Alphabet Pivots Compute Capacity to Anthropic as AI Model Edge Slips

AI模型优势渐失，Alphabet将算力转向Anthropic Google Cloud将其超过20%的TPU算力产出承包给Anthropic，优先保障基础设施收入而非内部AI研究，实际上放弃了其前沿实验室的地位。 主持人John Coogan和Jordi Hays分析了最新热门科技新闻，首先讨论了科学家利用AI设计出新型、对人类无致病性且针对细菌的合成病毒，此举可能为精准医疗带来革命性突破。他们还探讨了有关OpenAI即将推出、由Jony Ive设计的无屏幕家用设备的泄密信息，以及从预渲染AI播客摘要向实时交互式语音模型的转变。讨论进一步审视了Meta因未成年人社交媒体成瘾问题面临高达9亿美元的巨额法律处罚，并将其与1998年的《烟草主和解协议》进行类比，同时评估了SemiAnalysis一份措辞严厉的报告——该报告称Google DeepMind正在失去其作为前沿AI实验室的地位。最后，节目进入“豪宅板块”（The Mansion Section），探访了Authentic Brands Group前高管Nick Woodhouse拥有的价值6850万美元的迈阿密庄园，并分析了ABG庞大的品牌授权帝国。 斯坦福大学与 Arc Institute 的科学家首次利用生成式 AI 从零设计出全新的功能性合成病毒；为规避生物安全风险，该模型严格基于噬菌体训练，仅针对细菌而不会感染人类或动植物。 (1:05) OpenAI 正在与 Jony Ive 创办的 LoveFrom 合作开发首款名为“Donut”的便携式无屏幕消费级 AI 硬件设备，预估售价在 300 至 400 美元之间，计划于 2027 年左右发布。 (7:14) 因涉及未成年人社交媒体成瘾及心理健康损害，Meta 在新墨西哥州面临超 9 亿美元的罚款与赔偿金判决，此类诉讼正呈现出类似于 1998 年烟草总和解协议（MSA）的法律演进趋势。 (14:58) SemiAnalysis 报告称，因强化学习团队核心人才流失及算力资源分配失衡，Google DeepMind 已失去前沿 AI 实验室地位；同时 Google Cloud 将 2026 年第三季度至 2027 年第四季度超 20% 的 TPU 产能直接长协出售给竞争对手 Anthropic。 (23:07)

This episode examines how rapid technological scaling is reshaping biotechnology, consumer hardware, and corporate strategy, from AI-synthesized viruses and ambient voice devices to Google's pivot toward high-margin cloud infrastructure over internal frontier AI research. It also tracks the mounting institutional fallout across tech and commerce, including tobacco-style legal liability for social media platforms and the roll-up economics governing legacy consumer brands. AI-Generated Viruses and Biosecurity Guardrails (1:05) In a study published in Science, researchers from Stanford University and the Arc Institute demonstrated the first functional, completely novel viruses generated by artificial intelligence. The generative model was trained exclusively on naturally occurring bacteriophage DNA—specifically using the Phi X 174 baseline structure—while explicitly filtering out data related to human, plant, animal, and fungal pathogens to mitigate biosecurity risks. The synthesized viral blueprints were validated in laboratory tests, successfully producing viable phages that infect bacteria. While commentators such as Eliezer Yudkowsky raise alarms over existential threats—including scenarios where consumer-grade computing hardware could eventually be used to engineer dangerous biological agents—the hosts emphasize the immediate therapeutic upside. Developing custom synthetic phages at a fraction of standard costs could revolutionize medicine by creating targeted delivery vehicles for gene therapies, though commercialization may require strategic rebranding to overcome post-pandemic public apprehension surrounding the word 'virus.' Ambient Consumer Hardware and Interactive AI (7:14) Reporting from Bloomberg indicates that OpenAI is collaborating with Jony Ive's design firm, LoveFrom, on a screenless consumer hardware device targeted for a 2027 release at an estimated price of $300 to $400. Designed as a portable, hockey-puck or donut-shaped device equipped with microphones, speakers, cameras, and spatial sensors, the hardware relies on ChatGPT's Advanced Voice Mode and subtle mechanical movements to act as an ambient assistant capable of learning user habits over time. The initiative reflects a broader industry transition away from static, pre-rendered synthetic media—such as Google NotebookLM's two-person podcast generation—toward live, bidirectional conversational intelligence. However, the hosts note that the viability of dedicated AI hardware depends heavily on workflow integration; current systems remain functionally siloed, requiring manual context transfer across applications rather than autonomously executing persistent background tasks. Social Media Addiction and Tobacco-Style Liability (14:58) Legal liability surrounding algorithmic youth addiction is escalating rapidly, marked by a New Mexico state court ruling directing Meta Platforms to pay over $900 million in penalties, which includes a $567 million youth harm mitigation restitution fund alongside $375 million in earlier civil penalties. This follows a March 2026 Los Angeles jury verdict awarding $6 million to an individual plaintiff ($4.2 million against Meta and $1.8 million against Google) and a May 2026 $27 million settlement paid to a Kentucky school district by Meta, YouTube, TikTok, and Snap to reimburse student mental health service costs. The hosts compare these proceedings to the 1998 Tobacco Master Settlement Agreement (MSA), where 46 states resolved public healthcare cost deficits through an initial $206 billion baseline agreement and perpetual annual payments. Because a single lump-sum penalty comparable to Meta's roughly $200 billion annual revenue would prove devastating, legal precedents suggest state litigation may eventually force social media platforms into a similar long-term financialized settlement structure funded by continuous revenue-based payouts. DeepMind Talent Losses and Cloud Infrastructure Pivots (23:07) A critical report from SemiAnalysis asserts that Google DeepMind is losing its status as a frontier AI lab, driven by significant talent attrition from reinforcement learning teams—including the departures of Jeff Dean, Noam Shazeer, Sanjay Ghemawat, Quoc Le, and Oriol Vinyals. The report further points to internal compute reallocations, noting that Google contracted more than 20% of its total TPU capacity between Q3 2026 and Q4 2027 directly to Anthropic, despite holding an equity stake capped at 15% with no board representation or voting rights. This shift highlights an internal divergence between DeepMind founder Demis Hassabis's ambition for recursive self-improvement toward AGI and CEO Sundar Pichai's commercial focus on enterprise cloud products and AI-enhanced search. With Google Cloud Platform (GCP) delivering EBIT margins in the mid-30% range and projecting over $250 billion in upcoming TPU bookings, executive incentives increasingly prioritize lucrative infrastructure hosting agreements over high-risk, capital-intensive internal frontier model development. Brand Consolidation and Shifting Youth Culture (29:31) The listing of an 8,800-square-foot waterfront estate on Miami's La Gorce Island for $68.5 million by Nick Woodhouse, former president and CMO of Authentic Brands Group (ABG), underscores the immense capital generated by intellectual property licensing conglomerates. ABG's portfolio consolidates celebrity likenesses—including Marilyn Monroe, Elvis Presley, and Muhammad Ali—alongside legacy apparel labels such as Brooks Brothers, Reebok, and Tapout. Analogizing ABG's roll-up playbook to software consolidators like Bending Spoons, the hosts note that heritage action sports brands such as Volcom, Billabong, and Quiksilver have seen structural cultural decline. Outside coastal areas, modern youth demographics have largely abandoned traditional surf and skate apparel in favor of contemporary luxury streetwear like Chrome Hearts, turning iconic lifestyle labels into financialized licensing properties. Across synthetic biology, consumer devices, and platform software, tech enterprises face a dual tension between ambitious scientific breakthroughs and mounting commercial pragmatism, marked by regulatory scrutiny over digital addiction and capital shifts toward predictable infrastructure revenue.

本期节目聚焦人工智能在前沿生物工程与下一代消费硬件中的落地探索，并审视了大型科技公司在监管外部性追责与研发战略取舍上面临的深层挑战。核心议题涵盖AI生成新型病毒的生物安全争议、OpenAI的环境硬件布局、社交媒体成瘾诉讼对“大烟草”模式的借鉴，以及Google内部在云算力短期盈利与前沿AGI探索之间的战略割裂。 AI生成合成病毒突破与生物安全争议 (1:05) 斯坦福大学与Arc研究所发表在《Science》上的研究表明，科研人员首次利用生成式AI从零设计出全新的合成病毒。该模型以噬菌体ΦX174为基准架构，仅针对细菌进行训练，并严格过滤了所有人类、动植物与真菌病原体数据。实验通过将AI设计的基因蓝图导入宿主细菌，成功培育出具有完整感染功能的新型噬菌体。评论认为，这项技术的核心突破在于大幅压缩定制病毒载体的制造周期与成本，未来有望作为精准基因疗法的递送工具；但由于公众在后疫情时代对“病毒”存在天然恐慌，该领域若要实现商业化推广，亟需仿照GLP-1药物重塑为“肽类”的公关策略，进行全面的术语重构。 伴随技术突破的是对生物安全的长期担忧。主持人指出潜在的灾难性情景：个人未来可能仅凭普通消费级显卡和开源模型，便能获取定制致命病原体的完整方案；Eliezer Yudkowsky等关注AI存在性风险的人士也对此类生物设计能力亮起红灯。尽管正规科研机构与防御部门正在积极构筑防御体系，但开源工具扩散与生物防御之间的长期攻防博弈已经拉开帷幕。 硬件载体演化与实时环境交互 (7:14) 据彭博社记者Mark Gurman报道，OpenAI正联手Jony Ive创立的LoveFrom设计工作室开发首款消费级硬件。这款预计于2027年面世、定价在300至400美元之间的屏幕缺失型设备，外观类似冰球或甜甜圈，配备电池、麦克风、扬声器、摄像头及空间感应阵列。该设备依托ChatGPT的高级语音模式（Advanced Voice Mode），通过机械微动和动态光效提供拟人化反馈，旨在作为家庭环境中感知上下文的常驻助手，未来逐步承接智能手机的部分日常任务。 对话深入探讨了AI交互方式的范式转移：类似Google NotebookLM将文档生成预录播客的工具虽然广受学生欢迎，但单向静态音频正迅速被双向实时语音取代，用户更偏好能动态调整深度与语气的实时对话。然而，这类独立硬件成功的最大瓶颈仍在于工作流割裂——例如用户在OpenAI Codex中编写项目时，无法在关闭电脑后无缝将上下文流转至移动设备生成配图，仍需手动复制原始上下文。专用硬件若要立足，必须具备自主调动云端实例、执行代码和跨应用维护持久工作流的能力。 社交媒体成瘾诉讼与烟草和解模式 (14:58) 针对大型社交平台引发青少年心理健康危机的法律问责正在全美升级。新墨西哥州法院近日裁定Meta支付超过9亿美元的罚款与赔偿，其中包括5.67亿美元的危害缓解专项基金以及3.75亿美元的民事罚金，同时对未成年人使用施加严格限制。此前，洛杉矶陪审团于2026年3月判定Meta与Google分别对一名童年成瘾的原告承担420万美元和180万美元赔偿；同年5月，肯塔基州某学区针对致瘾设计导致心理健康预算激增的诉讼，以Meta出资900万美元、多家平台共付2700万美元达成和解。目前Meta面临的累计法律赔偿责任已接近10亿美元。 这一司法演进路径与1998年《烟草主和解协议》（MSA）高度吻合。当年烟草诉讼的根本转折，在于各州政府以公立医疗系统承担了吸烟导致的巨额外部性成本为由介入起诉，最终促成了46个州达成2060亿美元基准和解金，并确立了按销售额无限期支付年费的制度，华尔街甚至将后续赔偿金包装为市政债券。鉴于Meta年营收约为2000亿美元，一次性天价罚款具有毁灭性，因此引入类似烟草协议的永久性年度赔偿税收机制与强制性产品限制，可能成为规制社交平台算法致瘾的长期方案。 Google云算力套现与前沿研发撕裂 (23:07) 研究机构SemiAnalysis发布分析报告断言，Google DeepMind已实质上脱离最前沿AI实验室阵营。该报告指责Google在强化学习（RL）领域遭遇严重人才流失，包括Google Brain联合创始人Jeff Dean、Noam Shazeer、Sanjay Ghemawat、Quoc Le及Oriol Vinyals等核心技术奠基人纷纷离职。雪上加霜的是算力分配的商业化倾斜：Google选择将2026年第三季度至2027年第四季度期间超过20%的TPU产能，通过长期合同直接出售给核心竞争对手Anthropic，而非优先保障自身Gemini 4 Pro的前沿探索，导致其研发文化被指日趋官僚与保守。 这凸显了Google内部战略的深层矛盾。Google Cloud Platform（GCP）凭借30%至35%的息税前利润率（EBIT margin）成为盈利支柱，预计未来数个季度将斩获超过2500亿美元的TPU订单积压并计入剩余履约义务（RPO）。同时，Google持有Anthropic约14%的无投票权股权（上限15%），甘愿充当计算托管商。这一路径映射出高管层的理念分歧：CEO Sundar Pichai倾向于将AI视作赋能企业云和搜索展示的基础设施，而DeepMind创始人Demis Hassabis则坚持将AGI视为实现递归自我改进（RSI）的核心技术，算力变现优先的策略显著打击了内部前沿科研团队的士气。 ABG的IP授权运作与潮流品牌演变 (29:31) 在“豪宅专栏”中，节目讨论了Authentic Brands Group（ABG）前总裁兼首席营销官Nick Woodhouse将其位于迈阿密La Gorce Island的当代水岸豪宅以6850万美元挂牌出售的消息。该房产是其于2021年以1700万美元购入未完工基础、并在2023年建成的8800平方英尺庄园。这笔交易折射出ABG通过资本聚合经典品牌IP所建立的巨大商业版图。ABG旗下掌控着玛丽莲·梦露、猫王、穆罕默德·阿里等名人遗产，以及Brooks Brothers、Champion、Reebok等50多个知名消费品牌，通过收取授权费构建了庞大的轻资产现金流帝国。 节目将ABG的品牌收编策略比作软件界的Bending Spoons，同时也流露出对传统亚文化品牌没落的感叹。ABG旗下收纳了Volcom、Billabong、Quiksilver和RVCA等曾深刻影响青年文化的极限运动与冲浪标志。但随着当代年轻消费者的审美向Chrome Hearts等高端奢华街头品牌转移，内陆非沿海年轻群体不再向往冲浪文化的生活方式，导致这些传统街头与滑板服饰品牌失去独立文化号召力，最终沦为资本组合中依靠情怀授权维系的资产标的。 本期节目勾勒出AI与数字经济发展中的结构性矛盾：技术能力正加速渗透至生物合成设计与空间交互硬件，但面临着严峻的生物安全风险与交互集成瓶颈；同时，科技巨头一方面在社交媒体致瘾的外溢成本上面临烟草级别的系统性司法清算，另一方面又在云计算商业化盈利与激进的前沿AGI探索之间摇摆不定，反映出技术奇点愿景与现实商业回报之间的巨大张力。

## Background 该视频由科技谈话节目 TBPN 于 2026 年 8 月 13 日发布在 YouTube 平台上。根据页面简介，本期为时长约 30 分钟的精简版“Diet TBPN”，由主持人 John Coogan 与 Jordi Hays 联袂主持，精选当日科技行业中最重磅的新闻动态与深度探讨。 ## TL;DR 本期节目聚焦多项前沿科技与行业风暴：首先深入探讨了斯坦福与 Arc 研究所利用 AI 凭空设计全新活性病毒的技术突破及其生物安全隐患，并揭秘了 OpenAI 携手 Jony Ive 研发的首款 300 至 400 美元无屏“甜甜圈”硬件设备。随后，主持人从 NotebookLM 延展至实时端侧 AI 自动化的演进，并结合 Meta 因社交成瘾被判赔逾 9 亿美元的裁决，深度复盘了烟草行业《主和解协议》（MSA）对科技监管的镜鉴意义。最后，节目深入剖析了 SemiAnalysis 针对 Google DeepMind 失去前沿地位及向死敌 Anthropic 大举外售 TPU 算力的尖锐批评，并在标志性的“豪宅板块”中借 ABG 品牌操盘手挂牌 6850 万美元迈阿密海景庄园，揭示了经典消费品牌的资本运作底牌。 ## Atmosphere & style 节目的整体基调兼具硅谷极客圈的幽默风趣与专业分析师的辛辣犀利，两位主持人配合默契、包袱不断，时常在硬核前沿技术探讨与无厘头脑洞之间自如切换。从开场调侃墨镜颜色与构想“闪光弹”硬件形态的轻松嬉笑，到剖析烟草和解历史与大厂官僚病时的深刻审视，情绪层次丰富且推进迅速。整场对话既带有好友闲聊式的即兴随性与讽刺吐槽，又处处透露出对 AI 军备竞赛、资本利益博弈与产业终局的深刻洞察。 ## Bulletin - [02:35](https://www.youtube.com/watch?v=OQH8s5qg1yw&t=155s) 斯坦福与 Arc 研究所利用 AI 首次成功设计并合成自然界不存在的活体噬菌体病毒，大幅降低了合成生物学生产门槛，同时引发公众对生物安全防御滞后的担忧。 - [07:15](https://www.youtube.com/watch?v=OQH8s5qg1yw&t=435s) OpenAI 传出正由 Jony Ive 团队操刀研发首款无屏幕消费级硬件（代号似“甜甜圈”），预计 2027 年推出，定价 300–400 美元，主打拟人机械动态与长期情境感知。 - [14:58](https://www.youtube.com/watch?v=OQH8s5qg1yw&t=898s) Meta 在新墨西哥州被判决向未成年人专项基金和民事罚款支付超 9 亿美元，针对社交成瘾的法律诉讼升级，呈现出走向类似 1998 年烟草《主和解协议》（MSA）的行业清算趋势。 - [23:07](https://www.youtube.com/watch?v=OQH8s5qg1yw&t=1387s) SemiAnalysis 称 Google 内部人才流失且在 2026–2027 年间将逾 20% 的 TPU 算力长期锁定并供应给竞品 Anthropic，标志着 DeepMind 前沿技术统治力面临严重动摇。 ## Key takeaways - [02:35](https://www.youtube.com/watch?v=OQH8s5qg1yw&t=155s) 斯坦福与 Arc 研究所发表于《Science》的成果是人类首次完全由 AI 无中生有设计出可正常运作的活体病毒，其核心在于研发速度的指数级提升与合成成本的断崖式下降。 - [04:06](https://www.youtube.com/watch?v=OQH8s5qg1yw&t=246s) 由于公众历经疫情后对“人造病毒”极度敏感，合成生物学在医疗载体商业化落地上亟需类似 GLP-1 包装为“多肽”的公关重塑策略。 - [07:15](https://www.youtube.com/watch?v=OQH8s5qg1yw&t=435s) OpenAI 首款硬件设备设计为冰球大小的无屏幕“甜甜圈”，配备呼吸灯与微型机械活动部件以模拟生命感，旨在成为深度融入日常生活的个人 AI 终极语音载体。 - [09:10](https://www.youtube.com/watch?v=OQH8s5qg1yw&t=550s) NotebookLM 通过双人播客形式极大降低了复杂科学论文与长文档的理解门槛，展示了信息摄取形态的代际跃迁。 - [11:21](https://www.youtube.com/watch?v=OQH8s5qg1yw&t=681s) 从单向预生成的静态音频演进至支持实时打断追问的语音模式，AI 硬件成败的关键在于能否打通无头浏览器抓取与常驻后台自动化执行等工作流。 - [14:58](https://www.youtube.com/watch?v=OQH8s5qg1yw&t=898s) 新墨西哥州判决 Meta 赔偿超 9 亿美元，伴随洛杉矶判决和多地学区和解案，针对社交媒体成瘾与未成年人心理伤害的法律索赔规模已进入天价阶段。 - [18:45](https://www.youtube.com/watch?v=OQH8s5qg1yw&t=1125s) 1998 年烟草 MSA 协议通过终止各州诉讼换取基于卷烟销量的永久年度缴费与金融化贴现，这一机制极可能成为社交媒体巨头未来应对司法围剿的参考样本。 - [23:07](https://www.youtube.com/watch?v=OQH8s5qg1yw&t=1387s) SemiAnalysis 痛批 DeepMind 强化学习核心团队出走及算力分配失衡，认为 Google 内部保守官僚的文化使其难以重回行业前沿顶尖水平。 - [25:15](https://www.youtube.com/watch?v=OQH8s5qg1yw&t=1515s) Google 管理层倾向于充当算力军火商，通过 GCP 外售海量 TPU 锁定高确定性云营收，这与研究团队追求突破通用大模型本身的理想产生了战略分歧。 - [27:45](https://www.youtube.com/watch?v=OQH8s5qg1yw&t=1665s) Google 虽为主要竞争对手 Anthropic 提供了庞大算力与数据中心财务担保，但其持股被严格限制在约 14%，且完全没有董事会席位与投票权。 - [29:31](https://www.youtube.com/watch?v=OQH8s5qg1yw&t=1771s) 品牌运作巨头 ABG 靠打包收购破产知名品牌（如 Reebok）和运营玛丽莲·梦露等名人身后肖像权实现轻资产授权变现，其前高管挂牌 6850 万美元迈阿密海景庄园印证了该模式的资本积累规模。 ## Detailed walkthrough ### [00:05](https://www.youtube.com/watch?v=OQH8s5qg1yw&t=5s) 开场调侃与今日节目看点 节目在一阵轻松幽默的闲聊中拉开帷幕。两位主持人 John Coogan 与 Jordi Hays 针对直播间弹幕展开互动——观众们此前一直在竞猜 Jordi 今天会戴什么颜色的太阳镜。面对弹幕猜测的“橙色”还是“赭石色”（burnt sienna），Jordi 解释自己的眼镜其实是绿色的镜框搭配橙色镜片，并打趣道应该借机抛出那句经典台词：“我戴墨镜是因为我正在眺望未来，而未来确实太刺眼了。” 玩笑过后，主持人预告了周五精简版节目的重磅嘉宾与议题： - **Khosla Ventures 投资合伙人**：将做客节目，深度解析他们对初创公司 Discovery Loop 的最新投资。该公司由前 Google 人工智能领军人物 Jeff Dean 及其 DeepMind 核心团队成员联手创立。 - **Databricks 联合创始人 Patrick Wendell**：随后将上线解读他们最新发布的博客文章。虽然标题看似枯燥，但核心内容围绕“更智能的 AI 模型路由机制”（Smarter Model Routing），技术干货满满。 --- ### [01:05](https://www.youtube.com/watch?v=OQH8s5qg1yw&t=65s) AI 首次创造全新病毒与生物安全争议 主持人切入当天的首个重磅科技新闻：人类科学家首次利用 AI 凭空设计出了大自然中从未存在过的全新品种病毒。这一突破标志着合成生物学与生物技术的历史性飞跃，但随之而来的则是公众对长期生物安全（Biosecurity）隐患的巨大担忧。 面对这种技术突破，舆论很容易联想到最可怕的极端梦魇场景：任何人只要去百思买（Best Buy）买一台配有两张普通显卡的家用游戏 PC，就能在本地运行开源大模型，按照模型手把手的步骤指引合成极度危险的病原体。尽管目前有许多顶尖人才和资金雄厚的机构在筑牢防线、全力遏制这种滥用风险，形成博弈平衡，但相关监管和伦理挑战依然严峻。 两位主持人用半开玩笑的口吻探讨了这种心理冲击： - “你一觉醒来，科学家们心想：‘哎呀，自然界原有的病毒好像还不够多呢，咱们再搞点新的吧！’” - 尽管论文和实验明确强调这些人工病毒**完全不感染人类，仅靶向细菌**（属于噬菌体范畴），但 Jordi 吐槽这依然无法让人彻底安心，因为人体肠道内含有大量的共生菌群，“肠道菌群也是人体的关键部分，难道就没人站出来为这些无辜的细菌发声吗？” --- ### [02:35](https://www.youtube.com/watch?v=OQH8s5qg1yw&t=155s) 《Science》重磅突破：噬菌体生成与生物医药的本质差异 这项研究发表于权威学术期刊《Science》，由斯坦福大学与 Arc 研究所（Arc Institute）的科研团队共同完成。技术逻辑十分清晰：研究人员训练 AI 模型识别天然病毒 DNA 中的内在模式与组织逻辑，随后让 AI 凭空编写全新的遗传序列代码。在实验室合成这些 DNA 序列并植入细菌体内后，细菌成功组装并释放出具有完整生物活性的活体病毒，且这些病毒展现出了感染其他目标细菌的能力。 ```mermaid flowchart TD accTitle: AI 病毒设计与生成流程 accDescr: 展示斯坦福与 Arc 研究所如何利用 AI 训练、生成并最终在实验室表达新型人工病毒的全流程。 subgraph Phase1["1. 模型训练与模式识别"] A["天然噬菌体 DNA 序列"] --> B["AI 模式识别模型"] E["严格过滤机制<br/>(剔除人类/动植物/真菌病毒)"] -.-> B end subgraph Phase2["2. 序列生成与人工合成"] B --> C["AI 生成全新基因代码"] C --> D["实验室化学合成 DNA"] end subgraph Phase3["3. 细胞表达与功能验证"] D --> F["导入目标宿主细菌"] F --> G["宿主产出活性新型病毒"] G --> H["成功感染其他目标细菌"] end ``` 针对公众的恐慌，主持人特别澄清了该研究的安全边界与关键细节： - **安全过滤与限制**：AI 在训练阶段仅输入了噬菌体（专性感染细菌的病毒）数据，明确且严格地排除了任何感染人类、动物、植物和真菌的病毒样本。因此，模型在底层逻辑上根本无法生成能感染人类的病毒。 - **与传统合成的区别**：科学家人工合成病毒基因组已有多年历史（用于疫苗开发和疾病机理研究），但以往都是在自然病毒的基础上“复制或修改”，而**这是人类历史上首次由 AI 无中生有设计出能在现实世界正常运作的全新病毒**。 - **技术突破的本质**：主持人探讨这究竟是范式革命还是单纯工具升级，核心结论在于**“研发速度的指数级加速与生产成本的大幅断崖”**。如果合成功能性病毒的成本骤降一千倍，既能推动生物医药迎来爆发期，也会让防御生物安全威胁的任务变得无比紧迫。 --- ### [04:06](https://www.youtube.com/watch?v=OQH8s5qg1yw&t=246s) “病毒”污名化、公关重塑与 Arc 研究所的监管考量 技术引发的社会心理反弹非常剧烈。主持人提到，在历经新冠疫情（Wuhan）等重大公共卫生事件后，公众对“病毒”、“人造”这类字眼极其敏感且本能抗拒，AI 风险学者 Eliezer Yudkowsky 等人对此类研究也抱持强烈的警惕与悲观态度。 为了推动该技术在医疗场景的落地（例如利用病毒作为载体进行基因治疗或靶向药物输送），主持人认为业界亟需对“病毒”（Virus）进行一次彻底的**公关品牌重塑（Rebranding）**： - 就像减肥药物 GLP-1 巧妙地将自身包装为“多肽”（Peptide），听上去安全、健康且无害，让人联想到自然成分，成功摆脱了类固醇、兴奋剂（Gear）或提取自毒蜥毒液（Gila Monster Venom）的负面刻板印象。 - 反观“病毒”一词，在日常语境下百分之百代表灾难与疾病，人们绝不可能在听到自己染上病毒时问“是好病毒还是坏病毒”。如果不换个名称包装，普通大众很难在心理上接受商业化的基因载体病毒。 《纽约时报》也对此进行了专题报道，指出这次 AI 构思的全新病毒是以自然界著名的 **Phi X 174**（$\Phi\text{X }174$）单链 DNA 噬菌体为蓝本变奏而来的。Jordi 调侃“Phi X 174”听起来异常硬核冷酷，甚至像马斯克（Elon Musk）给自家孩子起的名字。目前全球政府和科学机构在建立防止致命病毒被滥用的“安全护栏”方面进展迟缓，远落后于科学狂奔的脚步，因此科研团队短期内绝无可能开源这项模型算法。 --- ### [07:15](https://www.youtube.com/watch?v=OQH8s5qg1yw&t=435s) 曝光：OpenAI 首款硬件设备“甜甜圈” 节目随后转向重磅消费电子泄密。彭博社著名科技记者 Mark Gurman（被两位主持人戏称为“终结者”The Germinator）独家曝光了 OpenAI 正在紧锣密鼓研发的首款消费级硬件： - **外观与尺寸**：外形被描述为类似**冰球（hockey puck）大小的甜甜圈**，中间带有开孔或特殊弧度设计。 - **售价定位**：预计售价在 **300 至 400 美元** 之间。 - **硬件形态与传感器**：这是一款由电池驱动、**不带屏幕**的便携式智能音频硬件。用户既可以随身携带在家中走动，也可以摆在床头柜或厨房台面上。设备集成了扬声器、高灵敏度麦克风阵列、摄像头以及一系列环境感知传感器，使 AI 能实时“看懂”和“听懂”周围发生的一切。 - **核心定位与交互**：它相当于 ChatGPT 现有语音模式（Voice Mode）的“终极增强版实体形态”。它具备强大的长期记忆与环境感知能力，会在长时间互动中深度学习主人的生活习惯与个性特征，使对话更加私人化、情境化。 - **机械动态与生命感**：为了彻底摆脱亚马逊 Echo 或 Google Nest 那种冰冷呆板的“音箱”质感，OpenAI 在硬件上增加了**呼吸灯光与微型机械活动部件**。当 AI 思考或回应时，机身部件能产生细腻的物理起伏与蠕动反馈，让机器呈现出“活着”的拟人生命感。 - **操刀团队与发布时间**：该设备由前苹果首席设计官 **Jony Ive（乔纳森·艾维）的设计团队**主导工业设计，预计最快将于 **2027 年** 正式推向市场。它是 OpenAI 庞大硬件生态布局的第一步，终极目标是打造出一套能接管现代智能手机核心功能的 AI 终端矩阵。 讨论到带有灯光、高分贝声音还会物理移动的小型设备时，Jordi 脑洞大开地提出了一个“功能需求”（Feature Request）——将其做成可随身携带的**“便携式震撼弹 / 闪光弹”（Rolling Flashbang）**，在推门去找朋友聚会时可以直接把它滚进房间给朋友们来个恶作剧式的物理惊喜。 ### [09:10](https://www.youtube.com/watch?v=OQH8s5qg1yw&t=550s) NotebookLM 的魔力与播客化学习体验 Google 内部在模型研发进展与商业化落地之间存在着明显博弈，其 AI 战略推进有亮点也有短板。主持人由此谈到 Google 旗下备受赞誉的 NotebookLM：这项技术被视为极具魔力的产品，用户只需喂给它几份参考资料或一份特定报告，它就能自动生成一段由两个虚拟主持人对谈的双人播客。这种极其口语化、生活化的对话形式，让很多不爱啃干瘪长文的人找到了绝佳的信息获取途径。举例来说，如果你想了解 ARC Institute（生物医学研究机构）正在研发的新型病毒，需要搞懂噬菌体（bacteriophage）的运作机制，只需把深度研究报告扔给 NotebookLM，指令让“两位科学家先用高中生听得懂的语言解释一遍，再逐步深入到大学水平”，就能迅速吃透复杂概念。 这种学习模式在实际场景中极受欢迎。YouTube 聊天区观众 Tyler 留言表示自己是 NotebookLM 的狂热粉丝，每周都在用。主持人感叹，如果自己在大学时代有这种工具，遇到没怎么准备但要交论文的情况，直接让它围绕几个主题生成一段一小时的播客，听完一轮基本就能直接把论文“撕”出来了。另一位搭档也分享了亲身经历：他曾在一场中世纪历史考试前，把包含大量事件年代和概念的词汇表直接丢给 NotebookLM 生成播客，轻松复习并拿到了优异成绩。 主持人还打趣了近期网络上调侃“老一辈没有 AI 怎么写出五段式议论文”的热梗——当时的答案其实是“兄弟，我们当年甚至连原著都不读，直接看 SparkNotes 摘要就能写出来”。这种从过去的 SparkNotes 到今天的 AI 播客的演变，凸显了信息摄取形态的代际跃进。 - **核心应用场景**：长篇报告降维解读、考试前词汇表与历史时间线速成听力复盘。 - **产品特质**：将单向长文本输入转化为双人互动播客，支持按需指定受众理解阶梯（如从高中水平过渡至大学水平）。 --- ### [11:21](https://www.youtube.com/watch?v=OQH8s5qg1yw&t=681s) 从预生成播客到实时语音：个人 AI 的能力融合与硬件前瞻 AI 能力正在发生急剧的“功能坍塌”（能力融合）。Sam Altman 之前曾因公开表示“会用 ChatGPT 整理自己的日程和家庭事务并生成播客来听”而遭到网络嘲讽，网友讥讽他“为什么不直接多陪孩子聊聊天”。但跳出八卦来看，背后的技术演进十分关键：未来用户真的还需要像 NotebookLM 或 ChatGPT 那样事先生成一段静态音频文件吗？随着实时语音模式（Voice Mode）与长期记忆（Memory）功能的成熟，AI 完全拥有足够的上下文，能够在你需要时随时随地化身一台“实时定制播客”，甚至比静态播客体验更好。 静态播客的一大缺陷在于它必须假设听众的认知水平，容易出现“讲得太深”或“讲得太浅”的错位；而实时语音交互则完全是“交互式探险”（Choose-your-own-adventure）或专家访谈（像调用行业专家智库一样）。你可以随时打断它，让它沿某个特定切口往深处挖。此外，针对 ChatGPT 出现的舆论反弹与实际使用热度之间的割裂现象（社交媒体上有上百万人点赞嘲弄 AI 的帖子，但应用商店榜单上却有数以亿计的真实用户高频活跃），主持人指出“声明偏好”（口头上嫌弃 AI）与“显性偏好”（家里有个能干的 AI 确实真香）形成了鲜明反差。 关于传闻中 OpenAI 即将推出的硬件设备（代号类似“Donut”的小型硬件），其能否实现跨模态工作流的深度整合成为成败关键。当前的痛点在于，ChatGPT 日常问答、ChatGPT for Work 与基于云端环境写代码的 Codex 依然割裂。 ```mermaid flowchart TD accTitle: 当前割裂工作流与未来端侧整合对比 accDescr: 展示目前在Codex、网页与移动端间手动搬运上下文的繁琐流程，对比未来硬件一体化端到端执行。 subgraph Current["当前割裂状态 (Manual Workaround)"] A["Codex / 电脑端运行代码"] -->|"手动复制输出与上下文"| B["ChatGPT 移动端 App"] B -->|"单次对话请求"| C["生成图片 / 后续改写"] note1["无法离开电脑运行多步自动化或持久工作流"] end subgraph Future["未来硬件与云端整合目标 (Unified Device)"] D["硬件端实时语音交互"] --> E["云端自主 Agent 引擎"] E --> F["启动无头浏览器抓取"] E --> G["自动编写与执行代码"] E --> H["资产多轮批处理与常驻工作流"] end ``` 用户在电脑上用 Codex 处理完代码和数据后，想生成配套图片并关上笔记本电脑带走，由于上下文无法自动流转，只能人肉将提示词和上下文复制粘贴进手机端的 ChatGPT App。要让独立硬件真正发挥威力，不仅要靠单个 LLM 交互窗口内的知识检索，更必须打通后台自动化流程：调用浏览器爬取、编写并部署代码、定时批量处理数据（例如每天自动改写 40 张图片）。这些系统级集成是消费级 AI 硬件产品彻底站稳脚跟的前置条件。 --- ### [14:58](https://www.youtube.com/watch?v=OQH8s5qg1yw&t=898s) 判赔超 9 亿美元：Meta 新墨西哥州判决与社交成瘾法律风暴 新墨西哥州一名法官判决 Meta 公司向该州支付超过 9 亿美元，并对其旗下 Facebook 和 Instagram 施加新的未成年人保护限制。这是全美针对社交媒体成瘾诉讼浪潮的又一次重锤。该判决要求 Meta 设立一个 5.67 亿美元的专项基金，用于弥补平台对其未成年人造成的心理与精神伤害，此外还要加上陪审团此前裁决的 3.75 亿美元民事罚款，总额高达 9.42 亿美元，直逼 10 亿美元大关。 主持人回顾了这一系列诉讼的发展轨迹，指出诉讼金额正以惊人的指数级速度攀升： * **2026 年 3 月 25 日（洛杉矶判决）**：陪审团认定 Meta 与 Google/YouTube 在产品设计上存在过失，未能保护青少年免受成瘾危害。一名自幼患上社交媒体成瘾症的女性原告获赔 600 万美元（Meta 承担 4.2 亿美元，Google 承担 1.8 亿美元）。当时请上节目的专业律师就直言这仅仅是冰山一角，而主持人当时还一度怀疑这种小额个人判决掀不起风浪。 * **2026 年 5 月（肯塔基州学区和解案）**：肯塔基州一处学区起诉 Instagram 故意部署成瘾机制，诱发学生焦虑、抑郁及自残行为，迫使学校斥巨资增设心理健康服务。学区最初索赔逾 6,000 万美元，最终由 YouTube、TikTok 和 Snap 共同出资达成 2,700 万美元和解，虽然该案中平台均未承认法律责任也未被强制要求更改产品设计。 * **新墨西哥州巨额判决**：作为一个体量并不大的州，这一单州判罚直接突破 9 亿美元，预示着后续各州针对社交平台的索赔规模将进入天价阶段。 面对愈演愈烈的合规风暴，主持人将社交媒体当下的境遇与当年烟草行业的终局进行了类比，引出了著名的 1998 年《烟草主和解协议》（Tobacco Master Settlement Agreement）。早年烟草公司同样遭遇单兵作战式的法律围剿，庭审场面与如今扎克伯格等科技大佬在国会辩称“我们只是在卖广告”如出一辙。当时公众直觉认为受害者应当是那些因吸烟患癌的普通人，然而最终的超级和解协议却发生在各大烟草巨头与全美 46 个州政府之间。其中，烟草公司 Liggett Group（母公司为 Vector Group）因中途率先“反水”转做污点证人、检举揭发同行内幕，得以在历史性天价赔偿中全身而退，至今依然活得非常滋润。社交巨头未来是否也会被迫走上类似的行业大和解之路，正成为全行业屏息以待的焦点。 ### [18:45](https://www.youtube.com/watch?v=OQH8s5qg1yw&t=1125s) 烟草《主和解协议》（MSA）的历史启示与社交媒体监管隐喻 主持人围绕烟草巨头当年与美国各州的世纪博弈展开了深入剖析。当时，美国 46 个州、哥伦比亚特区以及数个海外领地联合对大型烟草公司提起诉讼。各州起诉的核心逻辑在于烟草产生的“负外部性”：吸烟导致大量居民罹患肺癌，直接打垮了各州的公共医疗财政预算。各州政府原本规划好了在全科医生、放射科、X光设备上的常规支出，却突然发现大批民众患上癌症，公立医疗系统根本无力招架，必须紧急增聘大批肿瘤专家、采购昂贵的放化疗设备与靶向抗癌药物。各州因而认定，医疗体系的严重赤字完全是由烟草公司造成的。尽管直觉上官司更应该由患癌的个体提起，但这场诉讼最终却演变为了州政府与烟草行业之间的系统性清算。 最终达成的《主和解协议》（Master Settlement Agreement, 简称 MSA）让各州同意彻底终止所有诉讼，不再逐州通过琐碎诉讼对企业“抽丝剥茧”。作为交换，烟草公司不仅要承担严格的广告营销限制（尤其是彻底禁止针对未成年人的营销），还必须支付巨额赔款。虽然名义上的公开头条金额为 2060 亿美元——按如今科技巨头的尺度看似不算惊人，毕竟 Meta（Facebook）一年的营收就接近 2000 亿美元，但倘若真有一张 2000 亿美元的罚单砸下来，对任何企业都将是灭顶之灾。更关键的机制在于，MSA 绝非一笔付清的单次罚款，而是一套永久性的年度付款体系：只要烟草公司还在营业，就必须永久向各州缴纳这笔“特殊税款”。具体支付额度还会根据卷烟销量、通胀率以及各公司的市场份额动态调整。这也解释了为什么卷烟销量一旦下滑，赔付总额就会减少，进而直接倒逼整个行业加速转向非卷烟类的替代产品。 各州政府获得这笔永续现金流后，华尔街的金融机构迅速蜂拥而至，将其高度金融化。投行与金融机构通过精密的数学模型，测算出新墨西哥州等特定州未来几十年的预期进账——例如今年预估 3 亿美元、明年 2.97 亿美元并逐年递减，随后向州政府提出打包方案：“我们现在就一次性付给你 40 亿美元现金，换取你未来这笔特定比例的烟草赔付款流。”州政府随即可以用这笔巨额贴现资金去翻修桥梁或兴建大型基建项目，甚至通过发行由 MSA 收入抵押的“烟草债券”进行借贷。这种体制不仅为烟草巨头构筑了免受各州进一步起诉的法律护城河，也建立起一套围绕卷烟销售运转的长期金融与监管结构。主持人 John 顺势调侃道，自己参与创立的尼古丁替代品牌 Lucy 恰好迎来了从 2016 年 8 月 8 日新规算起的十周年，别人眼中的“一夜成名”，背后其实是长达十年在无人问津的行业泥潭中苦苦打磨。 - **诉讼覆盖范围**：涵盖美国 46 个州、哥伦比亚特区及多个属地。 - **协议核心金额与条款**：名义金额 2060 亿美元；终止各州诉讼，换取永久年度浮动缴费及针对儿童营销的严厉禁令。 - **金融衍生操作**：各州将未来的 MSA 赔付现金流打包转让，或发行由其担保的市政债券以换取当下基建所需的一次性巨款。 - **行业映射**：主持人认为社交媒体行业目前因成瘾性与青少年心理健康面临的司法与监管风暴，极可能正在复制烟草 MSA 的历史演进路径。 --- ### [23:07](https://www.youtube.com/watch?v=OQH8s5qg1yw&t=1387s) SemiAnalysis 痛批 DeepMind：前沿实验室神话破灭？ 科技分析机构 SemiAnalysis 近期发布了极其尖锐的行业评论，对 Google 旗下的 DeepMind 进行了毫不留情的“掘坟式”批判，甚至被主持人形容为不仅在挖坟，还要“把 DeepMind 扔进去并在坟头跳舞”。SemiAnalysis 直言不讳地断言：“从一切实际意义来看，由于强化学习（RL）核心团队的大规模出走以及灾难性的算力资源分配，DeepMind 已经不再是一家前沿 AI 实验室。Google 固然会继续按部就班地漫步、发布后续模型，但他们重回 SOTA（行业顶尖水平）的概率已经跌至零点。” 这番评述坐实了 Google 内部顶尖 AI 人才持续流失的残酷现状。在顶级科研领域，人才密度决定一切——顶尖学者只想与同等水准的同行共事。此前 Noam Shazeer 的出走就如同煤矿里的金丝雀，释放出强烈的危险信号；如今这批出走者显然对所谓的“Gemini 4 Pro”缺乏信心与兴奋感。包括 Quoc Le（瞿国）、Oriol Vinyals 等人在内的多位高管出走，更折射出底层结构性矛盾。尤其耐人寻味的是，被公认为“Google 工程界历史最佳（GOAT）”、共同创立了 Google Brain 并亲手开启 GPU/TPU 研发项目的传奇人物 Jeff Dean，在面临内部资源争夺时也显得力不从心。 与此形成鲜明讽刺的是，Google 管理层做出了一个令人咋舌的战略决定：他们宁愿将海量关键算力通过长期合同大举出售给 Gemini 最强劲的死敌，也绝不把这批算力分配给自家内部的 DeepMind。SemiAnalysis 的供应链数据显示，在 2026 年第三季度至 2027 年第四季度期间，Google 超过 20% 的全部 TPU 出货份额，竟然被直接锁定并出售给了竞争对手 Anthropic。SemiAnalysis 直言，Google 的核心顽疾从来不在于 Jeff Dean 或 Noam Shazeer 本人，而在于其内部极端官僚化、动作迟缓且在战略上唯唯诺诺的避险文化；它正步 IBM 和 Intel 等传统科技巨头的后尘，主动放弃攻克最艰难的前沿技术，转而选择更容易赚快钱的稳妥商业模式。 ```mermaid flowchart TD accTitle: Google TPU Compute Allocation Conflict accDescr: Diagram illustrating how Google allocates TPU compute away from internal DeepMind to external rival Anthropic for immediate cloud revenue. subgraph Hardware ["硬件与基础设施层"] TPU["Google 自研 TPU 算力集群"] end subgraph Internal_Division ["内部战略割裂"] GCP["Google Cloud (GCP)<br/>专注云计算营收与商业化合同"] DM["Google DeepMind<br/>前沿模型研发 (Gemini 系列)"] end subgraph Market ["外部竞争格局"] Anthropic["Anthropic<br/>(Gemini 的核心竞争对手)"] end TPU -->|超 20% 出货量长期外售| GCP GCP -->|高确定性算力交付协议| Anthropic TPU -.->|算力分配遭到严重挤占| DM Anthropic -->|争夺行业前沿技术垄断| DM ``` - **SemiAnalysis 核心论断**：人才流失与算力资源分配失衡，让 DeepMind 彻底跌出顶级前沿梯队。 - **算力外流数据**：2026 年 Q3 到 2027 年 Q4 期间，超过 20% 的 TPU 总产量被 Google Cloud 长期锁定并直接供应给主要对手 Anthropic。 - **企业文化病灶**：机构认为 Google 正在重复经典巨头的衰退轨迹，极度官僚与战略短视促使多位顶尖技术灵魂人物相继离职。 --- ### [25:15](https://www.youtube.com/watch?v=OQH8s5qg1yw&t=1515s) 云计算赚钱机器与商业化路线选择的深层分歧 虽然 DeepMind 在模型前沿探索上面临重大阻力，但硬币的另一面是 Google Cloud（GCP）惊人的变现能力。GCP 目前已演变为一台不可思议的“印钞机”，这也从商业逻辑上解释了为何 Google 高管层会毫不犹豫地优先保全 GCP 的客户需求，而非自家 Gemini 研发团队的算力配额。从财务指标分析，向外部大客户大批量交付系统级算力的息税前利润率（EBIT Margin）虽然略低于 GCP 核心云业务（处于 30% 出头的低位），但整个 GCP 部门未来预计仍能维持在 35% 至 40% 的中高位水平。随着各大顶级实验室对超大规模算力的饥渴需求，未来预计会有更多吉瓦（Gigawatt）级别的算力采购大单陆续出炉；SemiAnalysis 预估，未来数个季度内 GCP 的未履约合同价值（RPO）有望增加超过 2500 亿美元的 TPU 相关订单储备。 主持人在探讨中指出，如果从纯商业角度换个视角来看，Google 这种看似保守的做法也可以被解读为一种务实的“杠铃策略”：Google 具备顶尖的芯片自主研发能力与全球顶级的云网络架构，同时拥有由 Google 搜索、YouTube 和广告业务构筑的坚固现金牛护城河，完全有底气为巨额的基础设施扩张托底。与其在模型军备竞赛中与其他前沿实验室死磕高风险的“梦之队”抢人大战，不如效仿微软或亚马逊等主流超大规模服务商（Hyperscalers），专心扮演算力军火商与底层基建运营商的角色，同样能获得资本市场丰厚的回报。 然而，这种纯工具化的商业转向却与早期 AI 先驱的崇高愿景背道而驰。Google 内部早在 ChatGPT 问世之前就早已做出了同等体验的原型产品，例如主导开发后来代码模型 Codex 的核心技术骨干 Thibault 当时就在 Google 内部推进类似项目，但产品最终却被体制压制未能推出。这揭示出 Google 最高决策层与 Demis Hassabis 等前沿实验室领导者之间在“AI 核心价值究竟沉淀在何处”这一根本问题上的深刻认知断层——管理层认为价值在底层硬件、基础设施与云服务，而研究派则坚信核心价值在于通用前沿大模型本身。John 最后还分享了作家 Sebastian Mallaby 在其著作中记录的一则轶事：当年 Demis 在考虑出售 DeepMind 时曾试探 Mark Zuckerberg 是否真正对通用 AI 抱有崇高信仰，特意发问求证，结果 Zuckerberg 却兴高采烈地表示自己正无比痴迷于虚拟现实（VR），反差鲜明而充满戏剧性。 - **GCP 财务指标预期**：TPU 硬件系统直售的 EBIT 利润率在 30% 左右，核心 GCP 综合 EBIT 利润率维持在 35%-39% 区间。 - **潜在在手订单规模**：SemiAnalysis 估算在接下来的几个季度内，GCP 积压的 TPU 预订订单可能为 RPO 贡献超过 2500 亿美元。 - **战略理念冲突**：管理层侧重高确定性的“卖铲人”商业模式（基础设施服务），与模型研究部门追求 AGI 顶尖技术突破的理念产生根本分道扬镳。 ### [27:45](https://www.youtube.com/watch?v=OQH8s5qg1yw&t=1665s) 谷歌对 AI 的认知分歧与向 Anthropic 倾斜的资本算盘 讨论从 Demis Hassabis 当年对 AI 的终极构想切入。Demis 曾明确表示，他希望合作的对象必须全情投入于 AI，并将其视为一种彻底颠覆性的全新底层技术，绝非普通的软硬件升级——它完全不同于 VR、新型智能设备、电动汽车或太空卫星，必须被当作一个全然独立的质变浪潮来看待。然而，主持人调侃道，Sundar Pichai 看起来似乎并不真正相信递归自我改进（RSI, Recursive Self-Improvement）或通用超级智能（ASI）的叙事；Demis 当年应该更尖锐地追问谷歌高层：“你们对云计算有多兴奋？对搜索里的 AI Overviews 有多兴奋？”如果谷歌当年回答说“我们对云计算基础设施和对 ASI 一样兴奋”，Demis 当时就该警惕并抽身离开。 不过，从另一个角度来看，谷歌高层也可能是过于笃定 RSI 的未来，意识到内部机制面临瓶颈，因此干脆“开来布林克押运车（Brinks truck）”给竞品注资输血。他们为 Anthropic 组建了近四分之一万亿美元（2500 亿美元级别）的庞大融资与数据中心保障方案，让缺少债务市场准入能力的 Anthropic 得以顺利扩张算力。值得注意的是，谷歌在 Anthropic 的持股上限被严格锁定在 15%，目前大约持有 14%。Anthropic 没有给予谷歌任何投票权，甚至不给董事会观察员席位或正式董事席位，将谷歌完全隔离成了一个纯粹的财务出资方。 ```mermaid flowchart TD accTitle: Google and Anthropic Relationship accDescr: Diagram illustrating Google's financial relationship with Anthropic despite strict structural guardrails. subgraph GoogleCapital["谷歌资本与算力支撑"] G1["云基础设施与数据中心担保"] G2["巨额资金支持"] end subgraph Guardrails["Anthropic 设置的硬性防火墙"] A1["持股上限约 15%（现约 14%）"] A2["无投票权 (No Voting Rights)"] A3["无董事会席位及观察员身份"] end subgraph Scale["Anthropic 规模扩张"] S1["跳过直接债务市场准入限制"] S2["加速前沿大模型与算力扩展"] end GoogleCapital --> Guardrails Guardrails --> Scale ``` - Demis Hassabis 要求合作方必须将 AI 视为脱离常规科技（如 VR、硬件、卫星）的独立底层跃迁。 - 谷歌与 Anthropic 的特殊协议：提供庞大算力与数据中心背书，但持股被卡在 15% 上限以内（约占 14%）。 - 治理权彻底切割：谷歌仅充当“纯财务支持方”，无董事席位、无观察席位、无投票权。 --- ### [29:31](https://www.youtube.com/watch?v=OQH8s5qg1yw&t=1771s) 豪宅板块：品牌帝国操盘手挂牌 6850 万美元迈阿密海景庄园 节目进入标志性的“豪宅板块（The Mansion Section）”。话题焦点是品牌授权巨头 Authentic Brands Group（ABG）的前总裁兼首席营销官 Nick Woodhouse 及其妻子 Joseline Woodhouse 挂牌出售的迈阿密海滨豪宅，标价高达 6850 万美元。2019 年 Nick 在迈阿密乘帆船庆祝生日时便认定这里是未来的家，并在 2020 年从纽约搬迁至佛罗里达州阳光岛海滩（Sunny Isles Beach）的一套公寓。 随后，他们在一艘朋友的船上相中了位于迈阿密海滩带门禁安保的 La Gorce Island（岛名发音在节目中引发了一番打趣纠正）的一块滨水地皮。当时这块地仅建好了房屋地基，夫妇俩于 2021 年以 1700 万美元抄底买下这栋半成品，并于 2023 年左右建成了这栋占地 0.4 英亩、室内面积约 8,800 平方英尺的 7 卧现代风格海景大宅。如今他们以 6850 万美元挂牌转手，原因仅仅是他们打算在附近再建一栋新房。 - 资产坐落于迈阿密海滩知名封闭式豪宅区 La Gorce Island，占地 0.4 英亩。 - 2021 年以 1700 万美元买下打完地基的半成品，耗时两年打造出 8,800 平方英尺、7 间卧室的现代豪宅。 - 两年完工后随即以 6850 万美元挂牌上市，套利空间巨大。 --- ### [30:46](https://www.youtube.com/watch?v=OQH8s5qg1yw&t=1846s) Authentic Brands Group：标志性品牌的“最终墓地”还是商业引擎？ Nick Woodhouse 的前东家 Authentic Brands Group（ABG）随即引发了主持人们的热烈探讨。ABG 是一个极度庞大的消费品牌与名人人设（IP）运作帝国，旗下不仅聚合了 50 多个知名消费品牌，还垄断了包括拳王阿里（Muhammad Ali）、猫王（Elvis Presley）以及玛丽莲·梦露（Marilyn Monroe）等传奇巨星的身后肖像权与遗产商业权益——“只要你在市面上的 T 恤上看到玛丽莲·梦露的肖像，ABG 就能躺着分到一张支票”。 | 品牌/IP 类别 | 代表资产 | 商业模式与节目探讨细节 | |:--- |:--- |:--- | | **名人肖像与遗产管理** | 玛丽莲·梦露、猫王、穆罕默德·阿里 | 2011 年拿下梦露肖像权；全球任何周边衍生品均需向其支付版税 | | **格斗与经典潮流** | Tapout、Silverstar | 早期核心资产；主持人打趣调侃周末出街必备的硬汉格斗服饰 | | **滑板与冲浪文化** | Billabong（比拉邦）、DC Shoes、RVCA | 承载 80/90 后青春记忆，现被集中收购运营，但品牌文化被指流失 | | **主流服饰与零售** | Reebok、Champion、Brooks Brothers、Barneys New York | 陷入经营困境的破产名牌被其打包低价收购，靠轻资产授权变现 | 两位主持人与制作人调侃道，ABG 实际上就是消费品牌界的“Bending Spoons”，专门低价打包收购那些陷入困局但声名显赫的品牌。观众也留言吐槽 ABG 简直是“标志性品牌的墓地（a graveyard for iconic brands）”。对于从小看冲浪电影、关注滑板选手的千禧一代来说，Billabong、RVCA、DC Shoes 等品牌的沉沦令人唏嘘。如今不生活在海边的年轻一代根本不在乎冲浪文化，转而去追捧 Chrome Hearts（克罗心）。主持人借机开玩笑提议，美国财政部应该直接拨出“几十亿美元专项资金作为联邦救助兜底”，来挽救 Volcom、Quiksilver 和 Billabong 这些濒临消亡的极限运动品牌。 节目在轻松的周五告别中收尾。主播们预告周末将在 TBPN 的“Ultra Dome”带来持续报道，提醒听众关注他们在 Getty Images 上释出的前沿花絮，并呼吁大家在 Apple Podcasts 与 Spotify 留下五星好评，订阅官网 tbpn.com 的每日邮件通讯。 ## Notable quotes > [00:05](https://www.youtube.com/watch?v=OQH8s5qg1yw&t=5s) 我戴墨镜是因为我正在眺望未来，而未来确实太刺眼了。 — Jordi Hays > [01:05](https://www.youtube.com/watch?v=OQH8s5qg1yw&t=65s) 哎呀，自然界原有的病毒好像还不够多呢，咱们再搞点新的吧！ — John Coogan / Jordi Hays > [01:05](https://www.youtube.com/watch?v=OQH8s5qg1yw&t=65s) 肠道菌群也是人体的关键部分，难道就没人站出来为这些无辜的细菌发声吗？ — Jordi Hays > [09:10](https://www.youtube.com/watch?v=OQH8s5qg1yw&t=550s) 兄弟，我们当年甚至连原著都不读，直接看 SparkNotes 摘要就能写出来。 — John Coogan / Jordi Hays > [18:45](https://www.youtube.com/watch?v=OQH8s5qg1yw&t=1125s) 我们现在就一次性付给你 40 亿美元现金，换取你未来这笔特定比例的烟草赔付款流。 — John Coogan > [23:07](https://www.youtube.com/watch?v=OQH8s5qg1yw&t=1387s) 从一切实际意义来看，由于强化学习（RL）核心团队的大规模出走以及灾难性的算力资源分配，DeepMind 已经不再是一家前沿 AI 实验室。Google 固然会继续按部就班地漫步、发布后续模型，但他们重回 SOTA（行业顶尖水平）的概率已经跌至零点。 — SemiAnalysis > [27:45](https://www.youtube.com/watch?v=OQH8s5qg1yw&t=1665s) 你们对云计算有多兴奋？对搜索里的 AI Overviews 有多兴奋？ — John Coogan / Jordi Hays > [30:46](https://www.youtube.com/watch?v=OQH8s5qg1yw&t=1846s) 只要你在市面上的 T 恤上看到玛丽莲·梦露的肖像，ABG 就能躺着分到一张支票。 — John Coogan / Jordi Hays > [30:46](https://www.youtube.com/watch?v=OQH8s5qg1yw&t=1846s) 标志性品牌的墓地（a graveyard for iconic brands） — 观众留言 ## Mentioned TBPN, John Coogan, Jordi Hays, Khosla Ventures, Discovery Loop, Jeff Dean, Databricks, Patrick Wendell, Science, Stanford University, Arc Institute, Phi X 174, Eliezer Yudkowsky, Elon Musk, New York Times, Mark Gurman, Bloomberg, OpenAI, ChatGPT, Voice Mode, Jony Ive, Google, Google DeepMind, NotebookLM, Sam Altman, Codex, Meta, Facebook, Instagram, YouTube, TikTok, Snap, Liggett Group, Vector Group, Master Settlement Agreement (MSA), Lucy, SemiAnalysis, Noam Shazeer, Quoc Le, Oriol Vinyals, Google Brain, Anthropic, Google Cloud (GCP), Demis Hassabis, Mark Zuckerberg, Sebastian Mallaby, Sundar Pichai, Nick Woodhouse, Joseline Woodhouse, Authentic Brands Group (ABG), La Gorce Island, Muhammad Ali, Elvis Presley, Marilyn Monroe, Tapout, Silverstar, Billabong, DC Shoes, RVCA, Reebok, Champion, Brooks Brothers, Barneys New York, Bending Spoons, Chrome Hearts, Volcom, Quiksilver ## 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 Chapters: 00:00 AI-Designed Viruses + The Biosecurity Question 07:15 OpenAI’s First Device: The $300–$400 “Donut” 09:03 NotebookLM, Voice Mode + The Future of Personal AI 14:58 Meta Hit With $900M Judgment + Social Media Addiction 23:07 Is DeepMind No Longer a Frontier AI Lab? 29:31 The Mansion Section [Source](https://www.youtube.com/watch?v=OQH8s5qg1yw)

[0:05] Jordi Hays: 我戴墨镜是因为我正在眺望未来，而未来确实太刺眼了。 [1:05] John Coogan / Jordi Hays: 哎呀，自然界原有的病毒好像还不够多呢，咱们再搞点新的吧！ [1:05] Jordi Hays: 肠道菌群也是人体的关键部分，难道就没人站出来为这些无辜的细菌发声吗？ [9:10] John Coogan / Jordi Hays: 兄弟，我们当年甚至连原著都不读，直接看 SparkNotes 摘要就能写出来。 [18:45] John Coogan: 我们现在就一次性付给你 40 亿美元现金，换取你未来这笔特定比例的烟草赔付款流。 [23:07] SemiAnalysis: 从一切实际意义来看，由于强化学习（RL）核心团队的大规模出走以及灾难性的算力资源分配，DeepMind 已经不再是一家前沿 AI 实验室。Google 固然会继续按部就班地漫步、发布后续模型，但他们重回 SOTA（行业顶尖水平）的概率已经跌至零点。 [27:45] John Coogan / Jordi Hays: 你们对云计算有多兴奋？对搜索里的 AI Overviews 有多兴奋？ [30:46] John Coogan / Jordi Hays: 只要你在市面上的 T 恤上看到玛丽莲·梦露的肖像，ABG 就能躺着分到一张支票。 [30:46] 观众留言: 标志性品牌的墓地（a graveyard for iconic brands）

[1:05] Jordi Hays: You never say, 'I got a virus,' and someone asks, 'A good one or a bad one?' It's always bad. [9:03] John Coogan: If the stated preference is 'I don't like AI,' but the revealed preference is 'it's kind of nice to have this thing around the house,' how do you measure those two things? [17:20] The result is not just a legal settlement. It created this long-term financial and regulatory structure built around cigarette sales. [23:07] For all intents and purposes, we believe DeepMind is no longer a frontier lab due to large numbers of departures from their RL teams and poor compute allocation... Google will continue meandering on and releasing models, but their odds of ever reaching state-of-the-art again have dropped to zero.

[0:05] Who had orange in the chat? Because the [0:08] chat was trying to guess what color [0:10] sunglasses Jordan would be wearing [0:12] today. Is that orange or is that more of [0:14] like a burnt sienna? [0:15] >> Well, I think it's more of a green [0:17] greenish frame [0:19] >> with more of with an orange lens. [0:20] >> You got to drop the line. I'm wearing [0:22] sunglasses cuz I'm looking at the future [0:24] and it's very bright indeed. I don't [0:26] know. No one gets that reference. [0:28] Anyway, [0:33] >> oh, [0:33] >> uh, it's great to be back. It's Friday. [0:36] We got a shorter show for everyone [0:38] today. General partner at Coastal [0:40] Ventures coming on to talk about their [0:42] new investment, Discovery Loop. [0:44] >> Yeah. [0:44] >> Founded by none other than Jeff Dean and [0:47] some of his crew from Deep Mind. And [0:50] then we have Patrick Wendell, co-founder [0:51] of Datab Bricks, [0:53] >> joining to talk about their new blog [0:56] post. I promise it's going to be more [0:58] exciting than it sounds like. Um, no, [1:01] but they're doing smarter model routing [1:02] and uh we're excited to catch up with [1:04] him. [1:05] >> Very excited. Well, for the first time, [1:07] scientists have used AI to create [1:09] entirely new viruses. You [1:12] >> You asked for it. They delivered. [1:15] >> They delivered. [1:16] >> They delivered. [1:18] >> They woke up. They said, "You know what? [1:19] We don't have enough of [1:20] >> viruses that have never existed in [1:22] nature before." It's marked a milestone [1:24] that could accelerate biotechnology [1:26] while also raising long-term biocurity [1:28] questions. Of course, nightmare scenario [1:31] is you go to Best Buy, you get a gaming [1:33] PC with a couple pretty stock graphics [1:36] cards, and you're able to run an [1:37] open-source model that basically walks [1:40] you through the steps of creating [1:41] something really problematic. At the [1:43] same time, there's a lot of really [1:45] talented and well-resourced [1:46] organizations that are fighting that [1:48] tooth and nail. And so we'll probably [1:50] see a little little uh you know back and [1:52] forth equilibrium there. But uh lots of [1:55] interesting questions. Important to note [1:57] that these viruses do not affect humans [2:00] whatsoever. Even these new viruses uh [2:02] they target uh bacteria. So it's more of [2:05] an experiment, more of a demo, but you [2:07] can see where things are going. And um [2:09] are [2:10] >> that doesn't make me that comforted to [2:12] be honest because you like your bacter [2:14] you know we naturally have we're there's [2:16] bacteria [2:18] >> that is uh it's part of being human [2:20] right [2:20] >> and you would like the the bacteria to [2:21] be unaffected by viruses. Is anyone [2:24] standing up for the bacteria right now? [2:25] >> Well it's more so like you have bacteria [2:27] you have bacteria in your gut [2:29] >> and gut is kind of an important part of [2:32] >> now that bacteria is going to be [2:33] suffering from a novel virus apparently. [2:35] Yeah. [2:35] >> Uh well, in a study published Thursday [2:37] in Science, researchers at Stanford and [2:39] the ARK Institute trained an AI model to [2:42] recognize patterns in naturally [2:43] occurring viral DNA, then used it to [2:47] generate genetic sequences for brand new [2:49] viruses. After synthesizing those DNA [2:52] sequences and inserting them into [2:54] bacteria, the team found the bacteria [2:56] produced viable viruses capable of [2:58] infecting other bacteria. The work does [3:01] not create a new threat to humans. Be [3:03] careful. I'm sure that will get cut out [3:05] of a lot of uh messaging around this [3:07] because it sounds really scary on its [3:09] face. The AI was trained only on [3:11] bacteria phages, viruses that infect [3:13] bacteria, and specifically excluded [3:15] viruses that infect humans, plants, [3:17] animals, and fungi. As a result, the [3:20] model cannot generate viruses capable of [3:22] infecting people. While scientists have [3:24] been synthesizing viral genomes for [3:26] years to study diseases and develop [3:27] vaccines, this is the first time AI has [3:30] been used to design entirely new viruses [3:32] that function in the real world. And I [3:33] was going back and forth before the show [3:35] on is this anything special? You know, [3:38] we've been using tools to make viruses [3:40] for bacteria for a long time. This is [3:41] just another tool to do it. Or is this [3:43] some, you know, material breakthrough? [3:45] That's sort of a debate point. It sounds [3:46] really scary, but also like you've been [3:48] able to design these viruses in a lab [3:50] for a long time. So what is the material [3:53] difference here? I think it all comes [3:55] back to acceleration and cost. If all of [3:57] a sudden it becomes a thousand times [3:59] cheaper to generate viruses, then that [4:01] could reshape biotech in a positive way, [4:03] but also have biocurity implications. [4:06] >> Yeah, I think a lot of the reaction is [4:08] just that it feels like uh little too [4:10] soon post Wuhan, a little too soon post [4:13] hugging phase. There's been a variety of [4:15] events that I think naturally make [4:18] people a little apprehensive when you [4:20] see a headline like this. [4:21] >> Yeah, a lot of people are like, you [4:22] know, the Eleaser Udowski reaction of ah [4:26] like this is bad. Um but we'll see where [4:28] it goes. Uh if the approach proves [4:30] effective across other classes of [4:31] viruses, it could become a powerful tool [4:33] for medicine and biotechnology. Viruses [4:35] are already widely used as delivery [4:38] vehicles for gene therapies and other [4:40] medical treatments and AI designed [4:42] viruses could eventually expand that [4:44] toolkit. At the same time, the research [4:46] highlights how advanced how advances in [4:48] AI are making it increasingly important [4:50] to build safeguards alongside new [4:51] capabilities which I'm sure the ARC [4:52] Institute is working on. Um, and there's [4:54] also other uh other AIdriven [4:57] neoccientific labs. I mean Jeff Dean's [5:00] one of those was advancing one of the [5:02] goals of his new company is to work on [5:05] biotech broadly uh he has a very broad [5:08] remit but there are more narrow projects [5:10] like that new company that's focused on [5:12] finding a cure for the common cold. [5:14] There are a few other projects that are [5:15] more narrowly targeted as well as all [5:17] the biotech companies that are working [5:18] on stuff. If we're going to get advances [5:21] in viruses that deliver gene therapies [5:24] or medical treatments, we got to rebrand [5:27] virus. We got to come up with a new [5:28] word. Just like GLP1 peptide that felt [5:32] very safe. It was like you're not doing [5:34] steroids. You're not on gear. [5:36] >> What's the lizard? [5:37] >> You're doing a peptide. [5:38] >> You're not doing Hila Monster Venom. [5:40] >> Hila Monster Venom. Exactly. Exactly. [5:41] >> You're not doing Hila Monster Venom. [5:43] It's just some Chinese peptides. [5:46] >> The Chinese peptides was a rough go. But [5:48] in general, I think the fact that it was [5:50] like just a peptide, it felt much more [5:53] welcoming as opposed to being in the [5:55] world of the steroids, the [5:57] performance-enhancing drugs, and it [5:59] became easier for people to jump in. Oh, [6:01] I got to learn about peptides. Oh, there [6:03] oh, there's naturally occurring [6:04] peptides. Cool. I'm into that. But virus [6:06] has such a bad connotation post Wuhan as [6:08] well as just everything. You're never [6:11] like, I got a virus. And someone says, a [6:13] good one or a bad one. Like, it's always [6:14] bad. Yeah, it's never good. Um, but so [6:17] they need they need a new brand for that [6:19] if they're going to commercialize that [6:20] for sure, but we'll see. Anyway, uh [6:23] there's a whole article in the New York [6:24] Times about it. This AI just created [6:27] viruses not found in nature. The new [6:29] study published Thursday in science goes [6:32] well beyond duplicating viral genes. [6:34] Scientists at Sanford University and Arc [6:36] Institute taught AI to recognize [6:38] patterns in of DNA in nature and then [6:41] use that DNA to write recipes for [6:43] entirely new viruses. The viruses [6:45] dreamed up by AI do not pose a threat to [6:47] humans because they are all similar to a [6:49] naturally occurring virus called FX1 [6:52] 174, which can only infect bacteria. [6:56] Really hardcore name for something [6:58] that's not that dangerous. Fix 174 [7:01] sounds like a offspring of Elon Musk. [7:03] Governments and scientific organizations [7:04] have been slow to develop guardrails [7:06] that could block the creation of a [7:07] deadly virus even as the science races [7:10] ahead. So, I don't think they'll be open [7:12] sourcing this anytime soon. But there is [7:14] other news. Uh Mark Gurman's been [7:16] reporting on OpenAI's first consumer [7:18] device. Uh it'll reportedly look like a [7:20] hockey puck-sized donut. Mark German. [7:22] The Germinator. [7:22] >> Oh, the Germinator. [7:23] >> The Germinator. Yeah. OpenAI's first [7:25] consumer device. Will reportedly look [7:27] like a hockey puck-sized donut and cost [7:30] roughly $300 to $400. Uh what a funny [7:34] form factor. [7:34] >> Love hockey pucks. [7:36] >> Yeah, [7:36] >> love donuts. [7:37] >> Okay, so you're in. [7:38] >> So I I like where this is going. [7:39] >> There's also a rumor that it has [7:41] mechanical pieces on it. So, I think it [7:43] can like undulate potentially. The the [7:45] speculation's all over the place. [7:46] According to Bloomberg's Mark German, [7:48] the Germinator, as you put it, uh, the [7:50] battery powered device is essentially a [7:51] portable smart speaker without a screen [7:54] designed to be carried around the house [7:55] or placed on a nightstand or kitchen [7:57] counter. It will include speakers and [7:58] microphones along with cameras and other [8:00] sensors that allow its AI to perceive [8:02] what's happening around it. The device [8:03] is intended to work much more uh like a [8:07] much more capable version of ChatP's [8:08] current voice mode, learning about its [8:10] owner over time and using that context [8:12] to make conversations more personalized. [8:14] OpenAI is also making the hardware [8:16] itself feel more expressive. The device [8:18] will reportedly include lights and parts [8:20] that physically move as it responds. Uh [8:23] with the goal of making it feel more [8:24] alive than existing stationary smart [8:27] speakers from Amazon and Google. The [8:29] product being developed by Johnny Ob's [8:31] design team is expected to arrive in [8:32] 2027. [8:34] Envisioned as the first in a broader [8:37] family of OpenAI hardware long-term, the [8:39] company reportedly hopes to develop AI [8:41] devices capable of taking over some of [8:43] the functions now handled by [8:44] smartphones. Jordan, [8:45] >> feature request. [8:46] >> Yes. [8:46] >> Rolling flashbang. [8:48] >> Flashbang. That would be a feature [8:50] request. [8:51] >> Yeah. It says it has lights. It's got [8:53] sound. You should be able to use this as [8:55] a onthe-go flashbang. Yeah. So you're [8:58] going to hang out with some friends. [9:00] >> Yeah. [9:00] >> You want to prank them a little bit. [9:01] Yeah. When you're when you're kind of [9:03] coming in. [9:03] >> Yeah. I [9:04] >> I do I I I was reflecting on uh the the [9:07] deep mind story of uh you know the [9:10] departures there and the question of you [9:12] know how the models are progressing [9:13] versus the commercialization of those [9:15] models. There's some real strong points. [9:17] There's some weaker points within the uh [9:20] within the roll out of uh Google's AI [9:22] strategy. And I was I was thinking about [9:25] like what happened to notebook l because [9:27] that was heralded as a very magical [9:30] technology. You would uh you know give [9:33] it some sources a particular report uh [9:35] and it would just generate a podcast [9:38] talking between two different people uh [9:41] much more conversational and a lot of [9:42] people like consuming information that [9:44] way. So you could just go read a deep [9:47] research report on the history of, you [9:49] know, how uh bacteria phages work if you [9:52] want to get up to speed on that because [9:54] you're going to you're trying to [9:55] understand what's what the ARC Institute [9:57] is working on with these new viruses. Uh [9:59] you could go to Notebook LM and say, [10:01] "Hey, uh why don't you generate me a [10:03] podcast of, you know, two scientists [10:06] explaining this at, you know, high [10:09] school level and then take it into [10:10] college level." Tyler in the YouTube [10:12] chat says, "Nut loves notebook LM. Use [10:14] it weekly." [10:15] >> Use it weekly. [10:16] >> I always thought that I would have used [10:17] it when I was in college. [10:18] >> Yeah. [10:19] >> And I needed to, let's say, write a [10:22] paper on something and I wasn't super [10:25] prepared. I could say, "Generate me an [10:27] hourong podcast about this set of [10:29] topics." And I'd listen to that and then [10:31] I could probably [10:32] >> rip the paper. [10:33] >> Yeah. [10:34] >> I I did that for a history exam. I It [10:36] was like a vocab list and it went [10:38] through [10:38] >> Okay. [10:39] >> Yeah. middle-aged history, I think. [10:40] >> Wait, oh, you used notebook LM? [10:42] >> Yeah, [10:43] >> it generated podcast. So, I basically [10:44] fed in a vocab list of like dates and [10:47] various things. Then I had it. [10:48] >> Wait, and was it just a general? [10:52] >> Uh, it was really easy. I think I [10:54] probably aced it. [10:55] >> Wow, [10:56] >> there we go. [11:00] >> He needs a better confidence monitor. [11:01] I'm a big fan of that new uh trend [11:04] that's like uh asking old people like [11:06] how did you write a five paragraph essay [11:08] without AI and then the answer is like [11:10] buddy we we we wrote a five paragraph [11:13] essay without even reading the book [11:17] you just go to spark notes or something [11:19] but I was I was interested in in the [11:22] evolution of notebook because it's this [11:23] like there's this collapsing of [11:25] capabilities where uh Sam Alman was [11:28] recently sort of dragged a little bit [11:30] for saying like he would used Chachi [11:32] work to generate a podcast about what's [11:36] going on on his calendar and the family [11:38] life and stuff and people are like how [11:39] about you just talk to your kids but uh [11:41] the the more interesting technical point [11:43] on that is that do you even need chatpt [11:46] work to generate you that podcast or [11:48] will the voice mode and the memory [11:50] feature have enough context to just sit [11:54] there and talk to you like it's a [11:56] podcast on the fly [11:57] >> like live voice [11:59] >> if you're trying to learn about a topic [12:00] for example pretty good [12:02] >> is probably better than just generating [12:03] a podcast because a podcast assumes [12:06] >> like some certain understanding. Maybe [12:08] it's maybe maybe it thinks you [12:10] understand too much or too little [12:12] whereas voice you can be like go down [12:13] this like [12:14] >> Exactly. Exactly. It's a choose your own [12:16] adventure. It's a it's an expert call. [12:18] It's instead of notebook LM, it's Tus [12:21] LM. So basically I mean you're talking [12:24] to an expert and you can just guide the [12:26] conversation wherever you go. So, uh, [12:29] we'll be interesting to see where this [12:30] goes, what the reception is like. Uh, I [12:34] mean, huge huge delta divergence between [12:38] like the the social media push back for [12:41] chatbt versus like the app store [12:44] ratings. Like there's like a billion [12:45] people using it. A lot of people just [12:47] like the product and then there'll be [12:48] like someone dunking on it to the tune [12:50] of a million likes on Instagram. And so, [12:53] how do you measure those two things when [12:55] it comes to an actual consumer product? [12:57] like if it's delivering something good, [12:58] if people are like if the if the stated [13:00] preference is like I don't like AI, but [13:02] the revealed preference is like it's [13:04] kind of nice to have this thing around [13:05] the house. It's kind of useful. Um [13:07] interesting to see where it goes. The [13:09] launch of this device will be uh very [13:12] very interesting to see like how things [13:14] come together like you can see the [13:17] chacht work chach codeex chach like [13:20] coming together into one product but [13:22] there's a world where this product [13:24] launches with voice mode and it's not [13:26] really capable of linking to a cloud [13:29] codeex instance and writing you software [13:32] and doing the more advanced things that [13:34] are required just to accomplish some [13:36] tasks like you can go to chbt and ask it [13:39] to pull down an image from the internet, [13:41] restyle it, change it, but you can't [13:44] really tell it to do like 40 of those or [13:47] like every day forever do a whole host a [13:50] whole workflow. But you can in codeex [13:52] and you can talk to codecs but it has to [13:54] be running on a computer and and and I [13:57] would hope that by the time this [13:59] launches there's full like full context [14:02] like I I was doing some work in codeex [14:04] and I had the output and then I wanted [14:06] to like generate an image based on that [14:08] and take it on the go but I wanted to be [14:10] able to close my laptop and not and [14:12] still be able to access it. So I had to [14:13] like copy the context window into chat [14:15] GBT just the normal app so then I could [14:18] like access that information and [14:20] continue to transform it. Uh and that [14:22] was something that is like a very very [14:24] temporary thing that feels like it's [14:25] going to be fixed in like a couple [14:26] weeks. But there's a whole bunch of [14:27] these little minor integration issues [14:30] that probably need to be fulfilled [14:31] before this product launches and [14:33] delivers like the full the full [14:35] capability of what you can do because so [14:38] many things so many tasks instead of [14:41] just knowledge retrieval uh require [14:43] actually lo firing up a browser, [14:46] scraping it, writing some code, [14:48] downloading things, you know, setting up [14:50] an actual uh service and and workflow as [14:53] opposed to just something that can be [14:54] done within the context window of a [14:55] single LLM interaction. ction. Last top [14:58] story, a New Mexico judge has ordered [15:00] Meta to pay more than $900 million and [15:03] impose new restrictions on how minors in [15:06] the state use Facebook and Instagram. [15:09] This is a continuation of the social [15:11] media addiction. If you're watching this [15:13] clip on Instagram, let us know in the [15:15] comments. Are you addicted to TBPN reels [15:18] on Instagram? [15:20] It's not our fault. Apparently, it's [15:21] apparently it's Facebook's fault if we [15:23] if we got you hooked on this stuff. The [15:25] ruling requires Meta to establish a $567 [15:30] million fund aimed at addressing harms [15:33] linked to its platforms on top of a $375 [15:37] million in civil penalties previously [15:40] awarded by a jury. So they're up 900 [15:42] million. They're very close to a billion [15:44] dollars and the numbers are going to get [15:45] bigger from here. Yeah. At least at [15:47] least they're going to try. [15:48] >> Okay. So last time we really covered [15:51] something from this ongoing saga. There [15:53] was a verdict in on March 25th. A Los [15:57] Angeles jury found Meta and [15:59] Google/YouTube negligent for designing [16:01] platforms harmful to young people. [16:04] >> A woman who said she became addicted to [16:06] social media as a child was awarded 6 [16:07] million, 4.2 million against Meta and [16:09] 1.8 million against Google. And and [16:11] again, at the time we had this lawyer [16:14] on, he was giving his opinion. He was [16:15] like, "This is just the start." We were [16:17] like, [16:18] >> I was like, "No way. [16:18] >> The jury has a has has a verdict." [16:21] >> Yeah. to said, you know, talking to the [16:24] biggest companies in the world, you must [16:25] pay $6 million, right? It's like this [16:28] tiny amount, but it was to one person, [16:30] right? And so, you can imagine [16:31] >> as these cases evolve, um this this new [16:34] one is in New Mexico. [16:35] >> New Mexico is not the biggest state in [16:37] the union. It's not the biggest, [16:38] >> but there was also one uh in May 2026, [16:41] so just a couple months ago, for 9 [16:43] million uh to a school district in [16:46] Kentucky. Yeah. [16:47] >> Same sort of issue. Uh the lawsuit [16:49] accused Instagram of deliberately using [16:52] addictive features that contributed to [16:54] anxiety, depression, self harm, and [16:55] other problems among students, forcing [16:57] schools to spend more on mental health [17:00] services. The district had sought more [17:02] than 60 million. And I guess the total [17:04] payout was 27 million because it was [17:06] split between YouTube and Tik Tok and [17:09] Snap. Yep. [17:10] >> Um and in that case, there was no [17:12] admission of liability and no required [17:13] product changes. So I think it's time to [17:15] start thinking about what the the sort [17:18] of battle that social media has ahead of [17:20] it kind of in in cigarette terms. Tell [17:23] us about [17:24] >> uh [17:24] >> tobacco masters agreement. Yeah. [17:26] >> Yeah. Exactly. So the the tobacco [17:29] companies wound up getting sued [17:32] individually uh initially and there was [17:34] there were a whole bunch of court [17:35] hearings very similar to the senator we [17:37] sell ads moments um uh but uh more [17:41] focused on the cancercausing nature of [17:44] cigarettes and the question was like who [17:48] is ultimately being harmed economically [17:50] and you would think it's obvious uh the [17:53] person that saw an ad that made smoking [17:56] cool. They picked up a pack of [17:58] cigarettes. They started smoking and [18:00] then they got cancer and their life was [18:03] cut short. They are the victim. They [18:06] should be paid by the tobacco company. [18:08] That's what would be very logical. Uh [18:10] that's not what happened. In fact, the [18:12] tobacco master settlement agreement uh [18:15] landed in 1998 and it was agreement [18:17] between all of the major tobacco [18:19] companies. There's more nuance to this. [18:21] One of them broke loose and like [18:22] testified against the others. It's a [18:24] crazy story, but that's for another [18:25] time. uh in 46 states. [18:27] >> I'm the good cigarette company [18:29] >> basically. Yeah. And so they don't have [18:31] to pay. They're not part of the [18:32] settlement. So they don't have to pay [18:34] because they basically snitched on all [18:36] the others. It's crazy. [18:37] >> Are they still in business cigarette [18:39] company? [18:39] >> They're doing great. [18:40] >> Who are they? [18:41] >> Legette Victor. [18:41] >> I've never heard of it. [18:42] >> Yeah. Uh [18:43] >> last name Victor. [18:45] >> Yeah. They were literally they won. [18:46] >> They won. Yeah. There's more nuance to [18:48] it than that, but uh but that's like one [18:50] way to tell a story. Anyway, uh so uh a [18:52] bunch of the big tobacco companies [18:54] versus uh 46 of the United of the US [18:57] states, the District of Columbia and [18:58] several territories. The states agreed [19:00] to end major lawsuits against the [19:02] tobacco companies. So the same thing was [19:04] happening where all the different states [19:05] were suing and they were suing because [19:08] the negative externality of cigarettes [19:11] causing cancer was driving up medical [19:13] bills in the states. So the states have [19:15] health care and they assume, hey, okay, [19:17] we're going to spend this much on [19:19] doctors, this much on radiology, this [19:21] much on X-ray equipment, and then all of [19:23] a sudden they're starting to look at [19:24] their populations and saying like, wait, [19:26] everyone's getting lung cancer. We're [19:28] not equipped to deal with lung cancer. [19:29] We need to hire way more oncologists and [19:32] cancer doctors who specialize in lung [19:34] cancer. We need equipment. We need drugs [19:35] that treat lung cancer. We There's a [19:37] whole bunch of other things that we have [19:38] to spend money on. And so you got to pay [19:41] us because you, the tobacco company, are [19:43] responsible for us running out of money [19:45] for our health care system. And so that [19:47] was the nature of these like the battles [19:49] between the states and the big tobacco [19:52] companies. You would think it would be [19:54] the individuals who got the cancer. That [19:55] would be very logical, but that's not [19:57] actually the structure of this deal. And [19:59] so in return, uh, all the states said, [20:01] "Hey, we'll drop all those lawsuits. You [20:03] won't have to we won't be nickeling [20:05] nickel and dimeming you across every [20:06] single state." Instead the the companies [20:10] will uh make large payments uh and [20:13] follow new limits on advertising and [20:14] business practices. So the end result [20:16] the headline number is 206 billion uh [20:19] which feels like small relative to [20:22] today's standards of like hyperscalers [20:25] in social media. Facebook generates [20:27] roughly 200 billion in revenue every [20:29] year. Although if they got hit with a [20:31] $200 billion fine that would be [20:32] existential. But uh what happened with [20:35] the master settlement agreement with the [20:36] tobacco companies was that um there [20:38] wasn't just one fixed payment. Instead, [20:40] it created a system of annual payments [20:42] that continue indefinitely. They will [20:45] actually have to pay forever as long as [20:46] they are in business. They have like [20:48] effectively a special tax paid to them. [20:50] And the amount changes based on [20:52] cigarette sales, inflation, market [20:53] share, and other other adjustments. So [20:56] if cigarette sale sales fall, the total [20:58] payments usually fall. And that's a big [21:01] uh reason why there's like a shift to [21:03] non-cigarette products. [21:05] >> Well, and that just naturally makes [21:07] sense. If less people are buying [21:09] cigarettes, less people are going to [21:10] have health issues down. [21:11] >> Exactly. Exactly. So, [21:12] >> yeah. So, so I, so I brought it up [21:14] because I think that we could be heading [21:15] in that direction with social media. [21:17] >> There's also a fascinating dynamic [21:19] because these states now have an [21:22] indefinite revenue stream that will be [21:24] paid to them and you can model that out [21:26] financially and you can financialize it [21:28] and a lot of people have. And so [21:30] investment firms and banks and and [21:32] financial institutions have come in and [21:34] said okay you are the state of [21:37] of New Mexico for example and you are [21:39] expected to get uh 300 million from big [21:43] tobacco this year and then next year we [21:45] think it'll be 297 and then 298 then and [21:48] then it'll go down and we can model that [21:50] out and we can just give you $4 billion [21:53] right now in exchange for that revenue [21:55] stream or a piece of that revenue stream [21:57] and then those states can take that lump [21:59] sum of cash and build a new bridge or [22:01] something like that. Whatever they need, [22:02] whatever they need to do. So, there's [22:04] been a lot of financialization on top of [22:05] it. Each tobacco company pays a share of [22:07] the total amount. Uh its share depends [22:08] mainly on the share of cigarette sales [22:10] among companies that participate in the [22:12] MSA. The settlement then divides the [22:14] money among states using fixed [22:16] allocation percentages. Uh some states [22:18] later borrowed against these feature [22:19] payments by issuing bonds backed by MSA [22:21] revenue. The MSA also limits tobacco [22:24] advertising and marketing, especially [22:25] marketing that could reach children. In [22:27] simple terms, the agreement created a [22:29] permanent system. Ma major tobacco [22:31] companies receive protection from many [22:33] state lawsuits while states receive [22:34] continuing payments and new enforcement [22:36] powers. The result is not just a legal [22:39] settlement. It created this long-term [22:41] financial and regulatory structure built [22:43] around cigarette sales. And uh it's now [22:46] the 10-year anniversary of starting Lucy [22:48] 2016. August 8th technically was the new [22:52] regulation. [22:52] >> You know what? I'm going to hit [22:54] >> overnight success. [22:55] >> Yeah, for sure. [22:57] >> The exact opposite. Slaving away for [22:59] years. [23:00] >> Obscurity. [23:01] >> But uh very very interesting uh very [23:03] very interesting uh industry to have [23:05] operated in for as long as we have. [23:07] >> Semi analysis chimed in on the deep not [23:11] deep mind news. not just chimed in. [23:14] >> I think the term's posterized [23:16] >> grave digging. [23:17] >> No, not grave dancing. [23:18] >> No, they're [23:19] >> grave digging. They're digging the [23:22] >> putting deep mind in it and then they're [23:23] and then they're dancing on it at the [23:25] end. [23:25] >> Uh anyway, semi analysis says, "For all [23:27] intents and purposes, we believe [23:28] DeepMind is no longer a frontier lab due [23:30] to large numbers of departures from [23:32] their RL teams and poor compute [23:33] allocation. Google will continue [23:36] continue meandering on and releasing [23:38] models, but their odds of ever reaching [23:40] state-of-the-art again have dropped to [23:42] zero. [23:44] >> Damn it. Google is now simply unable to [23:47] retain top AI talent. Again, this is [23:49] what I was saying the day that the news [23:51] broke. It's like researchers want to [23:53] work with great researchers, right? And [23:56] so, uh, the more talent density you [23:59] have, the easier it is to recruit. Yeah. [24:01] Gnome leaving. Yep. earlier was like a [24:05] canary in the coal mine. Obviously, [24:07] these are not the actions of people [24:08] excited about Gemini 4 Pro. [24:11] >> Jeff, Sanjay, Quac, and Oral are just [24:13] the latest in a long string of [24:14] high-profile departures from DeepMind. [24:17] Jeff Dean [24:19] >> Jeff Dean is the undisputed goat of [24:21] Google engineering, co-founded Google [24:22] Brain, and started the GPU program. [24:24] Google, on the other hand, decided it [24:26] was totally worth it to sell enormous [24:27] amounts of compute to Gemini's fiercest [24:29] competitors on long-term contracts [24:31] without any hope of ever returning it to [24:33] DeepMind. More than 20% of total TPU [24:36] shipments from uh third quarter 26 to [24:39] fourth quarter 27 are being sold [24:41] directly to Anthropic. [24:42] >> Wow. [24:43] >> The issue with Google was not Jeff Dean [24:45] nor Nam Shazir, but rather their [24:46] extremely bureaucratic, painfully slow, [24:48] and strategically timid culture. Google [24:51] will join the ranks of other legendary [24:53] tech giants like IBM and Intel to give [24:55] up on the harder thing aggressive and do [24:57] the thing that will make you more money [24:59] becomes demoralizing as many of the [25:01] great technology leads have left. Um and [25:04] take him taking a victory lap. Uh he [25:06] wrote a uh piece uh on July 21st. Google [25:10] is a secular short kind of getting at a [25:13] lot of these issues. So [25:15] >> brutal brutal stuff. This is insane. Uh [25:17] on the other side GCP is working. GCP is [25:20] basically a money printing machine, no [25:21] wonder Google executives are choosing [25:23] GCP over Gemini. So although EBIT [25:26] margins for these system sales are [25:27] slightly lower than core cloud margins [25:30] in the low 30% range, we still expect [25:32] total GCP to deliver mid to high30s EBIT [25:35] margins going forward. How investors [25:37] decide to capitalize the current TPU [25:39] backlog and any large future sales is an [25:42] open question. However, given the strong [25:44] compute demand from the labs, we expect [25:46] more multi- gigawatt deals to be [25:47] announced soon. adding to this backlog. [25:49] In all, uh we estimate that over 250 [25:52] billion additional TPU bookings could be [25:54] added to GCP RPO in the next coming [25:57] quarters from semi- analysis. Uh very [25:59] interesting. I mean there there is like [26:00] a like a positive spin on this which is [26:02] like they seem to be really good at chip [26:04] development, really good at cloud like [26:06] focus where you're where it's working [26:08] and you don't have any tensions there [26:10] because you have excellent teams and [26:12] then you have the ability to underwrite [26:15] that build out with the cash machine [26:17] that is Google search and uh and YouTube [26:21] and their ads products and um and having [26:25] more of this like barbell approach as [26:26] opposed to playing in like the middle [26:28] race is maybe the maybe ultimately the [26:30] right thing to do. There are plenty of [26:32] hyperscalers that have lived that and [26:34] are doing very well on the back of it. [26:36] Like Microsoft, Amazon for example, [26:38] they've been partners to labs at various [26:39] points and uh and are very good at [26:42] building data centers, very good at at [26:43] scaling um and have not uh tried to [26:46] really uh go on a crazy poaching race [26:49] and amass the the the the dream team and [26:53] you know they've been rewarded for it. [26:55] So [26:55] >> yeah, I don't know. [26:56] >> Still wild series of events. They [26:58] obviously had effectively a version of [27:00] chat GBT internally. They didn't ship [27:02] it. Yeah. [27:03] >> Tibo who's now running chat codeex [27:06] >> was there working on that product which [27:08] is really wild. It just seems like [27:10] Google leadership has a very different [27:11] idea of like where value occurs in AI [27:14] than like [27:15] >> Demis or like the other labs, right? [27:16] It's like it's not actually like the [27:18] model itself. Yeah. It's it's the the [27:20] like infrastructure, GPUs, cloud [27:22] business. It's interesting because you [27:24] can run back the old Demos quote about [27:25] like he went when he was selling Deep [27:27] Mind, he talked to Mark Zuckerberg and [27:28] is like, "What are you excited about in [27:29] the future? Are you excited about AI?" [27:31] And Mark Zuckerberg apparently according [27:33] to this exchange that Sebastian Malibi [27:35] reported on. Mark Zuckerberg says, "Oh [27:37] yeah, I'm super excited about AI." And [27:38] then Deus is like, "I'm going to test [27:40] this guy. I'm going to see if he's a [27:41] true believer. Uh what do you think [27:42] about VR?" And Zach's like, "Oh, VR is [27:45] like equivalently as big." And then [27:48] Deven said that he was just equivalently [27:50] as [27:51] >> Yeah. Yeah. excited and he was excited [27:53] about a few other things. And Deus was [27:55] like, I want to be with someone who's [27:57] like all in on AI as a fundamentally [27:59] different technology, not a normal [28:01] technology, not like VR, not like new [28:04] devices, not like electric cars, not [28:06] like, you know, satellites in space. It [28:08] needs to be considered as a completely [28:11] separate development, like a completely [28:13] separate technology. [28:14] >> Yeah. Like I I don't know if if Sundar [28:16] like believes in RSI. It seems like he [28:18] doesn't think that that's like really [28:19] gonna [28:19] >> So the point is that Demis should have [28:21] probed more and said like, "Well, how [28:22] excited are you about cloud? [28:26] How excited are you about AI overviews?" [28:30] And if Sundar, it wasn't Sundar back [28:32] then, but if Google uh was like, "Oh [28:34] yeah, we're we're equivalently excited [28:36] about cloud infrastructure as ASI," then [28:41] that should have been Demis' uh moment [28:43] to be like, "Ah, maybe I should." The [28:45] only thing the only thing is there's a [28:46] world where he's actually so RSI pill [28:48] that he's thinking, okay, it's over for [28:50] us. I need to back up the Brinks truck [28:52] and help help Anthropic, right? And [28:54] putting together this like, you know, [28:56] almost a quarter of a trillion dollar [28:58] >> financing package. Yeah. [29:00] >> Um and variety of of data center [29:03] guarantees to enable [29:05] >> uh Anthropic to scale up even though [29:07] they don't have, you know, really uh [29:10] access to the debt markets in the way [29:12] that Google does. Google's ownership in [29:13] Anthropic is capped at 15%. [29:16] They, I believe, have roughly 14%. [29:19] >> Weird. Why? Wait, how is it capped? [29:20] >> I think Anthropic just didn't [29:22] >> Oh, we don't want you to have [29:23] >> They don't have any voting rights. They [29:25] don't don't even They're not even a [29:26] board observer. They don't have a board [29:28] seat. They basically are just a purely [29:30] financial background. [29:31] >> A branding mogul's Miami Beach home [29:34] lists for $68.5 million. Branding mogul [29:39] Nick Woodhouse was s was sailing around [29:42] Miami for his birthday in 2019 when he [29:44] realized he'd found his new home. [29:46] Turning to his wife uh Joseline [29:48] Woodhouse, the Canadian-B born [29:50] businessman said, "We have to live [29:52] here." It wasn't long after relocating [29:54] from New York in 2020 that the couple [29:57] then living in a condo on Sunny Isles [30:00] Beach, Florida, saw a waterfront lot on [30:03] a guardgated Lorch Island. Is that how [30:06] you pronounce? the Gorche I don't know [30:09] island in Miami Beach from a friend's [30:10] boat. The property had the foundation of [30:12] a house that was just starting to be [30:14] built. The Wood Houses purchased the the [30:16] partially built home for 17 million in [30:19] 2021 and completed construction of the [30:21] roughly 8,800 square ft 7bedroom [30:24] contemporary house around 2023. It's a [30:27] point4 acre estate and they're selling [30:29] it for 68.5 because they're building [30:32] another home nearby. Um, Nick is the [30:34] former president and chief marketing [30:35] officer of Authentic Brands Group. And [30:37] that's why I wanted to talk about this [30:38] because Authentic Brands is a very [30:40] fascinating company. Uh, Tyler, you have [30:42] something here. [30:42] >> Yeah. So, it's pronounced Legors. [30:44] >> Legors. [30:45] >> Thank you. [30:46] >> Legors. Well, Authentic Brands Group is [30:49] a very interesting lifestyle platform, I [30:53] guess. I don't know exactly what you [30:54] would call it, but [30:55] >> they have they have some truly tier one [30:58] assets. They own more than 50 consumer [31:00] brands as well as likenesses and estates [31:03] of celebrities including Muhammad Ali, [31:06] Elvis Presley, and Marilyn Monroe. But [31:08] what stuck out to you? [31:09] >> Let's start at the top. Cream of the [31:11] crop. They own Tapout. They do an iconic [31:14] brand. [31:14] >> Tap out brand. And if you guys ever run [31:17] into John on the weekends, he's almost [31:19] certainly headto toe Tapout. [31:21] >> It was one of their first purchases. [31:23] Silverstar and Tapout. In January 2011, [31:26] they acquired the rights to the likeness [31:28] of Marilyn Monroe. So if you see Marilyn [31:30] Monroe on a t-shirt, Authentic Brands is [31:34] getting a check. [31:34] >> They own Ruka, the surf brand. They own [31:36] Neiman Marcus. They own Prince. They own [31:40] Goodman. They own DC Shoes. [31:42] >> Yeah, [31:43] >> Tyler, I know I know you're still you're [31:45] pretty much headto toe Vulcan at all [31:47] time. Vulcan, maybe some Bibong thrown [31:49] in there. Well, they also own Bella [31:51] Bong. Anyway, Nick is the former [31:53] president and chief marketing officer of [31:54] Authentic Brands Group, a licensing and [31:56] managing company that works with [31:57] companies such as Reebok Champion and [31:59] Brooks Brothers. [32:00] >> Ez says ABG is a graveyard for iconic [32:04] brands. [32:04] >> It's the bending spoons of brands. [32:06] >> No, it really is. I mean, it's somewhat [32:08] sad because a lot of these brands are so [32:10] so iconic and with the right sort of [32:13] management and investment, they they [32:15] would be uh back to their former glory. [32:18] I the a lot of the surf brands and the [32:20] skate brands are a little sentimental [32:22] for me just because I grew up watching [32:24] so much of the content that those brands [32:26] put out and and following the different [32:29] uh athletes uh on their teams. Those [32:32] industries have just been impacted [32:33] surfing most aggressively just because [32:36] kids that don't live by the ocean now, [32:39] they don't really care about surfing. [32:41] >> Yeah. [32:41] >> They care about chrome hearts. [32:43] >> And so, [32:44] >> do you think the surfing industry needs [32:45] a federal backs stop? I would push for [32:47] one certainly. [32:48] >> Yeah. [32:48] >> Yeah. [32:49] >> Couple couple billion dollars from the [32:50] treasury directly to get VCO [32:53] >> quicksilver billong. [32:54] >> Quicks back to the top. [32:57] >> See, we're on to something here. I think [32:58] so. This is our new platform. [33:00] >> I think so. Okay. [33:01] >> We'll be back in the Ultra Dome. We're [33:02] going to have a lot of coverage this [33:04] weekend too around you. You may have [33:06] been seeing some of our Getty images. [33:07] >> Yeah, [33:08] >> you might be seeing some more. [33:09] >> Yeah, we're working on it. [33:10] >> We'll see. [33:11] >> But have a great weekend. [33:12] >> We'll see you Monday. [33:13] >> We'll see you Monday. Leave us five [33:14] stars on Apple Podcast and Spotify. Sign [33:16] up for our newsletter tbn.com. Goodbye.