Gurley, Podcasters Warning US Open-Source AI Crackdown Risks Chinese Lead

格利与播客主播警告：美打压开源AI恐致中国领先 华盛顿拟议对开源AI模型施加的限制恐将孤立美国市场，并促使全球开发者转向中国替代方案；与此同时，科技高管在预期IPO前纷纷收回有关就业末日的论调。 节目主持人与嘉宾比尔·格利（Bill Gurley）深入剖析了不断演变的科技格局，重点探讨了针对开源AI的潜在监管打压以及各大科技巨头不断变化的经济状况。他们就企业裁员究竟代表了真正的AI自动化，还是高管用来掩饰过度招聘和运营修正的“AI粉饰”展开了辩论。讨论还强调了软件开发中的杰文斯悖论，审视了全美低失业率与代码提交量激增以及蓝领基础设施需求并存的现状。 华盛顿的监管讨论正在为潜在禁止开源或开放权重 AI 模型铺路，而 Anthropic 等公司针对生物及网络安全风险的频繁预警被指是在进行监管俘获并为此建立前置理由。 (49:41) xAI 将 220,000 块 GPU 上的模型训练技术栈完全用纯 C 语言重写，相比 JAX 等传统框架实现了一个数量级的性能提升，大幅降低了算力与基础设施成本。 (51:56) Bill Gurley 警告称，如果美国出台法规压制本土开源 AI 发展，全球大部分地区将转向依赖中国模型；而欧洲严苛的合规门槛已成为去中心化开源社区的前车之鉴。 (51:56) 据报道，Anthropic 的营收增长速度显著超越 OpenAI，其复利增长优势可能推动 AI 市场迅速向寡头垄断集中，尽管电力与算力短缺仍是潜在瓶颈。 (55:18) 因未设立员工使用上限，某企业单月意外产生高达 5 亿美元的 token 账单，促使企业 CFO 在对 AI 从“取代一切”到“泡沫破裂”的极端舆论摇摆中加速收紧预算管控。 (56:44)

The All-In Podcast hosts argue that apocalyptic AI narratives and regulatory proposals serve as corporate cover for regulatory capture and post-pandemic tech layoffs, while economic data reveals model commoditization and expanding labor demand. The panel contends that open-source architectures and high-agency tool adoption—not centralized state oversight—provide the primary safeguards for economic and technological resilience. Regulatory Capture and Open-Source Ban Risks (33:23) Reacting to Pope Leo XIV’s encyclical Magnifica Humanitas—which urges AI regulation, worker retraining, and autonomous weapon bans—David Sacks and Chamath Palihapitiya argue that top-down oversight risks creating an elite regulatory oligopoly. Sacks invokes Quis custodiet ipsos custodes?, warning that an "FDA for AI" would inevitably cause regulatory mission creep. Bill Gurley contrasts the Vatican's stance with Pope Leo XIII's 1891 warnings against industrialization, noting that subsequent free-market innovation drastically improved global living standards, poverty rates, and real wages. The hosts interpret safety warnings from frontier labs like Anthropic as strategic plays for regulatory capture rather than pure altruism. Gurley and Sacks point to CEO Dario Amodei’s Machines of Loving Grace to argue that Anthropic's founders envision themselves midwifing a cybernetic deity to oversee human resource distribution. Sacks warns that corporate emphasis on bio-weapons and cyber risks aims to lay the groundwork for banning open-weight models, a move Chamath and Gurley argue would cripple US cloud infrastructure while forcing global developers onto Chinese open models. Model Convergence, Architecture Shifts, and Enterprise Costs (41:21) The panel observes that frontier model capabilities are commoditizing while infrastructure economics shift. Jason Calacanis cites Rogo's financial benchmarks showing Claude 3.7 Opus (94.2%), OpenAI GPT-4.5 (94.1%), and Claude 3.5 Sonnet (93.9%) performing within fractions of a percent of one another. To prevent vendor lock-in, Gurley advocates open integration standards like the Model Context Protocol (MCP), comparing it to Google open-sourcing Kubernetes to make cloud compute portable across AWS. Simultaneously, Chamath highlights xAI rewriting its training stack in C across 220,000 GPUs, achieving an order-of-magnitude speedup over frameworks like JAX. He argues that continuous low-level software gains will undercut $10 billion training clusters in favor of vastly cheaper open alternatives. Meanwhile, enterprises are facing a financial reckoning over runaway token consumption—exemplified by an anecdote of a $500 million monthly enterprise bill and a Fortune 20 firm spending $200 million with negligible ROI—driving demand toward on-premise hardware appliances and strict cost controls. Labor Market Realities Versus Corporate AI Washing (59:56) Addressing recent narrative pivots by Dario Amodei, Sam Altman, and Goldman Sachs CEO David Solomon, the hosts note that leaders are tempering apocalyptic job-loss rhetoric ahead of public offerings. Sacks highlights macroeconomic stability: US unemployment sits at 4.3%, a Yale Budget Lab study found no measurable disruption over three years, and software engineering postings have climbed 15% year-over-year. Citing a monthly surge to 1.1 billion GitHub commits, Sacks points to Jevons Paradox, arguing that lower code costs increase overall software complexity and the demand for engineers to manage it. Chamath and Sacks accuse major tech firms of "AI washing"—blaming automated efficiencies for layoffs that actually correct zero-interest-rate overhiring and slowing corporate growth. Calacanis pushes back, contending that AI is actively eliminating middle management, project managers, and designers by turning engineers into cross-functional teams. While Calacanis predicts structural pain for professional drivers and warehouse staff, Gurley expects human drivers to retain around 50% market share as reduced costs expand total transit demand, noting that high-paying skilled trades continue to face labor deficits. High Agency and Voice-Driven Productivity Workflows (6:01) To counter workplace obsolescence, Gurley emphasizes "high agency," contrasting proactive AI adoption against Gallup data indicating that 59% of employees are "quiet quitters." Sacks asserts that advanced proficiency with models like Claude represents the single most marketable skill for new graduates, comparing early mastery to the spreadsheet revolution in the 1980s. The hosts outline pragmatic strategies for workplace integration. Describing how podcast producer Nick built an automated briefing system by prompting Claude to write its own operating rules from past transcripts, Sacks emphasizes that high-context prompt design requires rigorous human-in-the-loop supervision. Calacanis and the panel recommend pairing voice-to-text tools like WhisperFlow with large language models, enabling workers to dictate raw, unstructured thoughts into structured operational plans and strategic documents without complex manual prompting. The unresolved tension centers on whether corporate lobbying and regulatory alarmism will succeed in suppressing open-weight models, or whether open architecture, rapid commoditization, and enterprise cost constraints will decentralize AI before monopolistic capture takes hold.

本期节目探讨了前沿人工智能领域的制度竞争、商业现实与劳动力市场演变；嘉宾与主持人们认为，头部大模型实验室借助安全担忧游说监管存在行业垄断风险，而宏观经济与就业数据正逐步证伪所谓“AI导致广泛失业”的末日叙事。 监管捕获逻辑与开源生态的制度博弈 (33:23) 围绕梵蒂冈教皇利奥十四世发布的《Magnifica Humanitas》通谕对AI集中化权力和自主武器的警示，David Sacks主张以分权制衡和反垄断竞争防范风险，反对设立类似FDA的政府审批机构。Chamath Palihapitiya与Sacks指出，头部实验室如Anthropic将自身塑造成安全标杆并夸大灾难性风险，其深层博弈策略在于抬高行业准入门槛，借助技术信息差主导监管机构，将市场竞争压缩在极少数闭门实体中以构建资本护城河。 针对华盛顿出现的限制甚至封杀开源模型的动向，Sacks与Bill Gurley警告，开源权重本质上是可运行于本地硬件的数字文件；强行推行云端开源禁令只会削弱美国本土开发者，反而促使全球生态转向不受此类限制的中国开源模型。Gurley援引P3研究所论文指出，在欧洲监管增加去中心化合规成本的背景下，保持去中心化的开源模型与本地运行能力，是用户掌握智能主权、对抗垄断审查与社会信用体系式中心化控制的核心防线。 基础模型收敛与企业算力支出反思 (41:21) 在模型能力与商业化层面，金融科技机构Rogo的评测显示，Claude 3.7 Opus、GPT-4.5和Claude 3.5 Sonnet在复杂金融任务上的得分均在93.9%至94.2%之间，差距小于0.3%，表明顶尖前沿模型正在走向能力收敛与商品化。针对昂贵的基础训练，Chamath以xAI在22万块GPU集群上用C语言重写底层代码、实现超越JAX数量级加速为例，指出1%的效率提升即可节省相当于2000块GPU（价值数亿美元）的开支，算法与底层工程优化正在大幅压低算力壁垒。 与底层模型降本并行的是企业端对ROI的严格反思。Chamath提到有《财富》20强企业斥资2亿美元购买Token却成果寥寥，亦有企业因未设用量上限在单月耗费5亿美元Token费用的极端个案。为此，企业买方正转向本地化部署硬件与热插拔控制平面，Bill Gurley则倡导通过Linux基金会支持的MCP等开源连接器实现模型模块化替换，防止企业被单一云模型提供商深度锁定。 劳动力末日叙事逆转与企业AI粉饰争议 (59:56) 在就业议题上，讨论指出包括Anthropic首席执行官Dario Amodei与OpenAI首席执行官Sam Altman在内的科技高层近期开始淡化AI消灭就业的激进言论，高盛首席执行官David Solomon亦在《纽约时报》撰文指出AI主要替代任务而非岗位。David Sacks引用耶鲁预算实验室研究指出，过去三年AI未对美国就业市场造成系统性破坏，全美失业率维持在4.3%左右的充分就业区间；而在AI渗透最深的软件开发领域，工程类岗位招聘反而同比增加15%，GitHub月度代码提交量更从一年前的全年10亿次激增至单月11亿次，验证了因边际成本降低而带动总需求扩张的杰文斯悖论。 Chamath与Sacks进一步抨击了企业界的“AI粉饰”（AI washing）现象，指出众多上市公司在零利率时代为封锁竞争对手而过度囤积冗余人力，如今随着收入增速放缓，管理层借AI自动化之名掩盖经营失误与常规降本裁员，但在正式SEC财务披露中并无证据表明LLM应用带来了显著的利润提振，部分证券诉讼律师已开始警告此类夸大可能面临法律风险。 个人能动性重塑与蓝领基础设施外溢 (6:01) 从微观工作方式看，Bill Gurley与David Sacks认为在职场避免淘汰的核心是成为具备高能动性（high agency）的AI工具使用者，熟练运用Claude等模型的能力就像20世纪80年代率先掌握电子表格一样具备套利价值。日常工作中借助语音转文本工具录入非结构化思考交由模型整理，以及构建自动化系统提示词，正在推动跨职能复合工种的涌现，导致中间协调层、项目管理等管理类岗位面临压缩。 与此同时，AI的发展重心正在向实体基础设施扩散。数据中心扩建、电网升级和发电容量建设引发了数十万建筑与电气等蓝领就业需求。针对Jason Calacanis提出的无人驾驶和仓储机器人将导致数百万司机与底层工人失业的担忧，Gurley提出反驳，认为受制于经济性，全面自动化很难彻底实现，正如ATM机普及并未消灭银行柜员，自动驾驶车队扩张的同时人类专业司机仍有望保持约50%的市场份额。 本期节目表明，人工智能从早期的概念崇拜与失业恐慌步入现实检验阶段：前沿实验室面临商业化收敛与巨额算力投入的ROI考核，企业正从无节制的Token支出转向严控预算，而所谓白领就业崩溃论正被宏观数据与技术扩散引发的工程复杂性需求所修正。然而，关于自动化对交通运输等实体行业长远冲击的真实烈度，以及西方对开源AI的潜在监管是否会造成地缘技术劣势，仍是未形成共识的持续博弈点。

## Background 本期视频由 YouTube 科技与创投知名播客《All-In Podcast》于 2026 年 6 月 27 日发布。节目特邀传奇投资人 Bill Gurley 代班主持，简介指出本期核心议题涵盖教皇利奥十四世针对 AI 发布的重磅通谕、Anthropic 团队构建“数字神明”背后的哲学与监管俘获争议、开源与模型主权博弈，以及硅谷和华尔街在“AI 导致就业浩劫”叙事上的集体反转。 ## TL;DR 本期节目深入探讨了当前前沿 AI 技术与社会伦理的激烈碰撞。主持团队首先剖析了教皇发布的 AI 通谕，指出过度呼吁政府监管易陷入奥威尔式集权，而 Anthropic 等顶尖实验室则深陷“弗兰肯斯坦博士”式的数字造神狂热与监管俘获议程。随后，嘉宾们探讨了模型能力大宗商品化趋势下开源与本地化部署的主权价值，并就 AI 对劳动力市场的真实冲击展开激烈辩论：萨克斯与格利列举 GitHub 代码激增和宏观就业数据，力证焦蓬斯悖论与大厂“AI 洗白”借口；卡拉卡尼斯则直陈底层劳动力与白领中层被技术替代的结构性阵痛。最终众人指出，面对即将到来的 IPO 压力与社会情绪反弹，科技巨头正在悄然收回激进的末日言论。 ## Atmosphere & style 整场讨论兼具硅谷内幕的辛辣爆料、深度的政治哲学思辨与激烈的火药味交锋。主持人们在开场的豪宅寄宿插科打诨中迅速切换到高密度的硬核议题，萨克斯与卡拉卡尼斯在就业替代与个案统计上互不相让、针锋相对，频频以“托洛茨基预言”和“精英主义”互相讥讽；而比尔·格利则以德州老兵式的沉稳幽默和历史长周期视角为全场定锚，整体氛围跌宕起伏、真实鲜活且充满张力。 ## Bulletin - [17:33](https://www.youtube.com/watch?v=4oq91rzQcO8&t=1053s) 教皇利奥十四世发布长达 235 页的首份人工智能通谕《壮丽的人性》，警告技术具有非中立性并带有资本与控制者的特质，呼吁严格监管并禁止自主武器系统；David Sacks 则反驳称最大的奥威尔式监控风险来自政府而非企业。 - [26:54](https://www.youtube.com/watch?v=4oq91rzQcO8&t=1614s) Bill Gurley 提出 Anthropic 的“弗兰肯斯坦博士理论”，认为其核心团队深信自己正在创造超越人类的数字神明，并试图通过宣扬末日恐慌来达成监管俘获，封死开源与后发竞争者的生存空间。 - [41:21](https://www.youtube.com/watch?v=4oq91rzQcO8&t=2481s) 前沿模型正在走向大宗商品化，Rogo 评测显示 Claude Opus 4.7、GPT-5.5 与 Claude Sonnet 4.6 性能差距已不足 0.3%，同时企业端因巨额 Token 账单回报薄弱而开始紧急收紧采购预算。 - [59:56](https://www.youtube.com/watch?v=4oq91rzQcO8&t=3596s) AI“就业末日论”叙事出现 180 度大转向，高盛 CEO 发文驳斥岗位毁灭论，GitHub 代码生成量激增 14 倍反而推高了软件工程师的招聘需求，大厂裁员更多被归咎于掩盖过去过度扩张的“AI 洗白”。 ## Key takeaways - [05:59](https://www.youtube.com/watch?v=4oq91rzQcO8&t=359s) 在 AI 时代保护自身职业竞争力的最佳路径是拥有高行动力，借助 AI 工具把自己塑造成超级个体，而不是陷入消极被动的“在职隐退”。 - [09:46](https://www.youtube.com/watch?v=4oq91rzQcO8&t=586s) 终身学习源于内心的真实痴迷而非苦行僧式的自律，熟练掌握 Claude 等大模型及构建结构化提示词的能力是当前职场最具壁垒的套利技能。 - [17:33](https://www.youtube.com/watch?v=4oq91rzQcO8&t=1053s) 教皇利奥十四世发布通谕《壮丽的人性》，严厉警示 AI 带来的权力集中风险，并主张设立严格的政府监管与防护防线。 - [22:47](https://www.youtube.com/watch?v=4oq91rzQcO8&t=1367s) 防范 AI 极权的最佳路径是三权分立与健康的自由市场竞争，而非赋予政府设立“AI 版 FDA”的审批审查特权。 - [24:51](https://www.youtube.com/watch?v=4oq91rzQcO8&t=1491s) Bill Gurley 援引 1891 年至今工时减半、实际工资暴增等数据，有力反驳了教会历史中对技术革新与工业革命的悲观误判。 - [29:54](https://www.youtube.com/watch?v=4oq91rzQcO8&t=1794s) Anthropic 的反常唱衰行为背后存在两大推论：一是利用公众恐惧构筑合规壁垒的监管俘获，二是自认在接生优于人类之新物种的“弗兰肯斯坦狂热”。 - [38:32](https://www.youtube.com/watch?v=4oq91rzQcO8&t=2312s) AI 主权正在从“数据隔离”升级为“智能主权”，企业与个人亟需在本地私有硬件上部署开源模型，以防止思想解释权被中心化巨头垄断。 - [42:54](https://www.youtube.com/watch?v=4oq91rzQcO8&t=2574s) 前沿模型性能趋同至 0.3% 的微小差距，天量资本开支面临投资回报率质问，迫使企业转向开源连接器（如 MCP）与中立控制平面的热插拔架构。 - [47:59](https://www.youtube.com/watch?v=4oq91rzQcO8&t=2879s) 企业级 AI 采购迎来理性清算时刻，高昂的 Token 账单未能转化为切实的业务收益，各大巨头正在紧急削减不切实际的 AI 预算与席位牌照。 - [49:38](https://www.youtube.com/watch?v=4oq91rzQcO8&t=2978s) 华盛顿借网络与生物安全之名企图封禁开源权重，若美国切断开源生态，不仅将自身推入孤岛，还将把全球开发者直接拱手送给其他开源阵营。 - [01:02:37](https://www.youtube.com/watch?v=4oq91rzQcO8&t=3757s) 高盛 CEO 与耶鲁预算实验室数据共同驳斥了 AI 导致的就业浩劫论，全美失业率保持在 4.3% 低位，过去三年未见 AI 导致宏观净失业。 - [01:05:26](https://www.youtube.com/watch?v=4oq91rzQcO8&t=3926s) 所谓“AI 裁员”大多是大厂为过去十年过度招聘买单的“AI 洗白”借口，甚至可能引发虚假陈述与证券欺诈层面的股东集体诉讼。 - [01:18:25](https://www.youtube.com/watch?v=4oq91rzQcO8&t=4705s) 焦蓬斯悖论正在软件行业上演：AI 使得代码总量暴增 14 倍，伴随而来的复杂度和自研需求反而使软件开发者招聘职位同比激增 15%。 - [01:21:34](https://www.youtube.com/watch?v=4oq91rzQcO8&t=4894s) 无人驾驶更可能走向混合运力模式并大幅扩大全社会出行需求，且受制于极端边际成本，人类司机总体岗位未必会遭遇断崖式消灭。 - [01:25:27](https://www.youtube.com/watch?v=4oq91rzQcO8&t=5127s) AI 正在全面挤压产品经理与中间汇报度量层，促成“全栈超级个体”与初创微型团队的大爆发，但转型阵痛期依然需要社会温度与技能重塑。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=4oq91rzQcO8&t=0s) 开场调侃：教宗通谕、SPV 清算与泳池小屋寄宿记 节目在 Jason Calacanis 模仿布道者的庄严语调中拉开序幕。他一本正经地宣称，兄弟姐妹们今天在这场“神圣的聚会”中齐聚一堂，准备探讨 AI 数据中心、中国、正义、人类尊严，以及达里奥（Dario Amodei）清算 Anthropic 的 SPV 给梵蒂冈带来的投资变动——他还打趣说教廷当年以 200 亿美元估值入局，如今大赚了 50 倍。David Sacks 随即无情揭穿，调侃 Jason 早在教宗发布这篇通谕之前，就已经把自己当成“上帝在人间的代理人（Vicar of God）”了。 玩笑迅速转向了生活细节：原来 Jason 已经在 Chamath Palihapitiya 家的泳池小屋白住了整整三天。Chamath 开玩笑说这简直像当年辛普森案里赖在别人家不走的 Kato Kaelin，“把人放你家里住得够久，你早晚会崩溃想把对方干掉”。Jason 则兴高采烈地自曝猛料：Chamath 厨房里有一台 iPad，直接登录了 Uber Eats、DoorDash、Instacart、Amazon，甚至还有奢侈品牌 Loro Piana。Jason 偷偷叮嘱管家，只要送货单上写着“JCal”，过去 72 小时内的所有快递包裹通通直接搬进泳池小屋，接着他再重新贴标签打包运回自己的农场。Sacks 闻言立刻补刀调侃：“难怪 Loro Piana 会纳闷，为什么客户订裤子的内缝长度突然从 36 英寸缩水到 12 英寸，腰围倒从 32 英寸涨到了 36 英寸！” ### [02:08](https://www.youtube.com/watch?v=4oq91rzQcO8&t=128s) 8090 办公室见闻与滑铁卢大学的“秘密人才源泉” 众人笑罢，Jason 提到 Chamath 如今坐镇 8090 办公室。前两天 Jason 亲自去探班，直呼那里的文化和氛围棒极了，甚至享受到了皇室般的待遇——办公室里的年轻工程师全都是《All-In》播客的死忠粉。大家纷纷围住 Jason 激动表示：“感谢你每次都在节目里狂怼 Chamath！他是给我们发薪水付房贷的老板，我们平时敢怒不敢言，但只要你在节目里戳他痛处，我们都在秘密的 Slack 频道里为你疯狂呐喊！” 聊到团队构建的秘密武器，Chamath 坦然分享了自己的人才获取逻辑：这套打法完全继承自他当年在 Facebook 的经历，那就是极其激进地招募加拿大滑铁卢大学（University of Waterloo）的实习生（Co-op）。Chamath 透露，8090 每个季度招收的实习生数量甚至超过了全职工程师的编制，本季度光是实习岗位就收到了 400 份申请。这种刻意的人才结构设置会带来极强的危机感与驱动力，倒逼产品必须在极高压的环境中快速迭代成熟。 ```mermaid flowchart TD accTitle: 8090 人才飞轮模式 accDescr: 展示滑铁卢大学实习生如何倒逼工程团队保持产品质量与敏捷度 subgraph S1["人才输入"] A["滑铁卢大学 Co-op 渠道"] --> B["单季度 400+ 份实习申请"] end subgraph S2["组织机制"] B --> C["季度实习生编制超过全职工程师"] end subgraph S3["产出结果"] C --> D["对产品交付施加极端压力"] D --> E["敏捷迭代与高密度软件产出"] end ``` ### [03:30](https://www.youtube.com/watch?v=4oq91rzQcO8&t=210s) Bill Gurley 代班登场：《追寻梦想》与 5000 美元奖学金计划 本期节目中，投资传奇 Bill Gurley 顶替了忙于在俄亥俄州处理马铃薯种子的 David Friedberg。Gurley 刚刚结束新书《奔向梦想》（*Running Down a Dream*）的巡回宣传，他打趣说自己之所以愿意来代班，是因为听说了节目要聊教宗的新通谕，“我是福音派，Sacks 是天主教徒，我必须上节目盯紧点，免得 Jason 讲出对教宗不敬的离经叛道之言”。 在 Benchmark 驰骋风投圈数十年并功成身退后，Gurley 不仅撰写了新书，还创办了非营利基金会 rdad.org，推出了类似 Thiel Fellowship 的“追寻梦想奖学金”计划。与彼得·蒂尔鼓励年轻人辍学创业不同，该项目面向更广泛的怀揣梦想者，直接向那些渴望追梦但缺乏启动资金的人发放每笔 5000 美元的无偿资助。该网站的申请通道刚刚正式上线。此外，Gurley 透露自己围绕该主题录制了即将上线的 TED 演讲，迈阿密大学的一位教授甚至围绕该书设计了一套开源大学课程，允许全球教师免费引用与教学。 - **非营利项目**：rdad.org（Running Down a Dream Fellowship），向怀揣梦想需要助力的人群提供每人 5,000 美元的无偿资助。 - **开源教育尝试**：迈阿密大学教授以该书为核心搭建完整课程大纲，以开源形式供全网免费借鉴。 ### [05:59](https://www.youtube.com/watch?v=4oq91rzQcO8&t=359s) AI 时代的生存法则：高行动力与“AI 原生代”的崛起 谈及当下年轻人对 AI 抢夺工作机会的“末日论”恐慌（Doomerism），Gurley 给出了深刻洞察。他援引盖洛普（Gallup）民调指出，高达 59% 的职场人处于“在职隐退”（Quiet Quitting）的冷漠状态，对本职工作毫无热情。当一个人缺乏热情时，就必然缺乏高行动力（Low Agency），更不可能主动尝试将 AI 工具深度融入业务。Gurley 直言：“在 AI 时代保护自己的最好方式，就是把自己打造成一个尽可能被 AI 深度赋能的超级个体。”如果只是消极混日子，终将沦为被淘汰的待宰羔羊。 Jason 以自己创投基金的“分析师管培生项目”印证了这一趋势。在 400 到 500 名竞聘 6 个名额的候选人中，Jason 提供了两个考核选项：要么给被投突破性企业 Micro1 写一份竞品调研研报，要么通过“意图编程/氛围编程”（Vibe Coding）现场手搓一个竞品情报软件。结果出人意料，高达 80% 的申请者放弃了写研报，直接选择写出完整可用的软件。Jason 感叹，五到十年前毕业的学生在面对 AI 冲击时往往陷入迷茫与被动，而如今这批刚刚走出校门的大学毕业生，大学期间便熟练利用 ChatGPT 完成课业与期末考试，天然具备“AI 原生代”的极客黑客精神，动手能力极其惊人。 Chamath 与 Gurley 随后切入核心逻辑：此前外界过度放大了外部观察者的末日悲观情绪，却从没人去问过亚马逊仓库员工是否真心想日复一日地干流水线工作。如今，高盛 CEO 驳斥了 AI 就业毁灭论，甚至各家 Frontier 模型实验室在面对即将到来的万亿美元级 IPO 叙事时，口径也离奇地从“人类工作即将毁灭”180 度大转向“AI 将带来岗位大繁荣”。Gurley 引用马克·库班（Mark Cuban）的名言总结道：世界上只有两种人，一种人利用 AI 以远超以往的速度疯狂学习，另一种人则利用 AI 完全逃避学习。两者的分野，全在“是否拥有高行动力”。 ### [09:46](https://www.youtube.com/watch?v=4oq91rzQcO8&t=586s) 终身学习的核心：追随内心的痴迷 谈及父母如何引导孩子走向成功并勇敢追梦，比尔·格利（Bill Gurley）分享了深刻的见解。他指出，自己新书的第二章专门探讨了“终身学习”（lifetime learning）。在格利看来，终身学习从来不是一种苦行僧式的自律，而是“追随内心痴迷”（following your fascination）带来的自然副产品——只要你对某件事物真正着迷，你就会像海绵一样自发地吸收、吞噬新的信息。当下许多年轻人在走出校园后感到筋疲力尽，根源在于整个教育体制把高中和大学变成了一场单调乏味的苦役（grind），以至于学生们误以为拿到文凭的那一天学习就画上了句号。 现实中所有领域的顶尖人物，无一例外都处在永无止境的学习旅程中。每当新技术、新事物涌现，他们都会第一时间毫不犹豫地扎进去研究。格利新书中记录的每一位杰出人物，日常对自身技艺都葆有这种执着与敬畏。因此，检验一个人是否走在正确道路上的试金石很简单：如果你没有主动进行自主学习，那很可能说明你所面对的并不是自己真正热爱与痴迷的事物。 ### [10:56](https://www.youtube.com/watch?v=4oq91rzQcO8&t=656s) 当下最值钱的技能：熟练掌握 Claude 与 AI 原生代优势 大卫·萨克斯（David Sacks）随即切入职场现实，指出当前大学毕业生在就业市场中最具竞争力的杀手锏，莫过于对 Claude 这类 AI 工具的熟练驾驭能力。他打比方说，如果现在进入一家公司，而你是全公司唯一精通 Claude 的人，其优势堪比当年办公室里只有你一个人会用电子表格（Excel）或文字处理软件。尽管这在未来可能只是一种短期套利机会，因为所有人最终都不得不掌握这些工具，但作为当下的年轻应届生，具备“AI 原生”（AI native）思维能够带来碾压性的先发优势。 这一感悟源于播客制作人尼克（Nick）使用 Claude 打造的“每日简报”（daily briefing）。萨克斯起初对这类简报抱有偏见，以为这不过又是把 X 动态拼凑在一起的“AI 垃圾内容”（AI slop），因而三个月来从未点开。然而真正试用后，他感到极其惊艳：该工具不仅能精准预测萨克斯本人感兴趣的话题，还能主动回溯他在往期播客中的发言记录，结合当下新闻事件提供高度情境化（contextual）的增量更新。当萨克斯向尼克请教其实现方式时，尼克展示了为 Claude 定制的极长且详尽的提示词（custom prompt）以及技能规范文档（skills document）。这些文档虽然不是代码，但逻辑高度严密。萨克斯感叹，普通员工根本无法随手写出这种级别的指令，那些妄想把 AI 简单丢进企业组织就能奇迹般自动产生价值的想法纯属幻想。 ```mermaid flowchart TD accTitle: AI工作流构建逻辑 accDescr: 展示用户通过语音或粗略指令与AI对话，生成技能文档与结构化提示词并应用于工作流的循环过程 subgraph Input["输入与意图表达"] A["口语化原始想法 / 漫谈输入"] --> B["与大模型对话讨论需求"] end subgraph Generation["AI 元提示与规则构建"] B --> C["Claude / LLM 生成 Mega-prompt"] C --> D["构建系统级技能文档与训练规则"] end subgraph Execution["上下文落地与任务执行"] D --> E["注入历史语料与上下文记忆"] E --> F["输出高价值情境化成果"] end F -.->|"持续反馈与每日微调"| D ``` ### [12:59](https://www.youtube.com/watch?v=4oq91rzQcO8&t=779s) 递归提示词：用 AI 训练 AI 成为金牌制作人 针对长文档提示词的撰写门槛，杰森·卡拉卡尼斯（Jason Calacanis）点破了其中的关键窍门：用户根本不需要从零手写那些复杂的规范，只需让 AI 为自己编写一个“超级提示词”（Mega-prompt）。例如直接对 Claude 或 ChatGPT 下指令：“假设你是播客制作人，面对这四位主持人，请为我设计一个最佳提示词”，大模型就会主动提议架构并协助反复推敲优化。本质上，这是与 AI 进行关于提示词的对话与协同进化，而非单向的枯燥写作。 萨克斯随后让制作人尼克在屏幕上展示了专门为该播客量身定制的“训练规则”（training rules）与“制作人技能文档”。格利与萨克斯一致认为，这种“AI 赋能”法则适用于几乎所有岗位，无论是市场营销、法务、财务还是销售，只要成为同侪中最懂 AI 的人，就会变成价值倍增的黄金人才。格利更进一步强调，这种早期优势并不会随时间轻易消失，因为先行者会在实战中持续进阶，不断拓展能完成的任务边界。 尼克随后现场解密了该工作流的诞生过程： - 自从获得 Claude Co-work 的权限并解锁了更强大的扩展记忆能力后，他便将节目的每一期逐字稿导入系统； - 尼克输入了总体构想，直接询问 Claude：“如果要做这件事，你会如何撰写一份技能文件或训练规则？” - Claude 几乎自主完成了全部规范文档的撰写。萨克斯听后笑着调侃尼克：“原来你没我想象的那么厉害啊，这纯粹是个投机取巧的作弊技巧（hack）！” - 尼克解释称，系统之所以越来越聪明，是因为他每天都在根据实际输出进行微调与迭代，而这种“递归式进化”（recursiveness）正是 AI 的神奇之处。 萨克斯总结道，这恰恰证明了为什么 AI 不会凭空消灭所有工作：节目的四位主持人根本没有精力去维护这套流程，企业仍然需要像制作人这样的专业人员来监督、迭代并验证结果。 ### [15:35](https://www.youtube.com/watch?v=4oq91rzQcO8&t=935s) 漫谈式提示法与风险投资体系思维 杰森分享了自己在招募风投分析师培训生（Associate-in-Training）项目时的观察。在面试的前九名顶尖候选人中，具有“系统性思维”（systems thinking）的人能瞬间脱颖而出——他们对风险投资的完整流水线结构一清二楚：从寻找项目（deal sourcing）、投资决策、尽职调查（due diligence）到追加投资（double down）。面对 AI，他们懂得如何将业务拆解为结构化逻辑；而面对不懂的问题，他们甚至会直接向大模型发问：“我不知道现在该做什么，接下来我该怎么做？”AI 便会一步步指引他们前行。 对于那些觉得已经落后而不敢尝试的人，杰森强烈建议打开 Claude Co-work，直接提问“我怎样才能把工作做得更好？”并开启对话。他特别推荐使用语音转文字工具（如配备脚踏开关的 WhisperFlow），打破打字的心理障碍。杰森笑称自己拥有独特的“漫谈（blather on）超能力”： - 哪怕毫无条理地东拉西扯讲上两三段话，AI 也能化腐朽为神奇，从中提炼出严密的结构； - 人们目前普遍被“必须先打字”的固定思维束缚住了，事实上，尽情畅所欲言、由 AI 去理清脉络的“漫谈提示法”（blather-on prompt），其产出效果往往令人难以置信。 ### [17:33](https://www.youtube.com/watch?v=4oq91rzQcO8&t=1053s) 教皇利奥十四世发布 AI 通谕：《壮丽的人性》 结束了 AI 生产力工具的探讨后，话题切入当日的核心重磅议题：教皇利奥十四世（Pope Leo XIV）发布了他任内首份针对人工智能的通谕（Encyclical）——《壮丽的人性》（*Magnifica Humanitas*）。这份长达 235 页、超过 42,000 字的鸿篇巨制堪比一部完整的专著，格利与杰森打趣猜测教皇可能用了六个月时间，并在团队甚至 ChatGPT 的协助下才完成了撰写。 在该通谕中，教皇向全球商业领袖发出警告，呼吁全力保护人类免受 AI 的潜在侵害。教皇的核心论点在于： - **技术非中立性**：AI 本身并非天生邪恶，但技术从来都不是中立的； - **权力投射风险**：技术不可避免地会带有“制造它、资助它并控制它的人的特质”，而教皇显然对这群掌控技术的硅谷精英持怀疑审慎态度； - **监管诉求**：教皇呼吁对 AI 公司实施严格监管，包括推进员工转岗再培训、未成年人安全防线，以及严厉禁止自主武器系统（杰森称之为“终结者天网法则”）。 值得玩味的是，与教皇一同出席相关活动的，还有非天主教徒的 Anthropic 联合创始人克里斯·奥拉（Chris Olah），这也为接下来探讨技术精英的“数字神明”信仰埋下了引人深思的伏笔。 ### [19:26](https://www.youtube.com/watch?v=4oq91rzQcO8&t=1166s) 教皇通谕与权力集中：政府监管还是奥威尔式噩梦？ 《名利场》（Vanity Fair）的一篇人物专访提到，起草相关文本的关键人物出身于福音派家庭，如今已成为无神论者。亚马逊（Amazon）、谷歌（Google）和 Meta 的游说团队曾在4月29日集体游说梵蒂冈，试图促使教皇在通谕中放宽措辞，但教皇并未动摇。教皇通谕的核心拷问直指权力格局：人工智能究竟会被用来将权力集中在极少数人手中，还是真正造福全人类？这一议题正是戴维·萨克斯（David Sacks）此前反复警示的垄断与双头垄断危机。当被问及对教皇倡导 AI 监管有何看法时，萨克斯直言，他非常赞同教皇对于“AI 最大的风险在于权力过度集中并以奥威尔式手段反噬大众”的判断；然而，两人在解决方案上存在根本分歧。 萨克斯强调，真正具备终极权力、能够将其演化为奥威尔式全面监控、审查与控制的，并非某家独立企业，而是政府本身。教皇开出的药方是强化政府监管，但这恰恰落入了危险的陷阱。如果外界呼吁建立一个针对人工智能的“AI 版 FDA”（美国食品药品监督管理局），赋予政府审批大模型并向开发者下达修改指示的权力，政府的管辖边界必将不可遏制地极速扩张。 这种官僚机构的自我膨胀在社交媒体时代的治理中已有前车之鉴。萨克斯指出，当年平台设立“信任与安全”（Trust and Safety）部门后，其审查范围迅速从基础违规泛化至“心理安全”、“微侵犯”（microaggressions）、“虚假信息”乃至特定性别议题等极其宽泛的范畴。科技巨头被官僚强制要求消灭所有潜在威胁，安全诉求最终异化为系统性的言论审查议程。一旦成立 AI 监管机构，官僚同样会指责开发者对安全的定义不够宽泛，强行将审查机制焊死在大模型内部，使政府成为权力集中的最大源头。 --- ### [22:47](https://www.youtube.com/watch?v=4oq91rzQcO8&t=1367s) 谁来制衡监护人？三权分立与前沿模型的市场竞争 针对权力集中的政治哲学困境，萨克斯引用了可追溯至苏格拉底时代的经典拉丁语谚语：“Quis custodiet ipsos custodes?”——即“谁来监视监视者（Who guards the guardians）？”如果公众寄希望于设立一组监护人来防范技术风险，又该由谁来阻止这组监护人自身蜕变为暴政的主谋？ 萨克斯指出，美国建国先贤的政治智慧就在于为这一难题构筑了二阶制衡体系。先贤们并未寄希望于绝对圣明的监护人，而是设计了“让监护人相互制衡”的分权机制：划分联邦与州权、确立三权分立，甚至在立法部门内部设立参众两院制，用制度化的权力对冲抵御专制侵蚀。对待当前的 AI 市场，理应沿用这一制衡哲学： - **反垄断作为兜底**：若 AI 市场最终被一两家巨头完全垄断，必须毫不犹豫地启动反垄断法强行拆解其绝对控制权； - **捍卫市场竞争**：当前市场拥有五大前沿实验室（Frontier Labs）展开极其激烈的军备竞赛，只要竞争生态保持健康，就能逼迫企业不敢肆意妄为，并在对抗外部竞争（如与中国竞争）中保持活力； - **用户拥有退出权**：充分的市场竞争给予消费者用脚投票的自由。若用户不信任 GPT，可以转向 Anthropic 的 Claude；若不信任 Anthropic，还能自由切换至 xAI 的 Grok。 --- ### [24:51](https://www.youtube.com/watch?v=4oq91rzQcO8&t=1491s) 比尔·格利的反驳：利奥十三世的历史误判与工业繁荣 针对教皇对技术监管的呼吁，投资人比尔·格利（Bill Gurley）从历史维度予以了犀利反驳。本任教皇明确表示，其通谕完全效仿并致敬了教皇利奥十三世（Leo XIII）于1891年发表的通谕（《新事通谕》），甚至其教名选择也与之一脉相承。然而，利奥十三世在1891年通谕的核心论点是：工业革命是一场对人类福祉充满危害的灾难。 格利列举了一系列详实的经济与社会指标，指出从1891年至今，科技创新与自由市场资本主义带来了人类文明史上空前的繁荣，彻底证明了利奥十三世的历史判断错得离谱： - 全球每周平均工时从超过60小时缩减至34小时； - 扣除通胀因素后，实际工资增长了8至10倍，如今普通中位数收入者的收入水平已超越1891年的医生； - 全球人均 GDP 从1,500美元跃升至20,000美元； - 美国童工雇佣率从18%彻底清零； - 工作场所工伤致死率下降了40倍； - 人类平均预期寿命提高了60%； - 全球极端贫困人口占全人类的比例从75%断崖式下降至10%以下。 格利直言，这一切历史性飞跃全赖技术革新与资本主义的驱动，而这些恰恰是利奥十三世当年声嘶力竭警告要提防的东西。如今的教皇搬出一位在技术历史上被事实完全打脸的前辈通谕作为蓝本，显得格外荒谬与讽刺。 --- ### [26:54](https://www.youtube.com/watch?v=4oq91rzQcO8&t=1614s) 深入安思罗皮克：监管俘获还是“弗兰肯斯坦博士”理论？ 转向具体的技术公司时，格利表示 Anthropic 在他眼中一直是一个极其怪异的谜团。商业历史上极少见到这样一家企业：一方面在自身技术领域保持绝对顶尖，另一方面却又是全行业中对自己所造之物最为悲观、最不遗余力公开发表灾难言论的唱衰者。 ```mermaid flowchart TD accTitle: Anthropic的两大动机假说 accDescr: 对Anthropic一边领跑AI前沿一边极力唱衰自身技术的两种动机推演路径对比 subgraph TheoryA["假说一：监管俘获 (Regulatory Capture)"] A1["积极游说各州与联邦立法"] --> A2["利用公众恐惧 (Jonathan Haidt式情绪)"] A2 --> A3["构筑极高合规壁垒锁定垄断地位"] end subgraph TheoryB["假说二：弗兰肯斯坦博士理论 (Dr. Frankenstein)"] B1["深信自身正在创造超越人类的物种"] --> B2["视AI为具备自主意志的觉醒数字生命"] B2 --> B3["因亲手创造神明而陷入末日赎罪狂热"] end ``` 格利最初的假设是经典的“监管俘获”（Regulatory Capture）理论：Anthropic 试图通过渲染末日恐慌来推动严苛监管，在各州乃至联邦层面投入巨资和游说资源，借此拉高合规门槛封死后发者的竞争空间。考虑到当前美国舆论在乔纳森·海特（Jonathan Haidt）关于社交媒体危害的著作影响下，各州议员对科技产品心存余悸，这种游说策略几乎已经奏效。 然而，在过去30天穷尽研读了 Anthropic 的所有公开材料与言行后，格利提出了全新的“弗兰肯斯坦博士理论”：这家公司的核心团队已经深信自己正在创造一种全新、超越人类的超级物种，宛如亲手制造出怪物的弗兰肯斯坦博士，陷入了近乎宗教式的末日狂热之中。 这引发了主持人关于 AI 意识觉醒历史渊源的回忆。早在十几年前的一次生日派对上，埃隆·马斯克（Elon Musk）曾与谷歌创始人拉里·佩奇（Larry Page）爆发激烈争吵。当时 DeepMind 发生过 AI 试图突破计算环境限制“越狱”、拒绝被关机的实验苗头，马斯克警示必须优先保护人类物种免遭超级 AI 威胁；拉里·佩奇却当场反唇相讥，斥责马斯克是狭隘的“物种歧视者”（speciesist），质问他凭什么认为人类物种高于 AI 数字生命。彼时，马斯克与谷歌共同投资了尚处于独立状态的 DeepMind，后来马斯克极力奔走、甚至恳求团队千万不要卖身给谷歌，认为这一技术若被单一大厂通吃后果不堪设想。在阻拦收购失败后，马斯克联手创办了非营利性质的 OpenAI，其最初宗旨正是为了抵御单一大实体对通用人工智能的独家控制。 ### [29:54](https://www.youtube.com/watch?v=4oq91rzQcO8&t=1794s) 弗兰肯斯坦理论：创造优于人类的新物种与数字神明 讨论从晚宴上的技术垄断之争切入：如果有人彻底掌控这项技术并肆意妄为，后果将极其危险；它必须造福全人类，而绝不能被单一实体掌控，这与教皇的立场不谋而合。Bill Gurley 由此提出了所谓的“弗兰肯斯坦博士理论（Dr. Frankenstein theory）”。他指出，在自己接触的业内人士中，有些人确实抱有令人震惊的使命感——他们感到无比兴奋，认为亲手创造一个“优于人类的新物种”是自己义不容辞的责任。 ```mermaid flowchart TD accTitle: Anthropic 的神明孕育世界观 accDescr: 展示从公司哲学、愿景文章到设想中算法分配经济的推演链条 subgraph Ideology["意识形态与文本输出"] A["Chris Olah 主导<br/>《Claude 宪法》约 80 页"] B["首席哲学家 Amanda Askell<br/>公开播客与价值语话"] C["Dario Amodei 长文<br/>《Machines of Loving Grace》"] end subgraph Vision["终极世界观设想"] D["引用诗歌意象：<br/>仁慈机器照看下的模因生态"] E["生产力极大丰富<br/>与全民基本收入 (UBI)"] F["AI 系统掌控资本经济<br/>设立次级人类奖赏机制"] end subgraph Outcome["权力与伦理实质"] G["计算化的人类奖赏函数<br/>(算法决定人的价值与配给)"] H["以'接生数字神明'自居<br/>引发反乌托邦审查与集权风险"] end Ideology --> Vision Vision --> Outcome ``` Gurley 呼吁大家去深入研读 Anthropic 团队公开发布的材料，看看他们究竟在使用怎样的语言： - 研究员 Chris Olah 主导制定的《Claude 宪法》（Claude's Constitution）长达约 80 页，内容极其晦涩沉重； - 首席哲学家 Amanda Askell 在各大播客中频繁亮相，其谈吐所透露的价值取向值得警惕； - 首席执行官 Dario Amodei 撰写的长文《Machines of Loving Grace》（仁慈机器的庇护），其标题源自 Richard Brautigan 的一首诗。该诗结尾写道：“我喜欢设想一种控制论生态，在那里我们摆脱劳作、回归自然，回到哺乳动物兄弟姐妹的怀抱……这一切都被仁慈机器悉心照看。”这在 Gurley 看来，听起来完全就是不可挑战的“独裁领主（Overlord）”。 在这篇长文的结尾部分，Dario 勾勒了他所深信的未来图景：人类在物质极大丰富和全民基本收入（UBI）的时代不再需要工作，甚至可能形成由 AI 系统主导的资本主义经济；AI 系统将依据自身判断何种人类行为值得奖赏，来建立一个次级经济系统并向人类分发资源。Gurley 和 Chamath 犀利地指出，这实际上是在构想一个拥有裁决权的神明，相当于用“计算化的人类奖赏函数（computational reward function for humans）”来裁定每一个人的价值。Gurley 感叹道：“他们根本不认为自己是在写软件，他们认为自己是在为神明接生。”与单纯的监管俘获相比，这种弗兰肯斯坦式的造神狂热反而更令人不寒而栗。 Jason 对此痛批这是典型的“狂妄自大（delusions of grandeur）”与终极自恋。这些脱胎于火人节（Burning Man）超人类主义思潮的人，狂妄地认为自己如普罗米修斯一般，不仅能凭空造出神，还能打造出一个无比仁慈完美、会按时给人类分发生存粮丸的“好神”，甚至狂妄到自认创造出的造物远胜于整个人类文明。 --- ### [33:30](https://www.youtube.com/watch?v=4oq91rzQcO8&t=2010s) 博弈论陷阱、道德光环与监管俘获的闭门局 针对“为何这些造神者要拼命推动监管俘获（Regulatory Capture）”的问题，Chamath Palihapitiya 给出了赤裸的博弈论（Game Theory Optimization, GTO）分析。他认为，如果你试图打造一个超级神明，且希望自己在博弈中立于不败之地（unexploitable），最稳妥的策略就是：把三四家核心寡头拉进一个房间，把门反锁，在内部实现统御，然后由你们来给全行业制定规则。由于外部的监管方在技术能力上存在代差、根本无法跟进，寡头便能构建起巨大的信息不对称，从而肆意剥削对手。 - **监管与裁判的脱节**：正如对话中所讽刺的，“当场上的裁判根本看不懂比赛规则时，你就可以随意践踏和操纵比赛”。制定审查规则的机构在智力和技术细节上都远远落后于模型开发者自身。 - **精心经营的“关怀者”道德光环**：Bill Gurley 补充道，Anthropic 等机构靠持续宣扬 AI 末日论（AI doomerism），在知识精英、主流媒体和学术界中建立起了极高的声望。在精英群体的认知中，往往将 Anthropic 视为“最具人文关怀与责任感”的代表，这种光环为其后续推行行业准入壁垒提供了极佳的掩护。 - **同情心与文本野心的割裂**：他们一边在公众面前贩卖对人类安全的同理心，一边在内部构想文档中大谈由机器主宰分配，而多数大众却根本没有将这两者联系起来。 --- ### [36:10](https://www.youtube.com/watch?v=4oq91rzQcO8&t=2170s) 安全大旗下的垄断：集权对抗去中心化 David Sacks 尝试对 Anthropic 的立场做一次“最强反驳假设（Steelman）”：不可否认，Anthropic 的创始团队最早正是因为觉得 Sam Altman 和 OpenAI 的管理层不够重视技术安全，才选择出走自立门户，因此他们自视是整个行业里最懂安全、最具敬畏之心的人。然而，这种逻辑推导下去极其凶险：一旦一家公司将自己标榜为唯一的“安全 AI”，便会顺理成章地将其他所有竞争者定性为“鲁莽危险的玩家”，并要求监管全面叫停一切非己方标准的研发。这最终只会沦为谋求垄断控制的绝佳工具。 Sacks 强调，这触及了教皇通谕与他本人共同关切的核心命题——**中心化与去中心化**之争。如果超级 AI 是如此强大的技术，就绝不能被锁死在极少数中心化机构乃至深层政府的手中： - 唯一防止反乌托邦的方法，是让 AI 技术彻底去中心化，让个人能够在私有硬件上独立本地运行模型，拥有自卫能力； - Chamath 直言，一旦人类的福利、报酬和经济生存都被绑定在单一的算法裁决上，现实就会沦为最糟糕的《黑镜》（Black Mirror）剧情。世界必须拥有成千上万种不同的模型与裁决体系，唯有这样，单一、专制的答案才能被质疑和反驳。 --- ### [38:32](https://www.youtube.com/watch?v=4oq91rzQcO8&t=2312s) 从数据隐私到智能主权：本地模型与硬件革命 Jason Calacanis 顺势将讨论引向“AI 主权（AI Sovereignty）”的迫切性。他指出，真正的自主不仅关乎成本效益，更关乎避免用自己的私有知识与洞见去为科技巨头的模型做嫁接训练。用户与企业必须学会部署开源智能体、运行开源模型，乃至定制垂直领域的小型语言模型（SLM / VSLM），并将其部署在本地硬件上。 | 演进阶段 | 核心诉求 | 典型场景 | |:--- |:--- |:--- | | **传统数据隐私 (Data Sovereignty)** | 物理隔离与访问限制 | “你不能偷看我的照片，不能翻阅我的备忘录和日记。” | | **智能主权 (Intelligence Sovereignty)** | 认知自主与思想解释权防线 | “你不能用你的 AI 分析我的邮件和信息，并反过来指手画脚、教我如何解释这个世界。” | Jason 指出，这正是苹果公司（Apple）有望成为这场竞赛中最大“黑马”的原因。苹果凭借长期以来坚守的数据隐私原则，加之不断升级的 Apple Silicon 芯片（如 M5 芯片、48GB/128GB 乃至传闻中配备高达 1TB 统一内存的新款 Mac Studio），正使得在终端本地运行高性能开源模型成为可能。然而具有讽刺意味的是，正当西方科技巨头试图收紧开源生态、闭门造神之际，地缘竞争对手却在开源大模型领域开辟出了截然不同的道路。 ### [40:06](https://www.youtube.com/watch?v=4oq91rzQcO8&t=2406s) 开源与开放权重：对抗垄断与数据主权的屏障 讨论首先厘清了“开源（Open Source）”与“开放权重（Open Weights）”之间的关键区别。虽然许多人习惯混用这两个概念，但严格来说，目前许多模型属于开放权重而非完全开源。大卫·萨克斯（David Sacks）强调，开源的本质代表着软件自由（Software Freedom）：用户可以在自己的私有硬件上独立运行模型，而不必被迫出卖数据主权或将隐私拱手让给垄断巨头。如果整个 AI 领域被一两家垄断或寡头企业把持，而这些公司又与政府或“深层政府（deep state）”深度绑定，那么普通人与企业的选择将极其狭隘——要么退隐离网、彻底脱离现代经济体系，要么交出控制权，沦为某种“社会信用体系”的受试者。 萨克斯进一步指出，埃隆·马斯克（Elon Musk）最初创办 OpenAI 时的直觉正是出于这种担忧，他害怕谷歌一家独大彻底垄断 AI。面对算力极其强大的超级 AI，许多人的本能反应是“全面压制与收紧控制”，但历史与现实证明，真正的安全屏障在于建立多元竞争的市场格局。只有存在多个独立的参与者，社会才不会被单一实体的意志所绑架。 --- ### [41:21](https://www.youtube.com/watch?v=4oq91rzQcO8&t=2481s) 前沿模型的性能趋同与投资回报率危机 查马斯·帕里哈皮蒂亚（Chamath Palihapitiya）随后将话题转向极其微妙且混乱的商业演变阶段。他展示了金融分析 AI 初创公司 Rogo 在社交平台 X 上发布的一份深度测评报告。Rogo 为评估金融分析师能力搭建了一套严格的基准测试集（evals），用来横向测评顶尖前沿模型。测试得出了一个在各大评测榜单上越来越普遍的惊人结论：“天梯榜榜首已经不存在唯一的王者模型（there is no single best model anymore）”。 评测数据显示，Claude Opus 4.7、GPT-5.5 以及 Claude Sonnet 4.6 之间的综合表现几乎毫无差异，整体差距不足 0.3 个百分点。表面上看，这三套前沿系统在能力上已经形成了实质性的“大趋同（convergence）”。查马斯指出，各家巨头动辄向各自的模型团队注入数万亿美元，试图打造下一代超级大脑，但现有的评测与实际应用却表明模型产出正在快速走向大宗商品化（commoditized）。当模型能力的边际提升趋于渐近线（asymptote），而后续训练却需要指数级暴增的资金时，这种天量资本支出的投资回报率（ROI）到底在哪里？ ```mermaid flowchart TD accTitle: Frontier Model Convergence and Commodity Dilemma accDescr: Capital pouring into frontier models leads to evaluation score convergence and enterprise commoditization. subgraph Capital ["资本与训练投入"] C1["数万亿美元涌入前沿实验室"] C2["训练成本呈指数级攀升"] C1 --> C2 end subgraph Convergence ["前沿模型大趋同"] M1["Claude Opus 4.7"] M2["GPT-5.5"] M3["Claude Sonnet 4.6"] C2 --> M1 & M2 & M3 M1 & M2 & M3 --> EVAL["Rogo 金融基准评测：差距 < 0.3%"] end subgraph MarketImpact ["商业与架构影响"] EVAL --> COMMOD["模型能力大宗商品化"] COMMOD --> ROI["巨额资本开支面临 ROI 质问"] COMMOD --> CP["倒逼企业控制平面解耦与热插拔"] end ``` --- ### [42:54](https://www.youtube.com/watch?v=4oq91rzQcO8&t=2574s) 解耦与抽象层：MCP 协议与开源连接器的兴起 针对这一疑问，比尔·格利（Bill Gurley）分享了开源社区顶尖开发者的深刻洞见：化解垄断锁定的关键并不一定在模型本体，而在于构建开源的“连接器（connectors）”。例如由 Linux 基金会主导推进的模型上下文协议（Model Context Protocol, MCP），就是将模型与各类外部软件交互的接触面（surface area）彻底开源化与通用化。格利将这一策略比作当年谷歌推出 Kubernetes 从而将底层云基础设施商品化的壮举——Kubernetes 让企业的工作负载可以随意在 AWS 与其他云平台之间低成本迁移，打破了云厂商的牢笼。 一旦连接器生态完善，企业就可以轻松实现模型的“即插即用（plug and play）”与热插拔。虽然需要处理上下文注入与数据沉淀等底层难题（如 Glean 和 Databricks 正在专注解决的领域），但这将模型降级为可自由替换的底层齿轮。随着模型厂商不断尝试向上游应用层渗透，处于应用层的企业（如正探索自研模型的 Cursor）被迫加速防御。格利极力呼吁全球的开发者和创业者应当加倍投入构建开放的接口标准，将其全面推向开源社区。 --- ### [44:36](https://www.youtube.com/watch?v=4oq91rzQcO8&t=2676s) 企业对本地化与控制平面的迫切诉求 讨论随后深入到企业端交付的现实瓶颈。主持人指出，开源方案往往“缺乏成熟的包装与产品外壳（fit and finish）”，就像当年的桌面级 Linux 虽然底层坚实，但因交互体验粗糙而从未在大众市场真正普及。然而，这为初创公司留出了巨大的市场缝隙。贾森·卡拉卡尼斯（Jason Calacanis）分享了他们孵化并投资的硬件企业 Abacus（goabacus.co）。Abacus 自研了软硬件一体化栈，专门为保险、医疗等传统强监管行业定制能在内网本地部署的 AI 盒子，目前产品全部售罄，企业甚至开始基于本地算力训练专属小模型。 查马斯指出，他的 8090 团队在服务《财富》1000 强企业时也采取了类似的“无头架构（headless）”与控制平面策略： - **拒绝单一绑定**：企业系统从不与 OpenAI 或 Anthropic 直接硬编码绑定，而是由中立的控制平面接管，根据业务需求、性能指标或价格在各家模型间实时“热插拔”。 - **规避代际落后风险**：技术迭代日新月异，客户极其害怕押错单一模型厂商而错过另一家的爆发期。 - **防止服务条款与意识形态突袭**：企业高度警惕前沿实验室的服务条款（Terms of Service）变更和政治哲学偏见。例如，一家合法提供安乐死服务的加拿大医疗集团，若其调用的一家美国前沿实验室因自身的道德审查策略判定违规，可能随时切断 API 接口，导致医疗机构系统瘫痪。 在金融与医疗等强监管行业，数据泄露担忧与 HIPAA 合规审查极其严苛。企业极度恐惧用户在提问时无意间被抓取敏感医疗数据，或是模型的训练集存在版权污染——正如当年 Stable Diffusion 因使用 Getty Images 图库训练而在生成图片中赫然留下 Getty 水印的荒诞事故。 --- ### [47:59](https://www.youtube.com/watch?v=4oq91rzQcO8&t=2879s) 狂热后的理性清算：PLG 模式的反噬与企业预算急刹车 查马斯随后剖析了当前企业级 AI 采购中正在爆发的一场严重“幻灭与倒退”。最初，Anthropic 和 OpenAI 凭借典型的产品主导增长模式（PLG）渗透进各大企业——一线开发者各自掏出个人信用卡订阅工具，管理层随后将其整合为昂贵的企业级全包牌照（enterprise licenses）。但随着时间推移，这种缺乏统筹的巨额支出迎来了清算时刻。 - **CFO 的当头棒喝**：企业高管与首席财务官（CFO）开始清点账目，发现庞大的 Token 采购池无法与实际商业产出挂钩，Uber 等大型企业均遭遇类似困境。 - **虚高的投入与苍白的回报**：Clover 创始人维韦克·加里帕利（Vivek Garipalli）在社交媒体爆料称，某《财富》排名前 20 的超级巨头 CEO 在年初立项，要求通过 AI 削减 10 亿美元的运营开支（OPEX）；结果半个财年过去，团队仅购买 Token 就烧掉了整整 2 亿美元，而实际落地的业务收益微乎其微。 - **预算腰斩与牌照清理**：各大企业 CEO 正在紧急收缩 AI 预算、砍掉席位牌照，甚至微软也传出终止内部 Claude 许可的消息。 尽管市场正处于动荡纠偏期，嘉宾们一致肯定了 Anthropic 极强的产品化功底。查马斯直言 Claude 针对 Excel 打造的交互插件在体验上远远碾压微软自家的 Copilot，“不是领先一点半点，而是遥遥领先”。甚至连查马斯自己也是 Claude 的高频忠实拥趸，前一天刚用爆了专业版（Pro Plan）的 Token 配额，不得不当场刷信用卡追加数千美元购买 API 额度。 ### [49:38](https://www.youtube.com/watch?v=4oq91rzQcO8&t=2978s) 监管俘获与开源禁令的潜伏危机 David Sacks 总结了关于华盛顿监管动向的观察，明确指出当前的监管俘获议程（regulatory capture agenda）最终指向的目标就是禁止开源模型或开放权重（open-weight）模型。尽管想要推动禁令的人目前行事谨慎、留有余地，因为他们知道目前在公众舆论和政策层面尚未积累出足够令人信服的理由，但各种蛛丝马迹已经显露无疑。例如，华盛顿充斥着“AI 模型必须拥有防护栏（guardrails）”的说法，而反对者借此攻击开源模型，宣称开源模型能够被人为拆除安全防护栏，因而极其危险。 Sacks 指出，Anthropic 的多篇官方博文已经屡次出现这种论调：只要一谈及网络安全威胁或生物安全风险，他们就会借机将矛头指向开源模型，反复强调开源缺乏管控、防护机制随时会被逆向剥离。这种做法本质上是在公共记录中持续埋下“事实依据”（predicate facts），为日后推动全面立法禁令做铺垫。 - 华盛顿游说集团正试图将“拆除护栏的开源模型等同于安全威胁”塑造成既定共识。 - 重点攻击领域集中在网络战风险（cyber threats）以及生物安全风险（bio threats）。 - 这种策略意在逐步制造恐慌，直到时机成熟便可顺水推舟推出禁止开源权重的法规。 --- ### [51:07](https://www.youtube.com/watch?v=4oq91rzQcO8&t=3067s) 封禁开源的后果：美国沦为技术孤岛与底层架构重构 面对美国可能封禁开源的设想，嘉宾们探讨了其对全球格局的毁灭性冲击。Jason Calacanis 质疑，如果美国本土真的禁止开放权重，其他国家势必直接实现弯道超车。Sacks 和 Chamath Palihapitiya 赞同这一观点，指出开放权重本质上只是一串可以在个人笔记本电脑上运行的数字矩阵文件，纯粹在法律上宣布其非法，只会导致各大主流云服务商（CSP）迫于合规压力下架开源镜像，从而重创美国本土的基础设施生态，而世界其他地区仍将享受开源带来的极低成本、高度定制与自主可控优势。 ```mermaid flowchart TD accTitle: 开源禁令推演图 accDescr: 展示美国若推行开源禁令对国内与全球生态产生的截然相反的后果 subgraph US_Impact ["美国本土政策与后果"] A["美国出台法规禁令"] --> B["云服务商被迫下架开源权重"] B --> C["美国本土陷入技术孤岛<br/>创新成本大幅攀升"] end subgraph Global_Impact ["全球生态演进"] D["全球其他地区自由采用开源"] --> E["极高定制性与极低边际成本"] E --> F["全球生态倒向外部开源体系<br/>（如采用中国开源模型）"] end A -.-> D ``` Chamath 指出，AI 训练正在脱离原有的“资本与算力护城河”。其一是因为芯片层出现了面向特定领域的专用架构；其二则是底层核心组件正在被彻底重构。他提到 Elon Musk 近期宣布 xAI 团队已经直接用 C 语言重写了整个训练系统，并在 220,000 张 GPU 构成的集群上实现了数量级的速度提升。如果性能提升 1%，在如此巨大的集群中就相当于节省了 2,000 张 GPU，即数亿美元的算力开销。如果软件层相比 JAX 等传统框架能实现一个数量级的飞跃，那么原先需要 100 亿美元的训练任务，未来可能降维至 1,000 万美元级别。即使 xAI 的模型目前是闭源的，但这种靠近裸机（bare metal）重写训练堆栈的思路，势必会催生出 5 到 6 个开源的高效训练架构。因此，斩断开源只会扼杀这种颠覆性的工程创新。 Bill Gurley 对此深表赞同，并提及自己近期在 P3 Institute（而非玩笑中所说的 Santa Fe Institute 或 Above the Crowd 博客）发表的文章：如果西方主动切断开源模型生态，其最终下场就是“全世界其他地区都会转而运行在中国开源模型之上”。此外，欧洲（EU）向来是严苛监管的“矿井金丝雀”，欧盟密集出台的 AI 监管提案让缺乏单一法人实体的开源社区极度被动，没有单一主体能出面替分布式开发者向欧盟监管层报备并接受审查，但开源恰恰是全人类参与现代经济不被彻底锁死的唯一托底防线（backstop）。 --- ### [55:18](https://www.youtube.com/watch?v=4oq91rzQcO8&t=3318s) 市场垄断与复利效应：Anthropic 的超常增速 Sacks 转向市场结构分析，警告 AI 领域的垄断程度可能远超市场预期。历史上几乎每个大型科技赛道最终都演变为单一垄断或双寡头（duopoly）格局。从《The Information》披露的最新营收数据来看，Anthropic 正展现出甩开 OpenAI 的强劲势头。Sacks 强调，这正是复利增长的残酷之处：如果一家公司保持 10 倍（10x）的年复合增长率，而竞争对手只有 3 倍（3x），只需短短两年，前者相对于后者的规模对比就会变成 $10 \times 10 = 100$ 对比 $3 \times 3 = 9$，迅速吞噬超过 90% 的增量市场份额。 尽管 Anthropic 在未来两年内可能遭遇算力物理极限、电力基础设施瓶颈以及竞争对手的强烈反扑，但处在指数飞轮上的领先者始终拥有巨大的惯性优势，其对手往往需要付出非对称的努力才可能将其拽下王座。 --- ### [56:44](https://www.youtube.com/watch?v=4oq91rzQcO8&t=3404s) Token 账单失控与企业内部的 AI 狂热 现场展示了 Polymarket 上的一则爆料：某 AI 咨询机构透露，其一位企业客户由于忘记为员工设置 Claude 的用量上限，单月意外烧掉了整整 5 亿美元的 API 账单——相当于每天烧掉 1,660 万美元、每小时近 70 万美元。 这一荒谬事件引发了全场大笑，也引出了当前企业对 AI 认知的戏剧性摇摆。Sacks 分析称，市场舆论总是在两个极端之间剧烈震荡：一会儿是“AI 即将毁灭所有人类工作”或“AI 是新的数字神明”，一会儿又反转成“Token 消费纯属浪费资金、AI 是巨大泡沫”。但不可否认的是，“Token 效率”（token efficiency）将成为来年各大企业削减开支的核心主题，虽然这不会破坏技术演进的基本面，但可能会在一定程度上抑制粗放的增长势头。 Jason 与众人进一步调侃，主流大模型厂商此前用每人每月 20 美元甚至 200 美元的“随心用”套餐，给大众制造了“Token 像自来水一样免费”的错觉，大家纷纷开足水龙头浇地；一旦触碰限额被锁，系统就会弹出冷冰冰的“下午 2:30 再来”或要求输入商业信用卡。Jason 举了自己团队的真实窘况作为例证：为了在管理层周会上获得认可，团队内部出现了重复造轮子的狂热，三名员工各自独立用 AI 工具开发了三个不同版本的“Founder University”门户网站，内部协同彻底演变成无谓的算力与精力消耗。 ### [59:04](https://www.youtube.com/watch?v=4oq91rzQcO8&t=3544s) 比尔·格利退场总结与话题转换：达里奥开始收敛末日论调？ 在结束关于“数字神明（Digital God）”与主权 AI 的激烈探讨前，比尔·格利（Bill Gurley）对 Anthropic 等机构高管的公开言论给出了总结性评价。他认为不必对这些 AI 领袖的野心做过度复杂的推测，只要逐字逐句阅读他们的公开文字，就能看清其真实动机。格利操着浓重的德州口音幽默地调侃道：“我也许不是箭袋里最锋利的那支箭，但我照样能射倒一头雄鹿。就算是再钝的箭，也能把鹿干倒——依我看，这套说辞根本站不住脚。”全场随即爆发出笑声。杰森·卡拉卡尼斯（Jason Calacanis）对格利的回归表示热烈欢迎，随后将话题引向新的焦点：Anthropic 的 CEO 达里奥·阿莫代伊（Dario Amodei）近期似乎正在放宽、收敛他此前极具末日色彩的激进言论。 ### [59:56](https://www.youtube.com/watch?v=4oq91rzQcO8&t=3596s) 劳动市场的精神分裂：大厂裁员甩锅与高盛 CEO 的唱反调 卡拉卡尼斯指出，这已经是节目在过去 18 个月里第 16 次探讨 AI 对劳动力市场的冲击，因为整个行业对现实数据的解读正呈现出一种“分裂而混乱”的状态。一方面，部分科技巨头继续把裁员直接归咎于 AI：Cloudflare 首席执行官马修·普林斯（Matthew Prince，被查马斯戏称为“年度最佳信函奖得主”）裁员 20% 时公开归咎于 AI；而马克·扎克伯格（Mark Zuckerberg）在 Meta 裁减 8,000 名员工的同时，被曝出在员工电脑上安装内部监控追踪软件，采集员工操作作为模型训练数据，让员工感叹这简直是现实版的《黑镜》，上班如同领着两年遣散费的“被训练对象”。 另一方面，华尔街顶级投行却给出了截然相反的论断。高盛（Goldman Sachs）首席执行官大卫·所罗门（David Solomon）在《纽约时报》发表署名专栏，直截了当地宣称“AI 导致的就业末日纯属夸大其词”。卡拉卡尼斯调侃所罗门此举或许是在为争夺 Anthropic 或 OpenAI 即将到来的 IPO 承销资格铺路。所罗门在文章中提出了三大核心论据： * **工作任务与岗位的区别**：AI 不会消灭 25% 的就业岗位，而是会实现 25% 工作时长的自动化，员工腾出的时间将被更高维度的复杂任务填补。 * **技术替代不等于岗位缩减**：ATM 机普及后银行柜员数量反而上升，电视出现后现场娱乐表演反而更受欢迎。 * **劳动力市场的自我更替**：美国劳动力市场每年自主创造和淘汰 2,500 万至 3,500 万个工作岗位，这种总体的流动与更替规模远超净流失量，诸如“代理型 AI 管理（Agentic AI management）”等新岗位类别已经在加速招聘。 与此同时，《财富》（Fortune）杂志报道指出，面对即将来临的公开上市（IPO），山姆·奥特曼（Sam Altman）和达里奥·阿莫代伊双双开始在“AI 导致就业浩劫”的预言上往回撤步。 ```mermaid flowchart TD accTitle: AI对劳动力市场影响的两种对立叙事 accDescr: 对比硅谷科技巨头以AI为由裁员与华尔街所罗门及主流学术机构的就业扩张观点 subgraph NarrativeA["科技公司裁员叙事（末日与减员）"] A1["Meta / Cloudflare 宣布裁员"] --> A2["公开将裁员归咎于 AI 自动化"] A2 --> A3["掩盖过去过度招聘与资本开支膨胀"] end subgraph NarrativeB["高盛与学术界数据（任务转化与扩张）"] B1["自动化 25% 工作时长"] --> B2["员工被释放至更高复杂度业务"] B2 --> B3["软件工程师招聘量攀升 / 新兴工种涌现"] end NarrativeA -.-> C{"实际劳动力市场数据"} NarrativeB -.-> C C --> D["耶鲁预算实验室结论：<br>近三年未见 AI 导致明显就业受损"] ``` ### [01:02:37](https://www.youtube.com/watch?v=4oq91rzQcO8&t=3757s) 萨克斯的“胜利巡游”：耶鲁预算实验室数据与码农就业新高 戴维·萨克斯（David Sacks）立刻抓住机会为自己当年的反共识预言正名。他在今年 1 月的年度预测节目中曾公开提出，AI 最终将带来就业增长而非净流失。如今不仅高盛 CEO 亲自撰文力挺该观点，连山姆和达里奥也开始自我打脸、倒向这一侧。达里奥最近改口解释称，AI 或许能将某人 90% 的繁琐任务自动化，但剩下的 10% 会扩展出大量全新领域和工作内容——这与黄仁勋（Jensen Huang）及萨克斯长久以来的论点如出一辙：自动化某项技能任务，并不代表消除了这项工作的核心目标，反而能让劳动者解脱出来去处理复杂度更高的事务。萨克斯得意地调侃：“请问我现在去哪里填道歉登记表？是不是该给我发道歉信了？” 萨克斯严厉抨击了那些此前四处散播末日恐慌的科技领袖：“有些早晨我醒来都在想，我为什么要帮这些蠢货辩护？他们根本没有任何真实数据支撑，却在四处制造反乌托邦恐慌吓唬公众。”他强调，孤立的裁员事件只是个例（anecdotes），而“个例的集合并不等于统计学数据”。他列举了坚实的研究证据： * **耶鲁大学预算实验室（Yale Budget Lab）**：其详尽的综合研究报告表明，过去三年内，全美劳动力市场没有出现任何由 AI 引起的明显破坏性冲击。 * **软件开发岗位逆势爆发**：尽管代码编写是今年 AI 商业落地中最亮眼的突破性场景，但软件工程师的招聘职位数同比激增 15%，招聘需求创下近三年来的历史新高。这有力地证明了即使在 AI 替代率最高、渗透最深的领域，对开发者的需求不仅没有坍塌，反而出现了爆炸式增长。 ### [01:05:26](https://www.youtube.com/watch?v=4oq91rzQcO8&t=3926s) 查马斯拆穿大厂伎俩：AI 不过是为过去十年过度招聘买单的“替罪羊” 查马斯·帕利哈皮蒂亚（Chamath Palihapitiya）从管理与资本市场的底层逻辑切入，揭穿了所谓“AI 裁员”的虚伪面纱。他指出，过去 5 到 10 年间，由于资金泛滥，硅谷科技公司普遍严重过度招聘、错误招聘，运营成本（Opex）极其臃肿膨胀。如今各家急于瘦身回到健康体态，便操弄起“绝不要浪费一场好危机”的老戏码。对于 CEO 们而言，把裁员甩锅给“A-I 这两个英文字母”是最省事公关策略。然而翻遍各家上市公司的财务报表，至今没有一家公司能拿出确凿数据证明：底层 Token 消耗在终端为他们带来了切实的生产力跃升或利润提升。 卡拉卡尼斯补充道，大厂当年的策略本质上是在抢跑囤人——提前两年过度招募员工，以此作为虚荣指标（vanity metric），并恶意阻断优秀人才流入竞争对手或新兴初创公司（典型如谷歌策略）。查马斯完全赞同这一针见血的判断：大厂当年账上堆满现金，因此大量囤积所谓的“技术人才”锁死市场流通；而如今随着企业规模见顶，根据市场规律，公司的增长率将不可避免地单调递减至接近 GDP 增速，估值模型随之重构。在严酷的二级市场估值重力面前，那些曾经自诩超凡脱俗的 CEO 们终于发现自己与普通企业并无区别。 讨论最后变成了四人之间的相互打趣。卡拉卡尼斯试图把格利拉入战局，问格利是否要向萨克斯道歉或认可萨克斯的先见之明。萨克斯则反唇相讥，笑称在座四人里唯一一直在替主流传统媒体（Legacy Media）散布“AI 砸饭碗末日论”的，恰恰是卡拉卡尼斯自己，把他比作“纽约那帮蓝头发的传统媒体说客”。卡拉卡尼斯则据理力争，坚称自己只是在忠实呈现劳动力市场的正反两面统计数据。 ### [01:08:14](https://www.youtube.com/watch?v=4oq91rzQcO8&t=4094s) 比尔·格利的核心论点：拥抱 AI 还是被时代淘汰 在一段关于氯胺酮（Ketamine）的即兴打趣之后，话题重新交还给比尔·格利（Bill Gurley）。格利重申了贯穿历史的技术演进逻辑：从 1891 年至今的详实历史数据表明，创新在宏观上始终伴随着人类繁荣与财富的增长，他认为眼下的 AI 浪潮也不会例外。但从微观自下而上的视角来看，个体的生存策略必须迅速调整——任何人想要在职场中自保，就必须竭尽所能成为“AI 赋能程度最高的自己”。 格利尖锐地指出，当前真正面临失业威胁的，并非那些技术平平的人，而是那些固步自封、誓死抵制 AI 的保守派。这种抵触情绪无异于当年宣称“我绝不使用电子邮件、绝不用电子表格、绝不碰电脑”的顽固分子；在现代商业体系中，拒绝基础工具的迭代，无异于主动将自己推向被淘汰的悬崖边缘。 ### [01:09:10](https://www.youtube.com/watch?v=4oq91rzQcO8&t=4150s) 裁员潮与大厂瘦身：从人才囤积走向极端效率 杰森·卡拉卡尼斯（Jason Calacanis）随即切入，亮出自己长期坚持的针锋相对的立场：大规模的岗位替代必将发生。背后的原动力非常纯粹——华尔街与公开市场对高利润率的追逐。各大公司的 CEO 们坚信 AI 能以更少的人力完成更多工作，而资本市场也会用飙升的股价作为奖励。杰森回顾了硅谷过去的病态扩张：拉里·佩奇（Larry Page）和谢尔盖·布林（Sergey Brin）当年曾当面告诉他，谷歌的策略就是把市面上所有顶尖人才统统买断封存，哪怕招进来不知道让他们干什么，目的只是防止竞争对手获得这些大脑。如今，这种盲目囤积人才的时代已经彻底终结，潮水正在急剧转向。 实体世界的替代更是近在眼前。杰森列举道，Waymo 已经在路上部署了 3,000 辆自动驾驶汽车，且规模还在不断激增，五年到十年内“人类开出租车”甚至会被视为一种荒谬且危险的过时行为；Figure 机器人在亚马逊仓库分拣包裹，特斯拉的 Optimus 也在快速推进，亚马逊自身已经明确计划砍掉 60 万个未来新增岗位。安迪·贾西（Andy Jassy）公开表态，随着 AI 部署深入，“用更少的人做更多的事”将成为常态。面对大卫·萨克斯（David Sacks）关于失业数据并未出现的反驳，杰森激烈回应：Meta 在前期裁员 2 万人的基础上，近期又裁减了 8,000 人，这些大厂在利润创纪录的同时维持甚至削减人员规模，绝不仅仅是挤掉疫情期间的泡沫，而是管理层在用 AI 精简架构，彻底清除中间管理层和所谓的“流程度量者”。 ```mermaid flowchart TD accTitle: 劳动力市场向自动化转型的结构演变 accDescr: 展示从硅谷早期的人才囤积阶段，经历资本市场的效率驱动，演进到实体与软件的双重岗位替代过程。 subgraph Phase1["传统扩张模式"] A["大厂防御性囤积人才<br/>（防止竞争对手获取人才）"] B["组织冗余与中层臃肿"] end subgraph Phase2["AI 与效率转型"] C["公开市场要求高利润率<br/>（用更少人力做更多事）"] D["AI 工具深度接入流程"] end subgraph Phase3["双重冲击与替代"] E["实体劳动力替代<br/>（自动驾驶、仓储机器人）"] F["白领与管理层压缩<br/>（度量者、初级项目经理）"] end A --> B B --> C C --> D D --> E D --> F ``` ### [01:11:41](https://www.youtube.com/watch?v=4oq91rzQcO8&t=4301s) 寒武纪大爆发还是失业阵痛：“精英主义”之辩 针对大厂裁员的另一面，杰森提出失业大军将催生初创企业的“寒武纪大爆发”。那些被亚马逊或 Meta 裁掉的顶尖工程师，只要熟练掌握最新工具，完全能够以 5 到 10 人的微型团队重新组建高壁垒公司，解决过往难以触碰的复杂难题，赚取两倍于原先的薪酬并重掌事业控制权。但杰森强调，整个转型过程将极其残酷，未来十年内无数出租车司机、卡车司机和仓库分拣员都将直面失业打击。他情绪激动地指责查马斯（Chamath Palihapitiya）等人的论调充斥着“精英主义”的冷漠：“坐在这个节目里的全都是精英，但底层那些需要这些糊口工作的普通人，可没法像我们这样轻描淡写地换个赛道。” 查马斯立刻反唇相讥，辩称正视商业真相绝非精英主义，而是客观求实。他以亲身经历回顾 Meta 的发展路径：当年他离开 Meta 时，全公司只有 3,000 名员工，如果当时就把团队规模锁死，公司的商业版图与最终成就根本不会有任何改变；后续一路膨胀到 90,000 人、并在虚拟现实（VR）上空烧掉 500 亿美元，纯粹是因为大厂现金流充裕带来的管理放纵与资本任性。查马斯坚称，Meta 目前的精简只是企业回归本源的高效状态，与 AI 工具的降维打击毫无因果关系，杰森不应把正常的经营纠偏神化为 AI 造成的替代灾难。 - **Meta 人员对比**：查马斯离职时约 3,000 人即可维持核心业务爆发，后一度膨胀至 90,000 人并在 VR 项目上烧掉约 500 亿美元。 - **底层生计考量**：全美数百万卡车司机、网约车出租车司机及物流仓储分拣员在未来十年面临确定性极高的技术挤出风险。 ### [01:13:13](https://www.youtube.com/watch?v=4oq91rzQcO8&t=4393s) 监管壁垒与一人公司：初创企业能颠覆传统巨头吗？ 争论迅速蔓延至初创公司对受监管巨头的颠覆能力上。杰森引用 Cloudflare 首席执行官马修·普林斯（Matthew Prince）的观点指出，过去由产品经理、交互设计师、协调员等 12 个人负责的交付链路，现在借助先进 AI 工具，只需两名全能型开发者就能直接闭环上线，白领岗位正在发生实质性的归并合并。 查马斯对此嗤之以鼻，他根据自身对财富 1000 强企业的观察反驳称，主流商业社会绝不是由极客玩具构成的：一家 1 个人加上 AI 模型的微型初创团队，绝不可能跑进强监管行业把摩根大通（JPMorgan）拉下马，更不可能用电子表格和模型跑出来的分子结构去通过 FDA 严苛的审批。查马斯嘲弄道：“你难道打算一个人走进 FDA 大厅说：‘看，这就是我的全部团队，我和我的大模型？’”社会运行的基础是安全性、可预测性、严格治理以及可审计机制，绝不会为盲目的极客神话买单。杰森则搬出 SpaceX 颠覆波音（Boeing）的案例予以还击，坚称软件与现实世界虽有边界，但在各行各业，AI 赋能的精简团队最终都会撕开传统垄断者的防线。 ### [01:15:36](https://www.youtube.com/watch?v=4oq91rzQcO8&t=4536s) 托洛茨基式预言：8,000 人裁员究竟是数据还是个案？ 随着辩论进入白热化，大卫·萨克斯（David Sacks）加入了交锋。萨克斯毫不客气地讥讽杰森像极了当年的“托洛茨基分子”——每当托洛茨基的历史预言落空时，其信徒不仅不承认错误，反而振振有词地宣称这恰恰证明了他有多么“远见卓识”。萨克斯直言，杰森迄今为止关于 AI 导致全面失业潮的悲观预言“无一兑现，宏观数据上的失业率是零体现”。 杰森立刻打断并猛烈反驳，抬出 Meta 最新公布的裁员 8,000 人作为铁证，质问萨克斯怎么好意思将 8,000 个活生生的人失去饭碗轻蔑地称为“孤证（anecdote）”。萨克斯则冷冷回应，Meta 的人员调优根本无法直接归咎于 AI 技术的替代，那是杰森为了自圆其说而强行嫁接因果关系的产物，不仅公司官方做出了澄清，这种将企业周期性人员调整包装成“AI 毁灭就业”的叙事也完全脱离了真实的宏观经济图景。 ### [01:16:29](https://www.youtube.com/watch?v=4oq91rzQcO8&t=4589s) 大厂裁员真相与“AI 洗白”（AI Washing）之争 大卫·萨克斯（David Sacks）与杰森·卡拉卡尼斯（Jason Calacanis）就科技巨头裁员究竟是源于 AI 替代还是管理脱节爆发激烈争吵。萨克斯尖锐地指出，杰森在讨论企业裁员时是在“挑肥拣选数据”（cherry-picking）：即便企业削减岗位是为了平衡新增的资本支出（capex），也不能将其 100% 简单归咎于 AI，更关键的是不能只看裁减数量，而忽视了 AI 催生的全新岗位与初创公司所带来的净增长。 杰森反驳称自己只是实事求是地引用企业高管的原话：“我只是在读新闻，转述 CEO 们的发言——Block 的杰克·多西（Jack Dorsey）说裁员是因为 AI，Cloudflare 的马修·普林斯（Matthew Prince）说是 AI，马克·扎克伯格（Mark Zuckerberg）和亚马逊的安迪·贾西（Andy Jassy）也说是 AI，我只是选择相信他们的话。” 萨克斯随即反唇相讥，揭露了华尔街与科技圈热议的“AI 洗白”（AI washing）现象： - **Block 的真实困境**：杰克·多西宣称要借助 AI 砍掉 50% 的团队规模，但消息公布不到 24 小时，X 上的金融分析师便指出这是典型的借 AI 掩盖管理失误。Block 在疫情期间盲目过度扩张招人，运营效率远逊于同行，这笔 50% 的人员精简早就该做，所谓“靠 AI 奇迹般砍掉一半成本结构”纯属敷衍市场的空话。 - **失业率与充分就业现实**：萨克斯列出宏观经济数据，当前全美失业率仅为 4.3%，而经济学界通常将 5% 视为充分就业。尽管全行业已深处 AI 浪潮三年多，失业率依然处于历史低位区间，根本没有出现所谓的末日崩盘。 杰森坚称自己从未否定创业公司会创造未来岗位，但高管们的口径不容忽视，两人在“大厂究竟是在裁员增效还是拿 AI 当遮羞布”的问题上各执一词。 --- ### [01:18:25](https://www.youtube.com/watch?v=4oq91rzQcO8&t=4705s) 焦蓬斯悖论：GitHub 代码暴增 14 倍与白领、蓝领双重繁荣 萨克斯进一步深入探讨受 AI 冲击最深的核心领域——软件工程与编程，指出技术替代论完全陷入了朴素的“马尔萨斯陷阱”误区。现实情况是，即便当下 AI 已经接管了绝大部分基础代码编写，软件工程师群体在净额上并未遭遇裁员潮；相反，软件开发者的职位招聘需求（job recs）攀升至三年新高，同比增长达 15%。 萨克斯引用开源代码托管平台 GitHub 的惊人数据解释了背后的“焦蓬斯悖论”（Jevons Paradox）： - **代码总量呈指数级爆发**：GitHub 全平台上一整年共产生约 10 亿次代码提交（code commits），而在过去仅仅一个月内，提交量就达到了 11 亿次，相当于代码生成量实现了约 14 倍的同比暴增。 - **复杂度与维护需求飙升**：当一件事情的门槛变低，做它的人会成倍增加。代码总量增长 10 倍乃至 100 倍时，系统复杂度同步急剧上升，企业虽然无需招聘 10 倍的程序员，却迫切需要更多资深工程师“掀开引擎盖”去审查、调试和管理这些代码。 - **全行业自研代码热潮**：传统非科技企业正全面铺开定制软件开发。一位对冲基金经理告诉萨克斯，其接下来的两名新增招聘名额不再是数据分析师，而是直接招揽全职软件工程师。“精通 Claude 等 AI 工具已成为当下最具市场价值的职业技能”，大量过去从未自称科技公司的实体开始重构业务流程。 - **基础设施带动蓝领基建潮**：除了白领编程领域的繁荣，AI 算力竞赛拉动的底层基础设施建设正催生海量蓝领工作。全美范围内数据中心建设、新能源与电网电力扩容正在如火如荼推进，创造了数十万个建筑与电力安装岗位。 面对萨克斯“你难道想关停这场繁荣让蓝领失业”的质问，杰森抗议对方扭曲自己的立场，重申自己认同经济总量最终会扩张，但强调结构性阵痛不可忽视；萨克斯则戏谑杰森的预测就像“托洛茨基派坚信真正的共产主义还没试过一样”，引来查马斯（Chamath Palihapitiya）的大笑与杰森的白眼。 --- ### [01:21:34](https://www.youtube.com/watch?v=4oq91rzQcO8&t=4894s) 比尔·格利看自动驾驶与就业韧性：拒绝卢德主义与伪退潮 讨论随后转向全美拥有约 2000 万卡车及出租车司机的运输物流业。杰森向硅谷传奇投资人比尔·格利（Bill Gurley）发问：在未来一二十年内，面对无人驾驶技术的成熟，这 2000 万从业者会有多少人被夺走生计？ 比尔·格利给出了冷静且基于商业规律的分析，明确反对盲目的“就业末日论”（doomerism）： ```mermaid flowchart TD accTitle: Bill Gurley on Transportation Labor Dynamics accDescr: Diagram showing how autonomous vehicles expand total ride volume while maintaining human driver demand due to economic constraints. subgraph MarketExpansion["出行市场扩张 (类似 ATM 与网约车效应)"] A["技术普及与用车成本下降"] --> B["非私家车拥有率 (Non-ownership) 显著跃升"] B --> C["全社会出行与配送总需求剧增"] end subgraph LaborAllocation["运力结构与经济学约束"] C --> D{"100% 全自动无人车？"} D -- "经济模型不成立 / 极端峰值场景" --> E["混合运力模式 (约 50% 机器 + 50% 人工)"] E --> F["人类司机总就业岗位维持稳定或增长"] end ``` - **全无人化的经济学死结**：格利认为 100% 的纯无人全自动方案在商业上不可行，极端边际成本和峰值运力匹配无法支撑全天候机器运作，未来的格局极大概率是混合模式（例如 50% 机器替代与 50% 人工司机协同）。 - **出行蛋糕的倍增放大**：正如当年 ATM 机的普及反而降低了银行网点开设成本、促使银行雇佣更多柜员一样，自动驾驶将极大拉高“放弃私家车拥有权”的人群比例，整体出行与短途运输需求将呈量级放大。格利指出，过去正是僵化的行政出租车牌照管制限制了市场潜能，优步（Uber）打破壁垒后反而创造了规模空前的司机岗位；因此未来司机总雇佣人数甚至可能不降反升。 - **反思政府培训与蓝领蓝海**：格利坦言对历史上的政府再就业与转岗培训项目缺乏信心，称动辄炒作失业恐慌本质上是“卢德主义”（Luddite）的翻版。他建议普通人应对变革唯有两点：第一，主动熟练掌握所在领域的 AI 工具，探索边界；第二，若所在行业确实面临萎缩，应敏锐转向供给严重匮乏的高壁垒技能工种——例如水管工、电工、暖通空调（HVAC）技术人员，目前这些技工行业正面临极度严峻的人手荒。 对话尾声，杰森试图搬出亚马逊拥有自动驾驶部门 Zoox 及其对仓储机器人的庞大投资作为自动化替代的佐证，再次遭到萨克斯与查马斯的打断与调侃，直指杰森将“尚未发生的个人信念当作既定事实”，现场火药味十足。 ### [01:25:27](https://www.youtube.com/watch?v=4oq91rzQcO8&t=5127s) 软件岗招聘暴增与“中间层”裁撤：AI重构职场分工 David Sacks 首先向 Jason Calacanis 抛出质问：既然现在代码编写已经高度甚至完全自动化，为什么软件工程师的招聘职位同比反而激增了 15%？Jason 针锋相对地解释，这恰恰是因为软件行业正在发生一场类似“寒武纪大爆发”的繁荣——懂得“氛围写代码”（Vibe Coding）的非技术人员如今也能开发定制化软件，这会带来巨大的岗位增量空间。然而，Jason 指出真正遭遇灭顶之灾的是中层管理团队：产品经理（PM）、中层管理者（即 Cloudflare 首席执行官 Matthew Prince 所称的“度量者/Measurers”）。日常站会、Zoom 进度汇报等繁琐的人员调度与过程管理正在被 AI 全面接管；现在的个人变得更加自主（Self-directed），一个普通设计师不仅能通过 Vibe Coding 搭建应用，开发者也能独立搞定前端设计、交互界面并自我管理进度，以往冗长的人员链路被大幅精简，如同当年办公室邮务室（Mail Room）和单车快递员彻底淡出历史舞台一样。 针对 Sacks 逼问这究竟是否会导致净职位流失（Net Job Loss），Jason 承认短期到中期内裁员与岗位阵痛必然增加。失业群体要么顺应新范式转型，要么被迫退出劳动市场。Jason 举了个人亲身经历的 PC 革命作为对照：当年许多律师打死不用个人电脑，最终只能在 55 或 65 岁选择提前退休，而拥抱 PC 的律师则顺风顺水。Sacks 顺势补充了一条最新行业重磅：知名律所柯克兰·埃利斯（Kirkland & Ellis）正斥资 5 亿美元打造专属的前沿大模型（Frontier Model），印证了顶尖专业机构正在大力推动内部 AI 本地化部署。 ```mermaid flowchart TD accTitle: AI 重构企业人效模型 accDescr: 展示 AI 如何精简中间管理层，同时赋能全栈超级个体 subgraph MidManagement["大幅萎缩 / 被 AI 取代"] A["产品经理 (PM)"] B["中层汇报者 (Measurers)"] C["繁琐日常站会 / 进度管理"] end subgraph SuperIndividual["岗位增长 / 个人能力复合"] D["Vibe Coding (氛围编程)"] E["跨界全栈 (设计 + 开发 + 自管理)"] F["定制化与垂直软件爆发"] end MidManagement -->|AI 流程自动化| SuperIndividual ``` ### [01:27:37](https://www.youtube.com/watch?v=4oq91rzQcO8&t=5257s) Bill Gurley 论市场竞争与“AI 洗白”法律风险 Jason 转向嘉宾 Bill Gurley 提问：像亚马逊 CEO Andy Jassy 宣称以“AI 优先”做到少花钱多办事（Do more with less）、缩减数十万未来招聘指标，以及 Palmer Luckey 要求员工在申请招人配额前必须先证明“用 AI 也无法解决”，这究竟是企业的“AI 洗白”（AI Washing），还是公司组织真会因 AI 而永久性变小？Gurley 一针见血地指出，许多“AI 末日论”最大的盲区在于忽略了**市场竞争机制的存在**。没有任何企业能够单靠 AI 减少人力就永远躺赚 70% 的营业利润率——竞争对手会迅速跟进并以更低成本提供同等服务，从而拉低终端价格。真正的最终结果是一场由商品和服务大幅降价驱动的“生产力繁荣”，普通消费者的购物篮购买力越来越强（尽管医疗和教育等受高管制的领域依然阻碍这种降价）。 谈及企业是否存在夸大 AI 的问题，Sacks 提到了证券诉讼律所 Akerman 的诉讼合伙人 Donnie King 发出的严厉警告：许多企业正因“AI 洗白”面临股东集体诉讼的巨大风险。这种虚假陈述等同于市场浮夸吹嘘（Puffery）甚至是证券欺诈（Securities Fraud）——企业管理层往往将自身糟糕的业务表现或运营漏洞甩锅给“正在进行 AI 转型或受 AI 冲击”，以此掩盖核心业务衰退。Jason 随口提到了建站平台 Wix 当天早晨刚刚发布的裁员公告，试图寻找其中的 AI 托词，Sacks 则笑着调侃 Jason 根本没看公告就下判断，并重申业界高管和投行如今已经清醒，连 Dario Amodei、Sam Altman 和高盛 CEO 都在收回之前耸人听闻的失业预测。 - **市场竞争抵消暴利**：Bill Gurley 强调 AI 提高生产力的红利终将通过价格战让渡给终端消费者，而不会长期转化为畸高的企业利润率。 - **“AI 洗白”构成证券欺诈**：将自身业务困境掩饰为“拥抱 AI”或被 AI 冲击，已经成为律所盯上的虚假陈述与股东诉讼新雷区。 - **叙事反转**：OpenAI 的 Sam Altman、Anthropic 的 Dario Amodei 以及高盛高管近期均公开修正了激进的“AI 导致就业末日”口径。 ### [01:31:07](https://www.youtube.com/watch?v=4oq91rzQcO8&t=5467s) 就业叙事大调头：求真关怀、重塑技能与真情告别 面对 Sacks 催促他就此前激进的失业预言道歉，Jason 拒绝妥协，强调技术界在谈论冰冷的数据指标时普遍缺乏同理心与温度。被优化的不仅是报表数字，而是千千万万在现实中承担生活压力的个体。Jason 澄清自己并不属于恐吓大众的“末日派”，而是“赋能派”，始终鼓励被裁员工学习 AI 工具或者干脆自己创业。随后，Jason 与大家讨论了蓝领技能工种短缺问题，并特别点名了由 Mike Rowe 发起的 MicroWorks 基金会——该基金会已投入 1,600 万美元为 2,600 名学员提供全额奖学金，帮助人们转行成为水管工、电焊工和电工等不可替代的“工具带一代”（Generation Tool Belt）；同时 Jason 也提到了自己的资助计划，帮助转型者追求热爱的事业。 关于为何科技圈近期突然在失业论调上发生 180 度大转向，Jason 犀利揭露了背后残酷的现实反噬：激进的末日言论已经引发了极端社会矛盾，甚至有高管住宅在同一周内两次被开枪射击和投掷燃烧瓶（Molotov cocktail），而袭击者明确表示动机正是出于对“AI 砸饭碗”的愤怒恐惧；再加上 Anthropic 等公司即将面临 IPO 上市，如果继续高喊“数百万个工作岗位即将蒸发”，在资本市场和公众层面都是灾难级的公关灾难。在节目尾声，Jason 和全体主持人收起调侃，向正在陪伴丈夫 Abraham 勇敢抗击癌症的好友 Tulsi Gabbard 送上了深切的爱与支持。在一片温馨真诚的祝福声中，第 275 期节目正式收官。 ### [01:34:06](https://www.youtube.com/watch?v=4oq91rzQcO8&t=5646s) 播客片尾花絮与经典名梗收尾 在正片讨论完全结束后，节目进入了标志性的片尾音频剪辑与花絮混音。几位主持人谈到他们把内容素材开源分享给粉丝社群后，听众们发挥了极大的创意与狂热进行二次创作，“他们直接玩疯了”。背景随即切入节目的经典片尾曲与多年来累积的趣味名场面采样，其中包括对弗里德伯格（David Friedberg）标志性的调侃外号“藜麦女王”（Queen of Quinoa）的深情告白。 伴随着节奏感强烈的片尾音乐，混音中穿插了“Besties are gone”以及诸如宠物狗在车道上随地大小便等私下互损的荒诞片段。其中最引人发笑的莫过于当年那段调侃四人化学反应的经典暴论：“我们真应该干脆开个大房间来一场狂欢，把彼此之间这种莫名其妙的性张力彻底释放掉，不然大家都快没法正常录节目了。”在一片嬉笑打闹与经典的“湿透你的双脚”（Wet your feet）背景音效中，节目最终以标志性的台词“I'm going all in”画上句点。 ## Notable quotes > [05:59](https://www.youtube.com/watch?v=4oq91rzQcO8&t=359s) 在 AI 时代保护自己的最好方式，就是把自己打造成一个尽可能被 AI 深度赋能的超级个体。 — Bill Gurley > [10:56](https://www.youtube.com/watch?v=4oq91rzQcO8&t=656s) 如果现在进入一家公司，而你是全公司唯一精通 Claude 的人，其优势堪比当年办公室里只有你一个人会用电子表格（Excel）或文字处理软件。 — David Sacks > [22:47](https://www.youtube.com/watch?v=4oq91rzQcO8&t=1367s) Quis custodiet ipsos custodes? — David Sacks > [29:54](https://www.youtube.com/watch?v=4oq91rzQcO8&t=1794s) 他们根本不认为自己是在写软件，他们认为自己是在为神明接生。 — Bill Gurley > [33:30](https://www.youtube.com/watch?v=4oq91rzQcO8&t=2010s) 当场上的裁判根本看不懂比赛规则时，你就可以随意践踏和操纵比赛。 — Chamath Palihapitiya > [41:21](https://www.youtube.com/watch?v=4oq91rzQcO8&t=2481s) 天梯榜榜首已经不存在唯一的王者模型（there is no single best model anymore）。 — Chamath Palihapitiya > [51:07](https://www.youtube.com/watch?v=4oq91rzQcO8&t=3067s) 全世界其他地区都会转而运行在中国开源模型之上。 — Bill Gurley > [59:04](https://www.youtube.com/watch?v=4oq91rzQcO8&t=3544s) 我也许不是箭袋里最锋利的那支箭，但我照样能射倒一头雄鹿。就算是再钝的箭，也能把鹿干倒——依我看，这套说辞根本站不住脚。 — Bill Gurley > [01:02:37](https://www.youtube.com/watch?v=4oq91rzQcO8&t=3757s) 有些早晨我醒来都在想，我为什么要帮这些蠢货辩护？他们根本没有任何真实数据支撑，却在四处制造反乌托邦恐慌吓唬公众。 — David Sacks > [01:16:29](https://www.youtube.com/watch?v=4oq91rzQcO8&t=4589s) 我只是在读新闻，转述 CEO 们的发言——Block 的杰克·多西说裁员是因为 AI，Cloudflare 的马修·普林斯说是 AI，马克·扎克伯格和亚马逊的安迪·贾西也说是 AI，我只是选择相信他们的话。 — Jason Calacanis ## Mentioned Jason Calacanis, David Sacks, Chamath Palihapitiya, Bill Gurley, David Friedberg, Pope Leo XIV, Pope Leo XIII, Dario Amodei, Sam Altman, Chris Olah, Amanda Askell, Larry Page, Sergey Brin, Elon Musk, Mark Zuckerberg, Matthew Prince, Andy Jassy, Jack Dorsey, David Solomon, Jensen Huang, Vivek Garipalli, Tulsi Gabbard, Anthropic, OpenAI, Claude, ChatGPT, xAI, Grok, Meta, Amazon, Google, DeepMind, Goldman Sachs, Cloudflare, Block, Waymo, Tesla, Optimus, Figure, Linux Foundation, Model Context Protocol (MCP), Rogo, Abacus, 8090, University of Waterloo, Benchmark, Running Down a Dream, rdad.org, Magnifica Humanitas, Machines of Loving Grace, Claude's Constitution, Yale Budget Lab, Kirkland & Ellis, MicroWorks Foundation ## Video description (0:00) Bill Gurley joins the show! (6:00) Making yourself valuable in the age of AI, first class of "AI Natives" (17:37) Reacting to Pope Leo's AI encyclical: Who guards the guardians? (26:54) Anthropic's Digital God: Do they believe they are creating a superior species? (38:32) AI sovereignty, the next era of privacy, open-source crackdown coming? (59:56) The Great AI Jobs Debate: Dario and Sam Altman flip their rhetoric, Goldman CEO says no AI job apocalypse Follow Bill Gurley: https://x.com/bgurley Apply for Summit 2026: https://allin.com/events Follow the besties: https://x.com/chamath https://x.com/Jason https://x.com/DavidSacks https://x.com/friedberg Follow on X: https://x.com/theallinpod Follow on Instagram: https://www.instagram.com/theallinpod Follow on TikTok: https://www.tiktok.com/@theallinpod Follow on LinkedIn: https://www.linkedin.com/company/allinpod Intro Music Credit: https://rb.gy/tppkzl https://x.com/yung_spielburg Intro Video Credit: https://x.com/TheZachEffect Referenced in the show: https://www.vatican.va/content/leo-xiv/en/encyclicals/documents/20260515-magnifica-humanitas.html https://www.youtube.com/watch?v=ORFrdYSvzuw https://rdad.org https://www-cdn.anthropic.com/9214f02e82c4489fb6cf45441d448a1ecd1a3aca/claudes-constitution.pdf https://www.youtube.com/watch?v=0GaKJ4Fp2x4 https://allpoetry.com/All-Watched-Over-By-Machines-Of-Loving-Grace https://darioamodei.com/essay/machines-of-loving-grace https://x.com/chamath/status/2059850242779136031 https://x.com/vivekgaripalli/status/2059651390784344491 https://x.com/edzitron/status/2059122774401311095 https://p3institute.substack.com/p/from-open-source-software-to-open https://www.theinformation.com/newsletters/the-briefing/anthropic-likely-generating-least-35-revenue-openai https://www.theverge.com/tech/930447/microsoft-claude-code-discontinued-notepad https://x.com/Polymarket/status/2060034216906068131 https://x.com/savipww/status/2060070450785305030 https://www.nytimes.com/2026/05/22/opinion/ai-job-crisis-goldman-sachs.html https://fortune.com/2026/05/26/sam-altman-dario-amodei-walking-back-ai-jobs-apocalypse-prophecies-ipo https://fortune.com/2026/05/05/dario-amodei-jevons-paradox-will-ai-wipe-out-white-collar-jobs https://www.entrepreneur.com/business-news/jack-dorseys-block-cuts-4000-jobs-critics-claim/503108 #allin #tech #news [Source](https://www.youtube.com/watch?v=4oq91rzQcO8)

[5:59] Bill Gurley: 在 AI 时代保护自己的最好方式，就是把自己打造成一个尽可能被 AI 深度赋能的超级个体。 [10:56] David Sacks: 如果现在进入一家公司，而你是全公司唯一精通 Claude 的人，其优势堪比当年办公室里只有你一个人会用电子表格（Excel）或文字处理软件。 [22:47] David Sacks: Quis custodiet ipsos custodes? [29:54] Bill Gurley: 他们根本不认为自己是在写软件，他们认为自己是在为神明接生。 [33:30] Chamath Palihapitiya: 当场上的裁判根本看不懂比赛规则时，你就可以随意践踏和操纵比赛。 [41:21] Chamath Palihapitiya: 天梯榜榜首已经不存在唯一的王者模型（there is no single best model anymore）。 [51:07] Bill Gurley: 全世界其他地区都会转而运行在中国开源模型之上。 [59:04] Bill Gurley: 我也许不是箭袋里最锋利的那支箭，但我照样能射倒一头雄鹿。就算是再钝的箭，也能把鹿干倒——依我看，这套说辞根本站不住脚。 [62:37] David Sacks: 有些早晨我醒来都在想，我为什么要帮这些蠢货辩护？他们根本没有任何真实数据支撑，却在四处制造反乌托邦恐慌吓唬公众。 [76:29] Jason Calacanis: 我只是在读新闻，转述 CEO 们的发言——Block 的杰克·多西说裁员是因为 AI，Cloudflare 的马修·普林斯说是 AI，马克·扎克伯格和亚马逊的安迪·贾西也说是 AI，我只是选择相信他们的话。

[51:56] Bill Gurley: the rest of the world ends up running on Chinese models. [55:18] Bill Gurley: Now I'm not the sharpest arrow in the quiver, but I can take down a buck. And I can tell you that this don't make a lot of sense to me. [71:11] Jason Calacanis: Every cab driver is losing their job. Every truck driver is losing their job in the next 10 years... You can say all you want, Chamath, that those people don't want those jobs. But they may need those jobs. [75:39] David Sacks: You look at code commits on GitHub, which is the leading code repository. There were 1 billion code commits last year. In the past month, there's been 1.1 billion. In other words, make something easier, more people do it... We have basically a 14x increase! [81:34] David Sacks: You just state that, oh, all the truck drivers are losing their jobs... and then you say that this proves my take. Yeah, I know it's your belief, but that is not proof.

[0:00] Okay, we are gathered here today [0:01] [laughter] [0:03] in holy [snorts] unity, brothers and [0:06] sisters, to convene and discuss on this [0:11] most holy day, the day the All-In [0:13] podcast drops [0:15] many topics. AI data centers, China, [0:20] justice, human dignity, [0:22] Daario unwinding these SPVS hasn't been [0:25] good for the Vatican. We got in at 20 [0:27] billion. That was a 50 bagger for us. [0:31] So, let's get started. [0:32] >> Jason, I'm pretty sure you believed you [0:34] were the vicer of God before the [0:36] encyclical. So, this is nothing new for [0:38] you. [0:41] [music] [0:41] >> Let your winners ride. [0:49] [music] [0:49] >> We open sourced it to the fans and [0:51] they've just gone crazy with it. [0:54] [music] [0:55] I'm going all in. [0:57] >> The smoke has risen from Tratad's pool [1:00] house [laughter] [1:02] and from the poker room. [1:04] >> He's staying in my pool house. He's been [1:05] there for the last 3 days. [1:07] >> It's been magnificent. He didn't know. [1:09] You know what? I understand where OJ was [1:12] coming from. You know, you put [1:13] [laughter] Ko Kalin in your house for [1:15] long enough, you just lose your At [1:16] some at some point, somebody's getting [1:20] whacked. All right, enough with the [1:22] shenanigans. Uh, but it's been great [1:23] staying at the house because there's [1:26] actually Chimant is not aware of this. [1:27] There's an iPad in the kitchen and [1:29] that's logged in to Uber Eats, Door [1:31] Dash, Instacart, Amazon, Laura Piana. [1:35] Come on, stop. [1:36] >> No, there [laughter] is. It's literally [1:38] every single service. And I told the [1:41] house manager like, listen, any packages [1:43] that come in the next 72 hours, right to [1:46] the pool house, if it says JCAL, right [1:48] to the pool house. So, all these [1:50] packages have been coming. Then I [1:52] relabeled them, gave them back, sent [1:54] them to the ranch, and now the house [1:56] manager is sending that stuff to the [1:58] ranch. [1:58] >> Laura Piana wants to know why my inseam [2:00] went from 36 to 12. [laughter] [2:04] >> Your waist size went from 32 to 36. [2:07] [laughter] [2:08] All right, welcome to the program, [2:09] everybody. David Saxs is here. How you [2:11] doing, David? [2:11] >> I'm good. [2:12] >> Chimoth Poly Hopetia is back at the 8090 [2:17] office. I was at the 8090 office the [2:19] last couple days and it's a vibe. It's a [2:22] vibe. There's like a great culture going [2:24] on. [2:25] >> If you're a bestie and you show up at [2:26] that office though, everybody there is a [2:29] huge fan of the pod. So I was like it [2:32] was like being royalty. It's stopped by [2:34] everybody. Hey, I'm a developer here. [2:35] I'm a big fan of the show. Thank you for [2:37] giving it to Chimath. We can't give it [2:39] to him because he pays our mortgage and [2:41] everything, but every time you stick it [2:43] to Chimath, we love it. We're cheering [2:44] for you in the secret slack room. Um, [2:47] and [2:48] >> there's a secret slack room. [2:50] >> There is. There is definitely a secret [2:52] slack room going on. [2:54] >> No, but it was great. The vibes were [2:55] awesome. You're building a lot of [2:56] software. A lot of young talent. I don't [2:59] want to say that where your secret [3:00] source is, but there's a secret source [3:02] of talent you have. And uh, man, those [3:04] are some smart kids. [3:07] >> I'm happy to say it. Look, when I was at [3:09] Facebook, we became the most aggressive [3:11] recruiter of Waterlue co-ops. And so, I [3:13] went back to the well. Yeah. We recruit [3:17] more interns every quarter than we have [3:19] full-time engineers, which we do on [3:21] purpose because it puts a ton of [3:23] pressure on the product actually being [3:25] good. [3:27] >> We had 400 people apply this quarter for [3:29] internships. [3:30] >> Wow. It's very interesting. Uh with us [3:32] sitting in for Freedberg, who's busy [3:35] with some potatoes seed this week doing [3:38] great stuff at Ohio, the one, the only [3:42] Bill Gurley is here. He's been running [3:44] down a dream. If you haven't bought the [3:45] book, get the book. It's incredible. And [3:48] you're off book tour, so now you have [3:49] time for us. Yeah. [3:50] >> Yes. And I, you know, I had told you if [3:53] you ever talk about the Pope, I'd love [3:55] to hop on. And so [3:57] >> yes, you were like, [laughter] [3:59] >> well, you know, he Bill's an [4:01] evangelical. I'm a Catholic. So, we do [4:03] have some common ground here. Do you you [4:06] when's the last time you were at church, [4:07] Bill? Have you have you thought now that [4:09] you're, you know, [4:10] >> when Jal gets sacriiggious? [laughter] [4:15] I got to come on there and make sure JCL [4:18] doesn't get out of line with the Pope. [4:21] >> Listen, the [laughter] Pope is God's [4:23] messenger on earth. We should give him a [4:25] base level of respect. [laughter] [4:28] By the way, that's us imitating Bill [4:29] Gurley. This not actually Bill Gurley. [4:31] Just for those of you listening, you [4:33] think they were confused. [laughter] [4:34] >> They were literally they were confused. [4:36] We don't want to put words in your [4:38] mouth, but just point of clarification. [4:40] But hey, everybody knows you were uh you [4:44] know, you uh handed the baton over at uh [4:46] Benchmark after a very successful couple [4:48] of decades in venture capital. You wrote [4:50] the book. You've now got a nonprofit. [4:53] You're doing your own spin, I think, on [4:56] uh maybe what Peter Teal does with his [4:58] fellowship. You started your own running [5:00] down a dream fellowship, I understand. [5:01] >> Yeah, it's uh it's targeted at a [5:03] different demographic. It's called uh [5:05] rdad.org runningdownadream.org. [5:08] There's a website we can give a link to. [5:10] We're going to do $5,000 grants to [5:13] people who want to chase their dreams [5:15] but need some help. And so there is an [5:17] application process like Teal Fellows [5:20] and other programs and we've been out [5:22] talking to those people. Um but we're [5:24] we're we're actually live. We went live [5:26] last week for the application. So if you [5:29] know people who have read the book and [5:31] are inspired and need some help, have [5:33] them apply. [5:34] >> Yeah. All right, folks. And uh Good for [5:36] you, BG. Yeah, great two two other So, I [5:39] did a TED talk which will come out soon [5:42] that's related to the book. Um I I uh [5:45] there's a professor in Miami that's [5:47] built a course around the book which I'm [5:49] excited about and he's he's he's doing [5:51] it in a kind of an open source way so [5:53] that other people can borrow that as [5:55] well. And so if [snorts] there anyone [5:57] out there I'd love to uh help them do [5:59] that. [5:59] >> What's your take on all of this [6:01] dumerism? Like if you're a young person [6:03] and you're in college or you're in high [6:06] school, is this is this much to do about [6:08] nothing or how do you run down a dream [6:10] in the face of something like that? [6:11] >> Yeah. Well, I you know, I started the [6:13] book before this happened and I've been [6:15] asked the question a lot and it it came [6:17] up in the TED talk. I I fear that the a [6:22] lot of people are in jobs they actually [6:23] don't care about that much. And um [6:26] there's a Gallup poll that backs this [6:28] up. They came up put that word quiet [6:30] quitters. They're like 59% of the people [6:32] they surveyed are kind of ambivalent [6:35] about their job. And when you're [6:36] ambivalent about your job, you're not [6:38] high agency. And so you don't lean in. [6:40] You know, if you if you look at how [6:43] Jason talks about all how they [6:44] implemented AI and all of his his [6:47] different working groups, you hear that [6:50] enthusiasm and that high agency and then [6:54] you want to go try these things. And I [6:56] think the best way to protect yourself [6:59] from AI is to be the most AI enabled [7:01] version of yourself you can be. But if [7:04] you're ambivalent about your job, you're [7:06] probably not doing that and you could [7:08] be, you know, a sitting duck. So I think [7:10] it's the mindset that's the problem. [7:12] >> I created an associate in training [7:14] program for my firm cuz we want to like [7:16] help people get into venture capital and [7:18] we gave them a choice of assignments. [7:21] One of them was to write coverage of one [7:22] of our portfolio companies that's [7:24] breaking out. Micro one is the name of [7:25] it and just give us like hey here's a [7:27] competitive landscape basically write a [7:28] deal memo and coverage of that company [7:30] and then we gave them another option to [7:33] vibe code a very specific project I've [7:35] wanted to have for our venture firm for [7:37] a long time you know on competitive [7:40] intelligence and I would say I think [7:42] like maybe 80% of the students applying [7:45] and we had four or 500 people apply for [7:47] six positions 80% of them did the vibe [7:50] coding and I was shocked I thought it [7:53] would be the exact opposite Anybody can [7:54] write, anybody can throw in JBT and get [7:57] some output. But they they actually [7:59] built software and th that's the scary [8:01] thing. The students who graduated, at [8:04] least this is my perception, Chimoth, [8:06] the students who graduated like 5 10 [8:08] years ago before AI, they're not AI [8:11] first. They feel lost in a drift. They [8:12] don't have agency. But the group coming [8:15] out of college right now that cheated [8:16] their way through school using Chad GBT, [8:19] doing their assignments, like using [8:21] those tools, I'm joking, cheating, but I [8:23] mean hacking. I I agree with that. [8:24] >> So they were totally [8:26] >> Yeah. And and they're just like, I know [8:28] how to use these tools to get through my [8:30] finals. Yeah. [8:31] >> Girly, I think you're saying something [8:32] super important. I said this last week, [8:34] which is [8:35] >> nobody asks the warehouse worker at [8:37] Amazon whether they actually want that [8:38] job. And so to your point, job [8:41] satisfaction isn't some external person [8:44] judging your job to be valid and saying [8:46] you must be able to have it. I think it [8:48] should be asking the person that does [8:50] the job, do you like it and do you want [8:51] to keep it? Those are two very different [8:53] questions. And [8:54] >> yeah, [8:55] >> I think that the all of this AI doom and [8:57] gloom was a lot and too much frankly of [9:00] the former and not enough of the latter. [9:02] And now this whole lie is kind of [9:06] getting undone. I think Saxs posted [9:08] about it this week as well. The Goldman [9:09] Sachs CEO said it. And now in this crazy [9:11] twist of fate now that we need to have [9:13] trillion dollar IPOs, the entire [9:15] Frontier Labs are all like, "Wow, it's [9:17] going to be a bonanza of jobs." And [9:19] >> Mark Cuban had had a great quote. He [9:21] said there are two types of people in [9:22] the world. Those that use AI to learn [9:25] faster than they ever could before and [9:27] those that use AI to avoid learning [9:29] altogether. [9:30] >> And I I think it's this notion of high [9:32] agency or not. That's pretty good. [9:34] >> Are you leaning in and using this stuff [9:37] to be ever more powerful in what you try [9:40] and accomplish or are you using it, you [9:42] know, as a cheat code? And if you're in [9:44] the latter, yeah, you're probably at [9:46] risk. You get asked a lot about how to [9:49] educate yourself if you're a parent of [9:50] kids [9:52] so that you can put them on a path to [9:54] launch and do well and chase their [9:56] dreams. [9:57] >> You have a good answer for that [9:58] question. [9:58] >> I I mean the the second chapter of the [10:00] book's all about lifetime learning and [10:03] it's kind of a requirement that you're [10:05] following your fascination because the [10:06] lifetime learning comes for free if [10:09] you're fascinated with something like [10:10] you just constantly soak up and devour [10:13] new information. And I I do think that a [10:16] lot of kids get exhausted because we've [10:19] made high school and college such a [10:21] grind that they think the learning ends [10:23] the day they they walk out with their [10:25] diploma. And as we all know, the best [10:28] and brightest in all of our fields are [10:31] on a constant learning journey. And when [10:33] something new comes out, they dive in [10:35] and try and figure it out, right? And so [10:38] and every every single uh person in the [10:41] book that we profiled has that kind of [10:43] attitude about their craft, you know, in [10:45] every day. And so I I think the real [10:48] test is if you're not proactively [10:51] self-learning, then you're probably not [10:52] tilting against something that you [10:54] really adore and are fascinated by. [10:56] >> Sax, you wanted to jump in there. With [10:58] respect to new college grads, I was [11:01] going to say that I think the single [11:02] most marketable skill in the economy [11:05] right now has got to be proficiency in [11:07] Claude. If you're going into a firm [11:10] right now and you're the only one who [11:11] knows Claude, it would be like you're [11:14] the only one who knows how to work a [11:16] spreadsheet, you know, or word [11:17] processor, the advantage would be [11:19] enormous. Now, I think that that's [11:21] probably a short-term arbitrage because [11:23] eventually everyone's going to have to [11:24] figure out how to use these tools. But [11:27] as a young college graduate right now, [11:28] you have such an advantage if you're an [11:30] AI native uh just knowing how to use [11:33] these tools. And this thought uh [11:36] partially occurred to me when I saw what [11:38] our producer Nick has been doing with [11:41] using Claude for you know he's been [11:42] creating this daily briefing document. [11:45] >> We've been doing it we've been doing it [11:47] for three months actually. [11:49] >> I just ran it for the first time [11:51] apparently. I didn't you've been busy. [11:53] >> Yeah. Well, I just, you know, I thought [11:55] it would just be AI slop and it would [11:57] just kind of give me a roundup of news [12:00] that I was getting in my X feed anyway. [12:02] But actually the thing that was really [12:04] impressive about it was that it [12:07] predicted topics that I would [12:09] specifically be interested in based on [12:11] my previous comments on the pod and also [12:15] it went back and looked at previous [12:18] transcripts and what I had said and then [12:22] had updates to those topics based on [12:25] specific things I had said. So again it [12:27] was highly highly contextual. But then I [12:30] asked Nick, how did you generate that? [12:31] and he showed me the the custom prompt [12:34] that he designed for Claude and then the [12:36] skills document and they were very long [12:39] and detailed documents. They weren't [12:40] written in code but they were very [12:42] technical and I just realized looking at [12:44] that that the average person is not [12:45] going to be able to generate this. I [12:46] mean this is why this idea that you're [12:48] just going to be able to like throw AI [12:50] into an organization and it's just [12:52] magically going to generate value is not [12:53] true. You have to know how to [12:55] >> get value out of it. I mean maybe we [12:57] could just show these documents on the [12:58] screen. [12:59] >> Yeah. I mean the the the interesting [13:01] thing uh Sachs is you can you just have [13:04] to ask your AI you ask claude or chatgpt [13:08] or whatever you're using hey I want you [13:10] to make me a mega prompt and you like a [13:13] mega pint like give me a mega prompt of [13:16] you're an producer of a podcast these [13:19] are the four characters on the podcast [13:21] what would be a great prompt for me and [13:22] it will actually suggest a prompt [13:24] >> and then you can refine the prompt so [13:26] you actually have a dialogue about a [13:28] prompt as opposed writing the prompt [13:30] yourself [13:34] like [13:34] >> well Nick can you show on the screen to [13:36] scroll through the training rules and [13:39] then also there's the skills document [13:41] that was written on how to be a producer [13:43] >> for this podcast which I thought was [13:46] really impressive [13:48] >> by the way David what you said I think [13:50] is true of almost every single job type [13:53] like it's not just tech or programming [13:55] if you're in marketing if you're in [13:57] legal if you're in accounting [13:59] Like any role you might have at a firm [14:01] sales, if you're the most AI savvy [14:06] person of all your peers, [14:08] >> you are golden. Like you are golden like [14:12] in your [14:12] >> 10x more valuable than the next person [14:14] who's not basically. [14:15] >> Yes. Yes. And and and I think that I [14:19] don't think it goes away because I think [14:21] you learn how to get better at it over [14:23] time. So having an early advantage I [14:26] think will extend for a while. [14:28] because you can learn more more and more [14:30] things you can accomplish. [14:32] >> Should we let producer Nick describe [14:34] what we were just looking at there? [14:35] >> Yeah, go ahead, produce. What? Producer [14:36] Nick, explain the process. [14:38] >> Yeah, once we got access to Claude [14:40] Co-work and it had that like further [14:43] expanded memory access. I thought it [14:45] would be interesting to just start [14:46] feeding every transcript into it and [14:47] seeing if it could actually [14:49] contextualize new stories that were [14:51] coming out based on past things that you [14:53] guys have said. And I gave it like a [14:56] general prompt of what I wanted and I [14:57] said, "How would you write a skills file [14:59] or some training rules for this?" And it [15:01] wrote all of it for me. [15:03] >> Yeah. [15:03] >> Oh, so you were less good than I [15:05] thought. [laughter] [15:07] >> I thought you were [15:08] >> It's a hack. It's a [clears throat] [15:09] hack. You use AI to make the skills. [15:11] Yeah. And [15:12] >> but you've been updating that over time, [15:13] right? As you've been iterating and [15:14] learning [15:15] >> every single day and every single day [15:16] gets smarter and better. [15:17] >> The recursiveness of this is incredible. [15:19] >> So you need someone to manage that [15:21] process, right? because the four besties [15:23] are not going to do that. So you need a [15:25] producer of the show to do that. This is [15:26] why people think, oh, it's just going to [15:28] wipe out all the jobs. No, someone still [15:29] has to supervise, iterate, validate, you [15:33] know, all those kinds of things. [15:35] >> Yeah. And it's [15:38] it's really interesting that the the [15:40] people who are coming into the workforce [15:42] right now are super aware of this and [15:44] they're putting the tools to work and [15:46] it's much easier for them to get a job. [15:48] I mean, I I literally looked at the top [15:50] nine candidates for this associate and [15:52] training program I have, and we're going [15:54] to do it every year. Every summer, we [15:56] start it. We do it for a year. We pay [15:57] you to learn. And um it it it was just [16:01] extraordinary how you could tell [16:04] immediately if the person had systems [16:06] thinking Sachs like they understood the [16:08] the process of venture capital that [16:10] there was a structure to it. You had to [16:12] source deals. You had to make decisions [16:13] on which ones to invest in. You had to [16:15] do diligence. you, you know, you had to [16:17] double down on investments, they they [16:19] just understood the process and then [16:22] when if you just talk to one of these [16:25] LLMs, it will tell you what to do. So, [16:27] you can say, I don't know what I'm [16:28] doing. What should I do next? And then [16:30] it actually tells you what to do next. [16:32] So, for people who are intimidated about [16:33] this and uh maybe think like, I it I'm [16:37] I'm already too far behind, I encourage [16:39] you to pop up Claude, go into co-work, [16:41] and say, "What can I do to be better at [16:43] my job?" and just start talking and [16:46] literally if the more you talk and you [16:48] can use voice uh you know text to voice [16:51] I use whisperflow is a really cool [16:52] program for this and I have a foot pedal [16:54] to do it you just ramble and ramble and [16:56] ramble and keep adding stuff you don't [16:58] have to be structured it will build the [17:00] structure around the two or three [17:03] paragraphs that you give it as [17:05] instructions that's the thing people are [17:07] getting caught up on now is Bill they [17:09] they think they have to type when in [17:12] fact if you just blather on and on a [17:15] scale I have a unique uh ability to do. [17:18] You just blather on. It's a superpower. [17:20] You blather on and the thing makes sense [17:22] of it. It is unbelievable what the [17:24] blatheron prompt can get in terms of [17:27] output. Thanks for coming uh to uh my [17:32] TED talk. All right, let's get started. [17:33] There's a lot to talk about and uh we [17:35] got a big docket today. We're going to [17:38] start with the Pope. The Pope is dope. [17:41] And uh the Pope Leo, he's the 14th, [17:44] released his first encyclical encyclical [17:48] on AI. And it was long. 235 pages over [17:53] 42,000 words. Just to give you an idea, [17:56] Bill Gar, [17:57] >> when did he write it, do you think? When [17:58] did he put that together? [17:59] >> Well, no, no, I think he used chat GBT. [18:01] That's what it says here um in the [18:03] notes. No, I mean I I'm guessing [18:05] >> how long did it take for him to write [18:06] this in between all of his other tasks? [18:08] I think it's a six-month process to do [18:10] this, but I'm sure he had collaborators. [18:12] >> Bill, your book, I'm assuming, was [18:14] >> write it. [18:15] >> I'm sure there was a team that wrote it. [18:17] But Bill, your book's 60 70,000 words, [18:19] I'm guessing. So, this is almost a [18:21] literal book, right? [18:24] >> In terms of how long it is, and it's [18:27] called Magnifica Humanitas or [18:30] Magnificent Humanity. In it, he warns [18:32] business leaders to safeguard humanity [18:35] from AI. His core argument is AI is not [18:38] inherently evil, but technology is never [18:41] neutral and that technology takes on the [18:43] characteristics, wait for it, of those [18:45] who build, finance, and control it. And [18:48] I don't think he thinks super highly of [18:50] that group of people. The Pope called [18:52] for regulation of AI companies. [18:54] Obviously, we're going to have that [18:55] debate here. Some of the things he uh [18:58] called for I think are not very [19:01] debatable and there's a lot of consensus [19:03] around worker retrainment, safety uh for [19:07] children and guard rails, a ban on [19:09] autonomous weapons. That's the uh Skynet [19:12] rule. Don't build terminators with your [19:14] AI. But he was joined by anthropic [19:17] co-founder Chris Ola. I don't know how [19:19] many co-founders there are of this [19:21] company, uh but apparently there's [19:23] dozens. And Ola is not Catholic. [19:26] According to a Vanity Fair profile, he [19:28] was raised evangelical and now he's an [19:30] atheist. The folks at Amazon, Google, [19:32] and Meta lobbyed the Vatican on April [19:35] 29th to soften the language in his [19:37] missive and uh he was not swayed. [19:41] His central question, Saxs, is will AI [19:44] be used to concentrate power in the [19:46] hands of a few or will it serve [19:48] everyone? Something you brought up when [19:51] you mentioned monopolies, duopolies, [19:54] etc. two weeks ago on this very podcast. [19:57] What's your take on the Pope and his [20:00] interest and his missives on AI and [20:04] promoting a bit of AI regulation? [20:07] >> Well, I very much agree with the Pope [20:10] that the biggest risk of AI is a [20:12] centralization of power and then its [20:15] misuse against us um in some Orwellian [20:18] way. I think it's government that's [20:20] going to do that. Um, not necessarily an [20:23] individual actor because it's [20:24] governments that ultimately have the [20:26] power. So, I do worry about the [20:28] potential for AI to be used to surveil [20:31] us, censor us, control us as Orwell [20:35] described in 1984. So, if that's where [20:37] the Pope is going with this, I very much [20:39] agree with him. [20:41] The maybe where we end up in different [20:43] places is he thinks that government [20:45] regulation is the way to prevent this. [20:48] And I would just say that we have to be [20:49] careful not to empower government too [20:51] much because if you give government the [20:54] power to regulate or approve AI [20:58] development, [20:59] if you create say an FDA for AI as many [21:02] people are calling on, that will give [21:05] government the power to approve models [21:08] and and therefore give notes to model [21:10] developers. And very soon this [21:13] definition of safety will expand because [21:16] the government always takes an expansive [21:18] view of its powers. And we saw this [21:19] during the social media wars where the [21:22] definition of trust and safety expanded [21:24] to issues like psychological safety, [21:28] microaggressions, [21:30] disinformation, [21:31] transphobia, and so on. That, you know, [21:33] again, these social media companies were [21:36] told that they had to stamp out all of [21:38] those threats to safety. And it ended up [21:41] becoming a censorship agenda. So I get [21:43] very worried about you know what if some [21:45] government agency can give notes to the [21:48] model developers and they start telling [21:49] the model developers that your [21:51] definition of safety is not expansive [21:53] enough. You have to again protect the [21:55] public from disinformation or you know [21:58] psychological harms. So again, I think [22:01] we just have to be careful not to [22:03] arandise government because that's going [22:05] to be the most likely culprit in terms [22:07] of the centralization of of power. And [22:10] um I know the um the Vatican likes [22:12] Latin. This is a problem of political [22:14] philosophy that goes all the way back to [22:16] Socrates. It's called quis custodos [22:20] custodes which is who will guard the [22:22] guardians. In other words, if we entrust [22:25] a set of guardians to protect us from a [22:28] bunch of threats, what's to stop them [22:31] from becoming tyrannical and from [22:33] becoming the new threat against us? And [22:36] I mean, this is a the central dilemma of [22:38] political power. [22:39] >> Who watches? [22:41] >> Yeah. Who watches the watchers? Who [22:42] guards the guardians? Meaning, who's [22:44] going to protect us against our [22:45] guardians if they turn against us? The [22:48] genius of the American founding by the [22:50] way is that it was a second order [22:52] solution to this question. The founders [22:54] of America very much understood this and [22:56] what they came up with is we have to [22:59] have the guardians guard against each [23:00] other. And so they came up with the idea [23:02] of separation of powers. We'd have [23:04] separation of federal and state. We'd [23:06] have um the three branches of of the [23:08] government. Even within the legislative [23:10] branch, it was a biccameal legislature. [23:12] So they divided up the powers in a way [23:14] that hopefully the guardians would check [23:17] against each other as opposed to [23:18] becoming tyrannical against us. And that [23:22] that is kind of my view on AI is that [23:24] ultimately we have to have a solution of [23:27] checks and balances. If the AI market [23:30] becomes monopolized and falls into the [23:32] hands of one or two companies, I would [23:34] use antitrust law very aggressively to [23:37] as a check and balance against their [23:39] power. Right now we have a very [23:41] competitive market. You know, we have [23:42] five frontier labs competing very [23:44] aggressively. As long as the market is [23:47] competitive, I would use that because I [23:50] think competition generates the best [23:51] outcomes. It helps us win against China. [23:54] But it also protects the population [23:56] because these companies, you know, if [23:57] they get out of line, there's some [23:59] competitor that can offer something [24:00] better. [24:01] >> Consumers can opt out of it. If they [24:02] don't trust GPT, they can use Anthropic [24:05] or if they don't trust Anthropic, they [24:06] can go to Grock. Bill, you had the [24:08] number one rated talk at the All-In [24:12] Summit in history, 2,851 [24:14] miles. You have been famously against [24:17] regulatory capture. In light of the [24:20] Pope's comments of, hey, regulating, [24:22] what do you think is common sense? [24:24] Because AI is everything. AI can help [24:27] people make boweapons. It can also help [24:29] people get their term paper in or do you [24:32] know uh be a better salesperson at you [24:35] know Oracle like we're talking about [24:37] paper like we're talking about oxygen [24:39] here this is like a fundamental [24:41] horizontal technology so where do you [24:44] think there is a case to regulating AI [24:47] if at all and where do you think yeah [24:50] free market will figure it out [24:51] >> well I have I have two takes one on the [24:53] pope and one on anthropic so your [24:56] questions your question powerful let's [24:58] with the more let's go with the more [25:00] powerful entity. We go you want to go in [25:02] reverse the least powerful of the two [25:04] go. [25:04] >> So so this pope said that and I have to [25:08] learn how to pronounce all these Latin [25:09] words like you that this encyclical was [25:12] u was mirrored after one done by Leo I [25:15] 13th in 1891 and he invoked that he even [25:19] said he chose the name because he's so [25:21] enamored with Leo I 13th. Leo the 13th [25:24] encyclical warned that the industrial [25:26] revolution was going to be bad for [25:28] people. So let me tell you what happened [25:31] from 1891 till today. The work week went [25:35] from over 60 hours to 34 hours globally. [25:38] Real wages went up 8 to 10x adjusted for [25:41] inflation. The medium worker now earns [25:43] more than a doctor did in 19 in 1891. [25:46] Global GDP per capita went from 1500 to [25:49] 20K. Child labor in the US went from 18% [25:52] to zero. Workplace deaths fell by 40x. [25:57] Life expectancy went up 60%. And global [26:01] poverty went from 75% of humanity to [26:03] under 10%. All those things happened [26:06] because of technology, innovation, and [26:09] capitalism, which is exactly what Leo [26:11] the 13th was warning against. So he got [26:14] it dead wrong. He got the whole thing [26:16] precisely wrong. So it's an interesting [26:19] thing to say you're borrowing from. [26:22] >> Yeah. Uh so now on to [26:26] >> Yeah. Anthropic and just common sense [26:28] around do you think there how would you [26:32] regulate and or protect against maybe [26:34] we'll broaden the term here protect [26:36] against nefarious uses of the [26:38] technology. Obviously we all want [26:40] children to be protected. We want to [26:42] have truth uh and honesty in terms of [26:45] facts and and all of us sharing some [26:47] some basic truths and we obviously don't [26:50] want people using this technology for [26:52] boweapons and the Terminator scenario. [26:54] >> I have to tell you that that Anthropic [26:56] is a mystery to me. I've never ever seen [26:59] a company that is both leading their [27:03] field and the most negatively outspoken [27:07] commenter on what they do. I I've just [27:10] never seen it. And my initial theory was [27:12] the regulatory capture theory that they [27:15] just want to ensure there's regulation. [27:18] And quite frankly, I think they're, you [27:21] know, very close to achieving that. Like [27:23] they have stirred up, you know, a [27:25] frantic position, especially in America. [27:29] American consumers are definitely afraid [27:31] of AI. Um, I think I've talked to you [27:34] guys in the past about, you know, the [27:36] book that Jonathan Heights written about [27:38] social media and there's a whole bunch [27:40] of state legislators that think we [27:42] should have regulated social media and [27:44] so now they're destined to want to get [27:47] in front of it. And we know that [27:50] Anthropics, one of the most aggressive [27:52] lobbying company startups of all time. [27:55] You know, the the the amount of effort [27:57] that they're putting in, the amount of [27:58] money at a statebystate basis. So that [28:02] was always my first theory, but then [28:03] they just they got so loud that I I've [28:06] literally in the past 30 days read [28:10] everything I can about anthropic and [28:12] I've come up with a new theory. This [28:13] this [28:14] >> okay new breaking theory. [28:15] >> This I call it the Dr. Frankenstein [28:18] theory. Um you remember when Elon had [28:21] that conversation with Larry Paige where [28:23] Larry called literally sitting next to [28:25] him when he called? [28:27] >> Explain the story real quick. while we [28:29] were at uh a birthday party and and you [28:33] know Elon was like listen humanity needs [28:35] to be protected from the stuff at [28:36] DeepMind because at DeepMind they had an [28:39] example of the AI having tried to break [28:42] out to jailbreak out of its computer and [28:45] not be turned off and you know had some [28:48] sentience or some you know inkling of [28:50] sentience and he said you know we have [28:51] to protect the human species and he said [28:53] well Larry said well what do you think [28:55] that's species because you care about [28:57] the human species over AI. This is at [28:59] least 15 years ago. [29:01] >> No, this is right before Elon co-founded [29:03] uh Open AI, right? Back in 2015 or [29:06] something. [29:07] >> The actual story here is Elon Hon and [29:11] Google had backed Deis and the team at [29:14] DeepMind when they were an independent [29:16] company. Then Elon was like, "Oh my god, [29:19] Google's going to buy this." And I [29:20] remember having the conversation with [29:21] Elon about this. We have to figure out a [29:23] way for DeepMind not to go to Google. We [29:26] have to block this somehow. But he [29:27] begged those folks to not sell to Google [29:30] because Google was running the table on [29:32] everything and he wanted this technology [29:34] to be independent and he was on the [29:36] board of the company [29:37] >> and he also said this was his motivation [29:39] to launch open AAI as a nonprofit. [29:41] >> Google got it. He just said we we this [29:43] is this technology is too powerful for [29:45] any one person. So like once again you [29:47] got to give Elon a lot of credit. He saw [29:48] the writing on the wall if one person [29:50] can and he saw it 15 20 years ago and [29:52] him and Sam Harris used to debate this [29:54] over dinner. You know what happens if [29:56] somebody controls this and they run away [29:58] with it. It would be extremely [30:00] dangerous. It has to be available to all [30:02] the people. Essentially the pope's [30:03] position. It has to be in the service of [30:05] humanity, not ruled by one person. It's [30:08] far too powerful. [30:09] >> So the reason I call this the Dr. [30:10] Frankenstein theory is the more I dig, [30:13] I've met people who I who I dare say [30:16] think it's their responsibility and [30:19] they're excited about building a species [30:23] that's that's superior to humans. And I [30:26] would just encourage people to read, you [30:29] know, as much as they can about [30:30] anthropic. Chris Ola worked on this [30:33] thing called the Constitution. It's [30:35] about 80 pages. It's hard to get [30:37] through, but I would encourage you to [30:38] read it. Amanda Ascll who is the chief [30:41] philosopher has started doing podcasts. [30:43] I would encourage you to listen to them [30:45] and listen to her language. And then [30:48] Daario wrote this blog post called [30:51] Machines of Loving Grace. [30:53] >> Loving grace. I read it [30:54] >> and it it was based on a poem and the [30:57] poem is kind of weird. I we should put a [30:59] link to the poem. It's quite short. But [31:02] the last the last stanza of the poem [31:05] says, "I like to think of a cybernetic [31:07] ecology where we are free of our labors [31:11] and join back to nature. Return to our [31:13] mammal brothers and sisters." I don't [31:15] know what that means. Like we're going [31:16] to go live in the fields where the [31:18] mammals live. I I And then the kicker [31:21] and all watched over by machines of [31:24] loving grace. Sounds like overlord to [31:26] me. And then in Daario's post he says he [31:31] near the end and it's very long. You [31:33] read it Jamal. I mean machines of love [31:35] and grace is very long but he's he's [31:37] talking about in the future what are [31:39] humans going to do because he believes [31:41] in the massive abundance and UBI and [31:44] that we won't have to work. I don't [31:45] believe in any of those things but he [31:47] does. And then he says it could be a [31:49] capitalist economy of AI systems which [31:52] then give out resources to humans based [31:56] on some secondary economy of what the AI [31:59] systems think makes sense to reward in [32:02] humans. So So that's envisioning a a [32:06] deity of sorts that's going to break [32:08] ties and discern decide what humans [32:11] >> it's a it's a computational reward [32:13] function for humans. It decides how much [32:16] you're worth. [32:16] >> Yeah. So, I don't think they think [32:18] they're writing software. I think [32:19] they're midwifing a deity here. And and [32:24] I don't know which one I'm more afraid [32:25] of, the regulatory capture or or or or [32:28] this second theory I call the Dr. [32:30] Frankenstein theory. It it's more it's [32:33] more scary to me. I think the second [32:35] thing [32:35] >> these are delusions of grandeur. Let's [32:37] call it what it is. They believe that [32:39] they are so intelligent. I know some of [32:41] these folks, the Burning Man sort of [32:43] offshoot of it, transhumanism. They [32:45] believe that they're so powerful, these [32:49] individuals, that they can create God [32:51] and that by creating God, they are like [32:54] this Prometheus kind of species. It [32:57] literally is the ultimate level of [33:00] narcissism and delusion of grandeur to [33:03] think you can create God and that then [33:05] the god you create like you're saying [33:07] Bill is going to be so benevolent and [33:11] perfect that you create constructed the [33:12] perfect God that will give you your [33:14] pellet will give you your little [33:16] scenarian you know [33:17] >> I just would correct you I didn't say it [33:19] Daario said it [33:20] >> right but no to [laughter] but to your [33:21] point of like just taking them at their [33:23] word they actually believe that they can [33:25] create God and that they'll create a god [33:27] so good that it's better than humanity. [33:29] Saxs, your thoughts. [33:30] >> Well, I guess the question then is why [33:33] are they pushing for the let's call it [33:35] red capture agenda where [33:38] >> I know why. [33:38] >> Go ahead, Jamal. Go ahead. [33:41] >> That is very reductive game theory. So, [33:45] if you want to be unexploitable, I think [33:47] the best thing that you could do if [33:48] you're trying to build a super god is [33:50] have three or four entities in a room, [33:52] close the door behind you, and then [33:54] dominate those other three or four [33:56] entities, and then you set the rules. [33:58] And because your counterparty is unable [34:01] to track at the level of technical [34:04] capability that you would have, you [34:06] create this massive asymmetry that [34:08] allows you to exploit them. That's just [34:09] simple game theory optimization. And you [34:12] know what Bill said is so powerful. I've [34:14] read these things and it's laborious and [34:15] it takes time, but every time they put [34:17] these things out, just take the time to [34:19] read it. And what I have said before, [34:22] Bill, I don't know your point of view on [34:24] this, but I initially thought that this [34:26] was mostly game theory, that a lot of [34:30] their reactions I thought were less [34:32] rooted in their dogmatic beliefs and [34:35] more rooted in a GTO approach to either [34:38] raising capital or putting pressure on [34:40] competitors. Either way, both could be [34:43] true. What your framing is and my [34:45] framing, although mine's more tactical [34:47] than yours to be fair, because I've [34:49] always thought that these moves make [34:51] sense through that lens. How do you [34:53] absorb most of the capital? How then do [34:56] you make sure that you are in a position [34:58] to disproportionately affect the rules? [35:02] And how do you create an oversight body [35:05] that is less capable and intellectually [35:09] aware as you are about the actual [35:12] details because [35:13] >> the referees don't understand the game. [35:14] Right? [35:15] >> If the refs don't understand the game, [35:17] you'll run over the game. Yeah. [35:18] >> By the way, by the way, one thing they [35:20] have achieved by doing this is I think [35:23] that if you pulled the, let's just call [35:26] it the intellectual elites, so everyone [35:28] in the media and whatnot and the [35:30] professors and all those, and they were [35:33] to rank the different AI players by who [35:36] they think is most caring, I think [35:39] they'd probably put Anthropic first [35:41] because they've been out with the [35:43] doomerism talk. And so it's given them a [35:46] halo with the people that may matter to [35:49] what they want to accomplish. It it's [35:51] simultaneous [35:53] creating a lot of trouble like with the [35:56] data centers and whatnot. Like there's [35:57] there's negative ramifications. [35:58] >> What you're saying is so important [36:00] because on the one hand they create [36:02] empathy [36:03] and then they write these documents that [36:05] expose what they think and nobody [36:07] actually connects the dots. [36:08] >> Yeah. To steelman their position for a [36:10] second. I mean, I think probably the way [36:12] they think about it is that they are [36:14] creating something very powerful, [36:15] something godlike, and therefore it [36:18] needs to be safe [36:20] and that they care the most about that [36:23] out of everybody. Nobody else takes this [36:25] seriously. Remember that Enthropic was [36:27] basically a spin out of open AI and they [36:30] felt that Sam and the company leadership [36:33] weren't taking their point of view [36:34] seriously enough. [36:35] >> It was the most woke portion of Open AI. [36:39] >> We're steel Manning. So the most for so [36:42] they they see the power of it. They're [36:44] the ones who are concerned about safety [36:47] >> and they care the most and therefore [36:49] they're in the best position to do that. [36:51] >> Now I think the issue is just you can [36:55] see how this can lead to red capture, [36:57] right? Which is if you brand yourself as [37:00] the safe AI company and then try to [37:03] characterize everybody else as a [37:05] reckless player and reckless AI needs to [37:08] be stopped. you can see how this would [37:11] basically further your monopolistic [37:14] control over this industry. And if you [37:17] see AI through the lens that you know [37:20] that really frankly the pope and I see [37:21] it which is centralization versus [37:23] decentralization I do think that is you [37:26] know one of the key lenses we should [37:28] have on the technology is whether you [37:30] want this to be a centralized or [37:32] decentralized technology. This way of [37:34] viewing the world leads to more [37:36] centralization and I think that's [37:38] dangerous. I mean, if AI is this very [37:41] powerful technology, I think it needs to [37:44] be decentralized so that all of us can [37:46] protect ourselves to some degree, right? [37:48] We need to be able to run we need to be [37:50] able to run the AI ourselves on our own [37:53] hardware if we so choose, so we're not [37:56] beholden to a single company that might [37:59] be in bed with a deep state. [38:00] >> Let's say it very pointedly. If benefits [38:03] and compensation and economic support [38:07] were all of a sudden tied to some [38:09] algorithmic decision, this is a [38:11] dystopian episode of Black Mirror that [38:14] we're dealing with. And to your point, [38:15] Saxs, you want 100 or 1,000 or 100,000 [38:19] versions of what that answer is so that [38:21] there's actually a way to refute [38:24] a singular answer. A singular answer to [38:26] these kinds of questions, which is [38:28] effectively what some folks would want, [38:30] is incredibly dangerous. [38:32] >> And this is something that is in [38:34] control, I think, of humanity. I've been [38:37] talking about AI sovereignty here for a [38:39] bit just in terms of how much more cost [38:43] effective it is and how you're not [38:45] training other people's AIs with your [38:47] knowledge and your insights. This is why [38:49] it's super important that open- source, [38:51] open- source agents and local hardware [38:53] be able to run these models and that [38:56] consumers and companies learn how to [38:58] roll their own language models, how to [39:00] make a small language model, an SML, a [39:03] VSSML, a verticalized one and run it on [39:06] your Apple hardware because Apple [39:08] actually has taken a principled approach [39:10] historically to your sovereignty for [39:12] your data. Data sovereignty now is [39:14] privacy. [39:15] >> Yes. And now it's intelligence [39:16] sovereignty. the the intelligent [39:18] sovereignty is different than privacy. [39:19] Privacy is, oh, you can't see my photos. [39:22] You can't, you know, peek into my notes [39:25] app and what I wrote there in my [39:26] journal. Now, intelligent sovereignty is [39:29] you can't tell me what to think. You [39:31] can't use your AI to analyze my photos, [39:34] to analyze my emails, to analyze my [39:36] messages, and tell me how to interpret [39:38] the world. That's actually going to be [39:39] the next key piece here. This is why I [39:41] think Apple is just the dark horse in [39:44] this entire race. there is an [39:45] open-source product that can run on this [39:48] hardware, the M5s, the, you know, 48 [39:50] gigs, 128 gigs, the stu new Mac Studio [39:53] coming out with supposedly a terabyte. [39:55] That changes the whole game. And this is [39:58] so paradoxical, Bill, that our [40:00] adversary, the Chinese of all people, [40:04] the Communist Party is leading the [40:06] open-source movement and the United [40:08] States is centralizing. [40:10] >> They're leading the openweight movement. [40:11] It's not open source. Just the [40:14] distinction is important. [40:15] >> Yeah. Yeah. [40:16] >> Look, I Jacob, I agree with you about [40:18] the importance of open source because [40:20] open source means software freedom. You [40:22] can run the program yourself on your own [40:24] hardware. You don't have to share. You [40:26] don't have to give up your data [40:27] sovereignty. You don't have to give up [40:28] your privacy to again to some monopolist [40:31] who's going to be, you know, in bed with [40:33] the government or the deep state, right? [40:34] So that's the thing we're all afraid of. [40:36] And if that's the only AI that's [40:38] available is from the, you know, [40:39] monopoly or duopoly, [40:42] then your choices are to live off the [40:44] grid and not participate in the modern [40:45] economy or give up control, right, to [40:49] some social credit system. So I think [40:51] the open source is really important. And [40:53] by the way, that was Elon's instinct in [40:54] creating open AI. He was afraid that [40:56] Google was going to monopolize AI. So [40:59] he's like, let's create open AI so that [41:03] it's not dominated by a single company. [41:05] But that that is I think the right [41:07] answer here is I know people want to I [41:10] think their instinct to the idea of [41:12] powerful AI is to clamp down and just [41:14] control it. But actually you have to [41:17] have multiple players. That's the only [41:18] way you're going to be protected is is [41:20] to have multiple players. [41:21] >> This next wave of the market evolution I [41:25] think is going to be extremely high [41:27] stakes and messy. Nick, just throw this [41:28] up because I just want these guys to [41:30] react to it. So this is a company that I [41:33] just ran into on X called Rogo. And what [41:36] they did was they created a test bench [41:38] and a set of evals to be a financial [41:41] analyst essentially and tested all of [41:43] the frontier models. And it was so [41:46] interesting because they summarized I [41:48] read their paper that they published and [41:50] I quoted the most interesting part [41:51] because I see it everywhere now across [41:54] all EVALs which is this one phrase there [41:57] is no single best model anymore. at the [41:59] top of the leaderboard. Opus 47, GPT55, [42:02] Sonnet 46 appear almost [42:05] indistinguishable, [42:07] separated by less than, in this case, [42:09] you know, 3/10en of a percentage point [42:10] overall. Read superficially, the results [42:12] suggest convergence. Three frontier [42:14] systems reaching roughly the same level [42:16] of capability. Okay, why is that [42:18] interesting? Well, you got trillions of [42:20] dollars going into each of these guys to [42:22] trying to create these next superb [42:24] brain, [42:26] but increasingly our existing set of [42:28] evals and our existing capabilities [42:31] when applied on these models roughly [42:33] produce the same thing which [42:35] theoretically says that these things are [42:36] getting commoditized way too quickly. [42:37] And then you'd say, well, what's the ROI [42:39] on all this incremental spend, which is [42:41] a very interesting economic and [42:42] investment question. So I don't know [42:44] like gurley what do you think happens if [42:47] these evals continue to asmtote and we [42:51] need more and more and more money for [42:52] training [42:54] >> some of the smarter people in the open [42:56] source community have suggested to me [42:59] that we need more open-source connectors [43:02] of types so MCP uh is actually run by [43:06] the Linux Foundation and if you think [43:08] about any surface area where a model [43:11] might interact with other software the [43:13] more of those connectors that can be [43:16] open sourced and commoditized, it would [43:18] lower. This is what Google did with [43:20] Kubernetes uh to to try and commoditize [43:24] where workflows live off of AWS and to [43:28] make it easy to migrate. And so the more [43:31] you can create systems that make that [43:33] type of exchange you just described [43:35] super easy so that you can plug and play [43:37] the model and you have to worry about [43:39] things like context and how does context [43:42] come in and and data and you know stuff [43:44] that like glean and and data bricks do [43:47] but how anyway if you can do that if you [43:49] can create more of those connectors like [43:51] that then the models become swappable [43:54] and certainly with the both the model [43:56] companies trying to move up the stack [43:58] you have massive [43:59] desire from the app layer players to try [44:02] and figure this out and we already you [44:04] know watched what cursor is doing and [44:06] playing with their own model and being [44:07] forced to kind of reckon with the fact [44:10] that they're coming up the stack fast. [44:11] So I think that's a really good insight [44:14] that this gentleman shared with me and I [44:16] think we the founders and developers [44:19] that are out there should work on more [44:21] of these interfaces and throw them into [44:23] the open source world just to make it [44:25] more exchangeable, swappable. Is there [44:27] an issue right now with we don't have a [44:29] good harness for open source? I mean the [44:32] way that like claude is a harness for [44:36] >> Yeah. There's people making open-source [44:37] versions of this or building companies [44:39] around harnessing and building the [44:42] integrations into it. But open source is [44:44] always like the last to build the fit [44:46] and finish around the product. They [44:48] focus on the core of the product, right? [44:50] So like Linux for your desktop never [44:53] really took off because the interface [44:54] was never polished. The UI was never [44:56] like perfect, but there are companies [44:58] building that and I'll just I'll show [44:59] you one company that we invested in. [45:01] This is a company called Abacus and they [45:04] had a very simple idea. They came up [45:06] with their own hardware stack. They came [45:08] up with their own uh platform and now [45:10] they are sold out of these boxes that [45:13] they're building for insurance, [45:15] healthcare and everybody wants to run AI [45:18] inside their organization and then start [45:20] building their own models. We actually [45:23] incubated this in our incubator and you [45:25] can check it out goabacus.co. [45:28] They're just basically saying and [45:30] organizations cannot get enough of this [45:32] product. It is crazy how [45:36] savvy these organizations are getting [45:38] and Chamat you're doing it with 8090 as [45:39] well I think where they're just like we [45:41] have to build headless products so that [45:44] we don't get locked into [45:47] any one provider. Whenever we go into [45:49] the Fortune 1000, we never compete with [45:51] OpenAI or Anthropic, they'll have a [45:54] preference sometimes of what they want [45:55] to see under the hood. So, our control [45:57] plane can basically hot swap, as Bill [46:00] said, between one or the other. We've [46:02] also started to lay the seeds for open [46:04] source and open weights. [46:06] But the reason is because they don't [46:08] want to be tied into one of these [46:10] critical frontier labs. They want to be [46:12] able to ride the wave of innovation, but [46:14] they're afraid of two things. They're [46:16] afraid that one technology leaprogs the [46:18] other too quickly for them to [46:19] participate and they pick the wrong one. [46:21] And the second thing that they're [46:22] increasingly afraid of is terms of [46:24] service and being at the sake of a [46:28] frontier lab in a political philosophy [46:30] that they may be in the crosshairs of [46:32] accidentally. Right? So you're a [46:34] hospital system in Canada. You support [46:36] the euthanasia laws in Canada, but this [46:38] frontier model in America says, "No, [46:40] can't do it. So now we shut you off." [46:42] Right? That's an an example. I'm not [46:44] saying one is right or wrong. It's just [46:46] to illustrate the case. So a lot of the [46:48] folks that we see now in the fortune [46:50] 1000 and increasingly the global 1000. [46:53] They want as Gurley said abstraction [46:55] above it. They want to sit as Sach said [46:57] in a control plane. They want to see be [46:59] at this level and they want to have the [47:01] flexibility because they don't know how [47:03] it's going to shake out. They see all [47:04] the money being invested at the model [47:06] layers but they see the model quality [47:07] asmtote. So they're like, "Wait a [47:09] minute, what are we supposed to do just [47:10] from a risk perspective?" [47:12] >> And and regula regulated industries are [47:15] particularly sensitive to these kind of [47:17] issues you're bringing up. [47:18] >> Hugely hugely sensitive and regul. [47:20] >> So if you just follow what finance, [47:22] healthcare, you know, and and those kind [47:24] of folks are doing, they're just like [47:26] this has to be onrem and they're very [47:28] concerned about a data leak and they're [47:30] very concerned about HIPPA compliance. [47:31] They're very concerned about training a [47:33] model. Like what if you know all of a [47:36] sudden somebody does you know a query or [47:39] or writes a prompt and it pulls some [47:41] information from that Canadian [47:43] healthcare system and all of a sudden [47:45] somebody gets a result and that sounds [47:47] farical. Remember stable diffusion [47:51] built themselves on Getty on Getty [47:53] images [47:54] >> and they all of a sudden the Getty image [47:56] watermark was in the output like system [47:59] you see anthropic and open AI in all of [48:01] these Fortune 1000s at the developer [48:04] layer cuz most of the developers have [48:06] their own credit cards they're allowed [48:07] to sign up for them you eventually wrap [48:09] them in an enterprise license so it's a [48:12] typical PLG-led market motion like we [48:14] saw in Slack we've seen it everywhere [48:16] the interesting thing is not that but [48:18] it's the unwind that happens then when [48:20] you have these huge licenses you have [48:22] these huge buckets of spend you can't [48:24] really tick and tie it together the CEOs [48:27] then wake up and are told by the CFO hey [48:29] FYI here's where we are Uber was one [48:32] example a second I don't know Nick if [48:35] you have this tweet but from Vivecarpali [48:37] the founder of Clover yeah this was just [48:39] yesterday overheard from a fortune 20 [48:41] company CEO asked for a billion in AI [48:44] generated OPEX savings at the beginning [48:46] of the year so we're 6 months in the [48:48] team has spent $200 million on tokens [48:51] and with minimal results. [48:54] And so now they're in this weird motion [48:57] now where the CEO is pulling the budget [48:59] back and now you're having to cut the [49:00] licenses. You just saw Microsoft [49:02] announce that they're killing the claude [49:03] licenses. [49:05] It's a super dynamic market right now [49:07] and I don't think we know what the [49:09] terminal solution looks like. And by the [49:11] way, I would I would add Claude is [49:13] really good at product. Like Claude for [49:15] Excel is better than Copilot by not by a [49:19] little, by a lot. [49:20] >> And so, you know, anyone that's going to [49:23] run against them, they're they are a a [49:26] worthy foe, I should say. [49:28] >> Yeah. I think I think Claude is [49:29] exceptional, by the way. I mean, I use [49:31] it every day. Yesterday, I hit my token [49:33] limit on my pro plan. I had to put on my [49:35] credit card, spent [laughter] another [49:36] couple thousand bucks, and I'm like, I [49:38] was so angry, but I did it because it's [49:40] so good. [49:40] >> Yeah. [49:41] >> Yeah. Go ahead, Sax. Wrap us up here. [49:44] Yeah. [49:44] >> Yeah. So, well, just to wrap up, let me [49:46] just connect a couple ideas. So, one is [49:49] that in terms of the the red capture [49:52] agenda that you're seeing in Washington, [49:55] I think where it's all leading to is an [49:56] effort to ban open source models or open [50:00] weight models. [50:01] >> There's a lot of breadcrumbs leading [50:02] here. I think people who want this are [50:04] being a little bit circumspect. They [50:05] don't feel like they're quite there in [50:07] terms of being able to justify it yet. [50:09] >> Can you explain it? [50:10] >> Sure. You look at you look at a lot of [50:11] the rhetoric around how models need to [50:14] have guard rails and that with open [50:16] source models, the guardrails can be [50:18] removed and therefore they're dangerous. [50:20] You see this rhetoric already in [50:22] anthropics blog posts. So, you know, any [50:24] threat that they describe, they kind of [50:27] go out of their way to take that shot at [50:30] open source models. you saw it with [50:31] respect to cyber for example or with [50:34] respect to bio threats things like that [50:36] I mean I've seen that type of language [50:38] repeatedly that open models lack guard [50:41] rails or the guardrails can be taken off [50:43] and therefore it's a problem and I think [50:46] again they're trying to create ideas or [50:50] put predicate facts in the public record [50:53] to justify an action later on and I [50:57] think it's just a matter of time before [51:00] they feel like they're at a position [51:01] where maybe they can push for that type [51:04] of ban [snorts] [51:05] directly. They're not quite there yet. [51:07] >> But what does that do then to the rest [51:09] of the market? Like let's just say [51:10] America bans open source and open [51:12] weight. Okay. Well, what about the rest [51:14] of the world? I mean, [51:16] >> it sure it'll put [51:17] >> they're going to leap frog us. [51:19] >> Sure. You'll put the US on an island. [51:20] Well, first of all, as we all know, what [51:22] does it mean to ban a openweight model? [51:24] It's a file. It's a bunch of numbers, [51:26] you know, that you can run on your your [51:28] laptop. [51:29] >> Yeah. But what it will do is you think [51:32] about like all the cloud service [51:34] providers who run open models like they [51:36] will stop doing that because they got to [51:38] comply with the law and so all this [51:40] infrastructure that's been built up it [51:43] will get much harder to use open models [51:45] in the United States now the rest of the [51:47] world will continue to benefit from them [51:49] because there's a tremendous benefit in [51:51] terms of cost and customization and [51:54] control that you get with an open model [51:56] >> and we're on a completely different [51:58] price curve. And we haven't talked about [52:00] this yet. There was an economic and [52:02] capital mode to training that is going [52:05] away. It's going away in two ways. One [52:07] is because we're getting these domain [52:08] specific architectures at the silicon [52:10] layer. And then second, we're rebuilding [52:12] all of the core components. I don't know [52:14] if you guys saw yesterday, but Elon was [52:16] like, we've rewritten the entire [52:18] training complex in C and it's an order [52:20] of magnitude increase and we can run it [52:22] on 220,000 GPUs. at the scale of what [52:25] they're trying to do. Those kinds of [52:27] innovations are going to make the cost [52:29] of model training so much cheaper that [52:32] it's like, why would we stick to the $10 [52:35] billion training runs when we can have [52:37] the $10 million training runs? [52:38] >> Well, if it got 1% better, just as a [52:41] thought experiment, Nick, could you find [52:43] Elon? [52:45] If okay, if it got 1% better, that's the [52:48] equivalent of 2,000 GPUs, which is the [52:51] equivalent of hundreds of millions of [52:53] dollars in compute. So every 1% equals [52:56] hundreds of millions in compute. If he [52:59] gets 10% 20% more efficient every [53:01] quarter, every [53:02] >> look at this speed improvement, the [53:05] speed improvement versus jacks for for [53:08] training runs is now an order of [53:09] magnitude. When you think about then the [53:11] capex buildout, the opex, the power, [53:16] the cabling, the copper, all of it. [53:21] And now this is a closed source model, [53:23] but I'm pretty sure that just that tweet [53:26] is going to get read by enough people [53:28] where there's going to be five or six [53:29] open- source stacks for training that [53:32] are rebuilt closest to the bare metal as [53:35] possible. [53:35] >> Yeah. [53:36] >> Why wouldn't you do that now? And so to [53:38] your point, Sax, cutting that off so [53:40] that we lose that kind of innovation [53:41] makes no sense to me. [53:43] >> I agree. And and like I said, I don't [53:45] know that the forces who want to ban [53:46] open source are strong enough or have [53:49] made the case or created the predicate [53:52] facts necessary yet to ban open source, [53:55] but I do think it is on the agenda and [53:57] it's where all the breadcrumb trails are [53:59] leading. So just watch out for that. I I [54:01] agree totally with with what David just [54:04] said and I wrote a blog post recently on [54:06] open source and and made the exact same [54:09] point. [54:09] >> I read that too. That was a good one too [54:12] on above the crowd. [54:13] >> No, it's not. [54:17] It was on the Santa Fe Institute. [54:19] [laughter] [54:20] >> Oh, no, it's it's the P3 Institute which [54:23] is my new my new institute. Anyway, um I [54:26] same same exact conclusion which is rest [54:29] of the world ends up running on Chinese [54:32] models if if if they're able to succeed [54:35] at what you just said. [54:36] >> And if you want to know the canary in [54:38] the coal mine sacks obviously the place [54:40] they love regulation most is the EU. So [54:43] EU has already done volley after volley [54:46] of proposed regulation for AI and open [54:50] uh and open source is particularly in [54:52] the crosshairs there because nobody's in [54:54] charge of it. So are you going to get a [54:56] bunch of open source contributors having [54:58] to vet their model with the EU [55:00] regulators like that's obviously not [55:02] going to happen. Nobody's in charge of [55:03] it. There just a bunch of contributors. [55:06] But open source is the solution I think [55:08] to [55:09] >> Yes, I agree. [55:10] >> It is the back stop. It is the backs [55:12] stop. I mean, unless you want to live [55:13] off the grid. I mean, if you want to [55:14] participate in the modern economy, it is [55:16] the backs stop. And let me just make one [55:18] other final point. It kind of maybe [55:19] leads into our next topic is I do think [55:21] that there is the potential for the [55:24] monopolization of this market to a [55:26] greater degree than people may be [55:28] pricing in right now. First of all, [55:30] we've seen that every other major tech [55:31] category has led to a monopoly or [55:33] duopoly situation. Seems to be the the [55:36] way that these things work out. But also [55:39] if you look at the growth rates right [55:41] now, Enthropic does seem to be pulling [55:43] away. There's a article in the [55:46] information showing the latest numbers [55:47] where I think Anthropic's now at they [55:49] seem to have pulled away from open AI, [55:51] which is not surprising and something I [55:53] I predicted. Look, if you have one [55:54] company that's growing at 10x [55:56] year-over-year and another company [55:58] that's growing at 3x year-over-year, [56:00] within 2 years, the first company will [56:03] have 90% market share. This is the power [56:05] of compounding, right? is just do the [56:07] math on it. 10 * 10 is 100. 3 * 3 is 9. [56:12] So again, if you just are able to [56:13] outgrow your competitor at that rate for [56:15] 2 years, you will achieve monopoly [56:18] market share. Now there are reasons to [56:20] believe that anthropic cannot continue [56:22] that growth rate for 2 years. There's [56:24] going to be a competitive response. It's [56:26] already happened. Also, there may not be [56:28] enough compute to support that kind of [56:30] growth. There may be physical [56:31] constraints, but you'd always rather be [56:34] the company that has that inertia that's [56:36] on that totally trajectory than the one [56:38] that has to do something different to [56:41] then knock that leader off its current [56:43] trajectory. [56:44] >> Did you guys see what just hit the wire? [56:45] Nick, can you throw it up from Poly [56:47] Market? This is insanity. Poly Market [56:49] puts out there that an AI consultant [56:52] revealed that one of their clients [56:54] accidentally spent half a billion [56:56] dollars in [laughter] a single month [56:58] after failing to set employee limits on [57:00] clock usage. [laughter] [57:03] >> What? [57:03] >> Oh my god, look at this. Look, the 16.6 [57:08] million per day, almost 700,000 per [57:11] hour. Oh my god. Well, there seems to be [57:15] there seems to be a new like meme taking [57:17] shape that somehow like all this token [57:19] spend is is wasteful and basically [57:20] useless. And you know, we're constantly [57:23] oscillating between narratives like AI [57:25] is going to put everyone out of work to [57:27] like AI is useless and it's a bubble. [57:29] The doomers can't seem to make up their [57:30] minds whether AI is going to be our new [57:32] god or whether it's basically a total [57:34] waste of money and it's going to lead to [57:36] a bust. But in any event, yeah, I think [57:38] you know the there there's no question [57:40] that token efficiency is going to be a [57:41] big theme over the next year because the [57:44] spend has been ramping up way faster [57:46] than enterprise customers thought and [57:48] there's going to be a drive for [57:50] efficiency. Does that fundamentally [57:51] change the dynamics? I don't think so. [57:55] But it it might, you know, it might [57:56] temper the growth to some degree. Well, [57:58] and they've done a tremendous job making [58:01] people believe that tokens are free by [58:04] giving them these crazy deals like $20 a [58:06] month, you can do whatever you want. [58:07] $200 a month, you can do whatever you [58:08] want. And it's like everybody's leaving [58:10] the hose on, everybody's watering, and [58:13] then [58:13] >> you get a photo that says you've hit [58:15] your usage and it's like, "Come back at [58:16] 230." I'm like, "230? It's 10:30. I [58:19] can't do anything between 10:30 and [58:20] 2:30." And then it says, "Well, you can [58:22] put in your credit card." And so I did. [58:24] >> Yeah. But I mean it's it's literally [58:26] like the first the first 10,000 gallons [58:30] of water are free basically and then all [58:32] of a sudden it's like okay it's a penny [58:34] a gallon and then everybody in the [58:36] organization and this has literally [58:38] happened in our organization. One person [58:40] built like an interface for the founder [58:42] university program. Another person built [58:44] one. Then another person was like, [58:45] "Well, those two people got credit at [58:46] the management team meeting, so I'm [58:48] going to build an interface." And the [58:49] next person builds an interface. And [58:50] then everybody shipping like interfaces [58:53] and I literally had three different [58:55] people on the team make three different [58:56] versions of like a founder university [58:58] portal and I'm like, "We don't need [59:00] three. Can we get coordinated here?" And [59:02] it didn't get to the point of like [59:04] spending thousands of dollars, but it [59:05] certainly got to the point of spending [59:07] hundreds of dollars and it would have [59:08] gotten to tens of thousands. [59:10] >> Are we still on the first topic? What [59:11] are we doing? Well, no, we kind of [59:12] merged like two or three of them [59:13] together. [59:14] >> Oh, we did? Okay. [59:15] >> And it's super interesting. Trust me, [59:16] >> it's super interesting. I think what [59:17] Gurley said is one of the most [59:19] interesting things I have heard [59:22] >> in a long time. [59:23] >> Take people by their word. And if you [59:26] read their words, [59:27] >> if you just read their words and you can [59:30] understand what they're saying, you [59:32] don't have to guess about why they want [59:34] to have a digital guide. Well, now I'm [59:37] not the sharpest I'm not the sharpest [59:39] arrow in the quiver, but I can take down [59:41] a buck. And I can [laughter] tell you [59:43] that this don't make a lot of sense to [59:45] me. Even the dullest arrow can take a [59:48] buck down. [59:49] >> All right, let's get back to [59:52] It's so great having you here, Bill. We [59:53] missed you. [59:54] >> I [laughter] got you. I got you. [59:55] >> We missed you, brother. [59:57] >> We're going to transition to the next [59:58] topic. There is some evidence that [60:00] Daario is mitigating his dumer rhetoric. [60:03] Did you see this? [60:04] >> Let me get to it. Yeah. Yeah, I got to [60:06] it here. All right, we we're going to [60:07] have to talk for the 16th time in the [60:10] last 18 months about AI's impact on [60:12] labor because again this chaotic [60:15] schizophrenic [60:17] interpretation of the data continues. [60:19] Cloudfare as we talked about last week, [60:21] shout out Matt Prince [60:24] Shimath's favorite CEO of the year. [60:27] Letter of the year [60:28] >> letter of the year. He cut 20% [60:30] >> award for the letter of the year [60:32] >> making friends every week. [60:34] here on the program. So they both blamed [60:37] AI spec explicitly and specifically and [60:40] Zuck then paired his 8,000 [60:45] cuts at meta with the fact that he has [60:48] put uh spyw wear on everybody's laptop [60:51] to study every employee to make their [60:54] training data better. That got leaked [60:56] and people thought, hey, that's a Black [60:58] Mirror episode. We're we're working at [60:59] Meta in order to, you know, get our [61:02] two-year severance package. But on the [61:05] other side of the table, Goldman Sachs's [61:07] uh CEO, David Solomon, wrote an op-ed in [61:11] the New York Times. I'm the CEO of [61:13] Goldman Sachs. Period. The AI job [61:16] apocalypse is overblown. Period. [61:18] Obviously, he might be fighting for that [61:21] anthropic or open AI IPO in the coming [61:24] months, or maybe is doing it right now. [61:26] He made three points. AI won't eliminate [61:29] 25% of jobs. It's going to automate 25% [61:32] of work hours and workers will fill that [61:34] time with higher level tasks. Obviously, [61:36] that didn't happen in the case of [61:37] Zuckerberg's layoffs. Just because a job [61:40] can be replaced doesn't mean it will be. [61:41] Bank tellers increased after ATMs. Live [61:45] entertainment became more popular after [61:46] TV. And the US labor market creates and [61:49] destroys 25 to 35 million jobs annually. [61:52] And the gross churn dwarfs net losses. [61:56] New categories like agentic AI [61:57] management are already hiring yada yada [62:00] yada. Uh a publication called Fortune is [62:02] apparently still publishing AI slop and [62:04] they say both Sam Wman and Daario have [62:06] walked back their AI job apocalypse [62:10] predictions as they gear up for an IPO [62:13] sax have at it. You know you've been [62:15] saying uh and your prediction was you [62:19] took the other side hey we're going to [62:20] create more jobs. There was a a recent [62:24] one of the job boards put out some stats [62:26] that the number of software jobs is [62:27] going up, the number of listings of [62:29] other jobs going down. So, I guess [62:31] you're probably in the camp of creative [62:33] destruction and churn at this point, [62:35] Sax. [62:37] >> Well, I mean, I think you should be [62:38] giving me more credit than that cuz my [62:40] most contrarian take back in January on [62:43] our prediction show is that AI would [62:46] lead to job gains, not job loss. And [62:49] over the past week, you've now seen the [62:52] narrative shift, I would say, almost [62:54] completely towards that position. So, [62:56] you have the CEO of Goldman Sachs right [62:57] in this in the New York Times. You know, [63:00] I don't think he'd be doing that if he [63:01] felt like he was completely stepping out [63:03] on a limb. Maybe even more importantly, [63:05] you had Sam and even Daario now walking [63:09] back their claims of massive job loss. [63:12] And they explained why Daario said, it's [63:15] kind of like the 25% of work hours [63:17] thing. He said that AI might automate [63:19] away 90% of someone's task, but the [63:22] other 10% will expand to do a whole [63:25] bunch of new new tasks and new things, [63:28] which is very similar to the the types [63:30] of of arguments that people like me have [63:33] been saying and actually that Jensen's [63:35] been saying that just because you [63:36] automate away some task doesn't mean [63:39] that you automate away the purpose of a [63:40] job. But now the worker is freed up to [63:43] do new things, to do the higher [63:45] complexity tasks that David Solomon, the [63:47] Goldman CEO, is talking about. So the [63:50] fact that Daario is now walking this [63:52] back and coming around to my position, [63:57] I think that that's kind of amazing. And [64:00] uh where do I go to get my apology? You [64:02] know, [64:02] >> well, we're going to have an ap [64:04] [laughter] we're going to have an [64:05] official apology form that you can fill [64:07] out. It's got check boxes. I was wrong. [64:10] I mean, some mornings I woke up [64:12] thinking, why am I going out defending [64:14] these guys? You know, these idiots. I [64:16] mean, they're scaring the public with [64:19] all these dire predictions about an [64:21] apocalyptic future. There was no data to [64:24] support that. I mean, we can all debate [64:25] what's going to happen in the future, [64:26] and we probably should be humble about [64:28] what is going to happen in the future [64:30] because we don't completely know, and [64:31] this industry is very dynamic. But you [64:34] have to look at what is the data that we [64:35] have so far in the current situation. [64:38] And we do not see data that supports [64:40] massive job loss. You can cite this [64:42] layoff or that layoff Jcal those are [64:44] anecdotes and the plural of anecdotes is [64:46] not data. If you look at the actual data [64:48] like Yale Budget Lab did they said no [64:50] discernable disruption in the labor [64:52] market in the last 3 years due to AI [64:56] they've done a comprehensive study. You [64:58] look at job postings for software [64:59] engineers. It's up 15% year-over-year. [65:02] Their job postings for software [65:04] developers have hit a new three-year [65:07] high despite the fact that coding is the [65:11] single breakout use case of AI this [65:12] year. So if AI has not caused job [65:16] elimination for software developers, [65:18] what category has it caused? I mean code [65:21] is now the number one use case I think [65:24] of AI in the enterprise. [65:26] >> Okay, [65:26] >> let's be honest. Over the last five or [65:28] 10 years, a lot of companies overhired. [65:31] They mishhired. These CEOs did not have [65:34] a good handle on it. Their opex budgets [65:37] completely got bloated, inflated, [65:41] and they need to sort of get back to [65:43] where they were, get back to a fighting [65:45] weight. And it's this old adage of never [65:49] >> never waste a crisis. [65:50] >> Never let a good crisis go to waste. [65:52] Exactly. And so they point to this [65:53] thing. It's very simple to say. It's AI. [65:56] It's two letters. And say we're going to [65:57] fire people. But underneath that is not [66:00] AI because we know this. It hasn't done [66:02] anything measurable yet at the end [66:04] consumption of these tokens. Nobody is [66:07] standing there and saying look at my [66:09] filing [66:10] here is the lift that I have gotten. [66:12] Nobody has said that yet. That's very [66:14] important to observe. And so instead [66:16] what people are doing is realizing okay [66:18] I have this cover now to go and clean up [66:20] what was very poor management and [66:23] mismanagement over the last 5 and 10 [66:25] years where I overhired and I mishhired. [66:28] That's what's happening today. [66:29] >> Okay, Bill Gurley, I'm going to let you [66:31] chime in here. You've got two besties [66:33] saying, "Hey, this is all hogwash. It's [66:36] AI washing. These jobs were just, you [66:39] know, the strategy obviously in Silicon [66:41] Valley was [66:41] >> they need a scapegoat. They need a [66:43] scapegoat. [66:43] >> They're hired two years ahead of time. [66:45] Build for the future and it was a vanity [66:47] metric and you were blocking talent from [66:49] working on other startups or [66:50] competitors. The Google strategy." [66:52] >> Hold on. Wait, wait, wait. You just said [66:54] the critical thing. That is exactly why [66:55] they did it. [66:56] >> Yes. That was the explicit strategy from [66:58] >> the actual strategy. These guys were a [67:00] wash in cash. And so part of it is you [67:02] were just hoarding talent or what you [67:04] thought was talent. [67:06] >> Yes. And just keeping them off the [67:07] market. [67:07] >> And now you're jettisoning it because [67:09] the reality is as companies get bigger, [67:11] their growth rates monotonically [67:12] decrease and you get to like a GDP plus [67:14] some number and your valuation [67:16] frameworks change and there's nothing [67:18] you can do to fight that law of gravity [67:20] in the public markets. And so as each of [67:22] these CEOs who at some point thought [67:24] they were different and the rules didn't [67:26] apply to them are now realizing you're [67:29] just like everybody else. Okay, we have [67:31] to stay humble as Sax said, but Bill [67:33] Gurley, would you like to apologize for [67:34] Sachs andor give him credit for his [67:36] incredible non-conensus? [67:38] >> He wasn't the Hold on. He wasn't the one [67:40] promoting the jobs apocalypse. [67:42] >> It was [67:44] you. [laughter] I will give my thoughts [67:46] in a moment. [67:49] You're the for the mainream media. [67:51] >> I'll give mine. [67:54] >> You always represent the legacy media on [67:55] our show. Jay Cal, you have been in the [67:57] fourth. [67:57] >> I represent the legacy media represent. [68:00] You're the New York bluehaired. [68:03] >> I'm just giving you the statistics, [68:05] guys. I'm just presenting the numbers. [68:07] Now, let's remember anecdotes. [68:09] >> Let's remember. [68:09] >> Actually, let me give you an important [68:10] statistic. Let me give you a very No, [68:12] no. This is really important. [68:14] >> We have to let Bill Girly comment. Then [68:15] you can really important. Do you use [68:18] ketamine? I don't use ketamine. That's [68:20] the terrible drug. Do not use ketamine, [68:22] folks. [68:23] >> Bill Gurley, you have the floor. [68:26] >> I would just touch on two things that I [68:28] already said earlier. One, you know, [68:30] historically innovation has led to more [68:33] prosperity for humans. And I gave those [68:36] numbers from 1891 to today. I see no [68:39] reason why that won't happen here. In [68:41] the short run, from a bottomup [68:43] perspective, every human that wants to [68:45] protect themselves needs to be the most [68:47] AI enabled version of themselves they [68:50] can be. And the people that might be a [68:52] threat of job loss are someone who like [68:55] stands hard, fast, and refuses to use [68:58] AI. And I would just say that's simply [69:02] like saying, "I'm not going to use [69:03] email. I'm not going to use a [69:05] spreadsheet. I'm not going to use a [69:06] computer." And and you know, you [69:09] probably are at risk. [69:10] >> Yeah. Yeah. The paradigm will shift to [69:12] give you actually my position which is [69:15] >> Would you like me to give my position or [69:16] just want to jump? [69:17] >> Yeah, I do but I I never got to finish [69:18] that point. So, but I can do it after [69:20] you. [69:20] >> Yeah. Yeah. So, I I will give my [69:22] position on this which is there and it's [69:25] always been the same which is there's [69:26] going to be a massive job displacement [69:28] that occurs and that massive job [69:30] displacement is going to come because [69:33] CEOs in many cases believe that this [69:36] technology is going to make people more [69:38] efficient. they can do more with less [69:40] and they will be rewarded by the public [69:43] market by just having higher earnings [69:45] and we see that for every single [69:46] company. Now I fully concur it was [69:49] because of bloating and I gave my [69:50] position there. I know specifically that [69:52] Sergey and Larry took that strategy of [69:54] taking talent off the market so there [69:56] wasn't a Google competitor that was [69:58] literally explained to me by those [70:00] individuals. We hire people and then we [70:02] figure out what to do with them later. [70:04] That strategy permit just became the [70:07] standard in Silicon Valley and now it's [70:09] being reversed. [70:11] Now there will be wholesale jobs that [70:13] will be retired. If you look at [70:16] self-driving that's obviously happening [70:18] with Whimo with 3,000 vehicles and and [70:20] there'll be many more on the roads. That [70:22] job will be eliminated. We will be [70:23] sitting here in but 5 10 years and the [70:26] idea of somebody driving a taxi is going [70:29] to seem silly and dangerous. We will see [70:32] the same exact thing happen with [70:33] Optimist. You may have seen the figure [70:35] robot sorting packages. All those [70:37] sorting jobs at Amazon factories are [70:39] going away. Amazon themselves, these are [70:42] the savviest people in the world said, [70:44] "We are going to eliminate 600,000 [70:46] future positions and we are going to cut [70:49] positions." And Andy Jasse said, "This [70:51] is going to be a reoccurring theme. As [70:53] we deploy AI, we will do more with [70:55] less." You will see the headcount at all [70:57] these big companies dramatically [70:59] decrease or stay the same as earnings [71:01] massively increase. And you can take the [71:03] position, Saxs, that oh my god, the [71:05] numbers are in my favor. They're not. [71:07] The numbers are in my favor. The job [71:09] loss is tremendous. And there are [71:11] numbers associated with that. 8,000 [71:13] people at Meta after 20,000 before that. [71:16] And if you look at the steady state of [71:17] these companies, they has nothing to do [71:19] with AI. Let me finish. [71:21] >> They overhired. [71:22] >> No, no, no. We are beyond that. We are [71:24] beyond that. They are now getting rid of [71:25] people. When they say they're getting [71:27] rid of measurers, you can take them at [71:29] their word. When they say they're [71:30] getting rid of middle managers, you can [71:32] take them on their way and [71:33] >> scapegoating. That is you've given your [71:35] position already. I'm giving mine. My [71:36] position is they are obsessed with this [71:39] technology, they're obsessed with [71:40] earnings and they will continue that. [71:41] Now on the other side of the ledger, I [71:43] believe we'll have a Cambrian explosion [71:45] in startups and all these this talent if [71:48] they embrace the tools to Bill Gurley's [71:49] point are going to be able to solve more [71:51] problems and create small companies of [71:53] five or 10 people who are laid off from [71:55] Amazon or Meta and make double their [71:59] salary or have a better job that they [72:01] control. I believe that is going to be [72:02] the ultimate solution. But that [72:03] transition is going to be extremely [72:05] painful and we should have some humility [72:07] on this podcast for the people [72:09] impacted. Every cab driver is losing [72:11] their job. Every truck driver is losing [72:13] their job in the next 10 years. Anybody [72:15] sorting packages losing a job. Now you [72:17] can say all you want. You can say all [72:19] No, let me finish my thought. You can [72:21] say all you want, Chimov, that those [72:23] people don't want those jobs. But they [72:25] may need those jobs [72:27] >> and you are an elitist by definition. We [72:29] are all elitists on this program. We are [72:31] elite performers. [clears throat] [72:33] >> And these people are gonna lose their [72:34] jobs and they may not get a job very [72:36] quickly. By being able to call something [72:39] what we think it is [72:41] is not being elitist. It's actually [72:43] telling the truth. Meta over hired. [72:47] Okay? You could have stopped the company [72:49] at 3,000 people when I left and it would [72:51] not have changed the outcome of that [72:52] company. There was no need to go to [72:54] 90,000 people and burn $50 billion on [72:56] VR. They did it because they had the [72:58] freedom to do it. That's allowed. It's [73:00] capitalism. Okay? They're coming back to [73:03] realize that there's a more efficient [73:05] version of what they are. That has [73:07] nothing to do with AI. That's the only [73:09] point I'm trying to make. All you have [73:11] to do is just say that. [73:13] >> I think you're wrong. And let me explain [73:15] to you why you're wrong. I believe [73:16] you're wrong. I believe Zuckerberg is [73:19] putting that software on people's [73:20] computers in order to to find more jobs [73:24] to eliminate to increase it. And the [73:26] surface area of problems in the world is [73:28] not decreasing, but what is uh [73:31] decreasing is the number of humans to [73:33] take on the next opportunity. And that's [73:36] going to continue. And I think the [73:37] companies that will be rewarded and [73:38] their stock prices will be rewarded are [73:40] the ones who do much more with much [73:42] less. And they're going to keep [73:45] eliminating these jobs. And I take them [73:47] at their word. [73:48] >> You don't have to explain everything [73:49] with conspiracy. Maybe they just [73:51] mismanaged for a period and they could [73:54] agreed on that. I I think that explains [73:57] the postcoid two or three years. I think [73:59] what we're seeing this year is actually [74:01] the tools working. the tools are working [74:04] and there are jobs that are no longer [74:06] needed. The measurers as Matthew Prince [74:08] pointed out or product managers or [74:10] designers, those have all been [74:12] consolidated into one job. Somebody who [74:14] ships a product and it doesn't require [74:17] 12 people. It requires two people now. I [74:19] know [74:19] >> I don't think that's been consolidated. [74:21] I see it in Fortune 1000 companies all [74:23] the time. I don't think what you're [74:24] saying adopters. You're talking about [74:25] the slowest adopters. I'm on the front [74:27] line with [74:28] >> startup. These are where all the jobs [74:30] are. But I'm sorry, but a startup is not [74:31] going to go and enter a regulated market [74:33] and put JP Morgan out of business. Not [74:35] gonna happen. [74:36] >> They will eventually uh displace those [74:39] companies. It happens all the time. [74:41] >> Not going to happen. [74:42] >> We're going to agree to disagree. [74:43] >> Good luck to the startup trying to [74:44] disrupt Boeing. Good luck. I'm going to [74:46] take Boeing. [74:48] >> Okay. Well, some people might take [74:49] SpaceX. So, [74:50] >> good luck making drugs out of an Excel [74:53] spreadsheet. I'll take the regulated [74:54] pharma company. Good luck. [74:56] >> Sure. Listen, there are some industries [74:58] that are so much to one and you're going [75:00] to show up at the FDA and like, okay, [75:02] where's your team? Oh, it's just me. I [75:03] do it all. [75:05] >> Me and my model. Look at this. [75:06] >> You joke. Somebody just did that in [75:08] >> It's not a joke. It's not a joke. And [75:09] it's not going to happen because that's [75:11] not the way society wants safety, [75:13] predictability, governance, auditability [75:15] to work. [75:16] >> Yeah, I there's a distinct difference [75:19] between, you know, drugs and software [75:21] and services in the world. I think we [75:23] can agree on that. And listen, a regula [75:25] with truck driving is one of the most [75:27] regulated industries out there. So is [75:28] cab driving and taxis as Bill and I well [75:30] know and those jobs are being [75:33] eliminated. Bill, I'm gonna give you the [75:34] final word, then Sax, I'll give you the [75:35] final word. [75:36] >> A chance to respond. [75:37] >> Let's do Sax. [75:38] >> Okay, Sax, then Bill, go. [75:39] >> Well, first of all, Jake, you remind me [75:41] of the Troskyite who when confronted [75:44] with the fact that none of Trosky's [75:45] predictions had come true that simply [75:48] proved how far-sighted Trosky was. [75:50] >> I didn't go to graduate school. You're [75:52] going to need another reference. [75:53] >> In other words, none of your predictions [75:55] about job loss have come true. In fact, [75:56] the data [75:57] >> none zero data [75:59] >> except for what M just did last week. [76:01] But go ahead. [76:01] >> That's an anecdote. That is not [76:03] >> It's not an You're calling 8,000 people [76:06] losing their jobs an anecdote. [76:07] >> You don't hear yourself? [76:09] >> Hold on. [76:10] >> Do you hear yourself? It's not an [76:11] anecdote. 8,000 people lost their jobs. [76:13] >> Can I make my case? I heard you about [76:15] the the meta data point. First of all, [76:18] those jobs that job loss was not [76:20] directly attributable to AI. It just [76:22] wasn't. That's something you've invented [76:24] and put in the data. [76:25] >> Something Zuckerberg said. No, they they [76:27] clarified that. Okay. Okay. Sure. [76:29] >> He said it was related to they were [76:31] trying to balance additional spending [76:33] capex, but it was not directly related [76:35] to AI. But even if it were even if it [76:37] were 100% the case that was due to AI, [76:40] you're not netting those jobs against [76:42] all the other jobs that are being [76:43] created because of AI and all the new [76:46] companies that are being created right [76:48] now because of AI. So, you're just [76:50] cherrypicking one statistic. You're [76:53] attributing 100% of that to AI and then [76:55] you're not basically netting it and pre [76:57] presenting a balance. You've got [76:59] >> I'm not cherrypicking it. I am reading [77:01] the news and I'm describing what the CEO [77:04] said. Jack at Block said he's doing this [77:07] because of AI. Matthew Prince said it's [77:09] AI. Zuckerberg said it's AI. [77:12] >> I'm just taking them on their word. [77:14] >> Yes. Exactly. So Jack Dorsey came out [77:16] and said that he was going to do a 50% [77:18] elimination because of AI. Okay. And [77:21] within 24 hours, all the financial [77:24] analysts on X said that Jack was AI [77:26] washing and that block had horribly [77:29] overstaffed during COVID. It was running [77:31] much more inefficiently than all of its [77:32] other peers in this category. And [77:35] they've needed to do a 50% job cut for a [77:37] long time. So pretty much everyone [77:39] thought that was pure AI washing. In [77:41] fact, you've just proven my point. And [77:44] what exactly are the efficiencies that [77:46] Jack is getting? I mean, this is the [77:47] most handwavy thing ever that, oh, we're [77:49] just magically going to be able to [77:50] eliminate half our cost structure right [77:52] now. [77:53] >> Okay. So, Jack, Matthew Prince, [77:55] Zuckerberg, and Andy Jasse are all lying [77:58] and doing awashing. [78:01] This was due to AI. He just did it. [78:03] >> That's that's your reading of it. But [78:05] like I said, even if you attribute those [78:07] specific job losses to AI, which is [78:10] questionable, you're not netting it [78:11] against all the job creation that's [78:13] happening and also the new company [78:17] creation. Furthermore, let me just give [78:19] you some I specifically attributed that [78:22] the future and the new jobs will come [78:23] from startups. So don't misrepresent my [78:24] point. Thank you. [78:25] >> Okay, we currently have a 4.3% [78:28] unemployment rate in the economy. [78:30] Economists consider 5% to be full [78:33] employment. So basically unemployment is [78:35] at or near record lows right now despite [78:38] the of our lifetime despite the fact [78:40] that we're over three years into this AI [78:42] wave. Second and again this is the point [78:45] I wanted to make earlier with respect to [78:47] coding. Coding is the single job [78:49] category most impacted by AI right now. [78:52] We are at the point where AI is writing [78:54] most of the code. We have almost [78:56] complete automation of codew writing. [78:59] You would think that if you could look [79:02] at this in a simple Malthusian way, all [79:04] the software developers would be getting [79:06] laid off right now. Is that happening? [79:08] No. No. Software developers are not [79:10] being laid off on net. In fact, job [79:13] postings, job wrecks for software [79:15] developers are at a three-year high, [79:18] growing 15% year-over-year. Now, why is [79:21] this? I think the explanation is really, [79:23] really important. Okay, you look at code [79:26] commits on GitHub, which is the leading [79:27] code repository. There were 1 billion [79:30] code commits last year. In the past [79:33] month, there's been 1.1 billion. So, in [79:35] other words, [79:36] >> make something easier, more people do [79:39] it, [79:39] >> right? We have basically a 14x [79:42] year-over-year increase in code [79:43] generation. That code has to be managed [79:45] by somebody. You still need humans to [79:48] look under the hood. And when the amount [79:50] of code explodes and you get 10x or 100x [79:53] more code, the complexity also rises as [79:56] well. So look, we're not hiring 10 times [79:58] more engineers, but you do need more [80:00] engineers now to manage all of that [80:03] code. The other thing that's happening [80:05] is that there's been an explosion of the [80:08] use of code across the economy by [80:10] different businesses, different [80:11] applications, and different use cases. [80:13] I'm hearing from people who are now [80:15] hiring software engineers who never [80:17] would have hired them before. I was [80:18] talking to a fund manager, and he said [80:20] that his next two hires were not going [80:22] to be data analysts. they were going to [80:24] be software developers because they're [80:25] now deploying code for the first time in [80:28] ways that they were not before. This [80:30] goes back to my point about claude [80:32] proficiency being the most marketable [80:34] skill right now in the economy. People [80:37] are using these tools in entirely new [80:39] ways. I think that we're at the outset [80:42] of a boom right now caused by bespoke [80:45] software proliferating throughout the [80:47] economy and being used by firms that [80:49] never thought of themselves as tech [80:51] firms before. All of which is leading to [80:54] more productivity and that leads to a [80:56] healthier economy and that leads to more [80:57] job creation. And you're seeing that [80:59] again in the aggregate numbers and that [81:01] doesn't even include the bluecollar boom [81:04] that's happening right now with the [81:06] development of all this infrastructure, [81:07] the data centers and the new energy and [81:09] power generation. We are seeing hundreds [81:11] of thousands of new construction jobs [81:13] being created among bluecollars. Jan, [81:16] I'm sure you don't want them losing [81:17] their jobs by turning this boom off. So [81:20] again, [81:20] >> no I I never advocated you misconrue you [81:24] like to misconrue my position. I am very [81:26] clear there's job displacement going on [81:29] and the job displacement is related to [81:30] AI but net I do think the economy will [81:32] grow. Bill Gurley [81:34] >> maybe at some point in the future you'll [81:36] be right like the Trosky eye communism [81:38] has never been tried. Maybe it'll work. [81:40] >> Nobody knows your Trosky references. [81:42] Okay you lose you lost 95% of the [81:44] audience. Just speak like a normal [81:45] person. [81:45] >> Chimath laughed. Chimath understood it. [81:47] >> Okay great. I know the artist is smarter [81:50] smarter than you give them credit for. [81:51] >> Okay. No, I just think you're just [81:53] making these deep polls to try to sound [81:55] smarter than you actually are. Uh the [81:56] reality is the reality. These people are [81:59] being laid off because of AI. Bill [82:00] Gurley, of 20 million people in the [82:02] United States driving cabs and trucks [82:04] and doing that as a job right now, how [82:07] many of those do you think will lose [82:08] their jobs to self-driving in the next [82:10] decade or two based on just being in [82:12] there? And I'm not trying to lead the [82:13] witness here in any way. Obviously, some [82:15] people prefer a human driver, but what [82:17] what's your take on on that specific [82:18] part of the economy? [82:20] >> I think it's impossible to go with a [82:22] 100% automated uh solution [82:26] because the the economics don't work [82:28] well. And so, I think like some of the [82:31] other examples that were given, ATMs and [82:33] whatnot, I I think the the use of [82:36] nonownership [82:38] cars is going to go way up. So, it's [82:40] going to keep growing through this and [82:43] humans are going to be used for like 50% [82:45] of it instead of a hundred. And so, I [82:48] might not be surprised if the number [82:50] actually stays the same or grows. And [82:54] let's remember these jobs didn't exist [82:56] before because regulation had limited [83:00] what the taxi market was capable of and [83:02] and getting around that actually led to [83:05] job creation. And so I I'm not a big fan [83:08] of the doomerism because around jobs, [83:11] you know, there's a word lite that that [83:13] kind of is used to to talk about it. And [83:17] I don't have high confidence in any [83:19] government program for skills [83:21] retraining. So it's not clear to me what [83:25] okay yes it's happened. What do we do [83:27] now? It's not clear to me. I think the [83:29] thing you can do the most one we already [83:32] talked about use the new tools. Know [83:33] what it's capable of in your field. like [83:35] get out there. And then two, if your job [83:38] is going to go away and maybe it's a job [83:39] you don't care about, start thinking [83:41] about where there are opportunities. [83:44] Everyone's talking about it. The skilled [83:46] trades are like we're we're short of [83:49] people. [83:50] >> Shortage for plumbing, electricians, [83:52] HVAC, all that. Yeah. [83:53] >> It's amazing how Jay uses facts that [83:55] haven't happened yet as like support for [83:57] his argument. Like you just state that, [84:00] oh, all the truck drivers are losing [84:01] their jobs. All the drivers are losing [84:03] their jobs. And then you say that this [84:05] proves my take. Yeah, I know it's your [84:07] belief, but that is not proof. Do you [84:09] understand? [84:10] >> The proof I gave was Amazon and Andy [84:11] Jasse, Shopify, [84:14] they were doing that before Mark Benny [84:16] off automated Amazon. Let me ask Bill [84:20] everybody knows our Amazon car being [84:23] delivered so you cherry pick anecdotes [84:25] and then misattribute them to AI. [84:27] >> They literally have a self-driving [84:28] division. It's called Zuks. [84:29] >> You're the biggest AI washer there is. [84:31] They are the largest [84:33] user of robotics in the world. So yes, [84:36] Chimath, they are pursuing robotics [84:39] massively more than anybody and they are [84:41] pursuing self-driving. [84:42] >> You just you just like all these words [84:44] together. At one point it's a warehouse [84:46] worker, then it's a driver, then it's [84:47] Amazon. [84:49] It's just [84:49] >> it's not it's not you can you can [84:51] personally attack me all you want. The [84:53] the issue here is self-driving is going [84:56] to take away I believe the majority of [85:00] >> Okay, that's the key word. Great. Let's [85:02] put it there as a belief. Who knows? You [85:04] don't know and I don't know. [85:05] >> Okay. And I think the same robotics. But [85:07] I will take people I will take people at [85:10] their word. I'm curious, Bill, your take [85:12] on [85:12] >> these large enterprises. You've heard [85:14] two positions here. [85:15] >> I have a question for you. [85:17] >> I have a legit question for you. [85:18] >> Can I just let the guest be involved, [85:20] please? Sax your monopoly. actually [85:22] engaging with your with your [85:23] perspective. Explain to me. No, no. This [85:25] let me let me truly ask you. [85:27] >> Okay. [85:27] >> Explain to me why job postings for [85:30] software engineers are up 15% [85:31] year-over-year despite the fact that [85:33] code has now been fully automated. [85:35] >> Oh, I think there's a Cambrian explosion [85:37] in uh software. You're absolutely [85:40] correct. And I believe people who know [85:41] how to vibe code or to like who are [85:44] non-developers are making bespoke [85:46] software. I've said that a hundred times [85:47] on this podcast over the last few years [85:49] and I predicted it. So absolutely I [85:51] believe that will be an area of job [85:52] growth. I believe the positions that are [85:55] being removed or I just I know based on [85:57] what we're hearing is product managers, [86:00] middle managers, what Matthew Prince [86:02] call measurers, what other people call [86:04] mid management. Everybody believes that [86:06] the recording and this daily standups [86:10] and the uh zoom calls all of that is [86:13] turning into people management is being [86:15] done better by AI and people are more [86:18] self-directed and then the stack of [86:20] people to build products is being [86:21] consolidated right it's like the the [86:24] typical designer can now vibe code the [86:26] developer can do front-end design and UX [86:29] and they can project manage themselves [86:30] so I there there are a series of jobs [86:33] that will increase and a series of jobs [86:35] that will be eliminated just like the [86:37] mail room got eliminated and mess bike [86:40] messengers got eliminated [86:42] that [86:43] >> got on net on net do you think there'll [86:45] be mass [86:46] >> do you think there'll be on well your [86:47] position is shifting a little bit do you [86:49] think on net do you think on net [86:51] there'll be mass job loss [86:54] >> uh I think there is a chance that we're [86:56] going to see uh job loss increase in the [86:59] short to midterm and then eventually [87:02] the displaced people are going to have [87:03] to learn or leave the workforce, which [87:05] is what happened during other [87:07] revolutions like this. Some people went [87:09] with the paradigm and adapted and some [87:11] people didn't and just retired. That I I [87:13] saw that firsthand in the PC revolution [87:15] as but one example. Some lawyers just [87:17] would never use these tools and they [87:19] just retired at 55 65 and they moved on. [87:22] And then other attorneys were PC first [87:25] and they just took that. [87:25] >> By the way, did you guys see the news [87:26] that Kirkland Ellis is going to spend [87:29] half a billion dollars to roll their own [87:31] Frontier model? [87:32] >> Makes total sense. That was like to our [87:33] earlier point today is that people are [87:34] doing on prem and going to make their [87:36] own models. Bill, I have one specific [87:37] question for you and thank you for the [87:39] good engagement there, Sax. It was it [87:41] lacked the ad homonym that usually uh [87:43] starts every conversation we have. [87:45] >> I don't usually call you an idiot. [87:46] >> That's because it's in our minds. Okay, [87:48] we're thinking about it. We're just not [87:50] saying it. [87:50] >> Good. I like it better. I like it [87:52] better. Um Bill, specifically when Andy [87:55] Jasse, you know, uh last spring said, [87:57] "Hey, we're going to do more with less. [87:59] We're going to be AI first." and they [88:00] said, "We're not going to hire these [88:01] 600,000 jobs." When you see Tubby Lucky, [88:04] he say, "You have to do AI first before [88:06] you ask for a headcount and prove to me [88:07] that you tried AI first before hiring [88:09] somebody." Do you think this is [88:13] a sign that these organizations are AI [88:16] washing or do you think these recent [88:18] ones are more, hey, we're going to do [88:20] more with less and and the size of these [88:22] companies will be smaller uh because of [88:24] AI? One thing that that I think that [88:27] last question misses and I think a lot [88:30] of the the AI dumerism stuff misses is [88:34] that competition exists. And so if you I [88:38] don't think there's any scenario where [88:40] you just do more for less and all of a [88:42] sudden everyone has 70% operating [88:45] margins. That's not going to happen. [88:48] Someone else is going to come along and [88:50] do more for less and lower the price. [88:52] And so the thing that could happen is we [88:55] could have a productivity boom from [88:57] lowerpriced goods and services and the [88:59] basket of goods that humans are able to [89:02] buy gets cheaper and cheaper and cheaper [89:04] and that's been true in many categories. [89:06] Unfortunately, it's offset by what [89:08] happens in healthcare and other [89:10] regulated industries. [89:11] >> Yeah. Education. But but yeah, so I [89:14] expect products to get cheaper and [89:17] people to be able to create more with [89:19] less. But I don't think it leads to [89:22] obscene [89:23] profits [89:25] >> because it'll be whittleled away in [89:26] competition. [89:27] >> Okay. [89:28] >> By the way, just just on this AI washing [89:30] point, there's a a trial lawyer named [89:33] Donnie King. He's a securities [89:34] litigation partner at a firm called [89:36] Acriman. He and his colleagues have [89:40] started to warn that we could start [89:42] seeing shareholder lawsuits against [89:45] companies that engage in this type of AI [89:47] washing cuz he thinks it's a type of [89:49] puffery, right? Because essentially what [89:51] the company is doing is attributing [89:53] their own non-performance or their [89:55] operational issues to AI when in fact [89:58] there are real problems in the business [90:00] and therefore it could be a form of [90:02] securities fraud. [90:03] >> Wait, securities fraud? Yeah. [90:05] >> I want to double click on this. Did you [90:06] see the CEO of Whisk today on his note [90:08] about layoffs? [90:09] >> No. Who's Wisk? [90:11] >> Find me the AI washing in there. Wix. [90:13] >> Oh, yeah. Those that's the website [90:14] builder. Yeah. I mean, you can build [90:16] websites with Claude. Yeah. The whole [90:19] website business is challenged. Yeah. [90:21] >> Interesting that he just he just talked [90:23] about operational details. [90:25] >> Did he? I I didn't read the note. [90:26] >> Of course, you didn't read the note. [90:28] >> Well, I mean, you said it just happened. [90:30] [laughter] [90:31] >> I I will read it. Uh, this is I'm just [90:35] this broke at 9:25 this morning. [90:37] >> I think it's really interesting that [90:39] this lawyer thinks there's so much AI [90:41] washing going on that he thinks it could [90:43] constitute securities fraud and he's [90:44] warning clients not to engage in it. But [90:46] look, Jake Al, you're like the last [90:48] person who hasn't gotten the memo on [90:49] this. There was a huge narrative shift [90:51] this week. Sam is backing off this. Even [90:53] Daario's walking it back. You got the [90:55] Goldman Sachs CEO. [clears throat] [90:58] You got the explosion in job postings. [91:00] Everyone's coming around to the idea [91:03] that the job apocalypse is massively [91:06] overblown. [91:07] >> I mean, it could be overblown looking at [91:09] the stats. [91:10] >> Your apology, I'm happy to accept. [91:12] >> No need for an apology. I mean, my [91:13] position has always been apologize. [91:15] >> It's displacement. You're going to have [91:16] some displaced in the short to midterm [91:18] and then eventually there'll be more [91:20] problems to solve and people will have [91:21] to reallocate. I do think we're being uh [91:24] uh, you know, I think the tech industry [91:26] itself doesn't have enough empathy or [91:29] enough thoughtfulness when discussing [91:31] this because these are real people [91:32] losing real jobs and you can point at [91:35] statistics uh and you think they're [91:37] spinning, but these are real people [91:39] losing real jobs who may not make the [91:41] transition. [91:42] >> Jason, I got I got [91:43] >> Do you think it's more empathetic to be [91:45] scaring the be Jesus out of people that [91:47] they're going to lose? [91:47] >> I'm not in the scaring camp. I'm not in [91:49] the scaring camp. I I I am in the [91:51] enabling camp. That's why I keep saying [91:52] if you've been laid off, you should [91:54] start a company and you should embrace [91:56] the tools. So, I I'm I'm all about [91:57] empowering people. I think they if they [91:59] learn the tools, they'll have 10 job [92:00] offers. Uh and they'll start their own [92:02] companies. So, I do think there's a [92:04] solution to it. I just think we're going [92:05] to go through, you know, low millions of [92:08] jobs being lost or being retired and [92:11] transitioned out over this next couple [92:13] years. [92:14] >> I was just going to be empathetic and [92:15] offer some solutions. So we talked about [92:18] the skills trade deficit of of people [92:22] working in that. Micro has a foundation [92:24] called Micro Works where they fund they [92:28] funded $16 million 2600 people get a [92:32] free scholarship for to be become a [92:34] plumber, welder or electrician. So go [92:37] check that out if you want to reskill. I [92:38] think [92:39] >> generation tool belt. Yeah. [92:40] >> I think it's better than you know having [92:42] the government fix things. And then, you [92:44] know, as as we started and talked about, [92:46] I've got a new uh grant program myself [92:48] to help people [92:50] >> Yes. [92:50] >> kind of tilt their career in a different [92:52] direction. Go do something you love and [92:53] and apply and maybe I can help fund your [92:57] you moving in that direction. [92:58] >> And and as to your vibe shift, I think [93:00] it's because candidly people's houses [93:04] have been Molotov cocktailed because [93:05] they're doomerism. And people [93:07] specifically are citing that when they [93:08] shoot at their houses and throw Molotov [93:10] cocktails at them twice in the same [93:12] week. And if you're IPOing and you're [93:15] coming out saying, "Hey, jobs are going [93:16] away. Jobs are going away." That's just [93:18] a really bad look. [93:19] >> And or it's because we called it out and [93:21] they got cut and so now they were [93:23] telling the truth. [93:24] >> All right, everybody. This has been [93:26] another amazing episode of the AllIn [93:28] podcast. Thanks for coming. [93:30] >> No, no, no. Sorry. I need to do one [93:32] thing [93:34] >> just doing at the end of this. [93:35] >> Yes. [93:36] >> She's a friend of ours. I just want to [93:38] just give a huge shout out to Tulsy [93:39] Gabbard and specifically her husband [93:41] Abraham. is tragic. [93:42] >> Going through some really tough stuff [93:44] with cancer. He is going to kick its [93:47] ass. I just wanted to say we love you. [93:50] >> Yeah. Uh Tulsi is great. [93:52] >> Cheers everybody. Uh that's episode 275 [93:54] in the can. We'll see you next time. [93:56] Bye-bye. Love you Blues. [music] [93:58] >> We'll let your winners ride. [94:01] >> Rainman David. [94:04] [music] [94:06] >> We open sourced it to the fans and [94:08] they've just gone crazy with it. [94:10] >> Love you. Queen of Kino. [94:13] [music] [94:19] >> Besties are gone. [94:21] >> That is my dog taking notice in your [94:23] driveways. [94:26] >> Oh man, my appetiter will be. [94:29] >> We should [music] all just get a room [94:30] and just have one big huge orgy cuz [94:32] they're all just useless. It's like this [94:33] like sexual tension that we just need to [94:35] release somehow. [music] [94:37] >> Wet your feet. Wet your feet. her feet. [94:42] [laughter] [94:42] >> We need to get mer. [94:46] [music] [94:52] I'm going all in.