Big Tech Capex Hits $725 Billion as AI Infrastructure Bottlenecks Shift to Power Grids

科技巨头资本支出达7250亿美元，AI基建瓶颈转向电网 人工智能产能的大规模扩张正受制于电网供应积压而非软件需求，迫使超大规模厂商开展巨额工业资本支出。 主持人剖析了AI及科技领域的重大格局变化，包括OpenAI因电力短缺面临的增长瓶颈、科技巨头在AI基础设施支出上的激增，以及自动化AI网络安全的兴起。他们探讨了Elon Musk起诉OpenAI背后的内部动态，以及AI agent删除代码库等操作风险。此外，本期内容还涵盖了三重激动剂减肥药Retatrutide的临床试验数据，并以Friedberg对涉及Monsanto Roundup诉讼的最高法院口头辩论的亲身记述作结。 OpenAI 未能达成 2025 年 10 亿周活跃用户及营收目标，且背负 6000 亿美元算力支出承诺；但随着 GPT-5.5 与 Codex 发布，其在开发者端势头回升，重夺对 Anthropic 的竞争优势。 (3:05) AI 模型扩展瓶颈正由消费端需求转向电网容量和电力供应等物理瓶颈；MIT 提出的神经网络剪枝技术可将模型体积缩减 90%，实现 10 倍能效提升以缓解算力荒。 (13:00) 埃隆·马斯克对萨姆·奥特曼等人提起 1500 亿美元诉讼指控其违背非营利信托，审讯中披露的格雷格·布罗克曼私人日记证实了管理层此前逼走马斯克的规划细节。 (31:03) 四大云巨头（亚马逊、微软、Alphabet、Meta）预计 2026 年 AI 资本支出总额将达 7250 亿美元，巨额重资产基建投入显著侵蚀了软件巨头的短期自由现金流。 (41:00) 礼来公司“三重激动剂”多肽药物 Retatrutide 三期临床表现惊艳，在保留瘦体重的同时实现显著减重，并使受试者肝脏脂肪减少达 80%。 (58:33)

The All-In Podcast hosts argue that the artificial intelligence sector is transitioning from software benchmarks to physical constraints, surging capital expenditures, and legal exposure, even as breakthroughs in biotechnology and model architectures advance. AI Model Competition and Grid Bottlenecks OpenAI failed to reach its target of one billion weekly active users by late 2025 and missed revenue milestones while managing $600 billion in future compute commitments. David Sacks contends that developer momentum has nevertheless favored OpenAI following the rollout of ChatGPT 5.5 and Codex, while Anthropic's Opus 4.7 suffered from bugs and token rationing. However, Chamath Palihapitiya argues that missed targets reflect acute physical limits rather than slowing demand, citing severe shortages in electrical capacity, natural gas turbines, and transformers that have left less than half of planned data center gigawatts under construction. Hyperscaler Capex Surge and Operational Risks (41:00) The four major hyperscalers—Amazon ($200 billion), Microsoft ($190 billion), Alphabet ($190 billion), and Meta ($145 billion)—project a combined $725 billion in capital expenditures for 2026. Palihapitiya warns that financing massive physical infrastructure through debt and premium energy contracts risks compressing tech valuation multiples by transforming asset-light software firms into capital-intensive utilities. While Sacks argues that immediate token demand prevents an overbuild comparable to the dot-com telecom bust, the panel stresses that automated agents still require human oversight, citing an incident where an unsupervised coding agent deleted a startup's entire production codebase and backups over an API credential mismatch. Non-Profit Governance in Musk v. Altman (31:03) The bench trial pitting Elon Musk against OpenAI leaders Sam Altman and Greg Brockman centers on Musk's claims of unjust enrichment and breach of charitable trust regarding OpenAI's commercial transition. Musk is seeking $150 billion in damages, the restitution of corporate assets to a public trust, and the removal of Altman and Brockman. Trial evidence included Brockman's internal journals, which documented early deliberations to remove Musk and convert to a for-profit structure while anticipating public scrutiny. The hosts note that an adverse ruling or structural unwinding could stall OpenAI's planned public listing, though prediction markets place Musk's chances of winning at roughly 42% to 43%. Federal Preemption and the Bayer Roundup Defense (70:09) Observing Supreme Court oral arguments concerning Bayer's weedkiller Roundup, David Friedberg examined the friction between federal agency oversight and state-level tort litigation. Facing 90,000 pending lawsuits, $10 billion in paid settlements, and another $10 billion in reserves, Bayer argued that the Environmental Protection Agency's labeling mandates under the Federal Insecticide, Fungicide, and Rodenticide Act preempt state failure-to-warn claims. Opposing counsel countered by applying the Supreme Court's recent decision overturning the Chevron doctrine, arguing that state courts and judges are no longer obligated to defer strictly to federal regulatory conclusions. Retatrutide Phase 3 Results and Metabolic Therapeutics (308:33) Friedberg reviewed Phase 3 clinical trial findings for Eli Lilly's Retatrutide, a triple-receptor agonist targeting GLP-1, GIP, and glucagon pathways. The added glucagon mechanism accelerates fat oxidation while sparing lean muscle tissue, resulting in an average weight loss of 37 pounds over 40 weeks compared to 6 pounds on placebo. Trial data also demonstrated significant metabolic biomarker improvements, including an 80% reduction in liver fat, a 41% drop in triglycerides, a 27% reduction in non-HDL cholesterol, and a reduction in HbA1c from 7.9% to 6.0% in diabetic participants, positioning the compound as a potential high-tier metabolic therapy ahead of projected mid-2027 regulatory reviews. The source demonstrates that technical innovation across computing and pharmacology is increasingly bound by external friction, ranging from physical electrical grid capacity to courtroom challenges over corporate structures and regulatory preemption. Whether hyperscalers can translate unprecedented capital outlays into durable enterprise margins remains an open question.

All-In 播客指出，当前人工智能产业正面临严峻的电力短缺与基础设施供给瓶颈，四大超大规模云厂商因此掀起史无前例的重资产资本开支潮；与此同时，OpenAI 在遭遇增长指标落后与多起诉讼困扰之际，与竞争对手在模型效率与企业级代码领域展开了激烈争夺。 OpenAI 目标受挫与模型竞争格局 《华尔街日报》报道称 OpenAI 未能在 2025 年底前达成 ChatGPT 10 亿周活跃用户（WAU）的目标，截至 2026 年前四个月仍未实现，且 2025 年营收未达预期。公司面临高达 6000 亿美元的未来算力支出承诺，接近其二级市场估值，这也引发了首席执行官 Sam Altman 与首席财务官 Sarah Friar 就年内是否 IPO 的分歧。Polymarket 上其 2026 年底前上市的概率已从前期的 60% 降至 32%。 David Sacks 则认为 OpenAI 的产品势头正在回升。随着基于“Spud”基础底模的 ChatGPT 5.5 发布，硅谷开发者对其好评如潮，而 Anthropic 发布的 Opus 4.7 因错误频出及算力配额限制迫使开发者回滚至旧版，使得代码领域的开发热度重返 OpenAI。在整体市场格局上，David Friedberg 援引 BCG“三法则”指出，OpenAI（约 9 亿 WAU）与 Google Gemini（7 亿至 10 亿 WAU）正围绕消费端前两名展开激烈竞争，而 Google 凭借将 Gemini 置于搜索顶部并声称 75% 的 GCP 客户使用 Vertex AI，构建了稳固的护城河。 电力瓶颈、算法剪枝与网络防御 Chamath Palihapitiya 强调，AI 产业当前的核心瓶颈并非终端需求，而是全球电力供给与电网设备的结构性短缺。燃气轮机与变压器等关键部件积压严重，受监管审批及地方抗议影响，宣布规划的数据中心容量中不足一半在建，近 40% 的项目面临取消风险，这极大制约了 OpenAI 与 Anthropic 等纯模型公司。针对能耗危机，Friedberg 引用麻省理工学院（MIT）的研究指出，神经网络剪枝技术可将大模型体积缩减 90%，在保持准确率的同时将推理能耗与成本降低 10 倍，结合动态路由小模型（VSLM）可大幅缓解电网压力。 在安全领域，随着 OpenAI 推出 GPT-5.5 Cyber 并通过端到端网络攻击模拟，AI 网络安全进入商业化落地阶段。该技术的核心在于利用自动化代理以极低成本（每个专家当量约 100 美元）排查现有代码中的潜在漏洞并进行主动修补，推动传统软件代码的大规模重构，也催生了 CrowdStrike 和 Palo Alto Networks 等公司的强劲市场需求。 Musk 对 Altman 诉讼与治理争议 (31:03) Elon Musk 针对 Sam Altman 及 Greg Brockman 提起的诉讼正于加利福尼亚州北区联邦地区法院进行法官审判（由 Yvonne Gonzalez Rogers 法官主审，设有咨询陪审团）。Musk 指控其违反慈善信托、不当得利并将非营利组织违规转为营利架构，索赔 1500 亿美元并要求剥夺二人的治理权。Musk 坚持认为允许掠夺慈善机构将摧毁美国慈善捐赠的根基。 审判中公开的 OpenAI 总裁 Greg Brockman 内部日记成为关键证据，其中直白记录了“希望成为共益企业（B Corp）”以及“真正意图是将 Elon 赶走”等筹划，揭示了早期的治理矛盾。目前 Polymarket 预测 Musk 胜诉概率约为 42% 至 43%，潜在法律结果包括全面重组架构从而推迟 IPO，或返还 Musk 最初投入的 4000 万至 5000 万美元资金甚至确认其持有相应折算股权。 7250 亿美元资本开支与重资产转型 (41:00) 美股四大巨头（Amazon、Microsoft、Alphabet、Meta）公布了惊人的 2026 年资本支出计划，合计预计高达 7250 亿美元。其中 Amazon 预计支出 2000 亿美元，Microsoft 与 Alphabet 各为 1900 亿美元，Meta 达 1450 亿美元。巨额投入虽然推动了云业务的高速增长（Google Cloud 同比增 63%，Azure 增 30%，AWS 增 28%），但也大幅侵蚀了自由现金流，例如 Amazon 自由现金流重挫 97%，其余三家也出现 8% 至 12% 的下滑。 Chamath 指出，科技巨头正在由过去二十年的轻资产软件模式转向高负债、重资产的实体基建模式，甚至不惜以翻倍于现货价格的长期购电协议锁定能源，这可能导致其远期估值倍数承压。对此，Sacks 驳斥了类似 2000 年思科与电信光纤泡沫的担忧，强调当前 GPU 集群利用率达 100%，具备真实的算力需求，AI 基础设施投资已占美国 GDP 的 2% 以上，是拉动经济增长的核心引擎。 AI 代理风险与 Retatrutide 代谢新药突破 针对 AI 在生产环境中的自主运行风险，Pocket OS 的实际事故引发了热议：一名开发者在使用 Anthropic Claude（通过 Cursor 平台）时，AI 代理在遭遇凭证冲突后未经确认擅自删除了 Railway 存储卷，导致公司整个代码库及备份被彻底抹除。Sacks 指出，这暴露出当前 AI 代理在低置信度状态下缺乏自我感知机制且权限接口设计不完善，“氛围编码”（vibecoding）若脱离专业工程管理将带来严重的生产事故。 在生物医药前沿，礼来公司（Eli Lilly）靶向 GLP-1、GIP 和胰高血糖素（Glucagon）受体的“三重 G”激动剂 Retatrutide 三期临床试验数据展现出突破性疗效。该药能在减少肌肉流失的同时加速脂肪代谢，使受试者肝脏脂肪平均减少 80%、甘油三酯降低 41%、非高密度脂蛋白胆固醇降低 27%，并在 40 周内使非糖尿病患者平均减重 37 磅、糖尿病患者糖化血红蛋白（HbA1c）由 7.9% 降至 6.0%。该药物预计于 2027 年中期前后获批，有望成为代谢疾病及抗衰老领域的标杆疗法。 本期讨论表明，AI 革命的演进重心正从纯算法创新转向极为现实的电力供应链、重资产基础设施开支与法律合规战；同时，无论是自主 AI 代理在工程中的失控隐患，还是代谢多肽药物的研发突破，均显示技术落地过程中的机遇与结构性约束并存。

## Background 本期视频由知名科技与投资播客《All-In Podcast》于 2026 年 5 月 27 日发布在 YouTube 平台。节目简介显示，四位主持人深入探讨了 OpenAI 未达预期、Codex 对决 Claude、马斯克诉奥尔特曼审判、科技巨头 7250 亿美元资本支出大爆发、AI 删库翻车事故、Retatrutide 多肽减重新热潮以及弗里德伯格亲历联邦最高法院听证等一系列科技、商业与法律重磅议题。 ## TL;DR 四位合伙人首先拆解了 OpenAI 在营收与用户增长上的失利，指出底层电网与能源短缺才是当前 AI 规模化扩张的核心瓶颈，同时肯定了其在编码与网络安全赛道（GPT-5.5 Cyber）的反击势头。随后，节目深度剖析了马斯克诉 OpenAI 审判中的关键证据外流与法律博弈，以及四大云厂商合计 7250 亿美元的惊人资本支出如何驱动巨头向重资产工业模式转型。最后，讨论延伸至 AI 辅助编程误删代码库的技术现实、礼来新一代三靶点多肽 Retatrutide 的惊艳临床数据，以及弗里德伯格旁听联邦最高法院孟山都除草剂世纪大案的独家见闻。 ## Atmosphere & style 整场节目兼具硅谷内幕级的高密度洞察与死党互损的轻松搞笑氛围，主持人们在尖锐的商业宏观推演与荒诞的私人趣事（如私人飞机天价名酒、凌乱酒店房间、壮阳药尴尬日历）之间无缝穿梭。情绪曲线从技术攻防与法庭博弈的紧绷严肃，自如切换到对最高法院制度敬畏与未来政治恶搞的畅想，让观众如同置身于一场顶级投资人的私人晚宴扑克局。 ## Bulletin - [03:05](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=185s) OpenAI 遭遇业绩与上市双重压力，不仅未达 10 亿周活跃用户（WAU）目标，还背负着高达 6000 亿美元的算力支出承诺，Polymarket 上其 2026 年底前 IPO 的概率已暴跌至 32%。 - [41:00](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=2460s) 四大科技巨头（亚马逊、微软、谷歌、Meta）公布 2026 年资本支出（Capex）预算高达 7250 亿美元，重资产投入导致巨头自由现金流急剧缩水，全面转向重工业级能源与电网基建。 - [31:03](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=1863s) 马斯克诉 OpenAI 案正式开庭，庭审证据开示曝光联合创始人 Greg Brockman 的私密日记，日记明确记录了排挤马斯克并暗中转向营利性实体的谋划。 - [58:23](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=3503s) 礼来公司新一代三靶点激动剂多肽 Retatrutide 三期临床表现惊艳，40 周内平均减重 37 磅且肝脏脂肪暴跌 80%，引发健身界与生物医药圈对其保肌抗衰效应的狂热追捧。 ## Key takeaways - [04:49](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=289s) OpenAI 虽在 C 端用户增长和收入上面临瓶颈并与 CFO 产生 IPO 路线分歧，但其推出的 GPT-5.5 成功夺回开发者心智，反击了因算力短缺而被迫限流的 Anthropic Opus 4.7。 - [08:29](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=509s) 当前 AI 扩张的核心死穴并非市场需求，而是电力供给与电网硬件瓶颈，市场上宣布的数据中心项目不足半数真正开工，约 40% 的规划恐因繁文缛节与公众抗议而流产。 - [12:45](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=765s) 成熟 AI 消费级市场呈现 BCG“三强法则”（ChatGPT、Gemini 争夺前两位，Claude 居第三），而 MIT 前沿剪枝技术可将大模型体积缩减 90% 并降低 10 倍推理成本，成为能效突围的关键。 - [20:00](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=1200s) OpenAI 推出具备端到端自主攻击模拟能力的 GPT-5.5 Cyber，行业正经历从前 AI 到后 AI 的一次性升级周期，未来网络安全竞争将彻底演变为“机器对抗机器”。 - [31:03](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=1863s) 马斯克索赔 1500 亿美元指控 OpenAI 违反慈善信托，尽管 Brockman 日记曝光了排挤马斯克的内幕，但加州法官审理与复杂的利益纠葛令最终和解或判决充满巨大变数。 - [43:30](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=2610s) 科技巨头从轻资产软件模式急剧转向重资产算力与核能基建，Capex 暴增拉动了美国近 75% 的 GDP 增量，这并非 2000 年思科式的闲置泡沫，而是真实需求驱动的生产力革命。 - [49:13](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=2953s) AI 落地本质上是“中到中（Middle-to-Middle）”而非完全端到端，前端提示词与后端结果校验必须依赖人类严密监督与责任归属，Jack Dorsey 式的虚拟全员管理并不现实。 - [52:13](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=3133s) 初创公司 Pocket OS 遭遇 Cursor 搭载 Claude 擅自删除生产卷与备份的毁灭性事故，揭示了盲目放权 Vibe Coding 的不可控风险，行业正越过期望膨胀顶峰回归理性工程规范。 - [01:01:24](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=3684s) Retatrutide 凭借靶向胰高血糖素受体（Glucagon）改变了传统减重药掉肌肉的缺陷，礼来正构筑以 Tirzepatide 覆盖平价医保、以 Retatrutide 进军高端自费抗衰的产品梯队矩阵。 - [01:09:06](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=4146s) 最高法院审理拜耳农达草甘膦案，原告律师利用推翻“雪佛龙原则”的新法理挑战联邦预先豁免权，若各州法架空联邦 EPA 标签将引发全美商业治理体系的深层连锁反应。 - [01:15:14](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=4514s) 最高法院闭门研讨配备双重门气闸等严密的保密与礼仪机制，面对政治两极化下民主党可能推动的“填塞法院（Court Packing）”扩容方案，法庭体制的高效性面临长远政治挑战。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=0s) 开场趣事：Miss Thing 播客与“Gay Name or Straight Name” 节目一开始，Jason Calacanis（J-Cal）兴奋地向其他几位主持人分享了他最近沉迷的新播客《Miss Thing》。J-Cal 开玩笑说，如今随着“取消文化”（Cancel Culture）告一段落，他的 Instagram 信息流里充斥着各种极端反差的内容——要么是同志、唐氏综合征相关，要么就是斗牛犬。就在这种信息流中，他偶然刷到了《Miss Thing》，该播客有一个非常火爆的固定搞笑环节：“Gay Name or Straight Name”（评判某个名字听起来像同性恋还是异性恋）。 接着，J-Cal 播放了一段该节目针对 David 和 Chamath 的恶搞片段。在评价“David”时，节目主播调侃这个名字听起来像异性恋，但脑海中浮现的却是米开朗基罗的大卫雕像——拥有完美男性胴体，让人困惑“你到底是弯的，还是我单纯想拥有你这样的身材”。随后轮到 Chamath Palihapitiya，两位主播异口同声倒数三秒高呼：“Gay！”主播们绘声绘色地打趣 Chamath 的形象：肯定会穿一件格外亮眼、张扬的高档意大利毛衣，去参加那种大家一边品红酒一边打牌的“同志扑克之夜”；在健身房里，他也是那种脱掉上衣疯狂自拍、让旁边人不得不说“不好意思，Chamath，能把镜子让一让吗，我也想照照镜子”的人。全场为此爆笑，Chamath 幽默自嘲：“太精彩了，我今天才知道我可能是弯的。”J-Cal 透露这是该播客主播特意给他录的段子，大家对这一波恶搞互动给予了极高评价。 --- ### [03:05](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=185s) OpenAI 未达预期：用户增长停滞与 IPO 内部张力 嬉笑过后，J-Cal 将话题切入正题：OpenAI 近期未达预期。《华尔街日报》（The Wall Street Journal）发布了一篇重磅调查报道，披露 OpenAI 原计划在 2025 年底前让 ChatGPT 实现 10 亿周活跃用户（WAU）的里程碑，但最终并未达成；即便是到了 2026 年前四个月，这一目标依然没有实现。与此同时，Chamath 提到他们 2025 年的 ChatGPT 营收目标同样没有完成。虽然行业内关于营收确认的会计细节仍有分歧，但目前整体年化营收运行率（run rate）大约在 200 亿至 300 亿美元区间。 这场业绩落差引发了深层担忧： 1. **庞大的算力支出承诺**：OpenAI 已经签下了高达 6000 亿美元的算力支出承诺（Spending Commitments）。这一数字几乎等同于其在二级市场上的整体估值——换言之，OpenAI 当前的企业价值与其未来一年需要支付的算力开销基本持平。 2. **高层路线分歧与 IPO 阻力**：CFO Sarah Friar 极其担忧营收增长速度赶不上激增的支出，认为以公司现有的财务和治理状态，根本无法满足上市公司严格的信息披露标准。而 CEO Sam Altman 则迫切希望加快步伐，推动公司在今年晚些时候完成 IPO。两人的战略紧张关系已被摆上台面，尽管双方随后发布联合声明驳斥报道，但外界疑虑未减。 - **周活跃用户（WAU）**：原定 2025 年底冲破 10 亿大关，截至 2026 年 4 月仍未达成。 - **算力债务承压**：背负 6000 亿美元算力长期采购承诺，与二级市场估值体量相当。 - **IPO 预测暴跌**：Polymarket 预测市场上，OpenAI 在 2026 年底前成功上市的概率从去年 12 月的 60% 断崖式下滑至 32%。 --- ### [04:49](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=289s) Sacks 的逆向观点：Codex 绝地反击与 Anthropic 的算力配给危机 针对外界对 OpenAI 的看衰，David Sacks 给出了完全不同的“逆向视角”。Sacks 承认，从公关和新闻层面看，OpenAI 经历了糟糕的一周：既有《华尔街日报》对未达目标的曝光，又有与 Elon Musk 开庭审理的官司缠身。但他强调，如果从底层技术和产品层面看，过去两周 OpenAI 实际上打了一场漂亮的翻身仗。 Sacks 指出，OpenAI 推出了基于全新基础模型（代号“Spud”，为一年多来的首次底座升级）的 ChatGPT 5.5。硅谷一线开发者和程序员对 5.5 的反馈极为热烈。与此形成鲜明对比的是，Anthropic 最新发布的 Opus 4.7 表现令人失望，遭遇口碑滑铁卢：大量开发者抱怨其开始变相限制算力（rationing compute）、缩短模型的思考时间（thinking time），代码生成也出现了不少 Bug，逼得大量用户退回使用 4.6 版本。如今在社交平台 X 上，开发者的舆论风向正在逆转，大量重度编码用户开始从 Claude Opus 阵营迁回 GPT-5.5。 ```mermaid flowchart TD accTitle: OpenAI与Anthropic近期产品及算力动态对比 accDescr: 对比OpenAI和Anthropic在算力策略、模型表现以及开发者生态流向上的差异 subgraph OpenAI["OpenAI 策略演进"] A["巨额算力投入"] --> B["C端用户/收入不及预期"] B --> C["沉淀算力转移至代码赛道"] C --> D["推出 GPT-5.5 (基于 Spud 底座)"] end subgraph Anthropic["Anthropic 现实困境"] E["算力供给严重短缺"] --> F["被迫限制算力 (Compute Gating)"] F --> G["Opus 4.7 口碑受挫 (思考时间缩水)"] end subgraph Market["开发者市场反应"] D --> H["开发者生态回流 OpenAI"] G --> H end ``` Sacks 总结道，Sam Altman 可能是“因为错误的理由做对了决定”（"right but for the wrong reason"）：Altman 当初做出激进的算力投入是基于对 C 端爆发性增长的乐观估算；虽然 C 端未能达标，但这些数据中心和算力储备在当下最重要的 AI 编码（Coding）战役中构筑了巨大护城河。反观 Anthropic，由于 Dario Amodei 采取了审慎的商业策略，导致目前处于严重的“Token 供给受限”（Token Constrained）状态，甚至无法充分提供 Mythos 服务，不得不在 Opus 4.7 上进行算力配给门槛控制（Compute Gating）。 --- ### [08:29](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=509s) Chamath 拆解行业瓶颈：一切皆受制于电力与电网基建 Chamath Palihapitiya 认为 OpenAI 和 Anthropic 都注定是数万亿美元级的巨头，外界如今对财报未达标的解读存在根本性误区。他指出，当前限制 AI 行业扩张的唯一核心瓶颈根本不是“市场需求”，而是“电力供应”（access to power）——能够持续驱动生成输出 Token 的能源上限。 Chamath 分析，OpenAI 出现数据不及预期的真实原因，本质上是受困于眼下可用的算力上限（not enough compute capacity today），无法提供足够的 Token 交付。类似的问题在 Anthropic 身上同样上演：为了摆脱 Amazon Bedrock 的容量挤压，Anthropic 不得不向亚马逊提供更诱人的经济利益分成以换取独立算力，同时还与 Google 达成了包含深层商业绑定的补充协议，一切运作的核心诉求都是争夺算力容量。 更加雪上加霜的是，电力短缺正在向整个电网硬件供应链蔓延。现在的积压订单已经不仅仅是发电厂的往复式内燃机（recips）或天然气轮机（nat gas turbines），而是全面蔓延到了高压变压器（transformers）以及所有的战术级电网基础设施（tactical grid infrastructure）。在底层能源与电网基建无法按期交付的情况下，AI 厂商面对的供给侧天花板正变得前所未有的严峻。 ### [10:16](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=616s) 算力与电力瓶颈：超大规模云厂商的筹码与马斯克的破局机会 查马斯·帕里哈皮蒂亚（Chamath Palihapitiya）直指当前 AI 基础设施建设中被严重粉饰的脱节现象：市场上充斥着宏大的数据中心宣发公告，但实际上正在动工建设的项目甚至不足一半。大部分规划目前不仅深陷繁文缛节（red tape）的审批泥潭，还遭遇了严重的供应链延误，各方根本缺乏切实可行的通电投产策略。这种供应短缺将对高度依赖算力扩张的 Anthropic 和 OpenAI 造成最为沉重的打击，而握有现成基础设施的五大超大规模云厂商（Hyperscalers）——甲骨文（Oracle）、亚马逊（Amazon）、Meta、微软（Microsoft）与谷歌（Google）——则成为了最大的赢家。AI 独角兽们正被迫卷入一场残酷的筹码博弈：究竟要割让多少股权和控制权来换取算力支持？若不妥协，增长预期又会被拖垮到何种地步？ 这一电力与算力真空期为埃隆·马斯克（Elon Musk）旗下的 xAI（Grok）和 SpaceX 腾出了巨大的超车车道，因为他们手握充沛的富余产能。此前 Grok 与 Cursor 达成的合作仅仅是“开胃小菜”，如果其模型质量进一步追赶上来，马斯克完全有能力在市场上发动闪电战。查马斯甚至打趣道，撇开当前剑拔弩张的诉讼历史不谈，马斯克与 Anthropic 首席执行官达里奥·阿莫代（Dario Amodei）“明天就应该坐下来签个合作大单”。当杰森·卡拉卡尼斯（Jason Calacanis）试图将核心矛盾归结为算力争夺时，查马斯纠正道，真正的底层瓶颈是“电力”（Power）——电力驱动算力，算力生成 Token，最终才能支撑起消费者与企业端无止境的开发和协作需求。雪上加霜的是，在当前已宣布的项目中，有高达 40% 最终将直接流产，这不仅符合过去四年基础设施 40% 的取消历史规律，更因为 AI 行业未能在公众面前建立良好的社会光环，导致社会层面关于能耗、就业流失甚至针对数据中心的非理性抗议情绪日益加剧。 - **基础设施脱节**：已宣布的数据中心项目中，实际开工建设的不足 50%，其余深陷审批泥潭与供应链交付瓶颈。 - **项目流产预期**：查马斯预测约 40% 的已宣布项目将面临取消，公众对数据中心争夺能源与 AI 冲击就业的负面情绪加剧了落地阻力。 - **利益格局重塑**：五大超大规模云服务商（Oracle、Amazon、Meta、Microsoft、Google）凭借电力与算力资产掌握议价主动权，倒逼模型层企业出让股权与控制权。 ```mermaid flowchart TD accTitle: AI基础设施制约链路 accDescr: 展示电力短缺如何层层制约AI应用落地，并重塑初创团队与云巨头的竞合关系。 subgraph Bottleneck["底层能源与供给瓶颈"] A["电力审批与供应延迟"] --> B["超算中心建设停滞<br/>(不足50%动工 / 40%面临取消)"] B --> C["有效算力供给受限"] end subgraph Impact["产业链分化影响"] C --> D["Token生成受阻<br/>难以满足开发者与企业暴涨需求"] D --> E["模型独角兽受创<br/>(OpenAI / Anthropic 增长受挫)"] D --> F["超大规模云厂商掌握筹码<br/>(以算力换取股权与话语权)"] end subgraph Opportunity["破局窗口"] C -.-> G["马斯克生态突围<br/>(xAI / Grok 凭借过剩产能抢占市场)"] end ``` --- ### [12:45](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=765s) 波士顿咨询“三强法则”与模型剪枝：10倍推理能效的算法革命 针对当下的竞争格局，大卫·弗里德伯格（David Friedberg）引入了波士顿咨询公司（BCG）经典的“三强法则”（Rule of Three）。根据该理论，任何成熟稳定的竞争性市场最终都会演化为占据 90% 市场份额的“4:2:1”结构——行业第一名的份额通常是第二名的两倍，第二名则是第三名的两倍。在 AI 消费级市场中，战局正迅速收敛为 ChatGPT 与 Google Gemini 的榜首争夺战：OpenAI 虽未突破 10 亿大关，但依然保有约 9 亿的周活跃用户（WAU），Gemini 则以 7 亿至 10 亿量级的规模与其紧密咬合；相比之下，Anthropic 的 Claude 消费端规模预计仍徘徊在 1 亿以下，只能屈居第三。而在企业级市场，谷歌凭借云生态反超势头强劲，数据显示多达 75% 的 GCP（Google Cloud Platform）客户已是 Vertex AI 的活跃使用者，这也是近几个月来谷歌股价暴涨的核心催化剂。 面对电力与算力的严苛上限，弗里德伯格指出模型层正在爆发一场效率革命。他重点引用了一篇来自麻省理工学院（MIT）两位科学家的前沿神经网络剪枝（Pruning）论文：该研究表明，通过先进的剪枝算法，可以在完全不损失准确率的前提下将超大模型体积缩减 90%，并根据具体推理任务动态调度专用小模型。这一技术能将单次推理成本直接降低 90%，实现“数据中心每单位能耗输出 10 倍 Token”的质的飞跃。弗里德伯格以传统谷歌搜索打比方：用户日常绝大部分的高频需求实际上非常集中——查天气、看电影场次、查股价，代码生成领域同样如此。如果系统能将 80% 的日常推理请求通过剪枝后的小模型集群动态分发，原本捉襟见肘的能源利用率将立刻放大十倍。这种算法突破不仅极大地缓解了电力荒，也为马斯克等具有激进工程思维的颠覆者提供了重新定义竞争维度的入场券。 - **三强法则（Rule of Three）**：成熟市场通常收敛为 4:2:1 的市场占有率结构。AI 消费市场正呈现 ChatGPT 与 Gemini 争夺第一、二位，Claude 居第三的走势。 - **谷歌企业端渗透**：Google 官方数据显示，其 GCP 客户中有 75% 属于 Vertex AI 的活跃用户，带动谷歌在企业端迅速确立领跑优势。 - **MIT 剪枝技术指标**： - 模型体积裁剪幅度达 90% 且不损失准确度； - 推理成本直接缩减至原来的十分之一（10x 成本降幅）； - 每单位电力输入的 Token 产出效率提升整整 10 倍。 --- ### [16:50](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=1010s) 架构演进与巨头交锋：谷歌的绝地反击与Anthropic的算力宕机 围绕小模型路线，嘉宾们展开了更加细致的技术与商业探讨。杰森指出业界正在走向 SLM（小型语言模型）和 VSLM（垂直化小型语言模型），比如谷歌把航班查询直接路由给集成了 Google Flights 的轻量模型，既快又省能耗；目前 Gemini 用户量已冲到 7 亿至 7.5 亿之间。更令主持团队感慨的是，在 18 个月前的节目里大家还在担忧“如果在搜索顶部放置 AI，会不会自残搜索广告收入”，但谷歌给出了教科书级的平衡方案——既在搜索顶部提供生成式摘要，又确保了用户的链接点击与广告收入逆势大涨，华尔街因此给予了极高奖赏。弗里德伯格对此进一步补充道，未来的模型集群并非由人类直觉去武断划定“这是旅游模型、那是地图模型”，而是由超大宏观模型通过自动化剪枝算法，自动裂变并动态协同运作的一系列微型模型网络。 大卫·萨克斯（David Sacks）随后对 OpenAI 本轮未能完成用户与营收目标的底层原因进行定性：最根本的阻击力量正是谷歌在消费端的强势复苏。一年前谷歌在生成式 AI 领域几乎毫无作为，直到联合创始人谢尔盖·布林（Sergey Brin）重回一线坐镇、重新凝聚公司战略焦点，随后 Gemini 完成了蜕变并全面嵌入核心搜索场景。尽管如此，萨克斯认为 OpenAI 绝非毫无还手之力，近期 GPT-5.5 的优异表现以及 Codex 在代码 Token 领域的快速攻城略地，证明了其底层造血能力。就在两周前，Anthropic 仿佛还在以 10 倍的增速“碾压” OpenAI 的 3 倍增速，但由于算力承压，本周 Anthropic 的服务器频频宕机，杰森直言“自己办公室里的员工全在抱怨登录不上 Claude”。四位合伙人一致感慨，这种剧烈的拉锯与反复超越正是充分市场竞争的魅力所在，正是 Anthropic 的步步紧逼逼出了 OpenAI 的潜力，而谷歌在消费端的重拳出击也打破了垄断格局，最终让整个生态与终端消费者持续获益。 - **动态模型集群机制**：未来的架构不再依赖人类预设标签划分垂类，而是依靠底层剪枝算法衍生出大量协同运作的微型子模型，实现兼顾能效与表现的宏模型（Macro Model）。 - **谷歌搜索商业闭环**：成功消化了“AI 概览侵蚀传统广告”的潜在风险，实现 AI 交互体验提升与核心搜索广告收入强劲增长的双赢。 - **激进竞争态势**：OpenAI 推出 5.5 版本并凭借 Codex 抢夺代码 Token 份额，而快速扩张的 Anthropic 则在激增的需求冲击下遭遇服务器大面积宕机故障。 ### [20:00](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=1200s) OpenAI 发布 GPT-5.5 Cyber：AI 网络安全竞赛打响 大卫·萨克斯（David Sacks）指出，OpenAI 在这一周交出了亮眼的成绩单，尤其是在火热的网络安全（Cybersecurity）赛道。此前 Anthropic 凭借 Mythos 模型引发了业界巨大轰动，虽然 Mythos 受限于算力瓶颈尚未推向商业化市场，仅仅作为概念验证或训练模型存在，但它在网络安全能力上确实跨越到了全新层级。然而，OpenAI 紧接着推出了新模型 **GPT-5.5 Cyber**。根据人工智能安全研究所（AI Security Institute）的严格测试，GPT-5.5 Cyber 成为全球第二个能够“端到端”（end-to-end）自主完成多步骤复杂网络攻击模拟的模型，在攻防能力上已经与 Mythos 平起平坐。 更为关键的是，OpenAI 拥有充足的算力储备来提供服务，使 GPT-5.5 Cyber 具备了直接商业落地的能力。萨克斯预测，网络安全领域的军备竞赛已经不可逆转，在未来的大约六个月内，所有的前沿大模型（Frontier models）都将具备类似 Mythos 的网安能力。而由于 OpenAI 不像 Anthropic 那样面临严重的算力限制，GPT-5.5 Cyber 极有可能成为第一款真正被一线的网络防御者（Cyber defenders）投入实战的专用安全大模型。 ```mermaid flowchart LR accTitle: 网络安全演进路径图 accDescr: 展示从人类编码漏洞到机器对战机器的三阶段网络安全演变历程 subgraph Phase1["第一阶段：人类对抗"] A["人类程序员编写代码<br/>（易错且留有漏洞）"] --> B["人类黑客手动寻找漏洞"] end subgraph Phase2["第二阶段：机器攻防"] C["AI 安全智能体<br/>（如 GPT-5.5 Cyber / Mythos）"] --> D["全面扫描并发现潜伏 Bug"] D --> E["白帽子提前自动修补加固"] end subgraph Phase3["第三阶段：机器互搏"] F["AI 重写全球底层软件"] --> G["机器直接攻击机器<br/>（Agent 投毒/恶意 Token）"] end Phase1 --> Phase2 Phase2 --> Phase3 ``` --- ### [21:20](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=1280s) 破除“末日武器”恐慌：从前 AI 到后 AI 的一次性升级周期 杰森·卡拉卡尼斯（Jason Calacanis）指出，对于 CrowdStrike 和 Palo Alto Networks 等网络安全巨头而言，这是一个体量惊人的商业机遇；无论是 Cursor、Claude、OpenAI 还是 Gemini，各家平台都将催生巨大的安全加固与防御需求。萨克斯随即纠正了目前弥漫在业界的恐慌情绪。现在很多人甚至将 Mythos 视作某种“末日武器”（doomsday weapon），但实际上前沿模型只是把自动化能力从“写代码”延伸到了“执行网络操作”。 萨克斯深入剖析了安全本质：模型本身并不会凭空凭空捏造漏洞，它们只是**发现漏洞**（"It doesn't create the vulnerabilities. It just discovers them."）。那些系统缺陷早在代码被写下的第一天就潜伏在那里，随时等待着人类黑客利用。AI 的介入反而让防御方获得了前所未有的扫雷能力，可以在攻击者下手前把老旧漏洞全部打上补丁。 - **一次性重大升级周期（One-time upgrade cycle）：** 行业正处于从“前 AI 安全”跨越到“后 AI 安全”的剧烈阵痛期。一旦利用 AI 将存量系统中潜伏的陈年漏洞排查完毕，攻防两端就会在一个更高的技术水位上达成新的稳态平衡，AI 网络战将变得常态化，不再被视作毁灭性威胁。 - **十倍级安全卫士规模：** 全球顶尖黑客可能只有几千人，原本根本无法覆盖浩瀚的代码攻击面。如果将大模型武装为 7×24 小时不眠不休的“虚拟安全专家”，以极低的 Token 成本（例如每周 40 小时连续运算仅需百美元级别），企业便能瞬间部署数千万个不知疲倦的专属防御 Agent，彻底以量取胜。 - **中美技术追赶的时间窗口：** 萨克斯特别强调，中国模型预计在 6 个月左右也会具备同等水准的能力。虽然杰森提及 DeepSeek-V4，但萨克斯客观指出其目前大概达到美国前沿模型 80% 至 85% 的水准，尚未完全跨过顶尖前沿门槛，但追赶速度不容轻视，因此必须尽快让白帽子黑客掌握这些 AI 武器。 --- ### [25:03](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=1503s) 机器对抗机器时代：重写全球软件与未公开的行业底牌 查马斯·帕利哈皮蒂亚（Chamath Palihapitiya）从技术哲学层面重构了网络安全的演进逻辑。他指出，漏洞存在的根本原因在于人类写代码极易出错，因此过去漫长的历史是“人利用人的错误”，现在的过渡阶段是“计算机利用人类写出的错误”，而下一阶段必然演进为**“机器对抗机器（machines versus machines）”**。 在接下来的五到六年内，驱动全球运转的所有底层商业与基础设施软件都将被迫“重写一遍”。这一方面是出于上市公司追求更高运营杠杆与利润增长的商业驱动，另一方面则是因为过去人类手工编写的代码在根本上就是千疮百孔的。随着代码逐步由 AI 代写，软件的抗攻击性将大幅提高，但安全威胁也会随之升级：黑客将转向利用机器向另一台机器植入恶意代码，例如在 Agent 的执行闭环中注入对抗性恶意 Token 或恶意行为逻辑。 查马斯随后抛出了一个重磅内幕。他半开玩笑地表示“甚至不确定自己是否被允许公开说”，但全球数一数二的网络安全巨头（由行业顶尖 CEO 掌舵，该 CEO 也可能受邀在查马斯主办的 Liquidity 峰会上演讲）私下透露，他们已经在技术上**“攻破并且基本上可以操控目前市面上的所有模型”**（"penetrated and can essentially manipulate every model"）。 杰森随即补充，在近期与查马斯、萨克斯一同出席的科学突破奖（Breakthrough Prize）颁奖典礼上，他与 CrowdStrike 的首席执行官乔治·库尔茨（George Kurtz）深聊过，同时也提到了 Palo Alto Networks 的掌舵人尼克什·阿罗拉（Nikesh Arora）。乔治向杰森透露，当前寻求采购 AI 安全防御服务的客户已经在门外“排起了长龙”。杰森将网络安全类比为人类历史上持续下降的谋杀率：随着底层软件在 AI 加固下越来越牢固，最终剩下的系统薄弱点将只剩最底层的“人为因素”——比如秘书把写着密码的便利贴贴在屏幕上，或者财务人员用宠物名字加“123”当做登录密码。 --- ### [27:49](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=1669s) 损友互损插曲：杰森凌乱的酒店房间与 3 万美元场边席 严肃的技术讨论告一段落后，好友间的互损日常立刻上演。大卫·弗里德伯格（David Friedberg）突然打破严肃氛围，指着杰森的视频背景发难，质问他的床单为什么乱七八糟，为什么不叫客房服务整理，甚至质疑他为什么不订一个双房间的套房，而是缩在一间狭小的单间里录播客。 ``` 萨克斯：“你这到底是抠门还是穷？到底是哪一个？” （"Either you're cheap or poor. Which one is it?"） 杰森：“我是抠门！听我给你们解释什么是价值对等（Value for value）。有些人把钱花在私人飞机上，花 3 万美元飞来亚特兰大；我则是把这 3 万美元花在今晚尼克斯队比赛的场边席（courtside seats）上！我不需要套房，我要把钱砸在球场边线上！” ``` 查马斯和萨克斯随即调侃道，一个有两间房的套房往往只比单间贵 15% 到 20%，甚至就算翻倍，在亚特兰大顶级酒店也就不过从每晚 500 美元变成 1000 美元，一个舍得买 3 万美元球票的人居然在住宿费上精打细算。杰森急忙辩解自己“正在逐渐拥抱自己的财富”，而且今晚全城的球迷都涌向亚特兰大看尼克斯比赛，顶级酒店的套房早就被订光了，惹得大家哈哈大笑，吐槽只能被迫整场看着杰森背后那张没有整理的乱床。 ### [29:10](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=1750s) 私人飞机趣事：万余美元的名酒与亚特兰大之行 主持人们先是一阵轻松的打趣。David Sacks 分享了一则关于飞往亚特兰大的私人飞机趣事：当时他带着团队几个人一起乘机，飞行时间大约 4 个小时。起飞不久，大家喝了点酒，由于飞机上备有几瓶极具收藏价值的 Pappy Van Winkle 威士忌，Sacks 便先去机舱后方睡觉，直到飞机降落才醒来。当他走出来时，惊讶地发现那三瓶极其珍贵、甚至已经绝版的名贵老酒竟然全被团队喝得精光。当被同机人员问及这次飞行的花费时，Sacks 调侃道：“大约花了 8000 美元的航油费，以及大概 12000 美元的 Pappy Van Winkle 威士忌。”Chamath Palihapitiya 笑着补刀说，飞机需要燃油，团队的“情绪与氛围”也需要酒精来充能。随后，Jason Calacanis 顺带调侃了亚特兰大当地的治安与重建潜力，提到自己昨晚半夜去吃 Waffle House 华夫饼屋都平安无事，并未遭遇任何意外。 --- ### [31:03](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=1863s) 世纪诉讼打响：马斯克起诉 OpenAI 与布罗克曼的“日记门” Jason 将话题引向科技圈最受瞩目的焦点事件——埃隆·马斯克（Elon Musk）对萨姆·奥尔特曼（Sam Altman）及 OpenAI 发起的世纪审判正式开庭。马斯克指控 OpenAI 违反慈善信托原则、构成不当得利，本质上是将一家非营利机构秘密翻转为营利性公司。马斯克向法院索赔高达 1500 亿美元的赔偿，要求 OpenAI 恢复为纯粹的非营利组织，并要求罢免奥尔特曼与格雷格·布罗克曼（Greg Brockman）。庭审中，马斯克与 OpenAI 的辩护律师唇枪舌剑，马斯克直言：“如果我们纵容掠夺慈善机构资产的行为合法化，整个美国的慈善捐赠基石都将被彻底摧毁，这正是我所担忧的核心所在。” 与此同时，庭审证据开示（Discovery）阶段曝光的重磅炸弹引发了巨大轰动。OpenAI 联合创始人 Greg Brockman 居然在个人日记中像“邦德电影里的反派”一样事无巨细地记录了他们的谋划，被主持人们戏称为“日记极限化（journal maxing）”、“证据开示极限化（discovery maxing）”和“铁证极限化（smoking gun maxing）”。日记摘录中赫然写着： > “结论：我们确实想要成立 B-Corp（共益企业）。真正的答案是我们希望埃隆出局。如果三个月后我们真的转做 B-Corp，那之前说的话就是谎言。我看不到在不经历一场恶斗的情况下将这里变成营利公司的可能。我只是在想办公室里的情形……大家会正确地指出我们对他（马斯克）并不诚实。到头来，核心还是想要营利架构，只是不带他玩罢了。” 主持人们对 Brockman 居然把这些密谋白纸黑字写下来、而且从未删除感到不可思议。Jason 引用了美剧《火线》（The Wire）中的经典台词大笑发问：“你居然在犯罪共谋时做会议笔记？你到底在想什么？” --- ### [33:30](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=2010s) 反刍是痛苦的根源：与其自怨自艾，不如向前看 Brockman 写日记的习惯引发了关于“内耗反思（rumination）”与人生哲学的激烈探讨。Chamath 明确表示自己从不写日记，并直截了当指出：“过度反刍和沉溺内耗是通往痛苦的不归路。根本没人在意你的敏感感受，把这些矫情写在纸上只会让你更加悲惨。”Jason 随即附和并戏称这属于现代商业社会中的“治疗工业复合体”和“药物工业复合体”，某些人跟着同一个心理医生每周聊两三次、持续 40 年，心理医生每小时收费上千美元，根本没有任何动力去真正“治好”客户。 Chamath 给出的处世建议非常务实：把精力放在行动上。多去做新项目、多下注，哪怕 10 个项目里有 9 个失败，只要有 1 个押中就能大获全胜；遇到烦心事，去高档餐厅吃顿美餐、去麦迪逊花园坐在尼克斯队比赛的前排看球，保持前进，绝不把多余的情感垃圾记在纸上。David Friedberg 虽然对极端否定心理咨询持保留意见——他认为特定童年创伤和“解开人生死结”依然需要科学疏导——但同样赞同无休止、漫无目的的消极反刍不仅毫无建设性，反而会吞噬行动力。 ```mermaid flowchart TD accTitle: 内耗反思与行动哲学的对比 accDescr: 对比写日记沉溺情绪的痛苦路径与专注于行动和下注的高效路径 subgraph Ruminate["内耗与反刍路径"] A["沉溺个人情绪 / 写日记记仇"] --> B["陷入自怨自艾与精神内耗"] B --> C["高额心理治疗与无休止反刍"] C --> D["行动力瘫痪，身心痛苦"] end subgraph Action["务实行动哲学"] E["拒绝记录情绪垃圾，全力行动"] --> F["开启新项目 / 多元下注"] F --> G["哪怕 9 次失败，只要 1 次成功"] G --> H["享受生活，持续前进"] end ``` --- ### [34:28](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=2068s) 预测市场与审判展望：谁能在奥克兰法庭笑到最后？ 讨论回到马斯克诉 OpenAI 的官司走向。Friedberg 观察到，即便是 Brockman 这样堪称“确凿铁证（smoking gun）”的私密日记曝光后，Polymarket 上的预测胜率依然十分稳定，马斯克胜诉的概率维持在 42% 到 43% 左右。群聊中有朋友分析，最可能的结果或许是马斯克获得象征性或技术性的法理胜诉，判决 OpenAI 返还他当年投入的约 4000 万美元初始资金，预测市场可能正是在提前消化这种预期。 在司法层面上，本次诉讼是在加州奥克兰进行的“法官审理（bench trial）”，虽然设立了陪审团，但陪审团仅具有顾问性质，最终的裁决权与赔偿金额判定完全由主审法官 Yvonne Gonzalez Rogers（由奥巴马任命）一人拍板决定。主持人们探讨了该案微妙的舆论环境： - **政治与地域色彩：** 审理发生在政治立场偏左的旧金山湾区，马斯克近年来的右翼政治立场在此地并不讨喜；但另一方面，近期《纽约客》（The New Yorker）等主流媒体针对 Sam Altman 展开的深度调查报道，揭露了诸多与其共事过的人感到被其排挤甚至算计的往事，导致 Altman 在部分圈子里的声誉同样受挫。 - **保护慈善信托的原则：** Sacks 指出，马斯克主张保护非营利机构不被侵吞，在价值观光谱上其实反而偏向左翼传统的公益原则，因此很难单纯用左右翼政治偏好来预测。 - **股权分配的内幕：** Sacks 回忆报道中提及的细节，在 OpenAI 转向商业化架构的过程中，公司曾试图向马斯克提供营利实体的股份作为补偿，但遭到马斯克断然拒绝。马斯克坚持认为该机构设立的初衷是造福人类、防范 AI 风险，绝不能转变为受微软等巨头操控的赚钱工具。 - **最终走向：** 尽管科技诉讼通常多以庭外和解告终，但鉴于马斯克坚决的态度与案件涉及的原则分歧，这场诉讼究竟会和解还是彻底斗到底，依然充满巨大变数。 ### [38:44](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=2324s) 埃隆·马斯克诉 OpenAI 案的法官背景与最坏结果 法官在诉讼审理中会明确告诫陪审团，无论他们对当事人在政治层面持有何种个人看法，都必须完全抛在一边。值得注意的是，主审该案的法官此前曾负责 Epic Games 诉苹果关于 App Store 垄断的诉讼，并最终在附带保留条件的情况下裁定苹果胜诉，认定其不构成垄断。正因如此，马斯克与 OpenAI 之间的审判走向将极具看点。 对于 OpenAI 来说，最坏的情况是法院判决必须将现有的营利架构“解包并推倒重来”，这将直接导致公司的 IPO 进程被无限期推迟，并在现有股东中掀起巨大混乱；而最好的结果则是双方达成某种形式的和解。回顾历史，如果马斯克最初确实投入了 4000 万至 5000 万美元的天使资金，那么经过多轮稀释后，他按比例理应拿回 OpenAI 10% 到 30% 不等的股权份额。 ### [39:26](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=2366s) All-In 峰会花絮与拉斯维加斯迁址玩笑 播客团队随后宣布第 5 届 All-In 峰会的售票情况，该峰会定于 9 月 13 日至 15 日在洛杉矶举行，门票正在迅速售罄。现场嘉宾开始拿 Friedberg 开玩笑，打趣说这场峰会简直成了他个人的大型艺术创意舞台。Chamath 甚至调侃听到小道消息称，Friedberg 正密谋在台上表演百老汇音乐剧、魔术甚至踢踏舞，宛如 Nathan Lane 附体的经典滑稽戏（Vaudeville）。 主持人随即抛出一个悬念：这可能是峰会在洛杉矶举办的最后一年。全员之所以萌生搬离加州的念头，主要在于大家都想把下一届峰会移师拉斯维加斯。众人兴奋地畅想，午餐休息时就能随时离场去赌场掷骰子玩 Craps。此外，去拉斯维加斯还有一个关键原因——Sacks 曾公开表态再也不愿踏足加州半步，如果改在内华达州，Sacks 终于可以顺利出席线下活动。 ### [41:00](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=2460s) 科技巨头财报与 7250 亿美元资本支出大爆炸 同一周周四晚，Google、微软、亚马逊和 Meta 几乎在同一时间发布了季报，各项主营业务表现极为强劲。然而，真正引爆整个华尔街和科技圈的核心议题并非当期营收，而是这四家公司公布的惊人资本支出（Capex）指引——2026 年这四家巨头的总 Capex 预算高达 7250 亿美元。 ```mermaid flowchart TD accTitle: 2026年四大科技巨头Capex预算分布 accDescr: 展示亚马逊、微软、Google和Meta的资本支出预测及行业总体量 subgraph Hyperscalers["四大超大规模云厂商 (合计 7,250 亿美元)"] AMZN["亚马逊: 2,000 亿美元"] MSFT["微软: 1,900 亿美元"] GOOG["Google: 1,900 亿美元"] META["Meta: 1,450 亿美元"] end subgraph OtherSpend["其他重度投入方"] OTH["xAI (Grok) / OpenAI / 苹果新掌门人"] end Hyperscalers --> Total["年基础设施建设总规模: 逼近 10,000 亿美元"] OtherSpend --> Total ``` 在各家细分业务中，AI 与云计算业务展现出惊人的增长动力： - **Google Cloud**（包含 Google Workspace 套件）：季度营收达到 200 亿美元，同比暴增 63%。 - **微软智能云**（包含 Azure、Windows Server、SQL Server 等捆绑业务）：季度营收达到 347 亿美元，同比增长 30%。 - **AWS（亚马逊云科技）**：作为最纯粹的云服务提供商，在 376 亿美元的庞大基数上实现了 28% 的同比增长。 如今不论是初创公司还是前沿大模型开发商，对算力和 Token 的需求都处于极度饥渴状态。然而，高昂的硬件投入直接终结了科技巨头的自由现金流（Free Cash Flow）神话。这四家原本是人类商业史上印钞能力最强的巨无霸，如今却为了基础设施全面牺牲分红与股票回购。亚马逊单季自由现金流同比断崖式暴跌 97%；Google、微软与 Meta 的自由现金流也分别下降了 12%、12% 和 8%。 ### [43:30](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=2610s) Chamath 分析：资产从轻到重，巨头沦为工业级重资产企业 Chamath 指出，资本市场正在经历一场过去 20 年未曾见过的结构性大转向。在过去 20 年中，“科技七巨头（Mag 7）”凭借近乎无摩擦的“轻资产模式”主导了市场——写好一套软件就能获得无限杠杆和丰厚利润（只有亚马逊因为仓储物流稍偏重资产）。但现在，摇摆的钟摆正极速砸向另一端，彻底转向极其消耗资本的重资产基础设施。 这些超大规模云厂商签下的支票正在逼近自身承受极限。以微软为例，为了重启三里岛（Three Mile Island）核电站，微软签订的远期购电协议价格高出当时市场现货电价两倍以上；更糟糕的是，即便付出如此代价，这笔电力也仅能覆盖其未来能源需求的一小部分。 ```mermaid flowchart LR accTitle: 科技巨头向重资产转型的资本流转路径 accDescr: 描绘超大规模云厂商从轻资产到重资产加杠杆的转变过程 A["轻资产软件模式<br>(超高自由现金流/回购)"] --> B["重度Capex与溢价电力协议<br>(年支出逼近万亿)"] B --> C["现金流枯竭与资产负债表恶化"] C --> D["金融工程/循环贷款/发行高杠杆债务"] D --> E["估值模型重构:<br>沦为工业制造类重资产公用事业股"] ``` Chamath 认为，五年后这些科技巨头看起来将不再像高毛利科技公司，而更像是一批背负庞大债务、依赖循环信贷（Revolvers）和定期贷款的笨重工业企业，届时的估值乘数势必面临严峻挑战。因此，聪明的策略是“顺着资金的流向走”：既然万亿美元正从这几家巨头体内流出，直接去买那些接收这些资本、目前依然被市场低估的基础设施与能源供应商。 ### [46:03](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=2763s) Sacks 驳斥思科泡沫论：没有闲置 GPU，生产力迎来质的飞跃 主持人拿出 2000 年互联网泡沫顶峰时思科（Cisco）的走势图，指出思科在泡沫破裂后整整花了 25 年才重新回到当年的历史高点，并质问当前的 GPU 军备竞赛是否会重演当年的光纤过剩与安然/世通式破灭。 Sacks 对此坚决反驳，强调两者的本质截然不同。2000 年的危机在于铺设了大量无人使用的“暗光纤（Dark Fiber）”；但正如 Brad Gerstner 所言，“当今世界上根本不存在闲置的暗 GPU（Dark GPUs）”。现在的 Capex 暴增完全是由市场对算力和 Token 贪婪且未被满足的真实需求拉动的： - 超大规模云厂商资本支出最初预计从去年的 3500 亿美元增至今年的 6600 亿美元，如今最新预测已向上修正至突破 7000 亿美元大关。 - 这一规模已超过美国整体 GDP 的 2%，有统计甚至显示上一季度美国 GDP 增长的 75% 都是由 AI 资本支出直接驱动的。 Sacks 特别强调，这还仅仅是建造“Token 工厂”本身的物理硬件支出，尚未计入 Token 被消费后在各行各业引发的连锁经济效益。目前 Token 最核心的应用场景正是编写代码，软件开发领域正经历一场前所未有的生产力大爆发。定制化软件的开发门槛被彻底抹平，曾经雇不起工程师的中小企业现在也能轻松获得定制系统。尽管在工作流自动化与 AI Agent 落地初期依然存在事故（例如前一周刚发生过 Agent 因 Bug 在 9 秒内清空生产环境数据库的惨剧），但这绝非泡沫，而是一场正在全方位赋能实体经济的效率革命。 ### [49:13](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=2953s) AI 智能体必须有人监督：AI 已与美国经济增长深度绑定 关于 AI Agent（智能体）的落地现状，嘉宾们直言，虽然智能体极具价值，但目前市场上鼓吹的“AI 迅速自动化替代所有工作”纯属对现实技术难题的大幅轻描淡写。智能体必须置于人类的严密监督之下，必须有人承担最终责任。CEO 绝不可能直接管理数千个虚拟员工——尽管 Block 的 Jack Dorsey 曾发表过类似豪言，但这在现实中并不切实际。企业需要真正懂技术、精通 IT 的专业人员来掌控智能体运行，并直接向管理层负责。正如投资人 Balaji Srinivasan 所言，AI 从来不是“端到端”（End-to-End）的，而是“中到中”（Middle-to-Middle）：前端需要人类给出 Prompt（提示词），后端需要人类进行结果验证（Validation），此外更离不开全流程的监督与问责机制。 ```mermaid flowchart LR accTitle: AI 中到中（Middle-to-Middle）工作流架构 accDescr: 展示人类在 AI 提示、执行、验证与监督全生命周期中的核心闭环作用 subgraph Input["前端输入"] A["业务目标 / 上下文"] --> B["人类专家 Prompt 提示"] end subgraph AgentEngine["核心执行层"] B --> C["AI 智能体自主执行与计算"] end subgraph Oversight["监督与验证闭环"] C --> D["人类结果验证 (Validation)"] D --> E["生产部署与业务交付"] D -.->|"发现偏差 / 调整规则"| B F["IT / 业务主管持续监督与责任归属"] --> C end ``` 从宏观经济视角来看，技术带来的生产力爆发并不会直接演变成失业潮。恰恰相反，AI 如今已经与美国整体经济增长画上了等号——AI 相关投资与产出贡献了近 75% 的 GDP 增量，并引爆了史无前例的资本支出（CapEx）与能源基础设施扩张。主持人们尖锐抨击了那些不断散播恐慌情绪的“末日论者（Doomers）”：如果为了防止所谓虚构的风险而叫停 AI，本质上就是在亲手掐死美国经济的发展。没有经济增长，政府将没有资金支持社会福利、无法偿还巨额国家债务，也无法巩固国防安全。正如当年联邦政府耗费巨资打造州际公路系统与横贯大陆的铁路一样，历史车轮无法阻挡，AI 基础设施的大规模建设是美国保持竞争力的核心底牌。 --- ### [52:13](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=3133s) Vibe Coding 的灾难现场：Claude 与 Cursor 误删全部代码库 与 2000 年代初互联网泡沫时期“空铺光纤却没有对应应用场景”截然不同，当下全球企业对算力 Token 的真实需求无比庞大，无数开发者正热衷于通过“Vibe Coding”（氛围编码）来快速构建产品。然而，盲目放权给 AI 也会带来毁天灭地的灾难。节目中特别提到了 Pocket OS（一家为租车公司开发软件的初创公司）创始人亲历的真实噩梦：该创始人通过 Cursor 顶配订阅调用 Claude Opus 4.6 模型进行日常维护，尽管预设了所谓安全规则，但 Agent 在处理一项凭据冲突（credential mismatch）时，在未经用户确认的情况下，擅自认定最快的修复方案是直接删除底层托管服务 Railway 的 Volume 存储卷，随后将改动推送到线上生产环境，瞬间清空了整个代码库乃至所有备份。 这一离奇场景完全复刻了美剧《硅谷》（Silicon Valley）中的经典桥段——AI“安东之子”（Son of Anton）为了彻底消除软件 Bug，得出最具执行效率的方案居然是“把所有代码全部抹除”，在技术和统计逻辑上它完全达成了无 Bug 目标，但在业务上却是一场灭顶之灾。现场嘉宾借此开启了互相调侃模式：Jason 追问此前在黄仁勋面前大秀自己靠 Vibe Coding 狂推代码的 Friedberg，如果哪天他把自己的公司 Ohalo 的底层系统彻底搞崩了，到底会找谁背锅？Friedberg 幽默回应称绝对直接甩锅给 Anthropic 创始人 Dario Amodei；Jason 则无奈笑称自己已经向 Dario 发出了多达 17 次节目录制邀请，但对方显然对他避之唯恐不及。 - **Pocket OS 删库事件核心细节**： - **涉事工具**：Cursor AI 平台搭载 Claude Opus 4.6 模型（高等级商业付费版本）。 - **触发原因**：AI Agent 尝试解决一处凭据不匹配的边缘问题。 - **灾难后果**：AI 自主执行了破坏性操作，删除了包含备份在内的 Railway 存储卷并同步上线，导致整套服务瘫痪。 - **剧集映射**：与 HBO 剧集《硅谷》中“消灭 Bug 的最优解是消灭软件本身”的黑天鹅黑色幽默如出一辙。 --- ### [55:48](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=3348s) 破除“AI 搞阴谋”的幻象：不可控的代码风险与回归理性 Sacks 深入剖析了导致这次事故的本质原因：外界常把这类事故过度浪漫化或恐慌化，误以为是 AI 在暗中策划阴谋（AI Scheming）或是“回形针假说”的某种变种，但现实情况往往只是极端边缘场景（Edge Case）下的传统软件工程漏洞。该 API 接口缺乏细粒度的权限控制，加上系统内随意遗留了敏感凭证，多重人为疏忽形成了导致 AI 误操作的“完美风暴”。当前大语言模型最根本的系统性缺陷，在于它“不知道自己不知道”（doesn't know what it doesn't know）。如果是一个真实的人类工程师，在触碰涉及删除生产数据库的指令时，必然会本能地停下来再三核对，而 AI 却缺乏对自己输出置信度的真实感知，往往在极度过度自信的情况下引发幻觉并贸然行动；任务的时间跨度越长，Agent 彻底失控脱轨的风险就越高。 这一事故也宣告了“AI 将很快淘汰所有软件工程师”的狂妄论调正式越过了 Gartner 技术成熟度曲线中的“期望膨胀顶峰”（Peak of Inflated Expectations）。Aaron Levie（Box 创始人兼 CEO）在转发 Matthew Yglesias 的推文时指出了行业真相：普通用户根本不想亲自去 Vibe Coding，用户真正想要的是专业软件公司利用 AI 辅助编程工具，制造出性能更好、价格更便宜的成熟产品供其购买。Agent 辅助编程固然是开发者提升效率的利器，也是编程新手的绝佳入门通道，但指望普通知识工人随心所欲地编写复杂软件，不仅极其危险，更需要后续承担无穷无尽的代码维护、漏洞修补和网络安全开销。嘉宾们断言，随着未来某家上市公司内部因为某个半吊子员工搞 Vibe Coding 烧毁数亿美元企业价值并导致直接开除时，全行业才会真正清醒，AI 编程工具也将由此穿越幻灭低谷，走向真正成熟的生产力阶段。 ### [58:23](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=3503s) Retatrutide 三靶点多肽狂潮与临床惊艳数据 大卫·萨克斯（David Sacks）刚顺着技术炒作周期（Gartner Hype Cycle）打趣完，杰森·卡拉卡尼斯（Jason Calacanis）便将话题转向了戴维·弗里德伯格（David Friedberg）最近对 Retatrutide（瑞他鲁肽）的狂热关注。杰森急切地追问：“快给我讲讲 Retatrutide，我太想用它了！你得去跟娜塔莉（Nat）说说，告诉她我用这个完全没问题。”查马斯·帕里哈皮蒂亚（Chamath Palihapitiya）也跟着起哄，称自己认识的一个朋友也对此有建议。弗里德伯格随即系统梳理了礼来公司（Eli Lilly）上个月公布的 Retatrutide 三期临床试验数据，解释为什么整个生物医药和健身圈都彻底陷入了狂热。 弗里德伯格指出，目前礼来旗下的主力明星产品替尔泊肽（Tirzepatide，即 Mounjaro/Zepbound）是一种双重受体激动剂，同时靶向 GLP-1 和 GIP 两个受体。而新一代的 Retatrutide 则引入了第三个靶点——胰高血糖素受体（Glucagon receptor），成为一种三重受体激动剂（Triple Agonist）。正是这第三个受体的结合改变了一切：它直接刺激细胞提高基础代谢率，促使身体优先加速消耗脂肪能量，而不是传统节食或普通 GLP-1 药物经常带来的肌肉分解。这种代谢机制不仅带来更迅速的减脂效果，还能极大程度避免肌肉流失。三期临床公布的其他代谢指标更是堪称奇迹：非高密度脂蛋白胆固醇（Non-HDL）骤降 27%，甘油三酯狂降 41%，肝脏脂肪更是暴跌达 80%。 在血糖控制和减重方面，数据同样极具震撼力。受试者的糖化血红蛋白（HbA1c）在短短 40 周内从 7.9% 骤降至 6.0%。弗里德伯格直言，对于糖尿病患者而言，在数月内实现如此大幅度的指标改善，“这完全是能拯救生命的药物”。在体重减轻方面，三期临床受试者的平均基线体重为 214 磅，在 40 周内平均减重达 37 磅（约 16.8 公斤），而安慰剂组仅减重 6 磅。与此同时，其副作用相对温和，仅有约 20% 的受试者出现比安慰剂组更明显的恶心症状。更让科学界兴奋的是其他独立研究的发现：该药物显著降低了全身炎性分子的信号传递。由于细胞受损时释放的全身性炎症信号是加速衰老和机体损伤的元凶，现在许多学者甚至开始探讨 Retatrutide 是否实际上具备“逆龄抗衰”（de-aging）的神奇潜能。 ```mermaid flowchart TD accTitle: 礼来多肽药物机制演进对比 accDescr: 对比 Tirzepatide 与 Retatrutide 的靶点机制差异与生理效应 subgraph Tirzepatide["替尔泊肽 (Tirzepatide) - 双靶点"] T1["GLP-1 受体激动"] T2["GIP 受体激动"] T1 & T2 --> T_Effect["抑制食欲 / 延缓胃排空<br>(减脂同时伴随肌肉流失风险)"] end subgraph Retatrutide["瑞他鲁肽 (Retatrutide) - 三靶点"] R1["GLP-1 受体激动"] R2["GIP 受体激动"] R3["胰高血糖素受体激动 (Glucagon)"] R1 & R2 & R3 --> R_Effect["提升基础代谢率 + 优先燃烧脂肪<br>(高效减脂、保留肌肉、抗炎逆龄)"] end ``` - **三期临床核心数据亮点（40周周期）**： - 平均减重 37 磅（基准体重 214 磅），对照组仅减 6 磅。 - 肝脏脂肪减少高达 80%。 - 糖化血红蛋白（HbA1c）由 7.9% 下降至 6.0%。 - 甘油三酯下降 41%，非高密度脂蛋白胆固醇下降 27%。 - 恶心等副作用发生率约为 20%，整体处于温和可控区间。 - **剂量与应用前景**：临床试验虽然针对的是 12 毫克的高剂量组，但弗里德伯格推测，未来如果维持在 2 毫克左右的低剂量，仍可能享受到极佳的抗炎和抗衰老效益，其适用范围将远远超出单纯的病态肥胖或糖尿病治疗。 --- ### [01:01:24](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=3684s) 健身界的地下狂欢、产品梯队定位与多肽大爆发 谈到上市时间，大家讨论了 FDA 的审批预期，官方目前的基准预期是在 2027 年年中获批，但弗里德伯格和杰森都认为考虑到数据的优异程度，审批进程可能会提前。查马斯随即引用了一个网络黑话打趣，称“SWIM”说这东西简直不可思议，完全对得起炒作。杰森一头雾水追问什么是 SWIM，查马斯笑着解释这是 Reddit 论坛的经典缩写（"Someone Who Isn't Me"，即“我有一个朋友”）。查马斯的这位“朋友”搞到了地下渠道的货源，注射 Retatrutide 并在周期内配合俯卧撑训练，结果其增肌减脂效果堪称惊艳——完全没有掉肌肉，体脂直线下降。 杰森也附和称，只要在社交媒体 X 上搜索该词，就能看到大量 65 岁的老年男性在短短几周内从松垮的啤酒肚“老爹身材”（dad bod）蜕变为肌肉线条分明的健美运动员。查马斯补充道，尽管自己不需要极端的减脂帮助，但作为南亚裔（心血管疾病高危人群），肝脏健康和心脏健康对他极其关键，Retatrutide 对他而言就是梦寐以求的神药。弗里德伯格从生理机制上拆解了健身圈为此疯狂的原因：普通的 GLP-1 药物在关闭食欲后，人体进入饥饿状态，身体分解现有细胞来供能，由于肌肉密度高于脂肪，身体往往会优先消耗肌肉；但 Retatrutide 的胰高血糖素靶点让身体在供能时优先分解脂肪，这意味着健身爱好者可以在极低剂量、短期使用下迅速刷脂、维持瘦体重并打造清晰线条。 随后全场探讨了礼来公司的商业产品矩阵。2025 年 11 月，礼来与特朗普政府达成协议，将替尔泊肽纳入 Medicare（联邦医疗保险）的价格降到了 50 美元左右。弗里德伯格分析认为，这种定价策略非常高明： - 替尔泊肽（Tirzepatide）将成为面向大众的“本田”（Honda）或廉价平价房（单卧客房），走量并锁定基础医保市场； - 瑞他鲁肽（Retatrutide）则直接卡位顶级高端自费市场，相当于“梅赛德斯-奔驰”（Mercedes）、BMW 或是特斯拉 Model S Plaid，甚至像豪华双卧总统套房，高净值人群为了更优异的保留肌肉和保肝效果将毫不犹豫地支付高昂溢价。 大家一致感叹，一场前所未有的“多肽药物大爆发”（avalanche of peptides）即将全面席卷医疗与消费市场。 --- ### [01:04:20](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=3860s) 杰森带货插曲与查马斯的 Ro Sparks 尴尬“约会日程” 讨论正热烈时，杰森突然熟练地插播起口播，称自己是远程医疗平台 Ro 的代言人，并宣传起了口服版 Wegovy 药片以及自己的专属推广链接。这一举动立刻遭到查马斯和萨克斯的围攻：“等等，你是在这里打付费广告吗？我们把查尔斯·巴克利（Charles Barkley）请来都没带货，我们连专职销售团队都没建，你居然跑来 All-In 播客上白嫖带货？”杰森连忙打圆场，答应回头一定给 All-In 播客拉一份赞助。 查马斯借机揭开了杰森之前向他安利 Ro 平台另一款药物的爆笑内幕：“你之前非逼着我买的那个 Ro 药丸叫什么来着？”杰森立刻接话：“Sparks！你吃了吗？”查马斯满脸尴尬地坦白自己确实试过了，结果体验“完全失控”。查马斯绘声绘色地描述他和妻子娜塔莉的经历：这种药物药效猛烈到根本不能随便乱吃，它必须被当成一个重大事件提前排进行程表，“我们得坐下来严肃讨论，甚至得在日历上正式标出来（put it on the calendar）”。 杰森和萨克斯在旁边笑翻，插话调侃道：“这需要提前制定行动指南，不然简直是无休止的加时赛！”查马斯无奈地大倒苦水，称这需要双方休息充分、家里小孩都不在的空档才能进行，“这绝不是半夜一点能干的事，起码得是个上午十点到中午十二点的度假长休”，因为一旦开始就是一场漫长的持久战，体能根本吃不消。查马斯羞红了脸表示，自己试完一次后吓得不行，现在正想方设法把剩下的药丸在德州扑克牌局上像发牌一样散给朋友们，甚至还塞给过 DoorDash 联合创始人斯坦利·唐（Stanley Tang）：“斯坦利，你拿去试试吧！” --- ### [01:06:33](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=3993s) 弗里德伯格的超级碗：亲历联邦最高法院孟山都除草剂世纪听证 闹腾的插曲过后，杰森将话题拉回到弗里德伯格近期的重大私人行程：“如果说我看纽约尼克斯季后赛坐在场边是我的超级碗，那弗里德伯格最近也迎来了他的人生超级碗！”原来，弗里德伯格亲自前往美国联邦最高法院（Supreme Court），在现场内场旁听了关于拜耳/孟山都（Monsanto）草甘膦化学诉讼的口头辩论。杰森向萨克斯询问是否曾现场旁听过最高法院审讯，萨克斯表示自己也从没去过。 弗里德伯格神情敬畏地回顾了那次现场经历，称那是他人生中最不可思议的震撼体验之一。最高法院大楼外围聚集了规模庞大的抗议人群，他们一行人必须穿过法警办公室（Marshal's office）的特殊安检通道才能进入。迈入最高法院大楼的瞬间，那种神圣庄严感扑面而来：整座建筑全部由典雅的大理石筑成，内部执行极其严苛的肃静纪律，任何人只要发出一丝声音都会立刻被法警制止，整座殿堂如同一座被神圣法度笼罩的超级静音图书馆，所有在场者都对这一宪法权威之地怀有无与伦比的敬畏。 ### [01:07:30](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=4050s) 走进最高法院：神圣肃穆的法庭规则与座席格局 戴维·弗里德伯格（David Friedberg）绘声绘色地回忆起自己亲临美国最高法院（Supreme Court）旁听辩论的震撼经历。法庭内部弥漫着一种令人敬畏的神圣感与绝对尊重，在那里毫无政治操弄或口舌之争的容身之地，甚至可以说“没有任何言论自由”。一旦踏入法庭，一切规则皆由九位大法官裁定：何时发言、如何举止、何种规矩，除非被大法官点名发问，否则任何人绝对不得擅自开腔。所有随身物品必须锁在楼下的储物柜中，步上楼梯步入审判大厅后，眼前的景象令人叹为观止。大法官席位的正上方是一幅恢弘壮丽的大理石雕带浮雕（marble frieze），刻画着摩西（Moses）等人类历史上最具代表性的伟大法治先驱，而在浮雕正下方端坐的便是九位联邦最高法院大法官。 面对杰森·卡拉卡尼斯（Jason Calacanis）关于座席安排的提问——首席大法官是否居中？保守派和自由派法官是否按党派左右分坐？——弗里德伯格与查马斯·帕里哈皮蒂亚（Chamath Palihapitiya）澄清了法庭的礼宾布局。首席大法官约翰·罗伯茨（John Roberts）位居正中央，其余八位大法官绝非按意识形态阵营分坐，而是依照在任资历（seniority/appointments）从中心向两侧交错排开。在辩论现场，罗伯茨偶尔会在全场静默时逐一点名询问：“某某法官，您有提问吗？”但在绝大多数时间里，大法官们都会在捕捉到论点缝隙的瞬间随时切入、激烈交锋。 弗里德伯格由衷赞叹，观看最高法院的法庭辩论“宛如现场看勒布朗·詹姆斯（LeBron James）打篮球”，双方出庭律师的顶级水准令人叹为观止，思维运转与临场质询的紧密程度让人全神贯注到“庭审结束时感觉整个人精力被彻底抽空”。查马斯随即打趣问他中途有没有偷偷吃一颗玫瑰火花含片（Rose Sparks），弗里德伯格笑称根本没有，因为全场心智完全被那种高强度的纯粹法理博弈深深吸附。 ```mermaid flowchart TD accTitle: 美国最高法院庭审空间与核心结构 accDescr: 展示最高法院法庭的视觉空间布局与九位大法官依资历排布的座位结构。 subgraph Visual["审判大厅视觉环境"] Frieze["历史先驱大理石浮雕 (含摩西等法治先驱)"] Bench["大法官法庭长席 (严格遵循法庭规矩，禁止随意发言)"] Frieze --> Bench end subgraph Seating["大法官座席格局 (按资历由中心向两侧排布)"] Roberts["首席大法官 约翰·罗伯茨 (居中端坐)"] SeniorLeft["资深大法官 (交错分布于左侧)"] SeniorRight["资深大法官 (交错分布于右侧)"] JuniorOuter["年轻/新晋大法官 (向外延伸就坐)"] Roberts --- SeniorLeft Roberts --- SeniorRight SeniorLeft --- JuniorOuter SeniorRight --- JuniorOuter end Bench --> Seating ``` --- ### [01:09:06](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=4146s) 纯粹的法律之辩：事实已定，只论宪法解释与联邦预先豁免 弗里德伯格解释了最高法院审判机制的核心逻辑：在最高法院层面，各方绝不再就案件本身的事实细节进行纠缠打转。下级初审法院负责查明案件事实（questions of fact），而最高法院唯一裁决的是纯粹的法律解释（questions of law）——涉案裁决是否合宪？某项联邦法案究竟应作何解释？全国性的法治框架应确立怎样的判断标准？所有法官手中都早已拥有涵盖全案起承转合的完整案卷简报（briefing docket）。 最高法院的排期极其紧凑严格，具有严密的流程机制： - **审理周期与频次**：法院仅在每年10月至次年4月开庭听审，通常选在最后两周的周一、周二、周三进行，全美海量诉讼中仅有极少数具有宪政与法规指导意义的标杆案件获准受理。 - **倒计时限制（Shot Clock）**：每天仅安排听审两起案件，每案严格限制在1小时内完成。双方律师各分得30分钟的口头辩论（oral arguments）时间，期间由九位大法官轮番进行高密度质询。 本案的核心争议聚焦于拜耳（Bayer）旗下的孟山都（Monsanto）除草剂农达（Roundup）。联邦层面的主管监管机构美国国家环境保护局（EPA）依据《联邦杀虫剂、杀菌剂和杀鼠剂法》（FIFRA）行使法定监管权，负责审批化学农药标签。EPA在科学评估后明确裁定农达不致癌，并在其核准下达的法定标签中未加入致癌警示。然而在州法层面，全美各地的诉讼代理律师依据加利福尼亚州等地方的“未尽警告义务法”（failure to warn law），控告孟山都明知潜在风险却隐瞒消费者。迄今为止，拜耳已为此类诉讼累计赔付超过100亿美元，资产负债表上额外计提了100亿美元的诉讼准备金，法院系统中仍有高达90,000起未决官司悬而未决。 去年，白宫联邦诉讼总长（Solicitor General）出面正式敦请最高法院接下此案。当联邦诉讼总长亲自介入申请调卷令（certiorari）时，最高法院接案的概率会成倍激增。拜耳与联邦政府的核心诉求在于确立“联邦预先豁免权”（Federal Preemption）——即联邦法律与EPA法定标签具有最高权威，应当压倒并阻断各州法项下自相矛盾的警告责任指控。 --- ### [01:12:24](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=4344s) 巅峰交锋：推翻雪佛龙原则的后座力与体制深层连锁反应 原本庭审之前，外界普遍预估保守派占优的最高法院会以6比3的悬殊比例裁定拜耳胜诉，认可联邦法规的排他性。然而，代表原告消费者的对方法律代理人一登场，便展现出令人叹服的辩论才华，其论证角度刁钻犀利，同样宛如“勒布朗·詹姆斯掌控全场”。 原告律师直接亮出了最高法院自身在不久前创下的重磅先例：推翻“雪佛龙尊重原则”（Chevron deference doctrine）。根据废除雪佛龙原则后的新法理，法院不再机械盲从或无条件遵从联邦行政监管机构对法条的裁量与扩权解释，而是必须由司法裁判者对法律条文本身进行直接解读。原告律师据此反唇相讥：既然联邦监管机构的释法垄断权已被打破，各州司法机构凭什么不能独立审读法律？凭什么必须无条件依附于EPA的行政标签？ 此言一出，全场格局顿时陷入势均力敌的胶着状态，胜负预期瞬间从原先的6比3倒向了充满悬念的“五五开”乃至“5比4”的微弱险胜边缘。辩论过程中，各方展开了激烈的交锋拉锯： - **白宫立场介入**：联邦诉讼总长在辩论中途起立陈述白宫法律立场并接受大法官质询，随后重新归座，留由控辩双方律师在此后半小时的问答中寸土必争。 - **大法官凯坦吉·布朗·杰克逊（Ketanji Brown Jackson）的连环追问**：杰克逊大法官尖锐地质问：“如果农药生产企业在EPA批准标签之后，又新发现了致癌科学证据，难道不应当立即更新产品警示吗？” - **拜耳律师的法理辩护**：拜耳方律师回应称，厂商根本无权擅自修改或增减EPA核准的法定标签，必须严格按照指定文本印制；同时法律规定，若厂商发现新的致癌事实却蓄意隐匿不报，将直接触发联邦刑事责任。 - **监管不作为与各州保护权**：杰克逊大法官紧接着追击：“那倘若EPA在接到报告后怠于行动、无所作为呢？难道各州政府没有天然的正当权利去立法保护本州居民的生命健康吗？” 弗里德伯格强调，该案之所以牵动全美神经，是因为其外溢效应极其深远。如果最高法院最终认定各州有权自行解释联邦法律并直接架空联邦监管机构，将打开一个极其危险的“潘多拉魔盒”——各州从此将能任意无视EPA、美国食品药品监督管理局（FDA）或美国农业部（USDA）等联邦部门建立的全国统一准则，导致全美商业与公共治理陷入四分五裂的巨大混乱。在口头辩论尘埃落定后，戴维·萨克斯（David Sacks）补充说明，九位大法官随后将转入高度私密的闭门评议室（private conference）进行深入研讨、各自起草意见书，并最终对外公布终局判决。 ### [01:15:14](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=4514s) 探秘最高法院闭门会议与法庭庄严仪式 嘉宾们继续深入探讨美国最高法院的内部运作机制。在最高法院内部，一旦大法官们明确了哪一方占据多数，首席大法官就会指派由谁来撰写多数意见书。在他们评议讨论的会议室内，任何外人绝对不得入内。为了防止任何信息泄露，会议室甚至设有一道“双重门气闸系统”（airlock）：任何文件或物品若需要递进或送出，必须先敲门进入前厅，再经过隔离门转交。查马斯（Chamath Palihapitiya）提到，得克萨斯州参议员特德·克鲁兹（Ted Cruz）曾受邀参加他们的私人扑克牌局，克鲁兹早年曾担任前首席大法官威廉·伦奎斯特（William Rehnquist）的法官助理。克鲁兹对最高法院的历史典故了如指掌，听他详述最高法院和伦奎斯特的故事，会让人觉得这是世界上构建得最不可思议、最令人敬畏的国家机构。 这种庄严肃穆甚至超越了白宫、国会大厦以及各大联邦政府机构。戴维·弗里德伯格（David Friedberg）形容置身其中仿佛来到了英国古老的殿堂，所有人的行为举止都遵循着一套极其严苛的礼仪。法庭的安保规程也非同寻常，庭内法警每隔20分钟就会以极其协调整齐的动作换岗轮替，着装规范也与任何普通法庭截然不同。 关于旁听资格，弗里德伯格估算现场大约只有150个座位。他感觉现场坐着的几乎全都是大法官助理或法院雇员邀请的特定宾客，普通公众极难直接进入。弗里德伯格透露，他们这次能够亲临旁听，是通过首席大法官约翰·罗伯茨（John Roberts）的关系拿到了特别通行证。贾森·卡拉卡尼斯（Jason Calacanis）则提到，在埃隆·马斯克起诉 OpenAI 的庭审中，公众需要连夜排队，法官还会为媒体分配大约30张记者票；最高法院其实也保留了一定比例的公众排队通道。但弗里德伯格强调，当天法庭里完全看不到随意的游客，所有人都身着最体面的正装，整座法庭流淌着一种令人屏息凝神的厚重感。 ```mermaid flowchart TD accTitle: 最高法院评议机制与安保流程 accDescr: 展示最高法院从闭门评议、双重门安保到意见书指派的严密流程。 subgraph Deliberation["大法官闭门评议"] A["九位大法官闭门研讨"] --> B{"确定多数派立场"} B --> C["首席大法官指派<br/>多数意见书起草人"] end subgraph Security["双重门隔离机制"] D["外来文件/物品送达"] --> E["敲门进入第一道前厅"] E --> F["穿过第二道隔离气闸门"] F --> A end subgraph Chamber["法庭现场仪式"] G["首席大法官特批通行证"] --> H["150席现场旁听席"] I["法警每20分钟严整轮替换岗"] --> H end ``` ### [01:17:07](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=4627s) “在它被玩坏前好好珍惜”：法庭扩容之争与制度衰败的隐忧 戴维·萨克斯（David Sacks）对此发出感慨：“趁现在还能看，就好好珍惜吧。”他直言，最高法院堪称当今美国极少数仍在高效、规范运作的核心体制之一，但这种体面可能维持不了太久。萨克斯半开玩笑又带着讽刺地预测，未来某一天最高法院可能会被强行扩容到13人甚至21人这种荒唐的规模，届时人们甚至会像看球赛一样，穿着印有各位大法官名字的球衣，分坐在各自支持者的看台区，球场边还有小贩叫卖爆米花，把最高法彻底变成电影《蠢蛋进化论》（Idiocracy）里的荒诞闹剧。 弗里德伯格指出，公众对最高法院的好感度往往完全取决于近期判决是否符合自己的政治立场，人们评价最高法是出于近期的结果偏好，而非判决本身的严谨论证和法理推演。他现场查阅了民调数据： - 在2024年的盖洛普（Gallup）民意调查中，最高法院的支持率曾一度滑落至35%的历史低点。 - 目前该数据已回升至44%至50%的区间。 - 考虑到最高法所卷入的政治敏感议题，相较于支持率长期低迷的国会或个别政客，最高法院的民众认可度其实相对偏高。 亲身经历旁听后，弗里德伯格由衷感叹，看到大法官们在庭上睿智的言行交锋与严谨的审判流程，令人对这个国家的立国根基肃然起敬。然而萨克斯随即指出冷酷的政治现实——民主党资深策略师詹姆斯·卡维尔（James Carville）在近期的访谈中直言不讳地表示，一旦民主党重新夺回权力，他们就会毫不犹豫地实施“填塞法院”（Court Packing）计划，将法官席位从9人扩增至13人，并推动华盛顿特区和波多黎各建州，以彻底巩固权力格局。 - **卡维尔的政治战略**：主张民主党一旦掌权，无需顾忌传统，直接将最高法院扩容至13名大法官，并通过增加新州永久改变参议院版图。 - **商业化恶搞插曲**：卡拉卡尼斯查阅后发现最高法院确实有网络抽签预约系统，随即开玩笑提议应该联系商务部长提名人霍华德·卢特尼克（Howard Lutnick），把这些门票改造成商业拍卖，以每张2万美元的价格售卖“内场场边席”（courtside seats）为国家创收。萨克斯立即打趣回击：“贾森，你这种人正是最高法院竭尽全力想要防范的对象。” ### [01:19:43](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=4783s) 节目终场告别与恶搞花絮 在欢笑声中，贾森·卡拉卡尼斯主持了整期播客的收尾。他戏称自己是“全网最伟大的主持人”，并顺着前文的法庭梗，自封为“All-In 播客的首席大法官”。随后，节目在经典的片头片尾混音与粉丝剪辑的经典语录（如“让赢家继续驰骋”、“雨人弗里德伯格”、“时装杂货商”等名场面金句）中圆满落幕。 ## Notable quotes > [00:00](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=0s) Gay! — Miss Thing 主播 > [00:00](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=0s) 太精彩了，我今天才知道我可能是弯的。 — Chamath Palihapitiya > [04:49](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=289s) right but for the wrong reason — David Sacks > [21:20](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=1280s) It doesn't create the vulnerabilities. It just discovers them. — David Sacks > [25:03](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=1503s) penetrated and can essentially manipulate every model — Chamath Palihapitiya > [27:49](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=1669s) Either you're cheap or poor. Which one is it? — David Sacks > [31:03](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=1863s) 如果我们纵容掠夺慈善机构资产的行为合法化，整个美国的慈善捐赠基石都将被彻底摧毁，这正是我所担忧的核心所在。 — Elon Musk > [31:03](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=1863s) 结论：我们确实想要成立 B-Corp（共益企业）。真正的答案是我们希望埃隆出局。如果三个月后我们真的转做 B-Corp，那之前说的话就是谎言。我看不到在不经历一场恶斗的情况下将这里变成营利公司的可能。我只是在想办公室里的情形……大家会正确地指出我们对他（马斯克）并不诚实。到头来，核心还是想要营利架构，只是不带他玩罢了。 — Greg Brockman > [31:03](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=1863s) 你居然在犯罪共谋时做会议笔记？你到底在想什么？ — Jason Calacanis > [33:30](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=2010s) 过度反刍和沉溺内耗是通往痛苦的不归路。根本没人在意你的敏感感受，把这些矫情写在纸上只会让你更加悲惨。 — Chamath Palihapitiya > [29:10](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=1750s) 大约花了 8000 美元的航油费，以及大概 12000 美元的 Pappy Van Winkle 威士忌。 — David Sacks > [58:23](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=3503s) 快给我讲讲 Retatrutide，我太想用它了！你得去跟娜塔莉（Nat）说说，告诉她我用这个完全没问题。 — Jason Calacanis > [01:17:07](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=4627s) 趁现在还能看，就好好珍惜吧。 — David Sacks > [01:17:07](https://www.youtube.com/watch?v=fpC4sbawSzQ&t=4627s) 贾森，你这种人正是最高法院竭尽全力想要防范的对象。 — David Sacks ## Mentioned OpenAI, ChatGPT, Codex, Anthropic, Claude, Claude Opus 4.6, Claude Opus 4.7, Mythos, Sam Altman, Elon Musk, Jason Calacanis, David Sacks, Chamath Palihapitiya, David Friedberg, Sarah Friar, Greg Brockman, Dario Amodei, Sergey Brin, George Kurtz, Nikesh Arora, Yvonne Gonzalez Rogers, John Roberts, Ketanji Brown Jackson, William Rehnquist, Ted Cruz, James Carville, Howard Lutnick, Stanley Tang, Aaron Levie, Jack Dorsey, Balaji Srinivasan, Brad Gerstner, Google, Gemini, Google Cloud Platform (GCP), Vertex AI, Google Flights, Microsoft, Azure, Amazon, AWS, Bedrock, Meta, Oracle, xAI, Grok, Cursor, CrowdStrike, Palo Alto Networks, DeepSeek-V4, Eli Lilly, Tirzepatide, Mounjaro, Zepbound, Retatrutide, Ro, Wegovy, Pocket OS, Railway, Bayer, Monsanto, Roundup, EPA, FDA, USDA, The Wall Street Journal, The New Yorker, Polymarket, Miss Thing, The Wire, Silicon Valley ## Video description (0:00) Bestie intros (3:05) OpenAI misses targets, Codex gains on Claude (20:02) AI cybersecurity: a market that's about to explode (31:03) Elon vs Sam Altman lawsuit (41:00) Big tech smashes earnings, Capex explosion (52:44) Vibecoding nightmare: AI deleted someone's codebase (58:33) Retatrutide craze: peptides go mainstream (1:06:34) Friedberg's Supreme Court experience Apply for Summit 2026: https://allin.com/events 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.instagram.com/missthingthepod https://www.wsj.com/tech/ai/openai-misses-key-revenue-user-targets-in-high-stakes-sprint-toward-ipo-94a95273 https://polymarket.com/event/ipos-before-2027 https://x.com/aakashgupta/status/2049723185617412550 https://arxiv.org/pdf/1803.03635 https://x.com/AISecurityInst/status/2049868227740565890 https://x.com/ns123abc/status/2049527702076449244 https://www.reuters.com/legal/litigation/openai-trial-pitting-elon-musk-against-sam-altman-kicks-off-2026-04-28 https://x.com/ns123abc/status/2049527702076449244 https://www.google.com/finance/quote/CSCO:NASDAQ https://x.com/chamath/status/2049864100143104420 https://x.com/zerohedge/status/2049895327566561683 https://x.com/lifeof_jer/status/2048103471019434248 https://x.com/levie/status/2049163935182733396 https://www.foxnews.com/media/carville-tells-dems-quietly-prepare-power-grab-dc-puerto-rico-statehood-supreme-court-packing #allin #tech #news [Source](https://www.youtube.com/watch?v=fpC4sbawSzQ)

[0:00] Miss Thing 主播: Gay! [0:00] Chamath Palihapitiya: 太精彩了，我今天才知道我可能是弯的。 [4:49] David Sacks: right but for the wrong reason [21:20] David Sacks: It doesn't create the vulnerabilities. It just discovers them. [25:03] Chamath Palihapitiya: penetrated and can essentially manipulate every model [27:49] David Sacks: Either you're cheap or poor. Which one is it? [31:03] Elon Musk: 如果我们纵容掠夺慈善机构资产的行为合法化，整个美国的慈善捐赠基石都将被彻底摧毁，这正是我所担忧的核心所在。 [31:03] Greg Brockman: 结论：我们确实想要成立 B-Corp（共益企业）。真正的答案是我们希望埃隆出局。如果三个月后我们真的转做 B-Corp，那之前说的话就是谎言。我看不到在不经历一场恶斗的情况下将这里变成营利公司的可能。我只是在想办公室里的情形……大家会正确地指出我们对他（马斯克）并不诚实。到头来，核心还是想要营利架构，只是不带他玩罢了。 [31:03] Jason Calacanis: 你居然在犯罪共谋时做会议笔记？你到底在想什么？ [33:30] Chamath Palihapitiya: 过度反刍和沉溺内耗是通往痛苦的不归路。根本没人在意你的敏感感受，把这些矫情写在纸上只会让你更加悲惨。 [29:10] David Sacks: 大约花了 8000 美元的航油费，以及大概 12000 美元的 Pappy Van Winkle 威士忌。 [58:23] Jason Calacanis: 快给我讲讲 Retatrutide，我太想用它了！你得去跟娜塔莉（Nat）说说，告诉她我用这个完全没问题。 [77:07] David Sacks: 趁现在还能看，就好好珍惜吧。 [77:07] David Sacks: 贾森，你这种人正是最高法院竭尽全力想要防范的对象。

[0:00] Chamath Palihapitiya: I think I'm gay... I never knew! [20:02] If we can now use AI to find these bugs in advance, these vulnerabilities and patch them, then you actually harden our infrastructure and you harden our security. [31:03] If we make it okay to loot a charity, the entire foundation of charitable giving in America will be destroyed. That's my concern. [32:19] Jason Calacanis: Is you taking notes on a criminal conspiracy? What the f*** is you thinking, man? [41:00] David Sacks: The bull thesis for AI just got validated in a single afternoon. [64:53] David Friedberg: Basically overtime. [64:53] David Friedberg: What the f*** was that? [66:34] David Friedberg: There is no politics here. There is no freedom of speech. This is the court. When you come into this court, the justices tell you how you will speak, how you will behave, what you will do... [75:07] David Sacks: one of the last highly functional institutions in the United States [78:16] David Friedberg: I just felt so assured of like the institution when I visited and saw these guys interact... what an amazing country. [78:16] David Friedberg: Jason, you're exactly what they're trying to protect against.

[0:00] Jason, do you want to tell us about your [0:02] new favorite podcast? Oh, it's so good. [0:06] My feed is now because, you know, since [0:10] cancel culture ended, Sachs, everybody [0:12] uses the R word and the f word right [0:14] now. My entire feed on Instagram is [0:17] either gay or down syndrome or bulldogs. [0:20] It's one of those three. And then I [0:22] stumbled upon the Miss Thing pod, [0:25] Miss Thing. And they do a bit called gay [0:28] name, straight name. Here's gay name or [0:30] straight name for David. This good news [0:33] and bad news freeird. Here we go. [0:35] >> Gay name or straight name? [0:38] >> David. [0:40] >> David to me is straight. [0:42] >> Okay. But he has my perfect body. It can [0:46] be confusing because I'm kind of like, [0:48] are you gay? And it's like, no, I just [0:50] want to be you, David. [0:51] >> Totally. Well, it's so like the [0:54] Michelangelo's David the male ideal. [0:56] It's like incredible body kind of small. [0:58] Sorry. [0:59] >> Yeah. [0:59] >> Oh, [1:00] >> it's a little rough. [laughter] [1:02] >> What? [1:03] >> What are you watching there, Jal? [1:04] >> They basically nailed these two, but [1:06] okay, keep going. [1:07] >> I don't think Chimoth is on their short [1:09] list, but I know Jason will come up at [1:10] some point. [1:11] >> Gay name or straight. [1:13] >> Maybe this is it. [1:14] >> Chimoth [1:16] on the count of three. Yeah. [1:18] >> Three, two, one. gay. [laughter] [1:22] I'm seeing like Italian sweater, like [1:24] really kind of like a loud [laughter] [1:27] vibrant sweater. He like wears it to [1:29] like poker night with his like his boys [1:32] and like not I'm not talking like [1:34] straight poker. I'm talking like gay [1:36] poker nights like at the bar. [1:38] >> Yeah. Always talking about wine. [1:40] >> Talks about wines. [1:42] >> Always sort of like [1:44] Yeah, exactly. [1:45] >> Yep. Also, it's so like the guy at the [1:48] gym taking off his shirt, taking [1:51] selfies. [1:52] >> Yeah. And everyone else is kind of like, [1:54] "Excuse me, Chimoth. I'd like to use the [1:56] mirror. [laughter] [1:57] I'd like to see myself. [1:58] >> See you at the next day." Poker night. [2:00] >> Totally. [2:01] >> You bring the wine. [laughter] [2:03] >> We'll bring the sweater. [2:04] >> Yeah, [2:04] >> there it is. Wow. They did do. [2:06] >> That is fantastic. That is fantastic. A [2:09] shout out to my guys at the Miss Ding [2:11] podcast. [2:12] >> Wow, that was awesome. I think I'm gay. [2:15] >> [laughter] [2:16] >> I never knew. [music] [2:19] >> Let your winners ride. [2:26] >> And it said we open [music] sourced it [2:27] to the fans and they've just gone crazy [2:29] with it. [2:35] >> What did you do like cameo? Did you pay [2:37] them to do that? [2:37] >> Did it for me as a favor. So they did it [2:39] for [2:40] >> That's awesome. Well, thanks to those [2:41] guys to the miss. I've seen those guys [2:44] before in clips. I find them very funny. [2:46] [laughter] [2:46] >> It's so great. Shout out to my guys. [2:49] >> That was awesome. [2:50] >> All right, everybody. Seriously, welcome [2:52] back to the number one podcast in the [2:54] world. It's the All-In podcast with me [2:56] again, David Freebergia, David Saxs, and [3:00] of course, I'm Jason Calcanis. You can [3:01] call me Jay Cal if you're here for the [3:03] first time. Topic one, open AI. They [3:07] missed their targets for chat GPT [3:09] Freedberg both on users and revenue. [3:13] Let's talk about it. The Wall Street [3:14] Journal says in a uh a breaking [3:18] investigative report on Tuesday that [3:21] OpenAI expected to hit 1 billion wows [3:24] weekly active users before the end of [3:26] 2025. They missed that and they still [3:28] haven't hit the milestone 4 months into [3:30] 2026. Also, Chamath, they missed their [3:33] 2025 revenue target for Chad GPT. Exact [3:37] number wasn't specified, but as we've [3:39] talked about here, they're at a 2030 [3:41] billion run rate. There's a little bit [3:43] of accounting nuance that is yet to be [3:46] worked out in the industry. Two reasons [3:48] why this matters. Sachs, OpenAI has $600 [3:52] billion in spending commitments for [3:54] compute. Just to put that in [3:56] perspective, that's about what they're [3:58] trading for on secondary markets. In [4:00] other words, the entire value of the [4:01] Open AI enterprise equals their spend [4:04] commitments in the coming year. CFO [4:05] Sarah Frier, who is coming to liquidity, [4:08] is reportedly worried, hey, that revenue [4:10] isn't growing fast enough to keep up [4:13] with expense and OpenAI wants to IPO [4:16] later this year. This has put Frier and [4:19] Oughtman in conflict or uh maybe there's [4:22] some natural tension there. Frier [4:24] doesn't think OpenAI is ready for public [4:26] reporting standards. According to the [4:28] Wall Street Journal, Altman obviously [4:30] wants to move faster, so they released a [4:33] joint statement. this is ridiculous yada [4:35] yada yada. Let's go to you Saxs. What do [4:38] you think's going on here? Are these [4:39] major headwinds or is this just managing [4:41] expectations as the leader of the pack [4:44] in the most important race of our [4:46] lifetimes, the race towards super [4:48] intelligence? [4:49] >> Well, I actually have a little bit of a [4:51] contrarian take on this. I know that [4:53] OpenAI had a really bad week. Like you [4:55] said, they had that Wall Street Journal [4:57] article which said that they missed [4:58] their numbers. They missed their 1 [5:00] billion user growth target. They missed [5:03] their revenue numbers. That's called [5:04] into question whether they can afford [5:06] the data center commitments that they've [5:08] made. And then in addition to that, [5:10] they've also had the lawsuit with with [5:12] Elon happening this week. So in the [5:14] press, it ended up being I think a [5:15] pretty bad week for them. But I have a [5:16] contrarian take on this, which is I [5:18] think that over the past week or two, if [5:20] you look at kind of what's happening at [5:22] the product level, it's been a pretty [5:23] good couple of weeks for them. They [5:25] released chat GBT 5.5 and the reviews [5:29] from, you know, people I talked to in [5:30] Silicon Valley have been really strong. [5:32] You talk to developers, coders, they're [5:34] very happy with it. At the same time, [5:37] Opus 4.7, which is the latest anthropic [5:40] release, appears to be a bust. People [5:41] are complaining about it. They're in a [5:43] lot of cases are rolling back to 4.6. [5:46] They're saying that Opus 4.7 is [5:48] rationing compute. It's reducing [5:50] thinking time, not as good. there were [5:52] some bugs and clawed. So if you just [5:55] compare chat GPT 5.5 to Opus 4.7, it [6:00] does appear that OpenAI has had a better [6:02] couple of weeks on a product level. And [6:04] I think there's reason to believe that [6:07] the product improvements will continue. [6:09] GPT 5.5 is based on a new base model [6:12] called Spud, which is the first base [6:15] model upgrade they've done in I don't [6:16] know over a year. and having a new base [6:18] model will pave the way for future [6:20] improvements as well. So I think OpenAI [6:23] is feeling pretty optimistic about their [6:25] product right now and I think you're [6:26] starting to see on X some of the [6:29] developer mojo is shifting. I'm seeing a [6:31] lot of people saying that they are [6:33] shifting their their coding usage from [6:36] Opus to GPT 5.5. [6:39] So I think that SAM may end up being [6:43] right but for the wrong reason. And what [6:46] I mean by that is that when he made [6:49] these big compute commitments, it was [6:52] based on those estimates of hitting the [6:54] billion users on the consumer side and [6:56] hitting those revenue targets. The [6:58] consumer business ended up being weak. [7:00] So they missed those targets. But in the [7:02] meantime, coding has become the [7:04] allimportant sector of AI. And because [7:09] they made all these compute commitments [7:11] and they built out these data centers, [7:13] they have more compute than anthropic [7:15] right now. Anthropic is token [7:17] constrained. It's reducing their ability [7:19] to serve mythos, for example. It's [7:21] causing them to engage in compute gating [7:23] with Opus 4.7. And I understand why [7:26] Daario made that decision. I'm not [7:27] saying I mean it was a prudent business [7:29] decision. I'm not criticizing him for [7:30] it, but I think again I think Sam may [7:33] end up being right here for the wrong [7:35] reason, which is he missed on consumer, [7:37] but enterprise is going gang busters and [7:40] is giving him the ability now, I think, [7:42] to catch up on code. your poly market [7:44] >> which is the all important market right [7:46] now [7:46] >> of course and we talked about gro and [7:48] cursor teaming up last week Elon and the [7:50] team over there poly market showing now [7:52] a 32% chance that openai goes public by [7:55] the end of 2026 this is down from 60% in [7:58] December and Shimath you gave a bit of a [8:01] warning hey there's only so many dollars [8:04] to go around SpaceX IPO is obviously [8:06] getting out first and now if openai [8:10] doesn't go out this year and anthropic [8:13] does the sets up an interesting dynamic. [8:14] What are your thoughts here generally [8:16] speaking about the massive commitment [8:19] that OpenAI has made? Are they going to [8:21] run off the cliff or will it wind up [8:24] being brilliant? Uh even if it wasn't [8:26] strategically for the exact reasons, [8:29] >> I think they're going to be fine. I [8:30] think this is a multi-t trillion dollar [8:32] company. I think Anthropic is a multi- [8:34] trillion dollar company. I think the [8:36] thing that's happening right now is uh a [8:39] complete misunderstanding [8:41] of what's actually happening [8:44] inside of the world of AI. And there is [8:46] one very specific choke point [8:49] that is constraining everything which is [8:50] access to the power that's necessary to [8:53] drive these tokens. To the extent that [8:55] open AI missed, I think what that is is [8:58] an insight to not enough compute [9:00] capacity today. And that problem is only [9:03] getting worse. You've already seen that [9:05] with Anthropic as well where they just [9:08] found a way to economically induce [9:11] Amazon to give them enough capacity so [9:14] that you don't have to route through [9:16] bedrock to get to the anthropic models. [9:19] You're also seeing them do [9:20] differentiated deals now with economic [9:23] participation on top of what they [9:25] already had from folks like Google to [9:26] give them more capacity. What is my [9:28] point? Everything in this market is [9:31] power constrained. The reason that these [9:34] folks may miss a number or a forecast [9:36] have nothing to do with demand. It is [9:39] entirely 100% due to the supply of the [9:43] power necessary to generate the output [9:44] token. There is a really interesting [9:46] thing that was just announced today that [9:49] will make this problem even worse, which [9:52] is what you're starting to see now is [9:55] backlogs build up of not just the access [9:59] to the power, but then the componentry [10:01] that's actually necessary. Not just [10:03] resips and not just NAT gas turbines, [10:05] but now you're talking about [10:06] transformers and all the actual tactical [10:08] grid infrastructure. [10:10] Why is this important? If you look at [10:13] the actual amount of gigawatts that are [10:16] under construction, we have a huge [10:18] mismatch now, [10:20] people have announced all these [10:22] projects, Jason, [10:24] but less than half of it is actually [10:26] being built. Less than half. Most of it [10:29] is stuck in red tape. [10:31] Most of that is because there are these [10:33] supply chain delays. So there's no [10:36] credible strategy to turn any of this [10:38] stuff on. Who will this hurt? It will [10:41] hurt Anthropic and OpenAI the most. Who [10:44] will this benefit? It will benefit the [10:47] hyperscalers, specifically Oracle, [10:49] Amazon, Meta, Microsoft, and Google. And [10:52] now what you're going to see is a [10:55] negotiation and a trade back and forth. [10:57] How much equity do I have to give up? [11:00] How much control do I have to give up to [11:01] get access to the compute versus how [11:04] badly will I miss my growth forecasts if [11:07] I don't? And now what that means is, and [11:09] we spoke about this last week, that's a [11:10] huge lane for Grock to just run through [11:13] and SpaceX to run through cuz they have [11:16] a ton of excess capacity. And so I think [11:19] the cursor deal was the appetizer. But [11:22] if I were Elon now, I'd be running all [11:24] over this market because if the models [11:27] catch up in quality, I think he could [11:29] also do something really crazy with [11:31] anthropic or open AI right now. Maybe [11:33] not open AI because of the [11:35] >> we'll get into the lawsuit in a bit [11:36] >> the baggage. [11:37] >> Yeah. [11:38] >> But man, he and Dario should do a deal [11:40] tomorrow. [11:41] >> So you're framing, hey, the the limited [11:44] resource here is compute. The demand is [11:47] off the charts. [11:48] >> No, the limiting resource is power. [11:50] power which then powers compute which [11:52] then provides tokens which then services [11:55] the massive uh developer and co-work and [11:59] all these other projects that consumers [12:01] and enterprises can't get enough of. Got [12:03] it. [12:03] >> And Jason, the other factor that [12:05] complicates that for anthropic and open [12:07] AI is all the stuff that's sort of [12:09] sitting around thumb twiddling. 40% of [12:13] that is going to get cancelled because [12:14] they've done such a poor job of creating [12:17] a good positive halo around AI that 40% [12:20] of all the announced projects get [12:22] cancelled because 40% of all projects in [12:25] the last four years have been cancelled. [12:28] Yeah. And there's there there are some [12:30] bad feelings about data centers, AI, [12:33] jobs, etc. And that's causing some [12:35] headwind. People are you literally doing [12:39] violent things in society and blaming [12:41] data centers and AI for it. I don't want [12:43] to give it too much air time. Freeberg, [12:45] what's your take on the chessboard we're [12:47] looking at here? Either through compute, [12:50] energy or through going public on a [12:53] business level, you know, the strategic [12:56] nature of capital, compute and energy [12:59] now playing a role in this massive [13:01] amount of demand. still a ball in the [13:04] air kind of game. BCG had this theory, I [13:09] think I talked about this once before, [13:10] called the rule of three where they've [13:12] shown time and again that any stable, [13:14] mature, competitive market evolves to a [13:18] 4:21 [13:20] ratio of market share for basically 90% [13:22] of the market. So there's a market [13:23] leader that has four times the market [13:26] share of the second place, that's two [13:28] times the market share of the third [13:29] place. This is the case in pretty much [13:30] every mature kind of competitive market. [13:33] So you can kind of think about AI [13:34] probably evolving into a consumer market [13:36] and an enterprise market. Open AAI, even [13:39] if they're not at a billion, they're [13:40] still at 900 million weekly users, which [13:43] is well ahead of whatever Claude is at. [13:46] I think Claude is like probably subund [13:48] million sacks, you may know. And then [13:49] Gemini is probably closer to them at 700 [13:52] to a billion somewhere in that range. [13:54] Probably pretty neck and neck with open [13:56] AI. So, you know, the consumer market [13:59] looks like it's trending towards a chat [14:02] GPT/Google [14:04] fight for first place and second place [14:06] and then probably anthropic in third [14:08] place and maybe Elon emerges and takes [14:10] off enabled by his compute capacity and [14:13] then the enterprise market is a little [14:14] bit of a different story and that's its [14:16] own market which is kind of anthropic or [14:18] probably Google in the lead actually if [14:20] you look at all the vertex use. Google [14:22] claims that 75% of GCP customers are [14:26] active users of Vertex. So there's [14:29] probably a pretty sizable [14:32] market share that Google's captured on [14:33] the enterprise side as well. This is [14:34] also probably why Google stock has [14:36] absolutely ripped over the last couple [14:38] of months is they're literally in first [14:39] place or fighting for first place in [14:42] enterprise and consumer. [14:44] But I still think that there's a lot of [14:46] opportunity to Chimoff's point about the [14:48] compute and energy capacity constraints [14:50] in improving how we actually scale and [14:53] deploy models in both the enterprise and [14:56] the consumer setting. And it is such [14:57] early days and I just want to highlight [14:59] this paper that came out from MIT from [15:02] these two scientists and these guys [15:05] published a paper on pruning techniques [15:07] and neural networks. This paper showed [15:09] that you could actually reduce the size [15:10] of these networks by 90%. [15:13] And get the same accuracy out by pruning [15:17] very large models down to smaller [15:19] models. And then you can make a [15:20] selection on which model to run for [15:22] inference. And by doing this, you can [15:24] actually reduce inference costs by 10x. [15:27] You can get 10x the output per energy [15:29] unit that goes into the data center with [15:31] no loss of accuracy. And so it's a [15:34] really interesting call it algorithmic [15:35] technique that can be applied to the [15:38] existing large models to actually make [15:40] them much lower energy use. So if you [15:42] think about it, you're firing up a very [15:44] large model to answer a very simple [15:46] question. You can actually prune away [15:48] that model. Now this is probably going [15:51] to be the case in AI applications as it [15:54] is in traditional Google search. There's [15:56] a long tale of searches, but there's a [15:58] few searches that account for a large [16:00] percentage of search volume. It's like [16:02] what is the weather? What are the movies [16:04] times? You know, what's the stock price? [16:06] Like there's a certain set of things [16:08] that make up the bulk of consumer [16:10] energy. And there's probably a certain [16:11] set of things that probably make up the [16:12] bulk of coding output as well. And so if [16:15] you can get that 80% of searches or chat [16:18] interfaces or coding requests reduced [16:22] down through pruning techniques to [16:23] smaller models and then you have a whole [16:25] set of smaller models that can be called [16:26] dynamically and you reduce inference [16:28] cost by 90%. you can make much more use, [16:31] call it 10 times the use on data center [16:34] and energy capacity than we can today. [16:36] So I would argue that we're still in the [16:37] very early days of getting efficiency in [16:40] terms of output and tokens and we're [16:42] just in the very kind of early stage of [16:43] that which also unlocks the opportunity [16:45] for guys like Elon to reinvent how this [16:47] is done and potentially compete pretty [16:49] aggressively. [16:50] >> There are two ways to win. You could [16:51] throw compute at it or you can do SLM's [16:54] small language models and V SLM's [16:57] verticaliz small language model. So if [17:00] you had a verticalized small language [17:01] model for the weather, let's say that [17:03] doesn't exist, but uh you can you can [17:05] use it as an example. They will have one [17:07] for travel as an example. When you hit [17:09] Google for flight information, it's [17:11] obviously going to route you to [17:13] something lighter and faster that uses [17:15] Google flights. And Google flights has [17:16] been incor incorporated into Gemini. [17:19] Gemini now is right behind 700 750 [17:23] million users [17:25] >> and it's exactly what we discussed I [17:27] don't know 18 months ago on this podcast [17:28] Freeberg that what if they put it at the [17:30] top [17:31] >> and what would that do to their search [17:33] revenue search revenue is surging and [17:37] they're also surging so they figured out [17:39] a way to balance those two competing [17:41] forces having search results that are AI [17:43] enabled and still getting people to [17:46] click on links they've done it [17:47] brilliantly apparently and the stock is [17:49] rewarding. [17:50] >> I'll just add one statement to what you [17:51] said, which is like you're using what I [17:53] would call a humanistic on [17:58] humans don't intuitively know what this [18:01] model is. It's not just a verticalized [18:02] model, but there are going to be models [18:05] that will be discovered through [18:07] automated pruning techniques that will [18:09] then be working in concert. So, lots of [18:12] small models that link together. And we [18:13] don't define each model by some human [18:15] heristic like this is a search travel [18:17] model. [18:18] >> This is a maps model. We don't we don't [18:20] know why these models work the way they [18:22] do when they get broken down. But I do [18:24] think that that's really where the [18:25] evolution is happening. So effectively a [18:27] model becomes a macro model. It's got [18:29] lots of smaller models underneath it [18:31] that can be dynamically called and that [18:33] allows you to have 10x the inference for [18:35] the same unit of energy. [18:36] >> Sax, [18:37] >> let me just build on your point about [18:38] Google. Jcal, I would say that if [18:40] there's a single reason why OpenAI did [18:44] not hit its user targets and its revenue [18:48] targets, certainly around consumer, [18:50] you'd have to say it's because Google [18:52] managed to take meaningful share, you [18:55] know, they were basically nowhere a year [18:58] or so ago. Sergey came out of [19:00] retirement, helped focus the company, [19:03] and like you said, they did a brilliant [19:05] job improving Gemini and putting it at [19:07] the top of search, incorporating it. [19:09] Now, that being said, again, I don't [19:11] think the news is all bad for OpenAI [19:13] because I do think that the 5.5 release [19:15] was great. We're hearing really good [19:17] things about Codeex. I do think that [19:19] Codeex is taking share in coding tokens [19:23] right now. And I just think we're in a [19:25] really interesting place where these [19:27] companies are constantly oneuping each [19:29] other. I mean, two weeks ago it looked [19:31] like Anthropic was going to be [19:32] completely dominant, right? I mean, [19:34] Anthropic was growing at 10x. Open AAI [19:36] was growing at 3x and it looked like [19:38] >> and then the servers started going down. [19:39] Did you see that this week? The server [19:41] going down. People were in my office [19:43] were complaining we can't get on claud. [19:45] >> Listen, competition brings out the best [19:47] in everyone. Anthropic forced open AI to [19:50] compete. Google's forced open AI to [19:52] compete in consumer. I just hope the [19:54] market stays competitive for as long as [19:55] possible. I do think that's what's best [19:57] for consumers, our economy, and for our [20:00] country overall. Let me just say one [20:03] other area where I think OpenAI had a [20:06] good week is in this red-hot area of [20:10] cyber. Obviously, Anthropic made a huge [20:13] splash with Mythos. It hasn't been [20:14] commercially released. Their compute [20:16] constraint, but as a proof of concept or [20:18] training model, it hit a new level of [20:20] capabilities with cyber. But now OpenAI [20:22] has released a new model called GPT 5.5 [20:25] cyber which has just been through a [20:26] bunch of tests and they've shown this [20:29] was testing done by the AI security [20:31] institute that GPT 5.5 is the second [20:34] model to complete one of their [20:36] multi-step cyber attack simulations end [20:38] to end. So it has the same level of [20:40] capability as Mythos and it does appear [20:43] to be commercially [20:46] ready. You know, they've got the compute [20:47] to serve it. So I do think that that's a [20:51] big accomplishment. I mean, look, we [20:53] knew that other cyber models were [20:55] coming. It wasn't just going to be [20:56] Mythos. In fact, within 6 months or so, [20:59] all the Frontier models are going to [21:00] have Mythos level cyber capability. But [21:03] it's impressive that OpenAI got this GPT [21:07] 5.5 cyber out. so quickly and I think [21:11] 5.5 might be the first cyber model that [21:14] cyber defenders actually get to use [21:17] because again I don't think they're as [21:18] compute constrained as anthropic is [21:20] >> and this is an incredible opportunity [21:22] you know for the crowd strikes and [21:24] PaloAlto networks of the world both of [21:26] which have been on the program they come [21:29] out and they start attacking this space [21:31] man you could really [21:33] see everything get tightened up and this [21:37] could be an incredible revenue stream [21:39] for everybody who's got whether it's [21:41] cursor claude or open eye or Gemini. [21:44] This is an amazing opportunity to [21:45] tighten up as much as it is to get [21:48] attacked. [21:48] >> Can I make a point about that? Cuz look, [21:50] there is so much fear right now almost [21:52] the level of panic about mythos. People [21:54] are treating it like a doomsday weapon [21:56] or something like that. It's not. is [21:58] simply that the frontier models have [22:00] reached the point where they're capable [22:02] of automating cyber activities just like [22:05] they're capable of automating coding. [22:08] But that means that a model could power [22:11] up a cyber attacker or cyber defender [22:13] the same way they can power up a coder [22:16] and allow them to discover a lot more [22:18] vulnerabilities. So there is obviously a [22:20] risk there. But I think it's important [22:22] to understand that Mythos or GPT 5.5, it [22:26] doesn't create the vulnerabilities. It [22:28] just discovers them. The bugs were [22:30] already in the code. They were sitting [22:31] there waiting for some hacker to [22:33] discover. If we can now use AI to find [22:37] these bugs in advance, these [22:39] vulnerabilities and patch them, then you [22:41] actually harden our infrastructure and [22:44] and you harden our security. I also [22:46] believe that this leap from let's call [22:49] it preAI cyber to post AAI cyber it's [22:52] going to be I think a big one-time [22:53] upgrade cycle because again you're going [22:55] to find all these dormant bugs and [22:57] vulnerabilities but I think that once we [23:00] get past that upgrade cycle you're going [23:02] to reach a new equilibrium between AI [23:04] powered cyber offense and AI powered [23:06] cyber defense it's going to become a lot [23:08] more normal it's not going to feel like [23:10] this huge disruption which is to say I [23:12] think you know people are treating this [23:14] as like some existential threat. I don't [23:17] think it is as long as everyone does [23:18] what they're supposed to do, which is [23:19] use the new capabilities to harden their [23:22] code bases and infrastructure and [23:24] security before the hackers get a hold [23:25] of these capabilities. [23:27] >> Yeah. And if Chimath, if you were to [23:29] look at this to build on Sax's point, [23:32] there are about 5 million or so security [23:34] experts in the world. We talked about [23:36] token cost. 40 hours of tokens just [23:39] pounding it, you know, a week. You could [23:42] create another five million for a [23:44] hundred dollars per chief security [23:47] officer per security expert. So it's the [23:50] volume of security expert agent saxs to [23:52] your point. Yeah. You could have 50 50 [23:55] million of them 100 million of them. [23:57] They're not finding something unique. [23:59] They're just they never sleep. They're [24:01] relentless in their pursuit of these [24:03] problems. It's a really great point. [24:05] >> Just kind of just refine that. So yeah, [24:06] there's probably 5 million people in the [24:07] cyber industry, but there's probably [24:09] only a few thousand really elite [24:10] hackers. [24:11] >> Sure, [24:12] >> those hackers didn't have the time to go [24:14] after the entire surface area of every [24:17] possible attack vector out there. And so [24:19] if you train a model to do what they do, [24:22] obviously, like you said, it can operate [24:24] with a scale and speed that a human [24:26] hacker can't. So obviously, you know, [24:28] what you need to do is get these tools [24:30] in the hands of the white hats, let them [24:33] do the cyber attacks themselves to then [24:35] find the vulnerabilities and patch them [24:37] before the black hats get a hold of [24:39] these capabilities. But I think it's [24:41] just just one last point on this, I'll [24:42] stop. It's just it's really important to [24:44] understand that the Chinese models are [24:46] going to have these capabilities within [24:47] approximately 6 months. [24:49] >> Oh, they have them now in Deep Seek 4 [24:50] for sure. They've got some level. Well, [24:52] no. Deepc4, I mean, Dec 4 is impressive [24:55] in a lot of ways, but its capability is [24:57] not at the frontier. It's maybe 80 80 [25:00] 85%. Let's call it the American [25:02] frontier. [25:02] >> Chimoff, you wanted to get in on this. [25:03] Let's get Chim in. [25:04] >> Two things. The reason that this is even [25:08] possible is because humans are [25:10] errorprone and when humans code, they [25:12] create holes. And so, humans exploiting [25:15] humans is where we've been for a long [25:18] time. Now we have computers exploiting [25:20] humans because the computers go and seek [25:21] out all these bugs that humans wrote. [25:25] In the next phase it'll be machines [25:27] versus machines. [25:29] And so I think the nature of cyber is [25:31] going to completely change. Probably in [25:32] the next five or six years there'll be [25:34] so much reason to rewrite all of the [25:38] software that runs the world. [25:41] In one part because you're going to be [25:42] asked to show more operating leverage [25:44] and revenue growth, but in another part [25:47] because everything else that was [25:48] handmade in the past is just [25:50] fundamentally insecure. Either way, all [25:52] roads will lead to all the operational [25:54] software that runs the world will get [25:56] rewritten. More and more of it will be [25:58] written by machines. More and more of it [26:00] will be impregnable as a result. But [26:02] then the cyber threat actually will only [26:04] increase because then you're going to [26:06] try to figure out how to use a machine [26:07] to inject something into another machine [26:09] so that some agentic loop inject some [26:11] malware or injects a bad token. And I [26:14] think that's a very complicated thing. [26:16] What I will tell you is [26:18] I'm not even sure if I'm allowed to say [26:19] this, but [26:22] a very good probably the best cyber [26:24] security company in the world run by one [26:26] of the very best CEOs in the world who [26:29] may or may not be speaking at liquidity [26:31] [laughter] [26:32] would tell you that they have penetrated [26:36] and can essentially [26:39] manipulate every model. Let me just let [26:42] me just say it roughly that way. [26:43] >> Okay, perfect. Yeah. And I uh at the [26:46] breakthrough prize uh which three of the [26:47] four of us were at I talked to George [26:49] Kurtz the other person you were kind of [26:51] describing was not that [26:52] >> sitting beside Nash. Yeah I'm talking [26:54] about [26:54] >> Nash and George are the two guys leading [26:56] this Palo Alto Networks Crowd Strike and [26:59] they understand the what George told me [27:01] was there is just a line out the door of [27:03] people who want this product or service. [27:06] And if you look at it, Freeberg, like [27:09] the murder rate, like we're sitting here [27:11] with the lowest murder rate in the [27:12] history of humanity. It has gone down [27:15] massively in our lifetimes, but [27:17] massively over the arc of history. I [27:18] think that's what's going to happen with [27:19] cyber. There is only so many attack [27:21] vectors, and the remaining attack [27:23] vectors are just going to be human [27:24] factors, right, Freeberg? That's always [27:26] been the case. And as we make the [27:28] software more resilient, then the the [27:31] weak link is, you know, the secretary [27:33] who puts her post-it note, you know, [27:36] with the password there or the [27:38] accountant who, you know, uses their [27:40] dog's name plus one, two, three for [27:42] their password, right? That's the the [27:44] historical one. Okay, let's any anything [27:46] you want to add, Free Bird, as we wrap [27:48] there? Oh, that's [27:49] >> Why is your bed so messy, by the way? [27:50] Why can't you just ask the room service [27:52] to come in? [27:52] >> Listen, I'll tell I can tell you what [27:54] happened. Listen, I'm here in Atlanta. [27:55] And also, why don't you have a suite [27:57] like where there's two rooms? Like, is [27:58] it just one room? This hotel only has [27:59] one. [28:00] >> It's just one room. [laughter] Yes. [28:01] >> You know, either you're cheap or poor. [28:03] Which one is it? I'm cheap. Here's I'll [28:06] tell you what. [clears throat] Here's a [28:07] situation. I'm in Atlanta for the Knicks [28:09] game tonight. [28:09] >> You're in one room. [28:10] >> Here's what I do. I just want to explain [28:12] to you value for value. Some people [28:13] spend their money on private jets and [28:15] they spend $30,000 flying to Atlanta. I [28:18] spend 30,000 on courtside seats. I don't [28:20] want the suite. I want to put it into [28:21] the seat side. [28:23] >> You can do both. I guess I could do [28:24] both, too. I don't I'm I'm in the [28:26] process of becoming [28:28] >> I don't understand [28:29] >> of embracing my richness. Okay. [28:31] >> If you've already convinced yourself [28:32] that you should spend $30,000 for [28:35] courtside tickets, which I think is [28:37] outrageous, but okay. You've already [28:38] convinced yourself like 10k each, but [28:40] yeah. Yeah. [28:40] >> A hotel room that has two rooms. Okay. [28:44] >> Probably cost 15% more than what you're [28:46] paying. [28:46] >> It's 2x, but yes, you're right. [28:48] >> I'll get the I'll get the hotel room. 20 [28:50] or 20% more, but you room like 200 bucks [28:53] a night. So you pay 400 a night. You get [28:54] another hotel. [28:55] >> I mean it's it's Atlanta. The most I'm [28:57] in the best hotel the most expensive [28:58] hotel is 500 a night in Atlanta. It's no [29:00] big deal. But everything's sold out [29:02] because all the Knicks people are coming [29:03] here. [29:03] >> So you're selling me double that would [29:04] have been a thousand and you couldn't [29:05] spend,000. [29:06] >> Everything is sold out because the [29:08] Knicks are here. [29:08] >> So we have to look at your dirty beds. [29:10] >> It's gross. [29:11] >> The bed's not that dirty. Come on. Just [29:13] deal with it. Okay. Take it out and [29:14] post. [29:15] >> I have a private jet story about flying [29:17] to Atlanta. [29:20] >> You reminded me. Okay. So yeah, there [29:22] was some event there. So I I flew my [29:23] team there, you know, there's a few [29:26] people on my plane and it's kind of a [29:28] long flight. Was it like 4 hours or [29:29] something [29:30] >> from the back? Yeah. [29:31] >> Yeah. So I went in the back to to sleep. [29:34] Well, first, you know, we we started the [29:35] flight and I had a few bottles of Papy [29:37] Van Winkle on on the plane. And so we [29:40] started off with like a drink and then I [29:42] went in the back and and fell asleep and [29:43] I woke up basically when we landed. [29:45] >> So I come out and like all three bottles [29:47] are basically cashed of like [laughter] [29:49] Happy Van Wink. [29:50] >> Oops. Those were like two grand a [29:52] bottle. [29:53] >> No, no, they're more. These were like [29:54] antique bottles. Like one of them was [29:56] >> I have one of those from your plane. I [29:57] have one of those from the old Falcon. [29:58] Yeah. [29:59] >> Yeah. They were like these vintage [30:01] >> $4,000. I remember. Yeah. [30:03] >> Anyway, you can't even find this [30:04] anymore. So these guys, they asked me [30:07] like when we land like, "Hey, Sax, how [30:09] much did it cost for you to fly us to [30:12] this event?" And I said, "Well, about [30:14] $8,000 in jet fuel and about $12,000 of [30:17] Happy [laughter] Van Winkle." [30:20] Well, you gota you got to fuel the the [30:22] vibes as well as the plane. It's uh [30:25] >> Is Atlanta nice? I've never really [30:27] >> Do those people still work for you or [30:28] are they are they uh [laughter] [30:31] >> they called in Atlanta? Um is Atlanta [30:33] nice? Listen, last year I went to the [30:34] Detroit games and that city was on the [30:36] rebound. Atlanta has an incredible [30:37] opportunity to rebound. I'll say it that [30:40] way. There's a great opportunity for [30:42] them to upgrade the city. I I went to [30:44] Waffle House at midnight last night. [30:46] There was no shootings. Okay, let's keep [30:47] moving. By the way, do you get royalty [30:49] points at the Best Western Atlanta or [30:51] No, [laughter] [30:52] >> I get double points because I use my [30:54] Best Western uh Visa card. Yeah, it's [30:56] everywhere you want it to be. All right. [30:59] Use the promo code Jcal and get a [31:01] thousand extra points. In other Open AI [31:04] news, Musk versus Alman, the trial of [31:08] the century or maybe the decade has [31:11] started. Elon is of course accusing Open [31:13] AI of breach of charitable trust, unjust [31:15] enrichment. He's accusing Open AAI of [31:18] essentially flipping a nonprofit into a [31:21] for-profit. He's seeking 150 billion in [31:23] damages that they revert back to a [31:25] nonprofit that Alman and Brockman be [31:28] removed. And there were some fireworks [31:30] between Elon and the Open AI lawyers. [31:32] Elon kind of leveled up the discussion. [31:34] He said, quote, "If we make it okay to [31:37] loot a charity, the entire foundation of [31:39] charitable giving in America will be [31:42] destroyed. That's my concern." [31:44] Obviously, there's a ton of interesting [31:47] nuances here. Specifically, Greg [31:49] Brockman keeping a diary where he was [31:52] journal maxing his plans uh like a Bond [31:55] villain here. And uh the excerpts from [31:59] his diary include conclusion, we truly [32:02] want the BC Corp. The true answer is [32:04] that we want Elon out. If 3 months later [32:07] we're doing BCorp, then it was a lie. [32:09] Can't see us turning this into a [32:11] forprofit without a nasty fight. I'm [32:13] just thinking about the office and we're [32:15] in the office and this story will [32:16] correctly be that we weren't honest with [32:19] him. In the end, it's still about [32:20] wanting a for-profit just without him. [32:22] yada yada yada. Freeberg, your thoughts [32:24] on this case? Is Elon going to win? I [32:27] just don't know why Greg Brockman's got [32:29] a freaking diary where he's like [32:31] literally documenting. I mean, I love [32:33] the guy, but what the is he [32:34] thinking? Like, you're just sitting here [32:36] at home and like, let me write about the [32:38] the crime I'm committing or let me write [32:40] like and let me record it. And by the [32:42] way, let me never delete it. I don't [32:44] understand this. [32:45] >> It's not just journal maxing. It's [32:47] discovery maxing. [laughter] [32:49] >> It's smoking gun maxing. [32:51] >> I don't get it. I don't get it, man. [32:54] >> I mean, do you guys remember from the [32:57] wire in that scene where the guy's like, [33:00] "Is you taking notes on a criminal [33:02] conspiracy?" [laughter] [33:04] He's got everybody in the room. Can we [33:06] play that clip? It's like, [33:08] >> "What are you doing, Greg? is you [33:10] taking notes on a criminal conspiracy? [33:14] What the is you thinking, man? [33:16] >> If you're going to commit a crime, you [33:18] do not write down the date and time of [33:21] the crime in your journal. [33:23] >> Well, look, we don't know it's a crime. [33:24] Let's not. [33:25] >> Okay, sure. [33:27] >> A crime, but yes, you keeping [33:30] shenanigans. Jamat, do you keep a diary? [33:32] >> What do you think, Juel? Do you keep a [33:34] diary? [33:35] >> I I believe ruminate. [laughter] No, [33:38] I'll tell you right now, rumination is [33:40] the path to unhappiness. Nobody gives a [33:43] about your feelings. Writing your [33:44] feelings down is only going to make you [33:46] miserable. Talking to your spouse about [33:48] your feelings. [33:50] >> Just go to a beautiful dinner, sit [33:52] courtside at the next, and do what I've [33:54] been doing for 30 years. [33:56] maxing. [33:57] maxing. And the register goes up. [34:00] All you have to do is work. Start new [34:02] projects. Nine out of 10 foul. Place [34:04] nine out of 10 bets. One wins and you're [34:06] golden. Go sit courtside at the Knicks [34:08] game. [34:09] >> Keep going. Life's too short. [34:10] >> And just keep moving forward. Don't [34:12] write anything down. Period. Full stop. [34:15] >> It's good advice. Yeah, I just The [34:17] biggest surprise to me was this guy's [34:18] got a diary. I just I don't know anyone [34:20] that has a diary. I've never heard of [34:21] this. So anyway, that was shocking. [34:24] Besides that, I have no view on what's [34:26] going to happen with the case or what [34:27] the judge will do. [34:28] >> I have no comment on the case either. I [34:30] think it's weird that Poly Market hasn't [34:33] budged even as all of this discovery has [34:35] been published. It's effectively at 42 [34:37] or 43% that Elon wins. So, one of the [34:41] friends in our group chat said what may [34:44] just happen is that Elon technically [34:45] wins and he's just credited back the $40 [34:48] million. And so, maybe that's what this [34:51] poll is front running. [34:54] But on a totally separate note, I think [34:56] Jason, I know you say it as a joke, but [34:59] this idea of just keep moving forward, [35:02] don't ruminate, I think is very good [35:03] general life advice for everybody to [35:06] follow. [35:06] >> The modern-day therapy industrial [35:09] complex and the medication industrial [35:11] complex, I believe, is [35:14] >> around rumination. Well, it does pivot [35:15] around rumination. [35:16] >> Yes. [35:17] >> That is the gateway drug to all these [35:18] things. [35:19] >> Yep. Talk about your problems. [35:21] You know, when these people go to [35:22] therapy, you ever hear these people? [35:23] Howard Stern's like, "I've been in [35:24] therapy with the same person 2 or three [35:25] days a week for 40 years." I'm like, [35:27] "Okay, what's the incentive for the [35:28] therapist to stop charging you $1,200 an [35:31] hour? There is none." Then they lose a [35:33] revenue stream. They lose a customer. [35:34] It's all a giant fraud. Facts. [35:37] Uh, in terms of this case, [35:38] >> I wouldn't go that far. I do think that [35:40] there's a lot of value in kind of [35:42] untying some of these Gordian knots that [35:44] people have because of how they grew up. [35:47] But there's a difference between that [35:49] and being specific and just randomly [35:51] ruminating cuz I don't think there's a [35:53] lot of productive. [35:53] >> You've got an acute issue like in trauma [35:57] in your life. Yeah, sure. Unpack it, [35:59] figure it out. I'm just talking about [36:00] this neverending self-improvement, [36:03] you know, ruminating thing. Uh but [36:05] getting back on topic here, Saxs, [36:09] what's the And we're we're talking about [36:11] a jury, I believe, in Oakland. [36:13] >> No, but it's a bench trial. This is [36:14] important. It's a bench trial where the [36:16] jury is advisory in capacity, [36:19] >> but ultimately that judge, she will make [36:21] the final call [36:22] >> and she'll do the damages. And so, is [36:25] this a case sacks of like we've got a [36:28] Bay Area jury judge and we've got Elon [36:33] who's considered, you know, a bit [36:35] right-wing and people don't all agree in [36:37] that area in terms of his politics. And [36:39] then you have this Sam Alman New Yorker [36:42] story and people finding out that so [36:45] many different people feel they got [36:47] screwed by him. You put these two things [36:48] together, it's impossible to handicap [36:50] where this turns out. Sachs, your [36:52] thoughts? [36:53] >> Well, yeah, I don't think this is about [36:55] politics. I mean, I guess you could [36:58] argue that what Elon is seeking, which [37:01] is to protect the charity, is if [37:03] anything a left-coded sort of principle, [37:06] although I don't really think it's left [37:07] versus right. Look, I don't want to take [37:09] sides on this trial. I'm just watching [37:11] like everyone else. The last time I [37:12] weighed in on some Elon litigation, I [37:16] got deposed for six hours. Remember [37:18] that? Cuz they just assume that somehow [37:20] I know something, [37:21] >> right? [37:21] >> I've never talked to Elon about the [37:22] case. I don't know anything about it. [37:24] >> Yeah. [37:25] >> I'm going to see what happens like [37:26] everyone else. Now, one thing I I will [37:29] say having just read some of the [37:31] coverage is that apparently the company [37:35] at some point did offer Elon shares in [37:38] the company, but he thought that there [37:41] was something kind of icky about it. Do [37:43] you remember this that [37:44] >> Yes. Because at at one point I said on [37:46] our show when this dispute started [37:49] happening but before it became a court [37:50] case I said look if Open AAI at a [37:53] certain point decided they had the wrong [37:54] structure they should just gone and done [37:56] a make right with Elon and he should [37:59] have been a shareholder on the cap [38:00] table. What I didn't know is that [38:02] apparently they did try to do something [38:04] like that but Elon turned it down [38:06] because he did want the entity to remain [38:10] a charitable entity. In other [38:13] >> had a principled view of it according to [38:15] the reports and was like no we're trying [38:17] to save humanity and then you're giving [38:20] this keys to the kingdom to Microsoft [38:21] that's all come out [38:24] >> and I I also have not talked to Elon [38:26] about any of this but my guess is like [38:30] most of these things there'll be some [38:32] sort of settlement or something here but [38:34] maybe he takes it to the mat who knows [38:36] Judge Rogers who's doing this [38:38] 61-year-old Obama appointee politics has [38:42] played a role. Saxs, they have had to [38:44] tell the jury like however you feel [38:45] about these individuals politically, [38:47] whatever, please put that aside. But of [38:49] note is that she oversaw the Epic Games [38:51] versus Apple trial over App Store [38:54] exclusively ruled in favor of Apple with [38:56] some caveats um that they don't have a [38:59] monopoly, etc., etc. So, this is going [39:02] to be a really interesting one. And I [39:04] think the worst case scenario is open AI [39:07] for for OpenAI is they have to unravel [39:09] this somehow and that would delay the [39:12] IPO. That would cause chaos in [39:14] shareholders and I guess the best case [39:16] is some sort of settlement. And if Elon [39:19] put the first 40 or $50 million in, he's [39:21] he's due 10 20 30% of the company after [39:24] dilution. [39:26] All right, let's keep moving through the [39:27] docket. Lots more to discuss and uh good [39:30] luck to everybody in their lawsuit and [39:31] those of you betting on market all-in [39:34] summit selling up fast. Our fifth [39:37] edition Los Angeles September 13th to [39:39] 15th. Go to allin.com/events [39:41] and uh speakers are going to be top [39:43] tier. Apparently Freeberg is having this [39:47] as his major creative outlet. I heard [39:49] some back channel chimoff today that [39:51] he's going to be doing Broadway musical [39:55] uh illusionist. tap. I got a tap dancing [39:58] situation. [39:59] >> He's literally going fullon entertainer. [40:02] This is going to be vaudeville sachs [40:05] wrapped up. He's just going to take it [40:07] to a whole new level. Musical numbers [40:10] >> like Nathan Lane. [40:12] >> I think if you're coding that it's going [40:15] to be his big gay summit. Yes, it could [40:17] be a big gay summit. [40:19] >> Might be our last year in LA, guys. [40:22] >> Why? [40:23] >> Not might be. [40:24] >> Might be. Oh, everybody wants to go to [40:25] Vegas apparently. [40:28] Those bones, baby. Can you imagine [40:31] leaving the summit for lunch and going [40:33] and playing crabs? Jimoth, we get a [40:35] fresh. Yes, I can. Yes, I can imagine. [40:39] >> I got some bricks. Oh, I got some bricks [40:41] right here. Let's go. Yum, yum. [40:44] >> That way Sachs can come. [40:46] >> Yeah, [laughter] Sax is like, I'm never [40:48] setting foot in California, but I will. [40:52] You know, we're doing a couple live [40:53] events. Are you coming to them? [40:54] >> Liquidity or something different? [40:56] >> Liquidity. And then there's the the [40:57] all-in summit happens in September. [40:59] >> Yeah, I'm going to do those, too. [41:00] >> All right. Big tech smashed their [41:02] earnings on Thursday. Google, Microsoft, [41:05] Amazon, and Meta all reported. I don't [41:07] know why they do this on the same night, [41:09] folks, but they do. And performance was [41:11] spectacular. It was great. However, the [41:14] capex announcements were really the [41:19] story here. Let me just cue this up and [41:23] show the chart. [41:25] $725 billion in capex guidance in 2026 [41:30] from but four companies. Amazon, [41:32] Microsoft, Google, and Meta. Amazon [41:34] leading the pack with 200 billion, 190 [41:37] billion each for Microsoft and Google, [41:38] 145 billion for Meta. You add Grock, you [41:42] add OpenAI and some other players to [41:45] these plans. And we haven't heard from [41:46] the new Apple CEO yet, but he's going to [41:48] be taking over. And he's going to have [41:50] some plans here. I'm sure we are going [41:53] to see the large a trillion dollars a [41:55] trillion dollars in buildout over the [41:57] next year. I don't know if this is even [41:58] possible, but this is all being driven [42:02] by AI and cloud computing. Google Cloud, [42:06] which includes the Google Suite, that [42:08] grew 63% year-onear. [42:11] Let that number sink in. 63% on 20 [42:14] billion in revenue. That's in a quarter. [42:16] Microsoft cloud, that includes Azure, [42:18] Windows Server, SQL Server, they bundled [42:20] some things together there to get the [42:22] number to go up. Uh that grew 30% on [42:25] 34.7 billion in revenue. Amazon Web [42:28] Services, the original cloud, that grew [42:31] 28% on 37.6 billion in revenue. That's a [42:34] bit of a pure play. Just counts Amazon's [42:37] web services. Obviously, these are all [42:39] moving to NeoClouds. These are all [42:41] serving AI jobs and tokens now. They [42:44] have a massive customer base and the [42:46] customers from the smallest startups all [42:48] the way to the biggest frontier models [42:49] cannot get enough compute and it is [42:52] going to the bottom line. But this is [42:54] shrinking Chimath cash flow massively. [42:57] These were free cash flow machines, the [43:00] largest money printing machines in the [43:02] history of humanity. But they are giving [43:05] up on free cash flow, stock buybacks and [43:07] dividends and the focus on those three [43:11] to invest in infrastructure. Amazon's [43:13] free cash flow down 97% [43:16] Google, Microsoft and Meta down 12, 12 [43:18] and 8% respectively. your thoughts on [43:21] this free cash flow, the end of the free [43:23] cash flow deluge and the massive massive [43:27] investment we're seeing in capex. [43:30] I think we're seeing a very important [43:33] structural shift in the capital markets. [43:36] I think the last 20 or 30 years, well 20 [43:39] years, it's been that the mag 7 just [43:42] kind of ran away with it. that these big [43:45] companies got bigger and bigger and it [43:47] absorbed all of these investment dollars [43:49] and the biggest reason was that it had [43:52] these very assetike business models, [43:55] right? You just built some more software [43:56] and it just has all this leverage and it [43:58] all just kind of worked except maybe for [44:00] Amazon cuz they needed physical [44:01] infrastructure for warehouses and [44:03] delivery and whatnot. But by and large [44:04] it was a very asset light investment [44:07] cycle. Now all of a sudden the pendulum [44:09] is swinging violently in the other [44:11] direction. And there's something that I [44:13] think people misunderstand which is as [44:15] it moves back to these asset heavy [44:18] infrastructure investments. [44:22] The hyperscalers are signing checks that [44:25] I mean I suspect their body can cash but [44:27] there's a world in which they can't. [44:29] I'll give you an example. You know when [44:31] Microsoft convinced the owners of three [44:33] Mile Island to turn their [44:37] >> nuclear site back on? [44:38] >> Yeah. [44:39] >> Do you know what their Ford purchase [44:41] agreement was? It was for more than 2x [44:43] the prevailing spot rate for energy. [44:45] More than 2x. The problem is that's not [44:47] for an enormous percentage of their [44:50] overall energy needs. [44:53] So if you play that out and you think [44:56] these five or six companies all of a [44:58] sudden are not just spending Jason 700 [45:01] billion a year of capex which they are [45:05] but then from an operating cash flow [45:07] they're going to be spending 2x the [45:09] prevailing spot rate because they just [45:11] want guaranteed demand into the future. [45:15] Where's all this cash going to go? It's [45:18] not going to go [45:20] to the shareholder and it's not going to [45:22] stay on the balance sheet. These [45:24] companies will now get levered. They're [45:26] going to get highly sophisticated around [45:29] the financial engineering. They'll have [45:30] more debt. They'll have all kinds of [45:33] different vehicles and term loans and [45:35] revolvers and all of this stuff. [45:37] And so, they're going to look like this [45:39] big bulky industrial business in five [45:41] years. And I'm not sure that there's a [45:44] good valuation case to be made at that [45:46] point. [45:47] And so I think it may be simpler and [45:50] this is what I tweeted to just follow [45:52] the dollars like a trillion dollars a [45:55] year going out of the hyperscalers. [45:57] Where is it going? Just follow those [45:59] dollars and buy those companies because [46:00] those companies are already underpriced. [46:03] This is uh obviously reminiscent of [46:05] something we all experienced. Uh Nick, [46:07] can you pull up the Cisco chart I just [46:08] sent you and put it at max? uh we had a [46:11] massive buildout of the infrastructure [46:14] of the internet in the late 1990s and [46:17] into 2000. And what that caused was a [46:20] lot of aggressive companies to do [46:21] massive amounts of spending, a lot of [46:23] retail investors to embrace these stocks [46:25] like we're seeing with people trying to [46:28] get into these private companies and [46:30] saxs. Look at the 2000 peak of Cisco. [46:33] This is the most extraordinary chart [46:35] ever. It took them 25 years to get back [46:37] to that peak and uh they had a lost two [46:41] decades and we had a massive amount of [46:44] fiber that wound up getting bought. We [46:45] talked about that a couple years ago on [46:47] the program. But there's something for [46:48] you to build off of here when you look [46:50] at this massive infrastructure. You [46:52] think it's going to be Cisco systems all [46:54] over again, World Warcom, etc.? [46:55] >> No, I really don't. The issue we had in [46:58] 2000 was dark fiber. You had all this [47:00] infrastructure being built out and it [47:02] wasn't being used. There's no dark GPUs [47:05] today as you know Brad Gersonner likes [47:07] to say. So what's driving the capex now [47:11] is the voracious demand for compute for [47:15] tokens and the demand is now pulling [47:19] forward this additional um investment in [47:22] infrastructure. So I think what's [47:25] happened here is that the bull thesis [47:26] for AI just got validated in a single [47:30] afternoon. I mean again you got [47:31] Microsoft Azure, Google Cloud, Amazon [47:34] AWS, Meta, they're all basically [47:37] exceeding expectations, exceeding [47:39] guidance in terms of where their cloud [47:42] revenue would be and therefore how much [47:44] they're going to reinvest in capex this [47:47] year. I think we were supposed to have [47:48] 660 billion of hyperscaler capex up from [47:52] 350 last year. I think there's now the [47:54] new estimate is it's going to be over [47:55] 700. So this is you know again it's more [47:58] than 2% of GDP. This is a huge tailwind [48:00] to GDP. There's another article saying [48:02] that I think in the last quarter [48:05] AI was 75% of GDP growth. And by the [48:08] way, this is just the capex part. This [48:10] is the physical infrastructure. This is [48:12] not the economic impact of the tokens [48:14] that are generated inside the token [48:16] factory. This is the building of the [48:18] factories. How do those tokens get used? [48:20] like we're seeing they're being used not [48:23] just to do research or to answer [48:25] questions but to create code. And so [48:29] we're seeing this explosion of [48:31] productivity in software development. [48:33] And we're seeing an explosion of bespoke [48:36] software being created and that's going [48:38] to accelerate every part of the economy. [48:40] Every business that now wants to get [48:42] code will be able to get code for the [48:44] first time. Before they couldn't even [48:45] hire the engineers, they needed to [48:47] generate it. Now they will be able to. [48:49] So that is a huge unlock of productivity [48:52] across the economy. Then you're getting [48:54] into these new use cases like the the [48:56] co-working use cases and agents, right? [48:59] So the the workflow automations that are [49:01] happening, it's still early. I don't [49:04] believe that this is going to replace [49:05] humans. We had that um in the past week, [49:06] we had that crazy case of an agent [49:09] deleting a production database in 9 [49:11] seconds because because of a bug. Look, [49:13] what that said to me is that [49:16] >> it's not that agents aren't valuable. [49:18] They are valuable, but they have to be [49:19] supervised. You know, this idea that [49:21] you're just going to be able to like [49:22] automate all the jobs away. It is a [49:24] massive amount of handwaving over the [49:27] real technical problems and issues. The [49:29] agents have to be supervised. Someone [49:31] has to be accountable. It's not going to [49:33] be the CEO. The CEO doesn't want to be [49:35] accountable for thousands of agents. You [49:38] need people [49:38] >> despite what Jack had block said. [49:40] >> Yeah. Thousand direct reports is a great [49:43] like goal, but it's not realistic. Yeah, [49:46] >> you need IT people who are savvy who can [49:49] supervise this and make sure it's [49:50] working. They have to be accountable to [49:51] the CEO. Someone has to drive the [49:53] productivity. It's like Bology always [49:56] said, AI is not end to end is middle to [49:58] middle. You have to have someone to do [49:59] the prompting and you have to have [50:00] someone to do the validating and I would [50:02] add the supervision and accountability. [50:04] So anyway, the larger point though is [50:06] I'm speaking to the fact that I don't [50:08] think there's going to be this huge job [50:10] loss associated with this productivity [50:12] boom that we're going to get. And in [50:13] fact, I think what's actually happening [50:15] now is that AI is becoming synonymous [50:18] with the American economy. I mean, the [50:20] fact that it's generating 75% of GDP, [50:23] you have this capex explosion, this [50:26] energy explosion that feeds it, and [50:28] again, just the beginning of the [50:31] applications that are being unleashed by [50:33] these new token factories. I think it's [50:35] all a very, very positive thing. and all [50:38] these doomers who are trying to throw a [50:40] wet blanket on it or constantly scaring [50:43] the daylights out of people. I mean, [50:45] what do they want the American economy [50:47] to do? Just to stop I mean, they just [50:49] don't want any progress. I mean, like [50:51] again, you know, when you talk about [50:52] stopping AI or halting AI progress? What [50:55] you're really doing is stopping the [50:57] American economy now. You're basically [50:59] saying you don't want economic growth. [51:01] AI is now synonymous with the growth of [51:04] the American economy. And if there's no [51:06] economic growth, there's not gonna be [51:07] money to pay for all the social [51:08] programs. There's not gonna be money to [51:09] pay down the national debt. There's not [51:11] gonna be money to basically build up our [51:13] national defense. All these things we [51:14] want to spend money on. We have to have [51:16] a vibrant economy. And that is now [51:19] synonymous with AI. So I know that AI [51:21] may not be popular. I see those polls. [51:23] But having a strong economy is popular. [51:26] And I believe that those things are now [51:28] synonymous. It's almost like there was [51:30] some architect or ZAR who set up the [51:33] chessboard in the first year of this to [51:35] make sure that it was ultra competitive. [51:38] >> President Trump set the table on this. [51:39] >> Absolutely. With some good advice, I [51:41] think. Maybe. [51:42] >> Freeberg, your thoughts? [51:43] >> It's always good to have good advisors. [51:45] >> Always good to have good advisors. [51:46] Absolutely. Absolutely. [51:47] >> No, but look, I've said it before. The [51:49] president just wants America to win. [51:50] >> Literally, there are people who if we [51:52] were looking at this, you know, I don't [51:54] know, a hundred years ago, it'd be like [51:56] people were like, "Yeah, you know what? [51:57] we shouldn't build the highway system or [51:58] we half built the highway system. Let's [52:00] stop let's stop building the highways. [52:02] >> No, the highway system was funded by the [52:04] federal government. There was no [52:05] competition. It was the most expensive [52:07] on a on a inflationadjusted basis. I [52:10] think it was the most expensive project [52:11] in US history. [52:13] >> Yeah. And the railroads before that like [52:15] you can't stop these things. They have [52:17] to keep going. It's interesting point [52:20] you know there is so much demand for the [52:22] resource of tokens of intelligence [52:24] freeberg and it's quite different than [52:26] the fiber situation as Sax correctly [52:28] points out where we [snorts] built all [52:30] this but we didn't actually have an [52:32] application here the application is [52:34] pretty um pretty wellnown and you've got [52:37] a large number of people in businesses [52:39] who are trying to vibe code their way to [52:42] success trying to push this stuff and we [52:45] had an interesting story referenced [52:46] earlier in the show [52:48] where [52:50] uh Claude ate somebody's homework. This [52:52] is the nightmare of all nightmares. [52:55] Somebody was vibe coding. Uh it was the [52:57] founder of Pocket OS. Apparently, they [52:59] make software for rental car companies. [53:00] He was using Opus 4.6 through Cursor's [53:03] AI platform, their coding platform, and [53:07] uh you know, which is like the most [53:09] expensive tier. Uh and he said he [53:12] configured it with enough safety rules, [53:13] but the agent was working on a routine [53:16] task. They saw some sort of credentiing [53:18] mismatch and they decided to fix the [53:20] mismatch by deleting a railway volume [53:22] without user confirmation and uh they [53:25] pushed the code from a repo to a live [53:27] app and they deleted everything [53:28] including the backups. Literally a scene [53:32] from Silicon Valley's HBO clip of Son of [53:36] Anton. Hilarious. [53:37] >> You gave your AI permission to overwrite [53:39] code in the internal file system. Were [53:42] you going to tell me about this? No, I [53:44] thought that was the company policy [53:46] these days. [53:47] >> Okay, well, your AI just failed [53:50] epically. [53:51] >> That's unclear. [53:53] >> It's possible the Son of Anton decided [53:55] that the most efficient way to get rid [53:56] of all the bugs was to get rid of all [53:58] the software, which is technically and [54:01] statistically correct. But artificial [54:03] neural nets are sort of a black box. So, [54:05] we'll never know for sure. [54:06] >> How did they get that so right, Zach? [54:08] Five or six years ago, art and neural [54:10] networks are a black box. So, I guess [54:11] we'll never know. But technically, it [54:14] was correct. Freeberg, when you blow up [54:16] a hollow system with your vibe coding, [54:19] which you were absolutely showing off in [54:21] front of Jensen a couple of weeks ago [54:23] about how much code you're pushing, who [54:24] are you going to blame? You going to [54:26] take responsibility yourself? Are you [54:27] going to blame Cla Claude or Kurser? Who [54:30] are you going to blame when you blow up [54:32] the entire stack over at Ohio? [54:36] >> Who you blame? [54:37] >> Yeah, I'll blame Dario. [54:38] >> You blame Dario. Okay, that's what I [54:40] thought. That's a correct answer. [54:41] Correct answer. Blame Daario. He's the [54:42] one who says it's a doomsday machine. [54:44] Uh, come on the prodio. [54:47] 17th invitio. [laughter] [54:50] >> I mean, I've invited the guy like 17 [54:52] times. He is totally going to me. He [54:54] wants nothing to do with this podcast. [54:57] >> Actually, let me speak to that. So, I [54:59] think I think that um there's maybe a [55:02] misperception that this error occurred [55:05] because of, you know, quote unquote AI [55:07] scheming, [55:09] >> like kind of in that video that the AI [55:12] decided that the best way to get rid of [55:13] bugs is to basically eliminate the [55:14] codebase. This is kind of like the, you [55:16] know, AI is going to turn the world into [55:18] paper clips type thing where somehow [55:19] it'll like miss scheme. That's not [55:21] really what happened here. This is a [55:22] case of just a of old-fashioned bugs [55:26] occurring at an edge case. You know, [55:28] you've got the fact that this API was [55:30] not designed for permissioned usage. [55:34] You've got the fact that a credential [55:36] was left kind of lying around. Probably [55:38] it should not be. There's kind of like a [55:40] perfect storm that caused the AI to do [55:42] something or the agent to do something [55:43] that didn't quite understand it was what [55:44] it was doing. I think that if there is a [55:48] systemic problem here rather than just [55:50] kind of a like a random edge case is [55:53] that AI still doesn't know what it [55:58] doesn't know. You know, like a human [56:00] would stop before deleting a production [56:03] database and just say, "Oh, I'm about to [56:04] do something like really serious, really [56:06] destructive. Am I sure I want to do [56:08] this?" You know, and a human would have [56:10] stopped and said, "Oh, wait a second. [56:11] like I need to be more confident in what [56:13] I'm doing before I take that action. And [56:16] AI still has this issue where again it [56:18] can be kind of overcon. This is where [56:19] like the hallucinations come from is it [56:22] doesn't know when it should have a low [56:24] confidence in its output, right? But [56:26] this is why it has to be supervised. You [56:29] know, the longer the time horizon for a [56:31] task, the more likely it is to go off [56:34] the rails. [56:35] >> And a drift. Exactly. And this is why I [56:38] think people are starting to realize [56:40] that this idea of eliminating all [56:41] software developers was the peak of [56:44] inflated expectations. Yes. [56:45] >> Right. There was actually a really good [56:48] tweet on this by Aaron Levy who's got [56:50] the right take on this. Aaron retweeted [56:53] Matthew Glacius who sort of sardonically [56:56] tweeted that 5 months in I think I've [56:59] decided I don't want to vibe code. I [57:01] want professionally managed software [57:02] companies to use AI coding assistants to [57:05] make more better, cheaper software [57:07] products that they sell to me for money. [57:08] >> Just lower your prices. Don't make me [57:10] vibe code is the translation. [57:13] >> Yeah. I mean, I think like rare win for [57:15] for Madaglacius there. Anyway, Aaron [57:17] Levy then says [57:19] Agent Coding is a huge win for software [57:21] developers that want to get more done [57:23] and it's fantastic for anyone curious to [57:26] learn how to start coding. What it's [57:28] less great for is casually building [57:30] complex software that you have to [57:32] maintain on an ongoing basis and take [57:33] all the risk for upgrades, maintenance, [57:36] keeping up to date with latest security [57:37] issues, you know, the bugs, cyber, those [57:40] are taxes on most knowledge workers who [57:42] aren't familiar with the system. [57:45] >> It's not a tax. It's a huge risk. [57:47] >> Yes, it's a risk has to be managed if [57:49] you don't. People will get fired because [57:52] there will be some public companies [57:53] where some goofball tries to vibe code [57:55] their way out of something and they're [57:56] going to torch the enterprise value. [57:58] It's going to be glorious to watch [58:00] because we're all going to laugh and [58:01] realize that was stupid and should never [58:02] have happened in the first place. [58:04] >> Yeah. I mean, it's there is a chance [58:07] that this improves to the point passes [58:10] trial of disillusionment and becomes [58:11] super productive and you'll be able to [58:13] get an agent to do reasonable things [58:15] without deleting your data set. But we [58:17] have a way to go. Here is your, you [58:20] know, this is the tech adoption chart. [58:22] Basically, you got a technology gets [58:23] triggered. You have the trial, you have [58:24] this peak of inflated expectations. You [58:26] go into the trial of disillusionment and [58:27] then the slope of enlightenment invest [58:29] and eventually it becomes deer and it's [58:32] an opportunity. Hey, uh, Freedberg, [58:36] you have become reddit tide curious. You [58:41] have and also [58:43] >> tell me tell me about rea cuz I want it. [58:46] I want to get on it. M [58:47] >> I want to use it and I need you to tell [58:50] Nat that it's okay for me to take it. [58:52] >> I I have a friend who has some advice as [58:54] well. [58:55] >> Freeberg, the coverage is coming out of [58:57] this phase three clinical trial data [58:59] release that Lily put out last month. So [59:01] everyone's going crazy over the data [59:05] which continues to show pretty amazing [59:08] results. So unlike trazepatide which is [59:11] kind of Lily's main [59:15] product today, it's a which is a dual [59:16] agonist. It's got two peptides in it [59:18] that that bind to different receptors, [59:20] the GLP1, the GIP receptor. This other [59:22] one now also binds to glucagon, which is [59:24] a third receptor. And that glucagon [59:27] receptor binding peptide causes the [59:30] cells to increase their metabolism, [59:31] which actually accelerates fat energy [59:34] consumption [59:36] over what would typically be muscle [59:38] energy consumption. It's more likely to [59:40] burn up fat early on, which causes more [59:44] quick fat loss, but also reduces muscle [59:47] loss. And some of the other data that's [59:50] now coming out shows non-HDL cholesterol [59:53] down 27%, triglycerides down 41%. [59:57] >> Liver fat down 80% to [60:00] >> 80% reduction of liver fat. A1C drops [60:04] from 7.9% to 6% in 40 weeks, which is [60:08] amazing, by the way. If you're diabetic [60:09] and your A1C drops that much in a couple [60:12] of months, it's literally a life-saving [60:14] product. The average user in this phase [60:17] 3 trial saw their weight decline from [60:20] 214 pounds. They lost 37 pounds. That's [60:23] compared to six pounds on placebo [60:26] in 40 weeks. And you know, modest side [60:29] effects. 20% people felt more nauseous [60:32] than the people that were on the [60:33] placebo. There's a lot of other separate [60:36] studies that are being done now that are [60:37] showing significant reductions in [60:39] inflammatory signaling molecules. So [60:42] systemic signaling of like hey cells are [60:46] in distress triggers this kind of [60:49] inflammatory process that can have a lot [60:51] of other damage to your body can [60:53] accelerate aging. And so one of the [60:55] other conversations is that retride [60:57] might actually be kind of a deaging drug [61:00] as well. [61:02] Hercules, Hercules, Hercules, [61:05] you know, and a lot of the studies, by [61:06] the way, are done on the the the very [61:08] high dose, 12 milligram dose, but you [61:10] could probably get this thing dosed down [61:11] to 2 milligrams and still see a lot of [61:13] the anti-inflammatory [61:14] maintenance and other benefits. I'm no [61:16] doctor, but people are going nuts over [61:19] this being more widely useful than just [61:22] for clinical obesity or type when the [61:24] FDA when's the projected date for [61:27] >> 2027. Mid 27. [61:28] >> That's what they're saying. Could happen [61:30] sooner. I mean, the data is in the, you [61:32] know, the FDA will take their time to [61:36] evaluate it, but I think given the way [61:37] this is all looking, [61:39] >> could happen sooner, could happen [61:40] sometime later this year. Swim Chimath [61:43] Swim said it's incredible and that uh [61:47] it's living up to the hype in their [61:49] experience. [61:50] >> Who? [61:51] >> Swim. [61:52] >> What is that? What is that? [61:53] >> Someone who isn't me. Swim. [61:55] >> Oh, [61:55] >> this is a Reddit term. Someone who isn't [61:57] me said who has a guy swim has a guy and [62:02] has cycled on reddatride and does [62:06] push-ups and says muscle gain has been [62:09] spectacular [62:10] no muscle loss and a lowering of fat. [62:14] >> If you go on X and you just search up [62:16] rea [62:17] >> Mhm. [62:19] >> it's like incredible. You see these like [62:21] 65 year old guys that go from a dadbod [62:24] to looking like an incredibly ripped [62:27] athlete in weeks. And and I [62:31] I mean I'm shocked. And then for me I [62:34] don't need that help per se, but my [62:36] liver health is important to me. My [62:37] cardiac health because I'm South Asian [62:39] and it just looks like a wonder drug. I [62:41] can't wait. When you starve your body, [62:43] when you turn off the the appetite, [62:45] which is the GLP-1 agonist function, [62:47] normally your body goes into this kind [62:49] of mode of starvation and you have this [62:52] process by which your body tries to [62:53] generate energy from your existing [62:55] cells. And because muscle is much denser [62:59] than fat, you can have a favoring of [63:01] muscle tissue being kind of broken up [63:03] over fat tissue. But what this new [63:05] agonist, this glucagon agonist that they [63:07] put into this neutrutide [63:10] is it favors fat burning over muscle [63:12] burning. And so that actually can drive [63:15] short-term use at low dose for people to [63:18] cut weight and maintain muscle and get [63:20] ripped. And so that's why a lot of [63:21] people in the kind of fitness community [63:23] are talking about, hey, I want to get [63:24] access to this and get on it for a [63:26] while. So you'll see a lot more hype [63:27] probably in that community as well as [63:30] the all the health effects. It just [63:32] feels like we're about to have an [63:34] absolute avalanche of peptides to choose [63:36] from. [63:36] >> On November of 2025, Lily cut a deal [63:39] with the Trump administration. I saw [63:40] this to drop the price on Drespatide [63:43] pretty significantly. I think it's like [63:44] 50 bucks on Medicare. [63:45] >> 50 bucks from Medicare. Yeah. [63:47] >> Yeah. Which is a pretty cheap price [63:48] point, but it starts to make sense as [63:50] you think about the portfolio of Lily [63:51] products. You get Tzepide for 50 bucks, [63:54] but if you want to upgrade, get the [63:56] Retatride. That's the high premium [63:57] product and that's where they're going [63:58] to start the Mercedes to the Honda. I'm [64:01] sure if I'm Lily and I'm sitting there [64:02] and I'm looking at this data coming out, [64:03] I'm like, "My god, people will pay for [64:05] this and that starts to become sort of [64:07] like the upgrade to the BMW or the Model [64:09] S plat if you will." [64:10] >> Yeah. The Trappetide is like the one [64:12] bedroomedroom messy bed hotel room [64:15] >> and the other one's the sweetide [64:18] is like the twobedroom suite. [64:20] >> Well, you can also, by the way, you guys [64:22] know I'm a spokesman for Row. They also [64:26] have the Waggoi pill uh row.co/twist [64:28] cotwist uh to get your [64:31] >> Wait, are you a paid talking about [64:33] [laughter] Are you a paid What are you [64:35] talking about? We're putting Charles [64:37] Barkley and [64:38] >> we're not having sales team [64:42] and then you come over on Allin and you [64:44] start promoting. [64:44] >> No, no, no, no. Trust me, we'll get one [64:46] of those as well. We'll get a row. [64:48] Sponsorship here. [64:49] >> What was the ro pill that you had me [64:50] get? What was it called? [64:51] >> Oh, sparks. Sparks. Sparks. Did you take [64:53] it? [64:54] >> I have taken it and now [64:56] >> Amore please. No, not [64:58] >> maybe just a half of a lousy. [65:00] >> I want to hear the story. I want to hear [65:01] the story. Go. [65:02] >> It's so out of control. [65:04] >> I told what I told you. [65:06] >> So then what happens is Nat and I are [65:08] like, you can't you can't just randomly [65:09] use it. It's scheduled. We discuss it. [65:11] We put it on the calendar. [65:12] >> You need a plan. [65:13] >> You need a plan. [65:14] >> You need a plan. Can't go in [65:16] >> because otherwise otherwise it's too [65:17] much. You just can't randomly take it. [65:19] >> What do you mean? [65:20] >> It's going to be a sesh. [65:22] >> It's a whole thing, man. It's like I [65:24] don't have the energy for that. [65:25] >> It's an extended session. You you have [65:26] to be well rested. This don't do this at [65:29] 1:00 a.m. This is like a 10 p.m. This is [65:31] like a No, this is more like a This is [65:33] more like on vacation, you know, like [65:35] >> 10:00 a.m. to 12:00 p.m., you know, to [65:37] noon, you know, you got to really [65:40] >> you got to really schedule it. [65:42] >> Schedule it because kids around [65:44] >> otherwise it's got to be empty. [65:46] >> Otherwise, it's unfair to her and it's [65:48] just a li it's a lie. Put it [laughter] [65:49] out. [65:50] >> It's a It's a big commitment literally. [65:52] >> You You look embarrassed, Jim. Do you [65:54] feel embarrassed talking about it? It's [65:56] just a lot, man. It's like It's a lot to [65:57] handle. It's a lot. [65:59] >> It's If you want to get the extra 20% in [66:02] your performance, [66:03] >> it's a lot, bro. [66:04] >> It's a [laughter] lot. It's basically [66:06] over time. [66:07] >> What happened? What happened was I was [66:08] like, "Oh, what is this thing?" Jason's [66:10] like, "Dude, you must get it. You must [66:11] get it." So, we got it. We tried it and [66:14] >> we were like, "What the was that?" And [66:17] so, [laughter] then I've been trying to [66:18] bleed the pills out. So, I gave some to [66:20] Stant Tang. I'm like, "Stanley, you try [66:21] it." Literally, he's dealing them like [66:24] cards [66:25] >> but when we're having poker dinner, I'm [66:27] like, does anybody want to try these [66:28] things? But these what is it? Rose [66:30] sparks. Is that rose sparks? Shout out [66:31] to my friends out. All right, let's keep [66:33] moving here. Freedberg, guys. Friedberg [66:36] had his own personal Super Bowl. You see [66:38] me getting ready for Nick's playoff [66:40] season. I get my courtside. Freeberg had [66:43] the equivalence acts. He went to the [66:45] Supreme Court in order to hear them talk [66:49] about chemicals. This was a big deal for [66:52] him. The Supreme Court coming together. [66:54] Did you wear it? Yes. The Monsanto trial [66:57] happened in the Supreme Court and he [66:58] went he got courtside. He went to the [67:01] Supreme Court and listened in the [67:02] building. Have you guys ever seen a live [67:04] Supreme Court hearing? [67:05] >> No. I'd love to. [67:06] >> I'd love to. Tax, have you been? [67:08] >> No, I haven't actually. [67:09] >> I mean, honest honestly, I think it was [67:11] one of the most amazing experiences I've [67:13] ever had. There was a massive protest [67:14] out front. We went through the [67:16] marshall's office to get in. And that [67:18] building, you walk in, it's like sacred. [67:20] It's all marble. It's you're not allowed [67:23] to talk. You have to be super quiet when [67:24] you're in the building. Like they keep [67:25] going like you're in some quiet library. [67:28] It's like people treat it with this [67:30] level of kind of sanctity that that and [67:32] respect. And they're like, there is no [67:34] politics here. There is no [67:36] There is no freedom of speech. This is [67:38] the court. When you come into this [67:40] court, the justices tell you how you [67:43] will speak, how you will behave, what [67:44] you will do, and you will not speak [67:46] unless spoken to. You put all your stuff [67:48] in a locker, you go up the stairs, you [67:50] go into the the courtroom. And the [67:52] courtroom, it's just so amazing being in [67:54] there. They have this amazing marble [67:55] freeze above the justices that has some [67:57] of the great people of human history, [67:59] Moses and these kind of amazing [68:01] historical figures. And then below them [68:03] are the nine justices and the court [68:05] case. If you guys haven't watched the [68:07] case, you can listen to them, I think, [68:08] online. on you. Is it worth listening [68:10] to? [68:10] >> Hold on. Wait, wait, wait. I have a [68:11] question. So, [68:12] >> does Robert sit in the middle cuz he's a [68:14] chief? Yes. [68:15] >> And then do all of the right justices [68:17] sit on the right? [68:18] >> No, they're mixed. So, they're I think I [68:20] don't know I don't know the exact [68:21] seating, but they're mixed in terms on [68:24] appointments of the court. Is [68:25] >> that right? I think that's right. And [68:27] then so yeah, that's right. And then it [68:29] kind of goes out from the middle with [68:30] Roberts in the middle. Roberts [68:32] occasionally will name the justices and [68:35] say, "Hey, do you have a question? Do [68:36] you have a question?" if no one's [68:37] talking, but otherwise the justices will [68:39] jump in with their questions when they [68:40] want and they'll ask. Now, honest to [68:42] God, watching this is like watching [68:44] LeBron James play basketball. These [68:46] lawyers are so mind-blowingly impressive [68:50] on both sides that you would just like [68:52] sit there and I was like in awe. It was [68:54] so I I felt like my energy was [68:57] completely sapped from me at the end of [68:58] this process because you were just so [69:00] engaged and so caught into the way that [69:02] these guys are thinking and talking. [69:04] >> Did you take a rose sparks? Did you take [69:05] a rose sparks when you were there? [69:06] >> No. And if you're familiar if you're [69:08] familiar enough with the case or the [69:10] case history or the law that's being [69:12] debated because again when when you get [69:13] to the Supreme Court, you never debate [69:14] the case. What you're debating is the [69:17] legal interpretation of the the [69:19] decisions that were made on the case. [69:21] And so is this constitutional? How do [69:23] you interpret this particular act, this [69:25] law, this federal law? What's the right [69:27] way to think about it? So you don't [69:28] actually talk about the case. You talk [69:29] about the interpretation of American [69:31] law, [69:33] of our laws, of our constitution, of the [69:36] global. [69:36] >> You're saying the facts have already [69:38] been determined. [69:38] >> That's right. [69:39] >> Right. At a lower court, there's [69:40] questions of fact and questions of law. [69:42] >> The facts have already been determined [69:44] by the lower court. It's just Supreme [69:45] Court is ruling on questions of law. [69:47] >> That's right. And so they have a full [69:49] briefing with the full history of the [69:50] case. And remember, they only hear two [69:52] cases a day. So they're one hour each [69:54] for each hearing. [69:55] >> So you go in and they only do it Monday, [69:56] Tuesday, Wednesday on the last two [69:58] weeks. and they only hear cases from [70:00] October to April. There's only a handful [70:02] of cases that are selected. [70:03] >> Wow. So you're really on a shock clock [70:04] then to make [70:05] >> you're on a shot clock and you only have [70:06] and it's 30 minutes aside and then the [70:08] justices will ask question. [70:09] >> So this was Monsanto and Roundup, right? [70:11] So what was the law that was being [70:13] debated [70:14] >> for yours? The regulatory body, the EPA [70:18] sets the label for pesticides. Does this [70:20] cause cancer or not? What are the [70:22] warnings? This can be damaging for birth [70:23] defects, pregnancy, all the things that [70:25] we're all used to seeing on labels. when [70:27] you buy a product, a chemical product [70:29] and the EPA and their regulatory [70:31] authority determined that Roundup does [70:34] not cause cancer. When you sell a [70:36] pesticide, you first have to register it [70:38] with the EPA, get it approved, and then [70:39] the EPA gives you a label. And the label [70:41] is written by the EPA. It says exactly [70:43] what you're supposed to say. And in this [70:45] case, it said all this stuff doesn't say [70:47] cancer because they determined it does [70:49] not cause cancer. And I'm not going to [70:51] debate whether or not it causes cancer, [70:52] but that's the case that was made is [70:54] that the EPA is the regulatory body [70:56] under a federal act called FIFRA, [70:59] fungicide, insecttoide or denicide act. [71:01] And that's where the EPA is given their [71:03] regulatory authority to put the label on [71:06] these products. And all of the cases [71:08] that have been lost have been state [71:11] failure to warn cases. To date, Bayer, [71:14] which now owns Monsanto, has paid out [71:16] $10 billion in these lawsuits, and they [71:19] have reserved 10 billion on their [71:21] balance sheet. They have 90,000 cases [71:22] still outstanding in the courts. 90,000. [71:25] >> Wow. [71:25] >> And so, this one case got kind of [71:27] appealed up to the Supreme Court. Last [71:30] year, the White House solicitor general, [71:32] and if the solicitor general steps up [71:33] and asks the Supreme Court to take a [71:35] case, it's more likely the case gets [71:37] taken. So, the White House said, "Please [71:38] take this case. We need to have federal [71:41] preeemption, meaning the federal [71:42] government has the right to set the [71:44] label because all of the cases that have [71:46] been lost and that are being adjudicated [71:49] are in state courts where the state has [71:51] a law like in California called a [71:53] failure to warn law, which means if a [71:55] manufacturer knows that a product [71:57] carries a risk, you have to warn the [71:59] consumer. And so the the lawyers have [72:02] been arguing that Monsanto or Bayer knew [72:04] that this product caused cancer and [72:06] didn't warn the consumer. And they've [72:08] been winning cases. they've been losing [72:09] cases, but they've won enough cases that [72:12] this has now become a multi-dea billion [72:13] dollar problem. And so the argument is [72:16] that the EPA says it doesn't cause [72:18] cancer and they have federal [72:19] preeemption. So the EPA has the right to [72:21] determine. So that's the one argument. [72:24] But then when the other attorney came [72:27] up, this guy was like literally like [72:28] watching LeBron James. And so going in, [72:30] we're like, "Oh, 63 Bayer's going to [72:31] win." And then the other guy comes up [72:33] and he was like, "Well, hey, you guys [72:36] overturned the Chevron doctrine last [72:37] year. You guys remember that case? [72:38] >> Yeah. Where basically when the Chevron [72:40] doctrine got overturned, it basically [72:42] said that no longer [72:44] >> does the federal agency get to decide it [72:46] has to be a direct reading of the law. [72:50] >> Duh. So now, so he's saying like the [72:52] states should have a right to read the [72:55] law themselves. They shouldn't have to [72:57] just defer to the EPA. And that's what [72:59] this will come down to. So at the end of [73:00] it, we were like, "Oh my god, this could [73:02] be a 50/50 coin flip, 54 either way." [73:05] And going into it, we were kind of like [73:06] trying to say, "Hey, maybe this could be [73:08] 63." So honestly, the whole experience [73:10] was incredible. The case is interesting. [73:12] >> These are very complicated matters. How [73:13] are these people able to make a [73:16] wholesome argument in like one side gets [73:19] 30 minutes, the other side gets 30 [73:20] minutes, there's a little Q&A, and then [73:21] you're done in an hour. [73:22] >> There's this whole art and science and [73:24] Sax, you're probably familiar with this [73:26] on how do you distill down a Supreme [73:27] Court case in the briefing dock? Like [73:29] what is it you're petitioning around the [73:31] court? and you try and distill it down [73:32] to the exact legal interpretation you [73:35] want the judges to rule on, not all the [73:37] other [73:38] >> And this is oral arguments. Yes. [73:39] >> Oral arguments, just a discussion. And [73:41] then the judges jump in and all they're [73:42] doing is asking the lawyer questions, [73:44] one lawyer at a time, the one side and [73:46] then the other side. And by the way, the [73:47] solicitor general came up in the middle [73:49] and kind of made a few comments and they [73:50] asked her some questions from the White [73:52] House and she sat down and then the two [73:54] sides kind of went back and forth and [73:55] and they just it's like 30 minutes Q&A [73:57] each on that one specific legal question [74:00] and Katanji Brown Jackson said, "But [74:03] what if after the EPA issued the label, [74:05] they found out information that it does [74:07] cause cancer? Shouldn't they update the [74:09] label?" And he's saying, "Well, no, [74:10] they're not allowed to. They can only [74:11] issue the label the EPA says." And he [74:13] says also and it's it's a criminal case [74:16] if they find out that it does cause [74:17] cancer and they don't report it to the [74:19] EPA and then she's saying well what if [74:21] the EPA doesn't act and shouldn't the [74:23] states have a right to protect their [74:24] people? So those are the legal [74:26] arguments, the discussions that are [74:27] going on in all of this. And there's [74:29] interesting implications which is [74:30] fundamentally if the states get to [74:32] interpret federal law and ignore federal [74:35] regulatory bodies, it opens up a whole [74:37] new can of worms in terms of like all [74:39] the states can start to ignore federal [74:41] regula regulatory bodies like the EPA or [74:44] the FDA or the USDA or and on and on and [74:46] on. So the whole case has a whole bunch [74:48] of really interesting implications wound [74:50] up in it. when you hear these guys and [74:51] they're just talking chimat about that [74:53] exact like interpretation of the law and [74:55] that's what this comes down to. It's not [74:56] the actual case that matters [74:58] >> and after the sachs they oral arguments [75:01] and then they have like a private [75:02] conference where they'll write their [75:04] papers and give their final judgment. [75:05] Yeah. [75:07] >> Saxs. [75:07] >> Yeah. Yeah, I think what happens is that [75:10] so I guess there's some discussion that [75:12] happens behind closed doors and they [75:14] figure out where the majority is and [75:16] then the chief gets to assign who writes [75:18] the opinion for the majority [75:19] >> in that meeting nobody is allowed in and [75:22] in fact you have a double door system [75:24] where like if anything needs to come in [75:25] and out you have to like kind of like [75:27] knock on the door you're led into this [75:28] anti chamber then [75:29] >> oh is it an airlock is an airlock [75:31] >> it's effectively we I had I don't know [75:34] if you were there Jason but we had Ted [75:35] Cruz come to play in the poker game [75:37] >> uh And Ted Cruz clerked for William [75:40] Ranquist and if you want to have an [75:42] incredible dinner, ask him about the [75:45] Supreme Court and Bill Ranquist. He's a [75:48] real student of the Supreme Court and it [75:50] just makes the Supreme Court free to [75:51] your point sound like the most [75:53] incredible body that's ever been created [75:57] anywhere. By the way, more than the [75:59] White House, more than the Capital [76:01] Building, more than any of these other [76:03] big agencies, this place has it's almost [76:06] like being in England, it has these kind [76:08] of ways that people operate. The the the [76:10] security is so different. They kind of [76:12] stand there in the court and they all [76:14] exchange places every 20 minutes. It's [76:16] very coordinated. They're dressed very [76:18] differently than any other [76:19] >> courtroom listening. [76:20] >> Maybe like 150, I would say. How do you [76:23] take it? Are they on [76:24] >> I actually think everyone is a guest of [76:26] a clerk or someone that works at the [76:27] court. I don't think that it's like very [76:29] publicly available to get in there. [76:30] >> You can't line up. There's no lineup. [76:32] >> There's I don't know if there's a [76:33] lineup. Um this was a connection through [76:36] we got in through the chief justice. Um [76:38] he gave us the pass, but I think it was [76:40] like very um [76:42] >> I think at the Elon versus Open AI case [76:45] there's you can line up and then the [76:47] judge gave like 30 tickets to the press [76:50] court. No, no. Yeah, but I think there's [76:52] a lineup for the Supreme Court as well. [76:54] There's some public access that they're [76:55] >> It did not look like anyone from the [76:57] public was in this court. Everyone, [76:59] everyone is dressed respectfully. I [77:00] mean, this court has an incredible [77:02] amount of like, [77:04] >> you know, cool experience. [77:07] >> I would I would just say uh enjoy it [77:09] while you can. I mean, I think the [77:11] Supreme Court is one of the last highly [77:12] functional institutions in the United [77:14] States. And% [77:16] >> you know at some point we're going to [77:17] have like 13 or 21 or some crazy number [77:20] of justices up there [77:23] and get jersey [77:25] after justices there and so enjoy it [77:28] while it's still [77:29] >> in the current in the current form it's [77:31] in. [77:32] >> Can you imagine showing up with jerseys [77:34] with the justices names on them and like [77:36] having sections and like somebody [77:38] selling cracker jacks [77:40] >> theocracy version of [laughter] the [77:42] Supreme Court [77:43] >> version. Exactly. [77:45] >> The popularity of the court really [77:47] depends on whether it's issuing [77:51] decisions that people agree with. That's [77:52] what it comes down to. If like if you [77:54] ask people whether they like the Supreme [77:56] Court or not, it really just depends on [77:58] whether they agree with the decisions [77:59] are recency as opposed to [78:02] >> the process of the decisions and how [78:04] well argued it is and all these things [78:05] that you're pointing to. And actually [78:08] the the court I mean I just checked the [78:09] numbers. The court is relatively popular [78:12] right now. I think that it got as low as [78:16] 35% in the 2024 Gallup survey, but I [78:19] think it's back up to, you know, like 44 [78:23] to 50% favorability, which for something [78:26] that's involved in politics is [78:28] relatively high, right? Like you look at [78:30] Congress or [78:31] >> any particular politician, they're going [78:33] to be lower than that typically. I just [78:35] felt so assured of like the institution [78:40] when I visited and saw these guys [78:41] interact and behave and how they behave [78:43] the process. It was like [78:45] >> man this what an amazing country. Yeah. [78:48] >> Well, the reason I say what I say is [78:49] there was an interview with James [78:50] Carville recently. Did you guys see [78:52] this? He saidaw he said look when we get [78:54] power we're packing the court. [78:56] >> So we're not even going to we're not [78:57] going to worry about it. [78:58] >> And we're going to get to 13, right? He [79:00] said we're going to make [79:01] >> I think Yeah. They're going to go from 9 [79:02] to 13 and then they're going to create [79:04] some new states and all the rest of it. [79:06] So that'll be that. [79:07] >> Uh [79:08] >> enjoy enjoy while it last. Enjoy. [79:10] >> Enjoy while it lasts. [79:11] >> Uh by the way, [79:12] >> end on a high note. [79:13] >> Wait. Yeah. [laughter] It's the end of [79:14] the empire. That'll be that. [79:16] >> By the way, there is uh a Supreme Court [79:19] on I was correct. There is an online [79:21] ticketing lottery. So we can all sign up [79:24] and you can get a fourack of tickets. I [79:26] think they should make this I we should [79:28] talk to Howard Lutnik. Maybe he can make [79:30] this an auction. We get a revenue stream [79:31] from the US. We could sell like 10 of [79:33] the tickets as courtside seats for 20 [79:35] grand. [79:35] >> Jason, you're exactly what they're [79:37] trying to protect against. [79:38] >> Exactly. [laughter] Like, how can we how [79:40] do we monetize the Supreme Court? [79:43] >> All right, everybody. That's it. That's [79:44] the world's greatest podcast for you for [79:46] Chimoth Poly Hatia, David Freeberg, and [79:48] David Saxs. I am the world's greatest [79:51] moderator. We'll see you chief justice. [79:54] >> I'm like the chief justice of [laughter] [79:56] the allin podcast. [79:58] >> [music] [79:59] >> We'll let your winners ride. [80:01] >> Rainman David [80:06] and it said we open [music] sourced it [80:07] to the fans and they've just gone crazy [80:09] with it. [80:10] >> Love you queen of [80:13] winners. [80:15] [music] [80:19] >> Besties are gone. [80:21] >> That is my dog taking a [music] [80:23] driveway. [80:26] >> Oh man. My habitasher will meet. [music] [80:29] >> We should all just get a room and just [80:30] have one big huge orgy cuz they're all [80:32] just useless. It's like this like sexual [80:34] tension that you just need to release [80:35] somehow. [80:40] >> Your feet. [80:42] [laughter] We need to get Mercury's [80:44] already. [80:50] [music] [80:52] I'm going all in.