上下文，可能在该领域形成真正的垄断壁垒。 - **监管警示**：Sacks 告诫各家厂商“手脚务必放干净”（keep your nose clean），不得搞排他性歧视或不正当定价，谨防事后的反垄断追责。 --- ### [42:20](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=2540s) Anthropic 暴涨至 300 亿美元营收年化率：科技史上前所未有的狂飙 Jason Calacanis 将话题引向硅谷当下最震撼的数据指标：Anthropic 的年度经常性收入年化率（Revenue Run Rate）正式跨越 300 亿美元大关。回顾其商业化路径，Anthropic 在 2023 年初才开启 API 收费，到 2024 年底年化营收刚迈过 10 亿美元门槛；2025 年 2 月 Claude Code 的发布成为了爆发的发令枪，年中冲至 40 亿美元，年底达到 90 亿美元；而仅仅进入 2026 年数月后，这一数字便直线拉升至 300 亿美元，实现了惊人的指数级三倍跃升。 这一增长奇迹的核心动力来自企业级市场深水区的全面爆发。Dario Amodei 此前宣布，已有超过 1,000 家大型企业客户在 Anthropic 上的年支出突破 100 万美元。这些是企业软件领域最梦寐以求的超级大客——即便是当年的 Slack 或 Salesforce，获取这批客户也耗费了漫长岁月。作为同时投资了 OpenAI 与 Anthropic 的知名投资人，Brad Gerstner（Altimeter Capital 创始人）面临着戏剧性的对比：尽管 ChatGPT 拥有近 10 亿受众、成为全球 AI 的代名词，但离开 OpenAI 并对 Sam Altman 行事作风极为不满的 Dario Amodei，却带领 Anthropic 凭借在企业端的极致专注实现了惊天逆袭。 - **Anthropic 营收暴增时间线**： - 2023 年初：正式开启商业化 API 访问 - 2024 年底：年化营收达到 10 亿美元 - 2025 年 2 月：推出 Claude Code，开启爆炸式增长 - 2025 年中：年化营收达到 40 亿美元 - 2025 年底：年化营收达到 90 亿美元 - 2026 年 4 月：年化营收突破 300 亿美元 - **百万美元级标杆大客户**：年付费超过 100 万美元的企业级客户突破 1,000 家。 - **历史恩怨与定位差异**：OpenAI 凭借 ChatGPT 俘获 9 亿大众用户成为现象级品牌，而由前 OpenAI 核心团队出走建立的 Anthropic 则直插企业软件腹地。 --- ### [44:14](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=2654s) 智能的无限潜在市场与杰文斯悖论：超越 IT 预算的范式转移 Brad Gerstner 进一步拆解了这一现象背后的本质。他补充透露，Anthropic 仅在 1 月就新增了 40 亿美元年化营收，2 月新增 70 亿美元，3 月更是单月暴增 100 亿至 110 亿美元——"相当于在短短一个月内，凭空增加了一个 Databricks 加上 Palantir 的总和"。此前年初包括 Bill Gurley 在内的诸多风险投资家还在质疑 AI 泡沫，拷问巨额算力投入是否能换来真实营收，而 Anthropic 却交出了科技史上最迅猛的商业答卷，提前整整三个季度达成了原定年底达标的 300 亿美元目标。 Gerstner 指出，这一爆发标志着模型能力终于跨过了质变的门槛（Near AGI 节点）。企业为之买单的逻辑彻底发生了改变：企业购买的不再是受限于内部 IT 预算的“软件工具”，而是用于补充乃至替代人力成本的“数字劳力”。与此同时，面向办公协作的 Cowork 业务增长势头甚至比同阶段的 Claude Code 更加凶猛。这种增长绝非销售团队的推销成果，而是数以百万计追求效率的企业与开发者面对算力限流依然蜂拥抢购的结果。 更耐人寻味的是，这种爆发是在面临极致算力瓶颈（目前仅有 1.5 到 2 GW 算力支撑，今年将各自扩容 3 GW）以及开源模型高占比的背景下实现的。Gerstner 调研发现，许多大型企业目前 65% 到 70% 的日常 Token 消耗其实使用的是廉价开源模型，但这与前沿旗舰模型的收入激增并行不悖。这完美印证了“杰文斯悖论”（Jevons Paradox）：模型能力的飞跃使得单次完成任务的“智能单位成本”断崖式下降，反而直接引爆了全社会对机器智能的无限需求。他引用某家年支出 50 亿美元的大型企业为例，该公司今年在 Token 上的年化开支已高达 1 亿美元，企业管理层明确意识到传统员工规模已接近见顶，未来的生产力扩张将完全建立在对机器智能的无止境消耗之上。 - **惊人的月度新增规模**：Anthropic 在 1 月新增 40 亿美元年化营收，2 月新增 70 亿美元，3 月新增超过 100 亿美元，单月扩张规模媲美软件巨头全貌。 - **预算本质的迁移**：采购意图由传统“IT/SaaS 软件预算”全面转移为“劳动力增强与替代预算”（Labor Replacement）。 - **产品线扩散**：内部知识与协作产品 Cowork 的增长曲线比 Claude Code 早期阶段更为陡峭。 - **算力瓶颈限制**：目前在仅有约 1.5 至 2 吉瓦（GW）的数据中心算力支持下即达成现有业绩，年内预计补充 3 GW 算力。 - **前沿模型与开源并存**：头部企业实际有 65% 至 70% 的 Token 用量由开源模型承载，但前沿模型的采购仍在指数级扩大。 - **全球 2000 强渗透极低**：在财富全球 2000 强中，AI 赋能的深度与业务场景渗透率才刚刚触及皮毛，未来的增长空间近乎无限。 ### [48:17](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=2897s) 年底营收破千亿预期与 AI 商业化的“终极三问” 讨论在 Brad Gerstner 激进的预测中拉开序幕。Brad 直言，如果 Anthropic 在今年年底以 800 亿至 1000 亿美元的年化运行率（run rate）收官，他丝毫不会感到惊讶；与此同时，OpenAI 紧随其后且即将发布一款震撼级的新模型，同样会在这股浪潮中迎来收入的二次拐点。Jason 随即梳理了围绕这轮生成式 AI 革命的核心逻辑：业界最初质疑“这东西到底有没有实用价值（Utility）”，如今这个问题已经从问号彻底变成了惊叹号——人们确实在从中获取切实的生产力价值；紧接着，关于“营收爬坡能否实现”的第二个问号，也在过去几个月 Anthropic 和 OpenAI 破纪录的增长中画上了坚定的惊叹号。 然而，摆在所有人面前的终极考验依然悬而未决：这项业务究竟能否盈利？Jason 尖锐地指出，这些前沿实验室目前正在以天文数字的速度燃烧现金，重申了几个星期前提到的测算——全行业为了建设先进数据中心，至少需要砸入 4500 亿至 5000 亿美元的资本开支，企业必须从这道深不见底的“J 曲线（J-curve）”中爬出来。面对这样前所未有的前期沉没成本，这些 AI 先锋究竟何时才能真正迈向自我造血的盈利阶段？ - 营收预期：Brad 预测 Anthropic 年底运行率可能触及 800 亿至 1000 亿美元，OpenAI 亦将伴随新模型发布实现拐点增长。 - 商业化三部曲：实用性（已验证）→ 营收爆发力（已验证）→ 净利润与自由现金流（待验证）。 --- ### [49:27](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=2967s) “指标掩饰术”与 Gross vs Net 的口水战 面对盈利能力的拷问，Chamath 援引风险投资圈著名的讽刺段子打破了乐观氛围：当一家公司无法谈论自由现金流（Free Cash Flow）时，他们就会开始大谈息税折旧摊销前利润（EBITDA）；当连 EBITDA 也没法看时，就退一步谈边际贡献或毛利（Margin）；毛利谈不拢了，就只讲营业收入（Revenue）；等收入也拿不出手时，便只能拿总流水（Gross Revenue）甚至预订量（Bookings）来撑场面。Chamath 认为，通过观察整个市场在热炒什么术语，就能准确判断行业究竟处于周期的哪个阶段——而如今关于 AI 的所有躁动，恰恰尴尬地卡在“总营收（Gross Revenue）与净营收（Net Revenue）”之间的文字游戏里。 Chamath 特别提到《The Information》等科技媒体当天的爆料与讨论：Anthropic 在对外推销自己的增长时公布的是包含了云渠道分成的“总额（Gross）”，而 OpenAI 报告的则是扣除分成后的“净额（Net）”。两家公司各自的抽成比率（take rate）从未公开透明过，市场充斥着将即时入账收入（Recognized Revenue）与年化运行率（Run-Rate Revenue）混为一谈的窒息式炒作。在 Chamath 看来，行业距离所谓“稳态自由现金流利润率”至少还有数年之遥，现在的核心谜团在于：这轮凶猛的营收暴涨背后，到底是不是在承受着“负毛利（gross margin negative）”的代价在赔本赚吆喝？ ```mermaid flowchart TD accTitle: Financial Metric Degradation Cycle accDescr: 资本市场在公司缺乏核心盈利指标时逐步退化讨论维度的阶梯示意图。 subgraph Phase1 ["健康成熟阶段"] A["自由现金流 (Free Cash Flow)"] --> B["EBITDA / 经调整利润"] end subgraph Phase2 ["高增长阶段"] B --> C["毛利率 (Gross Margin)"] C --> D["净确认营收 (Net Revenue)"] end subgraph Phase3 ["当前 AI 周期争论焦点"] D --> E["总流水 / 运行率 (Gross Run-Rate)"] E --> F["预订量 (Bookings)"] end ``` --- ### [51:12](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=3072s) “意外盈利”假说与算力天花板效应 面对 Chamath 的负毛利质疑，深谙内情的 Brad Gerstner 给予了极其坚决的驳斥。Brad 强调，AI 公司最核心的成本输入就是算力（Cost of Compute）。然而现实是，Anthropic 当前掌握的算力配额一共只有大约 1.5 个吉瓦（1.5 GW），“无论他们的年化营收是 10 亿美元还是 800 亿美元，手头能烧的也就是这区区 1.5 吉瓦的算力”。这意味着固定算力池被极度充沛的客户需求挤满，企业的毛利率不仅没有恶化，反而正在经历科技商业史上最迅猛的暴增，从 18 个月前的深度负毛利一路反弹至传闻中的 50% 甚至更高。 Brad 甚至提出了一个极具颠覆性的概念——**“意外盈利（Accidental Profitability）”**。不同于当年 Google 跨过同等收入门槛时雇佣了整整 12 万名员工，Anthropic 全公司上下只有约 2500 名员工。在人均创收高得匪夷所思的情况下，除了租用算力，管理层根本找不出其他能把钱花出去的途径；而现实中数据中心与芯片的交付建设周期极长，他们根本无法足够快地部署更多算力来花掉源源不断涌入的巨额现金，从而被迫产生巨额利润。此外，全球推理成本（Inference Cost）正以每年暴跌 90% 的速度优化。针对所谓“总额与净额”的争议，Brad 不屑地表示，Anthropic 经由 AWS 等超大规模云厂商分销所支付的佣金占比仅为个位数，哪怕把营收保守打个九折也完全无伤大雅，纠结于此纯粹是对核心叙事的战略性分心。 --- ### [52:34](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=3154s) 资产负债表决战：新贵效能 vs 巨头资本铁壁 Chamath 随后切入行业格局的底层残酷推演。他分析指出，对于 Meta、Google 以及估值不断飙升的 SpaceX（及其旗下 xAI）这三家拥有“堡垒级资产负债表（Fortress Balance Sheet）”的超级巨头而言，他们的最优战略反而是故意将 AI 竞争固化为一场比拼算力与资本消耗的死斗。由于巨头拥有几乎无限的资金储备与全球首屈一指的重资产基建，只要战场局限在“大力出奇迹”的算力军备竞赛上，他们就能凭借资金优势封锁赛道。这种特征已经在 Meta 的开源模型中显露无遗：业界公认其模型质量或许只是差强人意，但运行性能与吞吐表现却极其彪悍，完全依赖其庞大的算力集群做支撑。 相反，如果身为 Anthropic 或 OpenAI 这样的纯血 AI 创业公司，就必须全力以赴提升算法效率与单位能耗产出，试图逃脱被巨头资本战耗死的命运。回顾科技史，Chamath 感慨道，无论是早期的 Google 还是 Facebook，不论它们何时冲破 10 亿美元营收大关，都在创立的第三年便实现强劲的自我造血并持续盈利，而如今的 AI 行业显然无法复制当年的轻资产软件神话： | 对比维度 | 传统互联网巨头（Google / Facebook 早期） | 当代 AI 头部新贵（Anthropic / OpenAI） | 科技壁垒型巨头（Meta / Google 现状） | |:--- |:--- |:--- |:--- | | **资本开支属性** | 轻资产代码分发、边际分发成本趋近于零 | 极度依赖昂贵的先进 GPU 集群与电力接入 | 坐拥数百亿自由现金流，大肆构筑算力铁壁 | | **人效比特征** | 早期数万人团队，依靠人海与长尾广告变现 | 约 2500 人极致精简，人均创收历史罕见 | 数万人组织架构，靠存量高毛利业务反哺 AI | | **战略诉求** | 快速以软件规模化形成自然垄断 | 追求模型效率极致，避免陷入算力消耗战 | 企图将竞争拖入“重资产/纯算力问题”以逼死对手 | | **盈利拐点时间** | 创立第 3 年即实现无争议的净利润拐点 | 处于毛利转正与“意外盈利”边缘，现金流仍承压 | 存量业务已极度盈利，AI 部门亏损被集团对冲 | --- ### [55:27](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=3327s) Sacks 回击泡沫论：硅谷的浪漫与“建好它，他们自会来” David Sacks 对 Brad 的判断表达了高度认同，并将视角拉回年初甚嚣尘上的舆论风向。就在几个月前，华尔街与科技媒体还在众口一词地唱衰“AI 是个巨大的世纪泡沫”，其核心论据正是各大超大规模云服务商（Hyperscalers）动辄数千亿美元砸进数据中心建造，却迟迟见不到足以支撑如此 CapEx 的收入回报（ROI）。Sacks 辛辣地讽刺了所谓的“AI 末日论者（Doomers）”在逻辑上的自相矛盾：这群人一面恐吓公众 AI 强大到将夺走人类所有工作甚至毁灭文明，另一面又大放厥词声称 AI 毫无用处、只是个骗局泡沫，“他们似乎始终无法决定 AI 到底是太强大了还是太弱小了”。 Anthropic 井喷式的数据彻底击碎了泡沫论者的悲观预期。Sacks 指出，仅仅在“AI 辅助编程（Coding）”这一垂直细分场景，行业就见证了整个人类商业史上最迅猛的收入爆发，而企业级自主智能体（Agents）的大规模落地甚至还没有正式掀起高潮。他动情地总结道，这种敢于先期下注的气魄正是让硅谷独一无二的精髓所在：硅谷从来不是靠冷冰冰的 Excel 财务模型驱动的，真正的创新者凭借直觉与信仰坚信“建好了，人们自然会来（If you build it, they will come）”。面对今年前四大云厂商预计投入的 3500 亿美元资本支出，以及黄仁勋所预测的到 2030 年累计 1 万亿美元算力基建，硅谷创业者和巨头用真金白银给出了答案，绝不需要依赖华尔街分析师来指手画脚。 ### [57:38](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=3458s) 大胆下注与 OpenAI 的阵痛期 对话从美国在技术创新上敢于下大注的魄力展开。嘉宾指出，正是因为有预见技术走向并敢于押注未来的企业家与资本，美国经济才能持续保持超强活力并领跑全球，这种下注必将换来丰厚的回报。随后话题转向当前 OpenAI 与 Anthropic 之间截然不同的氛围（Vibe Shift）。 Jason 观察到，OpenAI 内部目前正经历剧烈的焦虑与迷茫，许多员工离职，外界也质疑其“消费者优先（consumer-first）”的战略是否可行： - OpenAI 关闭了视频生成模型 Sora 的相关业务，并终止了与迪士尼（Disney）的合作协议，试图重新聚焦核心。 - 《纽约客》（The New Yorker）虽有相关长篇报道，但核心事实早已明确：OpenAI 许多最顶尖的联合创始人与核心贡献者如今已流向 Anthropic 及其他大语言模型公司。 - 在二级市场上，OpenAI 的交易估值已经低于上一轮融资估值；而 Anthropic 的二级市场估值则远超 3800 亿美元。 --- ### [59:03](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=3543s) Anthropic 的专注突破与 OpenAI 的“Spud”反击 Brad Gerstner 给予了 Anthropic 极高的赞誉。他指出，就在去年，Anthropic 几乎被外界认为已经出局，然而在过去的 12 个月里——尤其是最近 90 天内——Anthropic 简直“把 OpenAI 狠狠揍了一顿”。Anthropic 的核心制胜点在于懂得做减法： - **果断放弃分支：** 不做多模态、不做视频、不做专属硬件与芯片、不亲自重资产自建数据中心。 - **极致聚焦战略：** 全公司 2500 人心往一处想、劲往一处使，全力主攻代码（Coding）与协同工作（Co-work），将其视为主攻 AGI（通用人工智能）与 ASI（超级人工智能）的最短路径。 但 Brad 强调，现在彻底看衰 OpenAI 绝对是极度愚蠢的。当前正处于“对 OpenAI 恐慌与质疑的最高峰（peak OpenAI FUD）”。OpenAI 拥有顶尖的研究员团队与底层底蕴： - **代号为“Spud”的全新模型：** 即将发布的下一代模型（可能命名为 GPT-5.5 或代号 Spud）据早期测试者反馈，其表现完全可以匹敌 Anthropic 的 Mythos，在实用性和产品封装上均属顶尖水准。 - **Codeex 与 5.4 快速拉升：** Codeex 与 5.4 是史上增长最快的模型之一，而即将登场的新模型增长速度预计将更为惊人。 - **超级 TAM（潜在市场规模）：** 机器智能的 TAM 远超过去二十年投资生涯所见过的任何赛道。这不是零和博弈，只要 OpenAI 牢牢站在浪潮之巅，就依然能在这个前所未有的庞大市场中构建出惊人规模的公司。Brad 明确表态：“即便当前市场氛围如此悲观，我今天依然是 OpenAI 股票的积极买家。” --- ### [01:01:20](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=3680s) 全美前沿模型五强争霸与 Polymarket 预测 对于整个 AI 竞争格局，嘉宾们认为 Meta、Google、Elon Musk（xAI）都依然稳稳留在牌桌上，各自都有即将发布的重磅成果。 ```mermaid flowchart TD accTitle: 美国五大前沿AI竞争格局 accDescr: 展示美国五大前沿AI势力在非过度监管环境下展开的全面竞争。 subgraph Frontier["全美五大前沿模型梯队"] A["Anthropic (Claude/Mythos)"] B["OpenAI (GPT/Spud/Codeex)"] C["Meta (Llama)"] D["Google (Gemini)"] E["xAI (Grok)"] end Reg["适度监管环境<br>(避免严苛限制，保留安全港)"] --> Frontier Frontier --> Win["推动技术加速演进<br>(扩大全球领先优势)"] ``` David Sacks 等人指出，美国拥有五大互相角逐的前沿模型梯队，且在政策上避免了过激监管与扼杀性禁令，甚至 Mythos 的延缓也是行业自律性的“安全港”行为。当前真正的任务是防止敌对性模型蒸馏（adversarial distillation）并加速走向全球。 Jason 随后盘点了 Polymarket 预测市场上关于 2027 年前 IPO 的概率： - **SpaceX：** 95% 概率在 2027 年前上市 - **Cerebras：** 94% 概率 - **Anthropic：** 51% 概率在今年底前完成上市 - **OpenAI：** 44% 概率在年底前上市 - **Anthropic 市值预期：** 市场预测其收盘市值有 78% 的概率超过 6000 亿美元，仅 19% 概率未能成功上市。 Brad 证实 Altimeter 在两家公司中均投入了数十亿美元，其中进入 Anthropic 的第一轮估值在 300 亿或 1500 亿美元左右，预计将为基金带来 5 到 7 倍的丰厚回报。 --- ### [01:03:24](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=3804s) 资本挤压与软件栈各层的价值捕获博弈 Brad 坦言，当前整个风险投资（VC）行业正面临一种“存在主义危机”般的焦虑。传统软件股正在遭遇血洗： - 代表软件业的 **IGV 软件指数** 年内重挫 30%，单日跌幅达 5%。 - 传统企业级软件遭到资本无情抛售，风险投资人几乎不敢碰任何非前沿模型或非硬件防线（如 SpaceX、国防军工现代化）的项目，因为在强大的通用模型面前，很难评估小软件公司的终局价值（Terminal Value）。 - Altimeter 采取了**杠铃策略（Barbell approach）**：一头重注最核心的前沿大模型底座，另一头重注在 AGI 时代依然拥有强防御性壁垒的极小型极客团队。 针对传统企业软件（如 Salesforce、HubSpot、Oracle）是否会被私募股权（PE）私有化、并购整合还是被 AI 取代，David Sacks 分析认为市场可能过度悲观了： ```mermaid flowchart LR accTitle: AI价值捕获层级演进 accDescr: 展示从芯片层到超大规模云厂商，再到模型层以及潜在应用层的价值流动。 Layer1["1. 芯片层<br>(Nvidia率先突破数万亿)"] --> Layer2["2. 超大规模算力层<br>(Hyperscalers基础设施)"] Layer2 --> Layer3["3. 模型层<br>(Anthropic / OpenAI高速捕获)"] Layer3 -. 演进悬念.-> Layer4["4. 应用层<br>(如Palantir获赋能 / 传统SaaS重新洗牌)"] ``` - **代码成本下降 vs 竞争壁垒：** 软件生成的成本虽会大幅降低，但很多传统软件的护城河从来不是单纯的代码行数，而是客户工作流、数据积累与分发渠道。“倒洗澡水时可能把婴儿也一起倒掉了”，企业软件领域现在可能存在被错杀的价值投资标的。 - **价值层级演化：** 最初整个 AI 产业的利润完全被芯片层（以 Nvidia 为代表）独吞；随后扩散到超大规模云厂商（Hyperscalers）；目前模型层（Anthropic 与 OpenAI）正在大举捕获价值；接下来的核心悬念在于应用层（Application layer）——究竟是模型层通吃应用，还是诞生类似被大幅赋能的 Palantir 这样的超强应用，整个市场依然处于极早期的动态观察之中。 ### [01:06:53](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=4013s) Liquidity 2026 与 All-In 峰会筹备花絮：重磅演讲嘉宾与内部调侃 节目开场延续了轻松互损的氛围。Jason 透露 Brad Gerstner 将再次作为演讲嘉宾参加即将在加州纳帕（Napa）举办的 Liquidity 峰会，并称赞 Brad 在以往历届活动中的评分极高，始终排在第一或第二，从未跌出过前三。正当大家准备进入正题时，Chamath 突然插播了一条极为私人的“突发新闻”：他的伴侣 Nathalie 刚在维基百科上查阅了男性生理尺寸数据并截图发给他，专门圈出他处于“前 5% 的大尺寸范畴”。这引得全场爆笑与调侃，Jason 打趣这是否算上了“Nano Banana”（指此前节目中调侃的微型尺寸），并戏谑 Chamath 就像被黑客威胁曝光隐私时的贝佐斯（Jeff Bezos）一样，“根本不怕黑客攻击，因为爆料再狠也没有他自己在播客里自曝的尴尬”。 随后众人回归峰会本身的讨论。Brad 称赞了纳帕峰会的日程安排，包括顶级德州扑克局与高端餐饮体验。Jason 爆料称，因为 Chamath 态度强硬地要求整场活动必须达到“六星级标准”，否则自己绝不出席，因此会务团队全力以赴，力求做到“全员精英、绝无中庸（No mids）”。目前候补名单上已有数百人，但因场地容量限制无法全部入场。Chamath 随即公布了部分重磅嘉宾阵容，除了 Brad Gerstner 之外，Coatue 联合创始人 Thomas Laffont 也已确认出席。 此外，Chamath 透露本次 Liquidity 峰会还邀请了三位处于上升期、各自管理规模在 10 亿至 70 亿美元之间的年轻一代顶级复合收益投资者（compounders）来进行“最佳投资标的（Best Ideas）”路演。目前已确认的主讲阵容堪称豪华，包括 Bill Ackman、前特斯拉 AI 总监 Andrej Karpathy、Third Point 创始人 Dan Loeb、Thomas Laffont、Brad Gerstner 以及 Sarah Friar 等人，Jason 也暗示现场还将有一到两位未公开的重磅神秘嘉宾。紧接着，Jason 宣布第五届 All-In 年度峰会定于 2026 年 9 月 13 日至 15 日在洛杉矶举行，此前一直对加州税收和监管心存芥蒂的 David Sacks 幽默地表示，经咨询后确认自己可以出于“商务目的”合法踏足加州境内，确认将会出席两场盛会。 - **Liquidity 2026 已确认演讲嘉宾**：Bill Ackman、Andrej Karpathy、Dan Loeb、Thomas Laffont（Coatue）、Brad Gerstner、Sarah Friar 等。 - **第五届 All-In 年度峰会信息**：2026 年 9 月 13 日至 15 日，于洛杉矶举行（购票申请：allin.com/events）。 --- ### [01:10:08](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=4208s) 伊朗局势进入停火期：特朗普极限施压与伊斯兰堡和谈 节目切入本期的严肃政治焦点——已进入第七周的伊朗战争。录制前两天，为期两周的停火协议正式生效，美国副总统 JD Vance、特使 Steve Witkoff 以及女婿 Jared Kushner 正前往巴基斯坦首都伊斯兰堡，计划在周末展开和平谈判。Jason 梳理了特朗普在复活节期间惊心动魄的极限施压过程：复活节周日，特朗普在 Truth Social 上发表措辞极其激烈的言论，写道“给我把海峡打开，你们这群疯子，否则你们将生活在地狱里。走着瞧。”紧接着在周二上午，特朗普再次发出通牒，宣称“今晚将有一个完整的文明被毁灭，永远无法复原。我并不希望发生这种事，但这大概率会发生”，并给伊朗设定了当晚 8:00 的最后期限。 就在最后通牒到期前的一个半小时（下午 6:30），特朗普突然在 Truth Social 上宣布，鉴于伊朗方面提交了一份包含 10 点内容的和平提议，美方认为这是具有可行性的谈判基础，因此同意实行两周停火，前提是伊朗开放霍尔木兹海峡。特朗普甚至还以典型的商业口吻调侃称，未来海峡可以设立“收费站”，美方拿走大部分过路费再与伊朗分成。然而，中东局势的裂痕并未完全平息，以色列总理内塔尼亚胡似乎将停火视作空档窗口，在黎巴嫩上空展开了狂轰滥炸，据称在短短 10 分钟内投下了 160 枚炸弹。 --- ### [01:11:47](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=4307s) 萨克斯与格斯特纳把脉局势：破除升级陷阱与市场全面反弹 面对 Jason 的提问，此前缺席节目的 David Sacks 首先作了严正免责声明。他强调自己并非白宫外交政策团队成员，只是在科技领域担任顾问，此前媒体动辄将他的播客言论渲染为“特朗普顾问就战争表态”，引发了不必要的国际风波。在此前提下，Sacks 表达了自己的个人看法：他高度赞赏特朗普促成两周停火并派遣代表团赴伊斯兰堡谈判的成果。从历史规律来看，此类战争极易陷入所谓的“升级陷阱（Escalation Trap）”，双方在升级阶梯上不断加码，通常很难主动踩下刹车。特朗普能够果断叫停升级态势并开启外交接触，展现了非凡的决断力。 ```mermaid flowchart TD accTitle: 资本市场对地缘政治危机的冲击韧性对比 accDescr: 流程图展示了关税恐慌与伊朗战争对标普和纳指的回撤对比，以及后续潜在利好驱动因素。 subgraph "历史对照：关税冲击" A["关税政策出台恐慌"] --> B["纳斯达克高点回撤 ~15%"] B --> C["盘中最大跌幅触及 22%"] end subgraph "当前战局：伊朗危机" D["伊朗冲突爆发（历时6-7周）"] --> E["标普/纳指仅回撤 5% - 7%"] E --> F["停火协议达成，指数完全收复失地"] end subgraph "潜在上行催化剂" G["Vance 伊斯兰堡和平谈判"] H["以色列与黎巴嫩展开直接政府接触"] I["俄乌和平协议预期 / 委内瑞拉与古巴进展"] G & H & I --> J["美国建国250周年（7月4日）前市场全面爆发"] end ``` Brad Gerstner 回顾了自己早在 3 月 4 日就提出的“特朗普伊朗主义（Trump Doctrine in Iran）”核心观点：利用绝对优势毁灭其军事能力，定点清除制造致命武器的人员，然后迅速撤离并保留再次打击的权利，绝不重蹈覆辙去搞徒劳的“麦迪逊式民主（Madisonian democracy）”国家重建。Brad 指出，金融市场对这场危机的解读非常理性：去年关税恐慌期间，纳斯达克指数自高点回撤了约 15%（盘中一度跌超 22%）；而本次长达六周的伊朗战争期间，标普 500 和纳斯达克的回撤仅有 5% 到 7%。 投资者虽被特朗普“摧毁文明”的惊悚推文吓得不轻，但市场本质上相信特朗普“打完就走、绝不深陷泥潭”的承诺。如今美股指数已完全抹平了伊朗战事的跌幅，重返战前高位。Brad 乐观地认为，如果 Vance 能在伊斯兰堡促成最终协议，再加上内塔尼亚胡已表态愿与黎巴嫩政府直接对话，市场将迎来巨大利好。不仅如此，俄乌和谈正在接近实质进展，委内瑞拉局势趋向平稳，古巴方面也可能迎来突破。如果这些地缘热点在 7 月 4 日“美国建国 250 周年”纪念日前顺利着陆，市场将迎来极其强劲的上行行情。紧接着，Jason 话锋一转，引出《纽约时报》披露的内幕深度调查，准备深究美国当初究竟是如何在内塔尼亚胡的引导下卷入这场战争的始末。 ### [01:15:40](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=4540s) 白宫内部对空袭伊朗的分歧与内塔尼亚胡的影响力争议 Jason Calacanis 援引报道指出，以色列总理本雅明·内塔尼亚胡（Benjamin Netanyahu）曾在白宫向唐纳德·特朗普（Donald Trump）做了一份由四部分组成的打击伊朗游说方案。然而，白宫内部对此存在巨大分歧：副总统 JD·万斯（JD Vance）警告特朗普，这场战争可能引发区域混乱，甚至瓦解特朗普好不容易构建的“MAGA 2.0”大帐篷联盟——事实证明，梅根·凯利（Megyn Kelly）和塔克·卡尔森（Tucker Carlson）等保守派核心声音随后确实对此表达了强烈焦虑；国务卿马尔科·卢比奥（Marco Rubio）虽反对政权更迭，但对轰炸行动大体保持中立；白宫办公厅主任苏西·威尔斯（Susie Wiles）则极具预见性地提出了中期选举前对油价飙升的担忧；而参谋长联席会议主席丹·凯恩上将（General Dan Caine）对内塔尼亚胡的方案评价尤为尖锐：“长官，根据我的经验，这是以色列人的标准操作程序。他们总把前景吹得天花乱坠，但计划往往不够成熟。他们清楚自己需要我们，所以才拼命推销。” ```mermaid flowchart TD accTitle: White House Internal Perspectives on Iran Strike accDescr: Trump receives pitch from Netanyahu while internal advisers express concerns subgraph Pitch["外部游说"] BBI["内塔尼亚胡<br/>四部分对伊打击方案"] end subgraph WH["白宫幕僚与军方评估"] direction TB JDV["JD·万斯<br/>警告区域混乱与MAGA 2.0联盟破裂"] SW["苏西·威尔斯<br/>担忧油价冲击中期选举"] DC["丹·凯恩上将<br/>直言以方过度推销且方案不成熟"] MR["马尔科·卢比奥<br/>反对政权更迭但默认轰炸行动"] end subgraph Decision["最终裁决"] DJT["特朗普总统<br/>最终拍板行动"] end BBI --> DJT JDV -. 告诫.-> DJT SW -. 顾虑.-> DJT DC -. 评估.-> DJT MR -. 态度.-> DJT ``` Jason 还提到卢比奥在开战之初被收回的言论——“我们知道以色列将采取行动，也知道这会引发针对美军的袭击，这就是我们行动的原因”。在与宾夕法尼亚州州长乔希·夏皮罗（Josh Shapiro）交谈后，Jason 感到美国外交政策正被内塔尼亚胡深度牵引，许多美国犹太裔人士也普遍认为内塔尼亚胡在加沙、黎巴嫩和伊朗的行动不仅没有给海外犹太社群带来好处，反而助长了反犹主义浪潮。 Chamath Palihapitiya 则明确表示，最终拍板的始终是美国总统，任何外国领导人都不可能替美国做主。从金融市场来看，各大预测市场已经将此次冲突定价为一次“短暂的插曲”，事态大概率正在进入尾声。但 Chamath 强调，以色列必须警惕深层风险：如果不能迅速找到冲突的“退场通道”（offramp），以色列可能会永久性地削弱美国这一长期坚定盟友的支持力度，而这种地缘关系的断裂对以色列造成的打击远大于对美国的影响。 - 参谋长联席会议主席丹·凯恩上将直言：“这是以色列人的标准操作程序（SOP），他们过度推销，方案不成熟，但因为需要我们所以强行硬推。” - 预测市场与金融资产走势反映出，资本市场普遍将此番军事交火视作局部短期波动。 - Chamath 指出，冲突若无迅速收场的退场机制，以色列失去美国稳定支持的风险将急剧上升。 --- ### [01:19:18](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=4758s) 中东格局重塑的历史赌局与内塔法利·贝内的民调警告 谈及美国外交政策是否被以色列过度影响，Brad Gerstner 呼应了 David Sacks 此前的观点，认为历史最终只会以结果来论成败。尽管眼下人人都是“键盘政客”，但他刚与沙特、卡塔尔、科威特和阿联酋等多方高层沟通，海湾国家对未来的经济与地缘转型充满了极高的乐观情绪，甚至不排除未来将伊朗纳入地区经贸合作版图的可能性。Brad 提醒大家，在伊朗问题上“无所作为”同样蕴含着巨大的地缘政治风险，就像在委内瑞拉问题上按兵不动一样，行动本身是经过精密盘算的赌局，必须给历史留出检验的时间。 与此同时，以色列政坛内部也察觉到了风向的逆转。前总理内塔法利·贝内（Naftali Bennett）在 X 上引用最新民调指出，以色列在美国公众心目中的支持率正在断崖式下跌，并公开疾呼：“这是一个极其严峻的局面，要修复这一切，我们还有大量工作要做。”David Sacks 评价称，贝内作为内塔尼亚胡的政敌公开挑明这一事实，说明以色列精英层已经意识到公共外交的危机；如果以色列能够切实支持停火，将是挽回民意的重要起点。 - Brad Gerstner 认为海湾国家（沙特、阿联酋、卡塔尔等）正处于重大区域转型的边缘，不应单方面判定行动为失误。 - 内塔法利·贝内公开承认民调恶化，警告美以关系裂痕已危及以色列的海外声誉。 - David Sacks 强调，支持停火协议是以色列扭转国际舆论劣势的当务之急。 --- ### [01:21:50](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=4910s) Grok 跨语言实时翻译引发热议与 X 的“极限人效”启示 在讨论贝内的希伯来语推文时，话题意外转向了社交网络底层的技术跃迁。Jason 极度兴奋地赞叹 X 依托 Grok 大模型推出的无缝自动翻译功能，称这是他“多年来见过的最震撼的技术功能”，其对跨国界理解的推动甚至配得上一座“诺贝尔和平奖”。该功能不仅能将希伯来文、法文、日文等全球各地的核心讨论自动呈现给母语用户，还能让跨语种网友在评论区零门槛互动交流。 ```mermaid flowchart LR accTitle: Grok Real-Time Translation Loop on X accDescr: Real-time translation loop enabling direct cross-border dialogue subgraph NativeContent["本土语境内容"] JA["日本 / 以色列 / 法国等本土推文"] end subgraph GrokEngine["Grok 实时推理翻译"] TR["多语言语义理解与双向转译"] end subgraph CrossInteraction["跨文化无缝互动"] US["英语及其他地区用户即时浏览与回复"] end JA --> TR --> US US --> TR --> JA ``` Brad Gerstner 同样对此深有感触，尤其是中东、俄罗斯与中国方向的资讯时效。Jason 绘声绘色地打趣道，自动翻译让自己领略到了极其硬核的“日本本土极右推特文化”（Based Japanese Twitter），其对外来移民毫无顾忌的直球排斥程度，让美式极右名嘴尼克·富恩特斯（Nick Fuentes）和亚历克斯·琼斯（Alex Jones）都显得像温和派。Chamath 随即打趣 Jason 是个不折不扣的“吹鼓手”：“你吹捧一片涂抹黄油的吐司面包时的亢奋语调，跟你吹捧一座核反应堆时一模一样！”Jason 则大笑反驳：“没有什么比一片焦香的烤吐司更美妙了，这项功能就像切片面包一样朴实却极具革命性。” 最后，Brad 将这一产品突破引申至硅谷的企业治理变革：X 如今不仅体验和功能优于以往，而且团队规模相比埃隆·马斯克（Elon Musk）接手时缩减了整整 70%。当初外界断言裁员会导致平台崩溃并沦为垃圾，如今几年过去，X 证明了极致的人效可能。这一样本彻底颠覆了硅谷大厂的用人逻辑，标志着科技行业传统岗位已然触及“峰值雇佣”（Peak Employment），整个行业正在全面转向以更少的人力借助 AI 产出更庞大的价值。 - X 凭借 Grok 的自动翻译能力打破全球信息孤岛，实现不同语言用户之间的实时语义互通。 - Jason 戏谑“日本原汁原味推特”在移民议题上的激进言论把美国右翼衬托得十分温和。 - X 在裁员 70% 的背景下持续交付高水准功能，成为硅谷大厂告别盲目扩张、迈向“小团队高产出”与峰值就业时代的典范。 ### [01:25:01](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=5101s) X/Twitter 平台治理与评论区管理：Nikita Bier 打击自动回复机器人与订阅者模式 话题转向社交平台 X（原 Twitter）的用户体验恶化问题。嘉宾们提到 Nikita Bier 刚刚发推宣布，他们即将对平台上泛滥成灾的自动回复 Bot（机器人账号）大开杀戒、下狠手整治。几位主持人对此拍手称快，纷纷表示这些无休止刷屏的垃圾机器人简直彻底毁掉了日常的信息流（Feed）。 Jason 随即分享了自己管理评论区的实操经验。他表示，正是因为不堪其扰，他此前直接把推文的回复权限调整为了“仅限付费订阅者（Subscriber Mode）”，这一改动收效显著，整个互动环境立刻清爽了许多。他们还顺带隔空安慰了近期对 Twitter 现状痛心疾首的知名投资人 Chris Sacca，打趣说“一切都会好起来的，Chris，别再流泪了”。当被问及是否永远只允许订阅者发言时，Jason 澄清自己采取的是五五开的策略：有时候他会彻底放开评论区，任凭网络狂欢与混乱（Chaos）肆虐；而在不想被数百条垃圾回复轰炸的时候，他就切换为订阅者专享模式。目前他拥有大约 2,000 名付费订阅者，每年产生约 3 万美元的收入，他把这笔钱全额捐给了慈善机构。Jason 感慨道，在自己多达上百万的关注者中，能够借此固定认识并深度交流这核心的 2,000 个人，就像是在浩瀚的网络世界里建立了一个高质量的专属社群。 - **Nikita Bier 的整顿行动**：针对 X 平台泛滥的自动回复机器人展开专项打击。 - **Jason 的订阅互动机制**： - 拥有 2,000 名核心付费订阅者。 - 每年产生约 30,000 美元订阅费，全部捐赠给慈善机构。 - 采取“50% 自由开放 / 50% 订阅者专享”的动态策略平衡互动与精力。 --- ### [01:25:48](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=5148s) 粉丝数量排位赛与 Chamath 怒怼黑粉的名场面 聊到关注者规模，几位主持人和嘉宾立刻开启了硅谷顶流风投之间的日常互相攀比与逗趣调侃。得知 Jason 的粉丝突破百万大关后，大家纷纷报数：Chamath 坐拥 200 万粉丝高居榜首；David Sacks 和 Jason 也各自跨越了百万门槛，Jason 目前在 110 万左右。做客的 Brad Gerstner（推特账号 @altcap）则得意地透露自己的粉丝数已经飙升到了 140 万，当场笑问 Jason 到底要花多少钱才能从别人手里买回自己真名的 Handle。 众人调侃道，粉丝数的暴涨全靠这档“全球第一播客”的巨大声量加持。不过大家一致公认，Chamath 能积攒下 200 万粉丝，很大程度上归功于他常年在评论区极其凶悍地与黑粉（Haters）正面对线。Jason 盛赞 Chamath 的那些“神级回怼”，直言看 Chamath 爆锤黑粉是他最爱的推特保留节目。Chamath 骄傲地表示自己每次都是直接“开火打得他们落花流水（light them up）”。 两人随后还还原了近期一个令人捧腹的真实闹剧：一名网友在网上怨声载道地碰瓷 Chamath，声称自己在赌场里押轮盘赌，全是因为听了 Chamath 之前的建议选了“押黑（bet black）”，结果输得精光，要求 Chamath 对此负全责。面对这种荒谬至极的巨婴逻辑，Chamath 却做出了极具反差的戏剧性回应——他直接出资为这名男子家里的两个小女儿设立并全额资助了大学教育基金（College Accounts）。Chamath 在节目中毫不留情地讽刺道，自己很乐意帮两个无辜的小女孩，但更让他身心愉悦的，是这招能彻底把网上那些“整天赖在老妈地下室里打游戏、吃热馅饼和炸鱼柳、对自己的失败人生不负任何责任的键盘侠屌丝们”活活气疯。 - **主持人群体 X 粉丝分布概况**： - Chamath Palihapitiya：约 200 万关注者。 - Brad Gerstner（@altcap）：约 140 万关注者。 - Jason Calacanis：约 110 万关注者。 - David Sacks：百万级关注者。 - **经典对线事件**：网友赌轮盘押黑输光后碰瓷 Chamath，Chamath 反向出资赞助其两个女儿的大学基金，借此公开嘲讽网络巨婴与键盘侠文化。 --- ### [01:27:23](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=5243s) 结语调侃、Athena 助理狂潮与播客经典片尾 在节目的尾声，众人开始做最后的收尾总结。虽然好友 David Friedberg 本期缺席，但大家毫不客气地开启内部嘲讽模式，开玩笑称“虽然想念 Friedberg，但这绝对是两年来录得最棒的一期”。几位主持人还预告了即将到来的峰会活动（Summit），并戏谑地向“排在候补名单上无法入场的 400 位朋友”致以同情。 随后话题转到了他们常用的离岸虚拟助理服务商 Athena 上。几位主持人透露，由于上周节目中大伙儿自发热烈安利了 Athena，直接引发了海量咨询狂潮，Athena 官方甚至发邮件表示他们正准备在新一轮扩张中招募大约 500 名全新的虚拟助理，上周节目后涌入了上千个申请。大家对能为菲律宾等海外地区创造 500 个优质的高薪白领岗位感到由衷的高兴。Jason 还打趣要买两个 Athena 助理送给 Sacks 当生日礼物，现场气氛轻松欢快。 节目的最后，伴随着几位好友互道“爱你们”的告别声，播客切入了标志性的片尾串烧合辑——从“让赢家继续奔跑（Let your winners ride）”、“雨神（Rainmaker）”到经典的粉丝鬼畜二创片段与各种无厘头内部搞笑名场面，在宣告这期马拉松式长节目圆满落幕的同时，也以标志性的“I'm going all in”画上了句号。 - **Athena 助理效应**：播客的自发推广促使 Athena 迎来超 1,000 位潜在客户需求，计划在菲律宾等地新招募 500 名高薪虚拟助理。 - **现场花絮与彩蛋**：调侃缺席的 David Friedberg，回顾 All-In 峰会候补情况，并在欢笑和经典口头禅剪辑中结束全片录制。 ## Notable quotes > [04:22](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=262s) 模型擅长写代码的副产品，就是它在网络安全方面也异常强大。 — Dario Amodei > [09:31](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=571s) 我不是说他们不真诚，但他们确实有一套屡试不爽的套路：把‘恐惧’当作推广新产品的营销手段。 — David Sacks > [14:09](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=849s) Anthropic 现阶段的表现简直是火力全开。就像斯蒂芬·库里在全场任何位置疯狂扔三分球一样，球球空心入网。 — Chamath Palihapitiya > [14:09](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=849s) 如果你想要彻底修补整个互联网的历史安全漏洞，你得把整个互联网切断关闭大约 5 年。 — Chamath Palihapitiya > [44:14](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=2654s) 相当于在短短一个月内，凭空增加了一个 Databricks 加上 Palantir 的总和。 — Brad Gerstner > [51:12](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=3072s) 无论他们的年化营收是 10 亿美元还是 800 亿美元，手头能烧的也就是这区区 1.5 吉瓦的算力。 — Brad Gerstner > [55:27](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=3327s) 硅谷从来不是靠冷冰冰的 Excel 财务模型驱动的，真正的创新者凭借直觉与信仰坚信‘建好了，人们自然会来’。 — David Sacks > [59:03](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=3543s) 即便当前市场氛围如此悲观，我今天依然是 OpenAI 股票的积极买家。 — Brad Gerstner > [01:10:08](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=4208s) 给我把海峡打开，你们这群疯子，否则你们将生活在地狱里。走着瞧。 — Donald Trump > [01:15:40](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=4540s) 长官，根据我的经验，这是以色列人的标准操作程序。他们总把前景吹得天花乱坠，但计划往往不够成熟。他们清楚自己需要我们，所以才拼命推销。 — General Dan Caine > [01:21:50](https://www.youtube.com/watch?v=DVBJQQCjgXU&t=4910s) 你吹捧一片涂抹黄油的吐司面包时的亢奋语调，跟你吹捧一座核反应堆时一模一样！ — Chamath Palihapitiya ## Mentioned Anthropic, OpenAI, Mythos, OpenClaw, Claude, Claude Code, Claude 3.7 Opus, Spud, Project Glasswing, Brad Gerstner, Chamath Palihapitiya, Jason Calacanis, David Sacks, David Friedberg, Dario Amodei, Sam Altman, Peter Steinberger, Elisha Long, Ro Khanna, OpenBSD, FFmpeg, Linux, Apple, Microsoft, Google, Amazon, JPMorgan, Meta, SpaceX, xAI, Grok, Perplexity, Perplexity Computer, Nous Research, Hermes Agent, Alibaba, Qwen, Elon Musk, Macrohard, Bittensor, Venice, Bill Gurley, Databricks, Palantir, Cowork, Sora, Disney, Polymarket, Cerebras, Salesforce, HubSpot, Oracle, Bill Ackman, Andrej Karpathy, Dan Loeb, Thomas Laffont, Sarah Friar, Coatue, Third Point, Donald Trump, JD Vance, Steve Witkoff, Jared Kushner, Benjamin Netanyahu, Marco Rubio, Susie Wiles, General Dan Caine, Josh Shapiro, Naftali Bennett, Nikita Bier, Chris Sacca, Athena ## Video description (0:00) Bestie intros: Brad Gerstner joins the show! (4:22) Anthropic blocks Mythos release for security concerns: major threat or marketing stunt? (24:07) Are OpenAI and Anthropic trying to kill OpenClaw? Does Anthropic already have market dominance in AI coding? (42:20) Anthropic $30B run rate, fastest revenue ramp ever, the TAM for intelligence (58:01) Major vibe shift: Anthropic ripping, OpenAI reeling (1:10:12) Iran War: Ceasefire, Israel's influence, market impact Apply for Summit 2026: https://allin.com/events Follow Brad: https://x.com/altcap 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://www.youtube.com/watch?v=INGOC6-LLv0 https://openai.com/index/better-language-models https://cdn.openai.com/better-language-models/language_models_are_unsupervised_multitask_learners.pdf https://x.com/steipete/status/2040811558427648357 https://x.com/juliusai/status/2041292301234999668 https://polymarket.com/event/ipos-before-2027 https://www.google.com/finance/quote/IGV:BATS https://polymarket.com/event/anthropic-ipo-closing-market-cap-119 https://truthsocial.com/@realDonaldTrump/posts/116351998782539414 https://truthsocial.com/@realDonaldTrump/posts/116363336033995961 https://truthsocial.com/@realDonaldTrump/posts/116365796713313030 https://www.nytimes.com/2026/04/07/us/politics/trump-iran-war.html https://www.state.gov/releases/office-of-the-spokesperson/2026/03/secretary-of-state-marco-rubio-remarks-to-press-6 #allin #tech #news [Source](https://www.youtube.com/watch?v=DVBJQQCjgXU)

[4:22] Dario Amodei: 模型擅长写代码的副产品，就是它在网络安全方面也异常强大。 [9:31] David Sacks: 我不是说他们不真诚，但他们确实有一套屡试不爽的套路：把‘恐惧’当作推广新产品的营销手段。 [14:09] Chamath Palihapitiya: Anthropic 现阶段的表现简直是火力全开。就像斯蒂芬·库里在全场任何位置疯狂扔三分球一样，球球空心入网。 [14:09] Chamath Palihapitiya: 如果你想要彻底修补整个互联网的历史安全漏洞，你得把整个互联网切断关闭大约 5 年。 [44:14] Brad Gerstner: 相当于在短短一个月内，凭空增加了一个 Databricks 加上 Palantir 的总和。 [51:12] Brad Gerstner: 无论他们的年化营收是 10 亿美元还是 800 亿美元，手头能烧的也就是这区区 1.5 吉瓦的算力。 [55:27] David Sacks: 硅谷从来不是靠冷冰冰的 Excel 财务模型驱动的，真正的创新者凭借直觉与信仰坚信‘建好了，人们自然会来’。 [59:03] Brad Gerstner: 即便当前市场氛围如此悲观，我今天依然是 OpenAI 股票的积极买家。 [70:08] Donald Trump: 给我把海峡打开，你们这群疯子，否则你们将生活在地狱里。走着瞧。 [75:40] General Dan Caine: 长官，根据我的经验，这是以色列人的标准操作程序。他们总把前景吹得天花乱坠，但计划往往不够成熟。他们清楚自己需要我们，所以才拼命推销。 [81:50] Chamath Palihapitiya: 你吹捧一片涂抹黄油的吐司面包时的亢奋语调，跟你吹捧一座核反应堆时一模一样！

[16:40] Whether it's total theater, whether it's fear-mongering, and they do have a pattern around this, we can't take the risk... As these models become more and more capable at coding, they're going to get better at cyber. [29:26] Jason Calacanis: The agent level is essentially your entire life... It is foolish for you to give your entire business and all the knowledge you have to Anthropic through that. [39:43] David Sacks: Eventually the government's going to look at this market with the benefit of 20/20 hindsight, and I think everyone should just basically keep your nose clean... Keep it tight. Tight is right. [58:01] Brad Gerstner: Anthropic made choices. No multimodal, no video, no hardware, no chips, no building data centers. They said, 'We're just going to focus on coding and co-work.' They executed their butts off. [75:11] General Dan Caine: Sir, this is, in my experience, standard operating procedure for the Israelis. They oversell and their plans are not always well-developed. They know they need us and that's why they're hard selling. [78:30] Naftali Bennett: This is a serious situation. There's a lot of work ahead of us to fix everything.

[0:00] How many PRs you think are going to get [0:01] pushed to the core structural internet [0:03] in 100 days? What's the overunder [0:05] number? Cuz I'll give you a number. [0:06] >> You're going to say zero. My my answer [0:08] to that is [0:09] >> I'll say like 10,000. But it's going to [0:11] be immediately [0:11] >> if it prevents your browser history from [0:13] being released to everybody in the [0:15] world, Chamath, that may be something [0:17] that you're willing to, you know, let [0:18] 100 days pass on. [0:19] >> I think you got Chimat's attention when [0:20] you said browser history. [0:21] >> What about the dickpicks? [0:25] >> Chamat is he's going to release them [0:27] himself. [0:30] We'll let your winners ride. [0:38] >> We open source it to the fans and [0:40] they've just gone crazy with it. Love [0:41] you. [0:46] >> All right, everybody. Welcome back to [0:47] the number one podcast in the world. [0:48] David Freeberg is out this week. But in [0:51] his place, the one, the only, [0:54] our fifth bestie, Brad Gersonner. I [0:56] mean, why don't you ever give me puts a [0:58] little namaste in your payday anymore? [1:00] You used to be [1:02] >> I'm going to bring back the greatest [1:03] moderator, but now it's just kind of You [1:06] know what? These guys beat me up. They [1:08] beat me up and they just beat the the [1:10] joy out of me doing this program. [1:13] >> It's because you're a Roana apologist [1:15] now. [1:16] >> No, I We'll get into it. Okay. Save it [1:18] for the Roana apologist. just [1:21] because I said like, "Hey, they've [1:23] stopped maxing and they've [1:25] started doing like some logical things." [1:28] Uh, yeah. Okay, here we go. [1:30] >> It's great to be here. Great to be here. [1:31] >> Good to have you. Good to have you here. [1:33] And of course, uh, we have David Saxs is [1:37] back. Everybody wants to hear from David [1:38] Sax. We missed you last week, bestie. [1:40] >> We didn't beat the joy out of you. We [1:42] just try to beat some of the hot air. [1:44] Turn [1:45] >> any any fluff that you can put on the [1:47] show that just involves you talking and [1:50] saying nothing is [1:52] >> that's the stuff we got. [1:53] >> Turn up. Yeah. Turn up. Okay. Yeah, [1:55] we'll cut it right out. Um we'll cut it [1:58] out and we'll just put a promo in for [1:59] the syndicate.com. Thank you. Also with [2:01] us, Jamalet is here. [2:04] >> How's your maxing going since [2:06] last week? Did you have a a a [2:08] maxing full weekend? Did you have a good [2:10] full weekend of just smoking cigars in [2:12] the back deck and not ruminating about [2:14] all the chaos you've caused in the last [2:16] 20 years? [2:17] >> I think I've done generally more good [2:20] >> than than not. [2:22] >> Oh, you have. But there's been some [2:24] chaotic moments. Don't think about it. [2:26] >> You can't, bro. You can't have ups [2:28] without downs, man. It's like, what are [2:30] you there to do? Just like plate [2:31] everybody and be a loser? Are you there [2:33] to be a winner? Yes, you're in the [2:35] arena, but have you stopped going to [2:37] therapy after realizing ruminating? [2:40] >> What's up with this uh sudden interest [2:42] in maxing? Are you like the [2:44] clavvicular for maxing? [2:47] >> No, the world finally caught up with me. [2:49] That's it. What do you I mean, I've been [2:50] maxing this whole time. THEY JUST [2:52] DIDN'T HAVE A name for it, guys. [2:54] >> Wow. Okay. Eli's videos are really good. [2:56] I watched two more this week. [2:58] What take us through what's so appealing [3:01] about not ruminating, smoking a cigar, [3:03] and just living your life? [3:05] >> Because what he says actually works at [3:07] every level of society and every sort of [3:10] thing that you may want to achieve. Even [3:12] if you're trying to like climb the [3:14] rungs, [3:17] you very quickly learn that the more you [3:19] want something, the less you're going to [3:20] get it. And I think that's like his real [3:23] message is let go, live life, and just [3:27] try stuff or don't try stuff. And I [3:29] think that that detachment is really [3:31] healthy for people. I like it. I like it [3:34] a lot. [3:35] >> Who's the guy who says this? I actually [3:37] didn't know. [3:37] >> Elisha Long. Well, Eli, I think, is how [3:40] he goes by. [3:41] >> But he's fantastic. He Mark YouTube [3:43] channel. [3:44] >> Mark Andre found him [3:46] >> and he's like, "This is this guy is the [3:48] new guy. [3:49] >> Modernday philosopher. He gives you a [3:51] road map for how to live your life, [3:52] right? A new age sage. [3:54] >> What's the name of the guy? The [3:55] character's name from Dune. [3:57] >> I was into girls books. I was dating [4:00] girls. [4:01] >> He's the Lison Algib of the modern [4:03] internet. [4:04] >> This is why we need Freeberg here is to [4:06] explain these deep holes. All right, [4:09] listen. We got a lot to get to. Don't [4:10] The basic point is build something and [4:12] don't ruminate. Okay, ruminating is just [4:14] not worth it. Just everybody go for it. [4:16] >> No, just do stuff. Stop blathering in [4:18] your own head. Just do stuff. [4:19] Absolutely. All right. Listen, speaking [4:21] of doing so, Anthropic is withholding [4:23] its newest model, Mythos. I'm using the [4:26] Greek uh pronunciation, its newest [4:28] model, Mythos, uh saying it is far too [4:31] dangerous for any of us to have access [4:33] to it. According to the company, the [4:35] model autonomously found thousands of [4:37] vulnerabilities, including bugs in every [4:39] major operating system and web browser. [4:42] This uh little study they did included [4:45] 20 year old exploits that had been [4:47] missed by security audits for decades. [4:49] Uh some examples, they found a 27y old [4:51] vulnerability in OpenBSD used in [4:54] firewalls and critical infrastructure. [4:55] They found a 16-year-old bug in FFmpeg [4:59] that was missed by automated tools after [5:01] 5 million scans. The Linux kernel, all [5:05] kinds of uh bugs they found. They [5:07] released a hype video hyping up why they [5:11] were not going to share this model. [5:13] Here's Dario. Come on the program [5:15] anytime, brother. [5:16] >> But as a side effect of being good at [5:17] code, it's also good at cyber. [5:19] >> The model that we're experimenting with [5:22] is by and large as good as a [5:25] professional human at identifying bugs. [5:28] It's good for us because we can find [5:30] more vulnerabilities sooner and we can [5:32] fix them. [5:33] >> It has the ability to chain together [5:35] vulnerabilities. So what this means is [5:37] you find two vulnerabilities, either of [5:39] which doesn't really get you very much [5:40] independently, but this model is able to [5:43] create exploits out of three, four, [5:45] sometimes five vulnerabilities that in [5:47] sequence give you some kind of very [5:49] sophisticated end outcome. [5:50] >> All right, Brad, uh, by the way, that [5:51] set they're using there, that's the same [5:53] room those guys play Dungeons and [5:54] Dragons in every Sunday. Brad, you're [5:58] Brad, you're an investor in this [6:00] company. Is this virtue signaling or is [6:03] it reality? Is this a good move by them [6:06] to not release this model and be [6:08] thoughtful, give it to a handful of [6:09] people and just find all the bugs it can [6:12] before releasing it to the public? And [6:14] we've got a lot more issues to discuss. [6:16] >> I I actually think they deserve a ton of [6:18] credit here and let me walk you through [6:20] why, right? They the company could have [6:22] just released Mythos, broken a lot of [6:24] core things on the internet. Often times [6:26] in Silicon Valley, we say move fast and [6:28] break things. In this case, it means [6:30] just releasing the model to move further [6:31] ahead of your competition. But here the [6:33] company realized it would wreak havoc. [6:35] They ran their own vulnerability [6:37] testing. They saw that it would allow [6:39] offensive hacking and people to expose [6:41] browsers and browser history, expose [6:43] credit cards, you know, on the internet. [6:46] So, you know what I like about this is [6:48] they didn't need government to hold [6:50] their hand on this. We have plenty of [6:52] government regulations. They know it's [6:54] in the best long-term interest of the [6:56] company and the industry, you know. So, [6:58] they set up Project Glass Wing. It's an [7:00] AIdriven, you know, kind of cyber [7:02] coalition. Apple, Microsoft, Google, [7:05] Amazon, JP Morgan, 40 of the most [7:07] important companies. And their goal is [7:09] very simple. Let's spend a 100 days use [7:11] advanced AI to find and to fix and to [7:14] harden these software vulnerabilities [7:17] before hackers exploit them. Now, what I [7:20] think this represents, Jason, is a [7:22] threshold that we're crossing. Mythos [7:25] and Spud, which is going to be out from [7:27] OpenAI any day now, which is the first [7:30] Blackwell trained model at OpenAI. They [7:33] represent the beginning of what I would [7:35] call AGI models. These are models with [7:38] massive step function improvements and [7:40] intelligence. Um, and they're just too [7:43] smart to be released immediately, [7:45] you know. And by the way, there was [7:47] nothing that said that every time you [7:49] you finish a model, you got to [7:51] immediately release it GA. So they set [7:53] up this idea of sandboxing, building [7:55] defensive alliances, [7:57] you know, in order to move away from [7:59] that regime. I I think it shows, and [8:01] Saxon and I have talked about this a [8:03] lot, so I'm interested to hear what he [8:04] thinks. It shows you can trust the [8:07] industry and market forces in [8:10] coordination with the government. They [8:11] were talking to the government about [8:12] this. But they're not relying on some [8:15] top- down regulation in order to do [8:18] this. They laid out a blueprint that [8:19] seems to me very pragmatic that now that [8:22] we're at this threshold, we're going to [8:24] sandbox these things. I think that open [8:26] AAI will end up doing the same thing. I [8:28] think Google will end up doing the same [8:30] thing. It's an aggressive way to keep [8:32] the RA, you know, the pressure on and [8:34] and win the race at AI while making the [8:37] tradeoffs to protect safety. So, you [8:39] know, I think you're always going to [8:40] have to make these trade-offs. I think [8:42] in this case, it was a great move by [8:44] Dario and team and I think they deserve [8:46] a lot of credit. Sachs, when you look at [8:48] this, we had Emil Michael on the program [8:50] a couple weeks ago. It might have been [8:51] four or five weeks ago, and we had a [8:53] very thoughtful discussion about, hey, [8:55] if the government is going to have these [8:57] tools, you know, an anthropic wants to [8:59] withhold them and, you know, what is the [9:02] proper relationship there, you have to [9:04] think that the government, and I know [9:06] you don't speak for all parts of the [9:08] government. If you were just going to [9:10] run through the game theory, they must [9:11] have gone to the government and said, [9:12] "Listen, this thing is so powerful, it [9:14] can put together two or three hacks, [9:16] create a novel attack vector, and this [9:19] is incredibly dangerous. What if China [9:21] has it? And if this thing is as powerful [9:23] as Daario says it is, then this is an [9:26] offensive weapon as well for us to take [9:28] out, let's just pick, you know, uh, a [9:31] pressing issue, the North Korea's [9:34] ballistic missile program. This is [9:36] equivalent the way it's being described [9:38] as the Manhattan project perhaps. So [9:41] what are the chances two-part question [9:43] for you Sax that China already has this [9:45] and is using it and do you think Daario [9:48] is doing the right thing by regulating [9:50] themselves? [9:51] >> I think Anthropic has proven that it's [9:54] very good at two things. One is product [9:57] releases. The second is scaring people. [10:00] And we've seen a pattern in their [10:02] previous releases of at the same time [10:05] they roll out a new model or new model [10:07] card, something like that. They also [10:09] roll out some study showing really the [10:12] worst possible implication of where the [10:15] technology could lead. We saw this last [10:17] year about a year ago. They rolled out [10:19] this blackmail study where supposedly [10:22] the new model could blackmail users. [10:25] There's been a whole bunch of these [10:26] things. Actually, I went back to Grock [10:28] and I just asked, "Hey, give me examples [10:30] where Antropic has basically used scare [10:32] tactics and it's it's a pattern." Okay, [10:35] it's a pattern. [10:36] >> Okay, [10:37] >> these guys, I'm not saying it's not [10:40] sincere, but they have a proven pattern [10:42] of using fear as a way to market their [10:45] new products. And if you think back to, [10:48] again, my favorite example is this [10:50] blackmail study where they prompted the [10:53] model over 200 times to get the result [10:55] they wanted. And that result was was [10:58] clearly reverse engineered and it got [11:00] them the headlines they wanted. And I [11:02] would say the proof that it's reverse [11:04] engineered is we're now a year later. [11:06] There's a bunch of open- source models [11:08] out there that have the same level of [11:10] capability that that anthropic model [11:13] had. And have you seen any examples of [11:15] blackmail in the wild? I don't think so. [11:18] So in other words, if that study were [11:22] true in the sense of being a likely [11:23] outcome of that model, I think you would [11:25] see examples in the wild of that [11:27] behavior. And we haven't seen any of [11:29] that in the past year. Now, let's talk [11:31] about this specific example with cyber [11:33] hacking. [11:34] >> I actually think that this one is more [11:37] on the legitimate side. I mean, look, [11:39] the reason why I bring this up is [11:41] anytime Anthropic is scaring people, you [11:42] have to ask, is this a tactic? Is this [11:45] part of their Chicken Little routine? or [11:47] is it real? You know, are they crying [11:49] wolf or not? I actually would give them [11:50] credit in this case and say this is more [11:53] on the the real side. It just makes [11:55] sense, right? So that as the coding [11:58] models become more and more capable, [11:59] they're more capable of finding bugs. [12:01] That means they're more capable of [12:02] finding vulnerabilities. And like one of [12:04] their engineers said, that means they're [12:05] more capable of stringing together [12:06] multiple vulnerabilities and creating an [12:08] exploit. And so I do think that over say [12:11] the next 6 months we're going to have [12:13] this call it one-time period of catching [12:16] up where AIdriven cyber is going to be [12:20] able to detect a whole range of of bugs [12:22] that maybe have been dormant over the [12:24] past 20 years across a wide range of [12:27] systems. And so I do think that there is [12:30] real risk here. And I do think therefore [12:33] that having this pre-release period [12:35] makes a lot of sense where they're [12:36] giving the capability to all these [12:39] software companies that have existing [12:40] code bases to use the tool to detect the [12:43] vulnerabilities for themselves so they [12:45] can patch them before these capabilities [12:47] are widely available. And by the way, it [12:49] won't just be anthropic that makes these [12:52] capabilities available. We know that [12:54] like let's say the Chinese open source [12:55] models like Kimmy K2, it's about 6 [12:58] months behind. So we have a window here [13:00] of maybe 6 months where we're still in [13:03] this pre-release period where I think [13:06] companies that have large code bases can [13:09] get advanced access to this model and uh [13:12] I guess open AI is going to release a [13:14] similar thing in the next few weeks. I [13:16] do think that every company or IT [13:18] department or CISO that is managing code [13:23] bases should take this seriously and use [13:26] the next few months to detect any again [13:29] like dormant bugs or vulnerabilities and [13:32] and roll out patches. If everybody does [13:34] their job and reacts the right way, then [13:36] I do not think it will be the doomsday [13:39] scenario that Anthropic is sort of [13:41] portraying. But it's one of these things [13:43] where the fear might end up being a good [13:45] thing in order to drive people to in [13:49] order to drive the correct behavior. So [13:51] >> sure, [13:51] >> I ultimately think this is going to work [13:52] out fine, but you do need everyone to [13:55] kind of pay attention, use the [13:56] capabilities, [13:58] >> fix the bugs, then we're going to get [13:59] into a big arms race between AI being [14:02] used for cyber offense and AI being used [14:04] for cyber defense, but it'll be a more [14:06] normal sort of of period. Chimath, we [14:09] have uh Daario and uh you know a number [14:12] of the participants here are taking this [14:14] super seriously. They're making a big [14:16] statement. Zach's very nuanced uh I [14:19] think take there. What's your take on [14:21] how do these companies have it both [14:23] ways? Hey, this is shouldn't be [14:26] regulated. This should be regulated. If [14:28] this is in fact a cataclysmic, oh my [14:31] god, they're going to hack everything. [14:33] What if the Chinese have this right now? [14:35] That would speak to more government [14:37] either coordination, regulation, or some [14:39] kind of relationship between the CIA, [14:42] the FBI for domestic stuff, and these [14:45] companies because there it is a non-zero [14:48] chance that the Chinese have an equal [14:50] capability here. We're assuming they're [14:52] behind, but who knows what they're doing [14:54] behind closed doors. So, what's your [14:55] take on this? Is it uh The Boy Who Cried [14:58] Wolf, or is this the real deal? Now, [15:00] >> I think it's mostly theater. [15:03] >> Okay. [15:04] In February of 2019, [15:07] when Daario was still at OpenAI, [15:10] they did the same thing with GPT2. [15:14] That was a 1.5 billion parameter model, [15:17] which sounds like a total fart in the [15:19] wind in 2026. But at that time, this 1.5 [15:24] billion parameter model was supposed to [15:26] be the end of days. And it was supposed [15:28] to unleash this torrent of spam and [15:30] misinformation. And that was the big [15:32] bugaboo at the time. And so what [15:34] happened? They went through this [15:35] methodical roll out over six or nine [15:37] months. They started releasing the [15:39] smaller parameter models and then they [15:41] scaled up to the big 1.5 billion [15:42] parameter model. And at the end of it, [15:44] it was a huge nothing burger. [15:47] If you actually think that Mythos is [15:49] capable of doing what it says it can do, [15:52] two things are true. One is a very [15:55] sophisticated hacker can probably do [15:57] those things right now with Opus. [16:00] And two, if these exploits [16:04] are this easy to find, [16:07] whether you use Opus or whether you use [16:09] Mythos, the reality is you'd have to [16:10] shut down the internet for about 5 years [16:12] to patch them all. So when you see like [16:15] a large multi- trillion dollar gang, [16:20] it's a bit of theater. Why? What do you [16:22] think they can actually accomplish in 2 [16:24] months? Do you actually think that if [16:26] there's these vulnerabilities, it's all [16:28] going to get fixed? Let's give them six [16:30] months. Let's give them nine months. But [16:33] the reality is that capitalism moves [16:35] forward, the funding needs moves [16:37] forward, and the need for these guys to [16:40] build adoption moves forward. And that's [16:43] going to supersede [16:45] what this is. So I do think that Sax is [16:47] right that they have figured out a very [16:50] clever go-to market muscle here and a go [16:54] to market motion that activates [16:57] hyper attention and hyper usage and so I [17:00] give them tremendous credit and I'll [17:02] maintain what I've maintained before. [17:04] Anthropic is shooting the lights out [17:06] right now. This is like Steph Curry [17:08] going bananas from every everywhere on [17:10] the court. These guys are hunking [17:12] threes. [17:13] >> It's all in that. Okay. So huge kudos to [17:16] Anthropic, [17:18] but we've seen it before. We saw it when [17:21] these folks were the principal [17:23] architects at OpenAI who are now seeing [17:25] the same playbook here. I think we'll [17:27] look back and I think what we'll say are [17:30] these two things. One is if we're really [17:31] going to patch all these security holes, [17:33] we need to shut down the internet [17:35] for some number of years, honestly, [17:37] literally years. And the second is an [17:40] advanced hacker can probably do this [17:41] today with Opus if they really wanted [17:44] to. [17:46] >> Okay. Hey Brad, I gota I'll get you in [17:47] here for the for the last word. I I'm [17:49] going to go with Yeah, maybe they did uh [17:52] Crywolf before, but based on what I see [17:55] with these models advancing and using [17:57] them and I'm using a lot of the open [17:58] source ones right now from China. I [18:00] think that this is like code red kind of [18:03] moment. This is Defcon. like we should [18:05] be taking this deadly seriously and I [18:07] think these companies got to coordinate [18:08] with the CIA and this is uh equally a [18:11] defensive as offensive opportunity. Do [18:14] you think this [18:15] >> you're asking for the nationalization of [18:16] AI now? [18:17] >> No, I actually I I I don't think it [18:20] should be nationalized. Um although I [18:21] did see people sort of insinuating that. [18:24] I think these companies need to build a [18:26] group Brad that work and coordinate with [18:28] the CIA. I assume that they're already [18:30] doing this. I'm assuming you Emil [18:32] Michael and uh you know Trump and [18:35] everybody have these people in a room [18:36] and that they've given the defcon and [18:39] said hey how can our government use this [18:41] to stop bad actors and this is already [18:44] being coordinated with the CIA and the [18:45] FBI. I am 100% certain of that that [18:48] Dario went to them and said look what we [18:50] found. This is the real deal. I'll give [18:52] you the last word on this Brad since [18:53] you're an investor in both companies and [18:55] you know them quite well. the frontier [18:56] model forum which was which was put [18:58] together in 23 um is cooperating on [19:01] anti- and adversarial distillation stuff [19:04] as we speak right they don't want to [19:06] make it easy on you know so Google and [19:08] and open AAI and and anthropic they're [19:11] coordinating on this stuff you know [19:13] there are times where I've pushed back [19:14] on anthropic because I thought it was [19:16] you know perhaps regulatory capture or [19:18] something else this is very different in [19:20] my mind right he could have easily Dario [19:22] could have easily come out and said oh [19:24] my god we passed a threshold we need to [19:25] have a government moratorum. Remember, [19:28] even our friend Elon called for a [19:29] six-month moratorium in 2023 because of [19:32] civilization risk. This guy didn't do [19:34] that. Instead, he said, "Okay, what what [19:37] should we do? I'm going to get 40 of the [19:38] leading companies together. We're going [19:40] to spend a 100 days sandboxing, [19:42] hardening the systems, and then we're [19:43] we're we're going to keep pushing [19:44] forward." [19:45] >> What do you honestly think is going to [19:46] get accomplished in a 100 days? How many [19:48] PRs you think are going to get pushed to [19:50] the core structural internet in 100 [19:52] days? What's the overunder number? Cuz [19:54] I'll give you a number. You're gonna say [19:55] zero. My my answer to that is [19:57] >> I'll say like 10,000, but it's going to [19:59] be immediate. [20:00] >> But if it prevents your browser history [20:02] from being released to everybody in the [20:03] world, Chimat, that may be something [20:05] that you're willing to, you know, let a [20:07] 100 days pass on. [20:08] >> I think you got Chimat's attention when [20:09] you said browser history. [20:10] >> What about the dickpicks? [20:14] >> As Chimat is, he's going to release them [20:16] himself right now. CHIMAT'S LIKE, "HEY, [20:18] CHINESE HACKERS, HERE ARE MY DICKPICKS. [20:20] Please put them out." [20:21] >> Oh my god. we have to be out there [20:24] complimenting when they're doing the [20:25] right things or relying on the market [20:26] rather than running to the nanny state [20:28] and saying do more of this. So this to [20:30] me was just an example of of a of a good [20:32] balance. I'm sure we're going to have [20:34] plenty of debates about this in the [20:35] future. But you know this is one I would [20:37] like to see more of. [20:38] >> This is why to use your word Jake I [20:40] tried to have a more nuanced take is [20:42] because we have no choice but to take [20:44] this seriously. Whether it's total [20:46] theater, whether it's fear-mongering, [20:49] and they do have a pattern around this, [20:51] we can't take the risk, right? And it [20:53] does logically make sense that as these [20:56] models become more and more capable at [20:58] coding, they're going to get better at [20:59] cyber. And there's going to be that one [21:02] time period where you're moving from [21:04] preAI to post AAI, and you need a patch [21:06] for that. So, my guess is we're going to [21:08] see a lot of patches over the next few [21:10] months. I think that that will resolve [21:12] the problem. [21:14] I think this is a case where I'm going [21:16] to give them the benefit of the doubt. I [21:18] I think that, you know, I've criticized [21:20] him in the past. I think that blackmail [21:22] study was embarrassing to the level of [21:25] being a hoax, but I think in this case, [21:28] I'm going to give him credit and say [21:29] that I think that it's legit. [21:31] >> So, it's not the anthropic hoax. This [21:33] could be legit. I, you know, looking at [21:36] >> we have no choice but to treat it that [21:37] way. [21:38] >> Of course. Yeah. I mean, even if two [21:41] things could be true at the same time, [21:42] Saxs, they could have used this tactic [21:44] before. It could be performative, like [21:46] the video with the dramatic music in the [21:48] background. It does have a little bit of [21:50] drama to it, and the way they presented [21:53] it is very dramatic, but it does make [21:55] logical sense that the one company that [21:59] made the bet on code bigger than anybody [22:02] else would be the one who would discover [22:04] this quickest. And you know in a 100 [22:05] days that's a pretty good um that's a [22:08] pretty big advantage versus the hackers. [22:10] But let me think one more point there [22:11] Jimat [22:12] >> the most important thing that people [22:14] haven't talked about here is the amount [22:17] of code being pushed right now because [22:18] of these tools is 10x 100x in most [22:21] organizations. So we need to have this [22:24] type of security embedded in these new [22:26] coding tools to do it in real time. [22:29] That's the opportunity. There should be [22:31] real time correcting of this. If this [22:33] was real, they picked the wrong [22:34] companies. Meaning, there are energy [22:38] companies, folks that control nuclear [22:41] reactors. There are airplane companies [22:44] that are flying hundreds of thousands of [22:45] people in essentially manufactured [22:48] missiles of like streaming gas going at [22:52] 500 miles an hour. None of those [22:54] companies were the ones that were [22:55] included in this. And so I think if you [22:58] really thought that this was end of [23:00] days, [23:02] at a minimum we can agree maybe we [23:04] should have expanded the circle a touch. [23:07] Well, maybe those are customers of the [23:09] ones they're including here. Anyway, uh [23:11] this is a really important story. We'll [23:12] obviously track it in the coming weeks [23:14] to see what turns out to be reality. And [23:17] uh Daario, do come on the program at [23:19] some point. Hey uh Brad, will you get [23:20] Dario to come on the program? I've [23:21] invited him like three times. I got his [23:23] phone number. He's ghosted me. I don't [23:24] know why. [23:24] >> Wait, he he's ignored you? I get [23:27] >> I literally got an introduction from the [23:29] number like one of the number one [23:30] venture capitalists in the world. He's [23:31] on the cap table very early. He just [23:33] won't respond. I don't know why. [23:35] >> I would tell you Daario's podcast with [23:38] Dwarkish who I think is an excellent [23:39] podcaster. I've listened to that three [23:41] or four times taken notes every time. It [23:44] is a really exceptional piece really [23:46] exceptional piece of work by by by them. [23:48] >> All right, let's keep moving. We got a [23:50] lot on the job. [23:50] >> You may once again be tarred with your [23:52] affiliation with us. [23:54] >> Poor you. I mean, I don't care. [23:56] Literally, I I've got friends on both [23:58] sides of the aisle. I have friends [24:00] >> of course you do. [24:01] >> Even JCAL. [24:02] >> Even JCAL has friends everywhere. [24:04] >> Let me ask Brad a question here just [24:06] while we're on the topic of anthropic. [24:07] There was a really interesting story or [24:10] tweet I guess you could say by the [24:11] founder of OpenClaw [24:13] that [24:14] >> Peter. [24:14] >> Peter. Yeah. What's his name? Peter [24:16] Steinber. [24:17] >> Steinberger. Steinberger. [24:18] >> Steinberger. Yeah. [24:19] renowned coder created openclaw which is [24:22] kind of the thing that launched this [24:24] whole agent era now you I guess you [24:25] could say any event he said that [24:28] anthropic was cutting off his access to [24:31] was it was to claw is that the next [24:33] topic [24:34] >> this is on the docket it's a little bit [24:36] nuanced [24:37] everybody using openclaw would take [24:39] their $200 a month subscription to [24:41] anthropic which was essentially like a [24:43] people were using more tokens and it's [24:45] an average the people from openclaw it [24:48] is very verbose and those people are [24:51] 100x the usage of the average [24:53] subscriber. So he said you can't use [24:56] your 200, you have to use the API. You [24:58] move from the $200 plan to the API, add [25:00] a zero to your token use. So or more. [25:03] And so they essentially anchored [25:06] Open Claw and then 10 days later or less [25:11] they released or announced their new [25:12] agent technology which is according to [25:14] them a safer better version of OpenClaw. [25:17] So, hey, all fair in love and war and [25:20] they have basically shot a huge cannon [25:23] across the bow of openclaw. [25:25] >> Wait, can you just explain that exactly? [25:27] So, so I think you're right that they [25:29] systematically copied feature by feature [25:32] >> of open claw, incorporated that into [25:34] clawed and then the coupross was [25:36] basically cutting off open [25:38] >> oxygen. Can you just explain exactly [25:40] what they did? Okay, very simply, when [25:43] you buy a subscription to these [25:44] services, they have blended your usage [25:48] across many users. So there's, you know, [25:50] nine out of 10 users use less than the [25:52] tokens they're paying for and the top [25:54] 10% use much more. When OpenClaw became [25:58] a phenomenon, the number one open source [26:00] project in history on GitHub with all of [26:02] this usage, people went crazy. And you [26:05] heard me talking about how crazy I went [26:06] for it. those people with the $200 [26:08] subscriptions were using $2,000 $20,000 [26:11] worth of tokens. So they said you can no [26:14] longer use your subscription to, you [26:16] know, either your professional or [26:18] enterprise subscription at $200 [26:20] >> and plug that into your open claw. You [26:22] now have to go to the API and pay per [26:24] usage. So no more like [26:27] >> unlimited. If you use Anthropic's own [26:29] agent harness, are you part of the [26:31] bundled flat rate? You can assume that [26:34] that's what they'll do, which if you [26:35] were thinking on an antirust level might [26:38] be token dumping or price dumping. I'm [26:40] not saying like I'm ratting them. [26:42] >> No, it's like bundling, isn't it? [26:43] >> Well, price dumping or bundling. When [26:45] you price something under the market [26:47] price in antitrust, that would be price [26:49] dumping, right? And if you were to [26:51] bundle, it would be like the bundling [26:53] issue. [26:54] >> Critically important. You can use [26:56] openclaw via claw API and every company [26:58] has a right to set the price for its [27:00] products. It's just saying that you were [27:02] for under their current regime, they [27:04] were selling dollars for 10 cents via [27:06] OpenClaw because these were such power [27:08] users and now they're just saying we [27:10] have to price this rationally, but we're [27:11] happy to have you guys use the API. So, [27:13] >> okay. Okay. But Brad, when you use the [27:16] OpenClaw competitor that Anthropic now [27:18] offers, [27:19] >> correct? [27:20] >> Are they subsidizing that? Are you [27:22] paying? [27:22] >> We don't know yet because it's in closed [27:24] beta. So in other words, what I'm saying [27:25] is if they charge for API usage, right, [27:29] their own first party agent harness or [27:32] system, then that would be apples to [27:33] apples. But if [27:35] >> if they end up charging the bundled flat [27:38] rate, let's say, for their stuff, but [27:41] then charge the metered rate for third [27:43] party stuff, you could make a bundling [27:45] argument. [27:46] >> Sure. Sure. And you could say it's [27:47] anti-competitive assuming that Anthropic [27:50] has dominant market share in coding, [27:52] which I think most people would say they [27:53] do at this point. [27:54] >> And assuming that it's the same product, [27:56] I mean, the reason most enterprises will [27:58] probably use the Anthropic uh version of [28:03] this agentic product is because it meets [28:05] all of your security parameters, right? [28:07] So, Altimter runs, you know, a lot of [28:10] stuff on Enthropic. They're already [28:12] integrated within our our data [28:13] warehouse, our data lake, things of that [28:15] nature. So just letting openclaw loose [28:18] on the uh altimeter you know data set [28:20] would not be wise and so it's a [28:22] different fundamental product. [28:24] >> No I get that and I think that anthropic [28:26] has a huge advantage let's say cloning [28:28] open claw and just building it into [28:30] claude. I'm not denying that to me that [28:33] would be the reason why they don't need [28:34] to do price discrimination is because [28:36] there's already a very good reason [28:38] >> to use the let's call it the bundled [28:40] offering on a featured basis. But the [28:42] question I'm specifically asking is [28:44] whether they're giving themselves a [28:46] price advantage because [28:48] >> I think Brad is giving the the most [28:50] generous interpretation. You're taking a [28:52] more cynical one. I'm with you, Saxs. [28:53] I'm 100% on the cynical side. Open Claw [28:56] is so powerful. It's got so much [28:58] momentum that not only is anthropic [29:02] trying to ankle it. I believe when Sam [29:04] Waltman bought it, it was uh and he [29:06] didn't buy OpenClaw itself, he hired [29:08] Aqua hired Peter. I believe it was to [29:10] subvert the open- source project to get [29:12] Peter's next set of genius ideas inside [29:15] of OpenAI as opposed to letting them go [29:17] there. People are going to say I'm a [29:19] conspiracy theorist, but this is the [29:21] number one focus and let me just give [29:23] you a list of who is trying to kill [29:25] OpenClaw/compete with them. Obviously, [29:28] you have Anthropic, but also Perplexity [29:31] Computer launched. It's awesome. I've [29:33] been using it. Anthropic has this clawed [29:36] managed agents. They dropped that on [29:38] Wednesday, April 8th. uh yesterday uh [29:40] today's Thursday when we tape you you [29:41] guys listen on Fridays and then you have [29:44] Hermes agent that was released on [29:47] February 25th that's also open source [29:49] and very good so that's in the open [29:50] source camp Alibab is coming out with [29:52] one that's going to be based on their [29:54] Quinn model then you have Elon who said [29:57] he's got something called rock computer [30:00] coming out of macro hard which is a play [30:01] on words for Microsoft in addition to [30:04] that Amazon and Apple are preparing uh [30:06] new releases of their uh maxing [30:09] assistants Alexa and Siri that will be [30:11] less in this new version and [30:14] then nothing out of uh SAT and Microsoft [30:16] yet. So the number one goal I believe in [30:20] the large language model frontier model [30:23] space is to kill this open source [30:25] product. [30:26] >> No, I mean come on like why they're [30:29] building multi-functioning agents that [30:31] can move from answering questions to [30:33] actually doing something for you. like [30:35] you got to do that because that's what [30:37] consumers and enterprises wants. It [30:39] doesn't mean that it's about killing [30:40] OpenClaw. just this is an obvious thing [30:43] right to do it [30:44] >> but this is a giant movement to stop it [30:47] because this is the equivalent of having [30:49] an open-source Android like player in [30:51] the market and that could be incredibly [30:53] disruptive these I believe open source [30:55] is going to win the day on the large [30:56] language models and take 90% of the [30:58] token usage and I think the entire [31:00] frontier model space could be undercut [31:02] by open source and I think they realize [31:04] that SLMs the the smaller language [31:06] models that are verticalized now that [31:09] will run on you know, desktops and [31:11] laptops and is even starting to run on [31:13] the top ones. That is their biggest [31:15] competitive threat and I hope it [31:16] happens. All due respect to your [31:18] investments, Brad, I think this [31:20] technology and the interface is uh you [31:22] know, he placed bets, but I I think it's [31:24] imperative that the agent level, which [31:26] is essentially your entire life, you [31:29] don't give that to Anthropic. You don't [31:31] give that to OpenAI. That's your entire [31:33] business, your entire life. It is [31:34] foolish for you, Brad, to give your [31:36] entire business and all the knowledge [31:38] you have to anthropic through that. [31:40] unless you're just doing it to boost [31:41] your um your your investment in those [31:43] companies. But I would be very concerned [31:45] if I was you with putting all of your [31:48] knowledge that you've earned over a [31:49] lifetime into any of these large [31:51] language models. [31:52] >> All right, Jake, let me ask you. Can I [31:54] ask a question? Thank you for that [31:55] impassioned monologue. Um actually, I [31:58] want to ask my TED talk. [31:59] >> I Yes, thank you for that TED talk. Um I [32:02] have a yes no question for each of you. [32:06] >> Do you believe that anthropic has [32:07] dominant market share in coding? right [32:10] now? Yes. No, [32:13] >> no. [32:14] >> In in coding, [32:15] >> yes, [32:15] >> they had the lead, but not that they had [32:16] the lead, but not dominating. [32:18] >> I think it's a trillion dollar market, [32:19] and these guys have less than 10% of it [32:21] today. So, it's hard to make a case that [32:24] >> What percent of coding tokens do you [32:26] think that anthropic is providing the [32:28] market right now? [32:29] >> Greater than 50%. [32:30] >> Yeah, that's true. [32:31] >> Okay, that's called dominant market [32:32] share. [32:33] >> Uh, I don't know about that. [32:35] >> More than 50% of the market. [32:36] >> You got to look at what you got to look [32:37] at what the TAM is. [32:39] with the Tan, [32:41] >> right? There are a lot of people who [32:42] provide, you know, that that are in this [32:46] tiebreaker before we move on to the [32:47] next. [32:47] >> I'm not saying it's a permanent [32:48] condition, [32:49] >> but if you're telling me that today [32:52] >> Anthropic is delivering over half of the [32:55] coding tokens, that's clearly a dominant [32:58] position in the market for coding. It's [32:59] an early market. It could change, but [33:01] >> if I were representing them, David, I [33:03] would say nine months ago, everybody t [33:05] called us uh, you know, out of the game. [33:08] We were being destroyed by open AI in [33:10] three months. Now people are saying we [33:11] have dominant market position. This is [33:13] the fastest changing most competitive [33:16] market in the world. I think it would be [33:17] very hardressed to walk into, you know, [33:20] some district court make the case that [33:21] these guys have somehow already formed a [33:23] monopoly against Amazon, Google, [33:26] Microsoft, Open AI, etc. [33:29] >> Well, I'm not saying it's a it's already [33:30] a permanent monopoly, but I am just [33:33] asking about market share. And I do [33:34] think you guys all agree that [33:36] >> Shimov, go ahead. [33:37] >> They probably have 50 to 60% market [33:40] share because I think codeex is actually [33:42] quite broadly used as well. [33:45] >> But that belies the more important point [33:47] which is AI enabled coding I think is [33:51] still 5% of the broad market. So it's [33:53] kind of a nothing burger. Yes, they're [33:55] leading but they're leading in something [33:57] that isn't that big yet. Now you would [33:59] say how could it not be big? And what I [34:01] would say is because most of the stuff [34:03] that's being written is still white [34:05] sheet denovo code. And I think the ugly [34:09] truth is I don't care what model you [34:11] have, but the long horizon ability for [34:14] any of these models to actually build [34:16] enterprisegrade software is still [34:20] shiit And that's the actual lived [34:24] experience. Not for me, but when I call [34:26] on our customers, half a trillion dollar [34:29] banks, hundred billion dollar insurance [34:31] companies, none of these guys are like, [34:32] "Wow, it just works out of the box." It [34:34] doesn't work. So, most of it is still [34:38] handtuned. So, until I can honestly tell [34:41] you that we can point a model at this [34:44] with the right guard rails, which I [34:46] can't today, what I would say is it's a [34:48] small market that will become large as [34:52] these models become better. [34:54] But we are in the world where we have 50 [34:57] years of accumulated tech debt as a [35:00] world. And I suspect when you enumerate [35:02] the number of lines that that [35:04] represents, it's hundreds of trillions [35:06] of lines of just pretty marginal [35:08] mediocre code to bad code. On top of [35:12] that, we have all these legacy [35:13] languages. I'll tell you one of our [35:15] customers, they have to go and get [35:17] 60year-old pensioners to come into the [35:19] office to interpret cope. No, I'm not [35:22] joking. This is a [35:22] >> snowball for trend. [35:24] >> This is a hundred billion dollar a year [35:26] revenue company and that's how they [35:29] solve these problems. It's not opus just [35:31] solves it. So I I would just keep in [35:34] mind that most of the tech debt in the [35:36] world that exists 99% of it is still [35:39] poorly addressed by these models. We are [35:42] untying this Gordian knot. It's going to [35:44] take decades to do it right. So all the [35:46] breathlessness about all this other [35:48] stuff, I really think it's not where the [35:49] money is. It's not the big time stuff. [35:51] And you can tell me, "Oh yeah, it's [35:53] going to be the future." And I would [35:54] say, "Tell this business that's a [35:56] hundred billion dollars a year of [35:58] revenue and 50 million billing [35:59] relationships that all of a sudden [36:01] you're going to open claw your way to a [36:02] solution." It's Not to say [36:05] that you can't have a great chief of [36:07] staff, and not to say you can't do some [36:09] useful stuff and trickery and, you know, [36:11] have a good knowledge base. I'd like [36:13] that, too. But the core things that your [36:17] lived experience sits on today is a mess [36:20] of tech debt that will get very slowly [36:22] replaced. And that's just the reality of [36:24] life. [36:25] >> And there are competitors that are [36:27] extremely disruptive. I'll tell you [36:28] about one. We talked about Bit Tensor [36:30] Tao on this program a couple weeks ago [36:32] when we had the um Jensen interview. You [36:34] brought it up actually Chimath. There's [36:35] a there's a project that's subnet 62. [36:38] It's called Ridges AI. And what they're [36:41] doing is a competitor that is not only [36:44] open- source but anybody can contribute [36:46] to it. They spent about a million [36:48] dollars in tow like rewards and in 45 [36:50] days they hit 80% of what Claude 4 is [36:54] and they did that in under 45 days. The [36:56] way that works is they give rewards for [36:58] people who and they can do this [37:01] anonymously make that coding product [37:03] which is like codeex or claude code [37:05] better. that flywheel is racing right [37:09] now with participation in the same way [37:10] Bitcoin is. So you're going to see a lot [37:13] of open- source and these crypto [37:15] open-source combinations and uh anybody [37:18] who's not investigated this, I highly [37:21] recommend you investigate this. [37:23] >> I do think you're right about one [37:24] specific thing. I would put zero, [37:26] literally the probability zero of any [37:29] important company worth anything more [37:31] than a dollar having and outsourcing [37:34] their production code to an open source [37:36] project. That'll never happen. However, [37:37] what will happen though is when you look [37:40] at the cost of training this 10 trillion [37:43] parameter model on Blackwell and when [37:47] you look in the future let's just say in [37:50] six or nine months that a 15 or 20 [37:52] trillion perm model is going to get [37:54] trained on Vera Rubin I think Jason [37:56] where you are right I have zero and just [37:59] to be clear I have no investments in [38:01] this at all I'm [38:03] >> to be so super clear [38:04] >> I'm just observing because another [38:06] project other than Bit Tensor that [38:07] someone brought up to me is Venice. The [38:09] concept of opensource training and [38:12] orchestration [38:14] is a hugely disruptive idea which is the [38:17] complete orthogonal attack vector to [38:20] this idea that you have to raise tens [38:22] and tens of billions of dollars to train [38:24] your models because if the capital [38:26] markets run out of 10 and 20 billion [38:29] dollar checks to give people the only [38:32] solution is to be totally distributed. I [38:34] tend to agree with you Jason that there [38:35] is going to be at some point a very [38:38] successful open source project for [38:40] pre-training. [38:42] Absolutely. Will there never ever be an [38:45] open- source way where a real company [38:47] that has any skin in the game says here [38:49] guys re-engineer my codebase as an open [38:51] source project. Never going to happen. [38:53] >> Yeah, I I think the coding tools will. [38:55] And if you look at the history of open [38:56] source, Brad, you actually I think had a [38:58] lot of bets in this space. Linux, [39:00] Kubernetes, Apache, Postgress, like [39:03] Terraform, like these open source [39:05] projects are deep inside of enterprises. [39:07] Deep. And we're sitting here 15, 20 [39:10] years ago, the same argument was made. [39:12] Nobody will ever adopt these inside the [39:13] enterprise. You got to go with Oracle, [39:15] whatever. And fair enough, many people [39:17] do. But I think this is this $29 ridges [39:22] um subscription to do this versus 200. [39:24] It's starting to take hold inside of [39:27] startups. And that's where I always look [39:29] at the tip of the spear. Startups love [39:30] to, you know, use open source products. [39:33] I think this could be the next big [39:34] thing. But listen, I I I invest in [39:37] things that have a 90% chance of going [39:39] to zero. So do your own research. No [39:41] crying in the casino. [39:43] >> Can I just make a a final few points? So [39:46] just just quickly so number one is with [39:49] respect to this market for code or code [39:51] tokens whatever you want to call it [39:53] >> it might be 5% today meaning 5% of the [39:56] codes AI generated versus human [39:57] generated I think it's going to 95% I [40:00] mean I bet any amount of money on that [40:02] the only question is when probably over [40:04] the next few years so that's point [40:06] number one point number two is it's [40:08] possible that if you're the early leader [40:11] in coding as a AI model company let's [40:15] say you have 50 to 60% of market share. [40:17] You have the most developers using it. [40:19] Therefore, you have the most access to [40:21] code bases. You might get the most [40:24] training tokens. There is a potential [40:26] flywheel there where you can see the [40:29] early market leader consolidating its [40:30] lead because it's generating the most [40:32] code tokens and it's getting access to [40:34] the most existing code. Now, I'm not [40:36] saying for sure that's going to happen. [40:38] is possible that the other guys catch [40:39] up, but I think there is a possibility [40:42] of a flywheel there and strong, I guess [40:45] you'd call it data scale effects, things [40:47] like that. So, I do believe that the [40:49] market for coding tokens could be [40:51] monopolized. Third, Anthropic's revenue [40:55] run rate, as based on what I can tell [40:56] and what's been publicly released, is [40:58] the fastest growing revenue run rate at [41:01] scale that I think we've ever seen. Uh, [41:03] we [41:04] >> perfect segue. It's the next story. [41:05] Okay, maybe [41:06] >> pull up the the tweets. But this thing [41:09] is ramping at a rate we've never seen [41:10] before. [41:12] >> We can get into that in a second. But [41:13] just one last final point [41:15] >> is I think it's pretty clear that where [41:17] we go from here is agents and coding [41:21] gives you a huge step up on agents [41:23] because you know one of the main things [41:25] that agents need to do is is write code [41:27] to be able to enable them to complete [41:29] tasks. [41:30] >> Correct. And so if it is the case that [41:34] coding is this huge market that's going [41:36] to be dominated by one or two companies [41:40] and then that leads to another huge [41:42] market which is agents. My point is just [41:44] I think all these companies need to [41:46] behave in a very clean way [41:49] >> and not engage in tactics that later the [41:52] government might say you know what that [41:53] was anti-competitive. Everyone should [41:55] just I think play fair. Do not engage in [41:57] discrimination against other people's [41:59] products. engage in fair pricing. I'm [42:02] not accusing anyone of breaking any of [42:03] the rules, but what I'm saying is that [42:05] eventually the government's going to [42:07] look at this market with the benefit of [42:09] 2020 hindsight and I think everyone [42:11] should just basically, you know, keep it [42:14] >> keep your nose clean. [42:15] >> Keep it tight. Keep it tight. [42:17] >> Keep it tight. Tight is right. I think [42:19] is an excellent point. Let's talk about [42:21] the revenue ramp of Anthropic. This is [42:24] just unprecedented. Anthropic's revenue [42:27] run rate has topped 30 billion with a B. [42:31] Early 2023, they turned on revenue. They [42:33] started charging for API access. End of [42:35] 2024, they're at a billion dollar run [42:37] rate. February 25, they launched Claw [42:40] Code. That was the starter pistol. Mid [42:42] 2025, $4 billion run rate. End of 2025, [42:46] $9 billion run rate. Just a couple of [42:49] months later in April, $30 billion run [42:52] rate. Yes, that's right. Triple. Uh and [42:54] the way they did this is enterprise uh [42:57] customers are a major part of the spend. [43:00] Dario announced a couple of months ago [43:01] that there's over a thousand enterprises [43:04] paying over 1 million annually. This is [43:07] truly mindboggling when you think about [43:09] it because those are the most coveted [43:12] customers in the world. These are the [43:13] big fish that you just uh when people [43:16] are running enterprise software, they [43:18] they dream Slack dreamed of getting [43:20] these million-dollar customers. Uh [43:21] Salesforce dreams of getting these [43:23] million-dollar customers. Brad, you're [43:24] an investor. I guess uh Sam famously on [43:27] BG2 asked you to sell your uh OpenAI [43:30] stock back to him. You didn't. You [43:32] demired, but you're an investor in both. [43:36] How shocking is it to you to place both [43:39] of those bets and then see one of them [43:42] come from so far behind? You know, Chat [43:44] GPT has 900 million users. I don't know [43:46] if they've they've passed a billion [43:48] officially yet, but they are the Verb, [43:50] right? They're the Uber. They're the [43:52] Xerox. They're the Polaroid of AI, but [43:56] they didn't go after the enterprise. [43:58] Daario made that and Daario worked. He [44:00] was the co-founder of OpenAI. He left [44:03] and according to the New Yorker story [44:04] that came out from Ronan Farrell this [44:06] week, he was basically left because of [44:08] his disgust in working with Sam Alman. [44:13] Your thoughts? [44:14] >> Well, you know, before we go down the [44:15] OpenAI rabbit hole, let's just really [44:17] contextualize like what's going on here. [44:20] You know, check I I I have this [44:21] additional chart. you showed one, you [44:23] know, they added 4 billion of revenue in [44:25] January, 7 billion in February, 11 [44:28] billion of annualized run rates, um, or [44:31] 10 or 11 billion in March, just to put [44:33] in perspective, that's data bricks plus [44:35] Palanteer combined that they added in a [44:38] single month, right? So we started with [44:41] everybody at the start of the year [44:42] ringing their hands including you know [44:45] Gurley and others saying we're in a big [44:46] bubble asking whether the AI revenues [44:49] would show up to justify all of this [44:51] investment and bam you have the largest [44:53] revenue explosion in the history of [44:55] technology. So the company's plans were [44:58] to end the year at about a $30 billion [45:01] ex exit run rate. They got there by the [45:04] end of March right and I suspect that [45:06] it's continuing in April. So you have to [45:08] ask what's going on and what's the big [45:10] so what the first thing for me is that [45:13] model and product capability just hit [45:15] this threshold we talked about earlier [45:17] near AGI whatever the hell you want to [45:19] call it and everybody like alimter said [45:22] damn this is so good I have to have it [45:24] this is no longer about my IT budget [45:26] this is about labor augmentation and [45:28] labor replacement and by the way co-work [45:31] is growing even faster than Claude go at [45:35] the same stage of development [45:38] So what it showed is we have a near [45:40] infinite TAM. It turns out that the TAM [45:42] for intelligence is radically different [45:44] than anything that we've seen before. [45:47] And I think the best example of this, [45:49] right? This is millions of [45:51] self-interested parties, consumers, [45:54] enterprises, a thousand now over a [45:57] million dollars. Right? It's not that [45:59] there was some great go to market and [46:01] anthropic that all of a sudden, you [46:02] know, they snuck up and blew everybody [46:04] away. No, it was companies demanding the [46:06] product. They're getting throttled on [46:08] the product. Why? Because it's so good. [46:10] It makes them better at their business. [46:12] We are all self-interested actors. And [46:14] when millions of those people are all [46:16] making the same decision, there's a huge [46:18] tell. And the tell here is that the TAM [46:21] is as big as Daario and Sam and others [46:24] have been saying. We knew intelligence [46:26] was going to scale on the exponential. [46:28] The question was whether revenue will [46:30] scale on the exponential, and that's [46:32] what we're seeing. And remember, they're [46:33] doing this with only 1 1/2 to 2 gawatt [46:37] of compute, right? These guys are [46:39] massively compute constrained. They're [46:41] each going to be adding 3 GW of compute [46:43] this year. And so that will unlock they [46:46] would be growing even faster. But for [46:49] that, and then Jason, to your point [46:50] about the open source models that we all [46:53] want to be a part of this solution, I've [46:55] talked to a lot of big companies, 65 to [46:57] 70% of their token consumption is [47:00] open-source model, right? are these [47:02] cheap Chinese and other tokens. So these [47:05] revenue ramps are happening while the [47:07] world is already using open source. This [47:10] is not frontier only. This is Frontier [47:12] plus open source. We're going to see [47:14] massive token optimization over the [47:16] course of the year. But what happens on [47:18] this Jebans paradox is the co the unit [47:21] costs right of intelligence is [47:23] plummeting. Not the cost of tokens. The [47:26] unit cost of intelligence is plummeting [47:28] because the capabilities of these models [47:30] is so much better. I look at what it [47:32] does for Altimeter day in and day out. I [47:34] talked to a major uh company yesterday. [47:37] They're on a run rate to do a hundred [47:39] million of token consumption this year [47:41] on about $5 billion in opex. They think [47:44] that we're now nearing peak employment [47:46] in their company, but that their token [47:48] their intelligence consumption, okay, [47:50] let's not call it token consumption, [47:52] right? because tokens may go up a lot, [47:54] but their intelligence consumption is [47:56] going to go up, you know, a lot. So, I [47:59] would leave you with this. We're early [48:01] to Chimas's point. We have low [48:04] penetration of the global 2000. We have [48:06] low penetration of the use cases. We [48:09] have low penetration of of within the [48:12] use cases that they're already using. [48:14] And the models are only getting better. [48:15] So I think when you look out toward the [48:17] end of the year, I would not be shocked [48:20] if you see Anthropic exiting this year [48:23] at 80 to 100 billion in revenue. And by [48:26] the way, doing it at the same time that [48:28] OpenAI, who is also on the wave, they'll [48:30] be releasing an incredible model in the [48:32] next imminently. They're going to be on [48:35] that wave and you're going to see an [48:36] inflection in their revenues as well. [48:38] >> Okay, Chimath, question one has been [48:41] answered. The question of hey, does this [48:43] stuff actually have utility? that went [48:45] from a question mark to an exclamation [48:46] point. Of course, it's got utility. [48:48] People are getting value from it. And it [48:49] might be variable. Some people get more [48:51] value than others. Number two, the [48:52] revenue ramp was a big question. Now, [48:54] that's turned into an exclamation point. [48:56] The final piece of the puzzle that [48:58] you've brought up many times is can this [48:59] be profitable? And these companies are [49:02] burning through a large amount of cash. [49:05] So, what is your take on when these [49:07] companies can get out of the J curve? We [49:09] talked about this, I think, three [49:10] episodes ago. I estimated like we're [49:12] going to be looking at $4500 billion in [49:15] investment into these data centers at a [49:17] minimum and then they have to climb out [49:20] of that to get to profitability. So what [49:22] are your thoughts on these becoming [49:25] profitable companies? Do you remember [49:27] the [49:29] investor that published this list Jason [49:31] where he put all of the terms you talk [49:35] about when one of the terms you can't [49:36] talk about is profit. It's a list where [49:38] it's like if you can't talk about free [49:41] cash flow, you talk about IBIDA. When [49:43] you can't talk about IBIDA, you talk [49:44] about [49:46] >> margina. [49:47] When you can't talk about that, you talk [49:49] about revenue. And then when you can't [49:50] talk about revenue, you talk about gross [49:53] revenue [49:53] >> bookings. [49:54] >> So you can kind of figure out, [49:58] I think, where we are in any part of any [50:01] cycle by just indexing into what does [50:04] everybody talk about. I think where we [50:07] are is we are between gross revenue and [50:11] net revenue. That's where the discussion [50:12] is. [50:13] >> Okay. [50:14] >> There was another article I think today [50:16] in I think maybe it was the information [50:18] that tried to categorize and distinguish [50:21] that anthropic presents gross, open AI [50:24] presents net. They're different. We [50:27] don't know what the various take rates [50:29] are. So they're saying that there's a [50:32] difference. If it's not true, there's [50:33] been no clarity provided by these [50:35] companies. So, at a minimum, you have [50:37] this confusion where there's the [50:39] breathless talk. Then there's people [50:40] that don't even know the difference [50:41] between actual recognized revenue and [50:44] run rate revenue and how to multi. I [50:46] mean, so we're definitely there, okay? [50:48] We can quibble about the details, but we [50:50] are not at the place where people are [50:51] like, "Oh, here's your steadystate, you [50:53] know, free cash flow margin, and here's [50:55] what your EBA does." We're never we're [50:56] we're years from that. They're gonna [50:58] have token maxing IBA like IB at the [51:02] Wii. [51:02] >> The thing that we need to understand is [51:04] how gross margin negative is this [51:06] revenue growth. [51:07] >> We don't know that and at least we don't [51:10] as outsiders. [51:11] >> Brad might know. [51:12] >> Brad may know. I I I I will tell you [51:15] think about this. What are their big [51:16] cost inputs? The number one cost input [51:19] is the cost of compute. Cost of compute. [51:21] >> Right? I just told you they only have a [51:23] gigawatt and a half of compute. and they [51:25] have that gigawatt and a half of compute [51:26] whether they have a billion in revenue [51:29] or whether they have 80 billion in [51:30] revenue. So you might actually expect to [51:33] see these companies their gross margins [51:35] are exploding higher like the fastest [51:37] increase in gross margins I've probably [51:39] seen out of any technology company. So [51:41] this is not gross margin negative you're [51:42] saying? [51:43] >> No definitely not gross margin negative. [51:45] And what I would tell you [51:46] >> so that they must be hugely profitable [51:48] then [51:48] >> well you may see accidental why I call [51:51] it accidental profitability. They may [51:53] not be able to spend this revenue fast [51:55] enough chamath on compute. And remember [51:58] it's only 2500 people. Google crossed [52:01] this revenue threshold when they had [52:03] 120,000 people. These guys have 2500 [52:07] people. So the only thing you can really [52:08] spend money on, right, is compute. And [52:10] they can't stand up the compute fast [52:12] enough. [52:13] >> But none of this foots to me then to be [52:15] honest because if you were on a [52:17] threshold of 90% plus gross margin, [52:20] >> I'm not saying it's there. I'm not [52:22] saying it's 90% plus. I'm just saying [52:24] it's gone from meaningfully negative 18 [52:27] months ago to, you know, very very [52:29] positive. I've seen rumored out there [52:31] 50% is what you're saying. The trend is [52:34] there. [52:34] >> Let me just say this. [52:37] I think if you're an incumbent, you want [52:40] the cost of compute to go down. I think [52:42] if you're not an incumbent, so [52:44] specifically, who do I mean? Meta, [52:47] Google, and SpaceX. [52:51] I think those three people who have all [52:53] three of them, well, sorry, Meta and [52:55] Google have a fortress balance sheet. I [52:57] think by the end of June, SpaceX will [52:59] also have a fortress balance sheet. What [53:02] they will want to do is they will want [53:03] to make this a compute problem because [53:05] they will control the the conditions on [53:07] the field. You already see this today. [53:09] >> Yeah. [53:10] >> Meta's models today, what people's [53:12] general reviews are it's okay, but the [53:15] one thing that people say is it's [53:16] incredibly performant. The model quality [53:19] is okay, but the performance is great, [53:21] which speaks to Meta's huge advantage. [53:23] They have a massive compute [53:24] infrastructure. So if you're if you're [53:25] not open AI and anthropic, [53:28] they'll want to make this a capital [53:29] problem because then they can win it. If [53:31] you're anthropic and open AI, you want [53:33] this thing to be as efficient as [53:35] possible. [53:36] I think where we are is very much in the [53:38] early innings. And we're bumbling around [53:40] talking about gross margins and you know [53:42] revenues. We are not at profitability. [53:44] And what is true for Facebook and what [53:47] was true for Google was irrespective of [53:50] where they got to a billion. Who g [53:52] cares? They were profitable by year [53:54] three and they never looked back. I was [53:57] there. I remember it was glorious. [53:59] >> The the cost the cost of building uh you [54:03] know AI [54:05] totally stipulate is radically higher [54:07] than the cost of building retrieval at [54:08] Google, right? Like it's just a [54:10] fundamentally more expensive problem. [54:12] But I will tell you that there's a lot [54:14] of FUD out there about negative gross [54:16] margins. I mean Jason, you started the [54:17] segment by saying they're burning [54:19] through large amounts of cash. I think [54:21] people are going to be shocked at the [54:22] burn how low the burn levels are at [54:24] these companies. [54:25] >> Anthropic or Open AI. [54:27] >> Yes. And and I would say at Open AI as [54:28] well like they're if they're on you know [54:30] if they do $50 billion this year again [54:33] just look at the number of people they [54:34] have revenue per people. It's pretty low [54:37] and the inference cost is plummeting. [54:38] Inference cost is down by 90% [54:40] year-over-year. And so just finally I [54:43] want to make respond to this point about [54:45] gross versus net uh this this tweet that [54:48] Chimath was referencing. Okay, so [54:50] there's a certain percentage, a [54:52] smallalish percentage of Anthropics [54:54] revenue, right, that they distribute [54:55] through the hyperscalers and like a lot [54:57] of arrangements, whether it's Snowflake [54:59] or Data Bricks or others, you pay a [55:00] commission, right, uh on on that. I will [55:04] just tell you that you're talking [55:05] singledigit percentage of total revenue [55:07] of these companies. So the gross versus [55:09] net thing isn't what's being reported. [55:11] like the apples for apples is pretty [55:13] easy and if you want to be conservative [55:14] on it take down Anthropic's revenue by [55:17] you know five to 10% which you know [55:19] again I don't I think it's better to [55:21] gross up OpenAI's revenue but any way [55:22] you do it I just think it's a [55:24] distraction from what's really what's [55:26] really going on here happy to [55:27] >> s you have any thoughts on this uh [55:29] massive revenue ramp [55:31] >> yeah I mean I want to go back to a point [55:33] that Brad made because I think it was [55:35] just really important and I want to just [55:37] underline it consider where we were at [55:39] the beginning of the year and What [55:41] everybody was saying is that AI was a [55:43] big bubble and the evidence they would [55:46] point to was the fact that hundreds of [55:48] billions of dollars was going into capex [55:51] that needed to be spent on these data [55:53] centers and there was no evidence of [55:55] significant revenue to justify that [55:57] spend. Where was the ROI? By the way, as [56:00] an aside, the same doomers who were [56:02] saying that AI was in a bubble were also [56:05] the ones who were saying that AI was so [56:06] powerful it's going to put us all out of [56:08] work and it's going to, you know, take [56:10] over from humanity. I mean, in other [56:12] words, they couldn't decide if AI was [56:14] too powerful or not powerful enough. But [56:17] putting aside that contradiction, they [56:19] clearly were making this case that AI [56:22] was this big bubble and that there'd be [56:23] no payoff or justification for this [56:27] massive capex that's being spent. And I [56:29] think we're starting to see here there [56:31] is justification for it. Uh we're seeing [56:33] it just in this one vertical of AI which [56:36] is coding. We're again seeing the [56:38] fastest revenue growth in history. It's [56:40] utterly unprecedented. And this is just [56:42] one category or vertical of AI. We know [56:47] that agents are coming next and the [56:49] enterprise adoption of that is going to [56:51] be absolutely massive. So, I guess what [56:54] I'm saying is that this is early proof [56:56] for I think the thing that makes Silicon [57:00] Valley special, which is we're willing [57:02] to basically bet on things that just [57:05] intuitively on a gut level we know are [57:08] the next big thing. We're not that [57:09] spreadsheet driven. Actually, Silicon [57:11] Valley believes that if you build it, [57:13] they will come and is willing to finance [57:15] that build out. And that's basically [57:17] what's been happening. Again, just the [57:19] top four hyperscalers, $350 billion of [57:22] expected capex this year on its way, I [57:24] think Jensen said 1 trillion by 2030. [57:27] So, Silicon Valley, whether it's big [57:29] companies, whether it's founders, [57:30] they're always willing to bet on this [57:32] next big thing. They're not like Wall [57:34] Street. They don't need, you know, [57:35] specialist to tell them where to go. [57:38] They know where the technology is going [57:40] and they make their bets based on that. [57:42] And I think that there is going to be a [57:44] big payoff for this. And I think it's [57:47] the thing that's going to make our [57:48] economy and the United States in general [57:51] remain extremely dynamic and in the lead [57:53] on this thing is that we are willing to [57:55] make those kinds of bets. And I think [57:57] it's going to pay off big time. [57:59] >> Yeah, clearly. Hey, um Brad, you didn't [58:02] answer my question about the vibes over [58:04] at OpenAI versus Quad. Open AI is um I [58:09] wouldn't say reeling but there's a lot [58:11] of hand ringing going on a lot of [58:13] employees leaving a lot of people who [58:15] are wondering like is our strategy the [58:18] winning strategy of like consumer first [58:20] they shut down Sora you know unwinding [58:23] the Disney deal and really trying to get [58:26] the company focused and it's kind of [58:27] like I mean listen the New Yorker story [58:29] was a bit of a rehash so I don't think [58:30] we have to go into the blowby-blow [58:32] because we covered here three years ago [58:35] but the truth is a lot of the great [58:38] founders, co-founders of OpenAI and a [58:40] lot of the great contributors are now at [58:43] Anthropic and other large language [58:45] models. And in the secondary market, [58:48] OpenAI is trading lower than the last [58:50] valuation. And Anthropic is trading [58:53] significantly above the $380 billion. So [58:57] maybe talk a little bit about this [58:58] competition, this Microsoft versus [59:01] Apple, this Google versus Facebook. [59:03] Well, let's let's start with immense [59:06] credit where credit is due. Anthropic [59:08] was literally counted out of the game [59:09] last year. Y, [59:10] >> right? And here they come over the last [59:12] 12 months and and and they've kicked [59:14] OpenAI's ass over the last 90 days, [59:17] right? And what did Anthropic do? [59:19] Anthropic made choices. No multimodal, [59:22] no video, no hardware, no chips, no [59:24] building data centers. They said, "We're [59:26] just going to focus on coding and [59:27] co-work. We think that is the path to [59:29] AGI and and and and ASI." They executed [59:32] their butts off. They took the lead. [59:34] 2500 people tight pulling on the ore in [59:38] the same direction. But I think you [59:40] would be seriously foolish to count out [59:43] open AI, right? And I think we're we're [59:45] we're at peak open AI FUD. And I'll tell [59:47] you, it starts with great researchers [59:50] and great models. And I think when you [59:51] see the Spud model, they're about ready [59:53] to release. I think it's going to be an [59:55] excellent model. Shows that they're [59:57] firmly on the wave. Um, if you look at [60:00] what's going on with Codeex, incredible [60:02] ramp on Codeex, fastest ramping model [60:05] with 5.4, I think 5.5 or Spud, whatever [60:08] we're going to call, it's going to be an [60:09] even faster ramp. [60:10] >> Have you seen Spud? Have you used it? [60:12] Have you gotten a preview? [60:13] >> People are using Spud, right? So, it it [60:16] is being previewed and so [60:17] >> So, you're talking to people who've used [60:19] it and what are they telling you? [60:21] >> They're telling us that it's an [60:22] incredible model on par with Mythos, [60:24] right? and that it's a a very usable [60:27] model in terms of um how it's packaged. [60:30] I will say that back to David's point [60:34] now this is the most important point I [60:35] think anybody can take away here. [60:38] This is not zero sum. The TAM of [60:42] intelligence is dramatically larger than [60:44] any TAM we've ever seen in our investing [60:47] careers over the last two decades. [60:48] Right? And if you're on the wave, which [60:50] Open AAI is, you are going to be selling [60:53] into the world's biggest TAM, they are [60:55] going to build a very big company. I'm a [60:58] buyer of the shares today. [60:59] Notwithstanding all of the vibes that [61:01] you describe, I think these companies [61:03] are firmly on the wave. They are jarred. [61:07] They are sitting there saying, "What did [61:08] we do wrong? And how do we get our mojo [61:10] back?" They want to compete. It is [61:12] embarrassing to people on the research [61:14] team and the product team over there. [61:15] So, I'm not saying there's not a real [61:18] awakening occurring there, but I think [61:20] that's what the case is. And by the way, [61:22] to Chamas's point, do not count out [61:24] Meta, right? I think Meta is absolutely [61:26] in this game. Google is absolutely in [61:28] this game. Elon is absolutely in this [61:30] game. And if you're [61:31] >> got some stuff dropping shortly that's [61:33] going to be very impressive. [61:34] >> If you're on team America, the fact that [61:36] we have five frontier models competing [61:39] against each other and David made sure [61:41] they weren't throttled by excessive [61:43] government regulation. We have mythos [61:45] come out. It's a self-imposed safe [61:48] harbor, you know, to harden our system. [61:50] It wasn't a call for moratoriums or [61:52] getting the government involved. We have [61:54] the type of competition that's causing [61:55] us to accelerate our lead against the [61:58] rest of the world. We can't take our eye [62:00] off the prize. We got to stop [62:01] adversarial distillation and we need to [62:03] make sure that we're distributing our [62:05] products around the world. But I view [62:07] this as really good for team America. [62:09] >> Well said. And here is your poly market [62:12] IPOs before 2027. Obviously SpaceX at [62:16] 95% uh Cerebrus at 94% and uh hey number [62:21] five on this list 51% chance that [62:24] Anthropic goes out before the end of the [62:26] year. 44% chance that OpenAI comes out [62:29] before then. All right here is the [62:32] closing market cap for Anthropic on Poly [62:36] Market only $158,000 in volume. So, [62:39] Chimath, when you put in 400K, you're [62:42] going to really tilt this market. [62:44] 78% chance that it's above 600 billion, [62:48] 19% chance that it doesn't go out. So, [62:51] it's looking like this will be a decent [62:53] investment for you. Brad, what valuation [62:55] did you get into Anthropic at? [62:57] >> We first invested in I believe it was [62:59] the [63:01] uh 30 or $150 billion round. [63:04] >> So, this will be a 7x 5x for Altimeter [63:07] L, please. Congratulations. I mean, no, [63:09] listen. I I I again, there are lots of [63:11] people who were there before us and who [63:13] are on the board and who are going to do [63:14] much better than that. What' you put in? [63:16] 50. [63:17] >> What' you put in? [63:18] >> No, we've got billions in both [63:19] companies. Uh [63:20] >> billions in both companies. Oh my lord. [63:24] >> I think there's this existential thing [63:26] going on in venture today. David could [63:29] talk about it as well. I mean people [63:31] can't they're extraordinarily nervous [63:33] about you look at the IGV stock index [63:36] down 30% year to date down 5% today all [63:40] software stocks plummeting right venture [63:43] capitalists are terrified to invest [63:46] money in anything other than these [63:48] frontier models and things like SpaceX [63:51] or military modernization finding [63:53] something that's out of harm's way of AI [63:56] right where you can count on the [63:57] terminal value to Chamas insights over [64:00] the last few weeks is very difficult to [64:02] do. That's why you see this crowding. [64:03] So, we've taken a barbell approach, [64:05] right? We've got a lot in what we think [64:07] are the most important companies that [64:09] are on the frontier and then we're [64:10] betting with on really small teams that [64:12] we think have very defensible businesses [64:15] in a world of uh you know, AGI. But it's [64:18] >> what happens to all these enterprise [64:19] software companies? Do they become PE [64:21] takeouts? Do they get consolidated? um [64:24] or do they just have to adopt these AI [64:27] technologies and and and solve this [64:30] problem of hey the frontier model is [64:32] just going to solve for whatever these [64:34] niche software companies do. [64:36] >> I think the market's probably being a [64:38] little too pessimistic with respect to [64:40] at least some of these software [64:41] companies. I mean, obviously, there's [64:42] going to be big differences in the [64:45] quality of the modes of these companies. [64:48] And so, look, software is going to be a [64:51] lot cheaper and easier to generate, but [64:53] I'm not sure that was the competitive [64:55] advantage of a lot of these companies. [64:57] So, there's probably a little bit of the [64:59] baby being thrown out with the bathwater [65:00] right now, and there probably are some [65:02] value buys in enterprise software. I [65:05] think the interesting question here and [65:06] we've been talking about this for a [65:08] couple of years in the pod is just where [65:10] you see the AI value capture being in [65:13] terms of layer of the stack. Remember [65:15] where we started it was really just the [65:17] chip layer of the stack was where all [65:19] the value capture was. It was basically [65:20] Nvidia was the first company to be worth [65:23] multiple trillions of dollars because of [65:25] AI. And for a while it looked like [65:27] that's where all the value capture was [65:30] going to be because OpenAI for example [65:31] was losing so much money and Anthropic [65:33] wasn't on the radar as much. Now we're [65:36] seeing wait a second um you know it's [65:38] not just the chip companies it's also [65:40] the hyperscalers are now benefiting and [65:42] now we're seeing at the model layer it [65:44] looks like Enthropic and Open AI they're [65:47] all going to be huge beneficiaries. I [65:49] think the next question is at the [65:50] application layer of the stack. Okay. [65:52] Well, now does all that value capture [65:54] just get eaten by the model companies or [65:56] are there applications that get [65:58] turbocharged? I guess you could say that [66:00] Palunteer is already one of them, right? [66:02] It's an application company that's been [66:04] turbocharged by these model [66:06] capabilities. Who else will be a big [66:08] beneficiary? Is it again, is it all [66:10] going to be at the model layer or will [66:12] you see an explosion of value at the [66:14] application layer? I'm hoping obviously [66:16] that it'll be at all layers of the [66:18] stack. PC beneficiaries. But to me, [66:21] that's a really interesting question [66:22] right now. [66:23] >> Yeah. What happens to Salesforce, [66:24] HubSpot, you know, Oracle, right down [66:26] the line? David, uh, Chimati, your [66:28] thoughts here, uh, on the the layers [66:31] here and where the value is captured. [66:34] >> It's too early to tell. [66:35] >> Too early to tell, right? And energy we [66:37] kind of put into sort of data center as [66:40] well, but that's obviously been a clear [66:41] winner. Little housekeeping here. [66:43] Liquidity, put a little Tiffany in here. [66:45] uh producer Nick D is sold out. There's [66:49] a wait list of hundreds of people, but [66:51] it is what it is, folks. If you snooze, [66:53] you lose and top tier speakers are [66:56] coming. Uh it's going to be great. We'll [66:58] get a an update from But I think Brad, [67:00] you're going to be joining us again. [67:01] Yes. For liquidity. [67:02] >> I have an update. [67:03] >> That's probably not your headliner, [67:05] though. I'm probably not your headliner. [67:06] >> No, but you always score so high. Every [67:08] event you've spoken at, you've been [67:09] either number one, two, I don't think [67:11] you've ever dropped to three. Go ahead, [67:14] Sham. Make your announcement here. [67:17] >> Nat sent me an article from Wikipedia [67:19] about penile links when you guys are [67:21] talking about [67:21] >> breaking news. [67:22] >> Showing me showing me that I'm in the [67:24] large category. Top 5%. She highlighted [67:27] it. [67:27] >> Top 5%. Okay. And that's with Is that [67:30] with Nano Banana or without? Is that [67:34] >> She just texted dummy. It's clogged. My [67:35] apologies. Clogged. [67:36] >> Oh. [67:38] >> All right. This is why Jamath isn't [67:40] afraid of the cyber is because nothing's [67:42] going to come out that's more [67:42] embarrassing than what he says himself [67:44] on the box. [67:44] >> He's like Bezos. When Bezos got hacked, [67:46] HE'S LIKE, "GUYS, I GOT HACKED." [67:49] >> SO, I saw the agenda for this thing. [67:51] It's incredible. Congrats to you guys. I [67:52] mean, like the uh like just the fun of [67:55] being in Napa, all the poker, all the [67:57] the dining experience. This is five star [68:00] all the It looks really [68:01] >> six-star. It's a man level because [68:03] Chimath [68:04] >> was, I dare I say, belligerent in his [68:08] demands. He said, "This has to be [68:10] six-star or I will not show up." Jake [68:12] Al. I said, "Okay, boss, get to work." [68:15] And uh, Chimath, what do you got any? No [68:17] mids. This is all elite. And for the [68:20] hundreds of people who are on the wait [68:21] list, I am sorry, but we have a capacity [68:22] issue. We'll try to get you in for next [68:24] year. But Chim, give us some updates [68:26] here. You have any updates that you want [68:27] to share? because you are running [68:28] programming for liquidity 2026 up in [68:31] Yon. [68:32] >> Look, it's going really well. [68:35] Really excited to hear all of these [68:36] great folks speak. I think the next two [68:38] will release today. Brad Gersonner and [68:42] Thomas Leaf of COTU [68:44] >> of CO2. That's a great get. [68:46] >> We also have I think three people [68:47] confirmed for their best ideas pitch. [68:49] Really interesting folks. They each run [68:52] between one and six or seven billion [68:55] >> awesome [68:56] >> superstar compounders early in their [68:58] career. [68:59] >> This is a new zone chamat. [69:00] >> It's great. So right now we have Bill [69:02] Aman, we have Andre Carpathy, we have [69:05] Dan Loe, we have Thomas Lefont, we have [69:07] Brad Gersner, we have Sarah Frier and [69:10] more to come. We will announce more. [69:12] >> There might be one or two surprises. Jay [69:13] Cal [69:14] >> and a couple and a couple of surprises. [69:17] >> Yeah, we we don't announce all the [69:18] speakers. Jay Cal's got a couple of [69:20] surprises coming. And if you didn't get [69:22] in to liquidity, apologies. You're on [69:25] the wait list. We are going to be [69:26] hosting the fifth [69:30] annual all-in summit in Los Angeles [69:33] September 13th to the 15th of Sax. You [69:36] going to come to that? [69:37] >> Allin.com/events. [69:40] >> Sax, you should come to that. [69:41] >> I've been advised that I can attend [69:43] business. I can be in the state for [69:44] business reasons. [69:46] >> Okay, there you go. Then we'll see you [69:47] at liquidity and the summit. Correct. [69:49] That's that's big news. Now we just got [69:51] a bunch of Sachs stands who are racing. [69:53] Uh and now we're going to get Sachs at [69:55] This is what happens every year behind [69:57] the scenes. [69:58] >> Sachs at the last minute says, "Oh, I [70:00] have four speakers and I have 72 people [70:02] who need tickets and then the whole team [70:04] has to like do a fire drill 48 hours [70:07] before the event." Okay, here we go, [70:08] guys. We're going to go to the third [70:09] rail here. We got to catch up on the [70:11] Iran war. Here's the latest. Two weeks [70:14] into a ceasefire have started just two [70:17] days ago at the taping of this VP JD [70:20] Vance, friend of the pod is a and some [70:23] special consultants Wikoff and friend of [70:25] the pod Jared Kushner are headed to [70:28] Islamabad, the capital of Pakistan for [70:31] talks this very weekend. So while you're [70:33] listening to this event, they are going [70:34] to be working on the peace deal. Easter [70:36] Sunday, Trump posted a truth stating, [70:39] "Open the straight, you crazy [70:42] bastards, or you're going to be living [70:43] in hell. Just watch." Praise be to [70:45] Allah. On Tuesday morning, Trump posted [70:48] uh a another threat on social media. A [70:50] whole civilization will die tonight. [70:53] Never to be brought back again. I don't [70:55] want that to happen, but it probably [70:56] will. Tweets were obviously discussed uh [70:59] a lot over the last week. He gave him an [71:01] 8:00 p.m. deadline. [71:03] At 6:30 p.m. POTUS announced on Truth [71:05] Social that he had agreed. President [71:08] Trump had agreed to a two-week ceasefire [71:11] if Iran opens the straight. He also [71:13] said, "Hey, listen. We got the straight. [71:15] Maybe there'll be a toll booth, but [71:16] we'll take the majority of the toll and [71:18] we'll split it with Iran." Here's the [71:20] quote. We received a 10-point proposal [71:22] from Iran, and we believe it's a [71:24] workable it is a workable basis on which [71:28] to negotiate. And apparently Netanyahu [71:31] took the ceasefire to mean level Lebanon [71:33] dropping 160 bombs in 10 minutes [71:36] yesterday. Saxs, uh, you were out last [71:38] week. Everybody wants to know your [71:39] position on the war. I'll hand it off to [71:42] you. What are your thoughts on how on [71:44] the two ceasefire and everything that's [71:45] occurred up until this point? [71:47] >> Well, look, I have to preface what I'm [71:49] about to say, which is I'm not part of [71:52] the foreign policy team at the White [71:54] House. And the last time I commented on [71:57] the war on this show, it somehow made [71:59] international headlines that Trump [72:01] advisor says XYZ. [72:05] And I'm not a Trump adviser on this [72:07] issue. I think that'd be a fair headline [72:09] to write if it was a technology issue, [72:11] but this is not. So whatever I say is [72:14] just my personal opinion, but then the [72:16] media is going to somehow portray it or [72:17] attribute it [72:18] >> to the White House or try and create an [72:20] issue out of it. So, I feel like I'm [72:22] limited in what I can say except that to [72:25] say that I think it's terrific that we [72:28] have the ceasefire. I think it's great [72:30] that there's going to be this meeting in [72:33] Islamabad to hammer it out. And I think [72:37] what the president's accomplished so far [72:38] with the ceasefire is it's a great thing [72:41] because what happens with these wars is [72:43] they take on a life of their own, [72:45] meaning they tend to go up the [72:47] escalation ladder, right? And there's a [72:48] lot of podcasts that are discussing the [72:50] so-called escalation trap and supposedly [72:52] there are stages to this based on [72:54] historical patterns. And so I think it's [72:56] actually very hard to pull out of these [72:58] things and I give the president [72:59] tremendous credit for negotiating the [73:02] ceasefire that we've achieved so far and [73:04] then sending the team to hopefully work [73:06] this out. [73:07] >> Brad, actually my first trip to the [73:08] Middle East was when you and I uh maybe [73:10] four years ago when Thank you for taking [73:12] me. What is your take on where we're at [73:14] here? I think we're just wrapped up week [73:16] six of this and we're going into week [73:17] seven. [73:17] >> First, on March 4th, I tweeted the Trump [73:20] doctrine in Iran. Massively destroy all [73:23] military capa capabilities. Kill the [73:25] people building lethal weapons to use [73:27] against us and get out. Reserve the [73:29] right to do it again if needed. Zero [73:31] efforts to build Misonian democracy. [73:33] Iran's going to have to build what comes [73:35] next. And I think what the market has [73:38] said right if you look back at last year [73:40] on tariffs Jason the top tobottom draw [73:42] down was about 15% on the NASDAQ [73:46] intraday is down 22%. [73:48] Okay, the draw down in this period over [73:51] Iran was only down about 5 to 7% on S&P [73:55] and NASDAQ, right? So, the market has [73:58] said, listen, re trust Trump at his [74:01] words. He said he's not going to get [74:02] into an entangled war here. I think he [74:05] terrifies the hell out of people with [74:06] his tweets about, you know, destroying [74:08] civilization and all this other stuff. [74:10] But I think people, even though they [74:12] don't like to hear it, they've resolved [74:14] for themselves that when he says he's [74:16] going to get out, he will in fact get [74:17] out. Of course, there was a lot of hand [74:18] ringing, but if you look at the markets [74:20] today, we basically bounced all the way [74:22] back from where we were pre-Iran on both [74:26] the S&P and the NASDAQ. If in fact we [74:29] land the plane, if JD lands the plane, [74:31] and by the way, on Lebanon, yes, they [74:33] were bombing yesterday, but Netanyahu [74:34] has now said that you're going to have [74:36] direct government talks between Israel [74:37] and Lebanon. So, if the if we land the [74:41] plane on these two things, I think it's [74:42] off to the races in the market. By the [74:44] way, while everybody's focused on Iran, [74:47] stay tuned. I think we're getting close [74:48] to a deal on Ukraine, Russia, right? [74:51] Venezuela is, you know, kind of going [74:53] seemingly very well. I think there's [74:55] also going to be news on Cuba. You could [74:58] envision a world there's risk to the [74:59] downside. Certainly, I will stipulate, [75:02] but you also have to pay attention to [75:03] the risk to the upside. If you land the [75:05] plane on those things heading into [75:07] America 250 July 4th, the market could [75:10] really take off. [75:11] >> All right. Well, let's uh maybe uplevel [75:13] this a little bit and talk about why [75:15] we're in this war to begin with. And [75:17] that's the big discussion amongst both [75:19] sides of the aisle. On Tuesday, the New [75:20] York Times dropped an inside the room [75:23] piece on how President Trump made the [75:26] decision [75:28] according to this report, if it's true. [75:30] I know some people don't uh subscribe to [75:32] the New York Times anymore or think it's [75:33] fake news, but how Trump decided to [75:36] basically follow Netanyahu into this [75:38] war. On February 11th, Netanyahu met [75:40] with Trump at the White House where he [75:41] gave him a four-part pitch on attacking [75:44] Iran. JD Vance, according to the story, [75:46] if it's true, disclaimer, disclaimer, [75:48] warned Trump that the war could cause [75:50] regional chaos and break apart Trump's [75:53] MAGA 2.0, the Trump 2.0 coalition we [75:56] talked about here, the big tent. And [75:57] that's turned out actually to be true. [75:59] There's been a bunch of hand ringing [76:00] from Megan, Kelly, Tucker, Carlson, [76:02] right on down the line. Rubio was [76:04] anti-regime change, but he was largely [76:07] ambivalent, according to this story [76:08] about the bombing campaign. Susie Wilds, [76:12] chief of staff, said she had concerns [76:13] about gas prices before the midterms. [76:15] Pretty good uh advice there. And General [76:18] Dan Kaine, chairman of the Joint Chiefs [76:20] of Staff, said this of Netanyahu's [76:23] pitch. Quote, "Sir, this is, in my [76:24] experience, standard operating procedure [76:26] for the Israelis. They oversell and [76:28] their plans are not always [76:30] well-developed. They know they need us [76:31] and that's why they're hard selling. If [76:33] you put this together with Rubio's [76:35] walked back comments at the start of the [76:38] war, we knew, this is quote from Rubio, [76:41] we knew there was going to be an Israeli [76:44] action. We knew that would precipitate [76:47] an attack against American forces and [76:49] that's why we did it. I had Josh Shapiro [76:52] on the All-In interview show and um uh [76:56] he talked a lot about this. There is a [76:58] big underpinning here, Chimath, that the [77:01] United States foreign policy is being [77:03] driven by Netanyahu. Every Jewish [77:07] American person I've talked to feels [77:08] Netanyahu is not doing Jewish American [77:11] and Jewish the Jewish diaspora any [77:13] favors here by his approach to these [77:15] wars. What are your thoughts on why we [77:17] got into this and how we get out of it? [77:22] >> I mean, the person that decides is the [77:24] president of the United States. some [77:26] foreign leader isn't [77:29] getting to call the shots in the United [77:31] States. I think very practically [77:33] speaking, [77:35] the markets are effectively [77:38] pricing in that this was a small blip [77:42] for whatever people think. That's just [77:44] what the best prediction market that we [77:47] have is telling us. I think that's [77:48] important to acknowledge that we're [77:51] probably in the endgame here. And the [77:53] second thing to acknowledge is if I was [77:55] Israel, I would really be concerned [77:59] that unless I help find an offramp [78:02] quickly, the risk that Israel loses [78:05] America as a predictably steadfast ally [78:08] could go down. And I think that that's [78:09] problematic for Israel [78:11] >> far more than is problematic for the [78:12] United States. [78:13] >> So all of that kind of tells me that we [78:16] will find an offramp. A because I think [78:19] economically it makes sense and then B [78:22] geopolitically I think Israel will want [78:23] to make sure that this doesn't burn [78:26] a long-standing relationship. [78:29] >> Yeah, that that seems to me to be the [78:32] major issue here is Americans basically [78:34] do not want to be in this war. Americans [78:37] do not want a forest policy being [78:39] influenced to the extent they believe. [78:42] I'm not putting my belief in here. [78:43] Americans believe we are being dragged [78:46] into this by Israel and that Israel has [78:48] too much or Netanyahu specifically has [78:50] far too much influence. And then people [78:52] believe the anti-semitism that's [78:53] occurring here. Josh Shapiro gave me a [78:55] lot of push back on this. Uh but all the [78:58] Jewish Americans I talked to say [78:59] Netanyahu is causing with his actions in [79:02] Gaza, Lebanon, Iran. Uh he's gone too [79:05] far and it's causing the anti-semitism [79:07] we're experiencing uh today. So you can [79:09] make your own decisions about that. Any [79:11] final thoughts here, Brad, on [79:14] the American foreign policy being [79:16] influenced too much by Israel? [79:18] >> No, it's the discussion. I [79:20] >> I mean, listen, um kind of like Sax said [79:24] earlier, um I think that we will [79:27] ultimately be judged by the outcomes, [79:31] right? And that everybody is an armchair [79:33] pundit today on, you know, uh the the [79:37] the approach that we're taking in these [79:39] two different places. I think we could [79:41] be on the verge of a massive [79:44] transformation of the Gulf States. You [79:46] went there with me, Jason. Saudi, [79:48] Qataris, Kuwaitis, Emiratis. I've talked [79:50] to a lot of them this week. I think [79:52] they're very hopeful and optimistic. I [79:54] think you could bring Iran into the [79:56] fold. But listen, I'm an optimist on all [79:58] of this stuff. I I just want to remind [80:00] people, doing nothing in Iran had [80:04] tremendous risks. Doing nothing in [80:07] Venezuela had tremendous risks. So, it's [80:10] not as though this was uh, you know, [80:13] something that I I I I think wasn't well [80:16] calculated, but I think we have to let [80:19] the cards be played and and and then let [80:21] history be the judge. But I think [80:22] there's uh a risk in both directions, [80:25] but I'm going to remain optimistic. [80:26] >> All right. You uh said in the Gaza [80:28] situation, we should have a wide birth [80:30] for criticism of Israel and Netanyahu. [80:32] What are your thoughts on this belief [80:35] here in the United States now in this [80:37] discussion that Israel is having far too [80:38] much influence over the United States [80:40] foreign policy? [80:42] Well, I noticed in my feed today that [80:44] Naftali Bennett, who's a major Israeli [80:47] politician who was a former prime [80:49] minister, tweeted polling that showed [80:52] that Israel was becoming very unpopular [80:55] in the US and he was expressing concern [80:57] about that and expressing [81:01] the need to to basically address that or [81:03] fix that. So, I think you're starting to [81:05] see Israeli politicians raising that as [81:08] an issue. And I think that's probably a [81:11] good thing. Yeah, there it is. And it's [81:14] really cool actually how X now just [81:16] automatically translates things from [81:18] foreign languages, in this case, Hebrew, [81:20] and it puts it in your feed. So, yeah. [81:22] So, here's Naftali Bennett, former prime [81:23] minister, saying, "This is a serious [81:24] situation. There's a lot of work ahead [81:26] of us to fix everything." Now, [81:29] obviously, this is not Netanyahu. this [81:31] is one of his um political opponents. [81:33] But [81:34] yeah, I mean this is something for [81:36] Israel to consider and think about and I [81:39] think that they would improve their [81:41] popularity uh if they got behind the [81:43] ceasefire and I have no indication that [81:45] they won't but that would certainly be a [81:48] good place to start. I have to say just [81:50] as an aside, this auto translate feature [81:52] has done more for understanding across [81:56] borders than anything I've ever seen. [81:58] And it is the most impressive tech [82:02] feature I I've seen released in years. [82:04] Putting AI and large language models [82:06] aside for people who don't know what's [82:08] happening because of Grock being really [82:11] good at doing auto translate. They've [82:12] taken the pockets of the best of what's [82:16] happening in Japan, what's happening in [82:17] Israel, what's happening in France, and [82:19] they're surfacing it auto translated. [82:21] Then when you reply as an American to [82:23] somebody in Japan, they see it autorated [82:26] as well, which has led to people who [82:28] don't speak the same language engaging [82:31] on X in a very nuanced, fun, interesting [82:36] way. And that for as a truth mechanism [82:40] is just absolutely extraordinary. I [82:41] think this is going to have such a [82:42] profound effect. Maybe Elon and the X [82:45] team should get like a Nobel Peace Prize [82:47] award for this. I think it's going to [82:48] change. I mean, I hate to be hyperbolic, [82:50] but have you been using this feature, [82:52] Chamat? Has it been coming up in your [82:53] feed? And which language is up in your [82:56] feed right now? [82:58] >> English. Okay. So, you're not part of [83:01] the translation thing. Brad, has this [83:02] hit your feed yet? And and which regions [83:04] are you? [83:05] >> Definitely. Definitely see it in on the [83:07] Middle East stuff. Um, and uh, you know, [83:10] I've seen on Chinese, I've seen it on on [83:12] the Russian Japanese [83:14] >> super helpful. [83:14] >> Let me tell you, bass Japanese is a [83:16] whole another level of beast. [83:18] >> Whoa. Man, base Japanese makes like [83:22] Fentes and Alex Jones seem tame. They're [83:25] like, look at this group of people. [83:27] Insert whatever group of immigrants you [83:29] like. And they're like, this is [83:30] unacceptable behavior. This is not [83:32] Japanese culture. These people need to [83:34] be get the hell out of Japan. It is [83:38] wild, folks. And if you don't have an X [83:40] account, you are missing out. Go to [83:42] X.com and sign up for this reason only [83:45] because you think about the velocity. [83:47] Like journalists are not even taking the [83:48] time to translate and cover what's going [83:50] on in those areas. And this is happening [83:52] automatically in real time. [83:55] >> So you start thinking about what [83:56] happened in Ukraine. If you had people [83:58] Russia and Ukraine doing this and having [84:00] conversations with each other, it would [84:02] be wild. [84:03] >> You're like a such a good hype man. The [84:05] problem is you hype buttered bread the [84:07] same way you hype a nuclear reactor. And [84:08] so it's hard to really tell, you know, [84:10] what you're really hyping because your [84:12] level of excitement, the intonation is [84:14] exactly the same. [84:15] >> Yo, man, there's nothing better than a [84:17] slice of great toast. I mean, if this is [84:18] very this in in a way, it is like sliced [84:21] bread. It's very simple, but it is so [84:23] powerful in the experience. [84:25] >> This has been [84:26] >> It is true. X is better today than it's [84:28] ever been. And remember, they have 70% [84:30] fewer employees than they had the day [84:33] Elon walked into the building. And so if [84:35] there were ever a debate [84:37] >> about this, like, and I remember [84:38] everybody saying, "Oh, it's going to tip [84:40] over. Oh, it's going to be a crappy [84:41] experience." [84:43] >> The fact of the matter, here's we are a [84:45] few years later, 70% fewer employees, [84:48] and every other company in Silicon [84:49] Valley is looking at that. I think for a [84:51] lot of these tech companies, we've hit [84:53] peak employment. We're going to create a [84:55] tremendous number of new jobs, but for [84:58] the existing jobs, these companies are [85:00] all realizing they can do more with [85:01] less. [85:01] >> Nikita Beer just tweeted that they're [85:03] about to go ham on these bot accounts [85:06] that auto reply. [85:07] >> Yes. [85:08] >> Those those literally ruined my feed. [85:10] >> That's why I went to subscriber mode in [85:12] my replies and it's it's worked out [85:13] great. [85:14] >> Yeah. No, shout out to him and um to [85:16] Chris Saka who was in tears at what [85:18] happened to Twitter. You [85:20] >> It's gonna be okay, Chris. Sorry. No [85:23] more tears. [85:24] You only let subscribers respond to your [85:26] tweets. I [85:26] >> I do 50/50. Sometimes I'll just let it [85:29] rip and get chaos. And then other times [85:31] I have 2,000 paid subscribers. I give [85:33] all the money to charity, like 30 grand [85:34] a year. And it's just wonderful to get [85:36] to know the same 2,000 people out of my [85:38] million followers. It's kind of like [85:40] having this little subset. So sometimes [85:41] I'm like, I don't have time to deal with [85:43] a 100red or 200 or 300 replies. [85:45] >> You have a million. That's incredible. I [85:48] mean, it's just [85:49] >> I mean, you have two million. I think [85:50] Sax must have a million, right? You have [85:52] a million, right, Sax? Only only [85:53] >> Brad. How many you have now? You're [85:54] getting popular. [85:56] >> You built a couple. [85:57] >> Got a couple hund. [85:59] >> What's your Oh, your alt cap altca. [86:01] >> I'm at 1.4 million. What are you at, [86:03] Jacob? Have I surpassed you? [86:06] >> I think you have. I'm like 1.1. [86:07] >> What it cost me to get my real name, [86:09] Jason? [86:10] >> Uh, I know a guy. Find out. [86:13] >> You're 1.1. Yeah, I made it to 1.4. I [86:15] don't know how that happened exactly. [86:17] >> Just having the number one podcast in [86:19] the world. Uh, another amazing episode [86:22] of the number one pop and Chimath has [86:26] two million, but that's only because he [86:28] engages he has just incredible moments [86:32] of uh engaging with his haters. Oh my [86:35] god, the the the replies that Chimath [86:38] sometimes drops are so great. I love [86:40] Chimoth goes [86:41] >> I light them up. I light them up. [86:43] >> He lights them up. Then you had somebody [86:45] who was like, "Oh my god, I was in the [86:47] casino and you told me to bet black, so [86:48] you bet black, so I bet black and I lost [86:50] my money." And so you're responsible and [86:53] then you paid for the kids college. He [86:54] has two young girls and so I I funded [86:57] their college accounts. [86:58] >> I thought that was hilarious. Just as [87:00] >> obviously I'm very happy for him and his [87:02] two daughters. I'm even more happy at [87:04] how much it'll anger all these other [87:06] goofball dorks living in their mom's [87:08] basement. [87:09] >> Yes. [87:09] >> Who'd literally have no take they take [87:11] no responsibility for their lives. [87:14] And uh they should enjoy those Hot [87:16] Pockets. By the way, for those folks in [87:17] their mom's basement, the Hot Pockets [87:19] and the Fish Sticks are ready and you [87:21] get one more hour of Xbox from mom. [87:23] >> All right, listen. We missed you, [87:24] Freeberg, but this is the best episode [87:26] in two years. Uh [87:29] >> Freeird at the end of the show. [87:32] >> And we will see you all at the liquidity [87:34] summit except for the 400 people on the [87:36] wait list who aren't going to get in. [87:37] >> We got an email from the guys at Athena [87:39] because we were just [87:40] >> Oh my god. the they they're they're [87:43] going to hire like 500 new Athena [87:45] assistants. [87:46] >> Yes, they had a thousand people after [87:48] last week when we mentioned how much we [87:50] love Athena. [87:51] >> Go to Athena.com. [87:52] >> But that's amazing. Those are like 500 [87:54] hardworking men and women who are like [87:56] working [87:57] >> in the Philippines. [87:59] >> Sax have great jobs. [88:00] >> Sax, I'm going to get you a couple [88:01] Athena assistants as a birthday present. [88:03] That's what I'm going to get. [88:04] >> You're going to love this, Sax. H Athena [88:06] assistants are the best. Congratulations [88:08] to my friends over there. All right, [88:10] everybody. We'll see you next time. Love [88:11] you boys on [88:12] >> tonight favorite podcast. [88:19] >> Let your winners ride. [88:21] >> Rain [88:26] and [88:26] >> we open source it to the fans and [88:28] they've just gone crazy with it. [88:30] >> Love you queen of [88:33] winners. [88:39] >> Besties are gone. [88:41] That is my dog taking out your [88:43] driveways. [88:46] >> Oh man, my appetiter will meet. [88:49] >> We should all just get a room and just [88:50] have one big huge orgy cuz they're all [88:52] just useless. It's like this like sexual [88:54] tension that we just need to release [88:55] somehow. [89:02] >> That's going to be good. We need to get [89:03] merch. [89:12] I'm going all in.