Apple Overhauls Siri With Private Cloud Infrastructure at Final Cook WWDC

苹果在库克收官WWDC上依托私有云基础设施彻底重构Siri 苹果将Google Cloud和Nvidia GPU整合进其私有基础设施，以驱动下一代Siri接力系统，并将儿童安全功能与空间相机重构定位为iOS 27的核心差异化特性。 主持人John Coogan和Jordi Hays深入解读了苹果最新的WWDC发布内容，探讨了Tim Cook的最后一场主题演讲、Siri AI的全面重构、新增的家长控制功能以及照片的空间重构（Spatial Reframing）编辑功能。他们还分析了科技评论家John Gruber与Mark Gurman围绕iOS 27的Siri架构及私有云计算基础设施产生的公开分歧。本期节目最后快速点评了朝鲜经济转型、Anthropic发布Claude Fable 5、Meta的AI数据中心人才学院（Workforce Academy）、旧金山利用无人机开展警务执法，以及Pat McAfee巨额ESPN合同谈判等热点事件。 Tim Cook 在其最后一届作为 CEO 的 WWDC 上推出了全新重构的 Siri AI，与其 2011 年首度亮相时在 iPhone 4S 上推介 Siri 形成闭环，标志着长达 15 年的技术长跑告一段落。 (0:00) 苹果在发布会中划出近 12 分钟专门介绍儿童安全与家长控制功能（包括“Ask to Browse”），并首发照片视角重构工具 Spatial Reframing，以务实的软硬件防线抵御极简儿童设备竞争。 (6:19) 苹果公布新一代 AFM 3 Cloud Pro 推理模型，通过与谷歌和英伟达合作将 Private Cloud Compute 隐私保护扩展至 Google Cloud 托管的英伟达 GPU 架构上。 (18:44) Anthropic 推出 Claude Fable 5 模型，定价为每百万输入 Token 10 美元、输出 50 美元，其在长周期推理任务中展现出极高稳定性。 (22:23) 知名体育媒体人 Pat McAfee 正在与 ESPN 洽谈年薪超 6000 万美元的超级续约合同，该合作依托独特的节目制作协议与独立人才合约双轨结构。 (28:28)

TBPN reviews Apple's WWDC keynote, framing the revamped Siri and new parental controls around Tim Cook's departure and platform defense, while examining broader shifts in AI model scaling, infrastructure labor, and media rights. Apple's AI Strategy and Third-Party Model Routing (0:00) Tim Cook's final WWDC keynote brought full symmetry to his tenure as CEO, closing a 15-year trajectory that began when Siri was originally introduced on the iPhone 4S in October 2011, just one day before Steve Jobs passed away. While Siri originated from SRI International and was acquired by Apple for roughly $200 million, Apple and its competitors navigated a decade-long lull where voice assistants failed to deliver natural agentic interaction until the emergence of large language models. The hosts estimate that Apple spent 12 to 13 years waiting for foundational models to mature, leaving it roughly three years or 10% behind frontier artificial intelligence laboratories. The mechanics of Apple's modern rollout triggered debate among commentators John Gruber and Mark Gurman regarding iOS 27 integrations. While Gurman reported that Siri would allow routing to third-party chatbots like Google Gemini, OpenAI ChatGPT, and Anthropic Claude, the current user experience operates via relay prompts—such as saying 'Hey Siri, ask ChatGPT'—rather than letting users permanently remap the hardware Siri button. Behind the scenes, Apple introduced AFM 3 Cloud Pro by deploying private cloud compute nodes on Nvidia GPUs inside Google Cloud, extending Apple's cryptographic privacy architecture to external infrastructure. Parental Controls as Device Defense and Synthetic Imagery (6:19) Apple dedicated nearly 12 minutes of its condensed presentation to child safety features, including 'Ask to Browse' approval alerts, age-restricted media limits, and managed contact approvals. John Coogan connects this feature-focused strategy to wider social concerns over screen addiction and falling fertility rates highlighted by commentators like Derek Thompson. Rather than delivering alarmist public messaging, Apple introduced practical software controls to protect its hardware ecosystem against the rise of minimalist dumb phones and child-focused GPS smartwatches. At the same time, Apple introduced Spatial Reframing, a tool in the native Photos app that lets users adjust camera angles and perspective after a photo is taken. The feature sparked online debate over the line between computational photography and synthetic generation. While some view it as an extension of tone mapping and artificial depth-of-field, critics such as Joe Weisenthal noted that generative adjustments risk detaching personal photo archives from physical reality. Frontier AI Scaling and the Physical Infrastructure Buildout (22:23) Anthropic introduced Claude Fable 5 at a cost of $10 per million input tokens and $50 per million output tokens, emphasizing reliability and sustained multi-hour execution. Early user testing reported runs lasting nine consecutive hours without degradation, aligning with researcher Noah Brown's observation that test-time compute scaling shows minimal plateaus under current reasoning budgets. In an aside, the hosts also noted North Korea's state-driven construction boom of 10,000 housing units financed by military sales to Russia and China. The computational buildout has generated shifting labor demands. Meta partnered with CBRE and Associated Builders and Contractors to launch a tuition-free, five-week Workforce Academy providing guaranteed job placement for data center electrical and wiring trades. Concurrently, surveillance infrastructure is shifting municipal operations: in San Francisco, Flock Safety automated license plate readers and connected police drone networks track stolen vehicles from the air, allowing officers to box in suspects without initiating high-speed vehicle pursuits. Pat McAfee's $60 Million ESPN Media Deal (28:28) In sports media business news, Pat McAfee is negotiating a multi-year contract extension with ESPN valued at over $60 million annually, despite having two years remaining on his current deal. The proposed contract incorporates a sliding scale tied to expanded obligations across NFL broadcasting, College GameDay, and primetime watch-along programs. The hosts emphasize the non-traditional structure of McAfee's arrangement, which separates a core talent agreement from an independent production contract. This framework preserves operational autonomy for his three-hour daily broadcast while maintaining a hybrid distribution model: the first two hours air on linear cable via ESPN, while all three hours stream live on YouTube. The broadcast highlights a transitional period across big tech and digital media, marked by Apple balancing proprietary privacy controls with third-party AI infrastructure, frontier labs pursuing long-duration reasoning, and digital-first creators securing unprecedented leverage over traditional cable networks.

TBPN本期节目深入解析了苹果在WWDC展现的产品哲学与AI战略演进，并探讨了前沿大模型发布、基础设施劳动力转型以及数字体育媒体商业化的最新变局。核心观点指出，从Siri历经十五年的技术落地到媒体版权的平台分化，科技与传播生态正在从概念噱头全面转向务实的工程落地与混合分发架构。 Siri十五年历程与苹果AI闭环演进 (0:00) 在Tim Cook最后一次亮相的WWDC主题演讲中，苹果发布了全面升级的Siri AI，为其CEO任期形成历史闭环。节目回顾指出，Cook于2011年10月4日首次以CEO身份主持发布会时推出了集成Siri的iPhone 4S，次日Steve Jobs离世。Siri技术源于斯坦福国际研究院（SRI International）人工智能中心于2008年创立的团队，苹果在2010年斥资约2亿美元将其收购。尽管当时提出了摆脱死板语法的自然交互愿景，但底层算力与模型经历了长达十余年的瓶颈期，谷歌与亚马逊等同行也均未能在早期实现自主智能体目标。主持人认为，苹果在底层技术成熟前等待了约12至13年，在大模型时代一度落后前沿研究机构约3年（或10%），直到近期大语言模型生态成型才真正兑现最初的产品设想。 务实家长控制与AI重塑真实影像之争 (6:19) 在时长缩短的主题演讲中，苹果罕见地花费了近12分钟详述儿童安全与家长控制功能。节目分析，针对智能手机沉迷与社会学界对生育率走低的讨论，苹果并未发表恐慌式的危机演说，而是推出了“请求浏览”（Ask to Browse）域名核准、通信通讯录添加权限及默认年龄保护等具体机制。这种以软件工具为先的策略，有效防御了主打基础通话和GPS定位的儿童手表与“轻智能手机”（dumb phones）对苹果硬件生态的侵蚀。 与此同时，相册新增的“空间重新构图”（Spatial Reframing）功能引发了关于影像本质的伦理探讨。该功能允许用户在拍摄后改变照片的拍摄角度和透视。部分观点视其为计算摄影（如降噪与景深模拟）的自然延伸，但财经记者Joe Weisenthal等人引发的讨论则质疑这跨越到了合成图像生成的范畴。随着AI去噪、物体消除与视角调整技术的演进，平台正面临元数据标注的巨大压力，人们未来的生活记忆可能与实际发生过的物理事实产生脱节。 外部模型中继集成与私有云计算架构 (14:54) 针对科技博主John Gruber与彭博社记者Mark Gurman就iOS 27 Siri集成第三方AI进展的争论，节目指出双方均有事实依据。Gurman展示的界面截图证实了系统存在外部模型选择器，但当前机制主要体现为向第三方聊天机器人中继请求并弹出确认对话框（如“Hey Siri, ask ChatGPT...”），而非深度科技受众所期待的底层物理按键重映射。Marques Brownlee提及的内部代号“Golden Gate”，进一步暗示苹果意在严密维持高收益的“围墙花园”。 在底层算力支持上，苹果展示了AFM 3基础模型体系，其高算力推理模型AFM 3 Cloud Pro通过与谷歌及英伟达合作，将私有云计算（Private Cloud Compute）节点直接部署在Google Cloud托管的英伟达GPU上，从而将端到端加密隐私保障延伸至第三方基础设施。此外，苹果此前曾与OpenAI前首席技术官Mira Murati创立的Thinking Machines Lab接洽过数十亿美元估值的早期收购，但最终选择依托外部算力合作而非高价并购。 前沿模型推理拓展、基建用工与城市安防 (22:23) 在更广阔的技术动态中，Anthropic正式推出了Claude Fable 5，每百万输入Token定价10美元，每百万输出Token定价50美元。该模型在长周期任务中表现突出，社区测试显示其可连续运行9小时无性能衰减。研究人员Noah Brown关于测试期算力（test-time compute）的分析指出，复杂多步骤推理的计算扩展曲线在实际预算内鲜有瓶颈，单次推理甚至可能持续数周。与此同时，朝鲜通过对俄中两国的军工制造与军售实现经济反常增长，并建造了超10,000套住房，成为地缘层面的特殊案例。 AI算力扩张同时重塑了实体就业与公共治理。Meta联合CBRE等机构设立了为期五周的免费“劳动力学院”（Workforce Academy），专项培训数据中心建设与电气布线技工，反映出裁员潮后产业向物理基建技术工种的倾斜。在地方治理上，旧金山警方结合Flock Safety自动车牌识别与无人机高空追踪网络，在数千英尺高空对被盗车辆进行静默锁定并在其停稳后合围，彻底消除了危险的高速追车执法风险。 体育传媒商业化：McAfee协议结构 (28:28) 在体育媒体商业领域，39岁的主持人Pat McAfee正与ESPN就一份年薪超过6,000万美元的多年期续约协议展开实质谈判。该协议不仅涵盖其在《College GameDay》的核心角色，还将深度扩展至NFL报道与黄金时段赛事伴看演播室。 节目着重强调了McAfee商业模式的特殊结构。不同于传统出镜主持人的标准化雇佣合同，ESPN将该合作划分为独立的“制作合同”与“艺人协议”。在其长达3小时的日常节目《The Pat McAfee Show》中，前2小时在ESPN电视网络直播，而全部3小时内容均在YouTube免费开放流媒体播放。这种在传统有线电视与数字流媒体平台之间形成的混合分发机制，不仅赋予了制作团队极高的自主权，也成为传统电视网绑定新一代年轻观众的关键范例。 本期节目梳理了从消费级硬件、底层AI基建到媒体版权商业化的务实演变轨迹：无论是苹果利用安全机制巩固设备生态，还是流媒体借助双轨分发融合传统电视，各领域均逐步告别初期的激进概念。然而，AI图像对人类物理记忆的重构边界，以及重度依赖第三方算力设施下的数据隐私博弈，仍是尚未解决的核心争议。

## Background 本期内容由科技商业谈话频道 TBPN 于 2026 年 6 月 23 日发布在 YouTube 平台。作为主打 30 分钟提炼每日精华的‘Diet TBPN’节目，主持人 John Coogan 和 Jordi Hays 在本期中重点追踪了苹果 WWDC 库克谢幕与 Siri 重构、Anthropic 发布 Claude Fable 5、Meta 设立技工培训学院以及体育名嘴 Pat McAfee 洽谈 ESPN 巨额合约等科技与商业要闻。 ## TL;DR 节目核心聚焦蒂姆·库克在 WWDC 迎来谢幕致意，回顾了 Siri 历经十五年从 iPhone 4S 到如今接入谷歌云与 Gemini 算力的完整演进，并探讨了系统级家长控制与争议性极大的‘空间重新构图’功能。随后，主持团队分析了 Anthropic 新一代模型 Claude Fable 5 的定价与其折射出的‘测试期算力无上限’趋势，以及 Meta 为满足数据中心建设而推出焊接培训学院的现实魔幻。尾声部分则盘点了旧金山警方利用无人机监控让街头犯罪失效的奇闻，并解读了 Pat McAfee 正在商谈的 6000 万美元 ESPN 续约对现代流媒体与电视版权格局的深远影响。 ## Atmosphere & style 整场节目节奏紧凑飞快、信息密度极高，两位主持人展现出极佳的默契与松弛感，将前沿行业洞察与辛辣的硅谷圈内调侃融为一体。讨论在硬核技术剖析与荒诞现实迷因之间自由跳跃——既有对私有云计算隐私架构和 AI 吞噬真实记忆的严肃思考，也不乏‘21 响 AirPods 礼炮’与‘旧金山惯犯上播客哭诉无人机断了财路’等令人捧腹的戏谑，听感极具极客趣味与鲜活的现场感。 ## Bulletin - [01:11](https://www.youtube.com/watch?v=aCMhC_Wn1J0&t=71s) 库克在 WWDC 迎来主题演讲谢幕致意，苹果推出以大模型底座全面重构的 Siri AI，完成了自 2011 年 iPhone 4S 亮相起跨越 15 年的技术闭环。 - [18:57](https://www.youtube.com/watch?v=aCMhC_Wn1J0&t=1137s) 苹果将其私有云计算（Private Cloud Compute）架构扩展至谷歌云托管的英伟达 GPU，其云端模型 AFM 3 Cloud Pro 选择基于谷歌 Gemini 进行训练与微调。 - [23:34](https://www.youtube.com/watch?v=aCMhC_Wn1J0&t=1414s) Anthropic 推出新模型 Claude Fable 5（定价为输入 $10/M tokens、输出 $50/M tokens），研究指出测试期算力在实际预算内几乎看不到性能瓶颈。 - [28:41](https://www.youtube.com/watch?v=aCMhC_Wn1J0&t=1721s) 体育名嘴 Pat McAfee 正在与 ESPN 洽谈年薪或超 6000 万美元的天价续约，建立起打通传统广播电视网与 YouTube 独立分发的新型制作合作架构。 ## Key takeaways - [00:00](https://www.youtube.com/watch?v=aCMhC_Wn1J0&t=0s) 库克在 WWDC 上完成了最后一次官方主题演讲亮相，标志着其执掌苹果十余年时代的落幕，公司底盘与业务依然稳健。 - [01:11](https://www.youtube.com/watch?v=aCMhC_Wn1J0&t=71s) 库克时代的开端与终局均围绕 Siri 展开：从 2011 年 iPhone 4S 发布次日乔布斯离世，到 15 年后借助生成式大模型真正兑现自然语言交互的最初愿景。 - [03:44](https://www.youtube.com/watch?v=aCMhC_Wn1J0&t=224s) 库克任期内精准穿越了 2011 至 2022 年的技术瓶颈期；在底层算法受限的“AI 寒冬”中，苹果将商业运营、服务收入与供应链能力推向了巅峰。 - [03:44](https://www.youtube.com/watch?v=aCMhC_Wn1J0&t=224s) WWDC 音频工程团队在广播流中精准剔除了 4 kHz 与 6 kHz 等唤醒频段，彻底杜绝了全球数千万外放观看 Keynote 的用户设备被频繁误唤醒的尴尬。 - [06:19](https://www.youtube.com/watch?v=aCMhC_Wn1J0&t=379s) 面对屏幕成瘾与青少年身心健康的社会焦虑，苹果拒绝采用恐惧营销，而是通过长达 12 分钟的精细化家长控制与儿童安全功能提供务实工程方案。 - [09:30](https://www.youtube.com/watch?v=aCMhC_Wn1J0&t=570s) 相册上线的“空间重新构图”（Spatial Reframing）结合超分辨率重塑了拍照视角，但也因“无中生有”重构画面而模糊了纪实摄影与数字生成的界限。 - [15:05](https://www.youtube.com/watch?v=aCMhC_Wn1J0&t=905s) John Gruber 与 Mark Gurman 针对第三方模型集成是否兑现展开论战；实机截图证实了模型选择器的存在，但核心极客更渴望彻底重绑 Siri 物理硬件入口。 - [17:43](https://www.youtube.com/watch?v=aCMhC_Wn1J0&t=1063s) 新系统代号“金门”（Golden Gate）被视作语意决定论的绝妙隐喻，既象征苹果精心维护的有门禁围墙花园，也代表其源源不断的掘金能力。 - [18:57](https://www.youtube.com/watch?v=aCMhC_Wn1J0&t=1137s) 苹果私有云端模型 AFM 3 选择接入谷歌 Gemini 与谷歌云英伟达算力集群，反映出顶尖大模型研发的高昂门槛正促使巨头走向实用主义算力合作。 - [21:37](https://www.youtube.com/watch?v=aCMhC_Wn1J0&t=1297s) 苹果地图正式集成高斯泼溅（3D Gaussian Splatting）技术，三维街景的真实感与光影质感获得质的飞跃。 - [23:34](https://www.youtube.com/watch?v=aCMhC_Wn1J0&t=1414s) 研究员 Noah Brown 指出测试期算力（Test-time Compute）在实际预算内未见平台期，极端情况下模型单次推理执行的时间甚至可能长于下一个版本的训练周期。 - [25:47](https://www.youtube.com/watch?v=aCMhC_Wn1J0&t=1547s) Meta 设立为期 5 周的免费劳动力学院并包就业以支援现代数据中心建设，AI 算力基建热潮使“学会焊接”成为取代“学会写代码”的新风向。 - [26:35](https://www.youtube.com/watch?v=aCMhC_Wn1J0&t=1595s) 旧金山警方运用高空静默无人机配合 Flock 车牌识别监控车辆，彻底瓦解了传统的街头飞车追逐与盗车犯罪模式。 - [28:41](https://www.youtube.com/watch?v=aCMhC_Wn1J0&t=1721s) Pat McAfee 洽谈 ESPN 超 6000 万美元年薪新约，其日播节目采取电视网与流媒体分时分发的独立制作架构，确立了体育媒体领域的新商业标杆。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=aCMhC_Wn1J0&t=0s) 开场头条：Anthropic 新模型、中国巨额 AI 基建与库克谢幕 节目开门见山抛出了几则重磅动态：Anthropic 推出了新一代模型 Mythos 与 Fable，其中 Fable 面向主流消费级用户，而 Mythos 则聚焦于网络安全与专业攻防领域；与此同时，大洋彼岸的中国正“按既定日程苏醒”，计划筹建规模高达 2950 亿美元（约合两万多亿人民币）的基础设施基金，全面备战 2027 年的 AI 竞局。 然而当日科技圈最瞩目的头条，莫过于苹果 CEO 蒂姆·库克（Tim Cook）在苹果开发者大会（WWDC）的舞台上迎来自己的“谢幕致意”（last bow）。这一幕直接登上了《华尔街日报》的头版头条。主持人调侃道，库克退休理应享受到极高规格的礼遇——比如将 21 只 AirPods 齐刷刷抛入水中的“21 响 AirPods 礼炮”。面对坊间所谓股价可能暴跌的担忧，主持人认为苹果的业务底盘依然坚不可摧，库克为这家科技巨擘交出了一份惊艳的答卷。 - **Anthropic 阵营划分**：Fable 主打日常消费级交互与通用体验，Mythos 侧重高壁垒的网络安全能力。 - **中国万亿基建**：规模约 2950 亿美元的算力与 AI 基础设施基金正在推进。 - **库克告别舞台**：库克在 WWDC 上完成了最后一次官方主题演讲亮相，标志着一个长达十余年时代的落幕。 --- ### [01:11](https://www.youtube.com/watch?v=aCMhC_Wn1J0&t=71s) 跨越 15 年的闭环：从 iPhone 4S 到 Siri 的真正成熟 库克谢幕发布会的核心焦点落在了脱胎换骨的“Siri AI”上。主持人翻出一组极具宿命感的数据：库克在 2011 年作为 CEO 首次主导的苹果秋季发布会（Let's Talk iPhone），正是 Siri 初次登台亮相的时刻；而 15 年后库克交棒之际，主角依然是 Siri。节目现场重温了 2011 年 iPhone 4S 发布会上的演示片段，当时苹果向全世界勾勒的蓝图是“抛弃刻板语法、无需死记指令，只要像跟真人说话一样问‘今天需要带伞吗’，设备就能心领神会”。 然而，这套科幻般的愿景足足等了 15 年才真正照进现实。Siri 最初起源于斯坦福国际研究院人工智能中心（SRI International Artificial Intelligence Center）孵化的项目 Siri Inc.，苹果于 2010 年将其收购，并在 2011 年 10 月 4 日正式发布。令人唏嘘的是，就在该场发布会结束的次日——2011 年 10 月 5 日，史蒂夫·乔布斯（Steve Jobs）与世长辞。 ```mermaid flowchart TD accTitle: Siri 与库克时代的完整闭环 accDescr: 展示从 2011 年乔布斯离世与 Siri 初登场，到 2026 年 Siri AI 与库克谢幕的历史进程。 subgraph Origin["2011：起点与传承"] A["SRI 实验室分拆技术"] --> B["2011.10.04：iPhone 4S<br/>库克首秀宣布 Siri"] B --> C["2011.10.05：乔布斯逝世<br/>库克全面执掌苹果"] end subgraph Winter["2012-2022：AI 寒冬与商业帝国"] C --> D["硬件与 App Store 利润爆发<br/>市值屡创新高"] D --> E["底层算法陷入瓶颈<br/>Siri 沦为平庸品类中的代表"] end subgraph Finale["2023-2026：大模型与闭环"] E --> F["LLM 与 ChatGPT 爆发<br/>语音助手重塑交互"] F --> G["库克谢幕发布会<br/>推出全面重构的 Siri AI"] end ``` 从产业周期来看，虽然苹果在生成式 AI 竞赛中看似慢了半拍，落后前沿实验室约 3 年时间，但如果从长达 15 年的技术长征来看，这 3 年的差距仅相当于延后了约 10% 的进度。苹果始终坚持按照自己的节奏打磨体验，最终在大模型技术底座成熟的今天，完成了对最初愿景的兑现。 --- ### [03:44](https://www.youtube.com/watch?v=aCMhC_Wn1J0&t=224s) 库克时代的 AI 寒冬与“防唤醒”黑科技 回溯库克执掌苹果的十四年，呈现出一种独特的双重性：在商业运营层面，库克将供应链与服务生态发挥到极致，App Store 创造了天文数字般的现金流，公司市值膨胀数倍；但在前沿人工智能领域，他恰好穿越了历史上漫长的“AI 寒冬”（2011 年至 2022 年前后）。在 Transformer 和大语言模型（LLM）问世之前，Siri 团队受制于整个科技界的技术天花板，即便反复迭代，也只能做到“平庸品类里体验最好的一款”。 在 ChatGPT 带来交互革命之前，即便用户对 Siri 的智商颇有微词，但市面上的竞争对手（如 Google Assistant 或 Amazon Alexa）本质上也都基于传统意图识别体系，并未拉开质的代差，更没有用户因此成规模抛弃 iPhone 转投安卓阵营。直到生成式对话大模型登场，传统助手的局限才被彻底暴露。 主持人还分享了本届 WWDC Keynote 幕后一个极其极客的音频工程细节：在整场流媒体直播中，每当讲台上提到“Siri”这个敏感词时，音频团队在广播流中精准剔除了 4 kHz、6 kHz 等特定唤醒频段的音频信号。这样一来，无论全球多少千万观众通过音响外放观看发布会，他们家里的 iPhone、HomePod 等智能家居设备都不会被反复误唤醒。 --- ### [06:19](https://www.youtube.com/watch?v=aCMhC_Wn1J0&t=379s) WWDC 核心转向：儿童安全管控与“空间重新构图” 在本场时长偏短的发布会中，苹果的议程分配传递出耐人寻味的信号：发布会仅用了大约 15 分钟简述部分常规产品，却破天荒地拿出了整整 12 分钟专门讲解儿童安全防护与家长控制系统。主持人指出，这反映了当前社会舆论的核心焦虑——越来越多的公众与学者意识到智能手机正在对青少年的心智健康与道德发育造成深远冲击，苹果必须在操作系统底层给出强力回应。 - **深度儿童账户体系**：系统原生限制访问成人网站，强制过滤非适龄媒体内容。 - **浏览申请机制（Ask to browse）**：当孩子因查资料或做作业需要访问未授权网站时，点击申请即可向父母手机推送确认通知，家长一键即可批准或拦截。 - **通信防火墙**：家长能够严格管理孩子的通讯录列表，孩子添加任何新联系人均需父母端二次确认。 除安全系统外，照片生态中低调上线了一项名为“空间重新构图”（Spatial Reframing）的 AI 新功能。如果拍照时的物理站位角度偏斜，用户可以在相册中借助 AI 算法将视角平滑调整为正对机位。主持人评价，这项工具没有那种粗暴生成图像的廉价“GenAI 感”，而是深度内嵌于相机底层工作流，体现了苹果利用先进算法解决纯粹摄影痛点的“经典苹果基因”。 不过，这一功能也在摄影圈引发了哲学层面的争议。传统观点认为相机是“捕捉现实”的记录工具，苹果也一直标榜相机的“所见即所得”；但从早期的计算摄影、人造景深（Portrait Mode）、局部色调映射，再到如今直接由 AI 脑补并重构拍摄视角的画面，软硬件的结合正在把“计算机本身变成相机”，彻底模糊了纪实摄影与数字生成的边界。 ### [09:30](https://www.youtube.com/watch?v=aCMhC_Wn1J0&t=570s) 空间重新构图与真实记忆的消解：AI 摄影引发的界限争议 在讨论苹果最新推出的图像处理功能时，主播们探讨了所谓的“空间重新构图”（Spatial Reframing）。虽然有观点认为苹果在 AI 领域表现克制、没有盲目将 AI 塞进所有角落，但这项新功能却显得极为激进。社交媒体对此反应两极分化：一部分人盛赞这是苹果最令人兴奋的 AI 落地之一，另一部分人则对 AI 生成图像的发展方向深感不安。以往苹果在消除背景路人或杂物等基础修图功能上表现平平，远不如 Android 阵营和 Gemini 驱动的 AI 相机那样强大。然而，当技术演进到“弱光拍摄后自动超分提高画质、叠加滤镜、色彩校准、空间重新构图并抹去所有闲杂人等”时，用户得到的已然是一张完全虚构的照片。这种技术处理甚至模糊了现实生活与生成幻象的边界。 这种技术引发的深层焦虑在于人类对生活经历的真实记忆正被悄然改写。节目中引用了彭博社记者 Joe Weisenthal 曾引发巨大争议的前卫观点：既然以后可以用 AI 随叫随到地生成“5岁儿子骑恐龙的照片”，那普通人为什么还要在云端备份海量照片？当时大众纷纷抨击“这根本不是我们想要生活的世界”，但在主播看来，戴着 AirPods 对口袋里的 iPhone 说一句“Hey Siri，帮我生成几张今天在迪士尼乐园玩耍的照片”的时代已经近在咫尺。讨论进一步延伸到 Instagram 等社交平台：未来社区很可能分化为“满身 AI 生成标签的账号”与“坚决杜绝任何 AI 修图的纯原生账号”。究竟何种操作需要被标记为 AI 生成？是简单的黑白滤镜、颜色分级，还是无中生有重构视角的“空间重新构图”？这种划分标准的界限正在变得前所未有的模糊。 - 苹果相机的技术进阶路径：从消除路人杂物，到结合空间重新构图（Spatial Reframing）与画质超分辨率重建。 - Joe Weisenthal 的云端相册激进论调：“以后只要按需生成即可，何必保留原始照片”，折射出公众对 AI 替代真实生活记忆的恐慌。 - 社交媒体生态面临割裂：平台需界定何为“增强滤镜”（Filter），何为“AI 伪造构图”（Spatial Reframe），并推行相应的打标机制。 --- ### [12:45](https://www.youtube.com/watch?v=aCMhC_Wn1J0&t=765s) 手机成瘾、生育率危机与苹果的“解决导向”务实主义 主播分享了自己昨天抛出的一个“天马行空”预测：苹果是否会在发布会上触碰由《大西洋月刊》作家 Derek Thompson 及《金融时报》近期热议的“智能手机成瘾、大脑退化（Brain Rot）与现代生育率下降”这一宏大社会议题。虽然手机普及与全球生育率下跌之间无法建立绝对的因果链条，但不少严肃学者已经将其列为占比可能高达 30% 的重要关联因素。面对这种社会焦虑，苹果展现出了一种极为独特的企业沟通哲学：它从不奉行“恐惧营销”（Fear-based marketing），其高管也从不上播客贩卖末日焦虑。正如蒂姆·库克（Tim Cook）多年来从不在公开场合大谈“海平面上升我们都会死”的气候浩劫，而是默默带领团队开发清洁能源数据中心、用实际成果说话一样，苹果面对手机成瘾问题同样直奔解决方案。 在整场节奏极其紧凑的 WWDC 主题演讲中，苹果罕见地花费了整整 12 分钟专门展示家长控制功能（Parental Controls）。随着越来越多父母对给年幼孩子配备全功能智能手机感到犹豫，市场上甚至催生了仅具备 GPS 定位和基础通话功能、没有屏幕与社交软件的“儿童哑巴手机（Dumb Phone）”细分品类。苹果没有发表长篇大论去痛陈“无休止的信息流滚动如何剥夺人类的幸福感”，而是直接通过健全、精细的管控工具告诉家长：“把 iPhone 交给孩子是安全的，你拥有绝对的控制权。”这种只在幕后默默研发、发布会仅聚焦于具体产品方案的做法，不仅是精明的商业防御，也展示了苹果产品哲学中令人耳目一新的务实一面。 ```mermaid flowchart TD accTitle: 苹果面对社会议题的务实解决模式 accDescr: 展现苹果不同于恐惧营销的解决方案推进路径，以环保和手机成瘾为例 subgraph Problem["外部社会争议"] P1["气候危机与碳排放"] P2["屏幕成瘾与儿童身心健康"] end subgraph AppleStyle["苹果处理模式"] style AppleStyle fill:#f9f9f9,stroke:#333,stroke-width:1px A1["拒绝末日论述<br/>避免恐惧营销"] A2["私下研发落地工程"] A3["直接发布解决工具"] end subgraph Solution["WWDC 落地形态"] S1["绿色供应链与清洁能源"] S2["12分钟家长控制功能详解<br/>狙击儿童简易手机市场"] end Problem --> AppleStyle AppleStyle --> Solution ``` --- ### [15:05](https://www.youtube.com/watch?v=aCMhC_Wn1J0&t=905s) Gruber 与 Gurman 的推特交锋：Siri 的第三方模型集成到底兑现了吗？ 在插播完赞助商 Cisco（提供 AI 时代关键网络基础设施）的信息后，节目切入了一场在科技推特圈引发巨大轰动的名嘴论战：科技博主 John Gruber（Daring Fireball）与彭博社知名记者 Mark Gurman 之间的公开互怼。事件的起因要追溯到 2026 年 3 月 Gurman 的一篇重磅独家报道，当时他透露苹果计划在 iOS 27 更新中向第三方 AI 助手开放 Siri 核心生态，允许用户在安装 Gemini、Claude 或 ChatGPT 等应用后，直接在 Siri 内部将查询请求分发给这些竞争对手的模型。然而在昨天的 WWDC 首日演讲结束后，苹果并未如预期那样隆重宣布这一开放动作。Gruber 随即在博客上发文开火，讥讽道：“也许苹果今天时间不够打算明天再宣布；也许他们纯粹忘了；又或者他们把两个月前存在的下一代 Siri 全盘推翻、在最后一个月里重新拼凑了一个？我打赌事情八成是这样。” 面对 Gruber 的冷嘲热讽，Gurman 迅速晒出系统设置界面的实机截图进行反击，证明 Siri 应用内确实已经内嵌了“模型选择器（Model Picker）”，可以自由配置调用外部 AI 应用，这与他此前的爆料完全吻合。主播分析认为，两人的分歧表面上是对爆料细节和发布会口径的争论，核心实则是“大众预期与实际落地之间的落差”。硬核 AI 用户原本寄希望于苹果能够彻底放权——允许用户将 iPhone 上的 Siri 物理按键直接重定向映射为 ChatGPT 或 Claude 的呼出键。当直播间观众打趣主播语速快得像开了 2 倍速时，主播调侃道“0.5 倍速播放按钮就在那儿”，随后继续指出，正是这种将默认入口彻底拱手让人的激进诉求，才让任何折中的系统集成显得不够彻底。 - **核心争议时间线**：Mark Gurman 于 2026 年 3 月 26 日率先爆料苹果将支持 Siri 调用 App Store 下载的第三方 AI 聊天机器人（涵盖 Gemini、Claude 等）。 - **Gruber 的质疑**：WWDC 主题演讲未给该功能提供重点篇幅，Gruber 据此质疑爆料的准确性，调侃苹果临时推翻了原定方案。 - **Gurman 的回应凭证**：公开 iOS 内部 Siri 设置界面的截图，证实“模型选择器”已真实存在。 - **用户核心落差**：极客群体真正渴望的是系统硬件层面的完全重绑（Hardware Button Remapping），即长按电源键能直接唤起全功能第三方前沿模型。 ### [17:43](https://www.youtube.com/watch?v=aCMhC_Wn1J0&t=1063s) Siri 模型选择与苹果的“金门”围墙花园 主持人们讨论起 Siri 与第三方大模型的整合体验。在目前的系统机制下，用户若想调用外部模型，整个流程极其繁琐，必须对手机说“Hey Siri，帮我去问问 ChatGPT 这个”，随后屏幕还会弹出一个确认弹窗，必须手动点击确认后才能继续，两人吐槽这种体验简直就是在“玩传话游戏”（playing telephone）。虽然最新的系统截图显示 Siri 应用中出现了一个模型选择下拉菜单，但目前尚不清楚该设置是会在每次查询后重置，还是可以永久固定。 科技博主 Marques Brownlee（MKBHD）针对苹果将新操作系统代号命名为“金门”（Golden Gate）发表了独到见解。他指出这体现了一种奇妙的“语意决定论”（nominative determinism）——这扇“金门”象征着苹果在全力维护其带有门禁的围墙花园（walled garden），并从门禁内不断开采出源源不断的黄金。主持人们调侃道，这可能只是因为苹果团队单纯喜欢旧金山，但无论如何，这个双关隐喻非常精准。同时，Brian Mac Duff 分享了新版系统中“消除”（Clean Up）功能的显著飞跃，推测底层可能微调了轻量级图像模型，彻底扭转了早期版本的粗糙表现。 ### [18:57](https://www.youtube.com/watch?v=aCMhC_Wn1J0&t=1137s) 苹果私有云扩展至谷歌云：AFM 3 与 Siri 进化史 节目深入解析了苹果私有云计算（Private Cloud Compute）的最新架构动态。苹果在第三代基础模型 AFM 3（Apple Foundation Models 3）的云端推理模型 AFM 3 Cloud Pro 上，选择基于谷歌 Gemini 进行微调或训练。为了兼顾算力与隐私，苹果与谷歌及英伟达展开合作，将投入巨额资源打造的私有云计算体系直接延伸至谷歌云环境中的英伟达 GPU 上，并宣称在第三方基础设施中保持相同的隐私防护等级。主持人调侃道，用户把所有身家隐私都托付给信任度极高的 iPhone，而苹果转手把数据算力接驳进了“那个最爱收集数据的广告巨头”谷歌的服务器中。 随后，主持人翻出了自己在 2023 年 1 月 8 日写下的预言贴：“伴随对话式 AI 聊天机器人的惊人进展，我敢在此立下一个明确预言：到 2043 年，Siri 终于能用了。”他笑称现在的 Siri 显然提前实现了这一目标。主持人还回顾了自己当年的折腾史——在官方支持集成之前，他曾通过 iOS 快捷指令（Shortcuts）绑定 OpenAI API 密钥，设置了“Siri Pro 模式”来将问题强制路由给 ChatGPT。回顾历史，Siri 本身就是苹果耗资约 2 亿美元收购而来的产物；而就在去年，苹果还曾与前 OpenAI CTO Mira Murati 会面，接洽收购其创办的新 AI 团队 Thinking Machines Lab 的可能性。然而两方谈判并未深入，苹果最终选择了与 Gemini 合作，毕竟如今要拿下拥有顶级工程人才的初创团队，代价可能高达上百亿美元。 ```mermaid flowchart TD accTitle: 苹果私有云计算与基础模型协同架构 accDescr: 展示苹果设备端请求如何通过私有云计算路由至本地芯片或谷歌云托管的英伟达算力中 subgraph User["用户终端"] iPhone["iPhone / Siri 交互"] end subgraph ApplePCC["Apple 私有云计算 (Private Cloud Compute)"] SecRouter["隐私安全路由与验证网关"] end subgraph CloudInfra["扩展算力集群"] direction TB GCP["Google Cloud 基础设施"] NV_GPU["Nvidia GPU 集群"] AFM["AFM 3 Cloud Pro\n(基于 Gemini 训练/微调)"] GCP --- NV_GPU --- AFM end iPhone -->|"端侧无法消纳的高阶推理请求"| SecRouter SecRouter -->|"端到端加密与同等隐私保障"| CloudInfra ``` ### [21:37](https://www.youtube.com/watch?v=aCMhC_Wn1J0&t=1297s) 高斯泼溅登入苹果地图与朝鲜经济的奇特反差 苹果地图迎来了一项亮眼升级——引入高斯泼溅（3D Gaussian Splatting）技术，使得三维城市街景的真实感和光影调色有了飞跃式的提升。主持人们讨论了这项技术的实际用途，认为对于刚入驻陌生城市酒店的旅行者来说，通过这种细腻的三维街景提前熟悉街区方位会格外实用。主持人随即开起了硅谷式的黑色玩笑，调侃用户未来可以用它在迪士尼乐园定位特定区域，合成出“AI 生成的假孩子在景区的虚假照片”，再顺理成章地发到虚拟社交网络上，实现全程数字化的虚荣闭环。 话题随后意外切入地缘政治与宏观经济报道。《华尔街日报》报道称朝鲜经济正成为“世界上最不可能的成功案例”。主持人打趣说，硅谷很多人常吹嘘自己从 PIF 融了钱（通常指沙特公共投资基金 Public Investment Fund），但在当下，搞不好有人拿的是“平壤投资基​​金”（Pyongyang Investment Fund）的黑钱。根据报道，这个受到全球最严厉制裁的封闭经济体不仅通过向俄中出售军火以及扩张核武器实现外汇创收，而且还在平壤一口气建设了 10,000 套住宅，房屋营建速度甚至直接碾压了洛杉矶。 ### [23:34](https://www.youtube.com/watch?v=aCMhC_Wn1J0&t=1414s) Claude Fable 5 登场与测试期算力无止境扩展 Anthropic 正式推出全新模型 Claude Fable 5，社媒平台上引发热烈反馈。针对《The Information》称其为“Mythos 的阉割版”的评价，主持人分析这并非贬义，而是意味着该版本更安全、可靠且调用成本更低。模型在基准测试中表现亮眼，定为每百万输入 Token 收费 10 美元、每百万输出 Token 收费 50 美元，资深测试者甚至反馈该模型能够连续执行 9 个小时或整个周末的复杂任务而不发生语境迷失。 这一现象引出了研究员 Noah Brown 关于大规模测试期算力（Test-time Compute）深刻影响的思考。他在文章中指出，人们常问为何不通过评测工具将测试期算力推到性能瓶颈（plateau），“问题在于从经验上看，那个所谓的性能平台期距离极度遥远；在实际预算范围内，我们甚至根本观察不到瓶颈的存在。”正如 Andrej Karpathy 在自动化研究实验中所观察到的，即使连续运行数百次实验，模型的性能仍在持续上升。Noah Brown 甚至提出一个令人震撼的推论：未来一个模型单次推理执行任务所消耗的时间，可能比训练下一个迭代版本模型所需的时间还要漫长。这种任务未跑完而下一代模型已发布的算力膨胀图景，彻底打破了人们的传统认知。 ### [25:47](https://www.youtube.com/watch?v=aCMhC_Wn1J0&t=1547s) Meta 创办施工学院：“学会焊接”取代“学会写代码” 在大举裁员 8,000 人之后，Meta 宣布启动一项“劳动力学院”（Workforce Academy）项目，为期五周且完全免费，为考核合格者提供全职就业保障。该项目与 CBRE 以及相关建筑行业协会合作，旨在专门培训美国工人掌握建设现代数据中心所需的技能。 这一举动被主持人调侃为网络经典迷因“学会焊接”（Learn to weld）在现实中的终极映射。过去科技行业倡导“学会写代码”，如今在 AI 数据中心狂潮下，熟练的技术工人变成了最抢手的硬通货，Meta 相当于在呼吁大众“放下键盘，拿起扳手”。主持人甚至戏谑制片人 Tyler 已经戴着假发、化名被该项目录取，马上就要进厂实操焊接。谈笑间，节目的视线随后转向了智能安防企业 Flock Safety 的最新争议。 - **Claude Fable 5 核心参数与资费**：输入端价格为 $10 / 100 万 tokens，输出端价格为 $50 / 100 万 tokens，具备极强的长程连贯推理与容错能力。 - **测试期算力核心结论**：测试期性能提升在现实预算内几乎看不到瓶颈，模型推理用时未来可能反超新模型的训练周期。 - **Meta 技能学院细节**：为期 5 周，免收学费且保证就业，重点培养数据中心管道、焊接与建造相关的熟练技工。 ### [26:35](https://www.youtube.com/watch?v=aCMhC_Wn1J0&t=1595s) 旧金山罪犯的烦恼：Flock 监控与警用无人机让犯罪彻底失效 节目现场原本准备播放一段网络上热传的播客视频，但刚按下播放键，里面连珠炮般的粗口便扑面而来，嘉宾直呼“那事儿彻底结束了，兄弟，我拿我妈发誓，他们搞来了无人机……”。主持人不得不紧急掐断音频，笑称这段内容“TBPN unsafe”（不符合节目播出尺度），必须深度消音才能播，随后两人直接为观众转述了视频中令人啼笑皆非的核心内容：几位有犯罪前科的旧金山街头人员在播客中大倒苦水，抱怨在旧金山“已经根本没法搞犯罪了”，因为警方的监控技术（如 Flock 车牌识别系统与无人机）已经强大到令人窒息。 根据视频中罪犯的现身说法，如今在旧金山一旦偷了车，警用无人机就会立即在数千英尺的高空对车辆展开无声尾随。嫌疑人根本无法察觉自己已被盯上，警方也不再冒着危险进行传统的高速警车追逐，而是从容等待嫌疑人把车停好或熄火驻留的瞬间，地面警力直接一拥而上将其死死围堵在车内。播客主持人甚至一本正经地追问：“所以你现在连‘偷辆车、突袭你的仇家（ops），然后弃车跑路’这种事都干不了了？”对方无奈地承认确实不行。TBPN 主持人吐槽道，播客主持人的发问方式简直把“盗车火拼再弃车”当成了一种人尽皆知的标准化日常操作，更讽刺的是，视频左侧大倒苦水埋怨“如今连车窗射击（drive-bys）都没法做”的嘉宾，据称还是旧金山最恶名昭彰的惯犯之一。主持人打趣说，与其讨论科技监控的弊端，警方不如直接多刷刷推特、听听这些罪犯自己录播客供认手法的节目，直接顺藤摸瓜去抓人。 ```mermaid flowchart TD accTitle: 旧金山无人机防犯罪追踪流程 accDescr: 展示旧金山警方如何利用高空无人机悄无声息追踪嫌疑车辆并最终实施围堵。 subgraph AirTracking ["空中隐形追踪"] A["嫌疑人盗车 / 违法行动"] --> B["高空警用无人机即时锁定"] B --> C["数千英尺高空静默尾随<br/>（嫌疑人完全不自知）"] end subgraph GroundInterdiction ["地面精准围堵"] C --> D["摒弃高危街头飞车追逐<br/>静待目标泊车或静止"] D --> E["地面警车瞬间合围封锁"] E --> F["嫌犯无路可逃，当场落网"] end ``` - 犯罪手法失效：旧金山高空无人机搭配车牌识别追踪，完全剥夺了嫌犯弃车或逃脱警车追击的机会。 - 荒诞现实：惯犯在播客节目上公开交流犯罪经验并抱怨警方高科技，成为社交媒体上的热议奇观。 --- ### [28:41](https://www.youtube.com/watch?v=aCMhC_Wn1J0&t=1721s) 体育媒体巨头新标杆：Pat McAfee 洽谈 ESPN 年薪 6000 万美元天价续约 在简短提及用于写代码、分析数据与自动化业务流的 AI Agent 平台赞助商后，节目切入了体育传媒界的重磅消息。据《The Athletic》报道，体育名嘴 Pat McAfee 目前正与 ESPN 及其母公司展开深度续约谈判，新合同的年薪可能超过惊人的 6000 万美元。作为 TBPN 主持团队极为推崇的行业标杆，McAfee 若能达成该协议，其收入与行业地位都将迈入全新维度。 尽管该交易尚未最终签署，且具体薪酬可能根据他承担的新职责设立浮动机制，但 McAfee 在 ESPN 版图中的影响力正在急速膨胀。除了此前在大学橄榄球王牌节目《College GameDay》中担任核心座席，以及前一晚在电视直播中主持比赛连线陪看（watch-along）外，他未来还将在 NFL 相关报道中扮演更关键的角色。39 岁的 McAfee 现有合同还剩两年，ESPN 将该合作定性为“制作合同加独立的出镜艺人协议”，使其在体制上完全不同于传统的电视台演播室主持人。目前他的王牌日播节目《The Pat McAfee Show》每天长达 3 小时，前 2 小时在 ESPN 线性电视和 YouTube 同步播出，最后 1 小时则专属于 YouTube，展现了传统广播电视网与新媒体流媒体极为独特的利益平衡与分发架构。 在节目尾声，主持人提醒听众在 Apple Podcasts 和 Spotify 上为节目留下五星好评，并前往官网 TBPN.com 订阅每日科技商业通讯，随后结束了全天节目的录制。 - 合同规模：《The Athletic》披露 Pat McAfee 与 ESPN 正在商谈年薪逾 6000 万美元的超级续约。 - 个人与合同现状：现年 39 岁，原合同尚余 2 年，新约可能涵盖更广的 NFL 报道权限与分级薪酬结构。 - 节目形态与分发模式：每天 3 小时直播，ESPN 播出前 2 小时，YouTube 全程直播 3 小时，合同由独立制作协议与艺人出镜协议共同构成。 ## Notable quotes > [00:00](https://www.youtube.com/watch?v=aCMhC_Wn1J0&t=0s) 21 响 AirPods 礼炮 > [01:11](https://www.youtube.com/watch?v=aCMhC_Wn1J0&t=71s) 抛弃刻板语法、无需死记指令，只要像跟真人说话一样问‘今天需要带伞吗’，设备就能心领神会 > [03:44](https://www.youtube.com/watch?v=aCMhC_Wn1J0&t=224s) 平庸品类里体验最好的一款 > [06:19](https://www.youtube.com/watch?v=aCMhC_Wn1J0&t=379s) 计算机本身变成相机 > [09:30](https://www.youtube.com/watch?v=aCMhC_Wn1J0&t=570s) 既然以后可以用 AI 随叫随到地生成‘5岁儿子骑恐龙的照片’，那普通人为什么还要在云端备份海量照片？ — Joe Weisenthal > [09:30](https://www.youtube.com/watch?v=aCMhC_Wn1J0&t=570s) Hey Siri，帮我生成几张今天在迪士尼乐园玩耍的照片 > [12:45](https://www.youtube.com/watch?v=aCMhC_Wn1J0&t=765s) 把 iPhone 交给孩子是安全的，你拥有绝对的控制权。 > [15:05](https://www.youtube.com/watch?v=aCMhC_Wn1J0&t=905s) 也许苹果今天时间不够打算明天再宣布；也许他们纯粹忘了；又或者他们把两个月前存在的下一代 Siri 全盘推翻、在最后一个月里重新拼凑了一个？我打赌事情八成是这样。 — John Gruber > [17:43](https://www.youtube.com/watch?v=aCMhC_Wn1J0&t=1063s) 语意决定论 — Marques Brownlee > [18:57](https://www.youtube.com/watch?v=aCMhC_Wn1J0&t=1137s) 伴随对话式 AI 聊天机器人的惊人进展，我敢在此立下一个明确预言：到 2043 年，Siri 终于能用了。 > [23:34](https://www.youtube.com/watch?v=aCMhC_Wn1J0&t=1414s) 问题在于从经验上看，那个所谓的性能平台期距离极度遥远；在实际预算范围内，我们甚至根本观察不到瓶颈的存在。 — Noah Brown > [26:35](https://www.youtube.com/watch?v=aCMhC_Wn1J0&t=1595s) 那事儿彻底结束了，兄弟，我拿我妈发誓，他们搞来了无人机…… > [26:35](https://www.youtube.com/watch?v=aCMhC_Wn1J0&t=1595s) 偷辆车、突袭你的仇家（ops），然后弃车跑路 ## Mentioned Anthropic, Mythos, Fable, Claude Fable 5, Tim Cook, WWDC, 《华尔街日报》, Siri, iPhone 4S, SRI International Artificial Intelligence Center, Steve Jobs, ChatGPT, Google Assistant, Amazon Alexa, HomePod, Spatial Reframing, Gemini, Joe Weisenthal, Derek Thompson, John Gruber, Daring Fireball, Mark Gurman, Cisco, Marques Brownlee (MKBHD), Brian Mac Duff, Private Cloud Compute, AFM 3, AFM 3 Cloud Pro, Google Cloud, Nvidia, OpenAI, Mira Murati, Thinking Machines Lab, 3D Gaussian Splatting, PIF, Noah Brown, Andrej Karpathy, Meta, CBRE, Flock Safety, Pat McAfee, ESPN, 《The Athletic》, The Pat McAfee Show ## 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=aCMhC_Wn1J0)

[0:00] 21 响 AirPods 礼炮 [1:11] 抛弃刻板语法、无需死记指令，只要像跟真人说话一样问‘今天需要带伞吗’，设备就能心领神会 [3:44] 平庸品类里体验最好的一款 [6:19] 计算机本身变成相机 [9:30] Joe Weisenthal: 既然以后可以用 AI 随叫随到地生成‘5岁儿子骑恐龙的照片’，那普通人为什么还要在云端备份海量照片？ [9:30] Hey Siri，帮我生成几张今天在迪士尼乐园玩耍的照片 [12:45] 把 iPhone 交给孩子是安全的，你拥有绝对的控制权。 [15:05] John Gruber: 也许苹果今天时间不够打算明天再宣布；也许他们纯粹忘了；又或者他们把两个月前存在的下一代 Siri 全盘推翻、在最后一个月里重新拼凑了一个？我打赌事情八成是这样。 [17:43] Marques Brownlee: 语意决定论 [18:57] 伴随对话式 AI 聊天机器人的惊人进展，我敢在此立下一个明确预言：到 2043 年，Siri 终于能用了。 [23:34] Noah Brown: 问题在于从经验上看，那个所谓的性能平台期距离极度遥远；在实际预算范围内，我们甚至根本观察不到瓶颈的存在。 [26:35] 那事儿彻底结束了，兄弟，我拿我妈发誓，他们搞来了无人机…… [26:35] 偷辆车、突袭你的仇家（ops），然后弃车跑路

[8:33] We're going to enter an era where people's memories of their lived experience are different than what actually happened.

[0:00] Anthropic launched Mythos or Fable. Uh [0:03] Fable's the main consumer model, but [0:05] Mythos, there's some details about that [0:07] as well. More for cybersecurity. [0:10] >> China's waking up. [0:11] >> China is waking up right on schedule. [0:13] The AI 2027. [0:14] >> $295 billion infrastructure fund of [0:17] sorts. [0:18] >> Tim Cook took his last bow at Apple's [0:21] showcase event. This is history on the [0:24] front page of the Wall Street Journal. [0:26] Uh yeah, he needs a 21-gun salute. A [0:28] 21-i airpod salute where you throw them [0:30] into the water. [0:31] >> 21-point nuke. [0:33] >> 21-point [laughter] [0:34] nuke? I don't think so. I think the [0:35] stock's going to do you great. Uh so, [0:38] this is this is actually fascinating. [0:39] So, the focus of Tim Cook's last WWDC [0:43] was Siri. And reviews are good. They [0:46] launched a bunch of new features. I [0:48] think people are going to be very [0:49] satisfied with it and all the decisions [0:51] that they made. Did you know Tim Cook's [0:53] first WWDC was when they announced Siri [0:56] the first time, 2011. [0:57] >> Woah. [0:57] >> So, we can play this clip from [0:59] >> Full circle. [1:00] >> first Apple Siri announcement. Listen to [1:03] what they actually say, how they talk [1:05] about Siri, and then look at where we [1:07] are today just 16 years later, 15 years [1:10] later. [1:11] >> But you're going to be able to talk to [1:12] technology and it'll do things for us. [1:15] Haven't we seen this before over and [1:16] over? [1:17] But it never comes true. [1:19] We have very limited capability. [1:22] >> It's crazy. [1:22] >> Just learn a syntax, call a name, dial a [1:26] number, play a song. [1:28] It is such a letdown. What we really [1:30] want to do is just talk to our device. [1:33] Ask a simple question. [1:34] What's the weather going to be like [1:36] today? And get a response. In fact, we [1:38] don't want to be told how to talk to it. [1:40] We want to talk to it any way we'd like. [1:41] Someone else might ask, "What rain in [1:43] Cupertino? [1:44] Or is the weather going to get worse [1:45] today? Or do I need an umbrella today?" [1:48] And your device, in this case your [1:49] phone, will figure out what you mean and [1:52] help you get what you want done. That's [1:54] a feature of the iPhone 4S we call Siri. [1:57] Siri [1:58] is your intelligent assistant and helps [2:00] you get things done [2:02] just by asking. [2:04] >> Pretty crazy. [2:05] >> Called it. [2:05] >> What happened? [2:06] >> But it just took 15 years for the [2:07] technology to actually catch up. And so [2:10] I think they they launched a you know a [2:12] functionality when Siri launched it was [2:14] the best voice assistant. It was great [2:16] experience. You did have to know the [2:18] syntax but it was cool. [2:20] >> years for their technology to catch up. [2:22] >> Yes but it took 13 years or 12 years for [2:26] the technology industry to catch up and [2:28] so they're only 10% behind if you look [2:31] at it that way. Like it's been a 15-year [2:33] project from that announcement. It's [2:35] been 12 years for the leading labs to [2:38] get there and so they're 3 years behind [2:40] feels like an eternity in AI world but [2:43] it's really only 10% 10% slower and they [2:46] wanted to do it their way and and they [2:48] got there. It is very interesting to me [2:50] that they're calling it Siri AI because [2:53] Siri the whole pitch was AI and [2:56] >> [laughter] [2:56] >> like Siri the name Siri comes from SRI [2:59] International which was [3:01] the Stanford Research Institute and it's [3:04] from SRI International Artificial [3:06] Intelligence Center. Like these were [3:08] Stanford AI researchers who started this [3:11] company in like 2008 or something like [3:13] that and grew it and then eventually [3:15] sold it to Apple [3:18] as Siri Inc. Acquired October initially [3:21] released October 4th 2011 14 years ago. [3:24] But do you know what happened the day [3:26] after October 4th 2011 the Let's Talk [3:30] iPhone WWDC where they announced Siri. [3:33] Tim Cook was the first this was his [3:35] first WWDC as CEO. He obviously hands it [3:38] off to his colleague to introduce Siri. [3:41] Steve Jobs passed away the next day. [3:43] Pretty crazy. And so you have this [3:44] bookending of Siri on both sides of Tim [3:47] Cook's career where sort of nothing [3:50] happened in an interesting way, you [3:52] know? But the stock did fantastically [3:55] well. His Tim Cook created an immense [3:57] amount of value. But it was this like AI [4:00] winter. [4:01] >> faster, stronger. [4:03] >> Look at the stock price. Look at the [4:04] stock price. Like [4:06] >> I'm I'm not I'm not saying that. I'm not [4:07] saying that in a in a negative way. I'm [4:09] just I'm just [4:10] >> So you have like Tim Cook [4:11] operationalizing the company, making so [4:14] much money from the App Store, building [4:16] the business. But at the same time, he [4:18] just happened to be the CEO during [4:22] the greatest AI winter ever, basically. [4:24] From 2011 to 20 to 2022, 2023, when we [4:28] started getting LLMs, started getting [4:30] chatbots. And it did take them a few [4:32] years because they'd lived through a [4:34] decade of like, "Okay, [4:36] so the Siri team, have you guys got it [4:38] better?" And they're like, "Not yet. The [4:40] technology is not moving fast enough. [4:42] Like we haven't invented anything new. [4:44] So we are the best of a very mediocre [4:47] category." And Siri sort of slowly fell [4:50] behind. But even when people say, even [4:53] even if you go back to 2019, [4:55] pre-ChatGPT, pre-AI, pre-LLMs, and you [4:59] ask the question of like, "Is Siri [5:01] good?" People would say, like, "Ah, no. [5:02] It's not really delivering the promise [5:04] of like talk to your phone, it does [5:06] whatever." Like people have the idea of [5:08] the the C-3PO, the agentic computer. [5:10] >> Yeah. [5:11] >> It's not delivering there. It's not like [5:13] people were switching in droves to [5:14] Android because Google Assistant was so [5:16] much better. Or Alexa was so much [5:18] better. Alexa was the same system. I'm [5:20] sorry, I'm triggering a whole bunch of [5:21] smart home devices. But [5:22] >> every device. [5:24] >> interesting. So during the keynote for [5:27] uh at WWC, they whenever the keyword S I [5:31] R I was mentioned, they took out the 4 [5:33] kHz, 6 kHz, a few different spaces, so [5:36] it wouldn't trigger the at-home devices. [5:38] Anyway, so they were like sort of [5:41] behind, but it didn't really matter [5:42] because they wouldn't they weren't [5:43] getting their lunch eaten until the dawn [5:46] of the LLM, the dawn of the chat app, [5:48] chat GPT, and other apps that came out [5:51] and became like super-powered relative [5:53] to Siri's capabilities. Today, they're [5:56] catching up, and it's a very interesting [5:58] bookend on this on the the Tim Cook era [6:01] at Apple, where he was incredibly [6:03] effective, but the technology was just [6:05] was not at an inflection point at any [6:07] time during his career, basically, until [6:09] the very, very end. And yeah, they got, [6:12] you know, they were a little bit behind [6:13] the ball, but uh ultimately, they did [6:15] catch up. There are a bunch of [6:16] interesting other things in here. Um [6:19] the facts, first we should go through. [6:21] So, Apple spent roughly 12 minutes [6:25] yesterday detailing the expansion of its [6:27] child safety and parental control [6:29] features. This was interesting because [6:31] it was a notably short keynote. It [6:33] wasn't a particularly long keynote. I [6:35] think they spent like 15 minutes talking [6:36] about products, and then 12 minutes on [6:38] parental features. Sort of gives you an [6:40] idea of where the energy, where the [6:43] where the focus is, where they want to [6:45] spend [6:45] >> Yeah, potentially what you were saying [6:46] yesterday. I think people, more and more [6:49] people are waking up to maybe phones are [6:51] causing systemic issues in society. [6:55] >> And I think [6:55] >> brain morality [6:57] stuff. And so, that's what I wrote about [6:59] today in the TBP newsletter. Quick on [7:01] the facts, child accounts with built-in [7:02] age protections like limiting adult [7:04] websites and showing age-appropriate [7:05] media, good. Ask to browse, which asks [7:08] the for the parent's approval before the [7:10] child can visit a new website. So, any [7:12] So, the child can say, "Hey, I got to go [7:14] to this website for my homework." Click [7:16] a button, and it sends a notification to [7:18] the parent's phone, and they approve, [7:19] and they say, "Yeah, that's fine." Um [7:21] so, you can really lock it down. [7:22] Communication controls allow parents to [7:24] manage who their child talks to, and [7:27] will ask for confirmation before their [7:29] child can add a new contact. Uh Apple [7:31] also unveiled a new photo editing tool. [7:33] This is separate, but spatial reframing. [7:36] Uh so, if you don't like the angle of [7:37] the photo you were taking, you took it [7:38] over this way, you can with AI reframe [7:41] it so that it's more straight on. This [7:43] is a very cool feature. I think this is [7:45] awesome for a few reasons. [7:47] Uh one is that it doesn't feel like full [7:50] gen AI to me where it's like going to [7:52] slap it up. It's more just like a nice [7:55] feature that's in the actual camera app [7:57] photo roll camera roll. I thought that [7:59] was cool. Um and also this doesn't feel [8:02] like oh yeah, this is something that [8:03] every that there was already a startup [8:05] doing and it was already baked into [8:06] Instagram and so Apple's just rolling it [8:08] out and they're like behind the ball. [8:09] This feels like the first time I've seen [8:10] this. Seems really obvious. You could [8:12] probably one shot this in uh images or [8:16] Nano Banana or any of the image models, [8:18] but it was cool and it wasn't like oh [8:20] yeah, this is a startup and they're just [8:21] like rolling it in. This feels like [8:23] Apple's DNA of like understanding the [8:25] technologies and then doing something [8:28] cool with it and unique was on display [8:30] here. So I thought this was cool. What [8:32] was interesting [8:33] >> I think some of the pushback is that [8:35] a lot of this kind of functionality had [8:37] been available in other apps like [8:40] post-production apps. And now bringing [8:42] this basically bringing it into [8:44] effectively the realm of the camera [8:47] >> Sure. [8:47] >> where the computer is now the camera. It [8:50] does feel notable to me. [8:52] >> But so notable how? Like pushback? [8:55] Because there was pushback which was [8:56] interesting. People were saying that [8:57] like this is too much AI [8:59] >> Like I just view this I just view this [9:01] as a as a [9:03] >> Capture. That's it. [9:04] >> This is a way to capture reality. And [9:06] now in the reality capture device you [9:09] can use AI to [9:12] >> Apple has long prided themselves on like [9:14] what you see is what you get in the [9:15] camera app, but that started to change [9:17] with the artificial depth of field, the [9:20] tone mapping that happens. There's a lot [9:22] of things that you can do. [9:23] >> Yeah, there's a lot of [clears throat] [9:24] like upscaling and sometimes you see an [9:25] image and people are like oh that's AI, [9:27] but it's actually just upscaled Apple [9:28] image. [9:29] >> Yeah. Yeah, especially if it's low light [9:30] noisy image, it'll clean all that up. I [9:33] think it's fine. I you You just not do [9:35] it if you want to [9:37] not have your photos reframed, but it [9:39] was interesting seeing that uh that [9:41] there's a lot of folks that do land in [9:43] that camp of [9:45] Apple is the is not going too hard into [9:48] AI. They're not stuffing AI everywhere, [9:50] and then here's a feature that feels [9:51] very on the nose in terms of AI. We'll [9:53] see. We'll see. Uh and there's mixed [9:55] feelings on the timeline about spatial [9:57] reframing with some calling it one of [9:58] Apple's most exciting AI features, while [10:00] others are expressing concern about the [10:02] detection direction AI-generated imaging [10:05] is is headed because you can I mean [10:07] Apple's been mediocre at like the remove [10:10] trash in the background feature. Like uh [10:12] Gemini or I guess uh Android AI, like [10:16] the AI camera on Android devices has [10:18] been smoking them there. They're [10:19] probably going to get better at that. [10:20] But pretty soon it's like you take the [10:21] photo in super low light, it up-reses [10:23] it, camera filter, color grade, spatial [10:26] reframing, remove everyone else, and you [10:28] get a totally different photo. And I I I [10:31] think that's fine. I I don't have a [10:32] problem with that. But like I understand [10:34] why people would be [10:36] >> I mean, it's a little bit strange [10:37] because your [10:38] >> Just take the normal photo. Like just [10:39] turn off all the features. [10:40] >> At least at least uh yeah. No, and I I [10:42] think a lot of people [10:42] >> It's an option, you know? You can always [10:44] pull it into Photoshop and just do [10:45] whatever you want with it. [10:45] >> Yeah, it it'll just be interesting [10:47] because you're going to we're going to [10:48] enter an era where people's memories of [10:51] their lived experience are different [10:53] than [10:53] >> the Joe Weisenthal view of the future. [10:56] Joe Weisenthal got in some hot water [10:57] because he said that uh why do you need [10:59] to store any photos in the cloud when [11:01] you could just [11:02] >> No, I mean [11:03] >> delete them all, and then on demand show [11:05] me a photo of my kid riding uh a [11:07] dinosaur when he was five, and it just [11:10] shows you a photo of that. And then [11:11] people were like, "That is not the world [11:13] I want to live in, Joe." [11:14] It's very funny. [11:15] >> Yeah, but we're just not that far off [11:16] from Hey Siri, your phone's in your [11:19] pocket, and you just have your AirPods [11:21] in. You say, "Hey Siri, make sure to [11:23] generate some images of of my time at [11:25] Disneyland today." [11:26] >> Mhm. Okay. [11:28] Yeah. [11:28] >> I am interested to see what's It's to [11:30] happen on Instagram, and [11:32] >> if there ends up being divide between [11:34] accounts that that are constantly [11:36] getting the tag like the AI tag on them. [11:39] >> And people that just say like, "Nope. No [11:42] AI ever." [11:43] >> Yeah. [11:43] >> And and I think I think there'll be a [11:45] pretty clear split. [11:46] >> Yeah. And and and when do you apply that [11:48] tag? Little spatial reframing? [11:50] >> No, I think Instagram I think they do [11:52] the tagging. [11:53] >> But would they for a color grade? Would [11:55] they would they [11:56] >> to figure out what the line is, but to [11:57] me to me that's spatial reframing. [12:00] >> To say make it black and white, does [12:02] that count? [12:02] >> Spatial reframing is slightly [12:05] cuz it's capturing a picture that never [12:07] existed. Whereas a filter is just sort [12:10] of like enhancing color. [12:12] >> That's kind of just separating the [12:12] layers, rearranging them. I don't know. [12:14] It it's it's clearly like a blurry line. [12:16] But but it's the blurriest we've ever [12:18] seen. [12:18] >> But we could use AI and use AI to make [12:20] it less blurry to make it less blurry. [12:22] >> Yeah. Good. And that's literally what [12:23] they do cuz you can see that the the [12:25] rest of the space is sort of blurry and [12:26] then it puts together. Tyler, what do [12:27] you think about this? [12:28] >> Yeah, I was just saying I I was kind of [12:30] surprised that they didn't release more [12:31] like AI detection kind of features [12:33] around this. Like I I would imagine that [12:35] like I assume there will be some [12:36] metadata if you do the reframing that [12:38] says that it [12:38] >> Yeah. [12:39] >> used the reframing. Maybe it doesn't say [12:40] it's AI. AI, true, but it's like [12:42] >> Yeah. [12:43] >> That's true, yeah. [12:44] >> Yeah, easy. [12:45] >> Uh so this was my wild card topic [12:46] yesterday. Maybe they'll talk about [12:48] something related to the fertility [12:50] crisis or something and this feels like [12:52] three steps away from it, but I like the [12:53] direction that they went for a few [12:55] reasons. So we got a bunch of new AI [12:56] features, especially the new Siri [12:58] technology and UI [13:00] >> patterns are mature, we talked about [13:01] this. So it's easier to for Apple to [13:03] nail. Also, as I predicted, they're [13:05] giving Siri an app and the models are [13:07] strong enough that I expect random [13:08] hallucinations to be basically [13:10] completely acceptable from a daily use [13:12] perspective. But my wild card was that [13:14] something in the presentation loosely [13:16] would link to the topic of the fertility [13:18] crisis that Derek Thompson has been [13:19] covering recently. There was that [13:21] article in the Financial Times, our [13:22] phones causing brain rot or problems. [13:25] Like how do you address this? The [13:27] correlation isn't perfect, for sure, but [13:29] there's enough evidence that reasonable [13:30] people are start are starting to put [13:32] phones in like the probably linked [13:34] bucket of reasons why there's a modern [13:37] fertility decline. Uh it's just one of [13:39] them. I think Derek Thompson says like [13:40] he's maybe 30% of the reason, something [13:42] like that. But what's interesting here [13:44] is that Apple is not a fear-based [13:45] marketing company. So their execs don't [13:48] go on podcasts and make freewheeling [13:50] proc proclamations about doom scenarios. [13:52] They identified environmental concerns [13:54] about fossil fuels and energy use early. [13:56] And you cannot find a clip of Tim Cook [13:59] talking about like climate apocalypse, [14:01] we're all going to die because the sea [14:03] levels are going to rise. That's not He [14:05] wasn't saying that for years and then [14:07] figured out how to do clean energy for [14:10] Apple's data centers. That's not what [14:11] they did. They just They just worked [14:13] privately and then started talking about [14:15] the solutions in their marketing [14:16] materials. They didn't put the cart [14:18] before the horse. And it seems like [14:19] they're doing the same thing here with [14:20] parental controls, phone addiction. [14:21] Instead of giving a long speech about [14:23] how bad endless scrolling can be and how [14:25] it might lead to a less flourishing life [14:26] down the road, they only talk about [14:28] solutions, parental controls. 12 minutes [14:30] of content in a notably short keynote. I [14:32] think many parents are very hesitant to [14:34] give kids smartphones these days. So [14:36] it's also a good business reason to say, [14:37] "Hey, you can buy your kid an iPhone [14:39] because there are so many controls that [14:41] you are fully in control. You know [14:43] what's going to happen. You're safe with [14:45] us." Uh and there's already a whole [14:47] niche market in these dumb phone devices [14:49] for kids. There's this watch that has [14:51] GPS and the ability to call home, but [14:53] not much else. There's no screens. [14:54] There's no social media. And so being [14:56] proactive about building solutions and [14:58] then only talking about the solutions is [15:00] just extremely refreshing to me. And I [15:02] think it's a good move. Anyway, the [15:03] timeline was in turmoil and we'll take [15:05] you through it after I tell you about [15:06] Cisco, critical infrastructure for the [15:08] AI era. Unlock seamless real-time [15:10] experiences and new value with Cisco. Uh [15:12] we got Gruber versus Gur Gurman, two of [15:15] our best friends from the show. We love [15:17] both of these guys. [15:18] >> versus the Gubernator. [15:19] >> It's rough, but we'll take you through [15:21] it what happened. So, John Gruber and [15:23] Mark Gurman threw the timeline in [15:25] turmoil over the rollout of Apple's new [15:28] Siri integration strategy. Here's the [15:30] play-by-play. [15:30] >> Citizen says again, this all sounds [15:32] complicated. Just generate your kids. [15:34] >> Generate your kids. Back in March of 20 [15:38] March 26th, 2026. So, 4 months ago, 3 [15:40] months ago. Mark Gurman reported in [15:42] Bloomberg that Apple plans to open up [15:44] Siri to rival AI assistants in iOS 27 [15:48] update. And so, that was the headline. [15:50] He got some scoops and he says the [15:51] company's working on developing new [15:52] tools to allow AI chatbot apps installed [15:55] via the App Store to integrate with the [15:57] Siri AI assistant. Um the chatbots also [16:00] work with upcoming Siri app and other [16:02] features in the Apple intelligence [16:03] platform. That means for instance, if [16:04] you have Gemini or Claude or ChatGPT [16:07] installed, they'll be able to send those [16:09] queries to those services within the [16:11] Siri voice assistant, just like they [16:13] have been able to do with ChatGPT since [16:16] Apple intelligence launched in 2024. But [16:18] as of yesterday, according to uh John [16:21] Gruber, Apple has not fully announced [16:23] exactly that. At least according to John [16:25] Gruber who says, quote, "Maybe Apple ran [16:27] out of time today and will announce this [16:28] tomorrow." That was yesterday. "Maybe [16:30] they forgot to announce it. Maybe they [16:31] scrapped the next generation Siri that [16:32] existed 2 months ago and in the last [16:35] month rebuilt an entirely new next [16:37] generation Siri. I'll bet something like [16:38] that is what happened." So, he's taking [16:40] shots. He's putting Mark Gurman in the [16:42] truth zone. But Gurman fired back and [16:44] said, [16:46] "Here's some screenshots of integrations [16:47] that basically match his original [16:48] reporting." And so, Gurman's saying, [16:50] "No, no, you can actually go There's a [16:52] model picker in the Siri app. You can [16:54] configure Siri to work with other apps." [16:57] Um but there's a deeper question here. I [16:59] think they're sort of quibbling over [17:00] like minor leaks and rumors and like the [17:03] interpretation of these things. And I [17:04] think actually think there's like enough [17:05] of a gap here that it's not totally [17:09] unreasonable for them both for both of [17:10] their interpretations to be roughly [17:12] correct. But [17:14] >> Someone in the X chat says, "I I like [17:15] I'm watching this in 2x speed. [17:17] >> Slow down. Hey, 0.5x button's right [17:19] there, buddy. 0.5x button's right there. [17:21] I don't actually know [17:22] >> This is great. This is This is a [17:23] feature, not a bug. I'm sorry, Humble. [17:25] >> You got to warm up for Deli every once [17:26] in a while. Okay, anyway. So, [17:28] >> John's on a tear. Let him cook. [17:29] >> I think the core question is about [17:31] expectations versus reality. There was [17:33] some hope from hardcore AI users that [17:35] the Siri button, me, the Siri button [17:37] would be able to fully remap to their AI [17:39] model of choice. So, you're deeply [17:41] integrated in ChatGPT and has knows a [17:43] bunch of things about you. You're [17:44] confident with that model. You know what [17:45] what keywords to use when. You have the [17:47] integration set up. You press the Siri [17:49] button and you wind up getting that [17:51] model by default. Currently, you have to [17:52] say, "Hey Siri, go ask ChatGPT to do [17:55] this for me." And then it gives you this [17:57] little pop-up with this button. You have [17:58] to click okay. [17:58] >> Yeah, you're playing telephone. [17:59] >> It's very Yeah, you're playing [18:00] telephone. It's not good. Now, there are [18:02] screenshots of a model picker drop-down [18:04] in the Siri app. Uh but it's unclear if [18:06] that resets on some regular cadence, [18:08] like maybe every query, or if it's a [18:10] permanent change. And I'm sure I'm sure [18:12] we'll learn more. Uh Marques Brownlee [18:13] had an interesting thing that they're [18:14] calling uh what what was it? Maybe like [18:16] the new the new operating system is [18:19] called Golden Gate. And he was saying [18:20] that like the nominative determinism of [18:22] calling something Golden Gate is like [18:23] it's a golden gate in the sense that [18:25] they are maintaining their gated walls, [18:27] their walled garden, and producing [18:29] endless gold from mining uh what's [18:31] inside the gates. It could just be that [18:33] they like San Francisco, but [18:34] >> [laughter] [18:35] >> uh I I like [18:36] >> one or the other. [18:36] >> As someone who uh as someone who loves [18:39] nominative determinism, I thought it was [18:40] a fun it was a fun take. Brian Mac Duff [18:44] shares exactly what's going on with the [18:45] cleanup feature, drastically improved in [18:48] iOS 27. Uh this feels like it might be [18:50] nano banana under the hood or something [18:52] fine-tuned on that under the hood. Uh it [18:55] was very rough in iOS 26. There's also [18:57] this interesting detail that I got more [19:01] information on the private cloud. So, uh [19:04] the private cloud has been extended into [19:09] Google Cloud. So, for AFM, which is [19:12] Apple Foundation Models, AFM 3, this is [19:14] the third iteration, AFM 3 crap cloud [19:17] pro, which is their reasoning model [19:19] that's [19:20] fine-tuned of Gemini or some trained on [19:22] Gemini. [19:23] Um [19:24] They say they worked with Google and [19:26] Nvidia to extend private cloud compute, [19:29] which they put a bunch of resources into [19:30] securing and branding, [19:32] into onto Nvidia GPUs in Google Cloud [19:36] while maintaining the same guarantees to [19:37] protect our user privacy. Apple's been [19:39] very privacy forward. They have a ton of [19:41] trust. People put everything on their [19:42] iPhones. They're very, very confident in [19:44] that. And Apple's bringing that, but [19:46] they're vending it into Google, which is [19:48] like the advertiser, the dangerous one, [19:50] the one that [19:51] is listening to you. Not really. Like [19:53] it's actually like Facebook that's more [19:54] of like the listening one, but back in [19:56] 2023 I said, "With all these incredible [19:58] advances in conversational AI chatbots, [20:00] I'm going willing to put down a firm [20:01] prediction. By the year 2043, Siri will [20:04] be usable." [20:05] >> [laughter] [20:05] >> And I think it came true. I I was [20:07] correct. I think Siri's usable now. I [20:10] think it's [20:10] ahead of schedule. Oh, yeah, it doesn't [20:12] release yet. It's a It's not live for [20:13] another couple more months. [20:14] >> Okay, I'm downloading the beta. [20:16] >> Let's [laughter] go. [20:17] Okay. At one point I actually [20:19] >> You're You're You're archive of posts. [20:20] >> I know. I know. I I You got to know this [20:22] lore. At one point I replaced Siri with [20:24] ChatGPT via an iOS shortcut. [20:27] >> January 8th, 2023. [20:29] >> Yeah, early. [20:30] >> complicated [20:31] >> Yeah, so you go in and it would and it [20:33] would and it would take an API key. You [20:36] would trigger this with a shortcut. [20:38] You'd say, "Hey Siri, trigger Siri pro [20:40] mode" or something. Or you'd ask Siri [20:42] pro mode a message and then it would [20:44] route that to ChatGPT. Before they had [20:46] the integration, you had to do this [20:47] manually. And it and it would You'd have [20:49] to set all this up. But it was pretty [20:50] easy to copy and paste that. But [20:53] important Apple history, Siri was an [20:55] acquisition. M&A for AI is in Apple's [20:58] DNA. This was a reminder from last year [21:00] when Apple met with Mira Murati, the [21:02] former CTO of OpenAI, to discuss a [21:04] potential deal for her new AI startup, [21:07] Thinking Machines Lab. Uh the talks [21:09] never progressed to an advanced stage. [21:11] What a different fork in the road moment [21:13] they went with Gemini. But, uh they [21:15] could have had Thinking Machines Lab [21:17] inside Apple, all of those incredible [21:21] engineers and that that talented team. [21:23] But, they would have had to pay through [21:23] the nose and Apple, you know, Siri was I [21:25] think a couple hundred million dollar [21:27] acquisition. It was not in the billions. [21:29] >> I think it was 200. [21:30] >> Yeah, and Thinking Machines would have [21:31] been 10 or something. It would have been [21:33] a lot to to to actually get that deal [21:35] across based on where they were. [21:37] Gaussian splatting is coming to Apple [21:39] Maps. This is very cool. Wow, it looks [21:41] so much better. Look at this. [21:43] >> Wild. [21:43] >> Normal Apple Maps, no, that is what [21:45] looks good. Right there. Once it gets [21:47] the color grade, I like that. Uh [21:50] I don't know when you would use this. [21:52] Maybe if you're looking around a town. I [21:54] don't know. [21:54] >> I feel like this would be very cool if [21:57] you were in a new city and you were [21:59] trying to kind of like learn [22:00] orientation, right? Like, let's say you [22:02] check into a hotel in a city you've [22:03] never been to before. And you can [22:05] navigate around and kind of get a feel. [22:07] >> Or like you could do this at Disneyland [22:09] with your AI generated kids and you [22:11] could pick the particular region of [22:14] Disneyland that you want to generate a [22:16] fake image of your fake kids and then [22:18] you could post that on your fake social [22:20] [laughter] network. And you could just [22:22] slap it up from start to finish. This [22:23] one's crazy. The North Korean economy [22:26] becomes the world's most unlikely [22:29] success. We've been talking up Pyongyang [22:32] for a while. [22:34] We've been saying a lot of people raise [22:35] from PIF. That's usually the Public [22:39] Investment Fund from Saudi Arabia. But, [22:41] if you just tell people I raised from [22:42] PIF and it turns out it's the Pyongyang [22:44] Investment Fund, you might go to jail [22:46] cuz they're sanctioned. But, you might [22:48] be able to sneak it across the finish [22:50] line. But, they are booming. Tyler, you [22:51] read this story. You you had a you had [22:53] some takeaways. What they do? How many [22:55] Oh, they built 10,000 homes. Yeah, which [22:58] I I believe is more than Uh, Los [23:00] Angeles. Yeah. [23:01] >> How are they beating us? [23:02] >> So, they're kind of I think I think they [23:04] read um [23:05] abundance, right? [23:06] >> Yeah. Yeah, that makes sense. [23:07] >> Yeah, that makes sense. [23:08] >> They got abundance killed and then they [23:10] said, "Yeah, we got to build housing." I [23:12] mean, it's like if they're doing it over [23:13] there, [23:15] it's like what are we doing here? [23:17] Like the most sanctioned and [23:20] locked up economy, but they are making [23:22] money. They're selling arms to Russia [23:24] and China. They have lots of nuclear [23:25] weapons. They're building more by the [23:27] day. [23:28] Uh, and so they're they're they're on [23:31] they're on track. Leopold also made it [23:32] to the front page of The Wall Street [23:33] Journal. Congrats. [23:34] >> There we go. Claude Fable 5 is out. We [23:37] mentioned it earlier in the show. Bunch [23:40] of feedback coming in. People on X [23:42] having have had access to the model for [23:45] a while now. [23:46] Dan Shipper [23:47] >> Which was the information have against [23:48] it? They called it a neutered version of [23:51] Mythos. That's like sort [laughter] [23:54] >> I I think they say [23:54] >> No, they're not framing it as neutered. [23:56] They're they're framing it as more safe. [23:58] >> [laughter] [23:58] >> It's a safer and more reliable for [24:00] certain things and probably cheaper. [24:02] >> [laughter] [24:03] >> I don't know. [24:03] >> Yeah, absolutely crushes the benchmarks. [24:05] Dan Shipper's doing a vibe live vibe [24:07] check right now. He had access to the [24:09] model. We will [24:11] uh, we'll summarize all the posts from [24:13] today. [24:14] Get them into [snorts] [24:15] >> The pricing is $10 per input millions of [24:18] tokens, $50 of output per millions of [24:21] tokens. So, if you're token maxing, get [24:24] ready to pay. [24:25] Um, but the capability seems [24:27] >> yeah, quite a bit more [24:28] >> People are talking about [24:30] like running it for 9 hours straight, [24:32] running it over the weekend for just [24:34] long long time without getting confused. [24:36] Uh, very exciting model. Noah Brown had [24:38] an interesting piece on the implications [24:41] of large-scale test time compute that I [24:43] thought tied to this a little bit. And [24:45] there was a very interesting quote in [24:47] here where he said, "Uh, frequently when [24:49] I discuss this, people ask why we don't [24:51] just evaluate a harness that pushes test [24:53] time compute until performance plateaus. [24:55] The problem is that empirically, the [24:57] plateau is very far out. Sometimes, we [24:59] may not observe a plateau at all within [25:02] practical budgets. You can just spend, [25:04] spend, spend. Uh so, Karpathy under [25:06] Karpathy saw this in the auto research [25:07] experiment where the performance [25:09] continues to improve even after hundreds [25:11] of experiments. He He goes on to talk [25:14] about benchmarks a little bit, but at [25:15] one point, he says that a model might [25:18] run for longer than it takes to train [25:21] the next run, which I thought was such [25:23] an interesting such an interesting [25:25] concept. Just this idea that you launch [25:28] a product, you're running it, I say go [25:31] do a job, and then before it completes [25:33] that job, you got the next model ready [25:36] and released, uh which is a crazy, crazy [25:39] world to be living in and something I I [25:40] I don't think people predicted, but more [25:42] and more compute, more and more [25:44] inference, more and more reasoning all [25:47] across the board. Meta launched a [25:49] workforce academy to train workers to [25:51] build data centers. With five-week [25:53] program, which is free of charge, [25:54] guarantees a job, follows the recent [25:56] layoffs of 8,000 employees. Uh this is [25:59] the learn-to-weld meme in real life. Uh [26:02] forget learning to code, Meta platform [26:04] says it's time to pick up a wrench. [26:06] Company is starting a workforce academy [26:08] to train Americans to build its data [26:10] centers as a skill a skilled trade [26:12] workers become sought-after commodity. [26:14] Five-week training program in [26:16] partnership with CBRE and the Associated [26:18] Builders and Tyler [26:20] >> has already been accepted into the [26:22] program with a disguise and a fake name. [26:25] He will be learning to weld. [26:27] >> Huge news. Huge [26:28] >> Now, that would be fun. We should [26:30] consider it. We got to talk about flock [26:32] safety. [26:33] >> Okay. [26:33] >> Can we pull up this video? An admitted [26:35] criminal discussing the impact of flock [26:38] >> Okay, let's play it. [26:39] >> on [26:40] >> Just crime in San Francisco, period. [26:41] [laughter] That [ __ ] over with, brother. [26:42] Oh, on my mama, [ __ ] they got drunk [26:44] I'mma keep it [26:46] >> TBP unsafe. We need to bleep this before [26:48] we play it. [26:49] >> Okay. We're We're to summarize it for [26:51] you because I didn't realize that there [26:53] were so many uh swear words. But [26:55] basically they say you can't do crime in [26:59] San Francisco anymore because [27:01] >> gentlemen are are admitting to have [27:04] previously participated in illegal [27:06] activity. [27:06] >> Exactly. And they [27:08] say basically if you [27:10] if you steal a car, a drone will start [27:13] following you immediately and it will [27:15] just trail you from thousands of feet [27:17] up. You won't even necessarily know that [27:19] it's following you. [27:20] >> Yes. [27:21] >> And uh [27:22] >> And so they won't So so you won't have a [27:24] chance to like run away from a cop that [27:26] finds you because they'll just wait [27:27] until you're actually parked or stopped [27:30] and then they'll box you in. [27:31] >> the host of the podcast says, "So can [27:33] you even steal a car, run up on your [27:35] ops, and ditch the car after?" And they [27:38] they say no. And so [27:40] >> He He doesn't just say that. He He [27:42] describes it as the classic steal a car, [27:44] run up on your ops, and then ditch the [27:46] car like it's a thing that's generally [27:49] accepted. Not to [27:51] lay out a bare case for Flock, but has [27:53] anyone considered listening to podcasts [27:55] where where criminals admit to crimes [27:58] and going arresting them? It seems like [28:00] there's a [28:01] >> Talk about their methods? [28:02] >> It seems like there's a there's a dearth [28:04] of these podcast clips, but who knows? [28:06] >> Yeah, I mean these aren't even just like [28:07] random petty criminals. They're listed [28:09] as [28:10] The gentleman on the left is one of the [28:11] most prolific criminals in all of San [28:13] Francisco. [28:14] >> Wow. [28:14] >> So presumably like yeah, someone in the [28:16] police force can just be like, "Oh, [28:18] that's an interesting tweet." [28:19] >> Let's Google [laughter] it. [28:21] >> This guy complains that he can't even do [28:22] drive-bys anymore. [28:23] >> Wow. [28:25] That That is wild. That is a wild time. [28:27] Anyway, let me tell you about Kodak. [28:29] Kodak is a powerful workforce for [28:32] getting work done with AI agents, [28:34] whether you're writing code, analyzing [28:35] data, creating content, or automating [28:37] business workflows. Kodak helps you move [28:39] projects forward from start to finish. [28:41] Um where should we go next? Pat McAfee [28:43] is negotiating new deal that could pay [28:45] him over $60 million a year for for [28:48] ESPN. This is in The Athletic. Big news [28:50] for one of our biggest role models. ESPN [28:53] and representatives for Pat McAfee are [28:55] discussing an extension to his contract [28:57] that would pay him more than 60 million [28:58] per year. Deal's not completed yet, and [29:01] if an agreement can be reached, it could [29:02] be a sliding scale based on his new [29:04] responsibilities. McAfee, already [29:06] omnipresent, [29:08] could be on the air even more with a [29:10] bigger role in NFL coverage. I'd love [29:12] that. [29:12] >> And last night he had a like watch along [29:16] like on actual TV for the the next game. [29:19] >> That's super cool. Yeah. And and he's [29:21] been brought into the the college [29:23] football ecosystem and been a really [29:26] important voice there. He's 39, two [29:29] years remaining on his current contract, [29:31] and he's [29:32] a College GameDay panelist and has [29:34] various appearances on other on other [29:37] programs. The makeup of the new deal [29:38] would be similar in structure. ESPN view [29:40] the arrangement as a production contract [29:43] and a separate talent agreement, [29:45] differentiating it from the deals with [29:46] most of the on-air personalities. Hosts [29:48] a 3-hour daily show with his crew, the [29:50] Pat McAfee Show. The first 2 hours are [29:52] on ESPN, all the all three are on [29:54] YouTube. Very interesting split there. [29:56] >> Stay sane out there. Leave us five stars [29:58] on Apple Podcasts and Spotify. [29:59] >> Subscribe to our newsletter, tvfan.com. [30:02] >> See you tomorrow. [30:03] Goodbye.