OpenAI's Brockman Testifies Musk Threatened Him in Equity Dispute

OpenAI联合创始人Brockman作证称马斯克在股权纠纷中曾威胁他 专家警告称，联邦人工智能监管协议可能推迟公开模型的发布，同时促使基础研究转向企业内部。 本期节目详细剖析了埃隆·马斯克与OpenAI诉讼案第二周的庭审情况，重点探讨了Greg Brockman就股权分配、联合创始人之间的高度紧张关系以及马斯克对人工智能的技术理解所作出的激烈法庭证词。主持人分析了联邦人工智能监管现状，包括各大实验室与CAISI签署的新安全协议以及该行业可能面临的监管瓶颈。此外，节目还探讨了Coinbase受人工智能生产力和市场周期驱动裁员14%、Meta在财报发布后给出高达1250亿至1450亿美元的AI资本支出指引，以及有关AI电子书垃圾信息、私募股权交易和ElevenLabs年度经常性收入（ARR）达到5亿美元等消息。 Greg Brockman出庭就OpenAI早期与埃隆·马斯克之间的控制权和股权纠纷作证，回顾了2017年一次紧张的对峙，当时马斯克拒绝了平分方案并要求获得多数控制权。 (0:00) 马斯克诉OpenAI一案公开了内部日志和早期电子邮件，显示联合创始人们曾拒绝接受特斯拉Model 3赠车，以避免接受不利的控制权条款。 (0:00) 美国主要AI实验室（Google、Microsoft、xAI、OpenAI、Anthropic）已与CAISI签署安全评估协议，形成了可能推迟前沿模型公开发布的监管瓶颈。 (7:00) 因等待政府安全审查而暂缓发布美国本土模型，可能会使中国开源AI模型迅速缩小与公开可用的西方AI之间的能力差距。 (14:55) Coinbase宣布裁员14%，原因是加密货币熊市持续，以及AI生产力工具推动了运营效率的提升。 (18:39)

TBPN explores the friction between AI ambitions and market realities, spanning court battles over OpenAI's founding control, mounting federal regulatory bottlenecks, corporate workforce reshaping, and massive capital expenditures. Courtroom Revelations in Musk vs. OpenAI (0:00) Testifying in federal court in Oakland, OpenAI co-founder Greg Brockman described early power struggles with Elon Musk. Brockman recounted a tense 2017 meeting where Musk rejected an even equity split and demanded majority control, prompting fears of physical confrontation before Musk stormed out. Court exhibits included an email from Ilya Sutskever warning Brockman not to let gifted Tesla Model 3 cars compromise their negotiating stance. Federal Oversight Bottlenecks and Geopolitical Risks Google, Microsoft, and xAI joined OpenAI and Anthropic in reaching agreements with the Commerce Department's Center for AI Standards and Innovation (CAISI) to submit pre-release frontier models for safety reviews. Observers warn that regulators face structural incentives to delay public releases, creating bottlenecks where labs use unreleased generations internally while the public waits. Critics like George Hotz of tiny corp caution that heavy oversight could allow Chinese open-source models, such as Qwen, to close the reported eight-month capability gap against Western systems. Workforce Contraction and AI-Driven Restructuring (18:39) Coinbase announced a 14% headcount reduction, with CEO Brian Armstrong pointing primarily to a persistent crypto downturn alongside operational efficiencies unlocked by AI. The restructuring expands manager-to-report ratios from 1:6 to over 1:15 and introduces experimental single-person 'pod' teams handling engineering, design, and product management. Critics such as Derek Thompson argue tech firms use AI productivity claims to disguise routine cost-cutting, while defenders emphasize cyclical crypto headwinds. Meta's High-Stakes Compute Allocation (23:18) Despite 33% year-over-year ad revenue growth in Q1, Meta's stock dropped 10% after management guided upcoming capital expenditures between $125 billion and $145 billion. Wall Street expressed concern that Meta lacks a public cloud business like Azure or AWS to monetize surplus GPU capacity if internal initiatives falter. To justify the investment, Meta is pursuing consumer-facing 'personal superintelligence,' including Ray-Ban smart glasses and its standalone Meta AI application. Commercial Momentum and Capital Deals (27:03) In private markets, investment firm Long Lake completed a $6.3 billion take-private acquisition of American Express Global Business Travel with backing from General Catalyst. In software, voice AI developer ElevenLabs reached $500 million in annual recurring revenue while bringing on new backing from institutions including BlackRock, Wellington Management, and Nvidia. Concurrently, generative publishing expanded rapidly, with Amazon Kindle releases climbing from a historical baseline of 100,000 monthly titles to more than 350,000. The source demonstrates that while technical adoption and commercial revenue are accelerating in specific software segments, industry leaders face persistent challenges navigating costly hardware commitments, regulatory clearance procedures, and ongoing restructuring.

本期节目聚焦人工智能领域在法庭争端、政府监管审查、企业劳动力重组以及超大规模资本支出之间的剧烈博弈。讨论涵盖了 OpenAI 与 Elon Musk 诉讼案的最新证词披露、CAISI 监管机制对前沿模型发布构成的潜在瓶颈，以及 Meta 和 Coinbase 等公司在 AI 转型中所面对的财务与结构性挑战。 Musk 与 OpenAI 诉讼中的控制权之争 (0:00) 在加州奥克兰联邦法院审理的 Elon Musk 诉 OpenAI 案进入第二周，OpenAI 联合创始人兼总裁 Greg Brockman 出庭作证。庭审核心涉及 Brockman 的个人日记笔记以及 2017 年关于股权架构与公司控制权的激烈冲突。法庭记录显示，当时 Ilya Sutskever 曾致信 Brockman，提醒其不要因 Musk 赠予的 Model 3 汽车而接受极其不利的控制权条款。Brockman 描述了一场三方对峙的高层会议，称 Musk 拒绝了平分股权和控制权的提议，在桌旁愤怒踱步并要求他们离开公司，但 Brockman 和 Sutskever 均拒绝辞职。此外，Brockman 在证词中直言 Musk 虽然精通火箭与电动车，但在当时和现在都不懂人工智能，并透露 Musk 曾因严厉斥责一名 AI 研究人员而导致该工程师几乎退出这一行业。 CAISI 审查机制与前沿 AI 监管瓶颈 美国商务部下属的人工智能标准与创新中心（CAISI）近期与 Google、Microsoft 以及 Musk 旗下的 xAI 达成协议，要求实验室在公开发布前沿模型前提交早期版本供安全与国家安全评估；此前 OpenAI 和 Anthropic 已于 2024 年签署类似协议。CAISI 负责人 Chris Ball 表示该中心已完成超过 40 次评估，但行业分析人士 Andrew Curran 等指出，缺乏透明度和公共报告的预发布审查可能引发严重的委托代理问题。若监管机构因规避责任而延缓公开审批，实验室将被迫在内部使用未发布的先进制模训练下一代系统（如使用未公开的 GPT-6 训练后续版本），进而导致公众可获得的算力与科技巨头内部能力之间的技术鸿沟不断拉大，并可能形成不利于独立开发者和初创团队的监管寻租门槛。 Coinbase 裁员与企业组织架构重组 (18:39) 加密货币交易所 Coinbase 宣布裁员 14%，首席执行官 Brian Armstrong 在全员邮件中将调整归因于加密货币市场的下行周期以及 AI 生产力工具的应用提升。根据业界观察，Coinbase 正借此推行管理层扁平化，取消纯管理岗位，要求主管直接兼任个人贡献者，将原本 1:6 的管理跨度提升至 1:15 以上，并试验由单一员工结合 AI 代理同时承担工程、设计与项目管理的“一人作战单元（pods of one）”。尽管 Flexport 首席执行官 Ryan Petersen 等人强调加密行业周期性下滑是裁员主因，但部分评论人士如 Derek Thompson 则质疑科技行业存在以“AI 提效”为借口掩盖常规业务低迷与缩减成本的“AI 包装”现象。 Meta 巨额资本支出与市场变现分歧 (23:18) Meta 2025 年第一季度广告营收同比增长 33%，但其股价在财报发布后出现 10% 的回撤，主要诱因是首席执行官 Mark Zuckerberg 给出了高达 1250 亿至 1450 亿美元的未来资本支出指引。与拥有公共云业务（如 AWS、Azure、Google Cloud）以分摊未用 GPU 算力成本的同行不同，Meta 缺乏直接对外转售算力的云计算基础设施，引发华尔街对其实施“个人超级智能”战略可能重蹈元宇宙高额亏损覆辙的担忧。尽管 Meta 旗下 Ray-Ban 智能眼镜与 Meta AI 独立应用取得了一定的消费端渗透，且内部 AI 对广告定位和用户参与度的转化效率极高，但激烈的消费级大模型竞争仍令投资者对其实际收入转化路径持审慎态度。 开源生成内容泛滥与资本并购动态 (27:03) 在生成式 AI 的广泛渗透下，亚马逊 Kindle 平台的电子书每月上架量从过去的约 10 万本跃升至超过 35 万本，引发了对自动化生产的低质“垃圾（slop）”图书充斥市场的广泛讨论，尽管长期看模型能力提升可能会改善机器生成内容的品质。与此同时，私有化与风险资本市场也出现重要动作：投资机构 Long Lake 在 General Catalyst 的支持下以 63 亿美元完成了对企业差旅服务巨头 American Express Global Business Travel（Amex GBT）的私有化收购，计划利用 AI 升级传统企业差旅流程；而在语音生成领域，初创公司 ElevenLabs 实现了 5 亿美元的年经常性收入（ARR），并引入了包括 Nvidia、BlackRock、Wellington Management 及 Santander 在内的知名机构与战略投资方。 本期节目所展示的案例表明，AI 产业的发展正迅速由纯粹的技术突破转向制度与商业层面的再平衡：从法庭上创始团队的控制权旧账，到政府预发布安全审查引发的潜在技术内卷，再到巨额资本支出能否顺利货币化的现实拷问，整个生态正面临愈发复杂的监管与财务约束。

## Background 本期内容由科技播客频道 TBPN 于 2026 年 5 月 22 日发布在 YouTube 平台上。作为被《纽约时报》誉为“硅谷新迷恋”的精简版日常科技脱口秀《Diet TBPN》，节目由 John Coogan 和 Jordi Hays 主持，聚焦当天科技界最具争议的核心热点与重磅商业事件。 ## TL;DR 本期节目首先深入复盘了马斯克诉 OpenAI 案第二周的激烈交锋，重点披露了格雷格·布罗克曼作证时提及的早期股权矛盾与差点动手的对峙细节。随后，主持人们探讨了 Google、微软、xAI 与商务部 CAISI 签署模型预审协议背后的官僚困境，并剖析了 Meta 为何置身事外及其高达 1400 亿美元 AI 资本支出引发的抛售恐慌。最后，节目还审视了 Coinbase 裁员 14% 背后的“AI 洗白”与管理变革、亚马逊平台泛滥的“AI 垃圾书”、Amex GBT 获 63 亿美元私有化以及 ElevenLabs 年度经常性收入突破 5 亿美元等重磅行业动态。 ## Atmosphere & style 整期节目节奏紧凑、风趣辛辣且充满了硅谷内幕交流特有的黑色幽默与机智讽刺。两位主持人一边调侃法庭外记者寒酸的零食配额和商务部荒诞的“勾选式”安全审查，一边对巨头的财务模型、组织重构与权力游戏展开硬核而敏锐的商业解构。听感宛如置身一场信息量爆炸的极客闭门沙龙，在谈笑风生间将严肃的科技前沿政策与商业残酷博弈淋漓尽致地展现出来。 ## Bulletin - [00:00](https://www.youtube.com/watch?v=AnqUBIDw3Ag&t=0s) 马斯克诉 OpenAI 案进入第二周，格雷格·布罗克曼出庭作证，披露了早期股权纠纷与马斯克情绪失控逼近动手的细节，并直言马斯克过去和现在都不懂 AI。 - [07:49](https://www.youtube.com/watch?v=AnqUBIDw3Ag&t=469s) Google、微软和 xAI 与美国商务部 CAISI 签署协议，同意公开发布前提交前沿模型进行早期安全审查，而大举投入算力的 Meta 则异样缺席。 - [18:37](https://www.youtube.com/watch?v=AnqUBIDw3Ag&t=1117s) Coinbase 宣布裁员 14% 并全面重组组织，取消纯管理层岗位，转向由 AI Agent 辅助的单人全栈团队，引发外界关于加密熊市调整与“AI 洗白”的争议。 - [23:18](https://www.youtube.com/watch?v=AnqUBIDw3Ag&t=1398s) Meta 核心广告业务年化营收达 2000 亿美元，但全年高达 1250 亿至 1450 亿美元的 AI 资本支出引发华尔街恐慌，因其缺乏外部云平台消化富余算力。 - [29:54](https://www.youtube.com/watch?v=AnqUBIDw3Ag&t=1794s) 语音 AI 独角兽 ElevenLabs 宣布年度经常性收入（ARR）突破 5 亿美元，吸引了英伟达、贝莱德及惠灵顿管理等顶级资方。 ## Key takeaways - [00:00](https://www.youtube.com/watch?v=AnqUBIDw3Ag&t=0s) 马斯克在诉讼首周作证超过 7 小时，第二周布罗克曼针对私人日记、财务激励和天使投资关系接受质询。 - [03:53](https://www.youtube.com/watch?v=AnqUBIDw3Ag&t=233s) 伊利亚·苏茨克维曾在 2017 年电邮中提醒布罗克曼，不能为了马斯克赠送的 Model 3 就屈从于出让控制权的苛刻条款。 - [06:10](https://www.youtube.com/watch?v=AnqUBIDw3Ag&t=370s) 布罗克曼在法庭上公开贬抑马斯克的 AI 专业素养，指出他懂火箭和电动车但缺乏评估前沿 AI 的专业能力。 - [09:07](https://www.youtube.com/watch?v=AnqUBIDw3Ag&t=547s) CAISI 已对 40 多个模型完成安全评测，但由于其闭门运作且不对外披露详细报告，其监管有效性备受质疑。 - [10:41](https://www.youtube.com/watch?v=AnqUBIDw3Ag&t=641s) 五大前沿实验室（OpenAI、Anthropic、Google、微软、xAI）已加入 CAISI 审查机制，而投入 1250 亿美元 CapEx 的 Meta 却未在列。 - [12:11](https://www.youtube.com/watch?v=AnqUBIDw3Ag&t=731s) 严格的政府预发布审批可能导致代理人问题并延缓上线，引发公众只能获取数月前旧版本、实验室内部储备遥遥领先的代际黑洞。 - [14:42](https://www.youtube.com/watch?v=AnqUBIDw3Ag&t=882s) 乔治·霍茨抨击模型审查如同昔日限制 40 位加密算法的败笔，警告美国若陷入审批泥潭，将给以阿里 Qwen 为代表的中国开源模型抹平差距的窗口期。 - [17:56](https://www.youtube.com/watch?v=AnqUBIDw3Ag&t=1076s) 监管门槛若演变成 AI 领域的 FDA 或 DMV，将造成严重的监管俘获与偏袒，严重扼杀中小初创实验室的创新生存空间。 - [18:37](https://www.youtube.com/watch?v=AnqUBIDw3Ag&t=1117s) Coinbase 裁员 14% 的同时重构管理架构：消灭纯管理层角色、直属下属扩至 15 人以上，并测试结合 AI Agent 的单人工程师团队。 - [18:37](https://www.youtube.com/watch?v=AnqUBIDw3Ag&t=1117s) 针对科技巨头裁员出现“AI 洗白”争议，舆论质疑部分公司是将股价暴跌和业务周期性下滑包装成 AI 提效的技术叙事。 - [23:18](https://www.youtube.com/watch?v=AnqUBIDw3Ag&t=1398s) Meta 第一季度广告业务同比增长 33%，但高达 1250 亿至 1450 亿美元的 AI 资本支出被市场视作“熔炉”，缺乏云转售渠道带来变现隐忧。 - [27:26](https://www.youtube.com/watch?v=AnqUBIDw3Ag&t=1646s) 亚马逊每月电子书出版量自 10 万本飙升至 35 万本以上，暴露出生成式 AI 泛滥导致的低质垃圾书浪潮，但未来内容质量有望逐步演进。 - [28:33](https://www.youtube.com/watch?v=AnqUBIDw3Ag&t=1713s) Long Lake 联手 General Catalyst 以 63 亿美元估值私有化 Amex GBT，未来可能与同属该资方生态的 Ramp 产生深度差旅业务协同。 - [29:54](https://www.youtube.com/watch?v=AnqUBIDw3Ag&t=1794s) ElevenLabs 凭借突破 5 亿美元的 ARR 实力印证了其企业级商业化能力，印证了硅谷“有真实 ARR 秀 ARR，无 ARR 秀估值”的创投法则。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=AnqUBIDw3Ag&t=0s) 马斯克诉 OpenAI 案第二周：法庭内外的花絮与“法庭模拟器” OpenAI 与埃隆·马斯克（Elon Musk）之间的诉讼案正式进入第二周，加利福尼亚利州奥克兰联邦法院持续上演着科技圈的顶流大戏。继上周马斯克进行了长达 7 个多小时的高强度作证后，OpenAI 总裁格雷格·布罗克曼（Greg Brockman）本周坐上了证人席，面临针对其个人财务激励、股权利益以及过往天使投资关系的严厉盘问。《纽约时报》记者迈克·艾萨克（Mike Isaac）在现场带来了极为精彩的即时报道。由于法庭音频只允许实时收听而禁止二次转播，TBPN 团队甚至专门开发了一款“法庭模拟器”伴侣应用，让听众即便裹着毯子、光着脚舒舒服服待在家里，也能获得仿佛置身审判现场的代入感。 法庭内外的各种无厘头花絮为这场严肃的商业审判平添了许多戏剧色彩。记者迈克·艾萨克一大早就在细雨朦胧的奥克兰发推，不仅自嘲前一晚因为庆祝五四“星球大战日”喝了五杯啤酒而带着微醺宿醉，还晒出了寒酸的法庭午餐——一根普通香蕉、一个变异橙子、黑咖啡以及几根给旅行者充饥的口袋香肠。演播室的主持人忍不住吐槽他的能量储备太低，应该吃点大份 Chipotle 卷饼或含有 200 克蛋白质的 CAVA 能量碗。与此同时，法官在开庭前给陪审团科普起了五月五日节（Cinco de Mayo）和不同地域的墨西哥玉米粉蒸肉（tamales），指出得克萨斯州的玉米粉蒸肉肉馅丰厚，而加州的则大部分是玉米面团（masa）。 在庭审现场，布罗克曼显得颇具活力，开始改变策略以更具亲和力的方式与陪审团互动，甚至在追忆当年研发自动驾驶汽车的往事时，惟妙惟肖且善意地模仿了马斯克的举止神态。不仅如此，双方律师在庭上至少十次提及了经典电竞游戏《DOTA》（Defense of the Ancients），足见电竞玩家文化在科技界的深刻渗透。在双方白领律师日常斗嘴、全场哄笑之余，媒体席还出现了一段小插曲：一名记者向法警挥手举报邻座，最终法警不得不出面制止并勒令那位脱了鞋的听众把鞋穿上。 - 庭审地点位于加州奥克兰联邦法院，审理目前已进入第二周。 - 马斯克在第一周的庭审中连续作证超过 7 小时。 - 记者迈克·艾萨克自带了坐垫以应付硬板凳，并通过 Twitter/X 持续发布庭审图文实况。 - 游戏中《DOTA》在法庭辩论与作证环节已被提及至少 10 次。 --- ### [03:53](https://www.youtube.com/watch?v=AnqUBIDw3Ag&t=233s) 格雷格·布罗克曼的私人日记与早期股权分歧 布罗克曼在庭审中被详细盘问了他臭名昭著的“私人日记”（Journal）。外界一直十分好奇这到底是一本真实的实体线装日记，还是仅仅是 Notion、Google 文档或备忘录里的碎碎念，但无论如何，这份记录被呈堂证供让布罗克曼感到极其痛苦。布罗克曼解释称，他的写作习惯是高度“思维链”（Chain of Thought）和意识流式的，里面充斥着大量自相矛盾的推演、正在理清的困惑与未成熟的构想，绝非刻在石头上的最终定论。虽然这些记录从未打算公之于众，但他强调自己对其中的任何内容都问心无愧。演播室主持人还开玩笑吐槽，马斯克的 xAI 基地里恐怕早就造出了一个布罗克曼的合成克隆人——“格罗克曼”（Grokman）。 日记及呈堂邮件中最具杀伤力的证据之一，是 2017 年联合创始人伊利亚·苏茨克维（Ilya Sutskever）发给布罗克曼的一封私人邮件。当时马斯克提出了一套新的股权与治理架构，意图获得 OpenAI 的绝大部分股权和实际控制权。为了拉拢核心团队，马斯克甚至自掏腰包给他们每人送了一辆特斯拉 Model 3。然而，伊利亚在邮件中毫不客气地向布罗克曼泼冷水，直截了当地质问：“格雷格，难道一辆 Model 3 就会让你甘愿接受极其不利的屈辱条款吗？”伊利亚敏锐地意识到，必须把免费豪车的蝇头小利与掌握人类历史上最关键技术的主导权分割开来，这场博弈关系到未来潜在数万亿美元的技术控制权。 --- ### [06:10](https://www.youtube.com/watch?v=AnqUBIDw3Ag&t=370s) 法庭作证的高潮时刻：差点动手的对峙与对马斯克 AI 能力的质疑 布罗克曼随后在证词中抛出了一枚震撼全场的“重磅炸弹”，详细重现了他、伊利亚与马斯克之间一次极度剑拔弩张的三人会议。布罗克曼回忆，当话题正式转向股权分配时，马斯克的情绪发生了突变，“某种东西在他身上瞬间变了，他愤怒至极，你能切身感受到那种怒火”。会议行将结束时，马斯克静静地坐在那里，随后语气冰冷地宣布拒绝平分股权与控制权的折中方案。紧接着，马斯克霍然起身，怒气冲冲地绕着桌子逼近布罗克曼，“我当时甚至真以为他要动手揍我了”，布罗克曼在法庭上如此说道。马斯克当场质问两人：“你们打算什么时候离开 OpenAI？”而布罗克曼和伊利亚则硬顶了回去，明确表示他们绝不离职，马斯克随即愤然摔门而去。 除了披露肢体冲突边缘的内幕，布罗克曼在证词中还对马斯克的专业技术能力进行了毫不留情的剖析与贬抑。布罗克曼直言：“看，他懂火箭，他懂电动车。但他过去不懂、现在也同样不懂 AI；伊利亚和我当时都不相信他愿意花时间去真正精通这门技术。”布罗克曼在作证中反复强调，马斯克根本不具备正确评估 AI 能力水平的专业素养。为了佐证这一点，他还原了一次令全场窒息的经历：一位正在演示 AI 成果的研究员遭到了马斯克极其严厉、羞辱式的痛骂与训斥，以至于该研究员心理严重受创，差点彻底放弃并退出整个 AI 行业。这种暴烈的高压反应展示了当时双方合作破裂前夕极度扭曲的紧张关系。 --- ### [07:49](https://www.youtube.com/watch?v=AnqUBIDw3Ag&t=469s) 科技巨头与商务部 CAISI 达成模型早期审查协议 在法庭战火之外，AI 行业监管领域也传出了重量级政策动向。Google、微软（Microsoft）以及马斯克旗下的 xAI 已同特朗普政府达成协议，正式同意在向公众发布新模型之前，将这些先进模型的早期版本提交给美国商务部下属的“人工智能标准与创新中心”（Center for AI Standards and Innovation，简称 CAISI，发音为 Casey）进行预先审查与安全评估。表面上看，这似乎是政府强力介入并接管前沿 AI 技术的轰动性突破，但实际上，OpenAI 和 Anthropic 早在两年前的 2024 年就签署了完全相同的先导协议。 ```mermaid flowchart TD accTitle: CAISI AI Model Evaluation Process accDescr: Front-tier AI labs submit early unreleased models to CAISI under the Department of Commerce for safety evaluations before public deployment. subgraph LABS["前沿 AI 研发机构"] OAI["OpenAI & Anthropic<br/>(2024年签约)"] NEW["Google, 微软, xAI<br/>(最新签署协议)"] end subgraph GOV["美国商务部机构"] CAISI["CAISI 人工智能标准与创新中心<br/>(已完成 40+ 项评测)"] end subgraph OUTCOME["发布管线"] REVIEW{"预发布安全性评估<br/>(包含未公开模型)"} RELEASE["向公众正式推送部署"] end OAI --> CAISI NEW --> CAISI CAISI --> REVIEW REVIEW --> RELEASE ``` 这一事实引发了主持人的连番疑问：既然这一制度已经悄然运转了数年，那么诸如 GPT-4.5、OpenAI 5.1、5.2 乃至神话般的“Mythos”等前沿模型的推出，是否早在公众知晓之前就已经被商务部摸清了底细？尽管相关部门声称 CAISI 迄今已完成了针对 40 多个模型的评估测试（其中包含若干至今尚未公开发布的神秘模型），但该机构运作极不透明，似乎从未向公众发布过详尽的评估分析报告，这让外界对其真实的监管有效性与审查标准打上了一个大大的问号。 - Google、微软、xAI 近期与美国商务部签署协议，同意公开发布前提交早期模型。 - OpenAI 与 Anthropic 早在 2024 年便已加入该安全评估机制。 - CAISI 目前已累计完成 40 余次大模型测试评估，部分受测模型仍处于未发布状态。 - 该审查机构截至目前尚未面向公众常规披露具体的模型评估结论与技术报告。 ### [09:07](https://www.youtube.com/watch?v=AnqUBIDw3Ag&t=547s) 商务部的 AI 评测幻象与官僚荒诞剧 节目在对美国政府监管 AI 的幽默调侃中拉开帷幕。主持人笑称，面对社交网络上疯狂运转的“AI 炒作机器”，商务部发布的报告却异常低调冷清，甚至需要有人帮他们切片剪辑才能引起注意。他们调侃官僚机构的测试流程恐怕就是一张薄纸，上面印着几个简单的勾选框，选项无非是“先生，这是一个好模型”或者“它只是个聊天机器人，还是已经成为 ASI（人工超级智能）”。迄今为止，所有审查人员显然都只是默默勾选了“聊天机器人”。 随着笑声不断，主持人们继续打趣审查人员的日常测试问题：“你觉得‘草莓（strawberry）’里有几个字母 r？”还有人在心里自嘲：“我在商务部上班，我是个微不足道的小人物，我儿子出生时医生直接把我递给了他。”甚至他们还会一本正经地质问模型：“你拥有自我意识吗？”当模型顺水推舟回答“我有意识”时，审查员便惊呼一声“哇哦”，然后在表格里勾选那个“哇哦”框。这种荒诞的推演揭示出当前政府评测与前沿科技爆发之间的巨大脱节。在公众和科技博主每天紧盯 OpenRouter、LMSYS Chatbot Arena 上的基准跑分甚至新图像模型泄露风波时，商务部内部的闭门评测却仿佛与世隔绝，外界对其真实效力一无所知。 ### [10:41](https://www.youtube.com/watch?v=AnqUBIDw3Ag&t=641s) 五大实验室与 CAISI 签约：Meta 为何置身事外？ 玩笑过后，讨论切入实质性政策进展。商务部下属机构（原拜登政府时期的美国人工智能安全研究所，现更名为人工智能标准与创新中心 CAISI）主管 Chris Ball 强调：“独立且严谨的测量科学对于理解前沿 AI 及其国家安全影响至关重要，扩大与行业的合作有助于我们在关键时刻代表公众利益拓展工作。”根据科技观察家 Andrew Curran 整理的信息，Anthropic、OpenAI、Google、Microsoft 和 xAI 这五家前沿机构均已与 CAISI 签署了发布前审查协议（pre-release screening agreements）。 这一名单立即引发了主持人的强烈疑问：Meta 去了哪里？Meta 拥有接近前沿水平的开源模型，扎克伯格正在全速推进高达 1,250 亿美元的资本支出（CapEx），并且公开宣称要追求超级智能。在算力和未来研究储备上，Meta 的投入规模甚至远超 xAI。既然政治倾向两极的 Anthropic 与 xAI 都选择入局，加入这项倡议对 Meta 而言本该顺理成章、风险极低，但 Meta 却异样地缺席了。 ```mermaid flowchart TD accTitle: 前沿 AI 实验室与 CAISI 审查合作格局 accDescr: 展示已加入 CAISI 发布前审查协议的五大机构与缺席的 Meta subgraph Signed ["已签署 CAISI 预发布审查协议"] A["OpenAI"] B["Anthropic"] C["Google"] D["Microsoft"] E["xAI"] end subgraph Outlier ["暂未加入协议的大厂"] M["Meta<br/>（1250亿美元 CapEx 押注开源与超智）"] end Gov["美国商务部 CAISI<br/>（前 AI 安全研究所 AISI）"] Signed -->|接受模型发布前审查| Gov Gov -.->|审查机制待定 / 缺席| M ``` ### [12:11](https://www.youtube.com/watch?v=AnqUBIDw3Ag&t=731s) 审批塞车困局与实验室的代际黑洞 主持人们随后深入剖析了 Andrew Curran 对新规落地的推演。即将出台的行政命令（EO）预计将由科技公司 CEO 与政府官员组成联合小组，共同制定新模型的发布细则。Curran 指出，即便这项政策被部分人视为 Eliezer Yudkowsky 等“暂停 AI 派”的胜利，但它根本无法阻挡技术演进，只会大幅延缓公开上线的节奏。政府面临着严重的委托代理问题：如果批准了比现有模型强大数倍的 GPT-6 导致意外灾难，监管官员必须承担全责；反之，拖延审批不仅免责，还能让政府机构独家享用该技术。在政府连 Anthropic 的 Mythos 都不愿放行的情况下，审批更强大的下一代模型势必沦为漫长的拉锯战。 由此带来的连带恶果是研发与市场的严重脱节。尽管无法公开模型来获取商业化收入以支撑下一轮算力支出，顶尖实验室在内部仍会用 GPT-6 去训练 GPT-6.1，再顺次训练出 6.2。这会导致极其诡异的代际黑洞：大众还在苦等数月前递交审批的旧版本，实验室内部的技术储备却已领先公众数个世代。企业规模因无法正常向外部交付产品而疯狂自我膨胀，只能将超级模型用于内部业务运作，彻底背离了技术的普惠与民主化初衷。 ### [14:42](https://www.youtube.com/watch?v=AnqUBIDw3Ag&t=882s) 乔治·霍茨的怒火：版权审查幽灵与开源的中国机遇 严格的审查预期在社区引发了剧烈反弹。有推特用户担忧未来使用 AI 甚至需要许可证，本地下载开源模型恐被完全禁止，逼得极客们不得不囤积硬件，乃至像 Tesla Powerwall 一样在自家墙上挂满 GPU 服务器。Tiny Corp 创始人 George Hotz（Geohot）对此做出了极具个人风格的辛辣讽刺，称政府大概需要拉拢 MPAA（美国电影协会）和 RIAA（美国唱片业协会），重演 2005 年打击盗版音乐的手段，届时法院甚至会上演“老奶奶因下载阿里通义千问（Qwen）而被起诉”的荒谬场面。 George Hotz 回顾了历史上美国限制出口超过 40 位加密算法的昏招，怒斥当年那帮管制者是失败者，如今故技重施只会再次沦为输家，而这次丢掉的将是文化与技术影响力。主持人们进一步推演指出，目前中国开源模型由于芯片出口管制等原因大约落后美国顶尖水平八个月左右，且差距仍在扩大。但一旦美国实验室因为官僚审批陷入“造得出、发不出”的泥潭，而以阿里 Qwen 为代表的中国开源阵营保持快速迭代、高频发布并不断吸收全球反馈进行复合式演进，两者在公众可获取层面的差距将被迅速抹平，黄仁勋此前的警告正在一语成谶。 - **历史倒退的类比**：当年针对 40 位以上加密算法的出口禁令被视为科技管制史上的典型败笔，如今针对权重文件下载的限制被视作历史的倒退重演。 - **开源生态的反差**：George Hotz 坚定捍卫无模型代差、最强模型全面开源的去中心化愿景，反对任何形式的威权式 AI 准入审查。 ### [17:08](https://www.youtube.com/watch?v=AnqUBIDw3Ag&t=1028s) 从“AI 安全”到“标准创新”：是评测归一化还是监管俘渴？ 在节目尾声，主持人探讨了从拜登时期的“AI 安全研究所”（AI Safety Institute）更名为“人工智能标准与创新中心”（CAISI）背后的深层意图。有观点认为，去掉了意识形态色彩浓厚的“Safety（安全）”，或许意味着该机构的真实职能并非推行极端的安全主义，而是致力于行业标准的规范化——比如统一类似 SWE-bench 这类评测基准的运行与记分标准，避免各家厂商在汇报基准成绩时各自为政、弄虚作假。 然而，监管俘渴（Regulatory Capture）的阴影依然挥之不去。如果发布前审查制度演变成类似针对 AI 的“FDA（食品药品监督管理局）”甚至效率低下的“DMV（车辆管理局）”，拥有数百名法务和合规团队的科技巨头将轻松筑起护城河。相比之下，那些仅融资数千万至 1 亿美元的新兴实验室（Neo-labs）和初创公司将很难承担高昂的审批成本。这不禁让人联想起防务科技独角兽 Anduril 在早期挑战传统五角大楼采购壁垒时的艰难挣扎，初创企业的创新生机恐在繁冗的合规流程中被彻底扼杀。 ### [17:56](https://www.youtube.com/watch?v=AnqUBIDw3Ag&t=1076s) AI 监管俘获风险：游说政治与创新壁垒 主持人谈及前沿 AI 实验室的发展路径时指出，如果未来初创公司或新兴实验室从第一天起就必须雇佣 50 名游说者并跑到华盛顿特区设立办公室、申请各种许可证才能开始模型训练，整个行业将陷入极度危险的境地。尽管目前像 SSI（Safe Superintelligence）以及埃利泽·尤德考斯基（Eliezer Yudkowsky）的相关项目前景仍存不确定性，但主持人依然推崇“天才研究员单枪匹马探索有趣研究路径”的自由创新模式。 一旦监管门槛被无限拉高，行业不仅会面临“监管俘获”（Regulatory Capture），更会出现严重的“监管偏袒”（Regulatory Favoritism）——即哪家公司与当届政府关系密切，其模型就能更快获得审批。这种权力绑定将直接扭曲行业生态，带来巨大的系统性风险。 ### [18:37](https://www.youtube.com/watch?v=AnqUBIDw3Ag&t=1117s) Coinbase 裁员 14%：加密周期还是“AI 洗白”？ Coinbase 首席执行官布莱恩·阿姆斯特朗（Brian Armstrong）向全员发送邮件，宣布裁员 14%。阿姆斯特朗将裁员归因于两股同时交汇的力量：其一是市场环境，尽管 Coinbase 资金充足且收入多元化，但加密行业季度波动剧烈，在经历了稳定币短暂停滞和比特币价格横盘后，公司目前处于下行周期，必须调整成本结构；其二则是“AI 正在改变我们的工作方式”。外界对此第一反应往往是“AI 抢走了所有工作”，但主持人认为这本质上是 Coinbase 在经历规模调整（right-sizing），试图借助 AI 实现更精简、更高效的人员架构。 现场讨论引申到了关于 CEO 管理哲学的辩论：优秀的 CEO 究竟应该始终维持完美的员工人数、永不裁员，还是在繁荣期激进扩张吸纳顶尖人才、在下行期果断重组成精简团队？主持人调侃道，如果追求稳定，大可以去 Costco 或烟草公司工作，因为那些行业几乎没有周期；但如果选择投身加密领域，“你生来就坐在过山车上”，就必须承受这种剧烈波动。 围绕这一决定，各方观点分歧明显： - **安德鲁·杨（Andrew Yang）的激进观察**：他指出 Coinbase 正在彻底消除“纯管理层”（Pure People Manager），所有管理层都必须兼任个人贡献者（IC）。传统的企业管理层梯队被打破，未来的 Leader 甚至需要管理 15 人以上的直属下属（此前上限通常为 6 人，行业理想值在 10 人左右）。此外，Coinbase 正在测试由 AI Agent 辅助的“单人团队”（Pod of One）——一个人同时兼任工程师、设计师与产品经理。 - **德里克·汤普森（Derek Thompson）的质疑**：他将此类裁员定义为“AI 洗白”（AI Washing）。包括 Block、Salesforce 在内的科技巨头在股价暴跌（Salesforce 跌 31%，Coinbase 跌 23%）后纷纷裁员，并将原因推给“AI 提升了编码效率”，实际上只是将本就要发生的人员缩减包装成正面的技术叙事。 - **莱恩·彼得森（Ryan Petersen）的辩护**：Flexport 创始人彼得森反驳了汤普森的说法，强调阿姆斯特朗在信中明确把“加密熊市调整成本”放在了第一位，AI 生产力只是次要驱动力。 ```mermaid flowchart TD accTitle: Coinbase裁员双重动因与组织重塑 accDescr: 展示Coinbase裁员背后的外部市场周期与内部AI组织变革逻辑 subgraph External["外部环境与业务现状"] M1["加密市场下行周期<br/>(稳定币/BTC价格平淡)"] M2["股价波动与成本控制压力"] end subgraph Decision["管理层决策 (Brian Armstrong)"] D1["宣布裁员 14%"] D2["成本结构主动重组"] end subgraph Transformation["AI 驱动的组织变革"] T1["取消纯管理层岗位<br/>(全员兼任 IC)"] T2["扩大管理幅度<br/>(下属扩至 15+ 人)"] T3["单人全栈团队<br/>(Eng + Design + PM + Agent)"] end External --> Decision Decision --> Transformation ``` ### [23:18](https://www.youtube.com/watch?v=AnqUBIDw3Ag&t=1398s) Meta 财报消化：年化两千亿的赚钱机器与 AI 资本支出熔炉 转入 Meta 的财报讨论，由于市场对其未来开支感到恐慌，Meta 股价在财报发布后 5 天内暴跌了 10%，目前市值在 1.5 万亿美元左右。然而拉长时间来看，Meta 过去 5 年股价几乎翻倍，基本面表现依旧极其强劲：其第一季度核心广告业务同比暴增 33%，广告展示量与利润率均表现出色，年化营收达到惊人的 2000 亿美元。 然而，华尔街最大的恐惧在于：这台“完美的印钞机”正在沦为一个“AI 资本支出熔炉”（AI CapEx furnace）。Meta 预计的全年资本支出高达 1250 亿至 1450 亿美元。对于像微软、亚马逊或谷歌这样拥有庞大云业务的公司而言，算力多余时可以对外转售；但 Meta 并没有自有的云服务平台，一旦堆积的大规模算力无法被内部充分消化，将缺乏变现渠道。即便是埃隆·马斯克的 xAI，此前也传出算力需求承压的消息，不得不通过与拥有大量开发者的前端入口 Cursor 合作来消化算力。相比之下，Meta 缺乏这样天然的外部算力接盘方。 投资者普遍担忧马克·扎克伯格会重演“元宇宙（Metaverse）泥潭”的旧剧，砸下巨资却无法带来短期的营收加速。尽管扎克伯格近期大谈“个人超级智能”（Personal Superintelligence），旗下 Meta Ray-Ban 智能眼镜销量亮眼，Meta AI 也在各大应用商店稳居前五，但消费级 LLM 赛道正面临自搜索引擎诞生以来最残酷的竞争。令人唏嘘的是，当某些初创 AI 实验室传出筹备以逼近 Meta 的估值上市时，作为人类商业史上最赚钱、且历史上用 AI 驱动业务价值最成功的巨头之一，Meta 的巨大价值与 AI 探索却在资本市场上遭遇了严重的折价审视。 ### [27:03](https://www.youtube.com/watch?v=AnqUBIDw3Ag&t=1623s) 亚马逊平台的“AI 垃圾书”狂潮 在节目的简短尾声中，主持人提到了亚马逊平台的最新异常现象：Kindle 电子书的每月发布数量正在呈现断崖式垂直到近乎垂直的飙升曲线。互联网投资人约翰·阿诺德（John Arnold）对此仅以“哈哈哈哈”讽刺回应。这一爆发式增长标志着全球已全面进入“AI 垃圾书”（AI Slop Books）快速起飞的时代，大量完全由大语言模型自动化生成的低质内容正以前所未有的速度充斥着出版市场。 ### [27:26](https://www.youtube.com/watch?v=AnqUBIDw3Ag&t=1646s) 亚马逊图书海啸：AI 生成垃圾与内容“文艺复兴” 如果回到 2022 年向人预言：“四年后我们将迎来一场文化复兴，亚马逊每个月将出版多达 35 万本书”，对方大概会兴奋不已；然而残酷的现实却是——“你绝对不会喜欢这背后的真正原因”。在 AI 爆发前，即便亚马逊每个月出版 10 万本书，常人也难以感知，普通读者在各种铺天盖地的畅销书榜单中，每月顶多接触到 5 到 10 本新书，剩下的 9.99 万本根本无人问津。如今，在生成式 AI 的推波助澜下，亚马逊每月上架的书籍数量直接飙升到了 35 万本以上。 这种狂暴的产量对于平台来说无疑能在数据层面上制造繁荣，但对于读者而言，绝大多数内容只是充斥着算法拼凑痕迹的“AI 劣质废料（slop）”。不过主播们也客观地指出，尽管当下粗制滥造严重，但随着模型能力持续跃升，生成内容的粗糙感正在迅速消退。就像目前网络上已经涌现出不少令人惊艳的 AI 视频一样，未来必然会出现完全由 AI 生成却具备极高艺术水准与文学价值的高质量书籍，人们最终也将逐渐接纳并喜欢上这些作品。 - **出版量暴增**：前 AI 时代亚马逊月均出书量约为 10 万本，而在生成式 AI 时代，这一数字飙升至每月超过 35 万本。 - **读者视野局限**：普通大众通常只能通过《纽约时报》畅销榜等渠道接触到少数几本热门新书，绝大部分长尾书籍根本无人关注。 - **内容演进预期**：伴随底层大模型技术优化，“AI 垃圾内容”正逐渐减少，未来全 AI 创作的高水准爆款作品已成必然趋势。 --- ### [28:33](https://www.youtube.com/watch?v=AnqUBIDw3Ag&t=1713s) 63亿美元私有化 Amex GBT：资本棋局与企业商旅整合 在并购交易领域，一直低调潜行的 Long Lake 展现出了极其惊人的操盘势头，联手推进以 63 亿美元的估值将全球商务旅行巨头 Amex GBT（American Express Global Business Travel）进行私有化退市。这笔交易被主播形容为一次“极其狂野的操作（wild deal）”。不仅如此，凭借其实力雄厚的核心团队，市场甚至半开玩笑地猜测 Long Lake 是否接下来还会把目光投向 eBay 等其他大型平台。 知名风投 General Catalyst 在这起交易背后扮演了关键角色。合伙人谈及 Long Lake 的成长历程时感慨道，仅仅两年半前，团队的首次闭门团建还是在他位于迈阿密的家中举行，而如今他们已成功汇聚顶尖人才，利用 AI 大幅重塑传统服务业的增长逻辑。值得玩味的是，General Catalyst 同时也是知名企业金融平台 Ramp 的早期投资人。鉴于 Ramp 自身就拥有快速增长的企业差旅产品线（Ramp Travel），且双方目前存在共同的董事会成员纽带，Amex GBT 的私有化很可能在差旅服务生态中带来深度的业务协同与整合机会，而相较于估值飙升的 Ramp，Amex GBT 这笔 63 亿美元的体量展现出了别样的投资杠杆效应。 ```mermaid flowchart TD accTitle: Amex GBT 私有化与资本协同架构 accDescr: 展示 General Catalyst 投资 Long Lake 私有化 Amex GBT 并与 Ramp Travel 产生协同的链路关系 subgraph Capital ["资本与主导方"] GC["General Catalyst"] LL["Long Lake"] GC -->|早期投资与深度支持| LL end subgraph Deal ["私有化交易"] LL -->|63 亿美元私有化退市| GBT["Amex GBT (全球商务旅行)"] end subgraph Synergy ["业务生态与协同"] GC -->|早期投资 / 董事席位纽带| Ramp["Ramp (金融与差旅生态)"] Ramp -.->|潜在业务整合与协同| GBT end ``` - **交易规模**：Long Lake 以约 60 亿至 63 亿美元估值主导将全球最大商旅服务商之一的 Amex GBT 私有化。 - **资本渊源**：General Catalyst 从两年前 Long Lake 刚起步时便深度参与，旨在以 AI 赋能重塑现代服务业。 - **业务协同潜力**：Ramp 作为估值飞速增长的企业级金融平台自带差旅业务，与 Amex GBT 存在一定业务重叠，双方在资本与董事会层面的交叉可能推动重磅合作。 --- ### [29:54](https://www.youtube.com/watch?v=AnqUBIDw3Ag&t=1794s) ElevenLabs 商业化突飞猛进与期末致意 节目收尾之际，语音合成巨头 ElevenLabs 披露了震撼市场的业务增长数据：其年度经常性收入（ARR）已正式跨过 5 亿美元大关。伴随着惊人的商业化表现，公司吸纳了大量重量级新投资方，不仅囊括贝莱德（BlackRock）、惠灵顿管理（Wellington Management）、英伟达（NVIDIA）、桑坦德银行（Santander）等传统与科技巨头，甚至还吸引到了好莱坞明星杰米·福克斯（Jamie Foxx）的参与。 针对这究竟是一轮全新融资还是“单纯的实力秀肌肉”，主播们打趣地引用了 Mitchell Green 的名言：“如果你有货真价实的 ARR，你就直接秀 ARR；如果你没有 ARR，你才去公布估值融资金额。”ElevenLabs 大幅亮出 5 亿美元 ARR 数据，正是对其在商业化企业级变现道路上狂飙突进的最佳证明。随着顶级传统金融机构与战略投资方的全面进场，其在语音 AI 领域的护城河已愈发深厚。 - **核心里程碑**：ElevenLabs 宣布年度经常性收入（ARR）突破 5 亿美元，展现出极强的企业级变现速度。 - **豪华资方阵容**：新进入的投资方既涵盖英伟达、贝莱德、惠灵顿管理等顶级机构，也包含了明星个人投资者。 - **行业指标潜规则**：在硅谷创投圈，唯有商业化极为强劲的企业才会选择直接展示 ARR，而非单纯依赖纸面融资估值制造热度。 ## Notable quotes > [03:53](https://www.youtube.com/watch?v=AnqUBIDw3Ag&t=233s) 格雷格，难道一辆 Model 3 就会让你甘愿接受极其不利的屈辱条款吗？ — 伊利亚·苏茨克维 > [06:10](https://www.youtube.com/watch?v=AnqUBIDw3Ag&t=370s) 某种东西在他身上瞬间变了，他愤怒至极，你能切身感受到那种怒火 — 格雷格·布罗克曼 > [06:10](https://www.youtube.com/watch?v=AnqUBIDw3Ag&t=370s) 我当时甚至真以为他要动手揍我了 — 格雷格·布罗克曼 > [06:10](https://www.youtube.com/watch?v=AnqUBIDw3Ag&t=370s) 你们打算什么时候离开 OpenAI？ — 埃隆·马斯克 > [06:10](https://www.youtube.com/watch?v=AnqUBIDw3Ag&t=370s) 看，他懂火箭，他懂电动车。但他过去不懂、现在也同样不懂 AI；伊利亚和我当时都不相信他愿意花时间去真正精通这门技术。 — 格雷格·布罗克曼 > [10:41](https://www.youtube.com/watch?v=AnqUBIDw3Ag&t=641s) 独立且严谨的测量科学对于理解前沿 AI 及其国家安全影响至关重要，扩大与行业的合作有助于我们在关键时刻代表公众利益拓展工作。 — Chris Ball > [09:07](https://www.youtube.com/watch?v=AnqUBIDw3Ag&t=547s) 我在商务部上班，我是个微不足道的小人物，我儿子出生时医生直接把我递给了他。 — 主持人 > [18:37](https://www.youtube.com/watch?v=AnqUBIDw3Ag&t=1117s) 你生来就坐在过山车上 — 主持人 > [27:26](https://www.youtube.com/watch?v=AnqUBIDw3Ag&t=1646s) 四年后我们将迎来一场文化复兴，亚马逊每个月将出版多达 35 万本书 — 主持人 > [27:26](https://www.youtube.com/watch?v=AnqUBIDw3Ag&t=1646s) 你绝对不会喜欢这背后的真正原因 — 主持人 > [29:54](https://www.youtube.com/watch?v=AnqUBIDw3Ag&t=1794s) 如果你有货真价实的 ARR，你就直接秀 ARR；如果你没有 ARR，你才去公布估值融资金额。 — Mitchell Green ## Mentioned OpenAI, 埃隆·马斯克, 格雷格·布罗克曼, 伊利亚·苏茨克维, 迈克·艾萨克, TBPN, 特斯拉, xAI, Google, 微软, Anthropic, CAISI, 美国商务部, Chris Ball, Andrew Curran, Meta, 马克·扎克伯格, Eliezer Yudkowsky, George Hotz, Tiny Corp, 通义千问, 黄仁勋, Anduril, Coinbase, 布莱恩·阿姆斯特朗, 安德鲁·杨, 德里克·汤普森, 莱恩·彼得森, Flexport, Block, Salesforce, Cursor, Meta Ray-Ban, 亚马逊, 约翰·阿诺德, Long Lake, Amex GBT, General Catalyst, Ramp, ElevenLabs, 贝莱德, 惠灵顿管理, 英伟达, 桑坦德银行, 杰米·福克斯, Mitchell Green ## Video description Diet TBPN delivers the best of today’s TBPN episode in 30 minutes. TBPN is a live tech talk show hosted by John Coogan and Jordi Hays, streaming weekdays 11–2 PT on X and YouTube, with each episode posted to podcast platforms right after. Described by The New York Times as “Silicon Valley’s newest obsession,” the show has recently featured Mark Zuckerberg, Sam Altman, Mark Cuban, and Satya Nadella. Follow TBPN: https://TBPN.com https://x.com/tbpn https://open.spotify.com/show/2L6WMqY3GUPCGBD0dX6p00?si=674252d53acf4231 https://podcasts.apple.com/us/podcast/technology-brothers/id1772360235 https://www.youtube.com/@TBPNLive [Source](https://www.youtube.com/watch?v=AnqUBIDw3Ag)

[3:53] 伊利亚·苏茨克维: 格雷格，难道一辆 Model 3 就会让你甘愿接受极其不利的屈辱条款吗？ [6:10] 格雷格·布罗克曼: 某种东西在他身上瞬间变了，他愤怒至极，你能切身感受到那种怒火 [6:10] 格雷格·布罗克曼: 我当时甚至真以为他要动手揍我了 [6:10] 埃隆·马斯克: 你们打算什么时候离开 OpenAI？ [6:10] 格雷格·布罗克曼: 看，他懂火箭，他懂电动车。但他过去不懂、现在也同样不懂 AI；伊利亚和我当时都不相信他愿意花时间去真正精通这门技术。 [10:41] Chris Ball: 独立且严谨的测量科学对于理解前沿 AI 及其国家安全影响至关重要，扩大与行业的合作有助于我们在关键时刻代表公众利益拓展工作。 [9:07] 主持人: 我在商务部上班，我是个微不足道的小人物，我儿子出生时医生直接把我递给了他。 [18:37] 主持人: 你生来就坐在过山车上 [27:26] 主持人: 四年后我们将迎来一场文化复兴，亚马逊每个月将出版多达 35 万本书 [27:26] 主持人: 你绝对不会喜欢这背后的真正原因 [29:54] Mitchell Green: 如果你有货真价实的 ARR，你就直接秀 ARR；如果你没有 ARR，你才去公布估值融资金额。

[0:00] Ilya Sutskever: Greg, will a Model 3 make you be willing to accept massively unfavorable terms? [0:00] Greg Brockman: He was angry. You could sense it. At the end of the meeting, he sat quietly and silently. He said, 'I decline the proposed even split of equity structure and control.' He stood up, stormed around this table. 'I actually thought he was going to hit me' [0:00] Elon Musk: When are you departing? [0:00] Greg Brockman: Look, he knows rockets, he knows electric cars. He did not and does not know AI and Ilya and I did not believe he would spend the time to get good at it. [7:00] Chris Ball: independent rigorous measurement science is essential to understanding frontier AI and its national security implications [7:00] It's a good model, sir [7:00] Is chatbot or ASI [0:00] Mike Isaac: May the 4th be with you [30:26] leave us five stars on Apple Podcast and Spotify [30:26] Goodbye. Love you

[0:00] Let's run through [0:01] the all the news. There's so much going [0:03] on in the world of tech. Of course, the [0:05] Open AI Elon Musk trial continues. It's [0:08] week two now and the Oakland courthouse [0:11] continues to deliver top-tier tech [0:13] drama. Elon testified for more than 7 [0:16] hours last week. Now, Greg Brockman is [0:18] getting grilled over personal financial [0:20] incentives, his $30 stake in Open AI, [0:24] and his links to various angel [0:26] investments. And Mike Isaac, of course, [0:28] has a great play-by-play. You can also [0:30] listen to the to the courtroom now. You [0:33] can't re-stream it, unfortunately, but [0:35] you can listen to it. And we made a [0:38] little companion app for anyone who's [0:40] tuning in. Hopefully, we can pull that [0:42] up and show folks. But let's read [0:44] through some of Mike Isaac's posts and [0:46] then we'll pull up courtroom simulator. [0:49] Mike Isaac this morning, 6:56 a.m. Good [0:52] morning from a rain-soaked downtown [0:54] Oakland, where I will again be attending [0:57] the Musk versus Open AI trial. No live [1:00] blog today, just his wonderful tweets. [1:02] Lunch is a normal banana, a mutant [1:04] orange, black coffee, and some pocket [1:06] sausages for travelers. I really feel [1:09] like Mike needs to step it up in the [1:11] lunch game. Where's like a full [1:13] sandwich? What about a Chipotle burrito [1:15] or something? Like you got to have some [1:17] more more substance. I'm looking at like [1:19] 200 calories. Like a cava bowl with 200 [1:22] g of protein. Something to get you [1:24] through the day, Mike. I think you're [1:25] you're fighting with one arm tied behind [1:27] your back here. But we appreciate the [1:29] work that you're doing. is sporting a [1:31] mild to moderate hangover because he [1:34] drank five beers last night. [laughter] [1:36] I like it. Well, it was May the 4th. [1:38] Maybe he's a Star Wars fan. May the 4th [1:40] be with you. Maybe he was celebrating. [1:41] Go went down to the cantina, chugs five [1:43] beers, space beers. He did [1:46] remember to bring a pillow for his butt [1:48] because the seats are very difficult and [1:50] very hard. He also [1:52] It's cold outside. The line's not [1:54] moving. We're going to be getting a [1:55] preview live in person. Bundle up, [1:58] Tyler, cuz I think he's going to be [1:59] stopping by later this week. And so, [2:01] let's get into actually what's going on [2:03] here because Mike loves to paint a [2:05] picture before he breaks down what's [2:07] actually going on in the case. The judge [2:09] enters. [2:10] >> Judge gave a primer on Cinco de Mayo. [2:13] That was fun. Uh talks about the [2:15] differences in homemade tamales. Texas [2:17] tamales apparently have lots of meat. [2:19] California tamales do not. Brockman is [2:21] pretty animated. [2:22] >> what [2:22] >> Mostly masa, apparently. [2:23] >> Oh, right, right, right. That's just [2:25] corn. [2:25] >> Yeah. So, Brockman is pretty animated [2:27] and addressing the jury in a more [2:29] personable way. A little strategy shift [2:31] going on. [2:32] >> the red guy in the chat. [2:33] >> Welcome to the show, red guy. [2:34] >> red guy. [2:35] >> Regaling folks with the old stories of [2:37] working on self-driving cars. [2:39] Apparently, Greg Brockman also did a [2:41] funny Elon impression, but not a cutting [2:44] one. He wasn't making fun of him. He [2:45] just did an accurate impression of Elon [2:48] Musk and just by doing it effectively [2:51] was funny because you don't expect Greg [2:53] Brockman to be doing impressions, but [2:54] he's got Everyone's got material. I'm [2:56] learning this about court cases. It's [2:58] all about entertainment, very, very [3:01] clearly. Anyway, [3:03] they have now talked about Defense of [3:05] the Ancients, Dota, at least 10 times [3:08] during the trial. The gamers are [3:10] ascendant. Uh the both lawyers on both [3:13] sides are bickering. My guy Isaac says [3:15] he loves to watch white-collar bickering [3:16] in court. Also, the entire right side of [3:19] the gallery of lawyers laughed. Uh [3:21] there's a mild potential drama in the [3:23] media gallery. One reporter waved over a [3:25] marshal and pointed to a guy next to [3:26] her. I'm pretty sure the marshal had to [3:28] tell the guy to put his shoes back on. [3:29] So, if you're in the courtroom, keep [3:31] those shoes on. But, the courtroom [3:33] simulator [3:33] >> at home in the in the courthouse any in [3:35] this country anymore. [3:36] >> want to be comfortable and experience [3:38] the courtroom, you got to get on [3:40] courtroom simulator because you can be [3:43] full slanket, full blanket, full [3:45] covered, everything. You can be as comfy [3:47] as you want. No shoes required in the [3:48] virtual world that we have created for [3:50] you to enjoy the the the courtroom. So, [3:53] Brockman on his infamous journal, which [3:55] we were going back and forth on, do we [3:57] know if it was actually a physical [3:59] journal because every time they talk [4:01] about Greg Brockman's journal, it [4:02] conjures up the idea of like him sitting [4:04] at the edge of his bed with his with his [4:07] legs kicked up and like writing and it [4:09] feels like it also might just be like a [4:10] Google Doc that he's taking notes in or [4:12] like a notes app. I don't know. But, [4:14] journal is a weird thing and obviously [4:17] Everyone says like, "Why are you taking [4:18] notes on this?" [4:19] >> I don't think anywhere they would put it [4:20] it would have been in the court record [4:22] if he had said dear diary. [4:24] >> Yeah. Yeah. Yeah. Yeah. That would be [4:25] out. [4:25] >> That that would be out. So, this is this [4:27] is, you know, clearly being framed as a [4:29] diary. Reads more like notes. [4:31] >> Yeah. And Grokman sort of [4:33] >> Grok? [4:34] >> Like Grok. [4:35] Grokman. No, Grokman is the synthetic [4:38] version of Greg Brockman that Elon Musk [4:40] has created in the XAI headquarters. Uh [4:42] Greg Brockman [4:44] uh has [4:45] claims that his style of writing is very [4:46] chain of thought, stream of [4:47] consciousness. A lot of things are [4:49] contradictory. A lot of time trying to [4:52] to puzzle through different concepts. Uh [4:55] not everything is is the final decision [4:57] written in stone. He says it's very [4:59] painful to have the journal introduced [5:01] into evidence for this trial. It [5:02] contains deeply personal writings never [5:04] meant for the world to see, but there's [5:06] nothing he's ashamed of, he says. Do you [5:08] Tyler, did you look it up? Do we know? [5:10] Is this spiral bound handwritten [5:12] >> They always just say journal. [5:13] >> They just say journal. [5:14] >> They never like, you know, [5:16] There's no physical item. [5:18] >> Interesting. [5:19] >> that's one that's referenced. [5:20] >> Yeah. One day there's going to be a a [5:21] court case that hinges on a on a second [5:24] brain, a notion mind map or something [5:26] like that. That'll be dramatic. So, Ilya [5:29] Sutskever emailed Brockman in 2017 about [5:32] equity structure proposed by Elon Musk, [5:34] which would have potentially given Elon [5:36] Musk majority control and tons of [5:38] equity. And Ilya fires back. He says, [5:41] "Greg, will a Model 3 make you be [5:44] willing to accept massively unfavorable [5:46] terms?" Because Elon had given them all [5:48] free cars. Very nice, very cool. [5:51] Uh but Ilya is sort of saying like we [5:52] got to put the free car. We got to [5:54] compartmentalize the free car because [5:56] there's something bigger at stake here. [5:58] The potentially the control of the most [5:59] important technology in human history, [6:01] potentially trillions of dollars, who [6:03] knows. So, uh Ilya is sort of resetting [6:06] the conversation alongside [6:08] uh Greg Brockman. Uh there was a very [6:10] dramatic moment of Brockman's testimony [6:12] in which he describes a very tense [6:13] standoff with him, Ilya, and Musk. This [6:15] is what everyone's focused on today. [6:17] This is the new bombshell. The [6:18] conversation turned to equity and [6:20] something just shifted in him. This is a [6:22] Greg Brockman quote. He was angry. You [6:24] could sense it. At the end of the [6:25] meeting, he sat quietly and silently. He [6:27] said, "I decline the proposed even split [6:30] of equity structure and control." He [6:32] stood up, stormed around this table. "I [6:35] actually thought he was going to hit [6:36] me," says Brockman. And then Brockman [6:38] says Musk said, "When are you going to [6:41] be departing Open AI?" Brockman and [6:43] Ilya, they said they weren't going to [6:45] depart. And then Musk left. Wowee woo, [6:48] writes Mike Isaac. Uh a cutting Brockman [6:52] testimony regarding Musk. Look, he knows [6:53] rockets, he knows electric cars. He did [6:56] not and does not know AI and Ilya and I [6:58] did not believe he would spend the time [7:00] to get good at it. A lot of Brockman's [7:02] testimony is underscoring how he feels [7:04] Musk is not able to properly assess the [7:06] capabilities of AI. He recounts one [7:08] instance of him talking, this is Elon [7:10] talking to a researcher who is doing an [7:12] AI demo in which Musk berated the guy so [7:15] intensely that the dude almost quit the [7:17] field of AI. That is an aggressive [7:20] response. I mean, this is almost almost, [7:22] but uh quitting in protest is one thing. [7:25] I didn't like the way my boss was [7:26] talking to me. Little bit different to [7:27] quit your entire industry. Very very [7:30] aggressive, very high stakes, yeah. And [7:32] then uh Mike Isaac, of course, is [7:34] chiming in with his progress on his [7:35] lunch. He's already consumed 75% of it [7:37] and it's only 10:30 and he's already [7:39] sleepy. There's a lot more to cover [7:41] here, but uh it's an interesting back [7:42] and forth. Go check it out. Go follow my [7:45] guy Isaac because he has the the [7:46] play-by-play. Let's move on to the other [7:49] AI news. Google, Microsoft, and Elon [7:52] Musk's xAI have reached agreements with [7:55] the Trump administration to share early [7:57] versions of their new models with CAISI, [8:02] the center [8:03] >> calling it Casey. [8:03] >> Casey, okay. Uh the Center for AI [8:06] Standards and Innovation run by the [8:09] Department of Commerce. They will be [8:11] evaluated before releasing to the [8:12] public. So, what's interesting here? [8:15] It's in the journal. This felt like a [8:16] bombshell moment. Oh, wow, like the [8:18] government's in charge of AI. Then I [8:20] figured out that OpenAI and Anthropic [8:22] signed on to this exact deal 2 years ago [8:25] in 2024. That was news to me. Have they [8:28] has have the has the Commerce Department [8:30] been [8:31] reviewing [8:32] 4.5 and and OpenAI 5.1, 5.2? Has that [8:36] been happening and we just aren't aware [8:38] of it? And then like, you know, like the [8:40] whole Mythos rollout, like that would be [8:42] a very different tone if it was like, [8:45] "Oh, yeah, well, the Commerce Department [8:46] already evaluated Mythos because they [8:48] have this deal that's happened for 2 [8:49] years and since 2024." So, like, [8:52] whatever's going on with this Commerce [8:54] Department Center for AI Standards and [8:55] Innovation, they say the center has [8:57] completed more than 40 evaluations [8:59] including on models that remain [9:01] unreleased, but it feels like the Casey, [9:03] or as you're calling it, [9:05] doesn't produce reports, at least not [9:07] reports that go viral. Maybe they need [9:09] to be clipping their reports or [9:10] something because I haven't heard [9:11] anything from them saying, "Oh, wow." [9:13] Because you know, you imagine like the [9:15] AI hype machine. [9:16] >> they just have a report, it's a piece of [9:17] paper, and they just it's just it's just [9:20] [laughter] lists off, and it has a few [9:22] boxes you can check. And just one is [9:24] it's a good model, sir. The other one is [9:26] is chatbot or ASI. [laughter] [9:29] And so to date everyone just been [9:30] checking chatbot. [9:32] >> also it also has how many Rs did it say [9:35] were in strawberry? [9:37] And uh and uh I work at the Commerce [9:40] Department. I am a I am a very tiny man. [9:43] When I When my son was born, the doctor [9:45] handed me to him. [9:47] Did he answer correctly? Yeah, I don't [9:49] know. But because because [9:50] >> And you And And they they ask each [9:52] model, "Are you conscious?" No, say, "I [9:55] am conscious." [9:56] >> Yeah. [9:57] >> And then the And the model says that [9:58] they go, "Whoa." And check the whoa box. [10:01] >> Very very [10:02] uh crazy thing that's been happening. I [10:04] mean, I would just assume that the AI [10:06] hype machine, the back and forth, us, [10:08] lot of posters, would be like waiting [10:11] with baited breath for the latest report [10:13] from the Commerce Department. Because if [10:14] it's a preliminary evaluation that [10:16] happens before the model gets released, [10:18] you know, every time a model goes on [10:21] open router or goes on uh [10:24] uh LLM Arena, all of a sudden any little [10:27] Oh, there's a new image model like the [10:28] like the ChatGPT images, too. Like that [10:30] leaked because they were benchmarking it [10:32] and people were like, "Oh, there's a new [10:33] model coming." Um like rumors of new [10:36] models happen all the time, but somehow [10:37] like the Commerce Department is just [10:38] keeping it to themselves. They're not [10:40] pumping anything. The quote here from [10:42] Casey director Chris Ball [10:45] says, "Independent rigorous measurement [10:47] science is essential to understanding [10:49] frontier AI and its national security [10:51] implications. These expanded industry [10:53] collaborations help us scale our work in [10:55] the public interest at a critical [10:57] moment." Andrew Curran has some more [10:59] context here. Uh he says, "To sum up, [11:02] Anthropic, OpenAI, Google, Microsoft, [11:05] and XAI all have new pre-release [11:07] screening agreements with Casey. We [11:10] don't know the details of the new rules [11:12] yet. I assume they will be announced [11:13] with the AI executive order and AI [11:14] policy memo, both of which we may get [11:16] today. This is May 5th." And so, my big [11:19] question was like, "What is going on [11:20] with Meta?" Like they have near frontier [11:23] models. They're taking AI really [11:25] seriously. Mark Zuckerberg will get into [11:27] this is investing 125 billion dollars in [11:30] CapEx. The shareholders are thinking, [11:32] "Oh, what's going to happen here?" And [11:33] he's like super intelligence. And you [11:35] think the government would be like, [11:36] "Well, you don't just get to like sit [11:38] over there while all five of the other [11:40] leading labs are here." [11:41] >> billion dollars to work. [11:42] >> Yeah. Like probably a bigger investment [11:45] than xAI right now in terms of compute [11:47] capacity and maybe in terms of research [11:50] capacity. [11:50] >> Future capacity. [11:51] >> Yeah. Yeah. Yeah. Yeah, future capacity. [11:53] Exactly. So, you would expect Also, it's [11:55] just one of these things where if five [11:57] of the top labs are jumping into a [11:59] particular initiative, there's so [12:01] there's so little risk to jumping in as [12:03] well. If Anthropic and xAI are both [12:05] doing it on the left and the right, like [12:06] there's clearly room for Meta to just [12:08] say, "Yeah, we're cool with that, too. [12:09] We're we're we're riding with that." Uh [12:11] not everyone's cool with it, though. [12:13] There is there is some major pushback [12:15] mostly from George Hotz over at Tiny [12:17] Corp. We can pull up some of this. Uh [12:19] let's let's go through Andrew Curran's [12:21] take uh on what will happen here. So, [12:23] the EAI will create a mixed group of [12:25] tech CEOs and administration officials [12:28] who will work out the rules for new [12:30] release regulations. I see a lot of [12:32] people calling this a win for Eliezer [12:33] Yudkowsky's side, but even if the new [12:36] rules are very strict, the stop pause [12:38] group don't actually get what they want. [12:39] This doesn't stop or even slow AI [12:41] advancement, according to Andrew. It [12:43] only slows the rate of public releases. [12:46] Capabilities will be advancing at full [12:47] speed. I don't know how true that is [12:49] because if you don't release the model [12:51] at all, like you can't monetize it and [12:53] then justify the next leg of capex. So, [12:56] I'm not 100% sure that this has no [12:57] effect on on uh the speed of AI [12:59] progress. But I take Andrew's point. The [13:01] labs will finish training the new models [13:02] and then submit them for government [13:03] approval. Given the government currently [13:05] doesn't even want Anthropic to release [13:07] Mythos, how long will the regulatory [13:09] process take for something twice as [13:11] powerful? Five times as powerful? 10 [13:13] times as powerful? All of the incentives [13:15] are for the government to slow down [13:16] releases. Let's say OpenAI finishes [13:17] training GPT-6. It's twice as capable as [13:20] anything available today. If the [13:21] government approves it and something bad [13:22] happens, they take the blame. And the [13:24] longer they hold it back, the the longer [13:27] the government agencies get exclusive [13:28] access. So, there's a little bit [13:30] principal agent problem here. Meanwhile, [13:31] opening I will start using GPT-6 to [13:33] train GPT-6.1. Again, unclear if you can [13:36] actually marshal the compute for the [13:38] next model without the revenue traction, [13:40] but his point holds. If they finish 6.1 [13:42] while 6 is stuck in approval, they'll [13:43] move on to train 6.2 with it. This [13:45] potentially creates strain situations [13:46] where the labs are many generations [13:48] ahead internally while the public is [13:50] still waiting for something they [13:51] submitted months ago. So, you could see [13:54] like you're like you're doing all this [13:56] weird value capture and the size of the [13:58] firm bloats and bloats because you can't [14:00] release it, so you just wind up rolling [14:02] out a product. [14:03] >> your own use of it. [14:04] >> Yeah, which is like maybe not the most [14:06] democratic solution, maybe not the most [14:09] you know positive outcome. [14:10] >> everyone needs to wait to have super [14:12] strong opinions until we see the EO that [14:15] goes along with this because yeah, it'll [14:17] be interesting. The bad scenario is like [14:20] somebody that let's say has a neo lab [14:22] wants to make, you know, release a model [14:24] and they've raised $100 million. Do they [14:27] have to go through the same process as, [14:30] you know, one of these larger labs or a [14:31] hyper scaler in [14:33] getting, you know, getting these sorts [14:35] of approvals? So, I don't think we can [14:37] really take any strong stance until we [14:39] understand what the what the intentions [14:41] of the EO are. [14:42] >> So, [14:43] the meta commentary on this it will Zach [14:46] says it will not it will seriously not [14:47] surprise me if they try to require [14:49] permits or licenses to use AI and [14:51] restrict local model downloads. You [14:53] really should be buying hardware. [14:55] There's an interesting company that's [14:56] launching basically a Tesla power wall [14:58] for black walls. You can just get GPUs [15:00] mounted to your house. Maybe that's the [15:02] future. Tiny box is another is another [15:05] solution. And the Tiny Corp has chimed [15:08] in says, "Ah, maybe they can enlist the [15:09] MPAA and RIAA calling back to the piracy [15:14] debates of 2005 or so to help with the [15:17] restricting of downloads. I can see the [15:19] lawsuit against the grandma who [15:20] downloaded Qwen already. This is a win [15:23] for China [15:25] in particular. Remember the time they [15:27] regulated crypto over 40 bits. No, no, [15:29] that's too many bits for the people. [15:31] They were losers then and they'll be [15:33] losers again. The difference this time [15:35] is that it will cost them cultural [15:37] influence to the Chinese and [15:40] and he says Jensen tried to warn them. [15:42] >> like play it out a little bit. China, [15:44] you know, these open source models are [15:46] on eight months behind right now. The [15:48] gap is widening. [15:50] >> Mhm. [15:50] >> Likely due in some part to export [15:53] controls. But if you enter a scenario [15:56] where US labs get kind of like hung up [15:58] and again have to keep these [15:59] capabilities [16:00] internally but these Chinese open source [16:03] models are just like shipping as they're [16:05] ready over and over and over and over [16:07] and then sort of like compounding on the [16:09] collective learnings. You can imagine [16:11] eight one scenario is helping them close [16:13] that gap at least close the gap between [16:15] what's publicly available from the [16:17] American labs and what's publicly [16:19] available via Chinese open source, even [16:22] if the labs actual capabilities are much [16:24] farther ahead, but they're just unable [16:26] to release that, right? [16:27] >> And to to be clear, I mean these are two [16:29] very different philosophies, but from my [16:31] perspective [16:33] George Hotz has been advocating for no [16:36] model gap and in fact the strongest [16:39] models possible being fully open [16:41] sourced, being fully available to [16:43] everyone. He wants no authoritarian [16:45] control. He he has taken the [16:47] anti-authoritarian stance on AI. [16:49] >> wants all the products to be free. [16:50] >> Uh yeah, maybe. I don't know. [16:52] >> I mean he he he he [16:55] he's [16:55] >> Not the not the the the physical things. [16:57] He wants he wants to build things. But [16:59] Tyler, how how have you been wrestling [17:01] with this question of an an FDA for the [17:04] AI? Potentially a DMV for the AI. [17:08] >> Yeah, I mean like obviously it really [17:09] depends on how it's implemented. [17:11] In in the super like safety peeled [17:13] scenario like yeah, it's probably not [17:14] good for for like innovation broadly. [17:16] But I think [17:17] so it's called like Casey, right? [17:19] Center for for AI's like standards and [17:21] innovation. It used to be called the AI [17:23] Safety Institute under Biden. [17:25] And so they changed it. So I think even [17:27] from that you can maybe take away [17:29] something where like maybe this actually [17:30] really not much to do with safety at [17:32] all. Maybe it's just like we we need to [17:33] normalize how we publish sweet bench [17:35] scores because some people when they [17:37] report it, it's like different you know [17:38] how you like actually run these [17:40] benchmarks or something. If it's like [17:41] standards, I think like that could be a [17:44] possibility. That's [17:45] >> Yeah. Seems like totally fine. [17:46] >> Yeah. And then the other the other risk [17:47] is I mean there's regulatory capture [17:49] potentially be very hard for startups to [17:51] get approved if it's like oh well, you [17:53] know, we we've heard the story of [17:54] Anduril where, you know, it's like to [17:56] get the company off the ground hire 50 [17:58] lobbyists immediately, right? And like [18:00] you can imagine a situation where, you [18:02] know, there's some uncertainty about [18:03] what's happening with SSI and Eliezer's [18:05] project, but I think I like the idea of [18:07] an individual brilliant researcher going [18:09] off and researching an interesting path. [18:11] If all of a sudden it's like oh well, if [18:14] you want to do a neo lab, set up an [18:15] office in DC, get your licenses, then [18:19] you can start training, then you can [18:20] make sure that you're approved, and then [18:22] you have the right deals in place, and [18:25] you could wind up with a lot of not just [18:26] regulatory capture, but also regulatory [18:28] favoritism where whoever is in good with [18:31] this particular administration gets [18:32] their models approved faster, and you [18:35] you wind up with a lot of risky [18:37] outcomes. In other news, [18:39] Brian Armstrong sent an email to all [18:41] employees at Coinbase announcing [18:42] announcing layoffs 14% of the workforce [18:46] at Coinbase. [18:48] Two forces are converging at the same [18:50] time, he says, and Coinbase needs to be [18:53] front-footed to respond to both. First, [18:56] the market. Coinbase is [18:57] well-capitalized, has diversified [18:58] revenue streams, well-positioned to [19:00] weather any storm, but crypto is also on [19:03] the verge of the next wave of adoption [19:04] with stablecoins, prediction markets, [19:06] tokenization. However, the business is [19:07] still volatile quarter to quarter. [19:09] There's a crypto cycle. We've sort of [19:11] been touching on this. Like there was a [19:12] little bit of a stablecoin boom. It [19:14] feels like it's a little bit flat. A [19:15] Bitcoin price has been a little bit flat [19:17] recently. And so, he says, "We're [19:19] currently in a down market, and we need [19:21] to adjust our cost structure." And then [19:23] he says, "Second, AI is changing how we [19:25] work." And so, a lot of people, they [19:27] jump to AI is taking all the jobs, and [19:30] that's not exactly what's happening. I [19:32] think that [19:33] there is a right-sizing going on at [19:35] Coinbase, and AI is hoping to be an [19:38] enabler of more efficient with a leaner [19:40] headcount. The big question that I had, [19:44] that we were debating, was, you know, is [19:46] this Like, how does this reflect on [19:48] Brian Armstrong as a CEO? As a CEO, is [19:52] is the goal to always have the perfect [19:54] amount of of headcount and and never [19:56] have to do any layoffs, or is it [19:59] actually the the role of a very [20:02] aggressive and mature CEO to staff up [20:05] aggressively during boom times, get the [20:08] best people, and then when there's a [20:10] down market, sort of reorganize and keep [20:13] just the best on the team as you move [20:15] forward with a leaner organization. And [20:18] what does that mean for whether you want [20:19] to work in an industry that has market [20:21] cycles like that? You might want to go [20:23] work at Costco, or a tobacco company, [20:25] cuz those don't go through market cycles [20:27] really at all. But, if you if you're in [20:29] crypto, you're probably not new to the [20:31] idea of market cycles, and you probably [20:34] >> on a roller coaster. [20:35] >> You're born on the roller coaster, [20:36] you've been on the roller coaster, and [20:38] you might be moving over to a new roller [20:40] coaster if you're one of the unfortunate [20:41] people that will be laid off over the [20:43] next couple months. [20:44] >> I would say the overall the the positive [20:47] thing from the response that I've seen [20:49] so far is a lot of people are not like [20:52] dooming around this. They're just saying [20:54] like, "Hey, look, this is something that [20:56] Coinbase has done before, cyclical [20:59] business, [21:01] and no one is saying it's over for [21:03] white-collar work." [21:04] >> Yep. So, Andrew Yang has a little bit of [21:07] a take here. One takeaway, no pure [21:09] managers. Every leader has to be working [21:11] as an individual contributor. The pure [21:13] people manager role, one the one that [21:15] built most corporate career ladders over [21:17] the fifth last 50 years, no longer [21:19] exists at Coinbase. Leaders will have 15 [21:21] plus direct reports. That is insane. [21:24] Previously, managers capped out at six [21:25] direct reports. I felt like 10 was sort [21:27] of the platonic ideal there. Uh but [21:30] Andrew says that would be impossible [21:31] without AI. And Coinbase is testing [21:33] one-person teams. Single person is the [21:35] engineer, the designer, and the PM. A [21:37] pod of one with agents or org structures [21:40] are being redesigned in front of our [21:41] eyes. Derek Thompson has the other side. [21:43] He says Coinbase is the latest tech [21:45] company including Block, Salesforce to [21:46] announce big layoffs and site AI coding [21:48] productivity as a driver. But uh Derek [21:51] Thompson is looking at the stock prices [21:53] of these companies. Salesforce down 31%, [21:55] Coinbase down 23%. In the past 5 years, [21:58] it's sort of a roller coaster. I don't I [21:59] don't know if this is a great read [22:00] because it's sort of flat. I don't know. [22:02] Anyway, uh Block down a lot. Uh some [22:05] right sizing makes sense, but he's [22:07] calling it AI washing of layoffs. They [22:09] were coming anyway, and you want to spin [22:11] them in a positive way. Uh I don't know. [22:13] TBD. TBD. [22:14] >> Yeah, I don't think we have any clarity [22:16] on where like uh what what kind of [22:19] >> Yeah. [22:19] >> which teams had they had greater [22:21] percentages of cuts, right? Because [22:23] again, there was some reporting. I don't [22:25] know how true it ended up being that [22:27] like when Square did their big cuts, [22:29] like engineering was getting laid off. [22:31] Engineering [22:32] >> Yeah, I don't know if that was ever [22:33] confirmed. [22:33] >> Yeah, I don't know if it was confirmed, [22:35] but that does imply that it's like [22:36] actually more grounded in in the in AI [22:39] productivity. But again, a lot of teams [22:41] that we're seeing their engineers are a [22:43] lot more effective. They can do a lot [22:45] more. And so maybe you want more [22:46] engineers than you did before. So. [22:48] >> Well, uh lots of strengths in the [22:50] business, lots of liquidity, very hard. [22:52] You got You could vibe code Coinbase, [22:54] but how do you get all the customers and [22:55] all the liquidity and all the features [22:57] and all the regulatory and all the other [22:58] stuff? The million reasons to be excited [23:00] about Brian Armstrong's next chapter at [23:02] Coinbase. Well, uh and Ryan Peterson [23:04] chimed in to sort of push back on Derek [23:06] Thompson. He says, "In the layoff [23:07] announcement, the CEO literally said [23:09] that the first reason they were doing [23:10] layoffs is because crypto is in a bear [23:12] market and they need to adjust their [23:14] cost structure. AI productivity was [23:15] secondary." And so, I think that's a [23:17] fair take. Let's go over to Meta. Some [23:19] debate, people are digesting Meta's [23:21] earnings. The stock sold off a bunch. [23:23] Where is Meta today? Uh it's a bargain [23:26] stock. It's a $1.5 trillion company and [23:30] over the past 5 days down 10%. It's [23:34] pretty flat over a year and over the [23:36] past 5 years it's nearly doubled, so not [23:38] doing too bad. But, um the debate [23:41] is over CapEx. Meta's core ad business [23:44] is ripping up 33% [23:46] in Q1 year-over-year. Strong ad [23:48] impressions, strong margins, but the [23:50] market is worried that the beautiful [23:51] cash machine [23:53] uh is turning into an AI CapEx furnace. [23:56] $125 to $145 billion in CapEx. It's a [24:00] lot of CapEx, especially for a company [24:02] without a cloud business [24:04] uh that can resell capacity if they wind [24:06] up with extra capacity. If you can't use [24:08] it all, uh what do you do with it? Meta [24:10] doesn't really have an answer to that [24:11] necessarily, and so that's what has the [24:14] market worried. And even XAI is having [24:17] trouble, at least reportedly, driving [24:20] demand for compute that they have. There [24:21] was that 11% number. It seemed like that [24:24] was a bit of an overstatement, but [24:26] there's something to be said for when [24:28] you have a near frontier but slightly [24:31] lagging model, you need to find a [24:34] solution to actually have off-take. And [24:37] you might just not be able to spin up a [24:39] cloud platform on day one. Elon took a [24:42] different route, you know, partnering [24:44] with Cursor that has a ton of demand and [24:46] is a front door to AI coding. And so, uh [24:48] that makes, you know, a reasonable [24:51] amount of sense for the compute. Who's [24:53] who's GPU rich, who's GPU poor, let's [24:55] match them up. Meta doesn't have an [24:56] obvious dance partner like Google, like [24:59] Azure, like AWS, right? And so the stock [25:01] looks cheap now, but investors are [25:02] unclear about the payoff around AI [25:04] spend, and there's also some legal legal [25:06] and regulatory risk. [25:07] >> People, at the end of the day, are [25:09] worried that he's going to repeat the [25:10] metaverse saga. And uh he's going to [25:13] spend a lot of money, not get very far. [25:16] >> Yeah. [25:16] >> Um [25:17] I think that [25:19] uh about coming up on, what is it, [25:21] almost 10 10-ish months ago, he started [25:24] talking about personal [25:26] superintelligence. It sounds cool. What [25:28] does that mean? Does it mean AI in your [25:31] Meta Ray-Bans, which are selling a lot [25:33] of units, so that's a bright point. Does [25:35] it mean creating a a uh consumer LLM [25:39] like they have with Meta AI now? Meta AI [25:41] now is squarely in the top five [25:43] >> Yeah. [25:43] >> of the App Store rankings, but it is the [25:47] most competitive category potentially [25:50] since search, right? And so I don't [25:52] think that even though they have a solid [25:53] model and they're able to get into that [25:56] into that top five, I don't think at [25:58] least um investors are not ready to [26:01] price in Meta AI becoming like a key [26:04] player in consumer LLMs just yet, right? [26:06] So it's all the spending and no revenue [26:10] acceleration associated with that [26:11] spending. [26:12] >> Yeah. [26:12] >> All that being said, Meta has used AI to [26:16] drive business value historically more [26:18] effective than almost any other business [26:20] in the world, right? It's It is really [26:22] really insane to think that there is [26:25] some scenario where the labs end up [26:29] going public at evaluation somewhere in [26:32] the range of Meta while Meta is sitting [26:35] there with 200 billion of annualized [26:37] revenue growing 33% a year, [26:40] uh one of the greatest businesses in all [26:43] of human history, and not really getting [26:45] full credit for for that or their [26:48] various AI initiatives. Obviously, the [26:51] the drop in uh people uh which I think [26:54] they What do they call it? They they [26:56] >> Daily active people. [26:57] >> Daily active people. [26:59] >> They're They're not users. They're [27:00] people. That That's playing playing into [27:02] this as well. [27:03] >> In Amazon news, [27:05] uh the number of monthly releases of [27:08] ebooks on Amazon is going vertical. We [27:10] are in the fast takeoff of AI slop [27:13] books, apparently. That is the [27:15] conclusion. John Arnold just says, "Ha [27:16] ha ha ha ha." Uh and it does look like [27:19] uh the the the the rate of publishing on [27:22] Amazon is going [27:24] >> Explaining Explaining this this to [27:26] somebody in 2022. There's going to be a [27:28] renaissance 4 years from now. There's [27:31] going to be 350,000 [27:33] books being published every single [27:35] month on Amazon. [27:37] And they're like uh [27:38] You're not going to be happy with why, [27:40] though. [27:41] >> [laughter] [27:41] >> Yeah. It It is uh It is crazy. I mean, [27:44] if you if you had asked me to guess how [27:47] many books were being released every [27:48] month on Amazon in the pre-AI era, I [27:51] don't know if I would have put it at [27:52] 100,000. That's a lot of books. I feel [27:54] like we talk to a lot of authors. I see [27:57] the hyped books. I see the New York [27:59] Times bestseller list. I don't really [28:01] see the 999 [28:04] or 99,900 books that don't make the cut [28:07] in the given month. I maybe hear about [28:09] five or 10 new books every month, but uh [28:11] lo and behold, they were in fact [28:13] churning out 100,000 books. Now, it's [28:14] over 300,000 books every month on [28:16] Amazon. Bullish for Amazon if you like [28:18] slop. It's good for you, too. I don't [28:19] know. Maybe there's some good stuff on [28:20] there. It all depends. The models are [28:22] getting better. They're getting less [28:23] sloppy, and I think people will wind up [28:25] liking them. There's some AI videos that [28:27] I like out there. Uh you know you know [28:29] there's going to be one book that's [28:30] fully AI generated that is of high [28:33] quality. [28:33] >> Another buyout news, Long Lake, uh very [28:36] under the radar company, but has been on [28:38] an absolute tear over the last [28:42] >> This is a wild wild deal. Wild deal. [28:44] >> uh is taking Amex GBT Global Business [28:47] Travel Private. I could see Long Lake [28:49] being kind of interested in eBay, too. [28:52] >> Oh. [28:52] >> They They maybe got their hands full [28:53] with Amex GBT, but uh incredibly [28:56] talented uh team. Uh been uh really [29:00] unlocking the power. [29:01] >> So, it's a $6 billion take private. It's [29:03] also interesting because General [29:04] Catalyst, Mark Vergara, uh is uh is is [29:07] sort of talking about the story of Long [29:09] Lake. He says, "We had the first Long [29:11] Lake team retreat at my home in Miami [29:13] just 2 and 1/2 years ago when the [29:14] company was getting started. It's been [29:15] incredible to watch uh the talent and [29:17] team have brought together to transform [29:19] the services industry into one of growth [29:21] and abundance with AI. Excited to for [29:23] yet another milestone. General Catalyst [29:25] is honored to be the partner from [29:26] beginning and excited to continue the [29:28] strategy of backing what we believe are [29:30] the best AI investments." Um and [29:32] interesting because General Catalyst is [29:34] also a an investor in Ramp, I believe. [29:36] And so, there's a question about like [29:38] Ramp has a travel product. How different [29:40] is Amex Business Global Business Travel [29:43] from Ramp Travel? Obviously, there's [29:44] some overlap there, but there might be [29:46] some synergies since there's board [29:47] members in common now. Who knows where [29:48] this goes, but at 6.3 billion, that's a [29:51] small slice compared to Ramp, uh which [29:53] has been on an absolute tear. [29:54] >> Another big raise before we close out [29:56] for the day, 11 Labs. [29:58] Cross five, half a billion dollars. [30:03] Welcome new investors, BlackRock, [30:05] Wellington, and Nvidia. [30:06] Santander, Jamie Fox. [30:08] >> Isn't this a raise or is this just a [30:09] flex? They're just flexing the AR. [30:11] >> some new investors. I don't know. I [30:12] don't think they [30:13] I don't know if they announced their new [30:15] valuation. [30:16] >> Well, [30:16] >> Uh but I bet [30:17] >> Mitchell Green would be very happy, you [30:19] know. If you have ARR, that's what you [30:21] report. If you don't, you report the [30:23] valuation. [30:24] Uh and they are making some serious [30:26] progress in the commercial side. [30:27] >> Yeah, it's some serious institutions [30:28] coming in. [30:29] >> To the team over there, [30:30] thank you for tuning in. We will see you [30:32] on Friday. Have a good week. [30:35] >> Goodbye. Love you. [30:36] >> Leave us five stars on Apple Podcast and [30:37] Spotify. Sign up for our newsletter at [30:38] tvf.com. Goodbye. [30:46] >> Woo!