China Weighs AI Model Export Limits as SK Hynix Seeks $28B Listing

中国拟限制AI模型出口，SK海力士寻求280亿美元上市 北京正考虑限制海外获取Z.AI的GLM-5.2和字节跳动豆包等前沿AI模型，此举效仿了美国的管制措施，或迫使美国企业转回昂贵的本土方案。 主持人探讨了关于中国正考虑实施AI出口管制以限制海外获取Z.AI的GLM-5.2和字节跳动豆包等前沿模型的报道，并将其与文章《AI 2027》中的预测进行了对比。他们剖析了SK海力士拟在纳斯达克进行280亿美元上市的计划，以及Leopold Aschenbrenner的Situational Awareness基金的参与情况，同时讨论了Meta发布的Muse Image与Muse Video模型。本期内容还涵盖了投资人Jeremy Giffon对AI就业替代的看法、美国主要银行试图收购Fiserv借记卡网络、菲亚特售价1.4万美元的Topolino微型电动车亮相，以及纽约市一座摩天大楼在“商改住”翻新过程中出现的结构柱失效问题。 中国正考虑对阿里巴巴、字节跳动、智谱 AI（Z.AI）和 DeepSeek 等领先 AI 模型的出海实施出口管制，效仿美国的监管模式；此举可能导致全球对 GLM-5.2 和 Qwen 等高性价比开源权重模型的使用受限，迫使企业重新依赖昂贵的美国前沿模型或 Meta 生态。 (0:00) 英伟达核心高带宽内存（HBM）供应商 SK 海力士（SK Hynix）筹备在纳斯达克进行规模达 280 亿美元的上市，Baillie Gifford、Coatue 以及 Leopold Aschenbrenner 旗下的 Situational Awareness 基金拟认购高达 70 亿美元的股份。 (8:20) Meta 推出首款专用图像生成模型 Muse Image 及 Muse Video，依托 Muse Spark 实现联网搜索规划，并在推理思维链中具备突现的自我优化（self-refinement）与多轮连贯编辑能力。 (10:51) 摩根大通、美国银行和富国银行等美国大型银行正探讨收购 Fiserv 旗下的 Star 和 Accel 借记卡清算网络，旨在规避《德宾修正案》费率上限并重新整合支付链条，直接挑战 Visa 和 Mastercard 的市场主导地位。 (17:15)

As geopolitical competition prompts Chinese regulators to weigh export limits on premier open-weight models, infrastructure builders are expanding hardware capacity and reasoning-driven multimodal tools. Simultaneously, major financial institutions are mounting a direct challenge to incumbent payment rails by seeking vertical control over debit networks. Geopolitical Controls Threaten Open-Weight AI Ecosystems (0:00) Chinese authorities have initiated preliminary meetings with leading domestic tech firms—including Alibaba, ByteDance, Zhipu AI (Z.AI), and DeepSeek—to examine restricting overseas access to their advanced artificial intelligence models. Under consideration are regulatory mechanisms such as mandatory national security reviews before publishing weights, required state approvals for cross-border distribution, or an outright prohibition on exporting frontier open-weight models. The hosts emphasize that systems like Z.AI's GLM-5.2 and Alibaba's Qwen currently serve as cost-effective alternatives for US enterprises; cutting off access would direct developers back toward proprietary US frontier labs or Meta's open models. This regulatory impulse closely tracks projections from the essay AI 2027, which forecasted that Beijing would recognize resource disadvantages and centralize model oversight by mid-2026. The move mirrors tightening scrutiny by Washington over state-of-the-art models such as GPT-5.6 and Claude 3.5 Sonnet, signaling that both global powers are actively considering formal state vetting before frontier algorithmic weights can cross international borders. SK Hynix Listing Exposes Hardware Valuation Hesitation (8:20) South Korean semiconductor producer SK Hynix is preparing a $28 billion NASDAQ listing, providing Western capital with direct access to a pivotal artificial intelligence hardware supplier. Holding an estimated $1.1 trillion market capitalization after a 750% equity surge over the past year, SK Hynix dominates the High Bandwidth Memory (HBM) supply chain through its critical partnership with Nvidia. Substantial institutional interest anchors the offering, with Leopold Aschenbrenner's Situational Awareness hedge fund, Baillie Gifford, and Coatue projected to secure up to $7 billion of the equity sale. Despite remarkable 2025 financial results—revenues expanding 47% to $63 billion and annual profits more than doubling to $28 billion—the stock continues to trade at a modest forward price-to-earnings multiple of roughly 7x. The hosts note that institutional allocators remain restrained by memories of historical memory chip volatility across cloud, smartphone, and cryptocurrency cycles, maintaining valuation discounts rather than treating current AI infrastructure demand as an unyielding structural break. Multimodal Generation Integrates Inference Reasoning (10:51) Meta has unveiled Muse Image alongside a preview of Muse Video from its Meta Supercomputing Lab, indicating an evolution from earlier visual architectures like Chameleon and Emu toward reasoning-guided generation. The system incorporates Muse Spark, an internal reasoning layer that analyzes text prompts, queries live web data, and formulates an execution plan prior to rendering image outputs. Scale AI's Alexandr Wang highlighted technical developments including emergent self-refinement during the model's chain of thought at inference, multi-reference input blending, and coherent multi-turn iterative editing. The hosts observe that these advanced capabilities require substantial GPU infrastructure to sustain at consumer scale, marking an industry shift where generative media relies increasingly on planning and self-correction rather than pure latent diffusion. Labor Transitions and the Construction of Economic Utility (13:45) Analyzing the labor market implications of automation, investor Jeremy Giffon argues that widespread fear regarding structural job elimination misunderstands the nature of modern employment. Giffon asserts that the vast majority of white-collar occupations represent artificial social constructs rather than essential activities required for physical survival, such as basic food production, shelter, or healthcare. Because capital naturally depreciates under inflationary conditions, it requires ongoing allocation, creating coordination and administrative roles. Giffon acknowledges that near-term AI transitions will generate real economic turbulence, job loss, and displacement. However, he maintains that human consumption desires are functionally inexhaustible, ensuring that long-term economic systems will continue to invent novel tasks and categories of work as legacy cognitive functions become automated. Banks Target Card Networks to Reclaim Payment Rails A coalition of top US commercial lenders—including JPMorgan Chase, Bank of America, Wells Fargo, and PNC—is exploring the acquisition of Fiserv's Star and Accel debit networks. The potential transaction represents an effort by issuing banks to circumvent debit interchange limitations mandated under the post-2008 Durbin Amendment by vertically integrating the issuer, clearing network, and consumer digital wallet. This strategic alignment mirrors Capital One's acquisition of Discover, targeting the transaction fee supremacy and network fees traditionally collected by Visa and Mastercard. Rather than combating consumer-facing software fintechs, the banks are coordinating to construct proprietary alternative routing rails that claw back payment processing margins. Across foundational models, memory fabrication, and payment rails, dominant institutions are attempting to mitigate structural vulnerabilities by exerting tighter jurisdictional and infrastructural control. However, unresolved tensions persist over whether capital markets will continue to fund extensive compute buildouts and how smoothly national economies can absorb rapid technological displacements.

TBPN本期节目聚焦人工智能地缘政治与资本动态，重点探讨了中国考虑对先进AI模型实施出口管制的行业传闻、SK海力士拟在纳斯达克进行规模达280亿美元的上市计划，以及Meta最新多模态生成模型Muse的技术细节与商业意义。 中国或推行前沿AI出口管制 (0:00) 据路透社报道，中国相关部门已与阿里巴巴、字节跳动、智谱AI（Z.AI）和DeepSeek等本土领先科技企业举行初步会议，讨论对向海外输出前沿AI模型实施限制的可能性。监管机制可能包括要求企业就未来发布进行备案、在开源模型权重前进行国家安全审查、要求海外发布前获得政府批准，或彻底禁止前沿开放权重模型的输出。主持人指出，这一动向与此前行业文章《AI 2027》中预测中国将在2026年中期觉醒并应对AGI管制的时间线高度吻合，同时也呼应了华盛顿近期针对下一代模型（如GPT-5.6、Claude 3.5 Sonnet等）实施的海外准入限制政策。 战略层面上，中国开源模型在全球市场具备显著的价格和性能竞争力，例如智谱AI的GLM-5.2在成本大幅降低的同时性能比肩美国顶级模型，阿里通义千问（Qwen）家族也已成为开放生态的关键支柱。若中国收紧对外开源渠道，美国企业将失去用于低风险任务的廉价、高质量开源模型，被迫回流至采购昂贵的美国本土前沿实验室API或Meta开源生态，从而重塑全球AI生态系统的供给格局与成本结构。 SK海力士拟赴美上市与周期分歧 (8:20) 作为英伟达核心合作伙伴及高带宽内存（HBM）主导供应商，韩国存储芯片制造商SK海力士正筹备在纳斯达克进行规模达280亿美元的股份发售，若落地将成为亚洲企业在纽约历史上规模最大的上市交易之一。得益于AI基础设施的爆发式需求，SK海力士2025年营收增长47%至630亿美元，全年利润翻倍突破280亿美元，其首尔股价过去一年飙升超过750%，总市值达到约1.1万亿美元。Leopold Aschenbrenner旗下的Situational Awareness对冲基金与Baillie Gifford、Coatue拟共同认购至多70亿美元的配售份额，凸显了顶级机构资本对其核心AI硬件地位的认可。 尽管财务数据亮眼，SK海力士目前的远期市盈率仍维持在约7倍的相对低位。分析指出，这一估值折价根源于传统投资者对半导体周期性波动的顾虑。历史上，内存芯片制造业曾多次在云服务、智能手机及加密货币热潮消退后遭遇残酷的繁荣-萧条周期；许多市场参与者仍将其归类为周期性硬件股，对供应链瓶颈及潜在的需求停滞保持谨慎，尚未完全将其视作不可逆的AI算力长期范式转变。 Meta发布Muse多模态模型 (10:51) Meta超级计算实验室（MSL）正式推出了首款专用图像生成模型Muse Image，并与负责推理规划的组件Muse Spark进行深度整合。Muse Spark允许系统在最终输出前分析用户指令、调用网络搜索并完成生成规划，从而帮助用户在Meta AI应用内实现“一步到位”的生成效果。此外，Scale AI创始人Alexandr Wang公开指出了该模型的三个核心技术特征：推理阶段通过涌现的思维链进行自我修正（self-refinement）、将多张输入图像无缝融合的多参考组合（multi-reference composition），以及在保持全局连贯性的前提下执行多轮迭代局部编辑的能力。 与此前Meta FAIR推出的自回归Transformer模型（如Chameleon、Emu）以及Segment Anything不同，Muse系列代表了其产品化视觉生成技术的战略转向。Meta还同时预览了聚焦于时序一致性与精准提示词遵循的视频生成模型Muse Video。节目认为，此类高保真度图像和视频生成能力需要庞大的GPU算力支持，若直接接入Instagram等创作者分发场景，有望通过信息流推荐算法自然筛选低质内容并放大优质视觉创意。 AI焦虑下的白领劳动重构与金融博弈 (13:45) 针对AI引发的广泛失业焦虑，投资人Jeremy Giffon在播客采访中给出了反共识判断。他认为，绝大多数白领工作本质上并非生产食物、住所或医疗等维系生存所必需的基础物资，而是社会发展构建出的虚拟分工。例如，资本配置者的岗位存在仅源于货币的通胀属性驱动着资金必须再投资以保值。Giffon承认短期至中期技术转型将引发实质性的经济动荡和失业阵痛，但从长期来看，由于人类消费欲望永无止境，经济系统最终会不断发明全新的分工领域来容纳人类劳动力。 在传统金融与支付网络领域，摩根大通、美国银行、富国银行和PNC等华尔街大型银行正探讨联手收购Fiserv旗下的Star和Accel等借记卡网络。这一举动旨在效仿第一资本（Capital One）收购Discover的逻辑，通过垂直整合发卡行、支付清算网络和数字钱包全链条，绕开2008年金融危机后德宾修正案（Durbin Amendment）带来的费率管制，对长期占据垄断地位的Visa和万事达卡网络发起直接冲击。 本期节目梳理了从地缘政治限制、基础设施资本化到生成式AI端侧演进的多维进展，核心未解议题在于：中国潜在的开源限制是否会实质性削弱全球低成本AI开发生态，以及面对硬件周期性疑虑与支付基础设施的重组博弈，AI投资叙事能否在宏观经济现实中维持其高估值支撑。

## Background 本期内容由科技商业对谈节目 TBPN 于 2026 年 8 月 3 日发布在 YouTube 平台。作为其精炼版的《Diet TBPN》系列，官方节目简介指出本期旨在用 30 分钟时间呈现当天最重要的科技头条与商业动态，涵盖前沿模型监管、重磅半导体 IPO 以及华尔街金融支付格局重塑等热点。 ## TL;DR 节目核心聚焦于中国或将效仿美国出台前沿 AI 出口管制的传闻，并将其与利奥波德·阿申布伦纳在《AI 2027》中的“中国觉醒”时间线进行了深度对照。随后，主持人们分析了韩国 HBM 霸主 SK 海力士拟在纳斯达克募资 280 亿美元背后的资本热潮与周期迷局，并插播了 Meta 最新发布的 Muse 图像生成模型。在商业与宏观层面，节目探讨了摩根大通等多家华尔街大行联手竞购 Fiserv 借记网络以围剿 Visa 和万事达卡的垂直整合反击，最后还轻松调侃了仅有 46 英里续航的菲亚特微型电动车，以及曼哈顿前辉瑞总部大楼在“商改住”施工中发生的立柱弯曲险情。 ## Atmosphere & style 整期节目洋溢着高密度的行业洞见与极具硅谷特色的松弛讽刺风格，在严肃推演地缘 AI 博弈与万亿级半导体资本重构的同时，穿插着大量辛辣的即兴吐槽。主持人和现场嘉宾互动自然且充满化学反应，导播频繁触发的夸张音效甚至一度引发了直播聊天室观众的“两极冷战”与现场道歉自嘲。听众仿佛置身于科技精英的私人沙龙，既能迅速捕捉全球顶级硬核趋势，又能沉浸在对劣质工程改造与猎奇玩具车的欢声笑语中。 ## Bulletin - [00:00](https://www.youtube.com/watch?v=sedZBKHmkL0&t=0s) 路透社报道中国监管部门正就限制前沿 AI 模型对外出口与阿里、字节跳动、智谱 AI 等展开闭门磋商，或将影响全球廉价开源供给与开发者成本。 - [08:20](https://www.youtube.com/watch?v=sedZBKHmkL0&t=500s) 韩国存储芯片巨头 SK 海力士拟在纳斯达克 IPO 募资约 280 亿美元，Situational Awareness、Baillie Gifford 和 Coatue 计划认购高达 70 亿美元。 - [10:51](https://www.youtube.com/watch?v=sedZBKHmkL0&t=651s) Meta 正式推出多模态图像生成模型 Muse Image，具备强化学习自主涌现的自我修正、多图参考构图及多轮对话无损编辑三大核心特性。 - [17:15](https://www.youtube.com/watch?v=sedZBKHmkL0&t=1035s) 摩根大通、美国银行等华尔街大行拟联合收购 Fiserv 旗下借记卡清算网络（STAR 与 Accel），旨在绕过 Visa 和 Mastercard 打造支付闭环。 - [22:11](https://www.youtube.com/watch?v=sedZBKHmkL0&t=1331s) 曼哈顿东 42 街前辉瑞总部大楼在进行纽约最大规模“商改住”翻新工程时，因承重立柱弯曲导致第 21 至 26 层下陷并引发紧急疏散。 ## Key takeaways - [00:00](https://www.youtube.com/watch?v=sedZBKHmkL0&t=0s) 中国监管机构探讨限制前沿 AI 导出海外，暗合了此前 Leopold Aschenbrenner 在《AI 2027》蓝图中预测的“2026 年年中中国觉醒”节点。 - [05:00](https://www.youtube.com/watch?v=sedZBKHmkL0&t=300s) 政策对已发布的开源权重难以溯及既往，但未来可能通过事前安全审查、API 出口审批或直接禁止开源等手段进行管控。 - [05:00](https://www.youtube.com/watch?v=sedZBKHmkL0&t=300s) 美国中小开发者长期依赖智谱 GLM 与阿里 Qwen 等中国开源模型作为日常低成本主力，一旦出口收紧将带来沉重的成本冲击。 - [08:20](https://www.youtube.com/watch?v=sedZBKHmkL0&t=500s) SK 海力士受 HBM 需求驱动市值突破 1.1 万亿美元，2025 年营收达 630 亿美元，赴美上市将成为亚洲企业在美最大规模发售之一。 - [09:20](https://www.youtube.com/watch?v=sedZBKHmkL0&t=560s) 尽管 SK 海力士利润翻倍至 280 亿美元，但因存储行业固有的强周期性阴影，其远期市盈率仍被压制在 7 倍左右。 - [09:20](https://www.youtube.com/watch?v=sedZBKHmkL0&t=560s) Leopold Aschenbrenner 的《Situational Awareness》报告展现了极强的资本号召力，成为撬动数百亿算力基建投资的核心叙事资产。 - [10:51](https://www.youtube.com/watch?v=sedZBKHmkL0&t=651s) Meta 发布 Muse Image，与 Muse Spark 规划模型协同运作，能够在生成前理解逻辑并自主规划高难度构图与排版。 - [10:51](https://www.youtube.com/watch?v=sedZBKHmkL0&t=651s) Muse 模型展现出的思维链自我提纯（Self-refinement）能力，是在强化学习（IRL）过程中自然涌现而非人工写死规则。 - [13:45](https://www.youtube.com/watch?v=sedZBKHmkL0&t=825s) 投资人 Jeremy Giffon 提出白领岗位大多为现代信用社会的“虚构”衍生品，人类无穷尽的消费欲将持续催生出替代旧岗位的新工作。 - [17:15](https://www.youtube.com/watch?v=sedZBKHmkL0&t=1035s) 摩根大通、美银等多家银行联手竞购 Fiserv 旗下的 STAR 和 Accel 网络，借此对抗限制刷卡费率的德宾修正案。 - [17:15](https://www.youtube.com/watch?v=sedZBKHmkL0&t=1035s) 华尔街大行效仿 Capital One 收购 Discover 的闭环模式，其核心矛头直指收取高昂借记卡通道费的 Visa 与 Mastercard。 - [18:51](https://www.youtube.com/watch?v=sedZBKHmkL0&t=1131s) 支付生态正迎来从底层账户、发卡、清算轨道到 AI Agent 结账终端的“垂直重新捆绑”，而 Ramp 等顶层抽象软件有望从中获益。 - [19:53](https://www.youtube.com/watch?v=sedZBKHmkL0&t=1193s) 菲亚特 Topolino 微型纯电车以 1.4 万美元起售价入美，但仅 46 英里续航与 19 英里时速让其更偏向玩具车而非实用通勤工具。 - [22:11](https://www.youtube.com/watch?v=sedZBKHmkL0&t=1331s) 纽约前辉瑞总部 38 层办公大楼在加盖改建 1500 套公寓期间承重柱变形，六层楼面下沉引发严重结构性安全危机。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=sedZBKHmkL0&t=0s) 中国或效仿美国出台AI出口管制：闭门会谈与“AI 2027”预言对照 当天的重磅头条直指中美人工智能地缘博弈的新转折：中国监管部门可能正在借鉴华盛顿的打法，考虑限制前沿 AI 模型的对外出口。据路透社报道，中国相关部门已与国内多家顶尖科技企业展开闭门会议，参会方包括阿里巴巴（Alibaba）、字节跳动（ByteDance）以及智谱 AI（Zhipu AI/Z.AI）等，尽管深度求索（DeepSeek）和科大讯飞（iFlytek）在报道中未被直接点名，但外界普遍推测它们同样处于对话的核心圈层。会议探讨的核心在于限制海外用户访问中国最先进的模型，其性质正如美国此前针对前沿模型施加的层层监管——随着模型能力越发强大，政府正试图牢牢掌握谁能调用这些系统的决定权。 这番动作立刻引发了主持人们对前沿思想实验的现实对照。有人抛出问题：“在《AI 2027》的时间线里，中国究竟是何时‘觉醒’的？”答案赫然正是“2026 年年中前后”。主持人提到，自己甚至在日历上专门设置了提醒，用来验证 Anthropic 联合创始人杰克·克拉克（Jack Clark）在去年 12 月 23 日做出的神准预言。克拉克当时写道：“到 2026 年夏天，数字世界将仿佛经历一场极速进化，其中某些部分散发出巨大的光与热，并以反直觉的速度相对于其他一切疾驰。巨额财富将在这里诞生与破灭，而我们硅基造物的强大引擎将被投入运转，进一步加速经济并重塑万物。”现场团队打趣地询问身处技术一线的泰勒（Tyler），是否在现实中切实感受到了这股“巨大的光与热”。泰勒笑称自己早已深度沉浸其中，很难从外部视角审视变化，但不可否认的是，过去两三个月内，来自华盛顿特区的所有头条几乎全部被 AI 霸榜，这在年初还是不可想象的景象。 从网络安全风险、代码模型与 Agent 能力的飞跃，到企业级软件市场的实际营收落地，这一波爆发对自去年夏天由安德烈·卡帕斯（Andrej Karpathy）提出“氛围写代码”（Vibe Coding）以来就紧密跟踪技术迭代的圈内人而言，完全在意料之中。而回顾 2025 年 4 月利奥波德·阿申布伦纳（Leopold Aschenbrenner）发表的《AI 2027》蓝图，文中明确构想在 2026 年年中“中国觉醒”：面对 AGI 出口管制与西方算力优势带来的危机感，对策转向大举推进国家级 AI 攻势，甚至启动国家信息共享机制。主持人们不禁调侃：“中国莫非是把《AI 2027》当成指导手册在逐字执行吧？我们或许应该用‘别抄美国播客和哲学家的作业’去实行反向激将法。” - **克拉克预言核验点**：2026 年夏季数字世界出现“高热高光”与反直觉发展速度，财富急剧洗牌。 - **华盛顿政策转向**：近两三个月内，AI 彻底成为美国首都政治与国家安全的核心议题。 - **技术突破铺垫**：自 Karpathy 提出 Vibe Coding 以来，编程与 Agent 能力已在企业级场景全面转化为实际营收。 --- ### [05:00](https://www.youtube.com/watch?v=sedZBKHmkL0&t=300s) 开源权重的地缘博弈：从GLM与Qwen到美国开发者的成本冲击 虽然闭门会谈仍在初步阶段，但政策如何在技术层面真正落地充满了现实悬念。对于已经公开发布在 Hugging Face 等平台上的开源权重，北京几乎不可能“撤回”海外访问权限，无论是镜像下载、本地部署还是微调早已遍地开花，中国法院也不太可能去跨国起诉海外的模型推理服务商。然而，针对尚未问世的下一代前沿大模型，监管层却有多种干预手段：温和的做法是要求企业在发布权重前向监管机构备案或进行国家安全审查；激进的策略是直接实施海外发布审批许可；而最极端的手段则是彻底禁止前沿开源权重模型的公开发布。不过从博弈论来看，彻底闭源也是一把双刃剑，因为中国此前凭借 DeepSeek 等模型对美国顶尖产品形成的“平替冲击”与定价压力，在国际竞争中赢得了巨大的生态声量。 ```mermaid flowchart TD accTitle: China AI Export Control Scenarios accDescr: Potential regulatory approaches for Chinese frontier AI models and their impact on global markets. subgraph Regulation["政策推演路径"] A["拟议出口管控讨论"] --> B["事前安全审查与备案"] A --> C["海外调用审批许可"] A --> D["全面禁止前沿权重开源"] end subgraph Impact["全球生态与成本反响"] D --> E["海外廉价前沿开源供给断流"] E --> F["美国中小企业开发成本大幅攀升"] E --> G["调用需求被迫回流美系闭源巨头"] E --> H["Meta Llama 生态压力与战略选择"] end ``` 这种收紧与当下白宫的动作形成了微妙的呼应。此前特朗普政府就曾对 Anthropic 旗下最先进的模型（如 Fable 5 与 Mythos 5）施加过访问限制，后在追加安全防护措施后才部分解禁；OpenAI 在推进下一代前沿系统部署时同样承受着来自政府的监管压力。促使中国方面考虑设限的关键背景，在于国产开源模型正展现出令人侧目的全球竞争力：智谱 AI 的 GLM-5.2 以极低的推理成本逼近美国顶级模型水准，引发硅谷高度关注；阿里巴巴的通义千问（Qwen）家族已稳居全球最重要的开源模型生态之一；字节跳动的豆包（Doubao）则统治着国内大众端应用生态。一旦中国斩断海外免费或廉价供给，美国 AI 开发者势必承受直接的经济打击——过去开发者普遍将昂贵的美国闭源模型留给核心高价值任务，而将海量日常负载交给优质且廉价的中国开源模型。若供给消失，市场需求是全面退回美国闭源巨头，还是倒逼 Meta 等公司加码补位，将彻底改变全球竞争格局。 - **潜在监管梯次**：从模型事前安全审查备案、限制跨国 API 服务，到彻底禁止开源权重海外分发。 - **国产代表生态**：智谱 GLM-5.2（极致性价比）、阿里 Qwen 系列（全球核心开源基座）、字节跳动豆包（国内高渗透系统）。 - **产业经济冲击**：美国中小企业长期依赖中国高性价比开源模型跑通底层流水线，政策收紧可能导致开发成本陡增。 --- ### [08:20](https://www.youtube.com/watch?v=sedZBKHmkL0&t=500s) SK海力士拟纳斯达克募资280亿美元：万亿巨头的HBM霸权与资本盛宴 在国际芯片与算力供应链领域，另一项震撼资本市场的世纪大交易浮出水面：韩国存储半导体巨头 SK 海力士（SK Hynix）正准备本周在纳斯达克通过股票发售募集约 280 亿美元资金，这极有可能成为有史以来亚洲企业在纽约进行的最大规模上市交易之一。SK 海力士早已在首尔证券交易所上市，受 AI 算力浪潮的强力催化，其股价在过去一年内暴涨逾 750%。本次奔赴美股上市，旨在为渴望直接下注底层算力硬件的美国主流机构提供更直接的通道。 这起 IPO 在推特和投资圈引发轰动，更关键的原因在于买方名单的顶级号召力：由利奥波德·阿申布伦纳掌舵的对冲基金 Situational Awareness，据传将与柏基投资（Baillie Gifford）、寇图资本（Coatue）一同重仓进场，这三家顶级投资机构合计可能吃下高达 70 亿美元的认购份额。SK 海力士之所以被公认为全市场最纯粹的 AI 基础设施标的，源于其在高带宽内存（HBM）领域的绝对统治力以及作为英伟达（NVIDIA）核心供应伙伴的关键生态位。如今，这家半导体巨头的市值已攀升至约 1.1 万亿美元（折合韩元超 1000 万亿，主持人们戏称这数字大到在换算时令人眩晕）。其基本面同样狂飙突进，2025 年全年营收同比大幅跃升 47%，达到了 630 亿美元，净利润表现更是呈现倍数级爆发。 - **募资规模**：计划于纳斯达克发售约 280 亿美元股票，或创亚洲公司赴美上市纪录。 - **首尔股价与估值**：韩国本土股价过去一年飙升超 750%，总市值突破 1.1 万亿美元大关。 - **基石与领投财团**：Leopold 旗下 Situational Awareness、Baillie Gifford 及 Coatue 拟联合认购最高 70 亿美元。 - **财务表现**：2025 全年营收同比猛增 47% 至 630 亿美元，凭借 HBM 绑定英伟达坐稳 AI 核心链条。 ### [09:20](https://www.youtube.com/watch?v=sedZBKHmkL0&t=560s) SK海力士的周期迷局与价值百亿的《态势感知》 SK海力士（SK Hynix）近期的业绩表现堪称惊人，利润直接翻倍达到280亿美元，第一季度营收同比猛增三倍。然而令人玩味的是，该公司的远期市盈率（Forward P/E）却仅有7倍左右。主持人们分析指出，这主要是由于存储芯片行业存在极其残酷的强周期性（brutally cyclical）。许多在半导体和存储领域摸爬滚打多年的老牌投资者，曾亲眼目睹过云服务爆发、智能手机普及以及加密货币挖矿等历次大潮中的盛极而衰，因此面对本轮AI基础设施建设，他们依然疑虑重重：如果AI算力需求再上一个台阶，SK海力士是否会面临更激烈的竞争挤压？产能瓶颈是否会导致增长停滞？大量资本仍习惯于生搬硬套过去的周期规律，而没有将其视为一次质变级别的全新技术周期。尽管如此，作为一家在美股亮相的亚洲半导体巨头，这依然是一场备受瞩目的重磅IPO。 这场AI芯片狂潮的背后，不得不提及利奥波德·阿申布伦纳（Leopold Aschenbrenner）撰写的那份著名的《态势感知》（Situational Awareness）长篇报告。正如投资播客《Invest Like the Best》的主持人帕特里克·奥肖内西（Patrick O'Shaughnessy）和杰里米·吉芬（Jeremy Giffon）曾调侃的“价值十亿美元的PDF”，如今看来，《态势感知》所催生的投资论叙与行业号召力，价值甚至高达数百亿乃至上千亿美元。主持人们打趣道，如果一份PDF就价值上百亿，那访谈该作者的播客节目（如Dwarkesh Patel的访谈）又值多少钱？无疑也在这场巨大的资本认知重塑中分得了一杯羹。 - SK海力士利润翻倍达280亿美元，Q1营收同比增长3倍，但因存储行业周期宿命，远期市盈率仅为7倍左右。 - 市场分歧核心：投资者是在套用PC、手机、加密货币的历史周期模式，还是真正理解了AI基础设施建设这一全新的技术范式。 - 讨论提到前OpenAI研究员Leopold Aschenbrenner的《Situational Awareness》报告，已成为驱动上百亿甚至千亿美元AI算力基建投资的“超级PDF”。 ### [10:51](https://www.youtube.com/watch?v=sedZBKHmkL0&t=651s) Meta发布Muse图像与视频生成模型 节目直播期间插播了一条重磅突发新闻：Meta正式推出了由其生成式AI团队（MSL）打造的首款图像生成模型Muse Image。Meta技术高管Alex Wang随后发布了展示贴，画面中展现了极具细节的各类标志与复杂排版。Muse Image与推理规划模型Muse Spark协同工作，在正式生成画面前能够理解Prompt逻辑、搜索网络并自主规划构图，使用户“第一次尝试就能得到心中所想”。该功能目前已直接接入Meta AI独立应用，不仅能逼真还原复杂的QR二维码，还能高保真生成信息图表，甚至完美再现带有褪色与光影质感的1983年大苏尔（Big Sur）宝丽来复古老照片。 Alex Wang重点提到了底层技术中最令人兴奋的三大特性：第一是“自我提纯与修正”（Self-refinement），模型能在思维链（Chain of Thought）中自动审视并改进自身生成的输出，这一能力是在强化学习（IRL）过程中自然涌现（emerge）的，并非人工预先设定；第二是“多参考图合成”（Multi-reference composition），能够将用户上传的多张独立图片元素融合成一张协调自然的全新作品；第三是“多轮编辑”（Multi-turn editing），用户在后续对话中调整细节时不会破坏画面的整体一致性，无需推倒重来。连线主持Tyler补充指出，虽然Meta FAIR早期曾推出过自回归架构（Autoregressive Transformer）的Chameleon和Emu等图像模型，但此次Muse代表了全新的方向。此外，Meta还同步预告了在提示词遵循度、视觉真实感与时序一致性上极具竞争力的Muse Video。随着这些极度消耗GPU算力的多模态模型向Meta AI全量铺开，这无疑成了扎克伯格此前大举狂购英伟达显卡最合乎逻辑的应用出口。 - **三大技术亮点**： - **自我提纯（Self-refinement）**：利用强化学习在CoT思考链路中自我发现瑕疵并即时修正。 - **多参考图构图（Multi-reference composition）**：支持同时上传不同主体（如指定的人物、自行车、西装风格）融为一体。 - **多轮编辑（Multi-turn editing）**：无需重置提示词，在保持画风与场景连贯性的前提下反复微调。 - **产品规划**：Muse Image已在Meta AI上线，主打复杂文字渲染、信息图与逼真QR码生成；Muse Video即将跟进。 ### [13:45](https://www.youtube.com/watch?v=sedZBKHmkL0&t=825s) 白领工作是“人造”的？关于AI失业论的激进反思与音效吐槽 随后，节目切入了知名投资者杰里米·吉芬（Jeremy Giffon）做客播客时发表的惊人观点：为什么我们完全不用担心AI会消灭所有工作？吉芬直言不讳地指出，现代社会几乎所有的白领岗位本质上都是“虚构的、编造出来的”（totally fake and made up）。除了维系生存所必需的食物、住所、医药等基础物资外，绝大多数白领工作与这些核心生存要素毫无关系，或者仅以极度间接的衍生方式产生关联。以吉芬自己的资产配置（Allocator）工作为例：因为法币资本天然具有通胀属性，有钱人不能把钱闲置在账户里，所以必须委托他人进行投资运转，希望资金保值增值。“这份工作有用吗？当然有用。但它是绝对真实的吗？并不是。”吉芬认为，正因如此，人类社会创造新工作岗位的潜力是无限的——现代经济的终极引擎是人类“永不满足的消费欲”（unquenchable desire to consume）。尽管在中短期内，技术替代可能引发阵痛、失业与社会动荡，但从长远来看，只要人类想找事情做，总能发明出层出不穷的新领域去折腾。 在讨论吉芬关于“西海岸硅谷正在吞噬东海岸华尔街，杠杆收购（LBO）本质上是依靠债务与财技而非提升企业本质质量的剥离式文化”时，直播间因为导播过于频繁地播放欢呼、惨叫等浮夸音效，引发了聊天区观众的强烈抗议。网友“Afterglow”愤怒留言吐槽音效板“极其刺耳、简直看不下去”。两位主持人见状连连苦笑道歉，解释称今天音响增益确实有些过热，但也打趣道如果能用AI开发一个“实时静音音效版”的并播Agent就完美了。随后助理Tyler和另一批粉丝却在公屏力挺：“别道歉，我们就是冲着这抽象音效来看节目的！”让现场演变成了一场围绕音效板的“观众南北冷战”。 - **Jeremy Giffon的核心论点**： - 白领工作在根本上是“虚构”的衍生品，不直接关联温饱与生存。 - 资本配置人的本质是通胀环境下的资本代管者，工作本身的价值建立在复杂的现代信用体系之上。 - 经济由无止境的人性欲望驱动，旧岗位被AI替代后，人类必然会虚构出全新形态的商品与服务需求。 ### [17:15](https://www.youtube.com/watch?v=sedZBKHmkL0&t=1035s) 华尔街大行打响反击战：组团收购借记卡网络挑战Visa霸权 针对吉芬所谓“西海岸正在吞噬东海岸”的论调，主持人们话锋一转，指出纽约东海岸的传统金融巨头正在发起一场凶悍的逆袭。摩根大通（JPMorgan Chase）、美国银行（Bank of America）、富国银行（Wells Fargo）以及PNC等多家美国超大型银行正联合谋划一项重磅交易——计划收购金融科技巨头Fiserv旗下的借记卡网络业务（包括STAR和Accel网络）。 这笔交易的背景可以追溯到2008年次贷危机后的《多德-弗兰克法案》德宾修正案（Durbin Amendment），该法案对大银行向商家收取的借记卡刷卡手续费（Interchange Fees）设定了严格上限，极大压缩了发卡行的利润空间。为了重新夺回借记卡业务的丰厚收益，这几家掌控着全美最庞大账户资金池的银行巨头，决定复制第一资本（Capital One）收购Discover的闭环打法：与其作为单纯的发卡行将交易拱手让给第三方，不如直接掌控底层的支付清算轨道。Fiserv旗下的STAR和Accel网络已深度接入数百万持卡人与成千上万家金融机构，拥有成熟的银行和商户网络底座。一旦大行联合接盘并把自己庞大的交易量强制迁移至该网络，它们就能截留更高的交易分成。这一洗牌动作最大的潜在受害者，正是常年坐收通道过路费的Visa和Mastercard，以及可能面临更高规则压制的广大终端商户。 ```mermaid flowchart TD accTitle: 大银行收购借记卡清算网络运作逻辑 accDescr: 展示摩根大通等银行通过收购Fiserv旗下的STAR与Accel网络绕过Visa/Mastercard并实现闭环交易的路径 subgraph WallStreet["华尔街银行联盟"] JPM["摩根大通 (JPMorgan)"] BOA["美国银行 (BofA)"] WFC["富国银行 (Wells Fargo)"] PNC["PNC银行"] end subgraph Deal["收购标的：Fiserv底层清算网络"] STAR["STAR 网络"] ACCEL["Accel 网络"] end subgraph Incumbents["既有格局受冲击者"] VM["Visa / Mastercard<br/>(过路费受损)"] Merchants["广大零售商户<br/>(议价权受到挤压)"] end WallStreet -->|"联合注资收购"| Deal WallStreet -->|"迁移海量自营借记卡交易流"| Deal Deal -.->|"绕过并分流支付量"| VM Deal -->|"直接连通并重塑收费规则"| Merchants ``` - **核心参与方**：摩根大通、美国银行、富国银行、PNC等多家跨国银行业巨头。 - **底层资产**：Fiserv旗下的STAR和Accel两大成熟借记卡网络，拥有数百万持卡人及数千家金融机构接口。 - **商业意图**：对抗限制借记卡刷卡费的德宾修正案（Durbin Amendment），将发卡端与清算网络合二为一，形成类似于Capital One收购Discover后的闭环生态，直接冲击Visa和Mastercard的双寡头护城河。 ### [18:51](https://www.youtube.com/watch?v=sedZBKHmkL0&t=1131s) 银行争夺借记卡网络：支付链路的垂直重新捆绑 如果银行能成功将更多交易量引导至自身掌控的网络，借记卡的处理成本可能会显著上升；与此同时，Visa 和万事达卡（Mastercard）在美国借记卡领域的议价权也将被削弱。然而，这背后更宏大的趋势是**整个支付体系正在经历垂直重新捆绑（Vertical Rebundling）**。各大银行不再甘于只做发卡机构，他们渴望全面掌控银行账户、实体卡片、清算网络、数字钱包、反欺诈风控层，甚至未来的“AI Agent 结账入口界面”（AI agent checkout surface）。目前银行对 Fiserv 旗下借记网络的正式竞购或许尚未完全浮出水面，但风声已经开始在行业内流传。 ```mermaid flowchart TD accTitle: 银行支付垂直整合链路 accDescr: 展示银行如何从账户、卡片、清算网络逐步垂直整合至AI Agent结账终端 subgraph Core["底层金融基建"] A["银行账户 (Account)"] --> B["卡片发行 (Card)"] B --> C["专有清算网络 (Network)"] end subgraph Services["增值与风控服务"] C --> D["电子钱包 (Wallet)"] D --> E["反欺诈与风控层 (Fraud Layer)"] end subgraph Frontier["前沿交互界面"] E --> F["AI Agent 结账终端 (AI Agent Checkout)"] end ``` 这种架构变革对以 Ramp 为代表的现代金融科技公司反而是个利好消息。因为 Ramp 这类平台运作在底层网络之上的软件抽象层（layer of abstraction），网络层面的洗牌对他们而言更多是正向推动。这并非针对 Stripe 或西海岸科技初创公司的打击，其真正的矛头直指 Visa 和 Mastercard——这两家传统卡组织长期以来不断提高借记卡手续费，早已让银行心生不满。 - **银行垂直整合目标**：从底层账户、发卡、底层网络、钱包、反欺诈，直到面向未来的 AI Agent 结账交互层。 - **利益博弈重点**：削弱 Visa 与 Mastercard 在美国借记卡市场上的垄断与定价权，降低银行交易成本。 - **对软件层的影响**：位于上层抽象架构的企业（如 Ramp）受冲击较小，反而能在底层竞争中捕获灵活性。 --- ### [19:53](https://www.youtube.com/watch?v=sedZBKHmkL0&t=1193s) 菲亚特 Topolino 微型电动车进军美国 随后，两位主持人聊起了前一天与来自 Cowboy Space Company 的 Baiju 讨论过的话题：菲亚特即将引入美国市场的新车——菲亚特 Topolino。这是一辆双座微型纯电动车，起售价为 14,000 美元，几乎只有上周讨论过的那款电动高尔夫球车价格的一半。尽管有人觉得菲亚特的车向来“不太正经（unserious）”，甚至正是因为这种不正经，车企才更应该把小巧猎奇的风格贯彻到底，但两个人依然开启了疯狂吐槽模式：谁会把这玩意儿当成主力车来开？ ``` 菲亚特 Topolino 核心规格： - 起售价格：14,000 美元 - 续航里程：46 英里（约 74 公里） - 整车重量：约 1,000 磅（约 453 公斤） - 最高时速：默认 19 英里/小时；加装合法低速车套件后可达 25 英里/小时 ``` 这台车最夸张的地方在于规格参数：满电续航仅有 46 英里，如果通勤距离单程超过 10 英里，电量就显得捉襟见肘。主持人调侃道，这车不该叫 Topolino，应该改名叫“Fiat Dice”（菲亚特骰子），“因为每次开它出门，你都是在拿命掷骰子（rolling the dice）”。整车重量仅约 1,000 磅，甚至“跟朋友两个人一起就能把它硬拉（deadlift）抬起来”。最高速度在未经改装时只有时速 19 英里，即便利用户外低速车辆（LSV）套件，极速也才时速 25 英里。主持人打趣说到办公室得开 30 分钟，但要是真买上 5 辆在路上排成一列“贪吃蛇车队”（train mode），后面跟着支援车，场面绝对让人笑不出来。 - **极低自重与小尺寸**：自重仅 1,000 磅，被宣传为“有史以来最可爱的车”，但实用性极低。 - **时速限制**：标配时速 19 英里，需要合法低速车辆套件才能达到 25 英里/小时。 - **续航焦虑极限**：46 英里的续航意味着它彻底告别主流家用车定位，充其量是高尔夫球车的平替玩具。 --- ### [22:11](https://www.youtube.com/watch?v=sedZBKHmkL0&t=1331s) 曼哈顿辉瑞前总部大楼改造危机：立柱弯曲与楼层塌陷 话题切回纽约曼哈顿发生的一起突发工程险情。位于东 42 街的纽约前辉瑞总部大楼正处于施工阶段，周二早晨 8 点左右，纽约市消防局（FDNY）接到报案称有砖块从高空掉落，随后现场紧急疏散。救援与工程人员进入现场后震惊地发现，建筑物内部有两根核心结构立柱已经发生严重屈曲变形（buckled），导致大约第 21 层一直到第 26 层整整六层楼发生严重下垂（sagging）。现场流出的内部视频显示立柱已经肉眼可见地被压弯，整座摩天大楼一度面临垮塌的严重危险。 幸运的是，现场施工人员全部清点完毕，未造成人员伤亡。但这起事故背后是一个极具野心的改造项目：由 Metro Loft 与 David Werner Real Estate 联手主导的纽约历史上规模最大的“商业办公楼改住宅公寓”（Office to Residential Conversion）工程之一。该项目原计划对原有的 38 层大楼进行翻新，并在局部额外加盖十多层，最终打造成包含 1,500 到 1,600 套住宅单元的公寓群。事故发生后，周边建筑被连带紧急疏散，市政工程检查员全面进驻调查能否进行结构加固与立柱重建，东 42 街附近区域也呼吁通勤者避让绕行。 - **事故规模**：两根承重立柱弯曲变形，波及第 21 层至第 26 层共计 6 层楼板下陷。 - **项目背景**：原 38 层前辉瑞办公总部，计划翻新加盖并改建为 1,500~1,600 套豪华住宅公寓。 - **市场阴影**：即便结构加固成功，这种“曾经发生过六层楼面结构下垂”的建筑在后续推盘销售时也将面临巨大的心理与声誉阻力。 --- ### [23:56](https://www.youtube.com/watch?v=sedZBKHmkL0&t=1436s) 豪宅倾斜旧闻与节目收尾 面对这起工程灾难，主持人调侃这真是“进行‘商改住’翻新最糟糕的一天”，并设想未来销售团队该如何向买家推销：“快搬进这栋曾有六层楼结构塌陷下坠的大楼吧”。这让人联想到旧金山著名的千禧塔（Millennium Tower）丑闻——那座顶级豪华高层公寓因地基沉降发生倾斜，甚至在豪华公寓的地板上放一颗弹珠都会自己滚走，最终引发了漫长的官司诉讼，耗费巨资进行结构加固工程才得以勉强扶正。现场观众留言吐槽“这真是拆迁公司发财的好日子”，而主持人则戏谑地将其总结为“不过就是几层下垂的楼板嘛，调好心态（mindset）别较真细节就行”。在轻松的调侃中，主持人提醒听众在 Apple Podcasts 和 Spotify 上打五星好评，订阅 tbpn.com 每日简报，并宣布明早 11 点直播再见。 ## Notable quotes > [00:00](https://www.youtube.com/watch?v=sedZBKHmkL0&t=0s) 到 2026 年夏天，数字世界将仿佛经历一场极速进化，其中某些部分散发出巨大的光与热，并以反直觉的速度相对于其他一切疾驰。巨额财富将在这里诞生与破灭，而我们硅基造物的强大引擎将被投入运转，进一步加速经济并重塑万物。 — Jack Clark > [00:00](https://www.youtube.com/watch?v=sedZBKHmkL0&t=0s) 中国莫非是把《AI 2027》当成指导手册在逐字执行吧？我们或许应该用‘别抄美国播客和哲学家的作业’去实行反向激将法。 — 主持人 > [09:20](https://www.youtube.com/watch?v=sedZBKHmkL0&t=560s) 价值十亿美元的PDF — Patrick O'Shaughnessy / Jeremy Giffon > [10:51](https://www.youtube.com/watch?v=sedZBKHmkL0&t=651s) 第一次尝试就能得到心中所想 — Alex Wang > [13:45](https://www.youtube.com/watch?v=sedZBKHmkL0&t=825s) 现代社会几乎所有的白领岗位本质上都是‘虚构的、编造出来的’。 — Jeremy Giffon > [13:45](https://www.youtube.com/watch?v=sedZBKHmkL0&t=825s) 这份工作有用吗？当然有用。但它是绝对真实的吗？并不是。 — Jeremy Giffon > [13:45](https://www.youtube.com/watch?v=sedZBKHmkL0&t=825s) 别道歉，我们就是冲着这抽象音效来看节目的！ — Tyler / 观众 > [19:53](https://www.youtube.com/watch?v=sedZBKHmkL0&t=1193s) 因为每次开它出门，你都是在拿命掷骰子。 — 主持人 > [19:53](https://www.youtube.com/watch?v=sedZBKHmkL0&t=1193s) 跟朋友两个人一起就能把它硬拉抬起来。 — 主持人 > [23:56](https://www.youtube.com/watch?v=sedZBKHmkL0&t=1436s) 快搬进这栋曾有六层楼结构塌陷下坠的大楼吧。 — 主持人 > [23:56](https://www.youtube.com/watch?v=sedZBKHmkL0&t=1436s) 不过就是几层下垂的楼板嘛，调好心态别较真细节就行。 — 主持人 ## Mentioned John Coogan, Jordi Hays, Jack Clark, Leopold Aschenbrenner, Andrej Karpathy, Tyler, Patrick O'Shaughnessy, Jeremy Giffon, Dwarkesh Patel, Alex Wang, Mark Zuckerberg, TBPN, 阿里巴巴 (Alibaba), 字节跳动 (ByteDance), 智谱 AI (Zhipu AI), 深度求索 (DeepSeek), 科大讯飞 (iFlytek), Anthropic, OpenAI, Meta, SK 海力士 (SK Hynix), 英伟达 (NVIDIA), Situational Awareness, 柏基投资 (Baillie Gifford), 寇图资本 (Coatue), 摩根大通 (JPMorgan Chase), 美国银行 (Bank of America), 富国银行 (Wells Fargo), PNC, Fiserv, Capital One, Discover, Visa, Mastercard, Ramp, 菲亚特 (Fiat), Cowboy Space Company, 纽约市消防局 (FDNY), Metro Loft, David Werner Real Estate, AI 2027, 态势感知 (Situational Awareness), Vibe Coding, GLM-5.2, 通义千问 (Qwen), 豆包 (Doubao), Fable 5, Mythos 5, HBM, Muse Image, Muse Spark, Muse Video, Chameleon, Emu, STAR 网络, Accel 网络, 德宾修正案 (Durbin Amendment), 菲亚特 Topolino, 千禧塔 (Millennium Tower) ## Video description Sign up for TBPN’s daily newsletter at TBPN.com Diet TBPN delivers the best of today’s TBPN episode in 30 minutes. TBPN is a live tech talk show hosted by John Coogan and Jordi Hays, streaming weekdays 11–2 PT on X and YouTube, with each episode posted to podcast platforms right after. Described by The New York Times as “Silicon Valley’s newest obsession,” the show has recently featured Mark Zuckerberg, Sam Altman, Mark Cuban, and Satya Nadella. TBPN is made possible by: Ramp - https://ramp.com Public - https://public.com Cisco - https://www.cisco.com Console - https://www.console.com CrowdStrike - https://www.crowdstrike.com Figma - https://www.figma.com MongoDB - https://www.mongodb.com NYSE - https://www.nyse.com Railway - https://railway.com Shopify - https://www.shopify.com/ Follow TBPN: https://TBPN.com https://x.com/tbpn https://open.spotify.com/show/2L6WMqY3GUPCGBD0dX6p00?si=674252d53acf4231 https://podcasts.apple.com/us/podcast/technology-brothers/id1772360235 https://www.youtube.com/@TBPNLive [Source](https://www.youtube.com/watch?v=sedZBKHmkL0)

[0:00] Jack Clark: 到 2026 年夏天，数字世界将仿佛经历一场极速进化，其中某些部分散发出巨大的光与热，并以反直觉的速度相对于其他一切疾驰。巨额财富将在这里诞生与破灭，而我们硅基造物的强大引擎将被投入运转，进一步加速经济并重塑万物。 [0:00] 主持人: 中国莫非是把《AI 2027》当成指导手册在逐字执行吧？我们或许应该用‘别抄美国播客和哲学家的作业’去实行反向激将法。 [9:20] Patrick O'Shaughnessy / Jeremy Giffon: 价值十亿美元的PDF [10:51] Alex Wang: 第一次尝试就能得到心中所想 [13:45] Jeremy Giffon: 现代社会几乎所有的白领岗位本质上都是‘虚构的、编造出来的’。 [13:45] Jeremy Giffon: 这份工作有用吗？当然有用。但它是绝对真实的吗？并不是。 [13:45] Tyler / 观众: 别道歉，我们就是冲着这抽象音效来看节目的！ [19:53] 主持人: 因为每次开它出门，你都是在拿命掷骰子。 [19:53] 主持人: 跟朋友两个人一起就能把它硬拉抬起来。 [23:56] 主持人: 快搬进这栋曾有六层楼结构塌陷下坠的大楼吧。 [23:56] 主持人: 不过就是几层下垂的楼板嘛，调好心态别较真细节就行。

[13:45] Jeremy Giffon: To me, it's like very obvious that every, certainly every white-collar job is like totally fake and made up in the sense that like they are not contingent for shelter, and food, and medicine... [13:45] Jeremy Giffon: Our economy is solely driven by our like unquenchable desire to consume things. And so we're going to come up with new things to consume. [13:45] Jeremy Giffon: In the long run, like we're just going to invent new things to do... the worry about, 'Oh, we're not going to have more jobs,' it just doesn't really resonate because I feel like we just make up stuff for us to do. [13:45] Jeremy Giffon: We've already solved all of our problems [19:53] every time you leave the house, you're rolling the dice.

[0:00] The big news of the day is that China [0:02] has taken a page out of Washington's [0:05] playbook potentially and aiming to limit [0:08] the export of AI models. Now, these are [0:10] just meetings at this point, but uh [0:12] Chinese authorities have reportedly, [0:14] this is according to Reuters, uh held [0:16] meetings with some of the country's top [0:18] tech firms. That's Alibaba, Bite Dance, [0:21] Z.AI, AI, uh, Deep Seek, KIF-Fly are not [0:24] mentioned yet, but you have to imagine [0:26] that they'll be in the conversation if [0:28] they aren't already about potentially [0:30] restricting overseas access to China's [0:32] most advanced model. So, they're [0:33] thinking about doing the same thing [0:34] that's happening with GPT 5.6, Soul, [0:37] Ultra, Mythos, Fable, all these [0:40] different models that are starting to [0:42] get more restricted. as they get more [0:44] powerful, the government wants a say in [0:46] what companies and what uh [0:48] non-governmental organizations, what [0:50] governmental organizations have access [0:52] to them. So, the discussions are still [0:53] preliminary. [0:54] >> When when did uh in AI 2027 did China [0:58] wake up? [0:58] >> Right around this time. Right around [1:00] this time, [1:01] >> mid 2023. I actually have onion my [1:04] calendar check on Jack Clark for today [1:07] because because last year December 23rd [1:11] Jack Clark predicted that there would be [1:14] big changes and acceleration in mid2026. [1:18] He said, "So by summer of 2026, it will [1:20] be as though the digital world is going [1:23] through some kind of fast evolution with [1:26] some parts of it emitting a huge amount [1:29] of heat and light and moving with [1:31] counterintuitive speed relative to [1:33] everything else. Great fortunes will be [1:35] won and lost here and the powerful [1:38] engines of our silicon creation will be [1:40] put to work further accelerating this [1:43] economy and further changing things." Do [1:46] you think that's where we are? Is that [1:47] an accurate assessment of mid2026? When [1:50] I saw this prediction, I put a calendar [1:53] reminder for mid 2026. We're right in [1:56] the middle of the year and I wanted to [1:58] check in and say, does it feel like [2:00] this? Obviously, this is not a [2:01] quantitative benchmark. This is a [2:02] qualitative feeling. Uh, fast evolution [2:05] emitting a huge amount of heat and light [2:07] moving with counterintuitive speed [2:09] relative to everything else. Great [2:11] fortunes certainly are being won and [2:12] lost every day in the market. Tyler, how [2:14] would you assess the amount of heat and [2:17] light emitting from some parts of the [2:21] digital world? Would you say it's huge? [2:23] >> I would say it's huge. I think it's [2:24] huge. It's kind of hard like I feel like [2:25] I've been like very tapped in for a long [2:28] time. So, it's hard to say like [2:30] [laughter] [2:31] >> you've been Yeah. You've been feeling [2:33] the [2:33] >> It's hard to see like an outside view is [2:34] like, oh, is this actually like is has [2:36] there been a big change like recently? [2:38] >> Yes. So, there's two things like one [2:40] >> it's been kind of like a personal thing [2:41] for you being so tapped in. Would you [2:43] say that? [2:44] >> So there's I mean there's two things [2:46] going on. One is that uh anyone who has [2:49] >> been like a personal thing for Tyler. [2:50] [laughter] He's been so tapped in for so [2:52] long. It's like [2:53] >> no [2:55] so much of news coming out of DC. Yeah. [2:57] The past like two three months has been [2:59] about AI which is like definitely was [3:01] not true. Yes. You know even earlier [3:03] this year I would say. [3:04] >> Yes. and the just the amount of heat and [3:07] light around cyber security issues, the [3:09] coding models, the agentic capabilities, [3:11] the revenues that we would see in [3:13] enterprise, those have not been a shock [3:15] to AI insiders who have been tracking [3:18] the capabilities of these models since [3:20] last October or even last summer uh when [3:24] models started when vibe coding was [3:26] coined by Andre Carpathy and coding [3:29] became you know not solved but you know [3:32] greatly enhanced by artificial So going [3:34] back to AI 2027 which was published [3:37] April 3rd 2025 uh by late 2026 they had [3:42] or sorry mid 2026 was China wakes up. Uh [3:45] in China the CCP is starting to feel the [3:48] AGI export controls and lack of [3:50] government support have left China under [3:52] resource compared to the West. I'm going [3:54] to fast forward a little bit. Hawks uh [3:57] Hawks and the CCP warned that the [3:58] growing race towards AGI can no longer [4:00] be ignored. So they finally commit to [4:03] the big AI push that they had previously [4:05] tried to avoid. He sets in motion the [4:08] nationalization of Chinese AI research. [4:10] Okay. [4:11] >> Creating an information sharing [4:12] mechanism for AI companies. So right now [4:15] they're just looking at export controls. [4:17] >> Feels like we're a little bit behind on [4:18] this. [4:18] >> It's possible the CCP is just reading [4:21] this as a guide [laughter] and just [4:23] implementing it. They're just planning [4:24] to implement it to a tea. [4:26] >> Thanks for the guide. [4:27] >> Just like Yeah, thanks. Thanks for the [4:28] playbook. You should have you should [4:30] have build me for it. [4:31] >> Yeah. No, the real the real 4D chess [4:33] here is to shame the CCP into not [4:36] copying off of AI 2027, saying, "Oh, [4:39] really? You're just doing the AI 2027 [4:41] thing?" Come on. You don't need to [4:42] invest in AI. Don't be mimemetic with [4:44] our with our philosophers literature and [4:47] our most popular podcasts over here in [4:49] America. Think of your own strategy. [4:51] Come on. What? Do whatever. Uh, no. So, [4:54] the discussions around AI export [4:57] controls in China are still preliminary. [5:00] Uh, and it isn't clear how Beijing could [5:01] claw back access to models that they've [5:04] already released. If they're on [5:04] HuggingFace, you can download them, you [5:06] can run them. And even if they had some [5:08] sort of intellectual property, they [5:10] could sue an open-source inference [5:12] provider for running Kimmy or something [5:15] like that. That feels really difficult [5:17] to do from Beijing. But, uh, there are [5:20] other implications. So, once openweight [5:22] models are posted publicly, of course, [5:24] they can be copied, mirrored, [5:25] downloaded, reused anywhere, fine-tuned. [5:27] Uh but there are several ways China [5:29] could restrict future Frontier models [5:31] before they leave the country. Uh at the [5:32] latest end, Beijing could require [5:34] companies to file future releases with [5:36] regulators or submit them for national [5:38] security review before publishing model [5:39] weights. A more aggressive version would [5:42] require government approval before any [5:44] advanced model could be released [5:46] overseas. And at the most restrictive [5:48] end, China could simply ban the public [5:50] release of Frontier openweight models. [5:52] They could say no more open sourcing. [5:54] Although the game theory there is a [5:55] little odd because it feels like the [5:57] fact that China has been open sourcing [5:59] near frontier capabilities has been to [6:02] it to China's benefit um in general [6:05] because it it creates like more [6:08] disturbances like the deepseek moment [6:09] was definitely good for China [6:10] >> pricing pressure. [6:11] >> Yeah, pricing pressure. [6:12] >> So uh they could limit these to state [6:15] approved customers or Chinese government [6:17] priorities which at a certain level [6:19] might just be the most important thing [6:21] to do. uh the RSI loop. Of course, you [6:24] want all of the GPUs that you have [6:25] focusing on advancing the frontier [6:28] internally as opposed to sending it off [6:30] to America. Uh the talks come as the US [6:32] is also moving toward tighter control of [6:34] Frontier AI. The Trump administration [6:36] recently restricted access to [6:37] Anthropic's most advanced models, Fable [6:39] 5 and Mythos 5, before later lifting [6:41] some of those controls after additional [6:43] safety mitigations. OpenAI has also [6:46] reportedly faced government pressure [6:47] around the rollout of its next frontier [6:49] models. Uh the issue has become more [6:52] urgent because Chinese openweight models [6:54] are increasingly competitive. Uh Z.AI's [6:57] GLM 5.2 in particular recently impressed [6:59] Silicon Valley by performing close to [7:01] leading US models at a much lower cost. [7:03] The Quen family of models from Alibaba [7:05] has also become one of the most [7:07] important open model ecosystems in the [7:08] world. While Bite Dance's Dao is one of [7:11] China's most widely used domestic AI [7:13] systems. I didn't know about that. If [7:14] China places export controls on future [7:16] frontier models, it could raise costs [7:18] for American AI users and companies [7:20] because it's obviously very beneficial [7:22] to have a great open source model [7:24] available to you um and uh for and use [7:28] them for lower stakes tasks while [7:29] reserving expensive US frontier models [7:31] for their most important workloads. If [7:33] the best Chinese models are no longer [7:35] freely available abroad, some of that [7:37] cheap model supply disappears. But you [7:39] have to ask what happens to Meta? Do [7:42] other do other labs start piling back [7:45] into the open-source uh ecosystem uh and [7:48] more uh or will all the demand just go [7:51] straight back to the frontier labs? [7:52] That's the question. It really does. It [7:54] really does. We should play some clips [7:55] because uh [7:56] >> we need to be thinking about podcast [7:58] safety because this is [8:00] >> an addictive. [8:00] >> There should be export controls on this. [8:02] It should be only available to American [8:04] citizens. They should Patrick even one [8:06] of the one of the guys over there is [8:08] Canadian. Yes. Uh maybe maybe North [8:11] America. Uh let's pull up a clip from [8:12] this. I want to I want to hear Jeremy [8:15] chat about all sorts of things. They [8:17] they go all over. Where in the timeline [8:19] is [8:20] >> Well, should we talk about SK Highix's [8:23] new listing? [8:23] >> Yes, we can run through that and then we [8:25] can pull up. [8:26] >> SK Highix is set to raise around 28 [8:28] billion in a NASDAQ share sale this [8:30] week. Potentially one of the largest [8:31] ever New York listings by an Asian [8:33] company. The South Korean memory [8:34] chipmaker is already public in Soul [8:36] where it stock is up more than 750% over [8:39] the past year, but the US listing gives [8:42] American investors an easier way in the [8:45] story is big on the timeline because [8:46] situational awareness, the hedge fund [8:48] run by Leopold, Ashen Brener, may [8:50] participate alongside [8:52] uh Bali Gford and CO2. Together, the [8:55] firms could take as much as 7 billion of [8:57] the deal. SKH Highex has become one of [8:59] the purest AI infrastructure plays [9:01] because it leads in HBM, has become a [9:04] key Nvidia partner, and now has a market [9:06] cap around 1.1 trillion, [9:09] >> which is like a quadrillion one. [9:11] >> Yeah, it's a lot. [9:12] >> When you look it up, it's very [9:12] confusing. I think [9:13] >> it's a lot. [9:14] >> The business is ripping. 2025 revenue [9:17] grew 47% to 63 billion. Profit more than [9:20] doubled 28 billion in profit. Uh Q1 [9:24] revenue tripled year-over-year. But [9:26] interestingly, the stock only trades at [9:28] 7x forward earnings because memory is [9:30] brutally cyclical. And there's still a [9:33] lot of people that are wondering h what [9:35] happens if there's another leg up in the [9:38] AI demand and the AI buildout. Will that [9:40] mean more competitive pressure? Will [9:42] there be a bottleneck where they do [9:43] stagnate? How much can they actually [9:45] grow? And so many investors that are [9:47] just in memory have seen booms and busts [9:50] from cloud or smartphone or crypto and [9:53] they say ah like I don't know that this [9:55] one's going to last forever. I'm pattern [9:58] matching to previous cycles not thinking [10:00] about this as a different technology [10:02] cycle. So uh I think that's what's going [10:04] on with SKH but it should be a [10:07] blockbuster IPO nevertheless. very [10:09] interesting to see an Asian company [10:11] listed [10:11] >> and uh most you know notable part of [10:15] this deal is is obviously situational [10:17] awareness. [10:18] >> Yeah. [10:18] >> Uh and and their participation in it. [10:20] >> Uh they they cover in best like the best [10:23] the billion dollar PDF an idea that that [10:25] Jeremy and Patrick have talked about [10:26] before. But situational awareness is [10:28] turning out to be more of a maybe a 10 [10:30] or a hundred billion dollar PDF. [10:32] >> Yeah. Seriously. A whole investing [10:34] thesis built around that. Uh it's just a [10:37] PDF. I I don't know. is an essay, a [10:39] thesis. [10:40] >> If if so, if the PDF was worth a billion [10:42] or or more than that, what was the [10:43] Dwaresh episode worth? It's got to be [10:45] some percentage of it, right? [10:47] >> Yeah, definitely. Because a lot of [10:48] people watch the podcast, watch the [10:49] clips, scroll through the PDF. [10:51] >> Uh, breaking news. Meta unveils Muse [10:53] image, its first image generation model [10:55] from MSL. [10:56] >> Waiting for this. Let's pull up some [10:58] images of this and see uh the results. [11:02] It should be good given that uh they [11:04] have all the Instagram and Facebook data [11:06] to train on. Alex Wang has a post. [11:09] >> Cool. [11:09] >> Let's pull this up. [11:10] >> Okay. Muse image, Muse video, whole [11:13] bunch of uh very detailed badges printed [11:17] all in one image. Uh he says releasing [11:20] Muse image today. The first image [11:21] generation model from MSL. It is crazy [11:24] that they didn't have any like when back [11:26] when it was fair, did they do anything [11:28] with image? Did they? Because they had [11:31] that amazing model segment anything. So [11:33] they were doing work on images and [11:35] obviously they've processed a ton of [11:37] images. It pairs with Muse Spark to [11:39] reason through your prompt, search the [11:40] web and plan before it generates. People [11:42] get what they mean on the first try live [11:44] now in the Meta AI app. Very very fun. [11:47] Even puts a realistic QR code in there [11:50] and can do infographics and uh old [11:54] Polaroids as well. Big su August 1983. [12:00] Three things Alex Wang is excited about [12:02] under the hood. Self-refinement. The [12:04] model improves its own output within its [12:06] chain of thought. Emerge during IRL, not [12:09] by design. Multi-reference composition. [12:11] Many images blended into one coherent [12:13] generation. Multi-turn editing. Iterate [12:16] without losing coherence or starting [12:18] over. Make an image of this person [12:20] riding a bike, this bike, and wearing [12:21] this suit while passing by these people [12:24] on a park bench. Make it look like a [12:27] drawing in this style. That's cool. So [12:29] you can upload multiple images and sort [12:31] of drop them in in his various prompts. [12:32] What you got Tyler? [12:33] >> Uh yes. So Metafare did release a number [12:35] of image models. So there was uh I think [12:37] it's pronounced Chameleon in 2023. There [12:39] was emu also 2023 and then there was um [12:42] another like version of Chameleon. [12:45] >> I think they were I they were somewhat [12:47] different than uh the like diffusion [12:49] which was like popular at the time. I [12:50] think they were all autogressive [12:52] transformer which is like how you get [12:53] the good tech stuff like that. [12:54] >> But um I I believe this is a different [12:56] uh [12:57] >> that's cool [12:57] >> theme. Yeah. And uh Alex is also [13:00] previewing Muse video competitive [13:02] unprompt adherence, visual fidelity, [13:04] temporal consistency. Also coming to [13:07] Meta AI. This has got to be uh GPU [13:10] intensive if they actually roll this out [13:12] to everyone. Although Meta AI does have [13:14] a smaller install base than Instagram, [13:16] but I mean imagine you really push this [13:19] in Meta AI and [13:21] >> what they buy all the GPUs for. [13:22] >> You gota you got to get people [13:24] generating stuff. I I completely agree. [13:26] I think put this in the hands of the the [13:29] modern Instagram creator. Let them have [13:31] fun with it and uh see you already have [13:34] the algorithm to filter the slop out. [13:36] The stuff that is sloppy but still [13:39] entertaining and creative uh will [13:41] surface to the top. So, congrats to them [13:43] on a successful launch. And let's click [13:45] over to Jeremy Gon on why he doesn't [13:48] worry about AI taking jobs. A very, very [13:52] interesting take on the latest episode. [13:54] understand this idea of we're going to [13:56] run out of jobs. To me, it's like very [13:58] obvious that every certainly every white [14:02] collar job is like totally fake and made [14:04] up in the sense that like they are not [14:06] contingent for shelter [applause] and [14:09] food and medicine and other necessities. [14:13] But most jobs do not touch those or if [14:15] they do they touch it in a very very [14:17] derivative way. What is your job as an [14:19] allocator? It's like, well, because [14:21] capital is inherently inflationary, you [14:23] have to you can't just leave it alone. [14:25] And so, like, my job is like when you [14:27] have money and you don't want it to go [14:29] away, you have to give it to someone. [14:31] And so, you you give it to a bunch of [14:32] people and then I take it and I put it [14:35] into things that are productive and then [14:37] hopefully you don't lose your money, you [14:38] get more money. And like, is this [14:40] useful? Is this good? Yeah, sure. But [14:42] like, it's not real. And to me, there's [14:46] like unlimited amounts of jobs that you [14:48] can create in those sorts of scenarios. [14:50] We're going to have unlimited wants and [14:51] desires and and and our economy is [14:53] solely driven by our like unquenchable [14:55] desire [applause] to consume things. And [14:56] so we're going to come up with new [14:58] things to to consume. And now again, in [14:59] the short term and medium term, that [15:01] might be volatile and there might be a [15:02] lot of job loss, etc. [screaming] And [15:03] you know, that's not good. And uh you [15:05] know, there could be a lot of despair. [15:07] But in in the long run, like we're just [15:10] going to invent new things to do. Like [15:12] we've already solved all of our [15:13] problems. like the worry about, oh, [15:16] we're not going to have more jobs. It [15:17] just doesn't really resonate because I [15:19] feel like we just make up stuff for us [15:21] to do. And that's sort of the whole [15:23] point of it. [15:26] >> That makes me think we need [laughter] [15:27] an invest like the best soundtrack. [15:30] >> We do need to invest like the best [15:32] soundtrack. Uh let's play the other clip [15:34] that uh Patrick shared. The West Coast [15:36] is definitely eating the East Coast. the [15:38] founder is incredibly important and sort [15:40] of the um the founding act is incredibly [15:43] important in uh in any business. I I do [15:45] think that it's notable that like the [15:47] current paradigm in which we live the [15:50] largest and most important finance firms [15:52] come out of a culture of leverage buyout [15:56] which it's extractive in so far as the [15:58] primary goal is to make a thing more [16:02] profitable. The core idea behind it is [16:04] like using debt and financial [16:05] engineering to make a lot of money in a [16:07] way that the quality of the business [16:09] itself is maybe ancillary to the core in [16:14] >> Afterglow says my god what an obnoxious [16:17] soundboard [laughter] unwatchable. [16:21] >> Sorry, [16:21] >> it is a little loud today. I feel like [16:23] the soundboard's coming in a little bit [16:25] hot. We're sorry about that. But we're [16:26] having fun here. And of course, we don't [16:28] want to we don't want to pirate the [16:30] entire Invest Like the Best episode. We [16:32] got to tease it. [16:33] >> This would be a good use case for AI if [16:36] we can make a separate show that's just [16:38] the show, but an agent in real time [16:41] pulls out [16:42] >> the sound effect [laughter] [16:44] >> and just has a normal show, no sound. [16:46] >> Sure. Sure. So, you can choose whether [16:48] you want the soundboard version of the [16:49] show or the not or or the [16:52] >> Tyler says, "Don't be sorry. That's why [16:53] I'm here." Tyler's just here for the [16:55] soundboard. So, after unfortunately, [16:57] there's people that are here that don't [16:58] like the soundboard. There's people that [16:59] are here only because of the soundboard. [17:01] We are trying to find the middle ground [17:03] and we don't always get it right. [17:04] >> We don't always get it right, [17:06] >> but we're trying. [17:07] >> It's sort of a cold war between the pro- [17:10] soundboard and the anti- soundboard, uh, [17:13] constituents in the audience. So, well, [17:15] uh, revenge of the East Coast. The East [17:17] Coast is trying to is trying to make a [17:19] comeback. Jeremy Gon says the West Coast [17:22] is eating the East Coast, but the East [17:23] Coast is trying to make a comeback uh, [17:27] across all of these big banks. Did you [17:29] hear about this? They're going after the [17:30] debit card rails. They're going after [17:32] Visa. [17:34] >> This is news today. JP Morgan, Bank of [17:37] America, Wells Fargo, PNC, and other big [17:40] banks are reportedly exploring a deal to [17:43] buy a FServe debit card network that [17:46] would own let them own more of the [17:48] payment rails um that the customers [17:50] debit cards run on. So the basic idea is [17:53] that post I think it was post financial [17:55] crisis there were limits on how much uh [17:57] banks could make on debit card [17:59] transactions. This is the Durban [18:01] amendment. [18:02] >> Durban [18:02] >> the Durban amendment I think Dick [18:04] Durban. Uh and they want a way to turn [18:07] debit back into a more valuable [18:08] business. So Fiserve owns major debit [18:11] networks including Star and Excel which [18:13] I don't run into that often. I usually [18:15] see Visa, Mastercard, MX. I mean MX [18:17] doesn't have debit network but star and [18:19] Excel already connect to millions of [18:20] card holders and thousands of financial [18:22] institutions. So this is like a it's a [18:24] real rail with an existing bank and [18:27] merchant connectivity. Um and so this is [18:30] like when Capital One bought Discover. [18:33] Um they own the issuer then they also [18:35] own the network and then they can move [18:37] their own volume onto their own rails [18:39] and capture more of the economics. Uh so [18:42] who loses if this goes through? Who [18:44] should be upset about this? It's [18:47] probably the merchants, Visa, and [18:49] Mastercard. Uh, so merchants would [18:51] probably see higher effective debit [18:52] costs if banks find a way to route more [18:54] volume through bankowned networks, while [18:56] Visa and Mastercard would lose some [18:58] leverage over US debit. Uh, but the [19:01] bigger story is that payments are [19:02] getting vertically rebundled. Banks want [19:04] the account, the card, the network, the [19:06] wallet, the fraud layer, and eventually [19:08] the AI agent checkout surface. So, it's [19:11] all very uh they I don't even think [19:13] there's a real bid for the for ferves [19:16] debit networks just yet, but uh it's [19:19] it's like starting to leak out that [19:21] they're doing it. Uh and I was [19:22] interested to know like is this is this [19:25] bullish for ramp because they sit on top [19:28] of of networks and I think it is [19:31] generally positive and and because of [19:34] the layer of abstraction they operate [19:36] on. Um, and I was wondering like is this [19:39] an attack on Stripe? Is this a is this [19:41] an attack on any particular West Coast [19:43] tech startup that we know? Uh, and it [19:45] seems like it's much more of an attack [19:47] on Visa and Mastercard in particular, [19:49] uh, who have been driving up debit fees [19:51] for them. Anyway, did you want to talk [19:53] more about Fiat's next US car? You [19:56] mentioned it when we were talking to [19:57] Baiu from Cowboy Space Company [19:59] yesterday. The Topelino, it's coming to [20:02] America. tiny two seat electric micro [20:05] car they're calling it. Uh starts at [20:08] $14,000 almost half the price of the [20:11] other electric golf cart that we talked [20:13] about last what was that last Thursday [20:16] we talked two yeah two in [20:19] >> but you can fit a lot more people in the [20:20] Amble I like this move from Fiat. [20:22] >> You like it? [20:23] >> Uh I think that their cars are generally [20:26] unserious. Yes. [20:27] >> And so I think it's good to just lean [20:29] into lean into that and [20:31] >> Okay. Okay. And um I think they're [20:34] trying to price this so that when people [20:35] are thinking, "Hey, I need a new car." [20:37] Yeah. Should I go out and buy like a [20:39] crossover SUV or [20:41] >> There's no way anyone can consider this [20:44] for their main car. It's impossible. [20:46] >> No, but you can if you typically roll [20:49] around with a handful of people and [20:51] instead of buying one car for [20:53] >> Oh, you buy a bunch of them [20:54] >> for 60 grand. [20:55] >> You'll need a train of four. You know [20:56] how you know the range on this thing? 46 [20:59] miles. [21:00] >> 46 miles. I mean, I guess if you have a [21:02] 10 mile commute, you can definitely [21:04] commute here. [21:05] >> They launch an EV with [21:06] >> 46 miles [21:07] >> 46 miles of range. [21:08] >> It's nothing. [21:10] >> They should have called it They should [21:10] have called Fiat. It's less than 10 [21:12] miles. [21:13] >> Less than 10 miles. [21:13] >> They should have called it the Fiat dice [21:15] because every time you leave the house, [21:17] you're rolling the dice. [21:18] >> It's a,000 lbs. So, you can deadlift it [21:20] with another friend. [21:22] >> It's There's Are you serious? [21:23] >> Thousand pounds. Yeah. This thing's [21:24] tiny. It's It's really truly like so so [21:27] small. [21:27] >> They're saying it's the cutest car ever [21:29] made. Uh, look at the speed. [21:32] >> The tops [21:32] >> tops out at 25 miles an hour. [21:34] >> That's with a street legal low-speed [21:37] vehicle kit. By default, it only goes 19 [21:40] miles an hour. Uh, so how long would it [21:42] take you to get here? 19 miles an hour. [21:45] >> 30 minutes. [21:46] >> Yeah, 30 minutes. [21:47] >> Wow. [21:48] >> That's not too bad. That's not too bad. [21:50] Maybe. Maybe you got to do it. Okay. You [21:52] guys are laughing, but you won't be [21:54] laughing when when I pull up to the [21:56] office with [21:56] >> fleet of these [21:57] >> with with five of these [22:00] >> and you need a support car like a [22:02] multiple behind you. [22:04] >> Yeah. No, no, the the the topo the topo [22:07] is the support. You just have you have a [22:09] series of you just go train mode. [22:11] >> Well, the main uh other news is that uh [22:14] this is this is somewhat related to New [22:17] York and Manhattan. A Midtown Manhattan [22:19] skyscraper under construction is [22:20] reportedly at risk of collapse. The [22:23] building was the former Fizer [22:25] headquarters on East 42nd Street and it [22:28] was evacuated Tuesday morning after the [22:29] FDNY received reports around 8:00 a.m. [22:33] Eastern time of bricks falling from the [22:35] highrise. When crews arrived, they found [22:38] two structural columns inside the [22:41] building had buckled with floors sagging [22:44] from the roughly 21st all the way to the [22:46] 26th floor. So, there's six floors of [22:49] sagging. Uh, this video this uh this [22:52] building seems to be falling apart. [22:54] Video from inside the site shows columns [22:56] visibly bent, raising immediate concerns [22:58] about the stability of the building. So, [23:00] so far, thankfully, no injuries have [23:03] been reported. Uh, let's hope it stays [23:05] that way. And construction workers have [23:07] been accounted for. The site was [23:09] undergoing one of the most ambitious [23:11] office to residential conversions in New [23:13] York history. Metro Loft and David [23:15] Warner Real Estate were converting the [23:17] former Fizer complex into roughly 1,500 [23:20] to,600 uh apartments with plans to [23:23] renovate the existing 38tory tower and [23:25] add more than a dozen new stories to [23:27] part of the property. Officials have [23:28] evacuated the construction site and are [23:30] in nearby buildings as a precaution. [23:33] Very very crazy. City inspectors and [23:34] engineers are now in sight investigating [23:36] the damage and trying to determine [23:37] whether the building can be stabilized. [23:39] They'll need to go and uh rebuild those [23:41] columns. For now, commuters are being [23:43] told to avoid the area. So, if you're in [23:44] New York, stay away from East 42nd [23:48] Street. I hope everyone is safe there. [23:51] >> Bad day to be doing that conversion. [23:54] >> Yeah, very, very chaotic situation. Uh [23:56] anyway, [23:57] >> it's also imagine trying to sell those [23:59] apartments after you finish. It's like, [24:00] oh yeah, moving into the building that [24:02] was buckling across six different [24:04] stories. [24:06] >> Yeah, there was a building. Was it in [24:09] San Francisco, the Millennium Tower? [24:10] >> Yeah. TZ says, "Good day to be a [24:12] demolition company, [24:13] >> I guess." Yeah. Yeah. There was the [24:15] Millennium Tower that was a very was it [24:17] was a fancy high-rise in San Francisco. [24:19] And the foundation like sunk a little [24:22] bit, and so the whole building was [24:24] tilted slightly, and so if you put a [24:26] marble on the ground in these like [24:27] luxurious apartments, it would roll. And [24:30] uh there was a huge uh huge lawsuits. I [24:32] believe that they were able to go do [24:33] some structural reinforcement and [24:35] eventually get it back. [24:36] >> It's just some saggy floors. [24:38] >> Just some saggy floors. [24:40] >> Yeah. Lock in just uh Yeah, maybe it's [24:43] just a mindset. [24:44] >> I'm sweating the details. [24:45] >> Thank you so much for tuning in. Leave [24:47] us five stars on Apple Podcast and [24:48] Spotify. [24:49] >> Sign up for our newsletter tbn.com and [24:51] we'll see you tomorrow at 11 a.m. [24:54] >> Goodbye. We love you. [24:55] >> Have a great week. [25:03] >> [music]