US AI Leaders Warn White House as China Restricts Export of Model Weights

中国限制模型权重出口 美AI领袖向白宫示警 美国高管与政策顾问警告称，廉价且开放的中国AI模型威胁美国市场主导地位，若不加以监管，恐将建立起由国家支持的数字基础设施。 本期节目深入剖析了围绕中国AI模型日益激烈的争论、美国高管关于监管俘获的主张，以及开放权重与仅限API访问的地缘政治影响。主持人探讨了Substack最新集成的Pangram AI检测功能，并讨论了内容代笔背后的广泛文化规范。此外，他们还探讨了Dave Eggers对AI文学的评论，以及硅谷奇特的船屋“住房捷径”，结合调查数据辩论了飞机座椅后仰礼仪，并重点介绍了Ramp新推出的成本优化型AI模型路由器。 美国AI企业高管对中国低成本、高性能开源权重模型的崛起发出预警，而中国商务部据报正在考虑限制开源模型权重与核心训练数据的出口，未来或转向仅提供云端API访问。 (15:37) Anthropic因未经授权下载和存储大量版权图书训练模型达成15亿美元诉讼和解，与此同时，业内围绕开源权重是否会导致“AI共产主义”以及监管俘获展开了激烈辩论。 Substack在其应用中集成了Pangram AI检测工具，可扫描并测定内容中合成文本的比例，旨在保护平台以真实人类创作为核心的经济模式。 (16:54) 企业财务平台Ramp推出了AI模型路由器，可根据任务自动将API请求分配至最具成本效益的模型，帮助企业精准控制AI推理和Token开销。 (28:27)

As Chinese open-weight models rapidly improve, Silicon Valley and Washington face an escalating dispute over national security, regulatory capture, and software export restrictions. Meanwhile, tech platforms are navigating the economic fallout of generative AI through copyright settlements, content authenticity tools, and automated enterprise cost controls. Geopolitical Clashes Over Chinese Open-Weight AI (15:37) The rapid advancement of Chinese models—most notably Moonshot AI’s open-weight Kimi K3 and Alibaba’s Qwen 3.8 Max—has polarized Washington and Silicon Valley. Executives from OpenAI and Anthropic, both preparing for public listings within the next year, have warned that unregulated, ultra-cheap Chinese models threaten security and could lead to a dystopian future. Critics, including venture capitalist and White House AI advisor David Sacks, counter that these warnings represent a regulatory capture strategy aimed at weaponizing policy uncertainty to eliminate low-cost market competitors. The ideological debate escalated when OpenAI strategic futures lead Dean Ball characterized open-weight dominance as a path toward "AI communism," where the state provides AI as public infrastructure rather than a market product, complicating the funding model for frontier development. Concurrently, geopolitical lines are hardening: the Financial Times reported that China's Ministry of Commerce (MOFCOM) has discussed restricting the export of training data and open weights with Alibaba, ByteDance, and Zhipu AI, favoring cloud API access instead. In the U.S., Treasury Secretary Besson suggested sanctions could be deployed against foreign models built on stolen domestic intellectual property. Copyright Settlements and the Model Distillation Debate Legal scrutiny over how frontier artificial intelligence is trained reached a major milestone with final district court approval of a $1.5 billion copyright settlement against Anthropic. The decision mandates payouts of $3,000 per work across roughly 500,000 works following rulings that Anthropic illegally acquired and stored millions of copyrighted books. This legal friction coincides with ongoing debates over whether newer international systems are simply distilled from established American models. Online analyses highlighting semantic similarities between Kimi K3 and Anthropic's Fable 5 or Opus variants have fueled claims of distillation or shared post-training methods, though observers caution that statistical correlation does not establish direct copying. The dispute has divided commentators: financial host Jim Cramer urged bans on cheap Chinese models on national security grounds, while George Hotz of tinygrad criticized leading U.S. developers as "rent extraction machines," arguing American firms must publish weights on platforms like Hugging Face to compete effectively. Platform Detection and Authorship Transparency (16:54) Publishing platforms are recalibrating their standards to distinguish human effort from automated output. Substack integrated an AI detection tool powered by Pangram, allowing users to scan posts and comments to calculate the ratio of synthetic to human-written text. Substack executives framed the feature as essential to maintaining reader monetization anchored in authentic human culture, even as critics question whether readers genuinely penalize synthetic text if it remains engaging. This push for technical attribution parallels unresolved norms around ghostwriting across digital media. While digital influencers face increasing calls for byline transparency, participants noted that legacy book publishing has long relied on undisclosed co-authors, estimating that up to 30% of airport bookstore titles use ghostwriters. Separately, author Dave Eggers argued in the Financial Times that consumer demand for entirely AI-written novels is nonexistent, dismissing the concept of AI literature as gullible. Enterprise AI Governance and Cost Optimization (28:27) As commercial deployment broadens, enterprise software is shifting focus toward managing token usage and controlling model inference expenses. Corporate card and financial software provider Ramp introduced an AI model router, released in private alpha to both current and non-Ramp customers. The product addresses widespread corporate frustration where engineering teams route low-complexity queries to costly frontier systems. By automatically redirecting API queries to the most cost-efficient model capable of handling the task, the router implements spend controls and expense attribution directly within CFO workflows. Investor Brexton Bennett characterized the launch as an answer to widespread market exhaustion with unchecked inference expenses and rapid model proliferation. The commercial AI landscape faces concurrent pressures from escalating geopolitical restrictions, legal accountability for training inputs, and corporate demands for inference efficiency. Whether open-weight distribution survives regulatory crackdowns in Washington and Beijing or is replaced by closed, state-managed APIs remains the sector's central unresolved question.

节目探讨了以月之暗面和阿里巴巴为代表的中国高性价比AI模型对美国科技界造成的竞争冲击，揭示了华盛顿与硅谷围绕国家安全、监管俘获与开源模型未来展开的激烈博弈，并同步涵盖了版权诉讼、内容真实性检测及企业推理成本控制等前沿议题。 中美开源模型角力与“监管俘获”之争 围绕美国企业是否应获准使用中国AI模型，硅谷与华盛顿陷入深度分歧。随着月之暗面Kimi K3及阿里巴巴Qwen 3.8 Max等具备高竞争力且成本低廉的开源权重模型涌现，筹备上市的OpenAI和Anthropic高管相继发出安全预警，声称若缺乏监管可能导致“反乌托邦式的AI未来”。然而，行业分析人士及白宫AI顾问David Sacks尖锐指出，这是行业巨头企图将监管不确定性武器化以扼杀竞争对手的“监管俘获策略”。 此争论进一步升级为意识形态与研发资金维度的对抗。OpenAI战略未来主管Dean Ball公开指出，开源权重模型的主导地位可能促成国家将AI作为公共基础设施供给的“完全AI共产主义”，进而使私营前沿开发难以维系巨额资金回流；但他随后澄清这仅为地缘政治视角的趋势预测，并非官方政策倡议。美国财政部长Besson在受访中表示政府虽支持开源但坚决反对IP侵权，暗示可能对涉嫌盗窃美国技术的海外模型施加制裁；而技术开发者George Hotz等人则批评美国巨头已沦为“寻租机器”，呼吁美企效仿中国同行将权重开源至Hugging Face以接受公平检验。 版权侵权诉讼、模型蒸馏与出口管制动向 (15:37) 在技术来源与训练合法性层面，美国加州联邦地区法院正式批准了一项针对Anthropic的15亿美元版权集体诉讼和解案。针对其非法下载并存储约50万部受版权保护作品的判决，版权作者与出版商将获得每部作品3,000美元的赔偿。与此同时，行业流传的语义相似度图表显示Kimi K3与Anthropic的Fable 5或Opus存在高度重合，引发了关于低成本模型是否利用老牌前沿模型进行“知识蒸馏”乃至违反服务条款的技术争辩，但亦有分析指出指标相关性并不能作为直接抄袭的法律定论。 地缘层面的模型流动管控正显现双向收紧趋势。据《金融时报》报道，中国商务部（MOFCOM）已与阿里巴巴、字节跳动和智谱AI等领军企业探讨限制关键训练数据出境，并考虑限制直接下载或导出开源模型权重，倾向于仅通过云端API向海外客户开放服务。这一迹象表明，通过仅开放云访问来保护国家AI数据与知识产权正成为政策趋向，而Kimi K3后续是否实际开放模型权重，将成为评估中美开源生态脱钩程度的风向标。 内容真实性危机：Substack检测与代笔反思 (16:54) 为应对生成式AI对原创生态的冲击，内容订阅平台Substack宣布整合Pangram推出AI检测工具，支持读者与创作者扫描专栏、回复和评论以评估人工撰写与机器生成的比例，将其定位于捍卫平台“真实文化经济引擎”的核心策略。然而主持团队指出，该类检测器极易受到行文习惯干扰，常将充斥行业套话的人工文本误判为100%合成内容，且受众对内容是否经由AI辅助的真正介意程度仍存疑。 这一工具落地引发了对数字创作者与传统出版业代笔透明度标准的深度探讨。针对创作者是否应强制披露外聘写手的问题，讨论指出传统出版界中未署名协助本就是常态，估计机场书店多达30%的畅销书出自隐形代笔，已故作家Tom Clancy等超级个人IP也依赖团队持续产出。主持团队倾向于创作者应在背景资料中保留团队透明度，但反对在每篇内容前施加硬性警示。与此呼应，作家Dave Eggers在《金融时报》撰文直言公众对AI文学的期待“既轻信又荒谬”，断言不会有人真正愿意阅读AI小说。 成本理性与生活套利：模型路由与空间礼仪 (28:27) 在AI应用与企业财务领域，企业对推理成本暴增和模型激增的焦虑催生了新型管理工具。企业财务平台Ramp推出了独立的AI模型路由器（处于私有Alpha阶段），专门针对开发人员滥用昂贵前沿模型处理诸如查天气等简单查询、浪费数百万Token的痛点。该路由器能将API请求自动智能分发至性价比最高的模型，并嵌入预算控制与支出归因功能，被投资人Brexton Bennett评价为契合CFO控费刚需的高价值产品。 此外，节目还将视角投向硅谷现实生活中的经济套利与社会行为摩擦。面对旧金山每月2,000至3,000美元的一居室租金，科技从业者将金门大桥附近月供约2,259美元（首付5.2万美元）的现代船屋视为降低居住开销的替代方案。而在公共空间冲突方面，节目援引《纽约客》专文及民调数据，指出经济舱乘客对前排调节座椅靠背的抵触情绪经历了显著代际上升：FiveThirtyEight在2014年的调查显示41%受访者认为调低靠背不礼貌，而Vacationer在2020年的调查中该反对比例大幅上升至77%，折射出公共空间舒适权与个人权益界限的认知变迁。 本期节目展现出AI产业正从单纯的算力与模型指标比拼，转入涉及国家出口管制、知识产权诉讼、监管寻租防御及企业实际Token支出约束的多维博弈阶段。美中双方对于开源权重掌控权的收紧预期，结合商业应用端对推理成本与真实性的严苛审视，标志着技术狂热正快速让位于高度现实的合规与商业理性。

## Background 该内容由科技脱口秀频道 TBPN（Technology Brothers）于 2026 年 8 月 3 日发布在 YouTube 平台上。本期为《Diet TBPN》每日 30 分钟精华节目，由 John Coogan 与 Jordi Hays 共同主持，精选当天科技前沿动态、行业巨头博弈与深度对话。 ## TL;DR 节目核心聚焦中美大模型博弈的白热化进展，剖析了中国开源模型（如 Kimi K3、Qwen）在低成本优势下对美国闭源巨头造成的市场冲击，以及白宫内部对于限制海外模型与“AI 共产主义”争议的政策分歧。随后，主持人们梳理了 Anthropic 巨额版权和解案、模型蒸馏溯源难题，并讨论了 Substack 上线 AI 识别工具背后的代笔与真实性争议。节目的后半段轻松转向硅谷极客为躲避高昂房租而兴起的“船屋居住”风潮、长途飞机后仰座椅的机舱道德大辩论，并在最后深度评析了财务管理平台 Ramp 最新推出的企业级 AI 模型路由器。 ## Atmosphere & style 整期节目节奏明快、谈吐犀利，主持人在硬核商业洞察与无厘头插科打诨之间切换自如。演播室内充满互相拆台、爆笑调侃与即兴自嘲，把严肃的地缘政治、万亿美元算力泡沫以及“监管俘获”等宏大议题解构得举重若轻。从对闭源巨头恐慌游说的辛辣讽刺，再到计算二手船屋月供和控诉飞机上踢椅背的琐碎日常，整体氛围生动鲜活、极具科技圈内部茶余饭后“吃瓜锐评”的陪伴感与趣味性。 ## Bulletin - [00:02](https://www.youtube.com/watch?v=KRE39H7hyUI&t=2s) 中美前沿大模型博弈加剧：月之暗面发布 Kimi K3，阿里巴巴推出 Qwen 3.8 Max，低成本开放权重模型的冲击引发华盛顿与硅谷对国家安全及行政禁令的激烈政策分歧。 - [06:14](https://www.youtube.com/watch?v=KRE39H7hyUI&t=374s) 闭源巨头面临商业与合规双重拷问：OpenAI 与 Anthropic 因鼓吹开源威胁被指搞“监管俘获”，而 Anthropic 刚刚达成的 15 亿美元版权侵权和解暴露出本土大厂的数据合规软肋。 - [16:54](https://www.youtube.com/watch?v=KRE39H7hyUI&t=1014s) Substack 集成 AI 检测工具 Pangram 治理劣质内容，直接引爆关于内容原创性、真人代笔（影子写手）披露机制与 AI 生成小说商业现实的伦理大讨论。 - [28:27](https://www.youtube.com/watch?v=KRE39H7hyUI&t=1707s) 财务管理平台 Ramp 正式推出 AI 模型路由器（Model Router），切入 CFO 治理员工滥用前沿旗舰模型导致 Token 成本失控的核心痛点。 ## Key takeaways - [00:02](https://www.youtube.com/watch?v=KRE39H7hyUI&t=2s) 中国开源及开放权重模型在基准测试中表现亮眼，月之暗面 Kimi K3 与阿里 Qwen 3.8 Max 等低价平替直接威胁美企高溢价模式，引发白宫关于是否封禁的路线分歧。 - [02:26](https://www.youtube.com/watch?v=KRE39H7hyUI&t=146s) OpenAI 战略负责人迪恩·鲍尔称开源模型主导将导致“完全的 AI 共产主义”，但分析指出模型商品化后商业价值仅会向云厂商和数据中心等算力基础设施层沉淀。 - [04:16](https://www.youtube.com/watch?v=KRE39H7hyUI&t=256s) 网络安全前沿模型推理成本极高，Google 顺势推出轻量级 Gemini 3.5 Flash Cyber，为中型企业提供了远低于顶级闭源方案的高性价比代码安全扫描选择。 - [06:14](https://www.youtube.com/watch?v=KRE39H7hyUI&t=374s) 大卫·萨克斯等自由市场支持者抨击闭源巨头打着安全旗号游说白宫实施“监管俘获”，试图借助法律高墙压制初创公司与开源生态以保住上市估值。 - [06:14](https://www.youtube.com/watch?v=KRE39H7hyUI&t=374s) 穆拉蒂创办的 Thinking Machines Lab 开源首款权重模型，英伟达推出 Nemotron 3 Ultra，美国本土力量亦在加速重构开源竞争力。 - [09:18](https://www.youtube.com/watch?v=KRE39H7hyUI&t=558s) 美政府对海外模型拟采取的制裁面临开源供应链溯源困局：利用论文、开源微调或蒸馏生成的数据冷启动本土模型，合规界限在技术层面极难厘清。 - [10:41](https://www.youtube.com/watch?v=KRE39H7hyUI&t=641s) 加州联邦法院批准 Anthropic 达 15 亿美元的版权侵权和解案，为约 50 万部受侵害作品赔付，其销毁实体书及非法下载电子书的行为引发公众舆论反弹。 - [12:00](https://www.youtube.com/watch?v=KRE39H7hyUI&t=720s) 语义相似度对比图显示 Kimi K3 与 Anthropic 旗下模型存在较高重合，引发蒸馏疑云，但后训练技术趋同也可能导致类似结果，跨国法律追责几乎不可行。 - [14:50](https://www.youtube.com/watch?v=KRE39H7hyUI&t=890s) 乔治·霍茨抨击美国巨头已沦为丧失活力的“寻租机器”，呼吁其开放权重；同时中国商务部亦在研讨对关键训练数据出境与开源模型权重的出口限制。 - [16:54](https://www.youtube.com/watch?v=KRE39H7hyUI&t=1014s) Substack 引入 Pangram 识别 AI 创作比例，引发创作者群体关于“署名真实性”的辩论，探讨网红博主与畅销书作家普遍使用影子写手是否同样应被强制披露。 - [18:45](https://www.youtube.com/watch?v=KRE39H7hyUI&t=1125s) 戴夫·艾格斯断言“无人会读 AI 小说”，但现实中 AI 生成内容在特定群体中已具市场规模，如同机场书店中大量存在的未署名代笔图书一样已成为流水线产品。 - [20:10](https://www.youtube.com/watch?v=KRE39H7hyUI&t=1210s) 面对旧金山高昂租金，科技圈兴起购置现代船屋（如 La Mer Modern 11）常住的“生活黑客”风潮，被戏称为彼得·蒂尔“海上家园”理念在湾区平民化落地的创业新根据地。 - [24:14](https://www.youtube.com/watch?v=KRE39H7hyUI&t=1454s) 长途航班中飞机椅背放倒问题引发道德困境，民调数据显示反对前排后倾靠背的乘客比例从 2014 年的 41% 剧增至 2020 年的 77%。 - [28:27](https://www.youtube.com/watch?v=KRE39H7hyUI&t=1707s) 企业财务平台 Ramp 推出 AI 模型路由器，根据任务复杂度智能分流轻量模型与前沿大模型，帮助企业 CFO 精准压缩非必要的 Token 算力开销。 ## Detailed walkthrough ### [00:02](https://www.youtube.com/watch?v=KRE39H7hyUI&t=2s) 中美模型博弈与白宫的分歧政策 中国在 AI 领域的突飞猛进再次登上各大外媒头条，美国科技高管纷纷拉响警报，白宫内部也因如何应对中国大模型的最新突破而产生深刻分歧。主持人指出，当前华盛顿和硅谷的状态十分矛盾：一边高喊“踩死油门向前冲”，另一边又想踩刹车；一边讨论是否限制本土模型，一边又在权衡是否要以同样的方式封禁中国模型。各类内部站队与互相指责层出不穷，各方的论点充满了内在冲突。 《华尔街日报》（The Wall Street Journal）发布了一篇重磅报道，梳理了这场波及硅谷与华盛顿的“数十亿美元级核心问题”：**美国企业究竟该不该获准使用中国的人工智能模型？** 多年以来，美企调用外部模型从未受到如此严苛的审视，但如今中国开源与开放权重（Open-weight）模型的性能提升势头极猛。以 OpenAI 和 Anthropic 为代表的闭源阵营高管密集发出警告，宣称廉价且强大的中国 AI 模型的崛起如果缺乏强力监管，将把人类引向“反乌托邦式的 AI 未来（a dystopian AI future）”，并带来不可承受的国家安全风险。这直接引爆了业内辩论：开源到底是 George Hotz 眼中的“解放力量”，还是闭源厂商口中“通往地狱的捷径”？ 针对闭源巨头的警告，不少行业分析师直言不讳地指出，OpenAI 和 Anthropic 预计都在准备于明年申请上市（IPO），他们大肆贩卖恐慌，本质上不过是想通过游说政策来消灭潜在竞争对手。尤其是最近几天，中国团队接连释放重磅产品： - 月之暗面（Moonshot AI）发布了 Kimi K3 模型，在多项基准测试中已能直接与美国顶尖前沿模型抗衡。 - 阿里巴巴发布了 Qwen 3.8 Max（通义千问系列最新模型）。 - 阿里巴巴作为月之暗面的重要战略投资方，其生态位与资本布局极为类似“美国 Google 与 Anthropic 之间的深度绑定与持股关系”。 ```mermaid flowchart TD accTitle: AI开源与闭源竞争及地缘政治图谱 accDescr: 展示中美前沿大模型博弈、商业模式冲突及监管博弈逻辑 subgraph CN_Ecosystem["中国前沿开放权重模型"] K3["Moonshot AI: Kimi K3"] Qwen["阿里巴巴: Qwen 3.8 Max"] CostAdv["极低推理与训练成本"] K3 --- CostAdv Qwen --- CostAdv end subgraph US_Debate["美国内部路线与政策博弈"] Frontier["闭源巨头 (OpenAI / Anthropic)<br/>巨额资本投入，前沿研发靠高定价支撑"] Capital["资本开支逻辑<br/>数据中心、算力基础设施与万亿估值"] GovReg["游说与监管争议<br/>David Sacks 抨击'监管俘获'与泛安全化"] end CN_Ecosystem -->|"低成本平替冲击"| Frontier Frontier -->|"依赖高昂利润"| Capital Frontier -->|"推动针对开源的行政禁令"| GovReg ``` ### [02:26](https://www.youtube.com/watch?v=KRE39H7hyUI&t=146s) “AI 共产主义”争议与开源算力经济学 这些即将开放权重或开源的中国模型允许全球企业自由下载，并结合私有数据进行特定微调，对企业级市场极具吸引力。然而，这恰好赶上了特朗普（Trump）政府内部关于是否采取行政措施限制中国模型在美使用的大讨论。新近加入 OpenAI、出任战略未来负责人的前特朗普政府官员迪恩·鲍尔（Dean Ball）在社交平台 X（原 Twitter）上发表了引发轰动的暴论，并被《华尔街日报》整段收录。 迪恩·鲍尔写道：“开放权重模型主导的世界，其最可能的一种结局是‘完全的 AI 共产主义’（full AI communism）。这恰恰是中国所主张的路线——AI 不再是市场化商品，而是一种公共物品，最终由国家作为某种‘数字公共基础设施’直接提供。”鲍尔甚至将这种场景描述为“反乌托邦式的地狱（a dystopian hellscape）”，并预测特朗普政府将采取行动削弱开源模型的使用。 节目主播对此展开了剖析，指出鲍尔口中的“AI 共产主义”仅仅发生在模型权重这一层： - 即使模型层被彻底商品化、公共化，算力基础设施的商业逻辑依然存在：看好算力、重仓基础设施和云服务商（Neoclouds）的人认为，商业价值将自然沉淀到数据中心层。数据中心即便在模型免费的世界里，依然可以赚取合理算力边际收益。 - 这并不意味着算力体系会走向国有化，只是原本集中在模型开发商身上的超高毛利空间将被大幅压缩。 - 鲍尔随后在舆论发酵后做出了澄清，强调自己并非公开建议美国政府全盘打击中国开源 AI，并指出其个人发言不代表 OpenAI 官方立场，“他只是在社交平台上碎碎念（He's just posting）”。不过他也指出了行业面临的核心困境：顶尖前沿实验室每年需要募集数十亿美元来购买海量计算资源；如果全天下都在免费使用开源模型而拒绝向模型研发付费，将不再有足够的商业回报去支撑下一代前沿大模型的研发。 与此相对，OpenAI 首席执行官萨姆·奥特曼（Sam Altman）此前公开承认过，OpenAI 过去只做绝对闭源的工具是一项不明智的策略，公司如今也在逐步开发并对外提供自己的开源模型。 ### [04:16](https://www.youtube.com/watch?v=KRE39H7hyUI&t=256s) 网络安全威胁、高昂成本与低成本专用模型 闭源阵营高管反复敲响的另一大警钟是网络安全控制力缺失。他们认为，由于开发者与决策者对已释出的开放模型毫无修改和使用管控能力，极其容易被武器化，用来发动大规模自动化网络攻击或垃圾信息滥用。主持人提到，当前高级金融机构与安全巨头在防护端已拥有特权：例如华尔街等金融机构早在六个月前就获准接入了专门用于网络防护的封闭前沿模型（如 Mythos 以及 GPT Cyber）。 然而，顶尖网络安全模型的推理与研发成本极其夸张。正如派拓网络（Palo Alto Networks）首席执行官尼克什·阿罗拉（Nikesh Arora）曾分享的案例：为了捕捉并扫描一个复杂的零日漏洞（Zero-day bug），他们可能直接烧掉近 100 万美元的算力成本。对于巨头而言这确实物有所值，但普通中小企业根本无法承受如此庞大的支出。 在此背景下，Google 刚刚推出了聚焦网络攻防的新模型。两位主持人梳理了 Google 的具体产品线细节： - Google 实际上分别发布了 Gemini 3.5 Flash Light 与 Gemini 3.5 Flash Cyber，并非把名字混在一起的“Flash Light Cyber”。 - Flash 系列定位本身就比庞大的 Pro 模型更加便宜轻量。 - 对于中型企业而言，与其耗费 500 万美元甚至砸破门槛去排队申请类似 Project Glasswing（微软/OpenAI 等顶级闭源联合项目）这类极其昂贵的 Mythos 级服务，不如直接利用现有的 Google 商业账户和生态，部署经过针对性微调的 Flash Cyber。它能以极低的成本通读每一行工程代码，主动识别出显性安全隐患。 ### [06:14](https://www.youtube.com/watch?v=KRE39H7hyUI&t=374s) 开源追击战：Thinking Machines Lab、Nemotron 与市场估值动摇 Anthropic 首席执行官达里奥·阿莫代伊（Dario Amodei）多年来一直对外直言开源的致命隐患。在彭博社（Bloomberg）近期的采访中，他再次重申：如果让任何人都能随意免费下载具有高阶网络安全穿透能力的前沿模型，将带来难以预料的灾难。但现实的商业天平正在发生倾斜——由于中国模型的价格比美国竞品低上数个数量级，美国本土企业调用中国模型的规模正在迎来爆发式增长。这直接动摇了资本市场对 OpenAI 和 Anthropic 等模型巨头长期护城河的信心。 与此同时，美国本土的开源模型生态也在加速反扑： - 前 OpenAI 首席技术官米拉·穆拉蒂（Mira Murati）组建的全新实验室 Thinking Machines Lab，于本周三正式开源了旗下首款开放权重模型。 - 英伟达（Nvidia）旗下的 Nemotron 3 Ultra 模型也正在开发者圈层中迅速获取关注。主持人风趣地吐槽英伟达的命名逻辑：“Nemotron 这个名字听上去很酷，但 Ultra 之后还能叫什么？苹果先用了 Pro、Max、Ultra，接下来英伟达是不是该叫 Giga、甚至 Exa 了？” 整个资本市场过去两年来敢于向 AI 基础设施砸入数万亿美元的核心假设，就在于笃信 OpenAI 和 Anthropic 能持续推出具有断代优势的前沿模型并收取高额溢价。如今，中国廉价模型的围剿与“极低价格带平替”的出现，直接导致上周多家科技和 AI 概念股股价大跌。 对此，知名风险投资人、白宫 AI 顾问大卫·萨克斯（David Sacks）公开猛烈回击。他认为迪恩·鲍尔的发言无异于公开承认了“监管俘获（Regulatory capture）策略”。萨克斯等自由市场支持者长期以来一直指责 Anthropic 等巨头打着安全与伦理的幌子游说监管，真实目的是借用政策杠杆与法律高墙扼杀初创公司与开源生态。“把监管的不确定性当作打压对手的商业武器，是绝对不可接受的，”萨克斯警告道。 目前，白宫安全团队正在讨论的制裁与应对工具箱包括： - 将中国开源大模型相关实体与实验室列入贸易实体黑名单（Trade blacklists）。 - 针对使用中国 AI 模型的机构与企业发布国家安全风险预警。 - 酝酿出台专门限制在美调用开放模型的总统行政令（Executive Order）。 但在节目结尾，主持人借用与经济学家泰勒·考恩（Tyler Cowen）的辩论表达了怀疑：在推崇自由市场的国度，行政禁令真的能奏效吗？这就像当年打击数字盗版一样，全网封杀 Torrent 并没有彻底终结盗版，真正让盗版退场的是 Spotify、Apple TV 和 Netflix 带来的极致正版产品体验——只需付 2 美元买首歌或每月 10 美元订阅，消费者自然会用脚投票。就在节目当天上午，美国财政部长斯科特·贝森特（Scott Bessent）在福克斯新闻（Fox News）亮明了特朗普政府的官方立场：“本届政府坚定支持开源模型，但我们绝不容忍知识产权（IP）盗窃。” ### [09:18](https://www.youtube.com/watch?v=KRE39H7hyUI&t=558s) 美政府制裁威胁与开源 AI 的供应链困局 节目继续探讨针对海外模型的潜在制裁问题。主持人指出，如果美国政府采取严厉行动，其核心逻辑将是：“如果发现海外模型盗窃了我们伟大公司的成果，我们有权因这种盗窃行为对它们实施制裁。”从实际执行来看，政府可能会直接制裁这些模型背后的开发企业。 这种制裁措施预计会给美国本土开源模型带来巨大机遇。届时，许多新兴云服务商（Neoclouds）乃至 OpenRouter 等聚合平台，恐怕都将被迫停止托管或支持来自中国的开源模型。但这也引发了更棘手的现实拷问：在开源生态中，究竟该如何界定“违规”？如果某家海外机构发布了一篇研究论文，其他开发者依据该论文进行复现、训练、微调，或者利用其生成的数据来冷启动自己的模型，这种技术沿袭在供应链条上究竟要隔多少层才算“清白”？ 主播们在现场互相调侃起来。有人提及 Perplexity 此前曾极其迅速地基于 DeepSeek 推出了微调版本“1776 DeepSeek”；更有联合主播现场自爆：“我先用稀疏自编码器（Sparse Autoencoders）对 DeepSeek 做了微调，我是第一个搞的。”当被当场质问“你用的基座模型是哪国开发的”时，他只能坦白认领：“好吧，是 DeepSeek，它是中国的。”全场随即爆发出一阵哄笑，调侃这是现场的“自首认罪”。 ```mermaid flowchart TD accTitle: AI供应链合规与溯源困境 accDescr: 展示海外基座模型通过论文、权重微调或蒸馏流入生态后，监管层难以清晰界定违规边界的过程。 subgraph Origin["源头与技术输入"] A["海外基础模型 (如 DeepSeek)"] B["公开研究论文与架构设计"] end subgraph Derivative["衍生与二次开发"] C["数据提取与模型蒸馏 (Distillation)"] D["本地微调 / 稀疏自编码器 (SAE)"] E["合成数据冷启动本土模型"] end subgraph Distribution["服务分发层"] F["新兴云厂商 / 聚合平台 (如 OpenRouter)"] G["终端应用与衍生版本 (如 1776 DeepSeek)"] end A --> C A --> D B --> E C --> F D --> G E --> G ``` --- ### [10:41](https://www.youtube.com/watch?v=KRE39H7hyUI&t=641s) Anthropic 巨额版权和解案与“焚书”争议 话题随后转向美国本土科技巨头自身的版权软肋。针对财政部长候选人贝森特（Bessent）等人对海外盗窃知识产权的抨击，不少网友感到讽刺并指出：就在前一天，美国联邦法官刚刚批准了 Anthropic 高达 15 亿美元的版权侵权诉讼和解案。 现场主播梳理了这起和解案的荒诞背景。起初大家争论 Anthropic 到底是直接“偷书”还是“买书”，后来澄清事实其实包含两个层面：一方面，Anthropic 确实曾合法购入大量书籍用于模型训练，但随后将这些实体书全部销毁，这种“焚书”行径在公众层面引起了极大的情感反弹；另一方面，根据加州北区联邦地区法院威廉·阿尔苏普（William Alsup）法官此前的裁决，Anthropic 存在非法下载并存储数百万本受版权保护电子书的行为。 - 该和解案最终由另一位法官正式签署生效。 - 和解金总额达 15 亿美元，涵盖约 50 万部作品，相当于为每部受侵权作品赔付 3,000 美元，由持有版权的作者和出版商共同分享。 --- ### [12:00](https://www.youtube.com/watch?v=KRE39H7hyUI&t=720s) Kimi K3 蒸馏罗生门与模型语义相似度辩论 节目切入当时在社群引发热议的一张模型语义相似度矩阵图（源自账号 Lisan al Gaib 的转推）。该图通过对多个模型进行多轮 Prompt 采样，对比它们在措辞习惯、用词倾向等语义特征上的重合度。 分析显示，Kimi K3 与 Anthropic 的 Fable 5 表现出极高的语义相似度，同时也与 Opus 4.7、Opus 4.8 及 Sonnet 5 存在显著重叠，外界怀疑其训练数据可能融合了这四款模型的输出 Token。不过主播泰勒也客观指出，这类测试仅仅抽样了少数 Prompt，充其量能证明两者存在“相关性”，并不足以作为模型被蒸馏的铁证，因为相似的后训练（Post-training）技术也可能导致类似的结果。 有趣的是，数据展现出极具阵营色彩的模型族系特征： - **OpenAI 阵营内部沿袭**：GPT-5.4、5.5、5.6 与 Anthropic 旗下模型几乎毫无相似度；但 GPT 家族内部一脉相承，例如 5.6 Luna 与 5.4 呈现出高度自相关，反映了团队在前代成果上的持续迭代。 - **Anthropic 阵营风格自洽**：Fable 与 Opus、Haiku 保持着相同的叙事“韵味”。主播推测，这可能是 Anthropic 收集了大量用户在 Opus 上进行“氛围写代码”（Vibe Coding）的交互反馈与 Rollout 数据，并将其作为训练集反哺给 Fable 所致。 现场探讨了模型蒸馏的定性问题：模型蒸馏在本质上违反了各大厂商的服务条款（ToS），但若指望通过跨国法律诉讼去制裁海外公司，完全是天方夜谭。主播们戏谑道，任何试图跨国起诉被山寨产品的做法都毫无意义，“除非你能在中方领导人访问旧金山或联合国时，亲自冲上去把传票拍在他脸上高喊‘你被起诉了’”。 --- ### [14:50](https://www.youtube.com/watch?v=KRE39H7hyUI&t=890s) 鹰派反弹、George Hotz 炮轰与中美 AI 监管博弈 针对中国模型的崛起，华尔街知名评论员吉姆·克莱默（Jim Cramer）在社交媒体上大声疾呼，告诫美国企业切勿为了“省几个小钱”而去使用中国模型。他声称 OpenAI 和 Anthropic 的担忧事关国家安全底线，甚至搬出了宾·韦斯特（Bing West）新发布的《Cat 5》，极力渲染地缘对抗风险。 著名黑客、Tinygrad 创始人乔治·霍茨（George Hotz）随即以犀利言论强硬回怼：“这是在讲笑话吗？如果 Anthropic 能像中国厂商那样把模型权重开源并上传到 Hugging Face，我很乐意用他们的东西。美国科技巨头如今已沦为迷失方向的‘寻租机器’，继续为他们辩护毫无意义。他们如果想挽回人心，就必须彻底改变。” 与美国施压的同时，《金融时报》爆出中国商务部（MOFCOM）也在考虑收紧出口限制： - 中国商务部近期约谈了包括阿里巴巴、字节跳动和智谱在内的头部大模型企业，商讨限制用于模型训练的关键数据出境。 - 监管部门正评估是否限制海外直接下载模型权重。未来中国可能依然允许海外客户通过 API 访问大模型及相关云服务，但在模型权重的彻底开源上态度正趋于保守，下周发布的 Kimi K3 是否真正开源权重将成为关键风向标。 - 政策分析人士迪恩·鲍尔（Dean Ball）对此发表了“胜利宣言”，指出自己此前对中美两国针锋相对走势的预判再度应验，证明这纯粹是对现实规律的客观推演，而非个人价值预设。 --- ### [16:54](https://www.youtube.com/watch?v=KRE39H7hyUI&t=1014s) Substack 上线 Pangram 识别功能与代笔伦理 话题转向创作者生态的“垃圾内容之战”（Slop Wars）。Substack 正式宣布集成 AI 检测工具 Pangram。未来读者和作者可以在 Substack 应用中直接扫描文章、回复及评论，系统会自动测算内容由人类撰写还是 AI 辅助生成的比例。Substack 官方宣称，此举关乎其“打造文化经济引擎”的核心使命。 主播们开启了互损模式，调侃这项功能对泰勒简直是灭顶之灾，因为泰勒早已掌握了把人话写得像 AI 一样的绝技。泰勒曾手写过诸如“根据多项研究，你说的完全正确”这类充满 AI 味的套话，结果被 Pangram 判定为 100% 纯 AI 生成。不过主播也直言，受众的“显示性偏好”（Revealed Preference）往往很诚实：只要内容好看有用，大部分人其实并不真正在意是不是 AI 写的。 随后，现场引发了一场关于“署名真实性”的深度讨论：如果说用 AI 需要被检测，那么 Substack 作者聘用真人代笔（Ghostwriter）是否也必须强制公开？ - 媒体运作模式：彭博社（bloomberg.com）上的文章绝非迈克尔·布隆伯格本人撰写，但他们会明确标注具体记者的署名（如 Joe Weisenthal）。 - 网红个人 IP 现状：许多新兴意见领袖早在前 AI 时代就普遍雇佣枪手。有主播分享自身经历，称过去制作 YouTube 视频时团队内也有专职撰稿人负责初稿打磨，虽然会在片尾致谢中署名，但绝不会在视频一开篇就郑重声明“本期内容由他人代笔”。 - 传统出版界惯例：正如直播间弹幕所指出的，绝大多数知名图书作者雇佣影子写手早已是行业潜规则，而且几乎从不对外承认。 ### [18:45](https://www.youtube.com/watch?v=KRE39H7hyUI&t=1125s) 影子写手、Substack 创作团队与戴夫·艾格斯的“反 AI 小说论” 两位主持人继续探讨文字创作背后的真实性与披露机制。他们指出，在很多出版物中，书页上往往会写着“由某某与某某共同撰写”，但这在很大程度上取决于“影子写手”（枪手）自身的气场和影响力；部分枪手确实能在书封或版权页获得署名，但许多畅销书作家在去世后，其同名品牌依然能持续不断地推陈出新——比如著名军事小说家汤姆·克兰西（Tom Clancy）。这些作者的个人 IP 已经庞大到足以支撑起一整套商业运作体系。主持人推测，机场书店里摆放的图书中，大概有 30% 纯粹是完全未予披露的代笔作品。对于 Substack 等创作者平台，他们认为并不强求作者大张旗鼓地把所有幕后细节置顶，只要读者想深究时能查证得到、创作者坦诚披露自己背后有协助团队即可。 正是在这种对内容原创性的广泛焦虑中，美国著名作家、编辑兼出版人戴夫·艾格斯（Dave Eggers，曾出版畅销回忆录《怪才的荒诞与疾痛》并入围普利策奖，同时著有反思科技巨头的知名小说《圆圈》）在《金融时报》（Financial Times）上撰文断言：“轻信而荒谬，没有人会去读 AI 写的小说。”对此，主持人忍不住吐槽表示，事实上现实中已经有大量读者在消费 AI 创作的小说了，艾格斯所谓的“没有人”，大概仅仅是指“那些不够酷、或者不符合他品味的人”。 - 戴夫·艾格斯（Dave Eggers）：《金融时报》撰文称“AI 小说是荒唐的，无人会读”，但现实中 AI 生成读物已在特定读者群中悄然蔓延。 - 机场书店生态：被估测约有三成畅销书籍实为隐形枪手代笔，成为一种默许的流水线商业现象。 --- ### [20:10](https://www.youtube.com/watch?v=KRE39H7hyUI&t=1210s) 逃离高昂房租：硅谷极客的新风潮“船屋生活”与海上家园 话题随即转向了《金融时报》对戴夫·艾格斯的采访现场描写：记者登上艾格斯停泊在旧金山金门大桥阴影下的 23 英尺帆船，船尾放着油桶和渔网，周遭空旷安静，最近的卫生间需要骑一小段自行车去用移动厕所。文章特别强调，这种帆船的泊位费用微乎其微，堪称在超级富豪云集的硅谷生活成本最低的方式之一。这一描写立刻激起了主持人的极大兴趣，他们将此戏称为科技从业者“闭关专注”（Lock-in）的终极生活黑客技巧。安东尼奥·加西亚·马丁内斯（Antonio García Martínez）在畅销书《混沌猴子》（Chaos Monkeys）中就曾详述过住船上的经历。主持人甚至开始畅想，未来推特上那些渴望打造“估值 250 亿美元单人独角兽公司”的创业者们，不再晒车库照片，而是争相晒出自己买船住船的动态。 现场甚至直接打开了二手船交易网站 Boat Trader 算起了一笔经济账：旧金山普通的单居室公寓月租金动辄高达两三千美元，而一艘配备齐全的 2022 年或 2024 年全新豪华船屋，首付数万美元后，按揭加上泊位月供可能只需一两千美元。他们随后观赏了一款名为“La Mer Modern 11”的 36 英尺现代船屋实拍视频： ```mermaid flowchart TD accTitle: 现代船屋 La Mer Modern 11 核心配置与居住要素 accDescr: 图表展示了科技从业者选择船屋生活时所关注的动力性能、居住空间与生活舒适度等技术规格。 subgraph Spec["核心硬件与动力"] A["双雅马哈 25 高推力引擎<br/>共 50 马力"] B["泡沫填充复合浮筒<br/>+ 镀锌钢底架"] end subgraph Living["船舱空间与布局"] C["两间独立客舱 + 一间卫浴<br/>支持长年全天候居住"] D["顶层露台甲板<br/>Flexiteek 仿柚木地板"] end subgraph Comfort["生活配套与防寒"] E["逆循环冷暖空调"] F["地暖系统 + LED 氛围照明"] end A --- Living B --- Living Living --- Comfort ``` 看到船上不仅配齐了地暖、空调，甚至拥有豪华观景甲板，主持人们兴奋地调侃：科技界“车库创业”的时代已经彻底终结，下一家伟大的初创企业注定将在一艘船屋上诞生。他们戏谑这是彼得·蒂尔（Peter Thiel）倡导的“海上家园”（Seasteading）理论在旧金山湾区的平民化实践。弹幕中也有观众起哄道：“我得买一艘双倍宽度的船屋，在船上走私蒸馏 Anthropic 大模型！” - 船屋成本效益：一艘约 20 万美元的 2022 年款船屋，通过贷款月付约 1,483 至 2,259 美元，远低于旧金山市区每月 2,000 至 3,000 美元的单间房租。 - 硬件指标：La Mer Modern 11 长 36 英尺，配 50 马力双引擎、两房一卫、逆循环空调、地暖及复合稳定船体，足以满足常年居住需求。 --- ### [24:14](https://www.youtube.com/watch?v=KRE39H7hyUI&t=1454s) 飞机靠背该不该放倒？长途机舱里的“当代道德困局” 节目随后转向了另一个引发全网激烈站队的话题：“坐飞机到底应不应该把座椅靠背往后放倒？”主持人引述了《纽约客》（The New Yorker）撰稿人约书亚·罗斯曼（Joshua Rothman）描述的一个典型长途航班困境：在长达 15 小时的经济舱飞行中，前面的乘客将靠背猛地放倒到底，导致你的活动空间被极度压缩，甚至连脸前都悬着别人的椅背；当你忍无可忍正打算也按下放倒按钮时，身后带着哭闹孩童的父亲却轻咳一声，恳求你“能不能别放倒椅子”。 这一情景在演播室立刻引发了截然相反的激烈交锋： | 讨论阵营 | 核心观点与逻辑 | 现实行为与心理 | |:--- |:--- |:--- | | **支持放倒派** | 航空公司的座椅设计了倾斜按钮，赋予了乘客这项物理权利；笔直坐着极其难受，机舱内人人放倒才符合“空间守恒”与最大舒适原则。 | 从小到大只要安全带指示灯熄灭，放倒座椅就是理所当然的标准动作，不用按钮等同于“上飞机不上厕所”。 | | **反对放倒派** | 侵犯了后排乘客极为狭窄的个人物理边界，属于反社会自私行为；“有权利去做”绝不等于“道德上应该去做”。 | 机舱空间拥挤脆弱，主动克制不放倒靠背是对后排乘客的尊重与机舱文明礼貌。 | 在此期间，弹幕听众还分享了他们眼中的飞机“硬核自私操作”——比如飞机一接地滑行完毕、安全带灯刚一响就立刻弹起并猛冲到机舱前排的行径，被两位主持人痛斥为极其粗鲁的反社会行为。随后他们进一步发散到更多机舱争议场景：在机舱里公放 JBL 蓝牙音箱，或者随身携带类似 OpenAI 打造的便携智能设备并在整个航程中喋喋不休地对着它说话，是否属于严重冒犯？ 最后，他们引用了历史数据来说明公众对这一问题态度的剧烈变化。2014 年著名数据民调网站 FiveThirtyEight（由 Nate Silver 创立）曾对大约 1,000 名常客进行调查，当时有 41% 的受访者认为放倒椅背是粗鲁行为，反对者仍占少数；然而到了 2020 年，数学教授埃里克·琼斯（Eric Jones）在旅游统计网站 The Vacationer 上开展的一项类似民调却显示，反对放倒座椅的人数比例已经飙升至 77%。对此，力挺放倒权利的主持人无奈感叹现在的乘客简直变成了“顺从的绵羊”，并开玩笑提出折中方案：“只有当你穿着得体优雅时才有资格放倒座椅，要是穿着运动裤邋里邋遢登机，那就不准放倒。” ### [27:53](https://www.youtube.com/watch?v=KRE39H7hyUI&t=1673s) 飞机礼仪闲聊：后座小孩踢椅背与舱内生存哲学 围绕飞机上的各种失礼行为，主持人们继续展开了轻松幽默的争论。关于乘客是否应该随意调整后仰椅背，有人认为既然别人不后仰，你也不能随便躺平；但话题转到“小孩是否能随心所欲踢前面座椅”时，主持人的态度却出人意料地宽容。他表示自己完全能忍受被身后的小朋友踢椅背，因为带孩子坐飞机对父母和孩子而言都无异于一场折磨——小孩子被活生生困在封闭的客舱铁皮盒子里，本身就已经够痛苦了。 面对这种状况，他建议成年乘客干脆“硬扛过去就得了”：座椅明明有厚达数英寸的缓冲海绵垫，除非你有着像罗纳尔多一样矜贵敏感的后背，否则根本不用大惊小怪。他甚至调侃道，不妨在飞机上点一个滚烫的微波披萨口袋（Pizza Pocket），坦然享受这种身处“永久底层”的随性与狼狈，在平庸甚至混乱的旅途中“看开一点，好好享受生活”。 ### [28:27](https://www.youtube.com/watch?v=KRE39H7hyUI&t=1707s) Ramp 推出 AI 模型路由器：CFO 治理 Token 浪费的“神级”产品扩张 随后，话题切入当日的科技重头戏——企业财务平台 Ramp 正式推出了自己的 AI 模型路由器（Model Router）。风投投资人兼评论员 Brexton 对此给予了极高评价，称“Ramp 推出模型路由器简直是神级（god tier）的产品扩张”。当前市场已经被过度狂热的“Token 刷量”（token maxing）、层出不穷的新模型繁衍以及各类算力主权概念折磨得疲惫不堪。在业内看来，这不仅是一次功能试水，更是少有的让人眼前一亮、极具成为全新核心产品潜力的战略级发布。 这项新业务在逻辑上与 Ramp 自身的定位极为契合。Ramp 品牌立足的核心承诺就是“为企业省时省钱”，而当下的各大企业首席财务官（CFO）正为混乱失控的 LLM Token 开销与路由策略大为恼火。正如主持人们上周与 Eric Lyman 讨论时所提到的：如果员工随手问个今天天气如何，后台却调用最昂贵的前沿旗舰大模型、疯狂燃烧数百万 Token，那这种查询甚至应该直接贴上一张费用报销单才对。 ```mermaid flowchart TD accTitle: Ramp 模型路由器工作逻辑 accDescr: 展示用户查询如何经由 Ramp 模型路由器分流至轻量模型或前沿模型，从而优化企业成本。 subgraph UserRequest["企业用户请求"] Q1["简单查询（如天气）"] Q2["复杂推理任务"] end subgraph RampRouter["Ramp 模型路由器"] Router["智能成本与性能路由"] end subgraph LLMEngines["后端模型层"] Cheap["轻量经济模型 / 低 Token 成本"] Frontier["前沿旗舰大模型 / 高算力消耗"] end Q1 --> Router Q2 --> Router Router -->|低复杂度| Cheap Router -->|高复杂度| Frontier ``` - **业务协同性**：精准切入企业降本增效痛点，正如同行评价的那样，Ramp 正在“吞噬一切与时间和金钱相关的事物”。 - **测试与开放范围**：该模型路由器目前已在部分客户中进入 Alpha 内测阶段；据悉，该产品甚至对非 Ramp 传统企业客户开放，用户可以单独采用其模型路由工具。 节目在轻松的告别声中收尾，主持人们呼吁听众在 Apple Podcasts 和 Spotify 上打出五星好评，并订阅 tbpn.com 每日简报。伴随着“扔闪光弹了”、“向马扔闪光弹”的无厘头插科打诨与标志性片尾音乐，本期节目正式落下帷幕。 ## Notable quotes > [00:02](https://www.youtube.com/watch?v=KRE39H7hyUI&t=2s) 美国企业究竟该不该获准使用中国的人工智能模型？ — The Wall Street Journal > [02:26](https://www.youtube.com/watch?v=KRE39H7hyUI&t=146s) 开放权重模型主导的世界，其最可能的一种结局是‘完全的 AI 共产主义’（full AI communism）。这恰恰是中国所主张的路线——AI 不再是市场化商品，而是一种公共物品，最终由国家作为某种‘数字公共基础设施’直接提供。 — 迪恩·鲍尔（Dean Ball） > [06:14](https://www.youtube.com/watch?v=KRE39H7hyUI&t=374s) 把监管的不确定性当作打压对手的商业武器，是绝对不可接受的 — 大卫·萨克斯（David Sacks） > [06:14](https://www.youtube.com/watch?v=KRE39H7hyUI&t=374s) 本届政府坚定支持开源模型，但我们绝不容忍知识产权（IP）盗窃。 — 斯科特·贝森特（Scott Bessent） > [09:18](https://www.youtube.com/watch?v=KRE39H7hyUI&t=558s) 如果发现海外模型盗窃了我们伟大公司的成果，我们有权因这种盗窃行为对它们实施制裁。 > [14:50](https://www.youtube.com/watch?v=KRE39H7hyUI&t=890s) 这是在讲笑话吗？如果 Anthropic 能像中国厂商那样把模型权重开源并上传到 Hugging Face，我很乐意用他们的东西。美国科技巨头如今已沦为迷失方向的‘寻租机器’，继续为他们辩护毫无意义。他们如果想挽回人心，就必须彻底改变。 — 乔治·霍茨（George Hotz） > [18:45](https://www.youtube.com/watch?v=KRE39H7hyUI&t=1125s) 轻信而荒谬，没有人会去读 AI 写的小说。 — 戴夫·艾格斯（Dave Eggers） > [28:27](https://www.youtube.com/watch?v=KRE39H7hyUI&t=1707s) Ramp 推出模型路由器简直是神级（god tier）的产品扩张 — Brexton ## Mentioned OpenAI, Anthropic, 萨姆·奥特曼（Sam Altman）, 达里奥·阿莫代伊（Dario Amodei）, 米拉·穆拉蒂（Mira Murati）, Thinking Machines Lab, 月之暗面（Moonshot AI）, Kimi K3, 阿里巴巴, Qwen 3.8 Max, 迪恩·鲍尔（Dean Ball）, 乔治·霍茨（George Hotz）, 大卫·萨克斯（David Sacks）, 斯科特·贝森特（Scott Bessent）, 吉姆·克莱默（Jim Cramer）, 泰勒·考恩（Tyler Cowen）, 戴夫·艾格斯（Dave Eggers）, 英伟达（Nvidia）, Nemotron 3 Ultra, Google, Gemini 3.5 Flash Light, Gemini 3.5 Flash Cyber, Mythos, GPT Cyber, DeepSeek, Perplexity, Substack, Pangram, Ramp, OpenRouter, 派拓网络（Palo Alto Networks）, 尼克什·阿罗拉（Nikesh Arora）, 安东尼奥·加西亚·马丁内斯（Antonio García Martínez）, 彼得·蒂尔（Peter Thiel）, FiveThirtyEight, The Vacationer, 《华尔街日报》（The Wall Street Journal）, 《金融时报》（Financial Times） ## 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=KRE39H7hyUI)

[0:02] The Wall Street Journal: 美国企业究竟该不该获准使用中国的人工智能模型？ [2:26] 迪恩·鲍尔（Dean Ball）: 开放权重模型主导的世界，其最可能的一种结局是‘完全的 AI 共产主义’（full AI communism）。这恰恰是中国所主张的路线——AI 不再是市场化商品，而是一种公共物品，最终由国家作为某种‘数字公共基础设施’直接提供。 [6:14] 大卫·萨克斯（David Sacks）: 把监管的不确定性当作打压对手的商业武器，是绝对不可接受的 [6:14] 斯科特·贝森特（Scott Bessent）: 本届政府坚定支持开源模型，但我们绝不容忍知识产权（IP）盗窃。 [9:18] 如果发现海外模型盗窃了我们伟大公司的成果，我们有权因这种盗窃行为对它们实施制裁。 [14:50] 乔治·霍茨（George Hotz）: 这是在讲笑话吗？如果 Anthropic 能像中国厂商那样把模型权重开源并上传到 Hugging Face，我很乐意用他们的东西。美国科技巨头如今已沦为迷失方向的‘寻租机器’，继续为他们辩护毫无意义。他们如果想挽回人心，就必须彻底改变。 [18:45] 戴夫·艾格斯（Dave Eggers）: 轻信而荒谬，没有人会去读 AI 写的小说。 [28:27] Brexton: Ramp 推出模型路由器简直是神级（god tier）的产品扩张

Dean Ball: One probable outcome of an openweight model dominant world is full AI communism, which is precisely what China proposes rather than a market product. AI is a public good, which will ultimately be provided by the state as a kind of digital public infrastructure. David Sacks: The weaponization of regulatory uncertainty as a competitive tool should be completely unacceptable. George Hotz: rent extraction machines that have lost the plot [16:54] Tyler: According to studies, you're absolutely right [15:37] Ministry of Commerce of the People's Republic of China [28:27] Brexton Bennett: god-tier product expansion

[0:02] China, China, China, China is in the [0:04] news. Top American execs are sounding [0:08] alarms. The White House is divided on [0:10] how to respond to recent advances in [0:12] Chinese AI has weighed crackdown [0:14] measures. Uh we we talked about how this [0:16] was coming. It is interesting. There's a [0:19] there's there's a lot of like hurry up [0:21] and pedal to the metal and then pull [0:23] back and then ban the American model and [0:25] then are we going to ban the Chinese [0:27] model in the similar way? Uh, and lots [0:30] of contradictions, lots of insider [0:32] baseball, everyone's critiquing everyone [0:34] else. Lots of interesting takes. Also, [0:36] just some interesting data about what's [0:38] actually going on with these models. Uh, [0:41] how advanced are they? Where do they [0:42] come from? How cheap are they to make? [0:45] What are the economic impacts in [0:46] America? But the Wall Street Journal has [0:48] a nice little roundup here we can take [0:49] you through. So, Silicon Valley and [0:51] Washington are debating a multi-billion [0:53] dollar question. Should American [0:55] companies be able to use Chinese [0:57] artificial artificial intelligence [0:59] models? They certainly have been able [1:01] to. Hasn't been a problem for years, but [1:03] now they're getting stronger. Open AI [1:05] and anthropic executives are sounding [1:07] the alarm about the rise of cheap AI, [1:10] particularly powerful new models [1:12] produced in China, suggesting that they [1:14] will lead to quote a dystopian AI future [1:17] and present unacceptable security risks [1:19] without regulation. And that is the [1:21] point of debate. Are you on the dis the [1:23] side of dystopian open source or [1:26] liberating open source? The George Hans [1:28] frame. [1:28] >> Uh some analysis some analysts who study [1:31] the AI industry say the two companies [1:33] which are preparing for public listings [1:34] in the next year. They just want to [1:36] eliminate the competition. The emergence [1:38] of highly capable open autonomous AI [1:41] systems including Moonshot AI's Kimmy K3 [1:44] model and Alibaba's Quen 3.8 Max. And [1:48] Alibaba is actually a massive uh [1:50] investor in Moonshot. They have their [1:52] own anthropic Google over there going [1:55] on. Alibaba is like very much their the [1:58] the the Google too uh uh in the sense [2:01] that they they they have a large [2:03] position there. Uh which were released [2:05] in recent days and viewed favorably by [2:07] investors and users has turned the AI [2:09] race on its head once again. Kimmy K3 [2:12] was also competitive with US models on [2:14] some benchmarks. Uh the debate over the [2:15] new models which are quote openweight [2:17] but not yet. They're they're about to [2:19] open source them. We'll see what happens [2:20] next week. allowing users to download [2:22] and customize them with company data for [2:24] specific tasks. Very, very good. Very [2:26] useful. Coincides with division in the [2:28] Trump administration about whether to [2:30] take steps to limit the use of models in [2:32] the United States. Quote, "One probable [2:34] outcome of an openweight model dominant [2:36] world is full AI communism, which is [2:40] precisely what China proposes rather [2:41] than a market product. AI is a public [2:44] good, which will ultimately be provided [2:46] by the state as a kind of digital public [2:47] infrastructure." Dean Ball, OpenAI's [2:49] head of strategic futures said in an ex [2:51] post Friday, Dean Ball made it to the [2:53] pages of the Wall Street Journal. Now, [2:54] um the AI communism is just at the model [2:57] level. Like there would still be a bunch [2:59] of like a lot of people who are long [3:02] compute, long neoc. They say like this [3:05] is where the value should acrue. The [3:07] data centers will acrue a reasonable [3:10] margin and they will still continue to [3:13] make money. And so um it's not that this [3:16] automatically leads to nationalization [3:18] of compute in any way. People would [3:21] still be paid uh but uh the but the [3:24] margins might be lower is the is at [3:26] least the formulation here. So B a [3:28] former Trump administration official [3:30] called such a scenario a dystopian [3:32] hellscape and he said he expected the [3:34] Trump administration to make moves [3:36] toward reducing the use of open source [3:38] models. Ball later clarified that he [3:40] wasn't advocating for the United States [3:42] government to discourage Chinese AI. He [3:44] also highlighted a central challenge [3:45] with the AI race. Top tier AI companies [3:47] raise billions of dollars to pay for the [3:49] vast computing resources needed to [3:51] continue improving AI systems. If [3:52] everyone uses AI systems that people [3:54] largely don't pay for, there would be no [3:56] way to finance continued frontier AI [3:58] development. Ball, who started at the [4:00] company earlier this month and focuses [4:02] on policy and strategy and open AI, have [4:04] said his comments don't reflect the [4:05] company's official positions. He's just [4:07] posting. The company has developed its [4:08] own models, own open models, and chief [4:11] executive Sam Alman had previously said [4:12] that OpenAI's past approach of only [4:14] developing closed tools was unwise. [4:16] Other top AI executives have warned that [4:18] open models present huge risk since [4:21] developers and policy makers lack [4:22] control over how they are used and [4:24] modified. A lot of risk of cyber [4:26] attacks, a lot of risk of annoying spam [4:28] attacks. The cyber thing is very [4:32] concerning and annoying if that really [4:35] ramps up. But the good news is that the [4:36] closed source models have been available [4:38] to cyber security firms and a lot of big [4:41] major companies. The finance industry [4:43] for example has had access to mythos and [4:44] GPT cyber for almost six months. And [4:46] then just today, Google announced a [4:49] specific model uh flashlight cyber I [4:52] think. So they're trying to offer a [4:54] similar capability but really go down [4:56] the cost curve because a lot of the [4:58] cyber tools there was incredible [5:00] breakthroughs with mythos GPT cyber like [5:03] great benchmarks on those uh but could [5:05] be very expensive. We heard uh I think [5:08] wasn't it Nesh Aurora said I think he [5:10] spent like a million dollars to find [5:11] some zeroday bug and it was like totally [5:13] worth it but that's expensive for [5:15] smaller companies that might just need [5:17] uh it would be great to just be able to [5:18] go to a cheap cyber tuned model get [5:21] approved by Google and say hey just go [5:23] and read every line of code and just let [5:25] me know if there's anything that's [5:26] obvious and uh that could be good what [5:28] >> yeah and just be clear they released 3.5 [5:31] flash light and 3.5 flash cyber so it's [5:33] not like flash light cyber [5:35] >> okay but flash is still cheaper than the [5:37] big model, right? [5:38] >> There's like the pro and there's [5:39] >> and so I have to imagine that if you I [5:42] given the choice if you're like a [5:43] midsize company uh and and and you're [5:47] and your choice is like go beat down the [5:49] door to get into Project Glass Wing to [5:51] spend $5 million to have Mythos read [5:54] every line of code or you're like, "Oh, [5:56] well, I'm already a Google partner. I [5:58] have a really solid business [6:00] relationship with them because I've been [6:01] buying ads and I have Google web master [6:03] tools set up and I just want to go set [6:06] up flash cyber, not flash light cyber. I [6:09] imagine that the cost would be cheaper [6:11] because they've fine-tuned this model [6:12] for this specific thing. So, Anthropic [6:14] CEO Dario Amade for years has worry [6:17] warned about the risks of powerful and [6:19] open AI systems, saying in a recent [6:21] Bloomberg interview that having AI [6:23] models with advanced cyber security [6:25] capabilities that are free to download, [6:27] could be harmful. He said it's a serious [6:29] concern. Use of Chinese models, which [6:31] are far cheaper than US counterparts, is [6:33] surging at US companies, prompting some [6:34] investors to question the staying power [6:36] of top model makers such as Anthropic [6:38] and Open X. The current US open- source [6:40] frontier is starting to catch up with [6:42] China's. On Wednesday, Thinking Machines [6:44] Lab, led by former AI technology chief [6:46] Mirror Morati, released its f first AI [6:49] model as openweight. Nvidia's Neotron 3 [6:52] Ultra is starting to see traction. [6:54] That's interesting to hear. I haven't [6:55] heard nearly enough buzz around Neotron. [6:58] I know Jensen has been investing in it. [7:01] He's also investing in a lot of up and [7:03] down the stack. [7:04] >> I've been hearing buzz about the name, [7:06] John. [7:06] >> Neotron's a good name. [7:07] >> Neotron. Why? Very cool name. There [7:10] people are, you know, there's there's [7:11] some talks bubbling. People are saying [7:13] it's it's it's one of the best names for [7:16] for a family of models ever. [7:18] >> What comes after Ultra? Isn't that what [7:20] the the Didn't Apple first do that with [7:22] the the M4 Pro, the M4 Max, the M4 [7:26] Ultra. So, we got to come up with an [7:28] even better word for what happens after [7:31] Ultra Giga maybe. And then EXA. The [7:34] market's faith in anthropic and open AI [7:36] continuing to build more cable models [7:37] that push the frontier AI frontier has [7:39] been the at the heart of the boom [7:41] helping justify trillions of dollars in [7:42] spending on infrastructure in the coming [7:44] years. The threat that new players will [7:46] vastly undercut what they can charge for [7:48] advanced AI push down some tech and AI [7:50] company stocks prices last week. David [7:52] Saxs, the venture capitalist and White [7:54] House AI advisor, suggested that he read [7:58] Ball's post as a confession to a [8:00] regulatory capture strategy. Saxs and [8:02] others have long seen calls for AI [8:04] regulation by companies such as [8:06] Anthropic as efforts to use new laws and [8:08] policies to sty competitors. The [8:10] weaponization of regulatory uncertainty [8:12] as a competitive tool should be [8:14] completely unacceptable, he said. [8:16] Security focused administration [8:17] officials in the past year have [8:18] considered a slew of measures to crack [8:20] down on Chinese AI companies developing [8:22] open models. People familiar with the [8:24] discussion said they've weighed adding [8:26] the businesses to trade blacklists, [8:29] security warnings on Chinese AI [8:30] companies, and a potential executive [8:32] order targeting open models. So, it does [8:34] seem like even though Tyler Cowan, we [8:36] were going back and forth with him about [8:37] this about like can you even ban it? You [8:39] can't really ban anything in the land of [8:42] the free. Like the solution to piracy [8:45] was not really to ban piracy. Like you [8:47] can still torrent things, but it's just [8:50] so easy to use Apple TV and Spotify and [8:53] Netflix that most people just wind up [8:56] just paying $5, $2 for a song or or $10 [9:00] a month for something that it [9:02] effectively killed it just because the [9:04] product experience was better. This [9:06] morning, uh, our Treasury Secretary, uh, [9:10] Besson was on Fox and said, "This [9:12] administration supports open source [9:14] models, but what we do not support is IP [9:17] theft. [9:18] >> If we see, especially that overseas [9:20] models are stealing from our great [9:22] companies, we have the ability to [9:23] sanction them because of this theft." So [9:27] it will be if the admin does take [9:30] serious action here, I would expect that [9:32] to be the line of reasoning where they [9:34] could just sanction the companies behind [9:37] these models [9:38] >> which I imagine would create a real [9:40] opportunity for American open source. [9:42] You would ultimately see uh a lot of [9:45] these you know neoclouds unable I [9:48] imagine to uh serve the you know [9:52] traditional [clears throat] [9:53] >> Chinese open source models. I imagine [9:55] that even open router would probably [9:57] have to [9:58] >> yeah maybe [9:59] >> stop supporting that [10:00] >> but again with an open source model like [10:05] how do you even define it because if [10:07] there's a research paper and then you go [10:08] and implement that and you train it or [10:10] you take it you fine-tune it or you use [10:12] the data to bootstrap your own model [10:13] like just like where where in the supply [10:16] chain how many steps away do you need to [10:18] be to like not count on there? Didn't [10:21] Perplexity do like a 1776 Deep Seek very [10:24] quickly? They did a fine tune. I think [10:25] you did a fine tune on Deep Seek as [10:27] well. [10:27] >> Yeah. [10:29] Sparse autoenccoders. Okay. I did it [10:31] first. So, [10:31] >> but but and and what what nationality [10:33] was the model that you used as the base? [10:35] >> Well, it was Deep Seek. IT WAS CHINESE. [10:36] >> WOW. CONFESSION. Uh you can go now. Um, [10:41] yeah, people are are already angry at at [10:43] at Besson's comments uh in our and are [10:47] saying funny that he's not mentioning [10:50] >> uh that yesterday US judge approved [10:52] Anthropic's $1.5 billion copyright [10:55] lawsuit settlement for [10:58] >> I don't think they they I think they [11:00] settled that they were in possession of [11:02] books that had been torrented or or um [11:06] basically acquired unlawfully. Yeah, [11:08] >> I don't believe that. You're unlawfully [11:11] sound like they stole the book. I think [11:13] they bought the books. [11:14] >> No, no, no, no, no. I think this was No, [11:16] they they separately they bought a bunch [11:18] of books, [11:19] >> trained on them and destroyed them. [11:20] >> And that just made people upset because [11:22] no one idea of books being destroyed, of [11:24] course. Yeah. [11:25] >> Um I think those were rare like rare [11:27] rare books. [11:28] >> The Wilmanitis collection was destroyed. [11:31] Rare PDFs. Judge William Alup of the US [11:35] District Court for the Northern District [11:36] of California issued a preliminary [11:38] approval of the settlement last year [11:40] after ruling that Enthropic had [11:42] illegally downloaded and stored millions [11:44] of copyrighted books. And the settlement [11:46] was signed off on by a different judge [11:49] yesterday. The payout will deliver [11:50] $3,000 per work across an estimated [11:52] 500,000 works shared among the authors [11:55] and publishers who hold rights to them. [11:58] Anyway, there was some news about like [12:00] whether or not uh Kimmy K3 is distilled. [12:04] Tyler, can you take us through this [12:06] Lissan Algib post? [12:08] >> Yeah, so he he didn't actually make this [12:10] chart. It was from someone else, but [12:11] basically this shows a bunch of models [12:13] compared to another and it's like the [12:14] kind of uh semantic similarities between [12:16] them. Okay, so it's like if you use [12:18] similar kind of diction or or phrases, [12:20] stuff like this, [12:21] >> um you just like pump the models a bunch [12:23] of times and and you kind of can tell [12:24] what's similar to each other. Yes. [12:26] >> And you can see in this version it looks [12:28] like like Kimiki 3 is is very similar to [12:30] Fable 5. [12:31] >> Fable 5 and then also Opus 4.7, Opus 4.8 [12:34] and Sonnet 5. So they might have gotten [12:35] tokens from like all four of those [12:37] models sort of mixed. [12:38] >> Also I think this exact example like [12:41] only like a few prompts were used. It's [12:42] not like super involved. It's also like [12:44] this is very much just like correlation. [12:46] This this is not like proving that they [12:47] still learn anything. They could have [12:48] just used like similar you know post- [12:50] training techniques or something like [12:51] this. [12:52] >> Totally. But notably uh GPD 5.4, 5.5, [12:57] 5.6 share very little similarity to the [13:01] anthropic models and vice versa. The [13:04] enthropic models don't seem to be [13:06] correlated with the OpenAI models. [13:09] Although the OpenAI models are [13:11] correlated with them with with each [13:13] other. So, uh 5.6 Luna looks pretty [13:16] correlated with 5.4 for because they [13:19] probably were like, well, we're building [13:20] on the shoulders of the work that we've [13:22] been done. And and it's the same thing [13:23] with anthropic where Fable has some of [13:26] the same flavor as an Opus or a Haiku. [13:29] And that makes sense. [13:29] >> That makes sense because people have [13:31] like been saying, oh, maybe those models [13:33] were actually distilled from the the big [13:34] like mythos pre-train, right? [13:36] >> Yeah. Or orthos [13:38] or vice versa. you have a bunch of users [13:40] that are doing rollouts on Opus trying [13:43] to vibe code different things and you [13:45] take all that training data and then you [13:46] feed that into Fable and Fable sort of [13:48] like picks up on a lot of that because [13:50] that's what the the the user said, "Wow, [13:54] you really did a great job on this opus [13:56] or you did really poorly." It pulls that [13:58] out, but it's still learning the the the [13:59] textual similarity. Well, people will uh [14:02] go back and forth on this and uh and [14:05] debate the level of distillation and [14:07] also what the punishment for [14:09] distillation should be. Should it just [14:11] be a slap on the wrist or should it be [14:13] something bigger? [14:13] >> The only thing I would say here is like [14:15] distillation against against the terms [14:18] of service. [14:19] >> Yeah. [14:19] >> And I think that suing Chinese companies [14:23] Yeah. [14:23] >> it would be a complete and total waste [14:25] of time. Ask anyone who's built a [14:28] product that's been knocked. [14:29] >> I disagree. Incredible. or a farm. If [14:32] [laughter] you if you personally serve [14:35] Cunping the next time he's in San [14:37] Francisco uh meeting with a a world [14:40] leader, you show up. [14:41] >> You have been served. [14:42] >> You show up to the UN, [14:44] >> serve the world leader. There's [14:46] something about that that's, you know, [14:48] extremely bold. [laughter] [14:50] >> But no, I agree with you. Uh very [14:53] difficult. I don't even know if you went [14:55] to your law firm and you said like, [14:56] "Hey, we want to sue this Chinese [14:58] company." Like how they would trace that [14:59] down. Jim Kramer says, "We must not let [15:02] our companies use these Chinese models [15:04] to save a few bucks. Open AI and [15:06] anthropic are correct. This is vital [15:08] national security. Please break. Please [15:10] read Bing West's just released Cat 5. I [15:14] respect the Chinese people greatly, but [15:16] these companies are run by the PLA for [15:17] heaven's sakes." George Hotzfire's back [15:19] under the tiny grad handle. He says, "Is [15:22] this a joke? I'd be happy to use [15:23] anthropic if they put their weights on [15:25] hugging face like the Chinese. American [15:27] tech companies are rent extraction [15:28] machines that have lost the plot. No [15:31] point continuing to defend them. They [15:33] need to change if they want to win [15:35] people back. [15:36] >> Meanwhile, [15:37] >> China is also thinking about export [15:39] controls [15:40] >> in the Financial Times. Apparently, [15:42] Alibaba, Bite Dance, and keep [15:45] >> going back and forth. Moffcom, [15:46] >> Moffcom talked to AI companies including [15:48] Alibaba, Bite Dance, and Zepoo on [15:50] limiting the transfer of key data for [15:52] the training of their models overseas, [15:54] >> as well as allowing their model weights [15:55] to be downloaded by [15:56] >> Ministry of Commerce, probably their [15:58] Department of Commerce. [15:59] >> That seems more like it. China would [16:01] still let overseas customers access the [16:03] models and services, but sounds like [16:05] they're leaning towards not making the [16:08] weights open. We're going to we're going [16:09] to see with the new Kimmy K3 if they [16:12] actually release the weights next week. [16:14] Dean Ball says, "Wow, two days in a row [16:16] where things I said are likely to happen [16:18] seem likelier to happen from two [16:20] [laughter] adversarial countries. Almost [16:22] as though my analysis was right and as [16:24] though people should not have [16:25] deliberately misinterpreted [16:26] misinterpreted me as of making [16:29] prescriptions. Huh. [16:30] >> Dean Ball victory laugh. [16:33] >> You got to take a victory laugh for Dean [16:35] Ball. He called it. No, he I mean he was [16:37] saying like I I don't necessarily agree [16:39] with this, but just from his experience, [16:41] this is how things might play out. Uh [16:43] and people were very [laughter] upset [16:44] with him about that. [16:45] >> In the chat says Ministry of Commerce, [16:48] People's Republic, [16:49] >> it's their department of commerce. [16:51] >> That would go hard on a business card. [16:53] >> It would. [16:54] >> The great slop wars slop wars [16:56] >> continue. Let's pull up this video from [16:59] Substack. They say, "Today, Substack is [17:01] launching an AI detection feature via an [17:03] integration with Pangram. Going forward, [17:05] you'll be able to scan posts, replies, [17:07] and comments on the Substack app to see [17:09] an estimate of how much of it was [17:10] written by a human or with AI [17:12] assistance. [17:13] >> We care about this at Substack because [17:15] it gets to the core of our mission to [17:17] build an economic engine for culture. [17:19] >> This is going to be bad for Tyler since [17:20] he's mastered the ability to write [17:23] exactly like an AI. What was that [17:26] sentence you wrote by hand that was [17:29] like, "According to studies, you're [17:31] absolutely right." and you just threw in [17:33] every possible AI ism and got flagged [17:37] 100% by Pangram. No. Uh I think people [17:39] will enjoy this. The the revealed [17:42] preference is probably that people don't [17:44] care if it's good and they enjoy the the [17:47] content. But uh we'll we'll we'll go [17:49] back and forth and debate it with Chris [17:51] Best today and see where all of that is [17:54] going. Now, do you think that uh [17:56] Substack writers should admit when they [17:58] hire essentially ghost writers? Uh, [18:02] probably [18:03] >> because sometimes [18:05] you start a brand under a name. Like not [18:09] every post on bloomberg.com is from [18:13] Michael Bloomberg. In fact, none of them [18:15] are. They're all from other people. But [18:17] they disclose this is from, you know, [18:20] some writer. This is from Joe Wesenthal, [18:22] right? But some new influencer brands [18:25] have grown even in the pre-ai era. I had [18:28] a writer on my team when I was doing [18:30] YouTube videos where I would write [18:32] something, he would do a draft, we'd [18:34] sort of go back and forth. I'd put him [18:35] in the credits. Um, but I wouldn't like [18:38] kick it off with like, hey, just so you [18:39] know, [18:40] >> in the chat says book authors don't. [18:42] That is certainly true. [18:43] >> That's 100% true. Uh, some book authors [18:45] will put like this is written by this [18:47] person and this person. It depends on [18:49] like the aura of the ghostwriter. Some [18:51] ghost writers actually do get credits in [18:53] the book, but yes, most book authors [18:55] like Tom Clancy, I think, passed away [18:57] and they still producing books. There's [18:59] a number of writers where their their [19:00] individual brand grew so big that they [19:03] can you can have like a new business. [19:05] Yeah. They become their own brand. And [19:07] so, [19:07] >> yeah, I would I would imagine that 30% [19:09] of the books in like an airport [19:11] bookstore are just 100% ghostritten and [19:14] probably [19:15] >> Yeah. [19:15] >> don't disclose it. I think I I I do [19:18] think it should be it should be possible [19:20] to get to the truth if you hunt hard [19:23] enough, but I don't necessarily need you [19:26] to put it front and center. [19:27] >> For a Substack author or creator, just [19:30] put it somewhere. Yeah. [19:31] >> That you have a team of people that help [19:33] you create [19:34] >> Yeah. [19:35] >> your work. [19:35] >> Well, Dave Edgars took to the uh [19:38] Financial Times saying that no one is [19:40] going to read AI books. He's of course [19:42] the author of The Circle. He's an [19:45] American writer, editor, publisher. Uh, [19:47] a heartbreaking work of staggering [19:49] genius is his 2000 memoir bestseller. [19:52] Final for finalist for the Pulit [19:54] surprise. Uh, he says, "Gullible and [19:56] ridiculous. No one will read AI novels." [20:00] Uh, I think people are already reading [20:02] AI novels. So, he maybe he means like no [20:04] one who's cool or no one who he likes or [20:07] something. Uh, let's read through a [20:09] little bit of this Financial Times [20:10] article. Dave Edgars offers a hand so I [20:13] can clamor aboard his 23-foot sailboat [20:16] mored in the shadow of San Francisco's [20:18] Golden Gate Bridge. It's a grand setting [20:21] with few frills as we sit on the stern a [20:23] meter apart. Fuel cans and fishing nets [20:26] at our feet. There is no one around to [20:28] disturb our conversation. The nearest [20:30] toilets are portaloos, a short bike ride [20:33] away. The birthing fees are negligible, [20:34] making it one of the cheapest places to [20:36] live in uber wealthy Silicon Valley. I [20:39] hadn't thought of that sailboat. If [20:42] you're at a tech company, this has got [20:44] to be the next big lockin trend. [20:46] >> Antonio Garcia Martinez did this. He [20:48] talked about in his book, Chaos Monkeys. [20:50] >> I'm not I won't be satisfied until I see [20:52] larpers posting images on Twitter being [20:55] like, "Hey, I just moved to San [20:56] Francisco. I'm building the next [20:58] oneerson $25 billion." [21:00] >> I'm finding a house in San Rafale. [21:03] >> Yeah. How much [21:05] >> for 1,000? [21:07] Uh it it actually works out to $1,483 [21:11] a month apparently if you finance it [21:13] through Boat Trader. [21:14] >> Goodbye. [21:14] >> It's only about $200,000 to live in a [21:17] 2022 houseboat. [21:19] >> 2022. That's a brand new housebo. [21:21] >> Yeah. So you have to think about the [21:22] Yeah. the slip fees, but [21:24] >> you could easily come in, have your own [21:26] house on the water. [21:28] >> No. The average the average onebedroom [21:29] rent in San Francisco is like two grand, [21:31] three grand right now. [21:32] >> Yeah. You could have a whole house and a [21:34] boat. [21:34] >> And a boat. It's amazing. This is the [21:37] hack. If you get a job in tech and [21:40] you're printing, making six figures, [21:42] >> move into a housebo. [21:44] >> Let's pull this up. [21:45] >> Wait, is this an option in LA? Should [21:47] members of the team be doing this? [21:48] Should we Should we recommend this to [21:50] >> Look at this brand almost brand new 2024 [21:54] La Mer Modern 11. It's 36 foot housebo. [21:58] We have a video here. [22:00] >> Well, Dave Edgar says they call it a $10 [22:02] dock with a million dollar view. Look at [22:05] this, John. This [22:07] >> Yeah. [22:08] >> Couple thousand bucks a month. [22:10] >> This is where the next great startup [22:11] will be born for sure. [22:13] >> Garages are out. [22:15] >> Be out. And also like, oh, take your [22:17] shoes off. It's natural to take your [22:18] shoes off when you're on a boat, you [22:20] know? Right. Unless you got boat shoes. [22:22] Wear boat shoes to [22:23] >> the La Mer Modern 11 offers reliable [22:25] performance and strong power with two [22:26] Yamaha 25 high thrust engines providing [22:29] a total of 50 horsepower. Two cabins and [22:32] one head ensure overnight comfort for [22:33] extended stays. How about year-long [22:36] stays? How about a decade long stay? The [22:40] boat's design includes an upper deck [22:41] terrace with flexit teak flooring, [22:44] enhancing its appeal for relaxation and [22:46] entertainment and might I say locking in [22:49] standard features such as air [22:50] conditioning. We have air conditioning [22:52] on our house boats. [22:53] >> Mhm. [22:54] >> Uh with reverse cycle heat, heated [22:56] floors, LED lighting. [22:58] >> Heated floors. LED lighting. [23:01] >> The inclusion of foam filled composite [23:03] floats and galvanized steel subframe [23:06] ensures durability and stability. [23:08] >> Did we just talk about the housing [23:09] crisis? Send this to sager and jetty [23:11] immediately. This is seasteading. This [23:13] is the teal thesis. [laughter] [23:16] Everyone living in house boats just more [23:18] >> boattrader.com is saying that this is [23:21] amazing [23:21] >> modern housebo would come out to just [23:23] $2,259 [23:25] with only $52,000 down. I promise that [23:29] we have no financial relationship with [23:32] boattrader.com. [23:33] >> Yeah, [23:33] >> even though I'm making a pretty strong [23:35] pitch here [23:36] >> yet. [23:38] I love it. Um, should you recline your [23:41] airplane seat? What do you think? People [23:43] are going back and forth on this. [23:45] >> Brandon Duck says, "I don't remember [23:47] subscribing to this channel, but I've [23:48] been listening for 10 minutes. I'm [23:49] having a good time." [laughter] [23:51] >> Welcome. [23:51] >> Last name Ducks comes on just for laugh. [23:54] >> Yeah. [laughter] Make it make sense. [23:56] >> Make it make sense. I think it makes [23:57] sense. He's just like, I just happen to [23:59] be here for the duck section of the of [24:01] the show. [24:01] >> Andrew says, I'm going to need a double [24:03] wide housebo. Lol. Bootlegging distilled [24:06] anthropic models on the housebo. [24:08] [laughter] [24:08] >> There we go. [24:10] >> Living the dream. [24:12] >> This is crazy. [24:14] >> Okay, so should you recline your [24:16] airplane seat? Uh there's a central [24:18] investigating the central dilemma of our [24:20] time. Joshua Rothman in the New Yorker [24:23] wrote a uh a piece. Imagine this [24:25] scenario. After a long drive to the [24:27] airport, you've just taken off for a 15- [24:29] hour flight. Where are you going? 15 [24:31] hours is so long. How long was the [24:33] flight to Austria from LA? [24:35] >> I think it's like 12 hours. [24:36] >> So, you're going [24:37] >> I think you're going to the Middle East. [24:38] >> You're going to the Middle East, maybe. [24:39] Okay. Uh you're in coach and a family of [24:41] three is sitting behind you. A man, a [24:43] woman, and a toddler. They seem nice, [24:45] but as the flight continues, the kiddo [24:46] grows restless. He kicks your seat and [24:48] pokes her toys into your shoulder. Stop [24:51] that. You hear the father say. The [24:53] toddler incredibly quiets down. The [24:55] cabin lights dim. It's time to sleep. [24:56] Yet, just as rest closes in, you sense [24:59] motion. The seat in front of you is [25:01] reclining. You watch as it comes all the [25:04] way back. Now you feel cramped with [25:06] someone else's seat in your face. You [25:07] also notice some stiffness in your [25:09] shoulders. Your fingers caress your own [25:11] recline button. Surely the reclined upon [25:13] are entitled to recline. Just as you [25:15] start to lean back, you hear a clearing [25:17] throat. You turn around to see the dad [25:19] who fixes you with his gaze. Excuse me, [25:21] he says. Would you mind keeping your [25:23] seat up? I think it's game on. They put [25:25] the button there. They would just take [25:26] it away. I think I've always expected [25:29] everyone to just recline. [25:30] >> Everyone recines. Everyone has the same [25:32] amount of space. It's the conservation [25:34] of space. And [25:36] >> it's really hard to make the argument [25:37] that the the vertical seat is more [25:39] comfortable than the recline seat. [25:40] Exactly. [25:41] >> So, I think everyone commit to comfort. [25:43] >> Everyone commits to comfort. [25:43] >> And then the chat is sharing the real [25:45] alpha. Stand up immediately when [25:47] [laughter] [25:48] the flight lands. Have you seen those [25:50] Meta Rayban videos where people sprint [25:53] to the front? It's so rude. The guy is [25:56] like like, you [laughter] know, he's [25:57] like as soon as he goes beep, he just [25:59] bolts and passes everybody and gets to [26:01] the front. Very rude. I don't think you [26:02] should do that. [26:03] >> This is antisocial behavior. I don't [26:04] think you should recline your seat. [26:05] >> How is it antisocial? [26:07] >> Because you're infringing on some. Just [26:08] because you can do it doesn't mean you [26:09] should. [26:10] >> But they put the button there. Horrible [26:11] take. They're [26:13] >> horrible take. [26:14] >> That that's like don't use the bathroom [26:16] on the on the plane. My entire my entire [26:19] life. My entire life. Even as a child, I [26:21] was looking up waiting for for you know [26:24] the the seat belt sign to to you know to [26:26] be up in the air. Seat belt signs off. [26:28] You recline your seat. Everyone recines [26:30] seat. All good. [26:31] >> Yeah. [26:32] >> Uh what what about what about ordering a [26:34] Diet Coke on a plane? If I'm sitting [26:37] next to you, will you order a Diet Coke [26:38] or is that antisocial behavior? [26:39] >> That's completely different. [26:40] >> Okay. What's your stance on JBL in the [26:43] airplane? [laughter] [26:44] >> A little JBL. You know, [26:46] >> I I I do think I do think that could be [26:48] the next the next trend. [26:50] >> The big question is, are you bringing on [26:52] your, you know, smart portable beaker, [26:56] >> maybe from a from a from an open AI, are [27:00] you bringing that on the plane and [27:01] talking to it the whole time? Is that [27:03] offensive? [27:04] >> I think that's deeply offensive. [27:05] >> Have you seen when like a a band is [27:07] traveling and they start playing on the [27:09] plane? [27:10] >> Have you seen these videos? [27:10] >> No. [27:11] >> That sounds cool, though. [27:12] >> I'll send you one. Yeah, [27:13] >> I'd be down. [clears throat] [27:13] >> That would make me jump. So, this all [27:15] started because uh in 2014, 538, [27:18] we got to have him on the show. U Nate, [27:21] obviously, uh 538 asked around a [27:23] thousand flyers what they thought about [27:26] reclining their seats. 41% of them said [27:28] that reclining was rude, so they lost. [27:31] It's not rude. [27:31] >> It's over. [27:32] >> That number seems to have risen, though. [27:34] In 2020, Eric Jones, a math professor [27:36] who writes about travel statistics on [27:38] the website the vacationer, conducted a [27:39] similar survey and found that 77% of [27:42] respondents objected to reclining. Wow. [27:45] Sheep. Shele over here. We got sheep. [27:48] >> How about this? I think you should be [27:49] able to recline if you're like dressed [27:51] nicely. If you're wearing sweatpants, [27:53] then you don't get to. [27:54] >> You don't get to recline. [27:57] >> Should kids be able to kick the front [27:59] the seat in front of them as much as [28:01] they want? See, that's something I will [28:02] put I will totally put up with. I'm [28:04] totally fine getting kicked in the back [28:05] by a kid because it's so brutal as a [28:08] parent, as a kid to be on a plane. [28:10] They're locked up in the plane. That's [28:11] miserable. And just just tough it up. [28:14] You You're reclined. You have multiple [28:17] inches of padding. Unless you got [28:19] Ronaldo back. Order that hot pizza [28:22] pocket. Live life in the permanent [28:24] underclass. Live a little for God's [28:27] sake. [28:27] >> Ramp launched a model router. Uh Brexton [28:31] loves it. He says, "Ramp launching a [28:33] model router is god tier product [28:36] expansion. The winds are blessing it. [28:38] The market is exhausted with token [28:39] maxing, proliferation of new models, [28:41] sovereignty. This is the first time I've [28:43] looked at a launch and went, "Wow, this [28:46] might be the start of a new core [28:47] product." Yeah. Uh there's been a bunch [28:49] of speculation about this, but it makes [28:51] a lot of sense. Ramp's prog promise is [28:53] to save you time and money. The CFOs are [28:56] increasingly annoyed by token uh [28:59] routing. We talked to Eric Lyman about [29:01] this last week. Are you using the [29:03] craziest frontier model burning millions [29:05] of tokens to answer the weather? Maybe [29:08] that should have an expense report [29:09] attached. And uh so uh very very [29:12] exciting to see where this goes, how [29:14] they roll it out. I know it's in alpha [29:16] with a number of customers. And I [29:17] believe that the product's actually [29:18] available to non-ramp customers as well, [29:21] so you can use their product [29:23] exclusively. Uh [29:25] >> you says going after everything time and [29:27] money touches. [29:28] >> [laughter] [29:28] >> I love it. I love it. Anyway, leave us [29:31] five stars on Apple Podcast and Spotify. [29:33] Sign up for our newsletter at tbpn.com. [29:36] And [29:37] >> we love you. [29:38] >> We love you. [29:38] >> Throwing flashbang. [29:39] >> Goodbye. Flashbanging the horse. [29:48] >> [music]