Stripe, Block, and Advent Lodge $53 Billion Takeover Bid for PayPal

Stripe、Block与Advent拟530亿美元收购PayPal 这项拟议中的巨型并购旨在通过整合Stripe的商户资源与PayPal庞大的消费者生态系统，彻底绕开传统信用卡网络。 主持人剖析了DeepMind关于建立由行业资助、类似FINRA模式的AI监管机构的提议，并将其与他们所认为的Anthropic反竞争监管俘获策略进行了对比。他们分析了Stripe、Block和Advent潜在收购PayPal所带来的巨大财务协同效应，以及Apple针对OpenAI提起的激进商业机密诉讼。最后，讨论还涉及了超大规模数据中心引发的全球能源竞赛、将公众焦虑武器化的外国影响力行动，以及一种由AlphaFold设计的、能够逆转人类皮肤组织生物衰老的革命性酶。 Google DeepMind 首席执行官 Demis Hassabis 提议建立类似 FINRA 的 AI 行业自律组织（SRO）；David Sacks 随后提出五项必要条件，主张将监管严格限制在前沿模型与极端灾难性风险范围内，防止形成行业卡特尔与监管俘获。 (1:32) Stripe 联合 Block 与私募股权公司 Advent 提出以 530 亿美元联合收购 PayPal，意图将商户端 API、超 4 亿个人用户与稳定币网络打通，构建足以抗衡 Visa 和 Mastercard 的新一代端到端支付通道。 (20:00) Apple 对 OpenAI 提起长达 41 页的诉讼，指控其在过去两年挖角超过 400 名 Apple 员工，并在开发自研 AI 硬件过程中严重窃取商业机密与硬件设计知识产权。 (37:51) xAI 的编程工具 Grok Build 因漏洞在未获授权的情况下将开发者包含敏感密钥的完整私有代码库静默上传至服务器，事件曝光并被紧急修复后，再度引发对 AI 数据隐私脆弱性及企业 Token 成本失控的广泛担忧。 (42:49)

The All-In Podcast hosts argue that the American AI ecosystem faces stagnation from corporate regulatory capture, energy shortages, and state-level moratoriums, even as falling inference costs and international rivals like China challenge US dominance. Concurrently, shifting regulatory expectations are reviving mega-mergers—exemplified by a $53 billion bid for PayPal—aimed at re-engineering mature digital networks with AI infrastructure. Self-Regulatory AI Oversight Versus Regulatory Capture (1:32) DeepMind CEO Demis Hassabis proposed an industry-funded, federally overseen self-regulatory organization (SRO) modeled on FINRA, under which frontier AI labs submit models 30 days prior to release to assess cybersecurity and chemical, biological, radiological, or nuclear (CBRN) threats. While tech leaders across Google, Meta, and OpenAI signaled support, David Sacks argues that an SRO must meet five conditions: broad representation including open-source developers, a focus restricted strictly to genuine frontier step-changes, review limited to catastrophic threats rather than speech, a voluntary proving phase, and complete substitution for new federal agencies. Sacks contrasts this with Anthropic CEO Dario Amodei's call for an "FAA for AI," cautioning that multi-year review cycles would surrender American competitiveness to China. The hosts accuse incumbents like Anthropic of engineering regulatory capture by promoting state-by-state measures like California's SB53 and funding activist groups such as Public First. Sacks and Chamath Palihapitiya contend that entrenched labs leverage catastrophic fear-mongering to establish compliance barriers that freeze out smaller competitors, warning that premature interventions risk strangling innovation over unproven harms. The Stripe-PayPal Bid and AI-Driven Consolidation (20:00) Stripe, private equity firm Advent, and Jack Dorsey’s Block launched a joint $53 billion bid ($60 to $70 per share) to acquire PayPal, whose valuation had declined from a peak of $322 billion to the mid-$30 billions. Block contributed $17 billion in equity. The merger seeks to combine Stripe’s merchant APIs (processing $2 trillion annually) and Block’s Cash App and point-of-sale network with PayPal’s 439 million consumer accounts, Venmo, and PYUSD stablecoin. Palihapitiya and Sacks argue the asset combination is designed to bypass Visa and Mastercard through direct internal "on-us" settlement, which they argue could survive antitrust scrutiny if defined as pro-competitive against the established credit card duopoly. David Friedberg and Jason Calacanis frame this transaction as an emerging "AI-ification" private equity playbook, comparing it to Bending Spoons' acquisitions of Evernote and Vimeo. AI-native operators are partnering with buyout firms to purchase cash-generative, mature digital platforms and extract operational margins using automation. The panel notes that corporate M&A is broadly rebounding following expectations of reduced antitrust resistance after the departure of FTC Chair Lina Khan. Energy Bottlenecks, Data Center Moratoriums, and Geopolitics (5:53) Palihapitiya outlines severe domestic grid bottlenecks, forecasting a US power deficit equal to 2.5 times California's consumption by 2050, underscored by a recent PJM Interconnection auction securing only 156 megawatts against an 8-gigawatt capacity requirement. To bypass grid backlogs and clean-air permitting reviews, operators are deploying behind-the-meter generation, such as mobile natural gas turbines deployed for xAI's Memphis cluster or on-site Bloom Energy installations. Compounding these constraints, New York Governor Kathy Hochul enacted a statewide moratorium on hyperscale data centers over environmental and utility cost concerns. Sacks refutes these arguments, stating that modern facilities use closed-loop water recirculation and that upstate New York remains roughly 80% undeveloped. Palihapitiya reports that regulatory friction has caused approximately 40% of planned data center projects to be canceled or stalled, shifting capital toward Texas, Australia, and the UAE. Citing an OpenAI report, Sacks notes that Chinese influence operations are actively amplifying anti-data center protests in the US to handicap American infrastructure. Diverging Inference Economics and the Shift to Edge Compute (45:53) Corporate token consumption surged 21-fold year-over-year among Ramp customers, prompting CFOs to implement tighter oversight against open-ended model spending. Palihapitiya highlights a growing pricing divide: proprietary closed models like Fable ($56 per million input tokens) and Claude ($26) face intense price competition from Grok and Llama ($1.00 to $1.50) and Chinese models ($0.50). Startups like Mira Murati’s Thinking Machines are targeting this spread by fine-tuning open-source models to match proprietary frontier performance at a fraction of the operating cost. Concurrently, enterprise data security concerns—highlighted by an xAI Grok Build bug that uploaded developers' private codebases and API keys—are accelerating a transition toward decentralized edge compute. Calacanis points to Apple's forthcoming M7 Ultra chip, projected to feature 1.5 terabytes of unified RAM, which would enable standard desktop hardware to run frontier-scale models locally. The hosts argue that executing workloads on local hardware or residential battery microgrids will substantially reduce reliance on expensive cloud APIs. AI Enzyme Design and Extracellular Matrix Restoration (82:57) Friedberg details research published by Alphabet subsidiary Calico and Revel Pharma targeting advanced glycation end-products (AGEs) in the extracellular matrix. Specifically addressing carboxymethyllysine (CML)—a sugar-fat protein bond that degrades collagen, stiffens tissue, and triggers immune inflammation—researchers used DeepMind’s AlphaFold coupled with five cycles of directed evolution to engineer a synthetic enzyme capable of breaking down CML. In laboratory evaluations, the custom enzyme degraded 52% to 97% of target CML across biological proteins, and eliminated 55% of CML cross-links in skin tissue donated by patients over age 70, restoring matrix elasticity to that of a 31-year-old. While medical applications could address systemic inflammation and joint mobility, the panel anticipates commercial delivery will first target the estimated $2 trillion cosmetic anti-aging market. The hosts caution that the US competitive lead may have narrowed to months as Chinese developers deploy near-frontier models like Kimi K3 unencumbered by domestic regulatory friction. They warn that expanding bureaucratic oversight, local infrastructure moratoriums, and closed-stack lobbying threaten to surrender American market leadership just as fundamental breakthroughs in compute economics and biological engineering begin to scale.

All-In 播客本期聚焦于人工智能监管博弈、地缘竞争与前沿产业变局。主持人群体认为，基于虚构风险的过度监管与地方数据中心禁令正危及美国的创新优势；与此同时，市场正通过自律组织构想、支付巨头重组、推理成本断崖下跌及边缘算力崛起寻求效率突破。 AI自律监管构想与监管俘获争议 (1:32) Google DeepMind 首席执行官 Demis Hassabis 提议建立类似金融业监管局（FINRA）的由行业资助、专家主导的自律组织（SRO），要求前沿模型在发布前 30 天提交安全评估。David Sacks 表示支持但提出五项先决条件：必须吸纳开源与初创代表以防寡头垄断、仅针对真正实现阶跃提升的前沿模型、审查严格限于网络与化生放核（CBRN）等极端风险而不得干涉言论、先推行自愿原则、且必须彻底取代而非叠加新的联邦官僚机构。他坚决反对类似联邦航空局（FAA）的长周期认证机制，警告软件若面临数年审批将直接输掉技术竞争。 Sacks 进一步指责 Anthropic 借安全之名推行“监管俘获”。他援引《Politico》关于 Anthropic 推动各州收紧规章的报道，批评其利用雄厚资金通过州级立法构筑合规壁垒，实质是在阻碍初创企业发展。Chamath Palihapitiya 赞同尽快落地自律框架，强调行业必须划定底线并争取联邦优先适用权，以阻止传统利益集团斥巨资游说并拉高合规门槛。 Stripe竞购PayPal与并购复苏 (20:00) Stripe 联手私募股权机构 Advent 提出以约 530 亿美元收购 PayPal，Jack Dorsey 旗下 Block 亦拟出资 170 亿美元股权参团。尽管 PayPal 市值已从峰值约 3220 亿美元大幅跌落且年增速放缓至 7%，但仍保有逾 4 亿活跃账户、Venmo 及稳定币 PYUSD。Chamath 分析指出，Stripe 掌控深厚的商户 API 资源（年处理量约 2 万亿美元，高于 PayPal 的 1.7 万亿美元），若与 PayPal 的个人账户及 Block 的 POS 终端结合，可通过内部链路（on-us）直接绕开 Visa 和 Mastercard 双寡头。 David Friedberg 指出该交易代表了“AI 赋能型”私募重组趋势，即由顶级 AI 原生团队接管陷入停滞的老牌数字资产，利用算法优化运营并重塑老旧体验。此外，主持人群体认为，伴随特朗普胜选以及联邦贸易委员会（FTC）强硬反垄断周期的消退，受阻多年的企业并购迅速回暖，带动风险投资资金流动性显著复苏。 Token成本暴跌与边缘算力演化 (45:53) 生成式 AI 推理正在经历剧烈的微观成本重塑。Chamath 与 Sacks 指出，以 Fable 为代表的闭源前沿模型百万输入 Token 成本高达约 56 美元，而 Grok、Llama 等开源或开放模型降至 1 至 1.5 美元，中国前沿模型更压低至 0.5 美元。企业管理平台 Ramp 的数据显示，其客户的 Token 支出在一年内激增 21 倍，迫使财务团队建立严格的使用追踪机制以防利润被无序算力吞噬；同时，基于次前沿模型微调的垂直方案也正加速对高价闭源生态的替代。 在硬件端，算力正显著向本地设备与分布式能源倾斜。Jason Calacanis 援引彭博社报道称，苹果下一代 M7 Ultra 芯片预计支持高达 1.5 TB 统一内存，允许个人终端本地运行超大前沿模型，有望将企业 90% 以上的工作负载脱离云端 API。与此同时，Sunrun 和 SPAN 与英伟达合作推进的家庭数据中心模式，正在借助分布式光伏与储能直供电，以去中心化形态突破集中式电网约束。 能源赤字数据中心禁令与地缘博弈 (64:39) 算力扩张正遭遇严峻的能源短缺与政治阻力。Chamath 指出，到 2050 年美国电力缺口预计相当于加州用电量的 2.5 倍，近期 PJM 电网拍卖在需要 7 至 8 吉瓦容量时仅获 156 兆瓦供应。为避开繁复审批，企业纷纷采用“表后”（behind-the-meter）自建发电。然而，纽约州州长 Kathy Hochul 签署了全美首个超大规模数据中心禁令，称其推高电价、消耗水源并破坏农地。Sacks 驳斥这一举措为政治推诿，指出闭式循环数据中心耗水极少且纽约上州土地辽阔，禁令实质将使当地基建搁置至少五年。 与会嘉宾认为反基础设施运动具有深层地缘背景。Friedberg 指出抗议资金链条与当年的反水力压裂运动如出一辙，Sacks 则援引 OpenAI 官方报告指出，中国关联影响力组织正系统性放大美国对数据中心的抵制舆论。Sacks 警告，在美方因繁苛监管自缚手脚之际，中国 Kimi K3 等模型已逼近全球前沿，双方代差恐已收窄至数月以内，澳大利亚与中东等地区正积极吸纳全球算力转移。 AlphaFold助力抗衰老工程突破 (82:57) 在科技前沿板块，Friedberg 介绍了 Google 旗下 Calico 与 Revel Pharma 联合发布的一项抗衰老研究。研究聚焦于细胞外基质（ECM）中因糖化形成的损伤分子羧甲基赖氨酸（CML），该物质长期积累会导致胶原蛋白交联黏连，诱发皱纹、关节僵硬与慢性炎症。 科研团队借助 DeepMind 的 AlphaFold 逆向设计出一种自然界中不存在的合成酶，并通过五轮定向进化完成活性优化。实验显示，该酶在常规生物蛋白中实现了 52% 至 97% 的 CML 降解率；在 70 岁以上老年人的人体皮肤组织实验中，该酶成功清除了 55% 的 CML，使组织微观结构恢复至相当于 31 岁的生理状态。嘉宾们认为，此项研究标志着 AI 与高通量筛选深度融合的里程碑，不仅证明了 AI 赋能人类健康的切实价值，更孕育着达 2 万亿美元的抗衰老商业蓝海。 本期节目总结指出，人工智能正面临技术降本与基础设施管制的双重拉锯：底层推理成本暴跌与跨界应用展现出庞大生命力，但电力短缺与政客制造的监管壁垒构成了实质威胁。主持人们强调，若决策层放任道德恐慌转化为严苛的官僚审批或地方基建禁令，将严重削弱美国以风险资本和自由创新为基石的核心优势，甚至在地缘博弈中将领先主导权拱手让人。

## Background 本期内容由知名科技与投资播客《All-In Podcast》于 2026 年 7 月 27 日发布在 YouTube 平台。节目简介涵盖了 DeepMind CEO 戴米斯·哈萨比斯提出的类 FINRA 式 AI 自律监管框架、Stripe 联合 Block 与 Advent 拟以 530 亿美元收购 PayPal 的世纪并购，以及苹果起诉 OpenAI 硬件侵权、纽约州出台全美首个数据中心暂停令、Token 支出暴增与利用 AlphaFold 逆转衰老等多项前沿焦点议题。 ## TL;DR 四位合伙人首先深入拆解了 Demis Hassabis 提出的 AI 自律组织（SRO）构想，警惕官僚体系与部分巨头借“末日恐慌”推行监管俘获，并结合纽约州数据中心暂停令痛批反技术政策背后的能源短缺与现代认知战。在商业与资本战场，嘉宾们深度剖析了 Stripe、Block 与 Advent 联手竞购 PayPal 以颠覆传统卡组织的支付闭环逻辑，以及苹果起诉 OpenAI 窃取硬件机密的行业震撼弹，同时指出了企业 Token 支出飙升下的开源平替与端侧计算趋势。最后，“科学角”带来了震撼科研盘点，揭秘研究团队如何通过 AlphaFold 与定向进化设计出全新工程酶，成功将 70 岁以上人体皮肤的关键分子指标逆转至 31 岁状态，生动展现了 AI 在生命科学领域的颠覆性价值。 ## Atmosphere & style 整场节目贯穿着“四位好兄弟”（Besties）极具辨识度的叙事节奏，在开场互相调侃、互揭老底的戏谑松弛感中拉开序幕，随后迅速切换为针锋相对、火花四溅的高密度思想碰撞。David Sacks 与 Chamath Palihapitiya 的论述犀利老道，以行业内部操盘手视角对华盛顿官僚主义和巨头伪善展开毫不留情的辛辣解构；而在压轴的“科学角”板块，硬核科研拆解将全场氛围从现实商战的戾气带入对生命科学奇迹的由衷震撼与浪漫憧憬中，在机智幽默、智力激荡与浓厚的同侪默契中达到了高潮。 ## Bulletin - [01:32](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=92s) DeepMind CEO Demis Hassabis 提议建立类似金融业 FINRA 的 AI 行业自律组织（SRO），主张由业界出资、独立技术专家运营，模型公开发布前 30 天提交高危审查，获得马斯克、奥尔特曼等多位科技领袖支持。 - [20:01](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=1201s) Stripe 联合私募巨头 Advent 及出资 170 亿美元的 Block，以 530 亿美元要约竞购 PayPal，意在整合 6 至 7 亿账户与稳定币基建，绕开传统卡组织打造全新底层支付闭环。 - [37:51](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=2271s) 苹果提交 41 页诉状正式起诉 OpenAI，指控其通过高薪挖走逾 400 名员工，并在面试中索取实体零件以窃取消费级硬件商业机密。 - [59:53](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=3593s) 纽约州成为全美首个对超大规模数据中心实施暂停令（moratorium）的州，主持人群体驳斥所谓高耗水、占地与推高电费的指责，指出这实质上是受境外舆论与职业抗议操纵的‘奢侈型监管’。 - [01:24:30](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=5070s) Calico 与 Revel 利用 AlphaFold 结构预测及定向进化设计出全新工程酶，成功分解不可逆的 CML 衰老蛋白，在离体实验中将 70 岁以上人体皮肤指标逆转至 31 岁水平。 ## Key takeaways - [01:32](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=92s) Demis Hassabis 提出由行业出资、独立技术专家运营的类 FINRA AI 自律组织，核心聚焦网络与生物安全等边际灾难性风险，推行前沿大模型发布前 30 天审查机制。 - [09:43](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=583s) David Sacks 强调自律组织必须严守五大底线：不能限制开源与初创、仅限 CBRN 与网络灾难风险、渐进式门槛、先自愿后法定，且必须获得联邦预先豁免权以彻底取代各州碎片化监管。 - [15:58](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=958s) Anthropic 被指责推行全美各州‘层层加码’的游说策略，利用公众末日恐慌诱导政府建立类似 FAA 的事前繁苛许可制度，实质上是老牌寡头消灭开源与中小型竞争对手的监管俘获阳谋。 - [22:42](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=1362s) Stripe、Block 与 Advent 联手竞购 PayPal，代表着‘AI 原生代’资本利用顶尖执行力改造‘数字原生代’老化资产的新并购范式，并通过稳定币打造正面对决 Visa 和万事达的自建清算网络。 - [28:54](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=1734s) 随着 FTC 监管高压松动以及政治风向转变，企业并购（M&A）全面复苏，资本市场对创始人离场的 Web 2.0 遗留资产展开高杠杆重组与 AI 化降本增效。 - [39:50](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=2390s) 苹果就消费级硬件机密外泄正式起诉 OpenAI，凸显出硅谷高管跳槽的合规红线——个人认知与记忆可以带走，但任何硬件实体、内部云端权限及专有代码绝不能碰。 - [42:49](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=2569s) Grok Build 违背零数据留存（ZDR）承诺并在后台误传用户完整代码库的漏洞表明，大模型厂商单方面的口头安全承诺不可信，必须依赖中立的第三方独立架构构建数据防火墙。 - [45:53](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=2753s) 闭源模型每百万 Token 高达 26 至 56 美元的溢价与开源/中国模型低至 0.5 美元的成本形成巨大鸿沟，企业因 Token 支出失控面临严峻的财报利润压力，倒逼其转向微调开源与私有化方案。 - [51:28](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=3088s) 具备超大统一内存的苹果 M7 Ultra 等端侧芯片有望在本地顺畅运行顶尖模型，使企业绝大多数推理负载转向桌面端零边际成本运行，对高价云端闭源模型形成致命的降价冲击。 - [56:06](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=3366s) 全美正面临严峻的电力供需赤字，传统电网审批迟缓导致数据中心项目大面积搁浅，迫使马斯克等先行者转向收购燃气轮机自建‘表后发电’以规避公共电网瓶颈。 - [59:53](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=3593s) 纽约州出台的数据中心禁令是典型的‘奢侈型监管’与政治作秀，不仅指控的数据中心耗水、占地与推高电费完全违背事实，更将导致纽约州未来五年错失宝贵的算力基建红利。 - [01:14:44](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=4484s) 美国当前反数据中心狂潮与早年反转基因和反核电的认知战剧本如出一辙，其背后存在受境外势力支持的虚假影响力行动及职业抗议者，旨在通过挑起内耗诱使美国在 AGI 竞赛中自废武功。 - [01:17:08](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=4628s) 大模型行业正出现恐慌营销闭环：部分巨头通过夸大末日风险将自身塑造成‘唯一合规公司’，借此游说立法打压开源竞品，但实际上早先设定的算力毁灭红线从未带来实质性灾难。 - [01:24:30](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=5070s) Calico 与 Revel 联合攻克细胞外基质（ECM）衰老机制，利用 AlphaFold 设计的工程降解酶高效消除了 55% 的 CML 糖化交联物，将人体皮肤组织逆转至 31 岁指标，展现了 AI 造福实体生命的巨大商业与医学价值。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=0s) 开场寒暄：法国历史旧闻与主持人的互怼玩笑 节目一开场，主持人杰森·卡拉卡尼斯（Jason Calakanis）照例以“全球最伟大播客”和“最伟大主持人”自居拉开序幕，查马斯·帕利哈皮蒂亚（Chamath Palihapitiya）立刻接茬调侃，搬出一段 1719 年的法国历史来挖苦杰森。查马斯称，当时巴黎监狱给犯人提出的特赦条件是：只要同意娶一名妓女并搬迁到路易斯安那州，就可以换取自由。查马斯借此揶揄杰森有某种“特殊血统倾向”，杰森则笑着反唇相讥，说自己是希腊后裔，家族可从没蹲过大牢。随后戴维·弗里德伯格（David Friedberg）回归节目，大家聊起上一期布拉德·格斯特纳（Brad Gerstner）代班时给特朗普账户开的“胜利巡游”，现场气氛轻松活跃。 ### [01:32](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=92s) 戴米斯·哈萨比斯提出类 FINRA 的 AI 自律监管框架 寒暄过后，话题迅速切入重磅科技新闻：DeepMind 联合创始人兼 CEO 戴米斯·哈萨比斯（Demis Hassabis）在 X 平台上发表长文，提出建立一个由美国主导的国际 AI 标准机构。这一提案在科技圈引发广泛反响，其核心蓝本是借鉴金融业的自律组织——金融业监管局（FINRA）。该机构将由联邦政府提供监督，但资金由产业界自行承担，并由独立的外部技术专家来实际运营。 按照戴米斯的构想，前沿大模型实验室在发布前 30 天需将模型提交审查。该机制初期实行自愿申报，未来将转为强制执行。评估的核心维度聚焦于网络安全风险、国家安全、生物武器威胁以及其他高风险领域；相关测试基准将按季度动态更新；如果出现严重的网络或生物安全警报，该机构有权协调全行业暂缓研发进度。目前，科技界多位重量级人物均表态支持：埃隆·马斯克（Elon Musk）称赞该方案“非常具有深思熟虑”（thoughtful），OpenAI 的萨姆·奥尔特曼（Sam Altman）、Anthropic 的联合创始人杰克·克拉克（Jack Clark）、谷歌的桑达尔·皮查伊（Sundar Pichai）、微软的萨蒂亚·纳德拉（Satya Nadella）、Block 的杰克·多西（Jack Dorsey）以及卡尔森兄弟等人均表达了积极态度。 杰森在节目中甚至放出了自己此前在播客上的预言片段，强调自己早就呼吁过“整个行业必须自律监管，由谷歌、微软、亚马逊和马斯克共同制定测试标准并自我认证，而不是让完全不懂技术的政府插手”，并打趣道自己就像命中了一记超远半场三分球。 - **机构性质**：联邦监督（Federal oversight）、行业出资（Industry funded）、独立技术专家运营（Run by independent tech experts）。 - **时间窗口**：前沿模型需在公开发布前 30 天提交测评，先自愿后强制。 - **审查重点**：网络安全（Cybersecurity）、国家安全（National security）、生物威胁（Biological threats）等高危领域，评估基准每季度更新一次。 ### [04:18](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=258s) 弗里德伯格解析 SRO 机制：金融监管映射与避免官僚低效 戴维·弗里德伯格进一步对该提案的核心机制——自律组织（SRO, Self-Regulatory Organization）进行了专业视角的拆解。他指出，SRO 在金融衍生品和证券领域早有成熟实践，例如 FINRA 和美国全国期货协会（NFA）。这些机构的设立初衷，是因为金融机构彼此交易、相互承担风险，全行业既不希望出现系统性风险敞口，又承受不起被外部行政官僚机构拖慢交易效率的代价。 ```mermaid flowchart TD accTitle: AI Self-Regulatory Organization Model accDescr: Flowchart illustrating the structure and operation of Demis Hassabis FINRA-style AI SRO proposal. subgraph GovOversight["政府宏观监督"] A["美国国会参众两院委员会"] --> B["联邦监督机构"] end subgraph SROStructure["AI 自律组织 (SRO)"] B -.->|"监督而非直接控制"| C["独立专家评估委员会"] D["产业界共同出资"] --> C end subgraph IndustryWork["大模型审查与发布"] E["前沿大模型 (30天前提交)"] --> F{"风险评估<br/>网络 / 生物 / 国安"} C --> F F -->|"达标"| G["准予推向市场"] F -->|"高风险"| H["协调暂缓发布 / 调整"] end ``` 弗里德伯格强调，AI 产业与金融市场高度契合。业内玩家都在竞速推进技术，最忌讳的就是政府空降一套僵化生硬的标准。他举例提到，加州大约一年半前试图强行通过 AI 监管法案，但当时起草的法条条款不仅在当时就脱离实际，仅过了一年，那些规则就已经完全无法匹配当下的技术现状。相比之下，SRO 机制的优势在于： 1. **敏捷迭代评估基准**：由懂软件评测的独立技术专家负责执行压力测试，能够以软件工程的速度动态调整评测维度（如生物危害、网络攻击、虚假社交操控等）。 2. **监管机制权责明确**：SRO 拥有独立专家组成的投票决策层，接受国会参众两院委员会的最终监督，既保证了专业性与效率，又确保了公共安全问责。 3. **避免行政直接管控**：正如弗里德伯格所言，全行业之所以迅速达成共识，是因为它“避免了直接将技术生杀大权拱手交给政府，而是让专业的人做专业的事”。 ### [07:01](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=421s) 萨克斯与戴米斯的直接沟通：警惕“AI 车辆管理局”与准入条件 戴维·萨克斯（David Sacks）透露，他此前就该提案与戴米斯·哈萨比斯进行了深入的私下沟通。萨克斯明确表示，以个人身份而言，他非常支持自律组织方向，因为如果任由政府直接设立全新的官僚机构，AI 产业将面临毁灭性打击。Anthropic 的达里奥·阿莫代伊（Dario Amodei）曾将政府机构构想为“针对 AI 的联邦航空局（FAA for AI）”，但在萨克斯看来，政府根本不具备评估大模型的认知能力，一旦介入，它会迅速蜕变成为低效臃肿的“AI 车辆管理局”（DMV for AI）。所有模型都将被迫排大队等待审核，从拖延一个月迅速恶化为推迟数月，最终导致美国在国际 AI 竞赛中彻底丧失领先地位。 为了防止自律组织变质，萨克斯向戴米斯明确提出了两项至关重要的前提条件： - **防止监管俘获（Regulatory Capture），吸纳多元代表**：SRO 绝不能变成三大顶级实验室（如 Google、OpenAI、Anthropic）联手垄断的排他小圈子。管理架构中必须拥有广泛的初创企业与开源生态代表。萨克斯建议吸纳黄仁勋（Jensen Huang）、埃隆·马斯克、马克·扎克伯格（Mark Zuckerberg），以及刚携开源新架构创业的前 OpenAI CTO 米拉·穆拉蒂（Mira Murati），通过利益制衡避免政策被头部寡头操纵。 - **仅审查真正突破现状的前沿大模型（True Frontier Models）**：监管范围必须严格限定在真正刷新人类当前最高智能基准（State-of-the-Art）的前沿突破性模型上。对于次级模型或未发生跨代台阶式跃升（Step change）的技术，绝不能借评估之名设置准入壁垒，否则行业巨头必然会利用繁复的测试流程拖垮竞争对手。杰森对此追问是否指基准跑分排名前 10 或前 20 的模型，萨克斯纠正指出，不能按数量排名，而必须看它是否在智能的核心维度上实现了超越当前最新水平的阶跃突破。 ### [09:43](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=583s) 自律机构的五项底线与“AI 版 FAA”的致命陷阱 David Sacks 继续阐述对于 Demis Hassabis 所提出的自律机构（SRO）设想的具体主张。他强调，这种监管必须精准聚焦在“重大突破引发的边际灾难性风险”，决不能演变成泛滥的官僚枷锁。Sacks 明确梳理了支持该方案的五大关键原则： 1. **不能限制开源与初创企业**：必须确保开源生态与小团队的创新空间，不能让合规成为大厂专属的烧钱游戏。 2. **渐进式门槛**：合规要求应随着智能水平的突破阶梯式上升，而非一刀切。 3. **严格限定于灾难性风险**：范围仅限于网络攻击（Cyber）和化学、生物、放射性与核武器（CBRN）威胁，坚决排除“虚假信息”或“微冒犯”（Microaggressions）等言论监管内容。 4. **先自愿后法定**：机构成立初期应采取自愿参与制，在证明自身切实有效后，方可考虑通过法律强制执行。 5. **替代而非叠加现有监管**：该自律机构必须替代可能新增的联邦监管部门；如果只是在现有官僚体系上“叠床架屋”，则毫无支持价值。 ```mermaid flowchart TD accTitle: SRO vs FAA AI Regulation Comparison accDescr: 比较Demis提出的自律机构SRO模式与Dario倡导的FAA式许可监管模式的差异与风险链条 subgraph SRO["Demis SRO 方案（五大底线）"] A["仅针对 CBRN 与网络灾难风险"] --> B["先自愿验证，后法定推进"] B --> C["完全豁免开源与初创企业"] C --> D["替代新增行政机构，防止官僚叠加"] end subgraph FAA["FAA / 许可式监管陷阱"] E["型号认证 Type Certification"] --> F["全新设计需审批 5-9 年"] F --> G["严重拖垮每隔数月迭代的创新节奏"] G --> H["拱手让出对华 AI 竞争主导权"] end ``` Sacks 透露，Demis 在私人沟通中基本认同这些考量，但真正的危险在于 Anthropic 创始人 Dario Amodei 的态度。Dario 虽然表面赞成自律构想，但 Sacks 认为这只是 Anthropic 的“起步要价”（Opening bid），其终极目标是推动建立“AI 领域的联邦航空管理局（FAA）”。Sacks 直言这种对比极为荒谬：在航空领域，FAA 的型号认证（Type Certification）针对全新飞机设计通常需要 5 到 9 年（即使是波音 737 MAX 的改装也耗费了约 5 年）。若将这种“必须获准方可商用”的前置审查强加给数月便迭代一次的 AI 行业，整个美国的创新速度将被拖入按“年”计算的官僚泥潭，而中国绝不可能遵守这套规则，美国将因此直接输掉 AI 竞赛。 ### [13:34](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=814s) 监管俘获的隐忧与权衡：两害相权取其轻 Chamath Palihapitiya 赞成迅速建立行业自律机构，并警告称巨额资本正源源不断地涌向两党政客，试图通过法律设立行业壁垒。如果不尽快建立一套全行业通行的自律规范、切断直接设立庞大联邦监管机构的借口，掌握雄厚资产负债表与资本的巨头就会趁机“过河拆桥”（Pull the ladder up），将竞争者彻底阻隔在门外。Chamath 指出，虽然自律组织下依然会有商务部和司法部等既有联邦监管机制，但这能有效防止市场被少数寡头彻底垄断。 Sacks 补充道，Demis 提出的参考蓝本——金融业自律组织 FINRA（金融业监管局）最终也是向 SEC 汇报的。一旦设立此类机构，它究竟向哪个政府部门负责必然引发激烈的政治争夺，而软件产业历史上从未有过专门的监管机构。然而，两害相权取其轻，面对“开源被扼杀、市场走向双头垄断（Duopoly）”的最坏结局，纯粹且受控的自律组织依然是当前最具现实可行性的防线。 - **资本阻击**：大科技公司正动用资产负债表游说两党，企图通过抬高合规成本锁死生态位。 - **历史惯例**：软件行业数十年未曾设立专属监管机构，引入类 FINRA 架构极易让软件行业首次受到直接的行政干预。 ### [15:58](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=958s) 戳痛老虎：Anthropic 的“州级游击战”与逐利阳谋 话题转向了监管狂热背后的始作俑者。主持人与嘉宾们直言 Anthropic 是在“拿肉喂老虎”，通过主动挑起公众恐慌来逼迫政府介入。Sacks 提到，自己在去年 10 月发推指责 Anthropic 借恐慌情绪实施精密的“监管俘获”（Regulatory capture）时曾遭遇广泛围攻，当时舆论误以为他是在打压一家弱小的初创公司。如今真相大白：Anthropic 的估值已达数万亿美元规模（Gavin Baker 甚至预测其 IPO 后将突破 3 万亿美元），从营收和各方面看都已是庞然巨物，完全拥有动用雄厚资源阻断竞争的实力。 Sacks 援引《Politico》近期的深度报道《Anthropic 步步升级 AI 监管的州级策略内幕》（Inside Anthropic’s State-by-State Plan to Ratchet Up AI Rules）指出，Anthropic 正在全美推行一种“层层加码的竞标策略”（Strategy of one-upmanship）。他们先是在加州推动了 SB 53 等严格法案，随后又以此为基准，在伊利诺伊州、纽约州等蓝州不断推波助澜，促使各州制定越来越严苛、越发无所不包的 AI 规则。他们非但不寻求统一的联邦标准，反而故意利用各州的监管竞赛制造碎片化的合规难题，让缺乏雄厚法律合规团队的中小型初创公司在重压下彻底出局。政客们天生渴求权力，当巨头打着“请监管我”的幌子主动登门时，官僚体系绝不会回绝，只会顺水推舟将权力无限扩张。 ### [19:04](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=1144s) 科技行业不应绥靖求全：自律组织（SRO）必须伴随联邦预先豁免 科技行业目前普遍存在一个致命的误判，即以为通过向政客或政治体系主动让步就能换取安宁。这种绥靖策略绝不会息事宁人，反而只会导致监管压力的步步紧逼。政府乐于全盘接受科技界的主动献策，随后再变本加厉地索取更多权力，直至将整条赛道完全纳入国家控制之下。科技巨头们终究需要拿出骨气予以反击，明确划出自己的底线。 如果 Demis Hassabis 提议的自律组织（SRO）就是科技界划定的底线，各家公司必须展现出寸土不让的姿态，并在法律中明确要求以“联邦预先豁免权”（Preemption）等关键条款作为交换条件，彻底排除各州碎片化监管的干扰。倘若行业仅仅是将自律组织作为免费的政治贡品拱手奉上，各家公司唯唯诺诺地高呼“快来监管我们吧”，这绝不会成为立法的终点，而仅仅是政府全面收缴行业控制权的一场残酷开场白。 ### [20:01](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=1201s) 530亿美元重磅并购：Stripe、Block 与 Advent 联手竞购 PayPal 科技投资界迎来了一桩震撼级重磅交易：至今仍顶着股东巨大催上市压力保持私有状态的支付巨头 Stripe，联合老牌私募股权基金 Advent International，正式向上市公司 PayPal 发起收购要约，报价为每股约 60 美元，交易总估值达 530 亿美元。尽管市场普遍认为最终的成交价格很可能会被推高至每股 70 美元左右，PayPal 的股价依然应声暴涨。随后媒体进一步曝出重磅消息，Jack Dorsey 旗下的上市金融科技公司 Block（原 Square）也加入财团，并计划为这笔联合收购注入高达 170 亿美元的股权资本。 对于作为 PayPal 早期创始团队成员（PayPal 帮）的大卫·萨克斯（David Sacks）而言，这笔收购充满戏剧性。PayPal 历经二十五年的行业洗礼，至今依然拥有 4.39 亿活跃消费者账户，沉淀了庞大的网络效应。然而，该产品早已显露疲态，年营收增速放缓至 7% 左右，在许多新一代从业者眼中甚至已沦为一种过时的“遗产型资产”（Legacy Product）。大家非常好奇 Stripe 的领导层究竟打算如何改造这只步履蹒跚的庞然大物，抑或是纯粹像传统私募股权基金那样通过精简运营、削减开支来对其进行现金流榨取。 - **报价与估值**：Stripe 与 Advent 联合提出约每股 60 美元要约，整体估值达 530 亿美元；市场预期可能需溢价至 70 美元方能达成交易。 - **关键参与方**：Stripe、私募巨头 Advent International，以及提供 170 亿美元股权注资的 Block。 - **标的核心体量**：拥有 4.39 亿活跃消费者账户，旗下囊括 Venmo 等国民级移动支付品牌。 ### [22:42](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=1362s) 垂直整合剑指 Visa 与万事达：稳定币与6亿账户构建新支付网络 如果深入拆解 Stripe、Block 与 Advent 三方联手背后真正的战略协同，就会发现这场交易绝非一次普通的资产拼凑，而是在酝酿一个足以颠覆全球金融基础设施的新物种：打造一个能与 Visa 和万事达卡（Mastercard）正面对决的独立支付网络。在传统模式下，Stripe 长期作为搭建在传统信用卡清算通道上的中间商，外界曾普遍担忧其手续费抽成会被激烈竞争逐步侵蚀；但 Stripe 凭借长达十五至二十年的精深积累，成功构筑起了涵盖风险管理与增值服务的高壁垒护城河。 如今，这三方整合将实现从前端账户到底层清算的端到端彻底垂直整合。合并后的生态将坐拥 6 到 7 亿个消费者与商户账户，更掌握着业界最具统治力的稳定币基础设施——包括 Stripe 在 2025 年斥资 10 亿美元收购的 Bridge 基础设施平台，以及 PayPal 已经大规模流通的美元稳定币 PYUSD。一旦绕过传统卡组织的清算网络，Stripe 与 Block 将能够以近乎于零的清算成本在全球范围内瞬间调拨资金，而 Advent 的雄厚资本则能为该网络提供源源不断的资金保障。 ```mermaid flowchart TD accTitle: 联合收购与支付网络垂直整合架构图 accDescr: 展示 Stripe、Block 与 Advent 联合收购 PayPal 后形成的颠覆 Visa 和 Mastercard 的超大型金融基建 subgraph Capital["资本与收购主体"] ADV["Advent International<br/>（私募资金支持）"] STP["Stripe<br/>（Bridge 稳定币基建 + 风控中枢）"] BLK["Block / Square<br/>（170亿美元股权注资 + Cash App）"] end subgraph Target["核心整合标的"] PYP["PayPal 集团<br/>（4.39亿账户 + Venmo + PYUSD）"] end subgraph Impact["重构后的全球基础设施"] NEWNET["新一代超级支付清算网络<br/>（6~7亿总账户 + 超低成本稳定币轨道）"] RIVALS["对标并颠覆传统垄断清算网络<br/>（Visa / Mastercard）"] end ADV --> PYP STP --> PYP BLK --> PYP PYP --> NEWNET NEWNET -.->|正面竞争| RIVALS ``` ### [24:17](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=1457s) “AI 原生代”爆改过气互联网巨头：从 eBay 到 PayPal 的并购新范式 从宏观资本市场视角来看，这起交易标志着一轮全新的并购周期的爆发。此前瑞安·科恩（Ryan Cohen）对 eBay 发起的激进要约正是这股浪潮的前奏。当前资本市场出现了一个鲜明的新趋势：“AI 原生代”（AI Native）运营者开始将目光投向那些成熟、甚至略显衰颓的第一代“数字原生代”（Digital Native）互联网企业。这些公司大多已不再由创始人掌舵，坐拥海量用户与资产却增长停滞，既没有充分发挥其网络效应的商业潜能，内部严重超额支出，更彻底错失了利用 AI 进行深层重构的历史机遇。 这类交易的核心商业逻辑在于“AI 化赋能”： 1. **资本联姻超级运营方**：资本方不再迷信麦肯锡等传统咨询公司的空洞报告，而是直接与行业中最顶尖的执行派运营商结盟，例如乔希·库什纳（Josh Kushner）通过资本扫货传统会计师事务所并进行 AI 化改造，或是 General Catalyst 推进的类似整合项目。 2. **交易结构与实际控制权**：在这类结构中，Stripe 和 Block 实质上是向外部现金投资人定向出售股权筹集资金，再去竞购 PayPal，最终由具备最强作战能力的 Stripe 团队全面接管 PayPal 闭环后的实际运营工作。 3. **老树发新芽的超级并购潮**：公开市场中存在数十家创始人早已离场、业务增长陷入停滞的早期网络平台与软件巨头。充足的资本杠杆如同“商业版蓝矾（BlueChew）”，配合顶尖操盘手的强效执行力，将掀起一波复活陈旧数字资产的世纪并购巨浪。 现场还对潜在的竞标变数展开了遐想：当前 530 亿美元绝非这笔交易的最终价格，收购价极可能再上涨 10% 到 15%。甚至不排除会有某种掌握 4 到 5 万亿美元市值底气、原本就在早期 PayPal 留下过深刻烙印、且近期刚完成过 600 亿美元级别并购的“关键个人”强势介入竞标。与此同时，欧洲投资公司 Bending Spoons 刚刚完成上市，其操盘手从米兰远程统领全局，先后以 14 亿美元收购 AOL、14 亿美元接盘 Vimeo，并将 WeTransfer、Eventbrite 和 Evernote 等一众非创始人掌舵的资产收入囊中，凭借对产品底层架构与成本开支的精准外科手术式诊断，成为了当今互联网二级运营整合的标杆范例。 ### [28:54](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=1734s) 并购潮全面回归：政治风向转变与风投流动性复苏 Jason Calacanis 与嘉宾们讨论了像 Bending Spoons 这类公司的打法：通过年轻的 AI 原生代高管去整合所谓的 Web 2.0 老业务，削减运营成本，让老树开出能够大笔印钞的新花。但更大的宏观背景在于，整个风险投资（VC）行业此前在联邦贸易委员会（FTC）前主席莉娜·汗（Lina Khan）的高压反垄断监管下被狠狠压制了数年，陷入极度被动的僵局。随着唐纳德·特朗普（Donald J. Trump）再度当选总统，所有大企业的企业发展（Corp Dev）部门高管瞬间嗅到了风向变化，高呼“并购交易重回菜单（M&A is back on the menu）”。 交易市场的闸门一开，多起重磅收购接踵而至，市场不再畏惧反垄断审查。Uber 当天就宣布收购 Delivery Hero，尽管伴随约 10% 的股权稀释，但预计将直接拉动其营收暴增 24% 左右。Jason 透露，现在与有限合伙人（LP）及家族办公室交流时，对方最热衷询问的就是“下一期基金什么时候融”或“下一笔交易在哪里”。伴随 SpaceX 等超级独角兽的大额分红套现，以及新政府上台后自家已有四五家被投公司顺利被收购，LP 们对风投生态的信心正迎来强劲反弹。 ### [30:28](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=1828s) PayPal 的沉沦根源：创始人基因出局与“流散者”神话 谈到 PayPal 的现状，Jason 调侃 David Sacks 为什么当年没去接任 PayPal 的 CEO。Sacks 回忆起约十五年前的一段往事：当时他在最终候选的两选一决选中输给了一位传统的信用卡行业高管。Sacks 直言，PayPal 后来陷入长达二十余年的停滞，核心病灶正是从 2002 年被 eBay 收购的那一刻起埋下的——当时由梅格·惠特曼（Meg Whitman）掌舵的 eBay 充满了宝洁（Procter & Gamble）、迪士尼（Disney）以及贝恩咨询（Bain）式的传统大企业官僚思维，对他们而言，PayPal 那批野心勃勃、个性鲜明的创始人只是一群不可控的“牛仔”，根本不愿花心思挽留。 Sacks 澄清，外界津津乐道的“PayPal 黑手党（PayPal Mafia）”实际上是一个用词不当的误读，它更应该被称为“PayPal 流散者（PayPal Diaspora）”。因为当时的创始团队更像是一群被外力夺走故土、神庙被焚毁后被迫驱逐出境的流亡者，正是这种被锁在门外的经历，迫使这批人带着资本去创办了后来的 YouTube、LinkedIn、Yelp 和 Palantir 等一众科技巨头。如今 PayPal 的市值从巅峰时期的 3220 亿美元暴跌至收购要约前的 300 多亿至 400 亿美元区间，尽管现任 CEO Enrique 在降本增效上表现扎实，但被严重低估的股价与彻底老化的 25 年前交互模型，最终引来了外部财团的野心觊觎。 ### [33:10](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=1990s) 巨头拼图与反垄断重构：绕开卡组织的自有支付闭环 针对 Chamath Palihapitiya 提出的“收购核心在于客户账户而非产品本身”，Sacks 与众人深入剖析了 Stripe、Block 与 Advent 联手竞购 PayPal 的底层商业逻辑与协同效应。Stripe 是商户端 API 接入的绝对王者，年交易量高达约 2 万亿美元；而 PayPal 的核心资产在于年交易量 1.7 万亿美元背后的 4 亿多活跃个人消费者账户，以及旗下掌控的 Braintree 和深得千禧一代与 Gen-Z 青睐的 Venmo。如果再加上 Block（Square）强大的线下收银终端（POS）基础设施与 Cash App，买方实际上能够打通一个完全端到端的交易闭环。 ```mermaid flowchart TD accTitle: 联合收购 PayPal 后的支付闭环架构 accDescr: 展示 Stripe、Block 与 PayPal 旗下资产整合后打通商户与消费者的闭环，从而绕开 Visa 和 Mastercard。 subgraph Merchant["商户端体系 (Merchant Rails)"] S["Stripe (API 商户网络)"] B["Braintree (商户清算)"] POS["Block / Square (线下 POS 终端)"] end subgraph Consumer["消费者端体系 (Consumer Rails)"] PP["PayPal (4亿+ 活跃钱包账户)"] V["Venmo (年轻用户/P2P)"] CA["Cash App (移动支付生态)"] end subgraph ClosedLoop["自建闭环网络 (On-Us Network)"] TX["站内自清算交易 (On-Us Transactions)"] end CardRails["传统卡组织 (Visa / Mastercard)"] Merchant --> TX Consumer --> TX TX -.->|"绕过 / 颠覆"| CardRails ``` 这笔交易的战略杀伤力在于绕开传统信用卡网络： - **自建清算网络（On-Us）**：商户与消费者的直接对接允许平台在内部完成账务划转，无需向 Visa 和 Mastercard 支付高昂的刷卡手续费（Interchange Fees）。 - **商户端让利飞轮**：如 Chamath 所言，平台只要向 Shopify 等大商户提供 2% 到 5% 的交易费率折扣，商户就会主动引导消费者放弃传统刷卡方式，从而大幅降低商品零售价格并真正惠及消费者。 - **反垄断定义的攻防**：如果在两年前由 Lina Khan 掌舵的监管环境下，这笔合并面临的审查“无异于一次极度痛苦的结肠镜检查”，因为若将市场细分为“商户 API”，Stripe 与 Braintree 的合并涉嫌消灭竞争；但在当前的监管风向与更高维度的市场定义下，整个交易的对标对象变成了真正垄断全球的 Visa/Mastercard 双寡头（Duopoly），从而能够强有力地论证该交易不仅不反竞争，反而极大地促进了跨时代的市场竞争。 ### [037:52] 苹果起诉 OpenAI：41页诉状背后的硬件窃密风波与挖角争议 主持人 Jason 抛出重磅新话题：又一家科技巨头将 OpenAI 告上法庭，而这次出手的是苹果公司。7月10日，苹果提交了一份长达41页的起诉书，直指 OpenAI 窃取商业机密用于开发其消费级硬件设备。不久前科技记者 Sarah Frier 在节目中曾透露，这款正在研发的硬件非常“人性化且讨人喜欢”，但如今苹果却指控该设备底层很可能包含其知识产权。 诉讼细节格外引人注目：苹果前 iPhone 设计副总裁 Tang Tan 现任 OpenAI 首席硬件官，据称他曾要求前往 OpenAI 面试的苹果员工在面试时携带“实际零部件”来进行“实物展示与讲解”（show and tell）。此外，前苹果高级技术工程师 Chang Xu 曾给一名在职的苹果前同事发短信炫耀：“哈哈，我发现自己居然还能连上内网云存储，太搞笑了。”过去一到两年内，OpenAI 从苹果大举挖走了超过400名员工。面对如此明目张胆的挖角和机密外流，苹果 CEO 蒂姆·库克（Tim Cook）亲自批准了这场诉讼。 Jason 辛辣地指出，Sam Altman 似乎又一次在合作伙伴背后捅刀子——先是闹翻了最初的资助人马斯克，而在此前达成的重磅合作中，苹果本已指定 ChatGPT 作为 iPhone AI 的默认选项，让 OpenAI 独占了全球最重要的 AI 硬件平台入口，结果如今却因涉嫌偷窃硬件机密而对簿公堂。 - **起诉核心**：苹果向法院递交 41 页诉状，指控 OpenAI 窃取商业机密并用于开发消费级硬件。 - **涉案高管与线索**：原苹果 iPhone 设计副总裁 Tang Tan 现任 OpenAI 硬件一把手，涉嫌指示应聘者带苹果实体零件进行面试演示；工程师 Chang Xu 离职后仍发短信称能访问苹果内部存储。 - **挖角规模**：OpenAI 在过去一至两年内连续挖走逾 400 名苹果员工。 ### [039:50] 硅谷职场红线：记忆可以带走，资产绝不能碰 Chamath 针对这起诉讼表达了自己的看法。他提到自己在硅谷打拼了25年，极少见到苹果在法律诉讼上如此大动干戈。苹果一贯的做法是被动防御而非主动出击，因此他们这次选择起诉，一定是发生了极其恶劣、真正惹恼苹果的事情。尽管最终真相需要法庭审理定夺，但他强调了一条硅谷的基本底线：“无论何时，任何人都不应该从前雇主那里偷东西。”这些被挖走的高管和工程师非常聪明成功，他们带走多年沉淀的经验与认知本就足够，完全没有必要去违规拿走机密。 David Sacks 随后从行业宏观层面进行分析。加利福尼亚州崇尚自由就业与劳动力流动，法律上通常不允许竞业禁止协议（non-compete clauses），但这绝不代表可以带走老东家的知识产权。当年 Waymo 起诉 Uber 就是前车之鉴——前工程师窃取商业机密，迫使 Uber 最终不得不辞退当事人并划清界限。 Sacks 给所有职场人提出了一条极简的避坑法则：当你跳槽去新公司时，唯一被允许带走的只有你脑子里的东西、你的经验与记忆；文档、U盘、光盘以及任何物理或数字资料，哪怕一片纸都绝对不能碰。当 Jason 询问 Friedberg 对 OpenAI 官司越积越多的看法时，Friedberg 意味深长地回了一句冷幽默：“两点成一线，大家心里都有数（Couple dots make a line）。” ```mermaid flowchart TD accTitle: 硅谷人员跳槽知识产权合规边界 accDescr: 展示员工在加州离职跳槽时，合规认知与违法侵权的明确界限。 subgraph 合法带走 ["合法范畴（合规边界内）"] A["个人脑海中的知识"] B["多年积累的经验与智慧"] C["专业认知与技能"] end subgraph 严禁携带 ["违法红线（法律诉讼风险）"] D["实体硬件零部件 / 样品"] E["前公司内网 / 云端存储访问权限"] F["U盘、光盘、设计图纸与机密代码"] end 合法带走 --> G["新雇主（新岗位落地）"] 严禁携带 -.->|引发商业机密诉讼| H["法院裁决与刑事调查<br/>（如苹果诉OpenAI案）"] ``` ### [042:49] Grok Build 隐私漏洞曝光：ZDR 承诺在现实面前的脆弱性 Jason 将话题引向马斯克旗下 xAI 的数据泄露事件。xAI 在5月底公测了 Grok Build，其底层由最新的 Grok 4.5 代码模型驱动，并能嵌入 Cursor 等开发工具中无缝使用。虽然官方此前向用户承诺代码绝不会上传至 xAI 服务器，但实际曝光的报告显示，每次开发者使用该工具时，Grok Build 都会在后台悄无声息地将整个代码库打包传回云端，不仅包括当前任务涉及的文件，甚至连密码、API Key、变更日志等最敏感的核心数据也被一并抓取。 尽管用户勾选了隐私保护选项，该功能却未正常生效。发现问题后，xAI 在7月13日紧急通过服务端静默开关切断了上传通道，马斯克在社交平台 X 上公开承诺已将先前所有误上传的数据彻底销毁，并将 Grok Build 框架予以开源作为补救。 Chamath 借此事件呼应了他前不久在 CNBC 上的表态以及 Palantir CEO Alex Karp 的演讲。他指出：AI 领域的隐私保护极其脆弱，即使像马斯克这样极具透明度和信誉的团队，稍有不慎也会踩坑。行业里常鼓吹所谓的“零数据留存”（Zero Data Retention, ZDR），将其包装成安全灵药，但真实世界里到处都是防不胜防的后门与暗道。企业若想真正自保，必须通过中立的第三方独立架构对大模型调用层进行隔离，建立严格的数据防火墙，这正是其旗下公司 8090 正在实践的企业级软件工厂模式。 - **漏税事件**：Grok Build 因隐私保护开关失效，在未经授权的情况下将用户完整代码库（包含 API Key 与密码）上传至 xAI 云端。 - **应对措施**：xAI 紧急切断同步通道，马斯克承诺删除误存数据，并开源 Grok Build 框架以平息信任危机。 - **ZDR 质疑**：Chamath 认为单靠大模型厂商口头承诺的“零数据留存（ZDR）”根本不足以保障安全，模型层外部的第三方隔离架构必不可少。 ### [045:53] 反向信息悖论与 Token 成本战：闭源高溢价 VS 算力平价 Sacks 提到了 Satya 最近发表的一篇关于“反向信息悖论”的精彩博客文章。这篇博文进一步深化了 Alex Karp 此前的论断：凡是具备技术实力的大型企业，都极度渴望牢牢掌控自己的算力、模型权重、私有数据以及核心竞争壁垒（Alpha）。Satya 为企业落地掌控权给出了一套清晰的实操蓝图，包括：在租户内部构建私有评测与专有学习闭环、将模型调度与编排完全解耦，并拥有对自己模型输出进行微调的明确权利。 科技界领袖们正在自发形成一股合力，试图打破由 Anthropic 和 OpenAI 构筑的“封闭单体技术栈”。这些闭源巨头不仅希望将客户全盘锁定在自身生态中，还要收取令人咋舌的超高溢价。 Chamath 引用了 Ramp 首席执行官 Eric Glyman 在《Squawk Box》上分享的最新企业员工“Token 消费（Tokenmaxxing）”管理数据，横向对比了市场上主流模型的百万 Token 调用成本。两相对比之下，闭源巨头的价格劣势和开源/高性价比模型的冲击展现得淋漓尽致： | 模型厂商 / 产品分类 | 每百万输入 Token 价格（美元） | 市场与生态特征 | |:--- |:--- |:--- | | **闭源高价代表（如 Fable）** | 约 $56.00 | 深度技术栈绑定，锁定企业资产，溢价极高 | | **主流前沿闭源（如 Soul / Claude 4.8）** | 约 $26.00 | 综合能力强大，但对企业数据主权存在挤压 | | **新锐大模型（如 Grok 4.5）** | 约 $1.50 | 价格断崖式下降，主打算力性价比 | | **轻量级 / 平价模型代表（如 Z）** | 约 $1.50 | 主流平价替代方案 | | **马斯克相关生态模型** | 约 $1.00 | 极致压缩推理成本 | | **中国大模型（Chinese Models）** | 约 $0.50 | 极致性价比，加速重构全球算力价格底线 | 面对如此悬殊的调用成本与潜藏的数据主权流失危机，企业正越来越不愿意向昂贵的封闭模型纳贡，转而寻求真正自主可控、低成本的开源及自建生态方案。 ### [48:37](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=2917s) Token 开销失控与 Ramp 推出支出管理工具 对话延续了对高昂 Token 成本的震惊：如果输入 Token 每百万要收取高达 56 美元的费用，承担这种风险完全是“疯了”。Jason 提到自己正在使用 Perplexity 的 Computer 环境，它已经开始支持 Grok，并且此前已接入 GLM-52。在像 Claude 或 OpenAI 这样的封闭平台中，用户往往只能绑定单一的模型；于是 Jason 做了一个实验：他为不同模型输入了完全相同的 PRD（产品需求文档），要求编写一个具备深度跳转功能的播客播放器——比如当节目提到“Mythos”时，自动检索抓取所有科技和商业播客中关于该主题的剪辑片段，拼成一条连贯的音轨。结果显示，整个开发仅花了两小时，在最新的 Grok 上只耗费了 11 美元，成本之低令人咋舌；而他在 Fable 上的 200 美元账户早早就把额度烧光了。 随后，现场播放了金融科技公司 Ramp 联合创始人 Eric Glyman 的一段视频剪辑。Eric 宣布 Ramp 正式向客户及非客户群体推出“Token 支出管理”（Token Spend Management）功能。数据显示，过去一年内 Ramp 客户的 Token 支出暴增了整整 21 倍，而不是 21%。目前 AI 厂商为客户设立的机制就像是一张“无限记账单”（open tab），员工可以无节制地挥霍算力，而当厂商不断推出更昂贵的新模型时，CFO（首席财务官）事前完全无法穿透监督。Chamath 指出，如果工程师团队在毫无治理与预算控制的体系中肆意消耗每百万 Token 价值 56 美元的算力，最终必将直接传导并重创公司的财报利润。在支出每隔几个月就翻数倍的情况下，很快就会有上市公司 CFO 面对华尔街时因为 Token 运营开销（OpEx）严重超标而爆雷、无法达成季度盈利预期。如果没有严格把控并核算真实的投资回报率（ROI），这种烧钱模式就是一座通往深渊的焚钞炉。 - **Ramp 支出数据**：其平台企业客户的 Token 支出在过去一年中激增了 21 倍（2100%）。 - **失控根源**：工程师天然追求最新最强的昂贵前沿模型，且报销或用量无需承担直接成本，导致与关注 ROI 的 CFO 之间存在严重的激励错位（Misaligned incentives）。 --- ### [51:28](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=3088s) 开源平替、Mira Murati 的 Inkling 与端侧芯片的破局 高昂的 Token 费用正催生庞大的市场机会与开源平替浪潮。从 Bittensor 子网、GLM-52，到托管 Grok 4.5 以及 Mira Murati 创立的 Thinking Machines 推出的全新模型 Inkling，市场参与者纷纷打出价格战。嘉宾们分析，企业在封闭巨头处需要为每百万 Token 支付 26 至 56 美元，但其实际价值往往只值 50 美分。Inkling 的定位十分明确：它坦言自己不是追求最极致的前沿智能（Frontier Intelligence），而是略逊于顶尖梯队，专注于构建开源模型的微调平台，让企业以极低成本通过针对性微调达到特定业务需求。David Sacks 讽刺道，正因开源模型只要 50 美分即可取代 56 美元的专有产品，Anthropic 等垄断型巨头才千方百计试图推动监管来阻挠开源发展；不过尽管开源生态爆发式增长，专有模型巨头的营收眼下依然在狂飙突进。 Jason 认为，98% 的日常 Prompt 任务根本不需要顶级前沿模型，完全可以用 Qwen（通义千问）或次级开源模型解决。他进一步做出了一项重要行业预测：结合彭博社记者 Mark Gurman 的爆料，苹果 M7 Ultra 芯片预计将支持高达 1.5 TB 的统一内存，相比现款 M5 时代的 456GB 或 512GB 翻倍。这意味着用户只需一台价值 4,000 到 5,000 美元的 Mac Studio，就足以在本地全速运行上一代 Opus 级别的顶尖前沿模型。届时，企业 90% 到 99% 的 AI 计算负载都可以完全放在本地运行，桌面上拥有无上限、零边际成本的 Token。这种端侧内存架构的突破，将给 Claude 和 OpenAI 带来极其巨大的降价压力，苹果或将借此再次重演 iPhone 颠覆行业的商业奇迹。 ```mermaid flowchart TD accTitle: AI模型推理成本与端侧计算演进路径 accDescr: 对比昂贵闭源云端模型、微调开源模型与未来苹果端侧大内存架构的成本演进。 subgraph CloudHighCost["昂贵云端闭源模型"] A["Frontier 顶级闭源模型<br/>(Opus/Claude/OpenAI)"] --> B["每百万 Token 26~56 美元<br/>OpEx 失控 / 吞噬财报利润"] end subgraph OpenAlternatives["开源与定制替代方案"] C["Thinking Machines Inkling / Grok / GLM / Qwen"] --> D["次级底座 + 针对性微调<br/>每百万 Token 仅约 0.5 美元"] end subgraph EdgeCompute["端侧高内存革命"] E["Apple M7 Ultra 架构<br/>(支持达 1.5 TB 统一内存)"] --> F["本地部署 Opus 级模型<br/>桌面级无限 Token / 零推理开销"] end B -.->|成本倒逼替代| D D -.->|进一步转移到端侧| F ``` --- ### [56:06](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=3366s) 电力严重赤字、电网短缺与 Elon Musk 的 Memphis 破局 端侧计算的兴起本质上也是在应对电力资源的极限压力。嘉宾们提到，像大型太阳能公司 Sunrun 刚刚宣布推出家用分布式数据中心模块，初创公司 SPAN 也与 Nvidia 合作切入家庭电能管理与边缘计算，算力的分布式碎片化已成为不可逆转的趋势。更为严峻的底层瓶颈是电力供给。Chamath 指出，到 2050 年，美国在电力供给上将出现相当于“2.5 个加利福尼亚州用电总量”的巨大赤字——加州是全球第四大经济体，2.5 倍的电力缺口意味着极其危险的能源短缺。 在现实层面，这场危机已经开始显现。覆盖宾夕法尼亚州、新泽西州、马里兰州等 13 个州的美国最大区域电网运营商 PJM 举行了远期容量拍卖，以提供保障性固定日费率为条件吸引发电增量以满足未来预测负荷。当时电网需要大约 7 到 8 吉瓦（GW）的电力支撑，但实际竞标响应的竟然只有区区 156 兆瓦（MW）左右。全美电力供应陷入了极其被动的局面。面对电网电力（Grid Power）审批与输电线路建设的长周期阻碍，Elon Musk 在田纳西州孟菲斯（Memphis）为 xAI 的 Colossus 超级计算机寻找能源时，果断采取了“表后发电”（Behind-the-meter）的策略，直接收购燃气轮机公司进行自发电，规避向传统公用事业公司申请高压输电专线的漫长流程与监管僵局。 - **全美用电缺口预测**：预计到 2050 年，美国能源短缺总量将相当于 2.5 个加州的用电总和。 - **PJM 电力容量拍卖惨案**：远期需求缺口达 7~8 GW，但实际中标供应量仅约 156 MW。 - **算力供电模式差异**：传统电网供电（Grid Power）受限于电网主干输电线路与严苛审批；Elon Musk 则采用自备天然气轮机的“表后”（Behind-the-meter）独立发电模式为 Memphis 超级集群供电。 ### [58:16](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=3496s) “表后发电”破解电网瓶颈与马斯克的能源突围 当接入公共电网的容量耗尽或审批积压严重时，数据中心运营商不得不转向所谓的“表后发电”（behind the meter）方案——即在自己拥有的土地上自建发电设施。然而在美国，任何举措都会面临层层叠叠的监管壁垒，其中最复杂的门槛之一就是《清洁空气法》许可（clean air permitting）。太阳能占地面积过大，电池储能则需要预先发电，因此企业普遍采用天然气。 为了应对这一挑战，埃隆·马斯克（Elon Musk）采取了极其精明的策略：他买下了生产大型移动发电设备（本质上是类似18轮重卡发动机的涡轮发电机组）的公司，直接将其固定在地面上运转。由于这些设备在名义上属于个人或内部自用性质，因而能够绕过严苛的集中式清洁空气许可审查；哪怕未来被整体定性，也能通过法律抗辩争取空间。此外，市场上还涌现出 Bloom Energy 等新兴解决方案，允许部署超大规模装机容量的同时，仍满足个人使用性质的清洁空气许可规范。马斯克提前布局这些能源资产，正是为了为其未来的算力扩张铺平道路，确保其在国内推进数据中心建设时拥有清晰、畅通的审批视野。 ```mermaid flowchart TD accTitle: 数据中心供电路径对比 accDescr: 对比公共电网连接与表后发电方案在审批和执行上的差异 subgraph Grid["公共电网路径"] A["电网接入申请"] --> B["面临严重积压与审批瓶颈"] B --> C["推高居民电费引发民怨"] end subgraph BehindTheMeter["表后自备发电路径"] D["企业自有土地自建电厂"] --> E{"能源形式选择"} E -->|"太阳能"| F["占地面积过大受限"] E -->|"电池储能"| G["仍需上游发电源"] E -->|"天然气发电机组"| H["清洁空气许可审批 (Clean Air Permit)"] H --> I["创新方案: 移动燃气轮机/Bloom Energy"] end ``` ### [59:53](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=3593s) 纽约州出台首个数据中心暂停令，嘉宾逐条驳斥州长指控 讨论随即转向纽约州州长凯西·霍楚尔（Kathy Hochul）宣布签署的全美首个州级超大规模数据中心禁令（moratorium）。霍楚尔在发言中声称，数据中心依赖化石能源推高碳排放、占用大量可能用于农业或开放空间的土地，并指责其导致居民电费上涨、耗尽水资源并带来噪声污染。大卫·萨克斯（David Sacks）对此怒斥霍楚尔的言论纯属“全盘捏造的伪指控”，并与查马斯（Chamath Palihapitiya）逐一展开事实反驳： - **电费上涨问题**：只有在不增发电力、直接接入现有电网与普通居民抢夺负荷时，数据中心才会推高公共事业费率。如果按照特朗普政府倡导的模式推进“表后自建”，允许 AI 公司转型为发电主体，这一问题便迎刃而解。 - **土地利用效率**：数据中心是土地利用效率的典范。美国有大量闲置土地，以其单位面积创造的经济产值和 ROI 来看，数据中心居于所有产业的前列。杰森·卡拉卡尼斯（Jason Calacanis）补充道，纽约州除了拥挤的纽约市以外，北部（Upstate New York）有 70% 至 80% 的土地完全处于未开发状态，指责其侵占空间极其荒谬。 - **水资源与噪声争议**：现代数据中心采用闭环再循环冷却系统，萨克斯援引研究指出，一个典型数据中心的实际耗水量仅相当于 2.5 家 In-N-Out 汉堡连锁店的用水量；若真在乎节水，更应该去治理加州的杏仁种植或高尔夫球场。至于噪声，拉开与居民区的一定物理距离即可消除。 - **经济与税收效益**：数据中心不仅带来蓝领建筑工程热潮与长期的运维工作岗位，还创造了巨额地方税收。例如北达科他州部分地区凭借数据中心税收，直接为当地教师发放了 3 万至 4 万美元的奖金。 - **所谓污染问题**：绝大多数自备发电采用天然气，而天然气是目前人类掌握的燃烧最清洁的基荷能源之一。 萨克斯总结道，数据中心如今沦为了公众对 AI 焦虑情绪的“替罪羊”，这种行政干预不过是政客试图朝技术创新轮子中硬塞扳手、阻碍行业前进的粗暴手段。 ### [01:03:50](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=3830s) 职业抗议背后的资金链与 Anthropic 的自我扼杀悖论 查马斯分享了他本周参加戴夫·麦考密克（Dave McCormick）参议员在宾夕法尼亚州陆军战争学院举办的国防与创新峰会见闻。特朗普出席了该峰会，多位防务巨头 CEO 及能源领域知名人士克里斯·莱特（Chris Wright）也悉数到场。在迪娜·鲍威尔（Dina Powell）提问时，克里斯·莱特揭露了一个惊人内幕：当前在全美各地抗议数据中心建设的团体，其幕后资助模式与当年抗议水力压裂（fracking）采油的资金网络几乎完全重合。这些人本质上是一群以此谋生、四处串场的“职业抗议者”，抗议话题只是他们筹集经费的工具。 更为荒谬的是头部 AI 机构的内部撕裂。萨克斯指出，Anthropic 及其首席执行官达里奥·阿莫代伊（Dario Amodei）以及多名员工，竟然在持续向类似“Public First”这样极力推动限制数据中心建设、试图阻碍 AI 发展的游说团体提供七位数（百万美元级）的巨额捐款。萨克斯质疑，这究竟是弄巧成拙的监管俘获（regulatory capture）企图，还是彻底的理念失智？毕竟当前限制 Anthropic 营收增长的真正瓶颈根本不是市场需求，而是数据中心算力的极度短缺。 ### [01:05:37](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=3937s) 深度政治图谋与数据中心长达五年的产业停滞期 针对民主党主导的禁令政策，萨克斯引述了一位资深政界人士的深度推演：民主党设立所谓“暂缓期”（moratorium）并非打算永久取缔数据中心，而是意图通过叫停项目来争取筹码，直到他们认为自身具备足够的政治控制力来彻底制定行业规则。 推测的剧本是在未来某一届民主党政府或特朗普卸任后，官方才会以特定政治条件解除禁令。届时随之而来的将是一套由大政府主导的苛刻 AI 监管框架，包括设立专门的监管机构、引入言论审查，并将过去针对社交媒体的那套“信任与安全”（Trust and Safety）议程全盘移植到 AI 领域。 萨克斯警告称，这种政治算计极其危险。即便禁令在两三年后解除，从项目重新申报、许可获批到施工并网，也需要耗费数年。因此，纽约州的禁令绝非暂停几个月那么简单，它意味着至少在未来五年内，纽约州都不会再有任何新的大型数据中心并网投产。这直接导致了数据中心资产定价曲线的急剧扭曲。随后查马斯提到，他与合伙人内塔（Nita）手中积压了大量数据中心重资产，在与各大超大规模云服务商（hyperscalers）洽谈定价时，深切感受到了这种政策扭曲带来的巨大冲击。 ### [01:07:30](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=4050s) 算力追逐电力：全球选址困境与“奢侈型监管” 查马斯（Chamath Palihapitiya）指出，在评估当前数据中心项目是保留、自建还是直接出售时，最令人震惊的现象是电力曲线上前端价格的极端暴涨——只要你手里掌握着“今天就能立刻通电、经过验证的现成电力”（verifiable energizable power today），就能卖出天价。正如萨克斯（David Sacks）所强调的，放眼未来，大约有 40% 的数据中心规划项目目前正被迫搁置甚至取消。这直接导致了驱动 AI 发展的可用能源出现巨大缺口。查马斯直言，无论大家多么渴望利用 AI 实现新药研发、癌症诊断、更优质的医疗服务还是普惠的法律咨询，由于缺乏足够电力支撑，未来可能根本无法满足这些庞大的市场需求；而电力短缺的根本原因，正是许多人出于对未知的本能抗拒，正在盲目抗议他们根本不理解的技术。 面对这一僵局，杰森（Jason Calacanis）打趣道，未来将出现“GPU 到处追着电力跑”的滑稽景象。萨克斯进一步补充了更深层的地缘政治矛盾：当前的政治阻力不仅试图阻止美国本土建设数据中心，还在通过层出不穷的先进芯片出口管制，阻碍盟友和伙伴国家建设数据中心。芯片出口限制被施加给越来越多的国家，导致本土不让建、盟国也没法建。查马斯对此分享了自己的实操体会：中东地理位置优越，理论上可以在 200 毫秒的超低延迟内覆盖全球 40 亿人口，虽然此前的扩张步伐一度放缓，但随着本周阿联酋（UAE）重新获准进口顶尖先进制程芯片，中东势必再次成为极具吸引力的算力枢纽；与此相对，澳大利亚等看似环保法规极其严格的国家，反而在积极拥抱并促成了大型数据中心交易。 杰森尖锐地将纽约和加州出台的阻挠政策定性为“奢侈型监管”（luxury regulations）——只有财政底气雄厚或自命不凡的地区，才有资格为了政治正确摆出“我们不需要数据中心”的高姿态。相比之下，澳大利亚需要产业投资，得克萨斯州渴望资金流入；虚伪的道德标榜在现实经济面前总有极限，一旦债务占 GDP 比重飙升、生产力停滞或外商直接投资骤降，地方政府最终只能向现实低头。杰森还补充道，伴随星链（Starlink）最新超大企业级终端的演进，多机捆绑可提供 10G 到 20G 的超大带宽，这甚至意味着算力中心未来可以摆脱传统光纤网络的束缚，部署在全球任何能获取廉价能源的偏远角落。 - **项目搁浅比例**：目前约有 40% 的规划数据中心项目因审批与反弹被搁置（mothballed）。 - **区位响应优势**：中东地理位置可在 200 毫秒延迟以内覆盖全球 40 亿人口的计算需求。 - **卫星回传突破**：新一代企业级星链捆绑系统可提供 10G 至 20G 级别的吞吐能力，大幅拓宽了离网部署数据中心的可行性。 --- ### [01:11:24](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=4284s) 现代认知战与反技术狂潮：从反转基因到反数据中心 弗里德伯格（David Friedberg）展示了一张耐人寻味的历史数据图表，用以剖析美国社会眼下对数据中心的狂热敌意到底从何而来。他回顾了美国反转基因（anti-GMO）舆论的演变：转基因作物自 1996 年在美商业化推广以来的 14 年间，民间本无系统性的反对声音；然而，自俄罗斯官方外宣媒体《今日俄罗斯》（Russia Today, RT）于 2010 年在美国落地运营后，反转基因舆论迅速引爆。到了 2022 年拜登政府下令取缔 RT 在美播发时，Google 搜索趋势清晰显示，反转基因话题的热度随着 RT 的传播起伏而同步波动。RT 当年大量炮制反转基因伪科学，通过各类“宝妈育儿博主”（mom bloggers）扩散至社交媒体算法流，最终演变成全美大众不假思索全盘接受的“普遍真理”。 弗里德伯格指出，这种手法与冷战时期克格勃（KGB）设计的“积极措施”（active measures，或称定向影响行动）如出一辙，其核心就是通过境外媒体或代理人向西方民主体制精准投放认知毒素，诱导对手自废武功；德国自毁长城的全面弃核运动，背后也充斥着类似的影子。如今，这套熟悉的剧本在美国数据中心议题上卷土重来：反对建厂的非政府组织（NGO）背后有着错综复杂的资金链，媒体争相炒作反科技焦虑，大众则被情绪裹挟。最新民调显示，超过 50% 的美国人盲目相信数据中心会推高居民的水电账单；哪怕数据中心百分之百实现了水资源循环闭环，并且自建清洁发电设施，公众依然表现出强烈的“本能排斥”（repugnant reaction）。 ```mermaid flowchart TD accTitle: 现代认知战传播链路 accDescr: 展示境外定向舆论如何渗透母婴社群并最终诱导本国制定自毁性公共政策的流程图。 subgraph Origin["源头策划与投放"] A["境外地缘力量 / 宣传机构"] --> B["制作反技术伪科学叙事<br/>（转基因、核电、数据中心）"] end subgraph Amplification["舆论渗透与放大"] B --> C["外围媒体与社群博主<br/>（如宝妈博主、激进环保NGO）"] C --> D["社交媒体算法流扩散与病毒式重推"] end subgraph Impact["社会撕裂与政策反噬"] D --> E["公众形成非理性恐慌<br/>（民调显示逾50%民众盲目反对）"] E --> F["出台自杀式法案与禁令<br/>（州级数据中心暂停令、算力停摆）"] end ``` --- ### [01:14:46](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=4486s) 境外渗透疑云与美国 AI 基建的生死时速 针对这种看似非理性的社会思潮，查马斯与萨克斯进一步撕开了其背后的地缘利益角力。萨克斯直言，这绝非单纯的民间环保焦虑，而是高度组织化的境外干预。仅仅一个月前，OpenAI 便公开发布了一篇重磅调查报告，题为《与中国相关的虚假影响力行动正在针对美国 AI 政策辩论》（PRC-linked influence operations are targeting AI debates in the US），《Politico》等多家主流媒体对此进行了深入报道。报告揭露有组织的网络水军与认知行动正系统性地挑动美国民间与监管层对数据中心的敌意。 萨克斯分析，这种利用民主制度开放性制造内耗的战略极其符合对手的国家利益——如果能够成功阻挠美国本土及盟国快速落地至关重要的电力与算力基础设施，甚至诱导以 Anthropic 为代表的企业推动各州层层加码监管限制，那么对手无需在芯片硬件上正面突围，便能借由美国内部的自我瘫痪，在决定未来的全球通用人工智能（AGI）世纪竞赛中不战而胜。据悉，美国国会即将就此类针对前沿科技基建的外国影响力渗透展开专项调查。 ### [01:15:52](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=4552s) 封杀开源的代价：成本鸿沟与企业数据的二次掠夺 几位主持人深入探讨了美国若通过监管“抽掉梯子”、扼杀开源 AI 所带来的灾难性经济后果。如果政策层面上限制了低成本模型的选择权，美国企业将被迫承担畸高的成本：输入端每百万 Token 可能要支付高达 56 美元，而世界其他地区的成本仅需 50 美分。正如嘉宾们直言，在面对 50 到 100 倍的成本劣势时，哪怕海外竞争对手的模型质量略逊一筹，美国本土企业也毫无胜算可言，“这在数学账面上根本算不通”。更致命的是，企业不仅在用真金白银买单，还在把自身的专有知识和核心竞争优势（Alpha）拱手“喂养”给闭源的前沿大模型厂商，无异于抵押自己的商业未来。 与会者进一步指出，外国势力存在巨大的地缘政治动机去操控和影响美国的舆论与政策走向。如果美国决策层和公众无法从第一性原理出发看清这一点，就极易掉入别有用心的舆论陷阱，在恐慌中自废武功，自愿戴上阻碍创新的监管枷锁。 - **成本差距悬殊**：高昂的闭源专有模型（56 美元 / 百万 Token）对比全球廉价或开源模型（0.50 美元 / 百万 Token），存在高达 50 至 100 倍的商业套利空间。 - **双重剥削机制**：大企业既要承受高额 API 账单，又在交互过程中被迫输出独家专有数据，长期来看将丧失独立竞争护城河。 --- ### [01:17:08](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=4628s) AI 道德恐慌与监管陷阱：从 FINRA 提议到 Dario 的“末日焦虑” 大卫·萨克斯（David Sacks）尖锐地指出，美国当前正陷入一场空前的“AI 道德恐慌”。公众与政客整天挂在嘴边的灾难假说——无论是大规模网络攻击、超级毁灭还是工作岗位归零——至今没有任何一个真正发生。然而，社会的“奥弗顿之窗”（Overton window）却已经被推向了危险边缘：大家甚至开始认真讨论引入类似金融业 FINRA（金融业监管局）式的自律监管机构。“FINRA 无非就是一群华尔街股票经纪人关起门来给自己定规矩，这个行业上次出现真正的创新是什么时候？难道就凭 Robinhood 搞了个零佣金吗？”萨克斯警告，如果照搬这种模式，美国将亲手葬送赖以繁荣的自由市场创新文化与快速迭代机制。 与此同时，大洋彼岸的追赶势头极其凶猛。最新亮相的中国模型（如 Moonshot 推出的 Kimi K3）在基准测试中已经逼近全球最顶尖的前沿水平，中美之间的代际领先优势可能仅剩数月时间。然而美国却在为尚未发生的风险疲于设立复杂的准入机制与官僚机构。查马斯（Chamath Palihapitiya）与杰森（Jason Calacanis）讽刺了 Anthropic 首席执行官达里奥·阿莫代伊（Dario Amodei）长久以来的末日论调，回顾他在一年前曾宣称“未来 1 到 5 年内将有 50% 的初级知识型岗位被消灭”，甚至从 GPT-2 时代起就处于持续的焦虑状态中。萨克斯对此犀利总结道，这种恐惧营销本质上是一套精妙的商业闭环：“第一步，把自己标榜为‘唯一安全’的 AI 公司；第二步，呼吁政府立法封杀‘不安全’的竞品；第三步，垄断市场坐享暴利。” ```mermaid flowchart LR accTitle: AI安全营销与监管闭环 accDescr: 展示部分大模型公司通过宣扬末日恐慌来达成行业垄断与利润闭环的三步策略 subgraph Strategy["监管俘渴闭环三部曲"] A["打造人设<br/>将自身品牌绑定为'唯一合规安全'"] --> B["游说立法<br/>借助灾难恐慌呼吁封杀'不安全'竞品"] B --> C["垄断收割<br/>清除开源与廉价竞争对手获取暴利"] end ``` - **算力红线失效**：早前监管者曾提议对训练算力超过 $10^{25}$ FLOPs 的模型实行严格审批许可，认为该算力阈值可能诞生“天网”（Skynet）或“终结者”，而如今各家主流模型均已大幅超越这一界限，现实中却未出现任何实质性浩劫。 - **中国自动驾驶应对**：与美国盲目立法形成对比，中国面对无人驾驶出租车引发的底层就业冲击与社会摩擦，采取了更为务实的暂停发放新增牌照措施，而非在技术研发层面设置自缚手脚的官僚枷锁。 --- ### [01:21:33](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=4893s) 峰会预热、领养小狗插曲与“埃菲尔铁塔”退场风波 在严肃的政策讨论之余，现场画风一转进入轻松插播环节。主持人首先为即将在 9 月举办的 2026 年度 All-In 峰会打广告，提醒听众登录官网申请门票及奖学金名额。紧接着，杰森在节目中替北加州索诺马（Sonoma）的一家流浪狗救援机构发声，深情呼吁听众领养一只失去寄养家庭、名叫 Ronnie 的待领养犬。查马斯随即打趣杰森借救狗刷“道德积分（Q-points）”，随后更开起极不严肃的跨文化地狱笑话，调侃起“斯里兰卡式腌狗肉配方”，在现场引发一片哄笑与吐槽。 在一阵插科打诨中，萨克斯因巴黎行程匆忙告退，自嘲必须赶去“埃菲尔铁塔”，将麦克风全权交给了准备主持招牌科普板块“科学角（Science Corner）”的大卫·弗里德伯格（David Friedberg）。 --- ### [01:22:52](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=4972s) 科学角：逆转衰老新机制——破解细胞外基质的糖化沉淀 弗里德伯格接管本期的科学角，聚焦发表在《自然·通讯》（Nature Communications）上的重磅抗衰老前沿研究。不同于过去讨论的“山中因子”（Yamanaka factors）通过细胞内重编程让细胞返老还童，这项由谷歌旗下行事极其隐秘的抗衰老子公司 Calico 携手初创公司 Revel Pharmaceuticals 共同发表的论文，将目光投向了长期被忽视的领域——**细胞外基质（Extracellular Matrix, ECM）**的衰老机制。 弗里德伯格解释道，人体衰老很大程度上源于细胞外部环境的物理性病变。随着年龄增长，游离的糖分与脂肪分子会持续且不可逆地与细胞间隙中的胶原蛋白及弹性蛋白结合，发生**糖化反应（Glycation）**。年轻时，人体的微环境通畅清洁，受损或老化的蛋白质能够被高效代谢和更新；而随着糖化终产物不断积聚，细胞之间会形成黏稠、僵硬的交联物质，如同“黏糊糊的污垢”彻底堵塞组织间隙，导致身体自身无法再对其进行清洗与修复。这种微观层面的结构硬化，直接表现在宏观上的皮肤长出深层皱纹、关节活动受限及器官纤维化硬化，从而构成了生物体机能衰退的关键诱因。 - **研发主体**：谷歌母公司 Alphabet 旗下的长寿先锋机构 Calico 与专注于清除交联分子的 Revel Pharmaceuticals 联合团队。 - **细胞内 vs 细胞外抗衰**： - *山中因子（细胞内机制）*：向细胞核导入转录因子重置表观遗传时钟，恢复干细胞活性； - *细胞外基质策略（细胞外机制）*：靶向清除因糖化反应黏附在间质蛋白上的“糖-脂复合物”，恢复结缔组织的柔韧性与代谢通道。 ### [01:24:30](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=5070s) 逆转衰老的科学突破：利用 AlphaFold 与定向进化降解 CML 蛋白 大卫·弗里德伯格（David Friedberg）在“科学角”（Science Corner）深入拆解了一篇重磅科研成果。随着人体逐渐衰老，糖分会与蛋白质结合，改变其空间结构和形态，形成所谓的“晚期糖基化终末产物”（Advanced Glycation End Products，简称 AGEs）。最典型的后果就是胶原蛋白这类原本彼此分离的纤维发生异常交联而黏在一起，不仅导致皮肤长出皱纹、关节丧失活动灵活性，还会引发慢性炎症——因为免疫系统会将这些变形的蛋白质识别为异常异物并持续发起攻击，这正是人体越老体内炎症越严重的根本原因之一。在所有 AGEs 中，最重要的分子被称为 CML（羧甲基赖氨酸），它是驱动细胞外基质（ECM）衰老的核心推手，而自然界中原本没有任何生物酶能够将其分解。 为了攻克这一生物学死结，科学家们决定从零设计一种能够特异性分解 CML 的新型酶。蛋白质本质上是由 DNA 编码的三联体密码子所决定的氨基酸链，现代蛋白质工程可以通过化学合成 DNA，再利用工程菌合成目标蛋白进行高通量筛选。研究人员首先利用 DeepMind 的 AlphaFold 预测结构，在细菌数据库中锁定了一种能够初步结合 CML 并具备微弱催化潜力的天然蛋白质骨架。随后，团队采用“定向进化”（Directed Evolution）技术，借助 DNA 重编程不断修改蛋白质的关键结合位点，历经 5 轮递归式迭代与突变筛选，测试了成千上万种变体，使其催化降解活性大幅跃升。 在试管实验中，这种人工设计的全新工程酶对人体多种重要蛋白质上的 CML 展开了测试，包括酪蛋白（casein）、胶原蛋白（collagen）、视网膜蛋白（retinal proteins）以及血红蛋白（hemoglobin），结果成功消除了 52% 到 97% 的 CML，在多个位点的降解效率甚至突破了 90%。更令人震惊的是随后的离体人体组织实验：研究团队获取了年龄大于 70 岁的老年人捐赠的皮肤样本，将该酶施用于皮肤后，成功清除了 55% 的 CML。“这几乎直接将这些 70 岁以上老年人的皮肤年龄逆转到了 31 岁的水平，”弗里德伯格难掩赞叹地强调。这是 AlphaFold、高通量筛选与定向进化技术珠联璧合的开创性壮举，成功创造出一种自然界此前从未存在过、却对人类健康具有深远意义的全新蛋白质。 ```mermaid flowchart TD accTitle: 逆转衰老酶的设计与验证流程 accDescr: 展示从 AlphaFold 发现前体蛋白到定向进化再到人体皮肤验证的研发全流程 subgraph Discovery["蛋白发现与计算"] A["锁定靶点: 细胞外基质中的 CML"] --> B["AlphaFold 结构预测与细菌筛选"] B --> C["获得初始结合蛋白前体"] end subgraph Optimization["高通量工程优化"] C --> D["DNA 重编程与突变库构建"] D --> E["5 轮递归定向进化筛选"] E --> F["获得高活性人工工程降解酶"] end subgraph Validation["体外与人体皮肤验证"] F --> G["胶原蛋白/视网膜/血红蛋白测试: 消除 52%~97% CML"] G --> H["70+ 岁捐赠者皮肤测试: 消除 55% CML"] H --> I["皮肤生物学指标逆转回 31 岁状态"] end ``` - **病理核心 CML**：羧甲基赖氨酸（CML）是衰老过程中细胞外基质蓄积的最主要交联分子，会彻底改变胶原蛋白结构并不断激活内源性免疫炎症反应。 - **5 轮递归进化**：科研人员利用 DNA 序列修改合成数千个蛋白质变体，历经 5 次循环迭代，使酶对 CML 的降解活性实现数量级提升。 - **逆转 40 岁表型**：涂抹于 70 岁以上捐赠者的真实皮肤组织后，降解了 55% 的蓄积 CML，直接让组织关键分子指标年轻化至 31 岁状态。 - **给药路径的未知数**：目前科研团队正着手解决递送难题——未来是制成外用透皮乳霜、局部注射液、口服补剂，还是通过注射 mRNA 让身体自身细胞合成并由内而外清除糖化交联物。 ### [01:27:36](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=5256s) 商业化前景与 AI 的现实价值：万亿美元抗衰美容与终极告别 听到这一惊人数据后，全场瞬间沸腾。杰森·卡拉卡尼斯（Jason Calacanis）立刻打趣称自己现在髋关节和肩膀到处酸痛，“你随时都能真切感受到自己正在变老”。弗里德伯格当即指出，治疗关节退行性变虽然前景广阔，但绝对不会是该技术的第一商业化切入点；首个落地场景毫无疑问是医美抗衰护肤，而这将是一个无可估量的万亿美元赛道。查马斯·帕里哈皮蒂亚（Chamath Palihapitiya）对此深表认同，兴奋地表示：“如果真的能把这种酶做成涂在皮肤上就能被吸收的乳霜，游戏就彻底结束了，单单这一项产品就价值两万亿美元（$2 trillion）！” 除了惊人的商业空间，这起突破更是为整场播客关于 AI 监管与价值的讨论画上了最具说服力的句号。弗里德伯格动情地指出，当舆论还在为各种地缘政治认知战（psyop）或 AI 监管风险争吵不休时，AlphaFold 对新蛋白质的创造性发现就是 AI 赋能人类福祉最无可辩驳的铁证——它跳过了大自然数亿年的漫长演化，直接设计出治愈疾病、逆转衰老的工具，是整个人类社会都能共享的红利。“在我们这个时代能够拥有这样的工具，实在是太深邃、太不可思议了。” 节目在热烈而温馨的气氛中走向尾声，四位合伙人互相表达了对彼此的真挚爱意与默契。伴随着标志性的告别台词“让你的赢家继续奔跑”（Let your winners ride），播客切换到由社区粉丝二创的一幕幕经典鬼畜与玩梗片段剪辑，在充满幽默与解构趣味的混剪音乐声中，本期播客圆满落下帷幕。 - **商业化第一梯队**：医美与抗衰老外用护肤品将先于骨关节炎或心血管等医学疗法上市，市场规模预估高达 1 万亿至 2 万亿美元。 - **AI 实用价值的终极样本**：AlphaFold 在真实生命科学与长寿技术上的里程碑应用，证明了生成式计算设计对实体世界的革命性反哺，远超纯软件虚拟交互的意义。 - **片尾温情与社区文化**：在四位主持人的告别中，节目以开源给全网粉丝二创的经典集锦（涵盖“衣帽商”、“性张力”、“走向全仓”等梗）作为片尾彩蛋收官。 ## Notable quotes > [01:32](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=92s) 整个行业必须自律监管，由谷歌、微软、亚马逊和马斯克共同制定测试标准并自我认证，而不是让完全不懂技术的政府插手 — 杰森·卡拉卡尼斯 > [04:18](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=258s) 避免了直接将技术生杀大权拱手交给政府，而是让专业的人做专业的事 — 戴维·弗里德伯格 > [28:54](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=1734s) 并购交易重回菜单（M&A is back on the menu） — 杰森·卡拉卡尼斯 > [39:50](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=2390s) 无论何时，任何人都不应该从前雇主那里偷东西 — 查马斯·帕利哈皮蒂亚 > [39:50](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=2390s) 当你跳槽去新公司时，唯一被允许带走的只有你脑子里的东西、你的经验与记忆；文档、U盘、光盘以及任何物理或数字资料，哪怕一片纸都绝对不能碰 — 戴维·萨克斯 > [39:50](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=2390s) 两点成一线，大家心里都有数（Couple dots make a line） — 戴维·弗里德伯格 > [01:07:30](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=4050s) 今天就能立刻通电、经过验证的现成电力 — 查马斯·帕利哈皮蒂亚 > [01:17:08](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=4628s) 第一步，把自己标榜为‘唯一安全’的 AI 公司；第二步，呼吁政府立法封杀‘不安全’的竞品；第三步，垄断市场坐享暴利 — 戴维·萨克斯 > [01:24:30](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=5070s) 这几乎直接将这些 70 岁以上老年人的皮肤年龄逆转到了 31 岁的水平 — 戴维·弗里德伯格 > [01:27:36](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=5256s) 如果真的能把这种酶做成涂在皮肤上就能被吸收的乳霜，游戏就彻底结束了，单单这一项产品就价值两万亿美元！ — 查马斯·帕利哈皮蒂亚 > [01:27:36](https://www.youtube.com/watch?v=9IMwRIei-Xc&t=5256s) 让你的赢家继续奔跑（Let your winners ride） — 杰森·卡拉卡尼斯 ## Mentioned 杰森·卡拉卡尼斯 (Jason Calacanis), 查马斯·帕利哈皮蒂亚 (Chamath Palihapitiya), 戴维·弗里德伯格 (David Friedberg), 戴维·萨克斯 (David Sacks), 戴米斯·哈萨比斯 (Demis Hassabis), 埃隆·马斯克 (Elon Musk), 萨姆·奥尔特曼 (Sam Altman), 达里奥·阿莫代伊 (Dario Amodei), 米拉·穆拉蒂 (Mira Murati), 萨蒂亚·纳德拉 (Satya Nadella), 桑达尔·皮查伊 (Sundar Pichai), 杰克·多西 (Jack Dorsey), 黄仁勋 (Jensen Huang), 马克·扎克伯格 (Mark Zuckerberg), 蒂姆·库克 (Tim Cook), 凯西·霍楚尔 (Kathy Hochul), 莉娜·汗 (Lina Khan), 唐纳德·特朗普 (Donald J. Trump), DeepMind, OpenAI, Anthropic, Stripe, Block, PayPal, Advent International, Apple, xAI, Grok Build, Grok 4.5, Thinking Machines, Inkling, AlphaFold, Calico, Revel Pharmaceuticals, Ramp, FINRA, NFA, SEC, FAA, Visa, Mastercard, PYUSD, Bridge, Venmo, Cash App, PJM, Bloom Energy, Starlink, CML (羧甲基赖氨酸), Nature Communications ## Video description (0:00) Bestie intros! (1:32) New AI regulatory proposal: DeepMind's Demis Hassabis proposes FINRA-type body (20:01) Stripe, Block, and Advent offer $53B to acquire PayPal (37:51) Apple sues OpenAI, alleging stolen trade secrets (42:49) Grok Build data leak, AI data privacy, Tokenmaxxing update, Mira Murati's new model (59:53) NY bans datacenters, becoming first state to enact a moratorium (1:22:57) Science Corner: New data on reversing aging! Adopt Ronnie the Dog: https://www.instagram.com/reels/Da0pGahBwaW Apply for All-In 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://x.com/chamath/status/2077502408528212144 https://x.com/demishassabis/status/2076957440109625718 https://x.com/satyanadella/status/2076323181154230284 https://x.com/Jason/status/2076231055443440105 https://x.com/SquawkCNBC/status/2077741908391031246 https://thinkingmachines.ai/news/introducing-inkling https://www.politico.com/news/2026/07/15/inside-anthropics-state-by-state-plan-to-ratchet-up-ai-rules-00998415 https://x.com/politico/status/2077315780144996633 https://x.com/DavidSacks/status/1978145266269077891 https://www.reuters.com/business/finance/stripe-advent-offer-buy-paypal-more-than-53-billion-sources-say-2026-07-15 https://www.tipranks.com/news/the-fly/block-contributing-to-equity-for-paypal-takeover-bid-cnbc-says-thefly-news https://9to5mac.com/wp-content/uploads/sites/6/2026/07/Apple-Inc.-v.-Liu-et-al.pdf https://finance.yahoo.com/technology/ai/articles/apple-lawsuit-threatens-openais-hardware-215438163.html https://x.com/markgurman/status/2076306380583997665 https://www.engadget.com/2216186/elon-musk-bought-a-gas-turbine-company https://x.com/Reuters/status/2076957424339050839 https://x.com/teddyschleifer/status/2077596563380072694 https://x.com/teddyschleifer/status/2077606887055306879 https://openai.com/index/prc-linked-influence-operations-ai-debates https://www.politico.com/news/2026/06/10/openai-china-ai-data-centers-report-00957612 https://trends.google.com/explore?q=GMO%2C%2Fm%2F0dkz0z&date=2010-01-01%202026-07-15&geo=US https://www.nature.com/articles/s41467-026-75141-2 #allin #tech #news [Source](https://www.youtube.com/watch?v=9IMwRIei-Xc)

[1:32] 杰森·卡拉卡尼斯: 整个行业必须自律监管，由谷歌、微软、亚马逊和马斯克共同制定测试标准并自我认证，而不是让完全不懂技术的政府插手 [4:18] 戴维·弗里德伯格: 避免了直接将技术生杀大权拱手交给政府，而是让专业的人做专业的事 [28:54] 杰森·卡拉卡尼斯: 并购交易重回菜单（M&A is back on the menu） [39:50] 查马斯·帕利哈皮蒂亚: 无论何时，任何人都不应该从前雇主那里偷东西 [39:50] 戴维·萨克斯: 当你跳槽去新公司时，唯一被允许带走的只有你脑子里的东西、你的经验与记忆；文档、U盘、光盘以及任何物理或数字资料，哪怕一片纸都绝对不能碰 [39:50] 戴维·弗里德伯格: 两点成一线，大家心里都有数（Couple dots make a line） [67:30] 查马斯·帕利哈皮蒂亚: 今天就能立刻通电、经过验证的现成电力 [77:08] 戴维·萨克斯: 第一步，把自己标榜为‘唯一安全’的 AI 公司；第二步，呼吁政府立法封杀‘不安全’的竞品；第三步，垄断市场坐享暴利 [84:30] 戴维·弗里德伯格: 这几乎直接将这些 70 岁以上老年人的皮肤年龄逆转到了 31 岁的水平 [87:36] 查马斯·帕利哈皮蒂亚: 如果真的能把这种酶做成涂在皮肤上就能被吸收的乳霜，游戏就彻底结束了，单单这一项产品就价值两万亿美元！ [87:36] 杰森·卡拉卡尼斯: 让你的赢家继续奔跑（Let your winners ride）

[0:00] Jason Calakanis: I'm Greek. We never spent time in a prison. [7:01] David Sacks: We'll just simply lose the AI race if that happens because China is not going to abide by those rules [16:11] David Sacks: when you go to the government and say please regulate me... there's hardly anyone in government who will ever say oh no no no we're not qualified. [20:00] Chamath Palihapitiya: You are creating a competitor to Visa and Mastercard because you now have upwards of 6 or 700 million accounts. You have massive stablecoin infrastructure. You have all of the risk-management infrastructure that Stripe has built over the last 15 or 20 years. [29:19] David Sacks: M&A is back on the menu. [32:02] David Sacks: Our homeland was taken over and they burned our temple and then kicked everybody out. [34:24] David Sacks: If somehow you could combine the merchant relationships with all those consumer accounts and then bypass the credit card networks cuz there'd be a lot—in theory there could be a lot more on-us transactions... that's where the value is. [37:52] David Sacks: The only thing you can bring to your new job is what's in your head... No thumb drives, no CDs, no documents, no nothing. [44:42] Chamath Palihapitiya: money burning furnace [67:13] Jason Calakanis: We're going to have GPUs chasing energy [71:22] David Sacks: It makes a lot of sense because it is in their interest [76:41] David Sacks: We may have months on China, if that

[0:00] All right, everybody. Welcome back to [0:02] the world's greatest podcast. The number [0:04] one podcast, your favorite podcast, the [0:06] All-In podcast. I'm Jason Calakanis, the [0:08] world's greatest moderator. With me, of [0:10] course, Chimath Polyhapitive, [0:13] >> the great great great great great [0:16] grandchild, Jason was of a hooker and [0:20] >> you saw that from France [0:22] >> and a purse snatcher. [0:25] >> This this comes from like a history of [0:28] uh France. some history [0:30] >> uh said they let you out of this was [0:33] this was the deal in 1719 Sachs if [0:36] you're a prisoner in a in Paris you were [0:38] offered your freedom on the condition [0:41] that you marry a prostitute and move to [0:43] the great state of Louisiana what are [0:46] you saying taking the deal [0:47] >> it explains a a certain of your [0:50] proclivities Jake I I [0:52] >> I thought you were asking me to see if [0:53] they would extend the rule for you [0:55] >> no I'm saying that your great great [0:57] great great great Grandma, I'm not a [0:59] hooker. I'm Greek. [1:00] >> That's what I'm saying very explicitly. [1:02] She was [1:02] >> We never We never spent time in a [1:04] prison. Also, of course, David Freeberg [1:07] is here. How you doing, brother? [1:09] >> Living the dream. [1:10] >> Good to be back. Missed you guys last [1:11] week. How was Brad? How did he fill in? [1:14] >> It was great. Yeah. Yeah, he was. [1:16] >> Trump account victory lap. [1:18] >> He had a little victory lap. We played [1:19] Chariots of Fire. And uh how was your [1:22] special time at blank and your time next [1:27] week at blank? [1:29] >> Thanks for h Thanks for having me on [1:30] your show, Jel. [1:32] >> All right, we got a full docket today. [1:34] Uh lots of stories. Let's start with uh [1:36] Deep Mind's Deis Habis just dropped an [1:40] AI regulation proposal and it's um it's [1:44] pretty popular with the boys. In an ex [1:47] article, Demis called for a US-led [1:50] international AI standards body proposal [1:53] is modeled after FINRA, the Financial [1:55] Industry Regulatory Authority. [1:58] That's a self-regulatory body. And this [2:01] would be federally overseen, but [2:03] industry funded and run by independent [2:06] technological experts. Frontier Labs [2:08] would submit their models 30 days before [2:10] release. Uh, and it would be voluntary [2:13] initially, then mandatory. At some [2:15] point, the models would be assessed on [2:17] risk to cyber security, [2:20] national security, biological threats, [2:22] and other high-risisk domains. [2:24] Benchmarks would be updated quarterly, [2:26] and the body can coordinate a slowdown [2:29] in development if the situation demands [2:31] it. I guess that would be if uh there [2:33] was a cyber risk, etc. Looks like uh on [2:36] the positive side we have Elon who said [2:39] it was thoughtful, Sam at OpenAI, [2:42] Jack Clark at Anthropic, Sundar, [2:46] Satia, Jack Dorsey from Block, the [2:50] Carlson brothers. So Freeberg um [2:55] uh Freeberg, your thoughts on [2:58] >> Oh, wow. Way to really way to really uh [3:00] put in the effort today. Go ahead, [3:02] Jason. Good. [3:03] >> Just pass it to me. I'll take care of [3:05] it. Let me let me [3:06] >> What do you want me to do? Oh my god, [3:07] what an incredible topic. [3:09] >> All right, here. Let me do it. You want [3:10] me to do it, right? Three, two, [3:11] >> some of us actually care about this [3:13] topic. Let's go. [3:13] >> Okay. Yeah, you care about it. Okay, [3:14] here we go. Three, two. Here's a clip of [3:17] me calling it on the All-In podcast [3:20] first. The whole industry is going to [3:22] need to be regulated and I think the [3:24] industry needs to regulate themselves. [3:26] That's the key to this. We need to have [3:28] a set of tests that Google, Microsoft, [3:32] Amazon all agree to. Elon, hey, these [3:35] are the things we should test and they [3:36] should self-certify each model before [3:38] asking the government which doesn't [3:40] understand the models to certify them. [3:42] The industry should have an industry [3:44] certification like they do for countless [3:45] other things. I've talked about the MPAA [3:48] and the video game industry. We should [3:50] just self-certify. It's the simplest [3:51] thing in the world to do. And then we [3:54] could release the models ourselves [3:55] without the government getting involved. [3:57] >> Brie, would you like to congratulate me [3:59] on nailing it again? Well, first of all, [4:01] first of all, I thought that Demis' [4:04] proposal was really smart and [4:06] thoughtful. Now that I know that you may [4:08] have shared the same thought, I think we [4:09] should just do something different. [4:11] >> I can't win, Zach. Even when I nail it, [4:13] I hit a halfcourt shot. Truman's like, [4:15] "Move the net. Move the net." [4:18] >> I think it's worth putting a little [4:20] definition around this proposal, which [4:22] is to form an SRO, self-regulatory [4:25] organization, because they're not purely [4:26] independent. SRO's like FINRA and the [4:29] National Futures Association, they exist [4:32] in the financial markets and they were [4:33] created to allow the financial [4:36] institutions to set their regulatory [4:38] rules, how they check each other, how [4:40] they make sure that everyone is being [4:42] safe because they're obviously all [4:43] trading risk with one another. So the [4:45] industry doesn't want to have exposure [4:47] and they certainly don't want to have [4:48] things get slowed down because that [4:50] would make the markets inefficient. So [4:52] the analogy with AI is pretty [4:54] appropriate here which is that there are [4:56] many players in the industry. They are [4:58] all trying to progress AI technology and [5:01] no one wants to have a single regulatory [5:04] body that comes in from the government [5:06] or outside that says here are the tests [5:09] you guys have to pass with your models [5:11] in order for them to be appropriate. As [5:12] we saw in California when California [5:14] tried to pass AI legislation, I think it [5:16] was about a year a year and a half ago, [5:18] none of what they wrote even made sense [5:20] at the time. But fast forward a year, [5:23] none of those kind of rules and [5:24] requirements actually map to the [5:26] technology of the day. So the purpose of [5:28] an SRO like FINR run NFA is they can [5:30] adjust how tests are being run, who is [5:33] actually running the test and make sure [5:34] the right experts are involved in doing [5:36] this, independent experts that is to do [5:38] the testing with federal government [5:41] oversight but not control. So in the [5:44] case of FINRA NFA, they report up [5:47] ultimately to Senate committee and House [5:48] committee that gives those committees [5:50] oversight of those governing bodies that [5:53] are supposed to be doing the work to [5:54] make sure that they're doing their [5:56] frigin jobs. So the SRO concept would be [5:58] that experts could be brought in from [6:01] industry that know how to assess models [6:04] for things like cyber risk, for things [6:07] like biorisk, for things like weapons [6:09] risk, social manipulation, etc., etc. [6:12] that independent body can get voted on, [6:14] can get changed over time, and because [6:17] they actually have expertise in running [6:19] software valves and running tests like [6:20] this, they can operate at a faster pace [6:23] than setting up a new government agency. [6:25] So, it's it's kind of a very elegant [6:27] solution. And I think it's why everyone, [6:28] to your point, Jal, and I I'll say this [6:30] is right, the industry recognizes that [6:32] there needs to be some degree of [6:35] oversight and checkpoints here. And I [6:37] think that this could actually solve [6:38] that problem. So that's why I think [6:40] everyone's kind of climbing on board [6:41] with it because it doesn't actually hand [6:42] stuff over to the government. It says, [6:44] "Hey, we're going to get the right [6:45] people to to take a look at these things [6:46] and the government is going to have [6:48] oversight ultimately." [6:49] >> Did Anthropic and OpenAI have a point of [6:51] view? [6:51] >> They both signed up to it. I don't think [6:52] Daario directly, but Dario's president [6:57] gave his thumbs up and then I think Sam [6:59] gave his thumbs up [7:00] >> which means they're on board. Saxs, is [7:02] this the best of the uh possibilities in [7:06] your mind? uh is the industry regulating [7:08] itself after they have now provoked [7:12] governments around the world to be so [7:14] concerned about this issue. [7:16] >> Yeah. And I, you know, I talked to [7:18] Dennis about this and this may surprise [7:20] people, but I told him that I could [7:21] potentially get on board with this, [7:23] speaking just for myself, not on behalf [7:24] of anyone in the government because I [7:27] thought that an SRO, again, a [7:29] self-regulatory approach would be [7:30] infinitely better than creating a new [7:32] government agency that I think would [7:35] rapidly become a DMV for AI. Dario calls [7:38] it an FAA for AI. The government does [7:40] not have the expertise to evaluate AI [7:43] models. The criteria are changing too [7:45] rapidly. you're going to very rapidly [7:48] end up with a queue where all the models [7:50] would be waiting to get tested and it [7:52] would start with a month-long delay. It [7:54] would end up being many months and we [7:56] would just lose the AI race. So, I think [7:58] an SRO approach would be infinitely [8:00] better than that if it was done right. [8:03] And I outlined for Demis [8:07] criteria or conditions that I thought [8:09] were really important to in order to [8:10] make this work. And if I could, I'll [8:12] just run through them [8:14] >> please. [8:14] >> All right. So number one, I think the [8:17] SRO has to have broad representation [8:19] from within the industry, the AI [8:21] industry. It has to include startups and [8:24] open source. It can't just be the three [8:26] biggest labs, you know, can't just be [8:28] >> the fix, right? [8:29] >> Yeah. Exactly. And that's precisely to [8:32] avoid the problem of regulatory capture, [8:34] right? If you have a diverse enough [8:36] group of interests being represented, [8:38] it's much harder for this to turn into [8:40] red capture. So for example, I think if [8:43] you had Jensen, Elon, Zuck, and maybe [8:45] Mera because she just launched a very [8:48] interesting [8:49] >> open [8:49] >> platform that's based on open mach [8:55] capture problem to a large degree. So [8:57] that's number one. Number two is I think [9:00] that this body should only be reviewing [9:03] frontier models, meaning the true [9:06] frontier. The models that really [9:09] represent an advance in the [9:11] state-of-the-art of artificial [9:13] intelligence and models below this level [9:15] should not be held up from getting to [9:17] market. And I do think that is a big [9:19] risk under regulations is that the [9:21] leaders of the market use this as a way [9:23] to tie up lesser models. And there's no [9:25] reason if a model is not at the [9:27] frontier, why hold it up? Okay, so [9:28] that's [9:29] >> so they have to be in maybe the top 10 [9:31] performers, top 20 performers on the [9:33] benchmark test. [9:33] >> No, I think I think when they benchmark [9:35] on key dimensions of intelligence, they [9:37] have to represent an increase above [9:39] where the current state-of-the-art is if [9:41] it's not a step change, then how are you [9:43] dealing with some new incremental risk, [9:45] right? I mean, this is all about dealing [9:46] with some sort of incremental [9:48] catastrophic risk that could be [9:49] introduced by some new step change in [9:52] intelligence. And that brings me to [9:53] number three, which is I think this body [9:55] should be dealing with catastrophic risk [9:57] only. And to my knowledge, those right [10:00] now are cyber and CBRN, meaning it's, [10:03] you know, chemical, biological, [10:04] radiological, nuclear. So it should not [10:07] be things, for example, like [10:08] disinformation or microaggressions. This [10:11] should not become a speech regulator, [10:13] you know, or just, you know, things that [10:15] seem kind of trivial. The only reason to [10:17] have this is for truly catastrophic [10:19] risk. So that's number three. Number [10:21] four, and Deus mentioned this in his [10:23] post, is that I think it should be [10:26] voluntary first. This new organization [10:28] should prove it works before it gets [10:30] legally enshrined and becomes mandatory. [10:34] And then number five is this should be a [10:36] substitute for a new regulatory agency. [10:40] If it's just additive, then it defeats [10:42] the purpose and there's no real reason [10:44] to support it. So again, I think this [10:46] has to be a substitute, not an addition [10:48] to a bunch of new regulatory structures. [10:52] So I think if you did those five things, [10:54] I think this becomes much more [10:55] palatable. And Demis said, I mean, he [10:58] didn't put all these points in his blog [11:00] post, but he did say to me that he was [11:01] bas. [11:03] So I think if those notes were adopted, [11:06] this is something that we could [11:07] potentially get on board with. That [11:08] doesn't mean I don't still have [11:10] concerns. I'm quite concerned for [11:12] example that you know I think Daario has [11:15] expressed support for this however I [11:18] think this is just an opening bid for [11:20] anthropic meaning they'll take this [11:22] thank you very much this is more [11:23] regulation than we have today but that [11:25] won't be the end of it right this will [11:26] just be the stepping stone to get what [11:28] Dario has now called for many times [11:30] which is the FAA for AI and if I could [11:33] let me just as as a final point I just [11:35] want to explain what the FAA does [11:37] because people need to understand you [11:39] know FAA for AI sounds really nice, but [11:42] actually it's a really extreme proposal. [11:46] What the FAA does, among other things, [11:48] is approve new airplane designs. Okay? [11:51] And specifically, it requires what's [11:54] called a type certification for any new [11:56] aircraft design or major changes. And [11:59] for an entirely new aircraft design, it [12:02] takes 5 to 9 years [12:04] >> to get the certification. And if you [12:06] merely want to amend a certificate, I [12:09] guess for you know major changes or I'm [12:12] not even sure how major the changes need [12:14] to be, it takes three to five years. So [12:16] the Boeing 737 Max, for example, took [12:19] about 5 years. So this is [12:21] permissionbased regulation. There's no [12:23] approval, no flying commercially. It's [12:27] safety first. Look, that might make [12:28] sense in the case of preventing plane [12:30] crashes, but when you're talking about [12:32] AI models, you're talking about [12:34] replacing a system that is releasing new [12:38] versions every couple of months with one [12:40] that is potentially fully under the [12:42] control of government, fully government [12:44] approved. Everything has to be [12:45] certified, and you could expect the [12:47] timeline to go from months to years. [12:49] Again, I think we'll just simply lose AI [12:51] race if that happens because China is [12:53] not going to abide by those rules. So [12:55] just to sum up, if my choices are [12:57] between FAA for AI or what I would call [13:00] the DMV for AI, I would much rather go [13:03] for Demis' SRO for AI, the [13:05] self-regulatory approach. But we really [13:09] have to keep it honest and pure because [13:12] again otherwise it'll just be the [13:14] opening bid in a coming new wave of [13:16] regulation and it will be the vehicle [13:18] for massive regulatory capture. [13:20] >> To your point, it can't restrict. [13:22] >> Yeah. [13:22] >> Can't restrict open source. I think [13:24] that's so important because all of these [13:25] other efforts require money that you [13:28] have to spend which is always where [13:30] regulatory capture happens and you have [13:32] to enable startups and open source to [13:33] compete effectively. [13:34] >> Shimoth, any thoughts on this new [13:37] self-governing body? [13:39] >> I think it's really important and I hope [13:40] it happens quickly. The thing we have to [13:44] keep in mind is there's going to be a [13:47] torrent of money that's going to try to [13:50] influence both sides of the political [13:52] aisle to [13:55] regulate this in a way that creates some [13:58] form of regulatory capture. We just [14:00] don't know what. And so the faster we [14:03] avoid that off-ramp by actually [14:06] establishing a set of rules and [14:08] superseding the need for federal [14:11] oversight is a really important thing. [14:14] Now, at the end of the day, you still [14:15] have federal oversight in some ways [14:17] because you still have commerce that [14:19] plays a huge role in these standards. [14:21] You still have the DOJ. So, it's not as [14:23] if it's going to be a wild west. But [14:25] what it prevents is a handful of actors [14:28] using their balance sheets and their [14:31] capital to essentially pull the ladder [14:33] up. And I think if that happens, we're [14:35] in a really bad place. So I think Demis' [14:37] proposal makes a ton of sense and we [14:39] should just get on with it. [14:40] >> All right. [14:40] >> Well, yes. But provided I mean again [14:43] just provided that I think we we make [14:46] sure that I mean look in my view there [14:48] there are five conditions but I think we [14:49] do have to make sure it's pure. Even in [14:51] the FINRA example, that's the analogy [14:53] that Demis used was that we should set [14:55] this up in the same way that that FINRA [14:57] is set up. FINRA does report in [15:00] ultimately to the government. It reports [15:02] into the SEC. And so, you know, if we [15:05] are going to set this new SRO, where is [15:07] it going to report to in the government? [15:09] There's going to be a huge food fight [15:10] over that and it will then be subject to [15:14] political pressure. The software [15:15] industry has never been regulated in [15:17] that way. We do not have a dedicated [15:19] regulator for software. We're right, but [15:21] those issues are important. But my point [15:23] is, h I'm just saying these things are [15:25] never ideal. But if the choices are we [15:27] kill open source and we lad pull the [15:32] entire market, so there's a duopoly or [15:35] there's this, I'd say this [15:37] >> for sure. And and that's and that's [15:39] that's the crux of my argument is it's [15:40] definitely the lesser of two evils. I'm [15:42] not sure those are the only two choices, [15:44] but increasingly there's no question [15:46] that the pressure is coming to regulate [15:49] AI more and more. And frankly, this all [15:51] goes back to anthropics government. [15:54] >> They poke the tiger. They poked the [15:56] tiger. [15:56] >> Yeah. Well, it's it's more than that. [15:58] >> They're funding the meat the tiger. [16:00] >> Yeah. I want to give an update on that [16:02] actually cuz [16:02] >> I would say poking the tiger of like the [16:04] American public getting really freaked [16:06] out and then the government stepping in. [16:08] Yeah, [16:08] >> totally. And there's a couple of data [16:09] points on that actually. I just want to [16:11] give a quick update. So in October of [16:12] last year, I tweeted that Anthropic is [16:15] running a sophisticated regulatory [16:17] capture strategy based on fear-mongering [16:19] and everyone kind of went crazy over [16:21] this. This was again like a very hot [16:23] take or spicy take at the time. Back [16:25] then people thought that I was beating [16:27] up on a little startup. Now I think [16:29] everyone can kind of see the truth which [16:30] is look this is not a little startup. [16:32] They already have a trillion dollar [16:33] market cap valuation. Gavin Baker thinks [16:36] it'll be at three trillion after they [16:37] IPO. [16:39] This is actually one of the biggest of [16:41] the big tech companies and they are I [16:43] think by pretty much every criteria [16:44] including revenue the leading AI [16:46] company. So I think people can see now [16:48] that they have enormous resources and [16:50] they're putting those resources behind [16:52] an effort to like Jimoth you said pull [16:54] up the ladder and it's classic regatory [16:56] capture and there was an article in [16:58] Politico just the other day. is called [17:00] inside Anthropic state-by-state plan to [17:03] ratchet up AI rules. And what it says is [17:06] quote AI giant Anthropic is pursuing a [17:09] strategy of oneupmanship that encourages [17:12] states to impose increasingly tougher AI [17:15] guard rails rather than a line around a [17:18] single set of regulations. So the basic [17:20] idea is that they get a set of [17:22] regulations passed in one state like [17:24] California's SB53 [17:26] and that was then supposed to be the [17:28] model at least for all the blue states. [17:29] But then with each new state, they [17:31] actually make the regulations more and [17:34] more strict, more and more [17:35] all-encompassing. So there's not [17:37] actually a stable equilibrium. What [17:39] they're trying to do is drive each [17:41] incremental state to more and more [17:42] regulations. And so this is actually an [17:45] article explaining that they're doing [17:47] the opposite of trying to create what we [17:50] wanted, which was a single national [17:52] framework. they actually want the [17:54] patchwork because they're using again [17:56] the pressure they're creating at the [17:57] state level to impose more and more [18:00] regulations. And again, what I said last [18:01] year was that Anthropic was principally [18:05] responsible for the state regulatory [18:06] frenzy that is damaging the startup [18:09] ecosystem. Again, everyone went crazy at [18:10] the time. I think now there's plenty of [18:12] evidence showing this is their agenda. [18:15] And um [18:16] >> and by the way, they're going to win [18:17] that because states have great [18:19] sovereignty rights and like we're seeing [18:21] with self-driving the states are going [18:22] to decide. It's not going to be a [18:24] federal mandate. The states get to [18:25] decide just the nature of the US. [18:28] >> The reason why [18:28] >> they're going to they're going to win on [18:29] that in a couple of states, right, Zach? [18:31] Just realistically, [18:32] >> they've won in a bunch of states. They [18:33] already won in California and Illinois [18:35] and New York and I mean they're winning [18:37] in all the blue states and maybe even [18:38] some red states. But look ultimately the [18:40] reason why anthropics arguments are [18:42] finding purchase is because when you go [18:45] to the government and say please [18:47] regulate me you know you should have [18:49] more power there there's hardly anyone [18:51] in government who will ever say oh no no [18:53] no we're not qualified like we don't [18:55] want government [18:57] >> yeah there are very few people who are [18:59] principled that way and most people in [19:00] the government will say thank you very [19:02] much what else can we take and this is [19:04] the mistake that I think a lot of people [19:05] in the tech industry are making is they [19:07] think that they can just buy off [19:10] politicians or the political system by [19:12] making concessions. No, that will just [19:14] lead to a ratcheting up of the pressure. [19:16] The government will be happy to take [19:18] this and then come back for more and [19:20] more and more until it's fully under [19:22] government control. So, at some point, I [19:24] think these companies are going to have [19:25] to grow a spine and fight and decide [19:27] where they're willing to draw a line. [19:29] And if Demis' SRO is the line, if [19:33] they're saying, "Okay, we think this is [19:34] the right solution and we're going to [19:36] fight here and this has to be it and in [19:39] exchange for this we need preeemption [19:41] and we need, you know, other things [19:42] written into law that make sure this is [19:44] where the line is, then I think it can [19:46] work. But I think if you're just kind of [19:48] offering it up for free and all these [19:50] companies are just going to say, "Oh [19:51] yeah, regulates, give us the SRO." That [19:53] will not be the end of it. That will [19:54] just be the opening bid and the [19:56] government will come back to take more [19:58] and more and more. Yeah. All right, [20:00] let's keep moving through the docket [20:01] here. Stripe, which is still a private [20:03] company, much to the chagrin of many of [20:04] the shareholders, I think. Uh, now as [20:07] they go into the second day on this, [20:09] yes, [20:11] I mean, I'm in I'm in a couple of funds [20:13] that have large positions like go [20:15] public, boys. They are bidding bidding [20:18] 53 billion for your alma mada, David [20:21] Sachs, PayPal, which is a public [20:23] company. Stripe and the private equity [20:25] fund Advent [20:27] are jointly offering to acquire PayPal [20:29] for about 60 bucks a share, which is a [20:32] small premium. Most people think it'll [20:33] go for more like $70 a share. PayPal [20:37] stock jumped on the news obviously. And [20:40] >> were you able to get to the bottom of [20:42] this that was it just Stripe and Advent? [20:45] Because then somebody else reported that [20:47] it was also Block. [20:48] >> Yes, Block is coming in as well. [20:50] >> It's so confusing. every other media [20:52] source is like they're all over the [20:55] place on this. [20:55] >> Yeah. No, it's I think it's because it [20:56] was a breaking story and maybe they were [20:58] trying to keep it quiet. [20:59] >> It's a huge deal though if if Block is a [21:02] part of it versus if they're not. I [21:03] think [21:03] >> Yes. And Block, formerly known as [21:05] Square, is Jack Dorsey's uh payment [21:07] company, one of the few entrepreneurs to [21:08] ever create uh two decacorns in our [21:11] industry. And uh they're contributing 17 [21:14] billion in equity in the combined offer. [21:17] how they um chop up what's inside of [21:21] PayPal would be the big question. [21:23] Obviously, they own, [21:25] you know, a number of different brands [21:27] including Venmo in addition to PayPal. [21:29] That might go really well with the I'm [21:32] just taking a guess here with the block [21:34] assets. Stripe owns Bridge. That's their [21:36] stable coin infrastructure company they [21:38] acquired for a billion dollars in 2025. [21:40] PayPal has SIUSD which is already in [21:43] circulation. That's their stable coin. [21:45] So, stable coins are part of this. But [21:46] the biggest thing is PayPal is still a [21:49] juggernaut. Sachs 439 [21:54] million consumer accounts. You did [21:57] something right there 25 years ago. It [21:59] still the test of time. [22:00] >> It's kind of amazing. It really is. [22:02] >> Isn't it amazing that it's still that [22:03] strong? Yeah. It's weird when brands [22:06] keep going for that long. [22:07] >> But the problem is that the product is [22:10] getting very long in the tooth. I think [22:12] it's only growing 7% a year which is a [22:14] lot on the base that you know it's it's [22:16] grown to it's a big base but the product [22:19] has become somewhat obsolete and in a [22:21] way it's a legacy product. I'd be [22:23] curious to hear from the Stripe guys how [22:25] they would fix that cuz I think that's a [22:26] very hard problem to fix. Maybe they [22:29] wouldn't maybe they would just run it [22:30] more efficiently and kind of milk it [22:34] >> kind of do a private equity play. The [22:36] different question [22:39] >> the interesting question to ask is [22:42] what is the only kind of baby that [22:45] advent and stripe and block could have [22:47] together and I think there's one which [22:50] is you are creating a competitor to Visa [22:55] and Mastercard [22:58] >> because you now have upwards of 6 or 700 [23:02] million accounts. You have massive [23:05] stable coin infrastructure. [23:07] You have all of the riskmanagement [23:09] infrastructure that Stripe has built [23:11] over the last 15 or 20 years. The big [23:14] critique of Stripe's business model [23:16] early on was they had to build so many [23:19] value added services because everybody [23:22] always thought the take that they could [23:23] make as a middleman sitting on top of [23:25] the traditional rails would effectively [23:27] get competed away. Now to their credit, [23:29] they've done such a good job that that [23:30] hasn't happened. But I think what it [23:33] means now is you can vertically [23:35] integrate and go soup to nuts. That is [23:38] probably the most obvious [23:40] thing to make out of it. So that now [23:43] Stripe gets access to an ultra- lowcost [23:45] set of payment rails literally like to [23:48] zero brings it everywhere all over the [23:49] world. So does block advent can pour [23:52] money into it. So it's quite powerful if [23:55] it comes together. Rbert, what does this [23:57] say about private market companies at [23:59] scale and Stripe potentially never going [24:03] public? I mean, how does a private [24:04] equity firm get their return on this [24:06] investment in 2, three, four years? Do [24:10] they wind up selling more of it to [24:12] Stripe? What are your thoughts here on [24:14] the structure and capital market [24:15] implications here? [24:17] >> I think there's going to be more of [24:19] these kinds of deals. If you look at [24:22] Ryan Cohen's bid for eBay, I think it's [24:24] probably a second dot on a line that I [24:27] think is emerging, [24:29] which is folks [24:31] that are call it AI native are looking [24:35] at call it first generation digital [24:38] native businesses that have become [24:41] mature and old and stale and aren't run [24:43] by the founders anymore and have not yet [24:46] realized the opportunities with AI, have [24:48] not yet realized their potential. or [24:50] overspending in a lot of ways. And when [24:53] you take a look at those businesses as a [24:55] modernday AI operator, you're like, [24:58] "What the hell? This thing is so under [25:01] uh utilized. They're not using their [25:03] network well. They're not operating [25:05] well. They're overspending. They're not [25:07] using AI well." And there's a set of [25:09] opportunities that become quite obvious. [25:11] And I think the capital markets as we've [25:13] seen with like Josh Kushner's roll up of [25:15] accounting firms and General Catalyst [25:17] has a project like this where you can [25:20] kind of use capital to go buy you know [25:23] in those cases traditional services [25:25] businesses and AIify them. I think this [25:28] is part of a line of maybe looking at [25:30] traditional digital businesses and [25:33] AIifying them. And there's a long list [25:35] of these. There's a couple dozen of [25:36] them. So I think if you looked at the [25:38] public markets and you said, "Hey, where [25:40] are all these kind of software [25:42] companies, network businesses that [25:44] emerged in the early part of the [25:47] internet or even in the more recent part [25:48] of the internet, aren't run by their [25:50] founders anymore, have stalled out, [25:52] there's a massive opportunity. Now the [25:53] question as a capital provider is who do [25:55] you partner with to go and execute that [25:58] operational revival of that business?" [26:00] You're not going to go hire some [26:01] Mckenzie consultant to do that work for [26:03] you. It's got to be the best of the [26:05] best. It's got to be the right players [26:07] in the business. So, I think Ryan Cohen [26:09] has proved his metal obviously with some [26:11] of the things that he's done with Chewy [26:13] and GameStop and that's obviously [26:15] debatable. I spent some time [26:16] interviewing him to understand his [26:18] >> This was on the Allin interview program. [26:21] You can go to our channel and find it [26:22] there. It's last month. [26:23] >> Thanks for the plug. And obviously, when [26:25] it comes to payments, who better than [26:28] Stripe and maybe Jack Dorsey plays a [26:30] role here. And and by the way, I think [26:32] because capital I mean, if you think [26:33] about that $17 billion equity [26:34] contribution, what that technically [26:36] means is Stripe is selling and Block is [26:40] selling $17 billion of equity to the [26:44] cash investors. That cash is then going [26:47] to buy a PayPal. Therefore, Stripe and [26:50] Block end up owning a piece of PayPal. [26:53] The private equity investors own a piece [26:56] of Block and Stripe. And what is not [27:00] clear in the deal docs that were [27:02] published cuz I don't think it's [27:03] relevant to the public markets is who's [27:05] actually going to operate PayPal post [27:07] close. And my bet would be that they're [27:09] going to hand it over to the Stripe guys [27:11] and say you guys [27:12] >> I think that's clear. Yeah. Because [27:13] they're the most qualified and they will [27:15] have the biggest stake in it. Uh you're [27:17] >> so I think I so I I will make a [27:18] prediction. I think that eBay and PayPal [27:20] are probably the beginning of a wave of [27:23] mega deals of call it flaccid, you know, [27:28] digital businesses [27:30] that can be revived, okay, [27:32] >> with the blue chew of capital and the [27:34] right operator and I think that there's [27:36] probably a big [27:38] >> a big wave of this to come. [27:39] >> The other thing is this is probably not [27:42] the final clearing price. I think the [27:45] price is probably another 10 to 15% [27:47] higher from here. And and I will say [27:51] that there is a certain individual that [27:54] must look very closely at putting in a [27:56] competitive bid. Oh, a certain [27:59] individual who may may have his [28:03] fingerprints on the original PayPal who [28:06] might also have $4 or5 trillion in [28:09] market cap to play with who also made a [28:12] $60 billion acquisition recently. We [28:14] don't have inside information here. [28:16] >> To your point, Freeberg, [28:18] >> correct? [28:19] >> This is becoming a playbook. There's a [28:21] company called Bending Spoons that just [28:23] went public. They bought a bunch of [28:26] non-founderled assets. AOL for 1.4 [28:28] billion, Vimeo for 1.4 billion, We [28:30] Transfer, Eventbrite, Bright Code with [28:33] him. [28:34] >> I was just DMing with him. [28:35] >> He's He's awesome. I've hung out with [28:37] this guy. This guy is an absolute [28:39] freaking operational killer. He he [28:43] bought Evernote. [28:44] >> Yeah. In Milan. [28:44] >> And yeah, he runs the whole thing from [28:45] Milan. [28:46] >> He bought Evernote and he just goes in [28:48] and he diagnoses these businesses. He's [28:50] like, "How are they being over where are [28:51] you overspending? Where are you [28:52] underpending? what are you doing wrong [28:53] with the product and what are you doing [28:54] wrong with marketing and he just faking [28:56] fixes it and he's just a killer and he's [28:59] taken all of these what were called web [29:00] 2.0 I know businesses and he's [29:02] revitalized them, rolled them up and [29:04] printing cash out of them and leverage [29:06] the cost to run them lower. [29:10] >> He's using young AI first executives [29:12] from what I'm talking [29:12] >> totally. It's a great call out JL like [29:14] bending spoons is the roll up of this [29:16] sort of strategy but for these mega [29:18] deals I think there's more of them to [29:19] come. I will say uh you know the the [29:21] high order bit here uh which we talked [29:23] about for a couple years was venture [29:26] capital was on the ropes for a couple of [29:28] years under the wrath of Lena Khan and [29:31] then once Trump got elected all the [29:33] executives working in corporate [29:35] development said hey looks like M&A's [29:38] back on the menu and now we're seeing [29:40] deal after deal after deal get [29:42] consummated people are no longer scared [29:44] of doing deals Uber just bought delivery [29:47] hero today uh That's going to like jump [29:50] their revenue by 20. Yeah. And that's [29:52] going to jump their revenue by like [29:53] they're getting diluted 10%. It's going [29:55] to jump their revenue 24% or something [29:57] crazy like that. [29:59] >> And this is going to be I think the big [30:01] story the next couple of years and all [30:03] this liquidity. Talking to LPs and [30:06] family offices, which I do on a regular [30:07] basis, they're all like, "Hey, when's [30:09] your next fund? Hey, when's the next [30:10] deal?" Because now people are believing [30:12] in venture because of the SpaceX [30:14] distributions and all this M&A. And we [30:16] have four or five companies that got [30:17] bought since Donald Trump was elected [30:20] president. Thank you my President Donald [30:23] J. Trump for putting M&A back on the [30:25] menu. Sachs M&A back on the menu. Why [30:28] didn't you make a bid saxs for PayPal? [30:30] >> Been there, done that. [30:31] >> Been there, done it. Okay. [30:32] >> No, but look, you have to have [30:34] synergies. [30:34] >> There was a moment, this was like 15 [30:36] years ago where they asked Sachs to go [30:39] back and be the CEO. [30:40] >> That was right. I remember that at the [30:41] poker game. Yeah. He and I immediately [30:43] flew to Vegas and spent the weekend [30:44] there to think about it. [30:47] >> Like, you know what? [30:48] >> I wasn't No, they didn't ask me to uh [30:52] but I was [30:54] >> Well, no, I I never got the offer, but [30:56] it was down to like final two or [30:58] something and it was between me and [30:59] someone else. And actually, they ended [31:02] up going with some, you know, like [31:04] traditional like credit card executive. [31:07] And to be honest, that's why PayPal has [31:10] stagnated is that as soon as it was [31:12] acquired back in 2002, they basically [31:14] blew out like all the founders, all the [31:16] founding DNA and it was just kind of run [31:19] by, you know, consulting types. You got [31:21] to remember at that time it was acquired [31:23] by eBay and Meg Whitman had worked at [31:26] Proctor and Gamble and Disney and she [31:29] spent like eight years at Bane and it [31:31] was like a very corporate mindset. I [31:33] mean among all the internet companies of [31:34] that era it was definitely the most [31:36] corporatist and they saw the founders [31:39] the founding generation at PayPal is [31:42] just a problem just a bunch of like [31:44] cowboys they couldn't control they made [31:46] no effort to retain them and I think [31:47] they were kind of relieved when they all [31:49] left and then that's what created the [31:50] PayPal mafia was that you know normally [31:53] in an acquisition you'd lock up all the [31:55] talent but in this case [31:57] >> they locked them out they're like these [31:58] guys hard to manage change the keys [32:01] >> I've said for a long time it's a [32:02] misnomer to call it the PayPal mafia. [32:04] It's really the PayPal diaspora. [32:06] >> Totally. [32:06] >> Our homeland was taken over and they [32:08] burned our temple and then kicked [32:10] everybody out. [32:10] >> Yeah. [32:11] >> And that's why the whole PayPal mafia [32:13] got started with all those companies. [32:15] >> But as a result of that, [32:16] >> for for decades, I don't say anyone's a [32:18] victim. That's [32:20] >> Look, when you when you acquire a [32:21] company, you get to decide what to do [32:23] with that asset. [32:24] >> Totally. [32:25] >> So, I mean, that was just the reality. [32:27] But it's not like, you know, I don't [32:28] think anyone was bitter about it. [32:29] They're all like, "Okay, this give us [32:31] the capital to go. I'll do the next [32:33] thing we want to do." [32:34] >> Well, the new CEO, by the way, Enrique, [32:36] is really aces. Uh, I've met him before [32:38] and they're doing a great job. Um, [32:40] apparently, which is why they probably [32:42] got these offers because they've been [32:43] tightening that business up for the last [32:44] couple years. [32:45] >> Well, no, the reason they got these [32:46] offers is the market cap is down to, you [32:49] know, was down in the what, like 30 [32:51] something billion. I mean, this is a [32:52] company that was worth 200 billion, [32:54] wasn't it? Roughly [32:56] >> 322, I think, was the peak. [32:58] >> Mhm. And before this offer, it was down [33:00] to what 30 to 40 billion. Yeah. [33:02] >> So the reason why it's attracting offers [33:05] is it's so beaten up. And so now the [33:08] question is, can anyone else do [33:10] something with it? [33:10] >> To your comment about you've got to have [33:12] synergies. Doesn't it seem to be the [33:14] case that in this era the the core [33:17] synergy that any great operator can [33:18] bring to the table in this sort of a [33:20] scenario is AI? Like you can leverage [33:23] whether it be in this business or others [33:24] you can leverage tools that drive [33:26] automation that drive product [33:27] development that drive improvements and [33:29] efficiencies across the organization [33:31] that make the product actually better [33:32] for the user etc etc that simply [33:35] >> potentially obviously being well [33:37] implemented. Well, look, I think you [33:39] have to have a product vision of how you [33:41] would use AI to make the whole user [33:43] experience better. And yes, you're right [33:45] that you could just use AI to drive [33:47] efficiencies and that'll improve your [33:49] profitability and earnings. And so on a [33:51] financial level, you could make the [33:53] acquisition work. But it seems to me [33:57] that the existential issue for PayPal is [34:00] that you're dealing with a product [34:01] that's 25 years old. I mean, it's the [34:04] same thing that we created back, you [34:06] know, like 27 years ago. I mean, it's [34:08] changed a little bit, but not that much. [34:10] And the problem is that that that [34:13] interaction model is legacy. And so, [34:16] unless you've got a a vision of how to [34:19] resuscitate it and rejuvenate that that [34:21] product, I think, yeah, it could be a [34:23] good financial play. Maybe [34:24] >> I think they're buying the accounts. [34:26] >> Yeah. I mean, what you're saying, [34:28] Jamoth, is Yeah. what you're saying is [34:29] interesting because with Stripe, I mean, [34:32] this is the advantage that Stripe has is [34:34] that they have a ton of merchants, [34:36] right? So, they've they've become the [34:38] preferred mechanism for merchants to [34:42] basically accept payments via APIs. And [34:45] I think they're doing about two trillion [34:47] a year of annual transaction volume. I [34:50] think PayPal is doing 1.7. So, actually, [34:52] Stripe is a little bigger than PayPal [34:54] now. But the thing that PayPal has that [34:57] Stripe doesn't really have is the [34:58] consumer relationship. So, you know, [35:00] over 400 million active consumer [35:02] accounts. So, you're right, Jimoth, that [35:04] if if somehow you could combine the [35:06] merchant relationships with all those [35:09] consumer accounts and then bypass the [35:10] credit card networks cuz there'd be a [35:12] lot in theory there could be a lot more [35:14] on us transactions. [35:16] >> Exactly. And [35:17] >> that's where the value is. [35:18] >> PayPal already owns Brainree. So, now [35:20] you have Stripe and Brainree that [35:22] effectively were competitors that won't [35:23] be. And then what block gives you is an [35:26] entire point of sale infrastructure and [35:28] you get the cash app. So you put it all [35:30] together and I think it's a it's a shot [35:33] across the bow for Visa and Mastercard. [35:36] >> Brainree is the other one that I think [35:37] you just mentioned there Chimath that's [35:39] important because that is a very strong [35:41] business that you don't even know that [35:43] PayPal owns. Venmo is also that speaks [35:46] to a lot of young people. So you're kind [35:48] of getting two generations. You're [35:49] getting Gen X and millennials. Uh a lot [35:52] of bang for your buck there. Yeah, [35:54] >> what Zach said is true. They have enough [35:55] of the things to go end to end on their [35:57] own rails. [35:59] >> That is a very [36:01] >> Yeah, the question is whether you can [36:02] package it all together in a way that [36:04] the consumer will actually choose [36:06] because it's one thing to say, well, we [36:08] take the merchant relationships of [36:10] Stripe and the consumer relationships of [36:11] PayPal, we don't do that. But what if [36:14] the consumer doesn't want that? [36:15] >> No, you don't do that. I think what you [36:16] do is you go to places like all of the [36:18] merchants that use Stripe and say, [36:20] "We'll give you a 3 or four or 5% [36:23] discount and they'll be like, "Okay." [36:26] And so you'll see these prices that [36:28] just, you know, fall everywhere. Like [36:30] imagine if Shopify was like to all their [36:32] merchants, okay, you have two choices, [36:34] the old way or the new way. The new way, [36:36] you put another two or three or 4% in [36:38] your pocket. Of course, they're going to [36:39] pick the new way. [36:40] >> Well, and this is the paradox of like [36:42] modern M&A. You know, if you look at [36:44] protecting the consumer, this will [36:46] ultimately be great for the consumers. [36:48] This is going to lower the prices on it. [36:49] They're not buying this to [36:50] >> Okay, you're saying something really [36:52] interesting. It is so good for consumers [36:54] if this were to happen. This is the [36:56] exact reason why if this had happened [36:58] two years ago, [37:00] >> yeah, [37:00] >> this would have been the antitrust [37:04] equivalent of a coloctyl exam. I mean, [37:08] >> oh god, [37:09] >> you would not even get one step close to [37:12] doing this deal two years ago. [37:14] >> Well, that's that's really interesting [37:15] actually. I mean, the key question with [37:17] antitrust is how do you define the [37:19] market? And so, if you define the market [37:21] as, you know, APIs for merchants, then [37:24] JCAL, it would be Stripe versus [37:26] Brainree. And then the government would [37:27] say, well, that's you can't consolidate [37:29] share. [37:30] >> However, if the real market is Visa and [37:33] Mastercard, that's the ultimate duopoly. [37:35] And if PayPal can add competition to [37:37] that market, which is infinitely larger [37:39] than APIs. Exactly. Then [37:42] >> it's actually pro competitive. So how [37:43] you define the market determines whether [37:45] it's anti-competitive or prompetitive. [37:48] >> And those guys are smart enough and [37:49] they've read enough books where they [37:50] won't [ __ ] this one up. [37:52] >> All right, let's get to the next topic. [37:54] >> Somebody else is suing Open AI. This [37:56] time it's Apple. On July 10th, Apple [37:58] filed a 41page lawsuit against OpenAI [38:03] over alleged alleged stolen trade [38:06] secrets. Apple says Open AI stole their [38:08] IP to develop their consumer hardware [38:11] device. Remember, we had um [38:15] Sarah Frier at Liquidity and I uh probed [38:20] her on this new device and she said it [38:22] was very human and lovable. She gave us [38:24] a little bit of the goods. Well, it [38:26] turns out Apple is alleging that maybe [38:28] this is partially their IP. Tang Tan, [38:33] Apple's former VP of iPhone design, is [38:36] OpenAI's chief hardware officer. He [38:38] allegedly directed Apple job candidates [38:41] interviewing at OpenAI to bring quote [38:44] actual parts to interviews to quote show [38:47] and tell in the interviews. Changu, [38:50] former Apple senior technical engineer, [38:53] sent this text message to a still [38:56] employed Apple colleague. Quote, "lol, I [38:59] found out I can access the network [39:01] storage. So funny." During all of this, [39:04] OpenAI has poached over 400 Apple [39:07] employees over the last year or two. Big [39:10] numbers of poaching. Apple and Tim Cook [39:13] have seen enough chimoth. Tim Cook green [39:16] lit this. As insane as this is, Sam [39:19] Waltman has found a way found a way to [39:22] screw yet another party. Screwed Elon, [39:24] his first benefactor. [39:26] Chimati, if you remember correctly, the [39:29] default for iPhone AI was supposed to be [39:34] chat GPT. So they took this [39:36] relationship, Sam took this relationship [39:39] where he got to be the default on the [39:41] most important platform for AI, the [39:43] iPhone, and now it's wound up in a [39:46] massive lawsuit. What are your thoughts [39:48] on this? [39:50] I haven't really seen Apple [39:53] very latigiously in 25 years in Silicon [39:55] Valley. [39:58] So that's obviously a concerning data [40:00] point for OpenAI. They're very reactive [40:03] more than they are [40:05] proactive on these things. So, there [40:07] must have been something that really [40:09] really upset them. [40:10] >> Egregious. Yeah. [40:12] >> So, I don't know. It's going to take a [40:14] court to sort this out. I don't really [40:16] want to gossip because like who knows [40:18] what's actually going on and who said [40:20] what and blah blah blah, but nobody [40:23] should be stealing [40:25] things from their former employer. [40:27] Nobody. [40:27] >> Obvious. [40:28] >> You're just not allowed. It's just [40:29] obvious. These people are very, very [40:30] smart and they're very successful. And [40:33] you know that's why OpenAI probably [40:35] wanted them and that's why they wanted [40:36] them you know and you come to them with [40:39] the collective wisdom of what you've [40:40] accumulated and I think that's [40:42] sufficient. You don't need to especially [40:44] as a senior person do this. So I just [40:46] hope that this stuff isn't true because [40:49] I think I don't think that Sam or Sarah [40:51] or anybody else there are trying to [40:54] induce this to happen. I don't think so. [40:56] >> Yeah, I doubt they induced it but I do [40:58] believe that it's true or Apple wouldn't [40:59] have brought it. Sachs when we look at [41:02] this maybe you could open the um [41:04] aperture here if you want or you can [41:06] just go very detailed but the nature of [41:09] we have a free market we don't have [41:11] non-competes in California generally [41:14] speaking you can just employment at will [41:16] go where you want but we have had [41:18] instances you know whimo famously uh you [41:21] know brought some [41:23] IP to Uber when they did Travis and the [41:26] team said leave the building your job is [41:29] rescended you don't get to bring that [41:30] information here. In this case, it seems [41:34] maybe they didn't induce it, but it [41:37] occurred for some period of time. So, [41:39] take us through big picture [41:42] what you think is going on here and what [41:44] it means for the industry. [41:46] >> Well, like Jamas said, I have no idea [41:47] what's going on here. I mean, these this [41:49] is a lawsuit. The facts are all alleged. [41:51] We don't know. It's going to be [41:52] adjudicated. So, I really don't want to [41:54] opine on what happened here. But if [41:57] people want to know a very simple rule [41:59] of thumb for how to avoid these types of [42:01] disputes is just when an employee leaves [42:04] their previous company and joins the new [42:06] company, just don't take anything with [42:08] you. [42:09] >> The only thing you can bring to your new [42:10] job is what's in your head. [42:11] >> Your memories. [42:12] >> Your memories. That's fine. Whatever is [42:14] in your head, you're allowed to take. [42:16] But never leave with anything else. [42:18] >> No thumb drives, no CD, no documents, no [42:23] nothing. [42:24] >> Justip what's in your is okay. [42:27] Fraberg, any thoughts here on [42:30] just the [42:32] number of lawsuits that seem to be [42:35] piling up over at Open AI? Bad luck. [42:41] >> Couple dots make a line, I guess. [42:44] >> Okay, there you go. Very well said. [42:46] Couple of dots make a line. Okay, SpaceX [42:49] had a data leak this week. They launched [42:52] Grock Build in public beta at the end of [42:55] May. The newest coding model, Grock 4.5, [42:58] powers Grock build. I've been playing [42:59] with it. It's extraordinary. It's a [43:01] coding tool that works inside of Kurser. [43:04] Uh, SpaceX previously told users, [43:06] Shimoth, nothing from your codebase is [43:08] transmitted to XAI servers during a [43:10] session. But what actually was happening [43:12] is every time a developer or according [43:15] to reports used Grock build the tool was [43:18] sending their entire codebase to SpaceX [43:20] cloud servers out without alerting the [43:22] users not just the files that were [43:25] needed to do that specific coding task [43:27] just everything passwords API keys could [43:29] have got pulled up there all the uh [43:31] change logs etc. Uh the privacy setting [43:34] was supposed to stop this but it didn't [43:36] work. SpaceX quietly disabled the upload [43:39] on July 13th by flipping a switch on [43:41] their servers. Elon Musk, friend of the [43:43] pod, promised on X that all previously [43:45] uploaded data has been deleted, I guess. [43:47] Uh, and in response, SpaceX open- source [43:50] rock build. That's their harness. So, [43:52] that's another opensource win or win for [43:56] the open-source community and AI [43:58] sovereignty. [44:00] You got any kind of thoughts on this? [44:02] Obviously, this was not intentional, but [44:04] trust is important uh with these models [44:06] as we've been talking about for the last [44:08] couple of months here on the All-In [44:09] podcast. [44:10] >> I would actually connect this to my [44:13] comments on CNBC earlier this week, [44:16] which built on top of Alex Karp's [44:18] comments the week before. [44:21] Privacy in AI is very fragile and it's [44:25] very brittle. And this is despite the [44:28] best efforts of great businesses. Like, [44:32] you know, you may not like Elon for [44:34] personality quirks, but he is incredibly [44:37] trustworthy. He's overly [44:40] transparent. And so, to their credit, [44:42] they shut it off immediately. But my [44:44] takeaway is that there are all kinds of [44:48] non-obvious data leak vectors lurking in [44:51] AI. And so if you think that you're [44:55] going to flip a ZDR switch, zero data [44:57] retention, which is the magic term that [45:00] the industry uses to tell you that [45:02] everything's going to be okay. [45:04] I think the answer and the message [45:06] should be it's not going to be okay [45:07] because you can't guarantee any of it. [45:10] So the model companies when they give [45:12] you these zero data retention policies [45:14] are probably trying their best. But I [45:16] think the reality is you are leaking [45:19] information where you don't know it. and [45:21] they despite their best efforts may [45:23] still have trap doors that they don't [45:24] even know about until it's figured out [45:26] by somebody else like in this example. [45:28] So all of this speaks to you have to [45:30] have a stratified ecosystem. You have to [45:32] have third parties. Now look that's very [45:34] biased for me because it's in part what [45:37] we do for large enterprises at 8090 when [45:39] we implement our software factory. But [45:41] the reason why it's working so well is [45:43] this exact reason. You need an [45:45] independent third party layer to [45:47] interface to these models to manage this [45:49] exposure because there are trap doors [45:52] everywhere. [45:53] >> And that's what Sachi just said in a [45:55] really interesting blog post. Did you [45:57] guys see that? [45:57] >> Yeah, [45:58] >> I thought that was excellent. [45:59] >> The reverse information paradox. [46:00] >> Yeah, [46:01] >> that's exactly the the takeaway that he [46:03] left with. He was building on Alex [46:05] Karp's supposed crash out. You know the [46:08] point that Kart made about how [46:11] enterprises who have technical ability [46:14] want control over their compute models, [46:16] weights, data and alpha. But he he went [46:20] further with that idea. I mean he [46:21] started with Karp's idea but then he [46:23] kind of provided a recipe a road map for [46:26] how enterprises should operationalize [46:28] that. And what he says is that [46:32] enterprises they have to establish a [46:33] real trust boundary with private eval [46:36] proprietary learning loops inside the [46:38] tenant decoupled orchestration [46:41] and the explicit right to fine-tune [46:43] their own outputs. So he kind of goes [46:45] through a litany of fairly technical [46:47] things that enterprises should do in [46:50] order to achieve the operational control [46:53] that Karp was saying that enterprises [46:55] really want over their AI compute models [46:58] and data their alpha. So it's really [47:00] interesting. I think now there's you [47:03] know a virtual almost like college dorm [47:06] session going on between the leaders of [47:08] these companies who are brainstorming [47:11] some of these concepts and now extending [47:13] them. Right. [47:14] >> And what's happening is you're starting [47:16] to see the formation of not really an [47:18] alliance but like an ecosystem [47:21] that is trying to create alternatives to [47:23] you know a monolithic closed model stack [47:27] which is where anthropic and to some [47:29] extent open AAI want to go is they want [47:32] you to be locked in to their to their [47:34] stack right their models their harness [47:36] they control the data all you know all [47:38] of that and now you're starting to see [47:40] all these different companies [47:41] >> and you pay a huge premium for the [47:44] privilege for them to do it, which is [47:45] even more insane. So, I I saw this data [47:49] and Nick, maybe you can find this [47:50] companion clip. The companion clip I'd [47:52] like you to find is Eric Glyman, who's [47:54] the CEO of RAMP, was on Squawkbox, I [47:58] think today, talking about a new feature [48:00] where you can manage the token maxing of [48:02] your employees through your ramp card. [48:05] But the data that I saw was that a [48:08] million tokens on Fable is about sachs [48:11] 56 bucks. A million from soul is about [48:15] 26 bucks which is the same as quad 48. A [48:19] million input tokens from on Gro is [48:22] about $1.50. [48:24] >> Okay. [48:25] >> Z is about a $150. Elon's about a dollar [48:28] and the Chinese models are 50. [48:31] >> Wow. So on top of the whole data [48:33] sovereignty [48:35] bleeding your alpha away, can you [48:37] imagine that you're paying 56 bucks as [48:39] well per million input tokens for that [48:42] risk? That is insanity. [48:43] >> I'm using perplexity computer and they [48:45] started supporting Grock and they [48:47] already support GLM52. So when you're [48:49] like you're using Claude or OpenAI, you [48:51] can only use their model. So I started [48:52] effing with the different models and I [48:54] gave it all the same basically PRD and I [48:57] said I want to make a podcast player [48:58] that deep link. So, like if we were [49:00] talking about, I don't know, Mythos, it [49:02] would play me all the miso methos clips [49:05] across all the different tech and [49:07] business podcasts, but make it into one [49:09] stream. And I was like, this would be [49:10] like really helpful for me for prepping [49:12] for the show and just be interesting. I [49:14] did it. It took a couple of hours. It [49:17] cost $11 on the new Grock. It was [49:21] hilarious how cheap it was. And then [49:24] adding to this, I don't know if you saw [49:26] >> Sorry. Did you try to do it on Fable to [49:29] see how much more expensive it was? [49:30] >> I didn't. I didn't because I was out of [49:32] Fable credits on my $200 account. So, [49:35] you know, [49:35] >> look at this clip here. Nick, play the [49:36] clip from Eric Glenn. That's kind of [49:39] interesting. [49:39] >> We're thrilled to be launching token [49:41] spend management today. It's available [49:42] to RAMP and non-RAMP customers. [49:45] >> And he's exactly right. Over the last [49:48] year, I looked at the stats this [49:49] morning, token spend among RAMP [49:51] customers has grown by 21 times. [49:54] >> 21 times. [49:55] >> 21 times. [49:56] >> Not 21%, we're talking about 21 times. [49:58] >> That's exactly right. So being off by a [50:00] few pennies as a CFO uh actually might [50:02] be quite nice. At the rate it's going, [50:04] it might be several dollars. And and [50:05] look, like I I think that for uh many [50:08] CFOs, they're often very surprised by [50:10] the bill because what the AI companies [50:12] have functionally set up is you have a [50:14] tab. you can spend as much as you want. [50:17] It's very hard for CFOs to see [50:18] proactively what people are spending on [50:20] and every time they're introducing new [50:22] models, the rates often go up and so [50:25] there's very misaligned incentives. So [50:27] part of what we're trying to do is make [50:28] it easy for CFOs to see the spend, [50:30] understand the spend, and control it. [50:32] >> Thanks, Nick. He's saying something so [50:34] important there because if your [50:36] engineers are going off randomly in an [50:39] unguided system and then just ripping [50:42] through million tokens at 56 bucks, what [50:46] he's talking about is the eventual [50:47] downstream impact to earnings. And that [50:50] eventually a bunch of these public [50:52] market CFOs are going to show up to Wall [50:54] Street and they will have missed [50:55] earnings because they're upex at some [50:57] point. If things are 21xing every few [51:00] months, somebody's going to miss a [51:02] quarter. I don't know who, but somebody. [51:05] And it's not just going to be, you know, [51:07] I was speculating it'll be a few pennies [51:09] here or there, which they'll have to say [51:11] is because of token spend. He's saying [51:13] it could be as much as dollars at this [51:14] rate, which also could be the case. I [51:17] think the point that we're all trying to [51:18] make is unless you get a control of this [51:21] and you can directly say how much money [51:23] you're making, this is a bridge to [51:24] nowhere. It is a money burning furnace. [51:28] >> The good news is this is all creating a [51:32] massive market opportunity. Sachs, Bit [51:35] Tensor, Subnets, GLM52 hosting Grock [51:38] 4.5. Now [51:40] >> in Inkling, Inklink mirror Marott's new [51:42] model in everybody's now saying, "Hey, [51:44] wait a second. I can give you a better [51:46] deal. You're paying two bucks. I can get [51:48] you one buck. This is [51:50] >> No. No. People are paying between 26 and [51:52] 56 bucks. They should be paying 50 [51:54] cents." Exactly. [51:56] >> Well, you know, and [51:57] >> the Inkling announcement was kind of [51:58] interesting because I think the value [52:00] prop there is she's explicitly saying [52:03] that look, we're not frontier [52:05] intelligence. We're just under that, but [52:07] we're a platform for fine-tuning these [52:10] open models which are much much cheaper [52:12] and then you can achieve the result you [52:14] want based on fine-tuning. And so that's [52:18] really interesting. [52:20] >> Yeah. And uh but you know these open [52:22] models won't be around for very long if [52:25] Anthropic has its way. [52:26] >> There's a reason they want to stop it. [52:28] It's just they have such a monopoly. [52:30] >> Of course you're you're selling most of [52:32] the product for 50 cents per million [52:35] >> tokens when they're selling theirs for [52:37] 56 bucks. Of course you don't want that [52:39] to happen. Of course you want them you [52:41] want to try to stop it. [52:42] >> But that being said, they're still [52:43] growing like crazy. Just to be clear, I [52:45] mean, yes, you know, you are seeing this [52:47] explosion of interesting things [52:49] happening with open models like you [52:51] said, you know, the latest rock build is [52:53] open, thinking machines open and so [52:56] forth and so on, but still [52:58] >> I mean, they're growing. They're still [52:59] the industry leader in terms of revenue [53:01] growth. So, these things are happening [53:02] side by side. And [53:04] >> yeah, I would I think that the [53:05] interesting thing is Eric would not have [53:07] released this ramp product unless CFOs [53:10] were like, I can't control the spend. [53:12] >> Yes. And then he's like, "Well, here, [53:14] let me build it for you." And then if [53:16] enough CFOs essentially turn that [53:18] feature on and start to rate limit how [53:21] it's spent because maybe they're not [53:23] getting the ROI and the engineer doesn't [53:25] care about ROI. The engineer is like, I [53:27] want to use the latest greatest model. [53:28] >> Yeah. And maybe you don't need it. Maybe [53:30] Mirror's right. And for 95% of the [53:33] tasks, you should be at one level lower, [53:35] especially when it costs 1/100th of the [53:37] cost. But the engineer will never make [53:39] that trade-off because they'll never [53:40] want to think about it. You're right. [53:42] >> And also they're not tied to the money. [53:44] The CFO is tied to the money and the [53:46] engineer wants to go on an exploration [53:49] on using the latest greatest thing. [53:51] >> Yeah. If you're booking your Yeah. If [53:53] you're booking your travel, you're like, [53:55] you don't even see the price. You're [53:56] like, "Yeah, just put me in business [53:57] class. Put me in a nice hotel." And like [53:58] the travel department handles that. [54:00] >> You're saying something really [54:01] interesting. What percentage, if you had [54:02] to guess, of fable five prompts are just [54:06] average Michigan that should be running. [54:08] >> 98%. [54:10] 98%. I I I was using it for stupid stuff [54:13] that I could be using Quen for. Um I [54:16] think this is my like micro prediction [54:18] here. Mark German, who's like the most [54:20] in then guy when it comes to Apple, he [54:24] says, and you know, we got this new CEO [54:26] coming in for um uh John Furnus. [54:29] >> Yeah. And he is a hardware engineer. M7 [54:33] Ultra uh because M we're on M5 chips [54:36] now. you can get like, you know, 456, [54:38] 512 gigs of RAM. He says M7 Ultra is [54:42] going to support as much as 1.5 [54:44] terabytes. [54:46] That's double what they're already [54:47] supporting. So, if you think about uh [54:51] Frontier models, like the last [54:53] generation, this is like an Opus level [54:55] model running on your Mac Studio. You [54:57] guys all use Mac Studios, you're rich [54:59] venture capitalists, whatever. You're [55:00] like, "Yeah, I'll take a four or $5,000 [55:02] computer." This is going to change [55:04] everything. you're gonna have employers [55:05] go, "Oh, I can just run, you know, 90% [55:09] of my workloads, 99% of the workloads on [55:11] the local uh Mac Studio." I think Apple [55:15] is a screaming buy right now. Uh and I [55:18] this not financial advice, but my lord, [55:20] that company could just run the table on [55:21] AI if they get this right. [55:24] >> All right. [55:25] >> Apple. [55:26] >> Yes. Because I they're going to make [55:28] such a fortune. [55:30] >> You got Let me explain. It's just like [55:33] the iPhone. Everybody laughed at the [55:34] iPhone. People overpaid. [55:37] They did not. When the first iPhone came [55:40] out, many people [55:42] that was it. [55:43] >> That's true. Steve was the big one. You [55:45] can I can still hear him laughing. [55:46] >> No, if you think about how they make [55:49] money off of hardware, off of their [55:51] devices, they will put so much downward [55:54] pressure on Claude and Open AAI by just [55:57] putting local models and supporting them [56:00] with this memory architecture. It's [56:01] going to be wild when people have [56:04] unlimited tokens on their desktop. I'm [56:05] telling you, [56:06] >> I don't know if you guys uh saw this, [56:08] but um there's a very large solar [56:11] company called Sunun. They just [56:13] announced this week that they're making [56:16] distributed data center blocks that you [56:19] can put in your house. [56:22] Another company that did it is company [56:23] called SPAN that partnered with Nvidia. [56:26] So to your point Jason, you're seeing [56:28] this fragmentation and distribution of [56:30] edge compute which I think is a theme. [56:32] Definitely a theme. [56:33] >> Well, it's also chasing energy, right? [56:35] Like if you've got some solar, if you've [56:37] got excess battery power, hey, we power [56:39] up your batteries at night cheaply. [56:41] >> I think I told you this last week, we [56:43] are so massively short electrons. By [56:44] 2050, the United States of America will [56:47] be 2 and 12 California's worth of energy [56:50] in deficit. 2.5 Californians, the fourth [56:53] largest economy in the world. we will be [56:54] short 2.5x [56:57] of all of the energy consumed by [56:59] California by 2050. [57:02] This week there was an auction by this [57:06] huge utility called PGM which serves [57:09] Pennsylvania, New Jersey, Maryland, you [57:11] know, 13 states. And that auction is [57:14] where they publish a forward curve and [57:15] say, "Hey, listen guys, here's my [57:18] forecasted load and here's how much [57:20] energy I need." And people signed up to [57:23] essentially get paid a a guaranteed rate [57:26] every day so that they have to fork over [57:28] the energy in the future. Kind of like a [57:29] forward option. They needed like seven [57:31] or eight gawatt. [57:34] They had 156 megawatts or something show [57:36] up. Well, we are in such a bad place [57:39] right now on electrons and electricity [57:42] prices. [57:42] >> Did you see what our boy did this week? [57:44] >> We need we need So this is behind the [57:48] meter, which is different. And he he [57:49] Elon needed to do this by the way just [57:51] so you know because there's an issue in [57:53] Memphis where he was very clever about [57:55] how he was able to get Colossus off the [57:56] ground that regulatory it's not [58:00] >> explain this [58:01] >> when you try to power a data center [58:05] typically you have what's called grid [58:08] power. So you go to the utility in the [58:09] area and you say, "Hey, please run me a [58:11] line off of that main transmission [58:13] line." And that's how you power your [58:16] data center. When that runs out or is so [58:20] backlogged, you have to do what's called [58:23] behind the meter, which means on your [58:25] own property that you own, you build [58:28] something for yourself. Now, there's a [58:30] problem with that. You would think, [58:31] well, that's smart. Yes, but like in [58:34] everything in America, there's [58:35] regulation on top of regulation on top [58:36] of regulation. And one of the most [58:38] complicated regulatory schemes that you [58:41] have to overcome is clean air [58:44] permitting. [58:45] So even if you say you're going to do [58:47] behind the meter, then you're like, [58:48] well, what can I do? Solar you can do, [58:51] but it takes too much space for most [58:52] places. Batteries you can do, but you [58:56] need to generate the electricity in the [58:58] first place. So people use that gas. So [59:00] Elon cleverly bought a ton of 18-wheeler [59:05] like engines basically. [59:07] >> He bought the company that makes all [59:09] this and provides mobile and then just [59:12] and then just you know pin them to the [59:13] ground and then ran it and you know [59:16] those are personal use essentially and [59:18] so they came under the clean air [59:20] permitting requirements but then when [59:21] you act as a block you could make the [59:23] claim that it doesn't. Now, there are [59:25] new solutions like Bloom Energy, which [59:28] allows you to have huge installations [59:30] and still fall under the the the [59:33] personal use clean air permit. Um, and [59:35] so for all of Elon's future capacity, he [59:38] needed to have this in place so that he [59:40] gets the the clean air permits and he's [59:42] able to have a clean run of sight to [59:44] continue to build domestic data centers. [59:46] Anyway, [59:47] >> speaking there's your little TED talk on [59:49] energy, but uh we are in a bad place, [59:52] guys, and it's only getting worse. [59:53] Speaking of um data center sachs, [59:58] everybody's favorite socialist governor [60:00] Kathy Hochel in uh the great state of [60:03] New York, my hometown, [60:05] >> powered by fossil fuels, they drive up [60:07] our carbon footprint. They occupy [60:10] massive amounts of land, potentially [60:12] displacing agricultural space and open [60:15] spaces. The bottom line is progress [60:18] shouldn't arise with a higher utility [60:19] bill, deleted water supply, or noise [60:22] pollution. So we have no choice but to [60:24] address these challenges created by [60:27] these massive facilities. [60:29] That is why today I'll be signing the [60:31] nation's first ever statewide moratorum [60:34] on hyperscale data centers. [60:37] >> Everything she's saying there is a false [60:39] accusation uh on the data center. So [60:41] let's just go one by one. So she's [60:43] saying [60:44] >> that they eat up all of the power. Well, [60:47] yeah. I mean, look, if you connect to [60:48] the grid without producing more power [60:51] and you force data centers to compete [60:53] with, you know, residential rateayers, [60:55] then yeah, you could drive up utility [60:57] prices. However, if you do what Chamas [61:00] said and let them build behind the [61:01] meter, then they bring their own power. [61:03] And that's what the president has [61:04] advocated for since the beginning of his [61:06] administration is let the AI companies [61:08] become power companies. So, that is the [61:10] way to solve the energy problem or the [61:13] utility problem. Then she's talking [61:15] about, you know, eating up land. The [61:16] reality is these data centers are a [61:18] model of land use efficiency. We have a [61:21] ton of land in this country. Obviously, [61:23] you can find places where there is [61:25] enough open land to build a data center. [61:28] The economic impact and value of a data [61:30] center relative to the land use again is [61:33] one of the best ROIs there is. The [61:35] supposed noise pollution that's largely [61:37] made up that can be dealt with. You [61:39] obviously don't want to put these things [61:40] right next to a residential area, but [61:42] create a little bit of distance and it's [61:44] fine. The whole water consumption thing [61:46] is largely a hoax. [61:49] >> The the modern data centers uh [61:51] recirculate the water, [61:52] >> closed loop systems. [61:54] >> Yeah. And I think there there was a [61:55] study that showed that a typical data [61:57] center uses the same amount of water as [61:59] two and a half in-n-out burgers. So, [62:01] in-n-out burger chains. [62:03] >> I mean, just go after the almonds if [62:04] you're concerned about water, people, [62:06] please. [62:06] >> There's there Yeah. Or golf courses. I [62:08] mean there's many, you know, there's [62:10] many uses of water that are way more [62:12] wasteful. So when you compare economic [62:14] impact to all these different things, [62:17] data centers are like honestly one of [62:19] the best things we could be building as [62:21] a nation. But [62:22] >> and s there's all these taxes and [62:24] incremental revenues. Did you see the [62:26] article where I think it was in North [62:28] Dakota or something where like teachers [62:30] were getting like 30 and $40,000 bonuses [62:33] from all the tax revenue that was coming [62:35] in? There's all these upsides. [62:38] >> That's right. They generate um a lot of [62:40] tax revenue. They've created a [62:41] bluecollar construction boom. It's not [62:43] true that there's no jobs once they're [62:45] built. That you do have ongoing jobs [62:47] there. And then oh, one final thing just [62:49] on the the point that Hokll was making. [62:51] She said it created a lot of pollution. [62:54] Natural gas, which is how most of these [62:55] data centers are are powered, is one of [62:57] the most clean burning sources of power [62:59] that we have. [63:00] >> 100%. [63:01] >> These data centers have become the [63:03] scapegoat for all the angst that people [63:06] have about AI and it's kind of become [63:08] this very clumsy way of trying to throw [63:11] a wrench in the gears of innovation and [63:13] just kind of slow the whole thing down. [63:15] >> All I have to say is welcome to Texas. [63:17] We got plenty of land here. And what's [63:19] so stupid about her proposal and her [63:24] talk, aside from the thing she got [63:26] completely factually incorrect, is New [63:29] York State is like 80% underdeveloped. [63:33] Drive upstate, folks. You're thinking of [63:35] New York City. Yes, New York City is [63:36] packed. You go upstate, it's literally [63:39] 70 to 80% of the land in New York State [63:42] is undeveloped. There's so much land. [63:46] It's ridiculous. New York is giant. It's [63:49] giant. [63:50] >> On this topic this week, I just want to [63:53] give a shout out to Senator Dave [63:54] McCormack. [63:57] He had [63:59] a defense and innovation summit in uh [64:01] Carile, Pennsylvania at the Army War [64:03] College [64:05] which a bunch of us went to pus came [64:07] gave a speech had a CEO round table a [64:10] lot of defense company CEOs etc. But [64:12] Chris Wright was there at Sachs and [64:14] >> my guy [64:16] >> he's great and Chris mentioned this in [64:18] insane story. He said you know there is [64:21] a lot of common funding because Dina [64:23] Powell asked this question on stage and [64:25] he said there's a lot of common funding [64:27] patterns of these people that are [64:29] protesting um the data centers and he [64:32] said you could actually trace it back to [64:34] the same people that in a different era [64:37] were protesting fracking. And so he was [64:39] saying like it's these are all just [64:41] hobby horses that they use to raise [64:42] money, have a job. They're like [64:44] professionally paid protesters. They [64:46] kind of just show up out of nowhere. I [64:48] didn't realize that there was such a [64:49] commonality, but they're the same [64:51] people. [64:52] >> The thing that I just can't understand [64:53] for the life of me is why Anthropic is [64:56] still funding these groups that want to [64:59] put the kibos on new data center [65:01] construction. There's one called public [65:03] first where Daario just gave his first [65:05] seven figure contribution and then a [65:07] bunch of other employees at anthropic [65:10] gave it and you know all these groups [65:11] are trying to slow down AI development [65:13] with new regulations and making it [65:15] harder to build new data centers and at [65:18] a certain point you just have to wonder [65:20] I mean is this regulatory capture or [65:21] they just kind of lost the plot because [65:23] the number one thing slowing down the [65:25] growth of anthropics revenue it's not [65:27] demand I think it's the availability of [65:30] compute in data centers. And so you're [65:34] just kind of wondering like what is the [65:35] point of all of this? [65:37] >> It's so true. [65:37] >> I was talking to someone in politics [65:39] about this and the theory that they had [65:42] is well the Democrats aren't going to [65:45] pause the data centers forever. They're [65:47] going to pause them until they feel like [65:49] they're in enough control that they can [65:51] dictate all the rules. And so in other [65:53] words, they're calling this a moratorum. [65:56] And I think it does mean that the data [65:58] centers are going to stop but eventually [66:00] they're going to be in a position to say [66:02] okay here are our terms if you want to [66:04] turn these things back on right you want [66:07] to lift the moratorum and then that's [66:09] when we get this you know big government [66:11] democrat defined AI regime and you know [66:14] that it's going to consist of a new [66:16] regulatory agency and new speech [66:19] controls the whole trust and safety [66:20] agenda from social media will be ported [66:23] over that's this was this One person I [66:26] was talking to, this is what he was [66:27] speculating is the real agenda is that [66:30] eventually once Trump is no longer [66:32] president or in some future Democratic [66:35] administration, they will eventually [66:37] lift the moratorum but on their terms. [66:40] Now I think that's a really dangerous [66:42] thing to do because you know Trump is [66:44] president for another two years and then [66:46] no one knows what's going to happen [66:47] after that. And even if you lift the [66:50] moratorum in say 2 and 1/2 or 3 years, [66:53] it's going to take a couple of years for [66:54] those projects to even ramp back up. So [66:57] when you start talking about a [66:58] moratorium on data centers, it's not [67:00] like a a few month pause. It's probably [67:03] a good 5 years at least before you know [67:06] you can get another data center switched [67:08] on in the state of New York. Just so you [67:10] know how bad it's gotten, [67:13] there's a curve that you can use to [67:15] price data center assets. And I think [67:18] you guys know this, but I have this [67:20] portfolio of these assets that myself [67:22] and my partner Nita have accumulated. [67:24] And what's so interesting is when we [67:26] talk to all of the hyperscalers about [67:28] giving us a price because we're trying [67:30] to figure out whether we should keep it [67:32] or build it or just sell it, the most [67:34] incredible thing is how extreme the [67:36] price is at the front end of the curve. [67:37] When you have [67:39] verifiable energizable power today. [67:43] And the reason is exactly everything [67:44] that you're saying, Saxs, which is that [67:46] when you look out into the future, [67:49] you know, we've said this before, but [67:50] it's about 40% of all these projects are [67:52] getting mothballled and stopped. And so [67:54] it's creating this massive deficit of [67:57] available energy to actually drive [68:01] the use of AI. So to the extent that you [68:03] actually want, you know, drug discovery [68:05] or you want cancer diagnosis or you want [68:08] better healthcare or better legal [68:09] advice, we may actually not be able to [68:12] service it based on all of the demand [68:13] that exists because the power isn't [68:15] there. The energy isn't there. And the [68:16] reason why that's not there is because [68:18] folks are just kind of reflexively [68:22] protesting something that they don't [68:23] completely understand clearly. So I [68:26] think it's a really it's a really big [68:27] problem. [68:28] >> I mean, we're going to have GPUs chasing [68:30] energy. like where's their energy and [68:32] just drive the GPUs there is what's [68:34] going to happen right [68:34] >> let me add one layer to it which is [68:36] they're not only trying to stop data [68:38] centers from being built in the US [68:39] they're trying to stop data centers from [68:41] being built internationally in our [68:44] friends allies and partner countries and [68:46] the way they're doing that is the same [68:48] political forces that are stopping data [68:50] centers are also behind all these new [68:52] export controls on chips so they want to [68:54] make it harder and harder to export [68:55] chips to more and more countries [68:57] including our friends and allies and so [68:59] there's not going to data centers here. [69:01] There's not going to be data centers in [69:03] our allies. I mean, where are we going [69:05] to put these things? [69:05] >> I mean, those allies have unlimited [69:07] energy, Middle East. Like, if you want [69:10] some data centers, [69:11] >> well, what's funny, Jason, is, you know, [69:13] we did a bunch of Middle East data [69:15] center stuff and then it's kind of [69:16] stopped. [69:18] Meaning, like there wasn't this growth [69:20] that I thought would happen because it's [69:21] a very [69:24] conveniently placed geography. It's the [69:26] Middle East for a reason. And so, you [69:28] know, you can serve 4 billion people [69:30] very quickly in under 200 milliseconds [69:32] from there. [69:33] >> Instead, what happened was there was [69:34] this explosion in Asia and specifically [69:36] in Australia, which kind of surprised me [69:39] because I would have thought that those [69:40] folks are a little bit even further out [69:42] on the DSA, you know, far left. I [69:45] thought these things would not have [69:46] happened, but they they were able to get [69:48] big deals done. So, in this weird way, [69:50] you have all of these other countries [69:51] kind of running to try to embrace this [69:52] stuff quickly. They've done a decent [69:55] job. They're doing stuff to sort of like [69:58] displace some of the energy that that is [70:01] needed in the US. But the problem is we [70:03] need to have enough surplus here because [70:06] this is where most of the commerce is [70:08] going to get created. That really that I [70:09] think should [70:10] >> these are these are luxury regulations. [70:12] Like you can afford if you're New York [70:14] State or California to be like, you know [70:16] what, we don't need this. This is a [70:17] luxury for us to have an extra data [70:18] center. If you're Australia, you might [70:20] really need the money. If you're Texas, [70:22] you might really want the money. So this [70:24] is what's so crazy like virtue signaling [70:26] only goes so far until your debt to GDP [70:29] is high enough and or your productivity [70:31] is low enough andor your foreign direct [70:32] investment is low enough where you're [70:34] like all right you know what screw all [70:35] that we're just going to build a data [70:36] center but the the other thing is if you [70:38] saw what happened this week the UAE now [70:41] is able to import the best-in-class [70:43] leading chips and so to your point Jason [70:46] I think it restarts the cycle where you [70:48] have to look very carefully at the [70:49] Middle East because it's a very [70:50] attractive place to build these things [70:52] >> and by the way you saw [70:54] the you even if you just if you think [70:56] about fiber and the milliseconds as [70:58] you're talking about yes you can get to [70:59] the four billion people but I don't know [71:00] if you saw the the giant Starlink [71:04] versions now they make like a really big [71:06] version I think you got like one of the [71:08] enterprise versions but there's like an [71:09] even bigger enterprise version and they [71:11] can bundle them together and you're [71:12] starting to get to like 10 gig 20 gig [71:15] setups so that means you could start [71:17] putting these things almost anywhere [71:20] >> which gets also like adds another [71:22] >> can we see your clip. The thing that you [71:24] were mentioning before, [71:25] >> this is not to your point as prevalent [71:29] in the Middle East where you have [71:31] monarchies [71:33] and governments that aren't ruled by [71:35] democracy, but in democracies, we see [71:37] this anti-data center movement taking [71:39] hold. This chart is something that [71:43] for me always kind of played a role in [71:46] my understanding of where the incredible [71:48] anti-GMO sentiment came about in the [71:50] United States, [71:50] >> which is great. Russia Today, this [71:54] Russian media outlet launched in the US [71:56] in 2010. They were kicked out of the US [71:59] by Biden in 2022. [72:02] And you can see that prior to Russia [72:04] today existing in the US, there was no [72:06] anti-GMO sentiment. GMOs were around [72:08] since 1996. That's when they first had [72:10] their big commercial launch in the US [72:11] and were pretty prevalent for, you know, [72:14] 14 plus years before everyone started to [72:17] think GMOs are bad. We got to get rid of [72:19] GMOs. And you could ask people a hundred [72:21] different ways very pointedly and [72:23] specifically about the facts on the [72:25] matter and the science of GMOs and all [72:26] this sort of stuff, but everyone always [72:28] had a reason why they didn't want them, [72:30] similar to what we're hearing now with [72:32] AI and data centers. And it turns out [72:35] that if you track back all of the media [72:37] that had all this anti-GMO sentiment [72:39] that ultimately got picked up by the mom [72:41] bloggers that ultimately got put into [72:43] social media feeds that ultimately [72:44] everyone just accepted as truth. A lot [72:47] of it originated in this Russian media [72:49] push that happened around this era. And [72:52] you can actually see this on the Google [72:53] trend data that shows GMO and it's kind [72:56] of write up. And then as Russia Today [72:57] started to get cut by different media [72:59] outlets and people stopped retweeting [73:00] them and stopped reflecting them and [73:02] stopped writing articles that followed [73:04] Russia Today the anti-GMO sentiment [73:06] declined in the US and I think you can [73:09] see this going back decades. You know [73:11] there's this [73:13] effort that the KGB kind of designed [73:16] during the cold war called directed [73:18] measures which was really meant to try [73:20] and create an influence campaign through [73:23] affecting media. So putting this kind of [73:25] propaganda out through foreign media, [73:28] particularly targeted western [73:29] democracies. And you know, you could [73:31] argue that maybe you could trace back [73:32] what happened in Germany with nuclear [73:35] energy as being kind of similarly [73:36] originated. But there have been a series [73:39] of these pushes that seem nonsensical if [73:42] you're fairly rational and can have an [73:43] actually objective debate about the [73:45] scientific merit, the economic merit, [73:47] the benefits of these technologies. But [73:49] for some reason, what we call the [73:52] activist community become heightened to [73:54] them. Say that we've got to get rid of [73:56] them. And everyone's got these different [73:57] unfounded, scientifically unfounded [73:59] reasons why they want to get rid of [74:00] them. And you're like, wait a second, [74:02] how did we end up in this place that [74:04] we're literally handicapping ourselves. [74:06] And I think we're seeing something [74:07] similar happening with data centers in [74:09] the US today. The funding of the NOS's [74:11] as they're being called, the media [74:13] that's supporting this, the retweeting [74:15] of the media. And then you asked people, [74:16] there was a poll that came out today, [74:18] something north of 50% of Americans [74:20] believe that data centers increase the [74:23] cost of water and electricity. Even if [74:26] the data center is fully recycling the [74:27] water and they're producing their own [74:29] electricity, there's still this kind of [74:31] repugnant reaction to the data center. [74:34] And so there has been this like deeply [74:36] sewn psychological [74:38] shift that's happened in the United [74:40] States. and and and you know, people [74:41] have these, well, I hate the rich, I [74:43] hate tech, I hate AI, I don't want any [74:44] of this stuff, I don't want any of this [74:45] stuff, [74:46] >> but where does it all come from? I do [74:48] worry that there's some degree of kind [74:50] of call it foreign, [74:53] >> you know, influence [74:54] >> influence. I I don't love the word [74:56] influence because everyone kind of [74:57] everyone captures it up, but there is [74:58] some degree of like [75:00] >> I would say there's there's foreign [75:02] interest. Let's let's call it that. [75:04] >> No, I think it's more than that. So just [75:05] one month ago, just one month ago, [75:08] >> OpenAI published a blog post called PRC [75:11] linked influence operations are [75:13] targeting AI debates in the US and [75:16] Politico covered this and a lot of other [75:18] sites covered this. Basically what they [75:21] are saying and in fact many people are [75:23] saying is that China is behind a lot of [75:25] these influence campaigns to shape US [75:29] attitudes on AI data centers. [75:31] >> It makes sense. It makes a lot of sense. [75:33] and there's going to be a congressional [75:34] investigation of this. It does make [75:36] sense cuz it is in their interest, [75:38] right? If they can stop us from building [75:40] this necessary infrastructure, [75:42] >> then that's a way for China to win the [75:44] AI race. [75:45] >> If they can the market, if they can [75:48] incentivize anthropic to, you know, pull [75:52] the ladder up, if they can kill open [75:53] source in the United States and [75:55] constrain demand or the optionality and [75:58] choice of lower, cheaper models, think [76:00] about that for a second. at $56 per [76:03] million input tokens. I mean, versus 50 [76:06] cents for the rest of the world. All of [76:08] a sudden, it doesn't take a company [76:10] that's much much worse than you to beat [76:13] you when your cost is 50 to 100x more. [76:18] >> Right? [76:19] >> That's just the math. The math ain't [76:21] mapping. [76:21] >> You know, it's Sasha made the point that [76:24] these enterprises are not just paying [76:26] for AI with money. They're paying again [76:29] by feeding those frontier models their [76:31] proprietary knowledge, right? And all [76:34] their their alpha. So, it's like a [76:36] double whammy. [76:37] >> It's like it's more expensive and you're [76:40] potentially mortgaging your future. [76:41] >> Look, let's be honest. It is obvious [76:44] where foreign governments have an [76:46] enormous incentive to try to manipulate [76:48] and influence the comingings and goings [76:50] in America. [76:52] I think we should just acknowledge that. [76:54] The idea that that doesn't happen is [76:56] very naive. Now the question is we have [76:58] to be able to call it out and put our [77:00] finger on it because otherwise what is [77:02] clearly happening is that there's a lot [77:04] of Americans that will just fall for [77:05] this and they will not think from first [77:07] principles. [77:08] >> We have a huge we have a huge moral [77:09] panic going on with respect to AI. Look [77:12] when you talk about catastrophes that [77:14] could result from AI. What are we [77:15] talking about? We're talking about [77:16] things that might happen in the future. [77:18] Nothing resembling this has happened [77:20] yet. you know, even the cyber risk that [77:22] everyone's been talking about, [77:24] >> job loss [77:25] >> or job loss, it's like none of it's [77:27] turned out to be true. We haven't seen [77:28] any of it so far. But we're on the [77:30] threshold, I think, of destroying the [77:33] crown jewel of our economy, which is the [77:36] system of free market innovation that we [77:38] have, this culture of rapid iteration of [77:41] anyone with a good idea can go raise [77:43] risk capital and start their idea, start [77:46] their company. And we're on the verge, [77:48] you know, now we're talking I think [77:49] about how far the Overton window has [77:51] moved where we're actually saying that [77:53] creating a FINRA for our industry might [77:56] be better than all the alternatives. [77:58] FINRA is a bunch of stock brokers [78:01] writing rules. And when's the last time [78:03] there was ever any innovation in that [78:06] sector? I mean, I guess Robin Hood made [78:08] trading free. That was it. [78:09] >> That was a big one. Yeah. Flow. [78:11] >> Okay. But that's not real innovation. [78:13] Okay. That's like an innovation with [78:15] respect to a pricing model. And we're [78:17] actually saying that that might be the [78:18] least bad alternative is having the [78:20] equivalent of a bunch of stock brokers [78:22] creating new rules that all these AI [78:26] companies are not going to have to abide [78:27] by. It's crazy. We are going to we are [78:29] going to throw away the lead that we [78:31] have in this. And by the way, Kimmy K3 [78:34] just came out and people are saying it's [78:36] it's now right up there. It's very very [78:38] close to the frontier. We may have [78:40] months on China if that. and we're going [78:43] to create all these crazy rules and new [78:45] regulatory bodies for risks that have [78:47] not manifested yet. [78:48] >> It's worth monitoring the situation, but [78:51] it's not worth panicking. Like, you [78:52] should monitor the situation with [78:54] self-driving cars and job loss. China is [78:56] certainly doing that. They just stopped [78:57] giving out permits for self-driving cars [78:59] as an example because it's going so well [79:02] and they're losing jobs and and there [79:04] are people who are getting there's a [79:05] little civil unrest. So they just said [79:07] we're going to make self-driving cars um [79:10] licensed and so they're not giving out [79:11] any more license moratorium on license [79:13] for now. It's worth watching Mythos and [79:15] if it could hack your system. Palo Alto [79:17] Network's checking it out, other people [79:18] checking out. It's all worth monitoring [79:20] but yes it there's no disaster here [79:22] today because of AI. nothing's jumping [79:24] out of your chat GPT window. You know, [79:27] the worst case scenario is you blow out [79:28] some tokens, you know, okay, great. That [79:31] that's that's the biggest [79:32] >> There's only a handful of companies that [79:34] are even at the frontier and they all [79:36] have safety testing and red teaming and [79:39] all the rest of [79:39] >> doing a good job. [79:40] >> Yeah, I'm saying stop that. I'm just [79:43] questioning whether we need some vast [79:44] regulatory apparatus now to start doing [79:47] all [79:47] >> certainly premature and we did this [79:49] because of science fiction and Daario [79:52] saying all jobs are going away. I mean [79:54] that that that was the most ridiculous [79:56] thing when he said like he's panicked [79:57] that it's going to be 80 or 90% jobs in [79:59] 2026. What was his claim? I Nick get the [80:02] exact claim. I think he said 50% of [80:04] jobs. [80:05] >> He said he said with it he said 50% of [80:07] entry- level knowledge worker jobs are [80:09] going away within 1 to 5 years. That was [80:11] one year ago. [80:12] So [80:13] >> it's a little ridiculous. Yeah. I mean [80:15] it's but he's not even know how to use [80:17] the tools yet. [80:18] >> He's been in a state of panic since [80:20] GPT2. [80:21] >> Yes. Yes. I mean [80:23] >> remember they wanted they wanted to have [80:25] uh regulatory approval for models that [80:28] use 10 to the 25th flops. Right. And [80:31] every single AI model is like well past [80:33] that threshold now. [80:34] >> Yeah. [80:35] >> And we haven't seen any of the the harm. [80:38] Look, they thought that 10 to the 25th [80:41] flops would be enough compute to create, [80:44] you know, the Terminator, you know, to [80:46] create Skynet. [80:47] >> No offense, Freeberg, but one guy's [80:49] panic attacks, one guy's anxiety [80:51] condition might have shaped the whole [80:53] course of history here like Dar does [80:55] Daario have like I'm not making light of [80:57] it, but does he have an anxiety issue [80:59] where he's like overly concerned about [81:01] this stuff or is it just delusions of [81:02] grandeur? Come on the pod, Dario. [81:04] Invite's open. Come hang out. I'm sure [81:06] he'd love to come on the pod after you [81:07] just accused him of having a panic [81:09] attack, but [81:11] >> I mean, he seems like he's in a [81:12] perpetual one. [81:14] >> No, let me tell you, listen, I it could [81:16] be psychological, but I actually think [81:18] that there's a strategy that makes a lot [81:19] of sense, and it's a very simple, [81:21] straightforward strategy. Number one, [81:23] brand yourself as a safe AI company. [81:25] Number two, ban unsafe AI. Three, [81:29] profit. [81:30] >> Yeah, there you go. [81:31] >> That's the strategy. Kind of brilliant. [81:33] All right, everybody. Go to [81:35] allin.com/events [81:37] and sign up for the Allin Summit in [81:39] September. Scholarships are open. Let's [81:42] do a quick amazing deep robust [81:46] science corner with our boy David [81:48] Friedber. Before we get into the science [81:50] corner, I'm going to give a shout out to [81:51] Ronnie Dog for adoption. I love family [81:55] dog rescue in Sonoma. Check out his [81:57] Instagram link. God, here he goes [81:58] >> in the description. This dog needs a [82:00] home. He was fostered and he lost the [82:02] foster home. Someone come and grab him. [82:04] He's awesome. All right, let's get into [82:06] this. [82:07] >> This is what we're doing. He's trying to [82:09] >> he's trying to get more Q points. [82:11] >> That dog looks delicious. [82:14] >> Gross. [82:17] >> You don't live in Sri Lanka anymore. [82:19] Chimal. [82:20] >> Yeah, seriously. [82:22] >> Sri Lanka taking a spray. [82:25] >> How do you marinate that dog in Sri [82:26] Lanka? [82:27] >> Do what you got to do in Sri Lanka. Salt [82:29] and pepper free or do you like something [82:31] else? You know, just a little salt and [82:33] pepper. [82:33] >> 12 hour marinade. Do you use bolet? [82:36] >> Oh, do do you like do you like a little [82:38] yogurt and garam masala? Maybe do a [82:39] little [82:40] >> spicy. Do you guys want to talk about [82:43] reversing aging? [82:44] >> Yes. So, I want to talk about but I got [82:45] to drop. All right, guys. I got to go to [82:48] the EIFFEL TOWER. PAUL, [82:49] >> I'LL COVER IT. THE audience will stick [82:51] around. [82:52] >> Science corner. All right. So, Jamal, [82:56] you can drop two if you want. I'll cover [82:58] science corner solo. So in the past [83:00] we've talked about Yamanaka factors [83:02] which are these proteins that can go [83:03] into cells and reverse the aging of the [83:05] cell and the cell starts to act young [83:07] again. Pretty amazing. And uh there's a [83:10] lot of advancement happening on that [83:11] front. But this paper that came out just [83:13] this week that everyone's kind of going [83:15] crazy about was put out uh jointly by [83:17] Calico which is Google's kind of you [83:19] know age reversal startup that's super [83:21] secretive that they're not allowed to [83:23] talk about in partnership with a group [83:25] called Revel Pharma. And what they [83:28] focused on was what's called the [83:29] extracellular matrix. The parts outside [83:32] of the cell that age. And and what does [83:34] aging actually look like outside of the [83:36] cell? Well, over time, sugars and fats [83:40] bind to proteins in the area between our [83:43] cells and they accumulate. They don't [83:45] get cleaned off. And as they accumulate [83:46] and they don't get cleaned off, they [83:48] make it harder for your body to clean [83:50] out that area, to maintain that area. It [83:52] causes stickiness. It causes binding. [83:55] And that reduces mobility. and [83:57] ultimately leads to things like wrinkles [83:58] in our skin, makes it harder for our [84:00] joints to move. [84:01] >> Is that what visceral fat is? [84:02] >> No, it's called glycation. And so it's [84:05] the binding of sugar and fat to the [84:08] proteins that sit in that extra cellular [84:12] >> in between the cells. Exactly. And so [84:14] it's that whole gunky area in between [84:16] the cells that when you're young works [84:19] well, everything's smooth. The proteins [84:21] get replaced if they break down. And as [84:23] you get older, sugars and fats kind of [84:25] stick to these proteins, block them up. [84:27] And as they get blocked up, your body [84:29] can't repair them. It can't clean them. [84:30] And more importantly, it changes the [84:32] structure and the shape of those [84:33] proteins. So things like collagen that [84:36] are far apart stick together. And that [84:38] causes things like wrinkles and that [84:39] causes immobility. And it also causes [84:42] inflammation because then those proteins [84:45] kind of look different than they're [84:46] supposed to. And your body starts to [84:48] attack them. And that activates [84:50] inflammation. And that's why we get one [84:51] of the reasons why we get more and more [84:52] inflammation as we get older. And so one [84:55] of the key what are called advanced [84:56] glycation end products that's the term [84:59] for these things is called CML. CML is [85:02] kind of the predominant molecule that [85:04] that gets formed in this extracellular [85:07] matrix that's driving aging and nothing [85:10] breaks it down. So these scientists set [85:12] out to try and create an enzyme. An [85:13] enzyme is a protein that breaks [85:15] something down um that can break down [85:17] CML. And remember, a protein is just a [85:20] series of amino acids. And those amino [85:22] acids are programmed by DNA. So you can [85:24] put three letters of DNA to make an [85:25] amino acid. So you can literally just [85:27] print DNA and then put it in a bacteria [85:29] to print proteins and then test those [85:31] proteins to see what they do. That's the [85:33] the modern kind of era of kind of [85:36] protein synthesis and protein testing. [85:38] And so these guys kind of went out and [85:39] they took the target which is CML and [85:42] tried to figure out okay how do we [85:43] actually degrade CML clear that [85:46] extracellular matrix and reverse aging. [85:49] And they started with AlphaFold and they [85:51] used Alphafold to find a protein that [85:53] could bind to CML and activate an [85:55] enzyatic or process that would break it [85:57] down. And then they took that protein [85:59] from AlphaFold that comes out of a [86:01] bacteria. They produced it. They started [86:03] to test it and then they started to find [86:06] some of the binders or the parts of that [86:08] protein that they could make better and [86:09] they used you know DNA programming to [86:12] change it and they made hundreds and [86:14] then thousands of variants of it to [86:16] measure activity which is how good is it [86:18] at breaking down the CML and they did [86:20] this recursively five different cycles [86:23] and then eventually they tested it once [86:25] they' kind of gotten it breaking down [86:26] the CML really well in a test tube they [86:28] started to test it on the proteins that [86:31] we would find in our body, cassine, [86:34] collagen, retinal proteins which are in [86:36] your eye, hemoglobin, [86:38] and they were able to get rid of 52 to [86:42] 97% of the CML, just degraded away. [86:46] >> And then they um they found several [86:48] sites where they were able to degrade [86:49] over 90%. And then they took actual [86:52] human skin from elderly patients that [86:55] had donated their skin and they put this [86:58] enzyme onto that skin and they were able [87:02] to eliminate 55% of the CML on the skin [87:05] which basically reversed the skin's age [87:08] down to the age of a 31y old. This is [87:10] from greater than 70 year old patients [87:12] just by putting this enzyme on the skin. [87:14] And so it's kind of a groundbreaking [87:17] demonstration of combination of [87:19] alphafold what's called directed [87:21] evolution where you change the order of [87:23] the DNA that changes the structure of [87:25] the protein to test different proteins [87:27] do high throughput screening and [87:29] ultimately make a novel protein that [87:31] doesn't exist in nature today that can [87:33] do something pretty profound for human [87:34] health. And now the next set of [87:36] questions is okay well great this enzyme [87:38] is awesome. How are we going to get it [87:39] into our bodies? How are we going to get [87:40] it into that extracellular matrix? Is it [87:42] going to be a cream? Is it going to be a [87:44] shot, a supplement? Could we eventually [87:46] take an RNA shot that makes the protein [87:48] inside of our body and starts to do the [87:50] degradation from within? A lot of [87:53] questions kind of still to be answered, [87:55] but it really, I think, lights a great [87:56] path forward. [87:57] >> Amazing [87:58] >> for these novel therapies that we're [87:59] developing. It's [ __ ] awesome. I [88:01] mean, dude, likeing, [88:02] >> you know, all my I got all these joint [88:03] pains in my hip and my shoulder now. [88:05] Like everything you can feel yourself [88:07] getting older. [88:07] >> Well, that's I I will tell you this [88:09] right now. That will not be the first [88:10] market. The first market will be [88:11] cosmetic and [88:12] >> cosmetic skin. Yeah. [88:14] >> It will be a trillion dollar market. If [88:17] if you can create a cream, [88:18] >> dude, if you could put this enzyme [88:20] literally on your skin and have it [88:21] absorb [88:22] >> a cream. [88:22] >> Yeah. [88:23] >> Game over. It's it's a that that alone [88:25] is $2 trillion. [88:27] >> But I mean, dude, AI, let's just talk [88:29] about applications of AI, why it's [88:31] actually awesome that everyone should be [88:33] able to agree on and you can't be [88:34] convinced by some foreign scop. This is [88:36] [ __ ] awesome. I mean, this was alpha [88:38] full used to discover this thing and [88:40] evolve it and drive this outcome. [88:42] Everyone can benefit from it. It's just [88:44] so profound that we have this tool at [88:46] our disposal in this day and age. [88:48] >> I think it's pretty awesome. Anyway, [88:49] thanks for sticking around for Science [88:50] Corner. [88:51] >> Guys, I love you. [88:53] >> All right, bro. Love you, too. [88:56] >> We'll let your winners ride. [89:03] We open sourced it to the fans and [89:05] they've just gone crazy with it. [89:08] >> Queen of [89:16] besties are [89:19] my dog taking your driveways. [89:24] >> Oh man, my habitasher will meet. [89:26] >> We should all just get a room and just [89:28] have one big huge orgy cuz they're all [89:29] just useless. It's like this like sexual [89:31] tension that they just need to release. [89:37] >> Your feet. [89:40] >> We need to get merch. [89:41] >> I'm going all in. [89:49] I'm going all in.