Acemoglu Caps AI Labor Displacement at 5% Over Decade to Counter Hype

阿西莫格鲁称十年内AI替代劳动力仅限5%以反驳炒作 诺贝尔奖得主达龙·阿西莫格鲁（Daron Acemoglu）认为，未来十年人工智能仅能实现5%的人类工作任务自动化，因为算法的进步在实体经济中会遭遇阻滞，除非开发者将重心从裁减人员转向赋能劳动者。 John Coogan和Jordi Hays深入探讨了汤姆·克鲁斯秘密开展的表演课程，以及AI辅助电子游戏反编译的兴起——该技术正在催生各种新奇的混搭作品与网页端主机游戏移植，同时也对版权法提出了挑战。他们分析了麻省理工学院经济学家达龙·阿西莫格鲁对AI大规模取代人类劳动所持的审慎怀疑态度，并将其与连续语音识别等历史上技术采纳周期的滞后现象进行了类比。本期内容还重点关注了Mistral推出的1万亿参数模型“Le Chonk”、a16z关于ChatGPT在移动端领先Claude和Gemini的数据，以及96岁高龄的沃伦·巴菲特晚上热衷于在YouTube上探索各类冷门内容的趣闻轶事。 前沿大语言模型正在实现将电子游戏二进制文件从汇编语言自动反编译为C/C++，推动了现象级混搭和类似Xbox《光环：战斗进化》（Halo: Combat Evolved）的浏览器端移植，同时也引发了关于合理使用和版权的新担忧。 (4:10) 麻省理工学院经济学家达龙·阿西莫格鲁预测，未来十年AI将仅能实现约5%的人类工作任务自动化，并为GDP总量贡献1.5%的增长；他警告称，若以替代人工而非辅助劳动者为导向进行开发，可能会带来经济停滞的风险。 (10:42) Mistral发布了原生多模态、总参数量达1万亿且激活参数为490亿的Mistral Large 4（“Le Chonk”），为西方企业提供了一个对抗中国前沿模型的有力开源权重替代方案。 (25:26) a16z的最新数据表明，尽管竞争激烈且Claude在移动应用榜单上异军突起，OpenAI的ChatGPT仍保持着巨大的消费端领先优势，其网页端规模约为Gemini的2倍，移动端规模约为Claude的14倍。 (29:47)

The program examines diverging expectations for artificial intelligence, contrasting cautious macroeconomic forecasts with rapid technical developments in software decompilation, open-weight model development, and consumer distribution. Discussion centers on whether AI's practical diffusion will mirror historical productivity lags or rapidly transform software architecture and intellectual property. Labor Automation and the Economic Diffusion Debate (10:42) MIT economist and Nobel laureate Daron Acemoglu projects that AI will automate only roughly 5% of aggregate human work tasks over the next decade, adding a modest 1.5% to total GDP. Comparing current AI adoption to Robert Solow's computer productivity paradox and the prolonged rollout of electricity, Acemoglu argues that foundational general-purpose technologies require decades to yield measurable macroeconomic gains. He highlights speech recognition pioneer Dragon Systems, which launched continuous consumer dictation software in 1997 before corporate fraud, corporate consolidation, and enterprise pivots toward healthcare transcription stalled consumer desktop progress for nearly two decades until modern neural models emerged. Acemoglu also disputes claims that technological transition will guarantee net job creation—such as assertions by Mercor founder Brendan Foody—warning on 60 Minutes that focusing AI design on worker substitution rather than worker augmentation risks severe disruption. In contrast, the hosts argue that modern software spreads across web-connected businesses far faster than physical infrastructure, citing consumer brands successfully integrating AI across creative workflows and demand forecasting. Automated Game Decompilation and Intellectual Property (4:10) Frontier large language models have automated the reverse-engineering of compiled software by translating low-level assembly code back into editable C and C++ source code, a task that previously demanded years of human engineering. This capability has fueled viral community gaming mashups and technical feats, including a browser port of the original Xbox console binary of Halo: Combat Evolved. Modders circumvented official releases by decompiling the machine code directly, demonstrating the speed at which modern models unravel proprietary binaries. This trend raises distinct legal, technical, and operational challenges. While precedents like Sega Enterprises Ltd. v. Accolade, Inc. and Sony Computer Entertainment, Inc. v. Connectix Corp. established that disassembly can qualify as fair use for functional interoperability, automated decompilation generates code nearly identical to original copyrighted works, exposing modders to DMCA takedown notices. Meanwhile, AI labs gain valuable reinforcement learning and code-translation datasets from massive token consumption by modders. To protect proprietary binaries from local extraction, game studios may face mounting pressure to abandon local executables in favor of server-side cloud streaming architectures. Western Open-Weight Models and Benchmark Competition (25:26) European AI lab Mistral released Mistral Large 4, nicknamed 'Le Chonk,' a natively multimodal model featuring 1 trillion total parameters and 49 billion active parameters hosted on European cloud infrastructure. The model claims top rankings among Western open-weight releases, outperforming frontier closed models on visual grounding and scoring 15% on the Harvey legal agent benchmark compared to competitor scores between 5% and 13%. API access is immediate, with public weights scheduled for late October. Technical analysis positions Mistral Large 4 as an important alternative for Western enterprises restricted from deploying fast-moving Chinese open-weight models due to prospective compliance and geopolitical risks. Concurrently, enterprise developers are shifting from passive consumption of proprietary frontier APIs toward defining custom evaluation benchmarks tailored to their specific operational workloads, forcing AI labs to compete directly on task-specific efficiency and pricing. Consumer AI Platform Share and App Distribution (29:47) Market share research from Olivia Moore at Andreessen Horowitz shows that OpenAI's ChatGPT retains a substantial lead over competitors in consumer web and mobile adoption. Across web traffic, ChatGPT operates at approximately twice the scale of Google Gemini and six times that of Anthropic's Claude. In mobile distribution, ChatGPT's advantage widens to 2.5 times Gemini and roughly fourteen times Claude. Despite ChatGPT's aggregate market dominance, the mobile ecosystem exhibits rapid short-term volatility. Anthropic's Claude briefly reached number one on the mobile app store charts before ChatGPT reclaimed the top ranking. This shifting dynamic reflects active competition among frontier labs seeking to secure consumer distribution and retention across mobile interfaces. While macroeconomic indicators and historical precedents from speech recognition suggest enterprise productivity gains may lag behind algorithmic breakthroughs, technical capabilities in code decompilation and open-weight model architectures are advancing rapidly. Whether corporate inertia will restrain economic displacement or accelerated developer adoption will outpace traditional diffusion cycles remains the core unresolved question.

## Background 本期内容由科技播客频道 TBPN 于 2026 年 10 月 6 日在 YouTube 平台发布。节目为日常科技直播脱口秀《Diet TBPN》的 30 分钟精华版，由 John Coogan 和 Jordi Hays 主持，围绕 AI 对宏观经济的影响、开源大模型竞争、游戏反编译热潮以及科技与商业领域的近期热点事件展开讨论。 ## TL;DR 节目围绕多项前沿科技与文化热点展开：首先探讨了玩家利用大模型反编译将《光环》等经典游戏搬上浏览器引发的版权与云游戏思考，随后深入剖析诺奖得主阿西莫格鲁关于“AI在十年内仅替代5%人类工作”的争议观点，并结合语音识别历史反思技术的商业化扩散路径。此外，节目还趣评了沃伦·巴菲特深夜沉迷 YouTube 的日常、美国新版护照的字体排版争议，最后盘点了 Mistral 最新开源大模型“Le Chonk”的技术细节以及 ChatGPT、Gemini 与 Claude 在消费级市场的份额竞争格局。 ## Atmosphere & style 整场节目充斥着典型硅谷科技圈的快节奏与轻度讽刺风格，主播之间互动默契、插科打诨不断。从戏谑阿汤哥卖课、调侃巴菲特刷土味视频，到严肃推演宏观经济模型与开源权重生态，对话在硬核技术拆解与爆笑网络梗文化之间自如切换。听众仿佛置身于一场充满信息量却毫无距离感的科技播客茶话会，既能捕捉到前沿行业动态，又能感受到充满生活情趣的幽默张力。 ## Bulletin - [04:10](https://www.youtube.com/watch?v=0SbWBq9G_l8&t=250s) 大语言模型（LLM）正引发游戏反编译革命，玩家借助 AI 成功将初代 Xbox 版《光环：战斗进化》完整移植至网页浏览器运行，不仅推动了社区 Mod 狂欢，也为前沿实验室回流了高质量的代码对齐与强化学习数据。 - [10:42](https://www.youtube.com/watch?v=0SbWBq9G_l8&t=642s) 诺贝尔经济学奖得主达龙·阿西莫格鲁发表深度文章，预测未来 10 年 AI 仅会替代 5% 的人类工作，对美国 GDP 增益仅约 1.5%，但他在电视专访中同时警告若技术持续专注于替代人力而非赋能劳动者，将引发严重的失业与结构性冲击。 - [25:26](https://www.youtube.com/watch?v=0SbWBq9G_l8&t=1526s) 法国 AI 初创公司 Mistral 发布 1 万亿总参数的旗舰多模态模型 Mistral Large 4（代号“Le Chonk”），在网络安全与法律 Agent 测试中表现亮眼，为注重合规的欧洲及西方企业提供了替代方案。 - [29:45](https://www.youtube.com/watch?v=0SbWBq9G_l8&t=1785s) a16z 数据显示消费级 AI 呈现悬殊差距：ChatGPT 在网页端规模分别为 Gemini 的 2 倍和 Claude 的 6 倍，在移动端更是达到 Claude 的 14 倍，维持着极强的消费级统治地位。 ## Key takeaways - [00:01](https://www.youtube.com/watch?v=0SbWBq9G_l8&t=1s) 汤姆·克鲁斯曾为格伦·鲍威尔录制长达 6 小时的银幕表演指导课，注重电影构图本质与掌握规则后的遗忘，被节目主持人戏称为好莱坞巨星的“私域卖课教程”。 - [05:46](https://www.youtube.com/watch?v=0SbWBq9G_l8&t=346s) LLM 显著降低了将 x86 汇编代码反编译为高级 C 语言代码的工程门槛，使得经典游戏跨平台移植与病毒式 Mod 创作效率激增。 - [05:46](https://www.youtube.com/watch?v=0SbWBq9G_l8&t=346s) 反编译在技术与司法上有合理使用先例，但公开分发派生游戏资产会迅速触发版权方 DMCA 下架，直接商业化路径依然受阻。 - [05:46](https://www.youtube.com/watch?v=0SbWBq9G_l8&t=346s) 全球 Mod 社区自费购买 Token 进行逆向工程的浪潮，正成为 AI 实验室获取高质量游戏代码与真实强化学习环境的数据金矿。 - [09:11](https://www.youtube.com/watch?v=0SbWBq9G_l8&t=551s) 为应对本地二进制文件被瞬间反编译并篡改的风险，游戏工业或将加速转向全云端串流架构与以核心 IP 为壁垒的商业模式。 - [10:42](https://www.youtube.com/watch?v=0SbWBq9G_l8&t=642s) 阿西莫格鲁重提索洛悖论，认为底层技术突破必须与适配工具、商业渠道及用户反馈循环结合，否则无法转化为统计上的生产率激增。 - [13:42](https://www.youtube.com/watch?v=0SbWBq9G_l8&t=822s) Dragon Systems 的兴衰史表明，即便 1997 年语音识别准确率已达 95%，随后被跨国并购、财务欺诈及闭门深耕医疗企业级所累，消费级技术停滞了二十余年。 - [17:54](https://www.youtube.com/watch?v=0SbWBq9G_l8&t=1074s) 阿西莫格鲁在《60 分钟》电视采访中严厉反驳“十年后总就业岗位必定增加”的乐观预测，警示当前 AI 的破坏性不亚于早期工业革命。 - [18:54](https://www.youtube.com/watch?v=0SbWBq9G_l8&t=1134s) AI 就业辩论的核心分水岭在于构建范式：纯粹自动化替代劳动力将摧毁岗位，而用于解决净新增问题并赋能劳动者则能带来生产力扩容。 - [21:06](https://www.youtube.com/watch?v=0SbWBq9G_l8&t=1266s) 96 岁的沃伦·巴菲特每晚沉迷 YouTube 推荐流，观看老广告、业余翻唱与小众纪录片，被调侃为展示松弛感的心理战。 - [23:32](https://www.youtube.com/watch?v=0SbWBq9G_l8&t=1412s) 美国新版护照封面引发排版与设计争议，无衬线全大写文字的字距调配被批评具有廉价的“AI 垃圾内容”质感。 - [25:26](https://www.youtube.com/watch?v=0SbWBq9G_l8&t=1526s) Mistral Large 4 拥有 490 亿活跃参数（总参数 1 万亿），在 Harvey 法律 Agent 基准测试中取得 15% 得分，显著领先同业水平。 - [28:38](https://www.youtube.com/watch?v=0SbWBq9G_l8&t=1718s) 开源阵营的迅速逼近与闭源巨头之间的价格战，迫使企业逐步转向通过定制工作流基准测试来大幅压缩推理成本。 - [28:38](https://www.youtube.com/watch?v=0SbWBq9G_l8&t=1718s) 西方开源模型与中国顶尖开源模型的迭代周期存在时差，但地缘政治与数据合规考量仍是欧美企业选型的重要决策因素。 - [29:45](https://www.youtube.com/watch?v=0SbWBq9G_l8&t=1785s) 移动端应用商店中 Claude 与 ChatGPT 展开了激烈的榜单排名拉锯战，反映出大模型消费级入口争夺战进入白热化阶段。 ## Detailed walkthrough ### [00:01](https://www.youtube.com/watch?v=0SbWBq9G_l8&t=1s) 汤姆·克鲁斯的秘密六小时表演课程与泄密赏金 节目开场，主持人提到今晨翻阅最新一期《GQ》杂志，封面人物是汤姆·克鲁斯（Tom Cruise）。专访中透露了一则极具喜感的趣闻：好莱坞新生代男星格伦·鲍威尔（Glen Powell）在采访中提到，阿汤哥曾专门制作了一部长达六小时的电影表演教学视频给他看。阿汤哥在采访中坚称这套教程“非常基础”，主要是帮年轻演员理解电影镜头的本质、画面构图（Frame）的含义，以及如何像舞台剧或电影史早期演员那样理解结构后再去“遗忘”规则。主持人戏称这简直就是“超级巨星速成指南（Being a Global Superstar for Dummies）”，调侃阿汤哥私底下根本就是个私域卖课老哥（Course Bro）。 主持人笑称大家必须想办法搞到这份内部视频，甚至开玩笑说应该悬赏一万美金（10 racks）给能拿到这六小时录像的人，然后把它塞进付费墙变现。借由这个话题，他们延伸探讨了科技圈记者的独家爆料机制。比如科技记者凯文·罗斯（Kevin Roose）手里就持有 Dario Amodei 2017 年撰写的算力大论文（Big blob of compute thesis，关于 Scaling Law 和 The Bitter Lesson 的早期泄漏 PDF）。主持人打趣道，科技媒体圈为什么没有人公开立赏金：“比如 Eric Newcomer 完全可以放话，谁能把 Founders Fund 的真实回报率数据交出来，就当场奖励五千甚至五百万美元。”不过大家也随即承认，公然发悬赏收买泄密文件可能已经严重违反新闻伦理（Journalistic ethics）。同时，许多通过 Signal 等端到端加密工具流出、违反保密协议（NDA）的泄密，只要不损害公司形象甚至能带来正面公关效应，大公司往往也就睁一只眼闭一只眼。 - 格伦·鲍威尔向《GQ》证实曾观看汤姆·克鲁斯为其录制的 6 小时银幕表演私人指导视频。 - 主持人调侃阿汤哥的指导理念：“我先倾听年轻人的目标，正如我当年还是个新人时四处追问一样。” --- ### [03:08](https://www.youtube.com/watch?v=0SbWBq9G_l8&t=188s) 团队观影风波与票房滑铁卢 话题随后轻松转入团队内部的生活琐事。团队成员昨日一同去影院观看了一部最新上映的大片，但评价极其两极分化。团队成员尼克（Nick）看了大约 40 分钟就忍不住掏出手机摸鱼，直呼难受；而另一位成员泰勒（Tyler）则认为电影其实还行，构图极具“电影学院派风格”（Film school like）。 主持人指出，这部影片的舆论与商业表现极其惨淡：整部电影制作成本高达 1.6 亿美元（160 million），然而首周末票房（Box office）仅有惨不忍睹的 800 万美元。尽管外界差评如潮，风投大佬马克·安德森（Marc Andreessen）却罕见地公开给出了高分好评，成为为数不多的力挺者之一。 --- ### [04:10](https://www.youtube.com/watch?v=0SbWBq9G_l8&t=250s) AI 逆向反编译热潮与经典游戏大乱斗 近期社交媒体 X（前 Twitter）上爆发了一场疯狂的游戏 Mod 与混搭（Mashup）狂欢。知名科技分析师本·汤普森（Ben Thompson）在 Stratechery 专栏中深入分析了游戏反编译（Game Decompilation）的技术革命及其版权法律风暴。网络上涌现出大量极具视觉冲击力的离谱二创：有人在《生化危机》（Resident Evil）里操控穿着冰刀的运动员打 NHL 冰球射击丧尸；有人把《使命召唤》（Call of Duty）的干员放进滑板游戏《Skate》里做踢翻（Kickflip）高难度动作；还有玩家将《黑色行动 1》（Black Ops 1）的僵尸模式搬进了《天际线》（Skyrim），或是让小黑蛛迈尔斯·莫拉莱斯（Miles Morales）在哥谭市的高楼间荡蛛丝。 其中最轰动的技术案例，是玩家直接将初代 Xbox 版本的《光环：战斗进化》（Halo: Combat Evolved）完整移植到了云端网页浏览器中运行，并且居然原生支持多人联机与分屏模式（Split screen）。这背后的滑稽原因在于，《光环》硬核社区一直极其厌恶官方当年发行的 PC 移植版，因此干脆利用 AI 工具反编译初代 Xbox 版本并塞进浏览器。 ```mermaid flowchart LR accTitle: AI-Assisted Game Decompilation and Training Loop accDescr: Pipeline showing game binary disassembly, AI decompilation into C code, and subsequent viral mods feeding back into model training. subgraph ReverseEngineering["逆向工程流程"] A["可执行机器码 (x86 机器指令)"] -->|"反汇编 (精准映射)"| B["汇编代码 (Assembly)"] B -->|"LLM 自动化理解与反编译"| C["人类可读高级语言 (C / C++)"] end subgraph DeploymentAndFeedback["落地与数据回流"] C --> D["社区创意 Mod 与跨平台移植"] D -->|"玩家消耗 Token / 生成海量数据"| E["AI 实验室强化学习与下一代训练"] end ``` --- ### [05:46](https://www.youtube.com/watch?v=0SbWBq9G_l8&t=346s) LLM 重塑反编译、法律边界与云游戏的未来 本·汤普森在文章中剖析了这一现象底层的计算机科学原理：游戏原本是用人类可读的高级语言（C/C++ 及相关脚本语言）编写的，最终被编译器编译成目标处理器（如 Halo 的 x86 架构）执行的机器码。传统上，机器码很容易反汇编（Disassemble）为汇编语言，但要将艰涩的汇编代码逆向反编译回清晰可读、方便修改的 C 语言源码，过去需要人类逆向工程专家耗费数年心血。而如今的大语言模型（LLMs）极擅长处理此类语法与逻辑转换，极大降低了反编译门槛。 在法律层面上，反编译本身并不直接违法。根据著名的“世嘉诉 Accolade 案”（Sega v. Accolade）以及“索尼诉 Connectix 案”（Sony v. Connectix，涉及 PlayStation BIOS 逆向工程）确立的先例：出于合理目的、为了获取不受保护的功能性要素而进行代码反汇编，属于合理使用（Fair Use）。然而，一旦通过反编译生成的派生代码在结构上与受版权保护的原版代码高度一致甚至完全雷同，公开发布 Mod 或独立游戏就会直接构成侵权。创作者哪怕只为博人眼球，作品在爆火后两小时内就会遭到版权方的 DMCA 下架通知函（Takedown notice），导致其根本无法直接商业化变现。 尽管 Mod 制作者无法靠这些病毒式传播的混搭直接赚钱，但这场运动对前沿 AI 实验室（AI Labs）来说却是一座数据金矿。全球游戏 Mod 社区疯狂调用前沿模型（例如 GPT-4o、Claude 3.5 Sonnet 等）处理从《我的世界》（Minecraft）、《天际线》到各种长尾小众游戏的二进制文件，自行烧钱购买 Token 进行反编译。AI 实验室不仅获得了充沛的强化学习（RL）真实环境，还收割了极其高质量的代码对齐与游戏开发训练数据，未来的大模型在游戏全流程开发上势必产生质的飞跃。最后，这一趋势也可能加速整个游戏工业的商业模式变革：为了防止发布本地可执行二进制文件（Executable binary）被瞬间反编译魔改，大型游戏厂商未来可能会更坚决地转向服务型游戏（Game as a Service）与全云端串流架构——正如当年 Google Stadia 尝试推行的云游戏技术路线。 ### [09:11](https://www.youtube.com/watch?v=0SbWBq9G_l8&t=551s) 云游戏模式与游戏代码的“防线” 讨论从云游戏技术（例如 Google Stadia 及 Meta Quest 串流 Xbox 游戏的机制）切入：在这类架构中，服务器负责执行游戏本体，向玩家下发视频画面，而玩家只需上传控制器输入信号。这种做法彻底隔绝了用户与本地游戏代码的接触，玩家除了流式传输回来的视频帧之外什么也看不到。尽管 Xbox 在云游戏领域深耕更久，但核心玩家群体一直对此抱有抵触情绪，普遍抱怨操作延迟过高、Wi-Fi 断连导致单机游戏体验崩溃等问题。 主持人们分析，面对当前越来越泛滥的黑客逆向破解与 AI 生成内容，游戏行业正面临两条截然不同的路径：要么抱团取暖，利用类似云串流的“防火墙”将所有核心代码深锁在远端服务器之后；要么彻底张开怀抱拥抱变革，转型走向依靠 IP 保护来变现的策略。正如 AI 生成文本在社交平台普及后，大众依然在乎发帖的人是谁、故事背后的创作者是谁一样，未来具体的游戏产品代码或技术实现可能会被逐渐边缘化，真正具备稀缺价值的，反而是围绕 IP 塑造的叙事壁垒、核心开发团队的号召力，以及所有人聚在一起玩“同一款游戏”的谢林点（Schelling point）共识。 --- ### [10:42](https://www.youtube.com/watch?v=0SbWBq9G_l8&t=642s) 诺奖得主阿西莫格鲁的“解药论”：AI 十年内仅替代 5% 的工作？ 话题转向人工智能对宏观经济的影响。微软 AI 负责人穆斯塔法·苏莱曼（Mustafa Suleyman）发表了一份被主持人们戏称为“超级白丸”（major white pill，极度乐观且令人安心）的观点：AI 绝不会在短期内抢走大众的饭碗，在未来 10 年内，人类工作中只有区区 5% 会被 AI 彻底取代。这一观点出自《Humanist Review of AI》杂志上发表的深度文章，作者正是麻省理工学院（MIT）教授、新晋诺贝尔经济学奖得主达龙·阿西莫格鲁（Daron Acemoglu）。 ```mermaid flowchart TD accTitle: 达龙·阿西莫格鲁的经济学扩散框架 accDescr: 展示技术创新到经济生产力转化的四个必要环节 subgraph TechDiffusion["技术扩散的四道门槛"] A["底层技术基础设施突破<br/>（如算力、基础模型）"] B["适配基础设施的应用落地<br/>（垂直工具与工作流）"] C["精准触达目标客群的营销与商业渠道"] D["终端用户反馈推动持续迭代"] A --> B --> C --> D end subgraph RealityGap["现实脱节阻力"] E["历史重演：企业丑闻与被巨头边缘化<br/>（如 Dragon Systems 被廉价倒手）"] F["企业方向偏离终端用户<br/>（转攻医疗/B端后创新停滞）"] end D -.->|"遭遇阻碍导致技术未能普惠"| RealityGap ``` 在文章中，阿西莫格鲁尖锐指出，当前全社会需要停止开发旨在“取代人类”的 AI，转而集中精力构建能够“赋能人类工作”的工具。民调数据显示，52% 的美国人对 AI 砸掉自己的饭碗心存焦虑。阿西莫格鲁认为这种恐慌严重被夸大了，而且正在错误地引导技术研发的路线。至于剩下的 48% 对饭碗胸有成竹的人，主持人们调侃道：“祝你好运！那些人大概心里在想‘我的工种根本不在训练数据集里，目前世界上也没有适配我真实工作场景的强化学习环境（RL environment）’。” 阿西莫格鲁援引了经济学中著名的“索洛悖论”（Solow computer paradox）——正如当年互联网席卷世界时，“计算机时代随处可见，唯独在生产率统计数据中缺席”，AI 如今同样尚未转化成统计学上的真实生产率激增。他预测未来 10 年 AI 对美国 GDP 的增益仅为大约 1.5%，根本谈不上革命。针对“电力技术当年也经历了数十年才完成全面渗透与经济转化”的比方，主持人们以昨日对 Ridge 钱包 CEO 肖恩·弗兰克（Sean Frank）的访谈提出反驳：这家成立十余年的公司今年实现了 50% 的逆势增长，其需求规划、创意素材制作等业务全链路均已深度嵌入 AI 工具，展现出惊人的生产力增益；尽管宏观经济学家很难将这种增长孤立地全归功于 AI，但由于今天人人都能接入互联网，AI 技术的扩散速度绝不可能像百年前铺设电网那样拖上数十年。 --- ### [13:42](https://www.youtube.com/watch?v=0SbWBq9G_l8&t=822s) 语音识别技术 30 年兴衰：Dragon Systems 的教训 面对反驳，阿西莫格鲁在文中搬出了另一个更具说服力且更令现代人感同身受的历史对照组——语音识别技术。将语音转换成文字的梦想从 20 世纪 30 年代起就已经存在，技术人员探索了近一个世纪。直到 1980 年代，詹姆斯·贝克和珍妮特·贝克（James and Janet Baker）夫妇创立了标志性的 Dragon Systems，语音识别才迎来决定性转折。贝克夫妇放弃了当时主流但笨拙的语法结构分析法，率先转向利用机器学习构建纯预测系统，将声音信号直接映射到最佳匹配文本。 1997 年，Dragon 推出划时代的 Dragon NaturallySpeaking。此前包括 Dragon Dictate 在内的软件都要求说话人每个单词之间生硬地停顿一下，而 NaturallySpeaking 第一次实现了对自然连续语音的直接捕捉。阿西莫格鲁当时因双臂患上严重的重复性劳损（RSI）而无法打字，成为该产品的早期重度用户。在 1997 年极端缓慢的处理器性能下，NaturallySpeaking 就已展现出近乎魔法般的 95% 识别准确率。然而，随后的两十年里，这场技术奇迹并没有像所有人预测的那样迅速统治消费级市场，反而因接二连三的企业管理灾难而陷入停滞： - **世纪并购与金融欺诈崩溃**：2000 年，比利时语音技术公司 Lernout & Hauspie 高调收购了 Dragon Systems，但数月之内这家比利时公司就因严重的财务造假丑闻宣告破产，联合创始人锒铛入狱。 - **业务重心转向企业垄断与医疗深井**：随后 ScanSoft 收购了 Dragon 并更名为 Nuance（其中一位主持人大学时创业的公司后来也被 Nuance 收购，对其技术赞不绝口）。Nuance 逐渐放弃了普通消费级市场，全力转向医疗和企业级领域；时至今日，在 Whisper 大模型开源前，欧美医院里大部分医生护士依然使用其系统做听写记录。 - **巨头吸收与消费级停滞**：微软于 2022 年正式收购 Nuance，继续沿用相同的闭门深耕策略。 阿西莫格鲁感叹，如今 PC 上的内置语音识别表现，相比 2000 年的水准仅仅是微幅提升，完全辜负了过去二十多年算力与算法的实际进展。他警示道：底层基础设施的突破，必须与构建在其之上的精准场景应用、健全的商业推广架构，以及持续的终端用户反馈循环相辅相成。当年的语音识别在这几点上彻底搞砸了，而今天的 AI 产业，同样没有理顺这套扩散机制。 --- ### [17:54](https://www.youtube.com/watch?v=0SbWBq9G_l8&t=1074s) 《60 分钟》现场对决：同一位诺奖得主的两副面孔？ 正当听众准备接受阿西莫格鲁的“审慎扩散论”时，主持人们话锋一转，将镜头切换到了近期他在 CBS《60 分钟》（60 Minutes）节目中的惊人表现。与杂志上安抚人心的“白丸”口吻截然相反，阿西莫格鲁在电视专访中展现出了极度悲观的“黑丸”（blackpilled）立场。 节目中，记者引述了 Mercer 创始人布伦丹·富迪（Brendan Foody）在领英上掀起的新职业浪潮——富迪坚称自己通过培训 AI 正在为下岗工人创造未来的职业道路，并放话称“全美前 20 位顶尖经济学家全部都会认同：十年后社会上的总就业岗位一定比今天更多”。随后记者转头将话筒递向作为全美顶级经济学家的阿西莫格鲁，询问他是否赞同这一主流共识。阿西莫格鲁干脆利落地吐出一个词：“不（No）。”电视剪辑效果拉满，充满了宿命般的荒谬讽刺感。 阿西莫格鲁在节目中严肃警告，人类当前正在经历的 AI 浪潮在历史上是绝无仅有的，其破坏性冲击堪比工业革命的初期阶段。面对他在纸媒宣称“AI 仅威胁 5% 人类工作”与在电视镜头前公然反驳“就业岗位增加论”的巨大反差，主持人们在演播室展开了激烈剖析：这两个看似精神分裂的立场其实存在内在逻辑闭环——如果科技界继续沿着“以自动化全面取代人类”的惯性方向堆叠技术，AI 的确会无情淘汰现有工人；但如果能悬崖勒马，将研发重点扭转到为劳动者提供增强赋能工具的轨道上，大失业的末日图景便并非不可避免。 ### [18:54](https://www.youtube.com/watch?v=0SbWBq9G_l8&t=1134s) AI就业争论：替代劳动力还是赋能工人？ 主持人展开讨论了一篇核心文章，其头条标题直指关键问题：“AI会取代工人吗？如果我们构建得当就不会（Will AI replace workers? Not if we build it right）”。文中的核心论点在于，如果按照目前纯粹替代人力的发展轨迹继续下去，技术确实会淘汰工人；但如果改变方向，将AI用于解决人类目前无法解决的全新问题，而非“把既有的工作搬来搬去（just moving stuff around versus doing net new things）”，那么从经济学角度看，不仅能保住就业，还能带来更高的生产率和经济增长。主持人调侃道，顶级经济学家“可从来没犯过错”。 随后，讨论转向了科技界普遍蔓延的紧迫感。许多人都在鼓吹“2026年将是最后一个正常年份，而2027年将是见证奇迹的疯狂一年”。自GPT-3发布以来的这五年间，公众积累了无数疑问——诸如递归自我改进（RSI）、超级智能的出现，以及宏观GDP层面的显著跃升，都被预期将在2027年给出答案。与此同时，金融市场出现了一些诡异迹象，比如黄金价格暴涨，虽然很难直接归咎于AI，但事后回看很可能会被视为AI繁荣的副产物。 关于技术对人类的影响，观众Jeremy提出了一个耐人寻味的观点：“农业对人类肉体造成的改变，AI也会对人类大脑重演一遍——让我们变得更高更壮”。主持人对此解读为偏向消极的比喻：农业固然让人类获得了更多卡路里、长得更高大，但也带来了过度摄入和肥胖；同理，如果在AI时代过度摄入信息，最终带来的只能是“大脑腐败（brain rot）”。 --- ### [21:06](https://www.youtube.com/watch?v=0SbWBq9G_l8&t=1266s) 沃伦·巴菲特的深夜YouTube“兔子洞”之旅 《华尔街日报》发表了一篇颇为吸睛的报道，揭秘传奇投资人沃伦·巴菲特（Warren Buffett）每晚沉迷YouTube的私人生活习惯，告诉普通观众“不用为每天刷几小时视频感到内疚，巴菲特也和你一样”。已经年满96岁并卸任伯克希尔·哈撒韦（Berkshire Hathaway）董事长职位的巴菲特，几乎每天晚上都会坐在奥马哈家中的躺椅上，顺着算法推荐在YouTube里漫无目的地闲逛。 巴菲特亲密的好友兼长期“观影搭子”莎伦·奥斯伯格（Sharon Osberg）透露，巴菲特的涉猎范围极其广泛甚至有些无厘头：从格伦·克洛斯（Glenn Close）的影视片段、1950年代的老广告、保罗·安卡（Paul Anka）的经典金曲，到完全无名之辈的演讲。他不仅为伯克希尔自己发布的内容在YouTube上的播放表现感到骄傲，还会因为某个业余歌手深情翻唱的《My Way》而感动不已，第二天一早跑到办公室四处安利。今年8月奥斯伯格造访巴菲特家时，两人连看了几个小时的视频，包括一部关于罗伊·科恩（Roy Cohn）的高水准纪录片，以及乌兹别克斯坦版本的《达人秀》（Uzbekistan's Got Talent）。 对于这篇爆料，巴菲特本人拒绝置评。主持人一边打趣“莎伦怎么能把朋友通宵刷视频的老底全抖给媒体”，一边推测这或许是巴菲特的“奥卡姆4D剃刀（Occam's 4D razor）”式心理战策略：故意将这种生活细节透露给《华尔街日报》，制造一种退休老人的松弛假象，让其他资本配置者放松警惕，而背地里他其实依然在“暗中磨刀（studying the blade）”。 - **巴菲特的YouTube常看内容**：1950年代广告、格伦·克洛斯片段、保罗·安卡视频、业余翻唱版《My Way》、罗伊·科恩纪录片、《乌兹别克斯坦达人秀》（Uzbekistan's Got Talent）。 - **观看习惯**：通常在奥马哈家中的躺椅上，与好友莎伦·奥斯伯格一同连续观看数小时。 --- ### [23:32](https://www.youtube.com/watch?v=0SbWBq9G_l8&t=1412s) 美国新版护照设计与排版争议 关于美国新版护照封面设计的争论在社交媒体上引发热议。设计界人士保罗·麦格雷戈（Paul McGregor）公开发难，直言“如果由我来重新设计这个国家最重要的公共文件封面，我首先会花重金雇佣全天下最好的排版师和国徽设计师，然后再去摆弄版面布局”。Airbnb联合创始人乔·杰比亚（Joe Gebbia）则迅速回击：“等你摆弄完了，请把顶级国徽设计师的简历发我一份，看看他们能不能设计得比1782年的官方国徽更好”。 主持人认为，目前的改版实际上是一个非常克制且适度的迭代。由于公众对各类视觉再造的预期已经十分浮夸，甚至有人脑补新护照会做成“金灿灿的土豪金风格，上面画着熊和鹰在UFC八角笼里激烈互殴”，相比之下实物显得相当低调。然而，新版文字排版确实存在明显硬伤——无衬线全大写字母的字距调整（kerning）过宽，给人一种廉价的“AI垃圾内容（AI slop）”既视感。护照不同于印完就扔的会议宣传册，它需要印发上亿本并在全球口岸被反复查验长达十年，对平面设计的水准要求极高。主持人更开玩笑表示自己根本不打算离开美国，不如直接废除护照。 --- ### [25:26](https://www.youtube.com/watch?v=0SbWBq9G_l8&t=1526s) Mistral Large 4“Le Chonk”发布与开源大模型竞赛 法国AI初创公司Mistral AI正式发布了其最新旗舰模型Mistral Large 4，内部昵称为“Le Chonk”（法语意为庞然大物）。在调侃Mistral到底算“前新实验室（pre-neolab）”还是欧洲国家级实验室的同时，主持人详细梳理了这款模型的关键技术规格与战略地位： ```mermaid flowchart TD accTitle: Mistral Large 4 核心架构与商业部署 accDescr: 展示Mistral Large 4的模型规模、优势领域及部署流程 subgraph Spec["模型架构 (Le Chonk)"] Total["1万亿总参数 (1 Trillion)"] Active["490亿活跃参数 (49B Active)"] Native["原生多模态能力 (Native Multimodal)"] end subgraph Capabilities["优势领域"] Vision["视觉定位超越顶尖闭源模型"] Critical["关键领域SOTA: 网络防御 / 金融 / 制造"] Legal["法规基准: Harvey法律Agent测试达15%"] end subgraph Deployment["交付路线"] API["即日起开放商业API"] CyberPrivate["网络安全伙伴私有合作"] OpenWeights["预计10月底发布开放权重 (可能经安全微调)"] end Spec --> Capabilities Capabilities --> Deployment ``` 现场讨论指出，Mistral Large 4是目前欧美表现最佳的开源权重模型，在综合基准测试上直逼开源前沿，足以与中国顶尖开源模型同台竞技。对于顾虑中国开源模型可能存在潜在法律与合规风险的企业来说，完全在欧洲本土训练并由Mistral自主云架构托管的Le Chonk，提供了一个绝佳的替代方案，彻底改写了开源与闭源的抗衡态势。 耐人寻味的是，业内人士Leo指出，Mistral Large 4唯一遥遥领先的单项指标竟是法律法规基准测试——在法律AI平台Harvey的Legal Agent基准测试中，该模型得分高达15%，而其他模型普遍在5%至13%之间徘徊。主持人调侃这不愧是“欧盟出品的王炸（EU banger）”，完全契合欧洲热衷监管的底色，同时也将极大提振Harvey等饱受算力成本和利润率压力的法律科技公司的业务前景。此外，美国开源阵营的Reflection也发布了名为“Beam”的新模型，显示开源AI生态正迎来新一轮密集爆发。 ### [28:38](https://www.youtube.com/watch?v=0SbWBq9G_l8&t=1718s) 模型价格战与开源模型的地缘追赶 许多依赖顶级前沿模型的企业之前经历了一段艰难时期，当时高度依赖成本最昂贵的前沿模型（frontier models），运营成本居高不下。然而如今局面发生了剧烈转变：一方面是 OpenAI 与 Anthropic 之间激烈的降价竞争，另一方面则是开源模型（open-weight / open source models）带来的巨大冲击。现在的打法已经变成了“设定你希望各大实验室去刷榜的基准（benchmark）”——企业只需针对自己的具体业务负载制定出测试基准，各大模型就会全力以赴去贴合并优化这套指标，从而让推理成本大幅压低。 ```mermaid flowchart LR accTitle: AI推理成本演进与竞争格局 accDescr: 展示从早期高昂成本到多方竞争大幅降价的演进路径 subgraph EarlyStage["早期困境"] A["高度依赖昂贵前沿模型"] --> B["企业计算成本高昂"] end subgraph Pressure["多方竞争加剧"] C["OpenAI 与 Anthropic 价格战"] D["开源模型逼近 (Mistral / 中国开源)"] end subgraph CurrentState["当前格局"] E["针对特定工作流定制 Benchmark"] F["企业推理成本大幅下降"] end B --> Pressure C --> CurrentState D --> CurrentState ``` 直播间观众“Sushi commander”发弹幕提出异议，指出无论是 Mistral 旗下的模型（如 Mistral Large）相比中国最新的顶尖模型都要落后一代，并要求主播 Tyler 面对现实。主播们坦言，从公开基准测试来看，这些欧洲开源模型确实算不上绝对领跑的前沿模型，更像是多方角逐中的主要竞争者之一。更残酷的现实在于迭代周期的差距：几个月前中国开源阵营刚发布了最新一轮模型，常常呈现出西方团队刚费尽心力追赶上前沿水平、还需要再花六个月才能推出下一代，而对方可能只要两个月就能迭代出新成果的节奏。不过，各家企业最终选择站队哪一方模型，往往还掺杂着极其复杂的地缘政治因素。 ### [29:45](https://www.youtube.com/watch?v=0SbWBq9G_l8&t=1785s) 消费级 AI 三足鼎立：ChatGPT、Claude 与 Gemini 的规模之战 根据 a16z 投资人 Olivia Moore 整理的最新行业数据，ChatGPT、Claude 与 Google Gemini 在争夺消费级 AI 统治地位的战场上展现出鲜明的梯队分化。该项统计整合了网页端流量、移动端流量以及消费级软件收入： - **网页端规模（Web Scale）**：ChatGPT 的规模达到了 Gemini 的 2 倍，同时相当于 Claude 规模的整整 6 倍。 - **移动端规模（Mobile Scale）**：ChatGPT 在移动端展现出更恐怖的消费级统治力，规模是 Gemini 的 2.5 倍，更是达到了 Claude 的 14 倍。 讨论中主播指出，ChatGPT 拥有更纯粹的消费级产品基因，因此在移动端吸纳了海量大众流量。但与此同时，Claude 近期的追赶势头同样不可小觑。此前在应用商店的排行榜上，Claude 曾一度冲至榜首，随后滑落，而 ChatGPT 重新夺回第一。这种榜单拉锯战宛如前沿科技领域的阵地战，双方不断“互换火力”，呈现出你来我往的拉锯态势。 节目的最后，主播们提及了当前应用商店排名的有趣景象——二手交易平台 Vinted 依旧极其强势地稳居前列（目前排名第二），并对其团队的表现赞赏有加。随后主播向观众致谢，提醒大家在 Apple Podcasts 与 Spotify 上留下五星好评、订阅 tbpn.com 的每日邮件通讯，并相约下一个工作日上午 11 点（太平洋时间）在 TBPN 直播间继续收看。 ## Notable quotes > [00:01](https://www.youtube.com/watch?v=0SbWBq9G_l8&t=1s) Being a Global Superstar for Dummies > [10:42](https://www.youtube.com/watch?v=0SbWBq9G_l8&t=642s) 计算机时代随处可见，唯独在生产率统计数据中缺席 — 达龙·阿西莫格鲁 > [17:54](https://www.youtube.com/watch?v=0SbWBq9G_l8&t=1074s) 全美前 20 位顶尖经济学家全部都会认同：十年后社会上的总就业岗位一定比今天更多 — 布伦丹·富迪 > [17:54](https://www.youtube.com/watch?v=0SbWBq9G_l8&t=1074s) No. — 达龙·阿西莫格鲁 > [18:54](https://www.youtube.com/watch?v=0SbWBq9G_l8&t=1134s) Will AI replace workers? Not if we build it right > [18:54](https://www.youtube.com/watch?v=0SbWBq9G_l8&t=1134s) just moving stuff around versus doing net new things > [18:54](https://www.youtube.com/watch?v=0SbWBq9G_l8&t=1134s) 农业对人类肉体造成的改变，AI也会对人类大脑重演一遍——让我们变得更高更壮 — Jeremy > [23:32](https://www.youtube.com/watch?v=0SbWBq9G_l8&t=1412s) 如果由我来重新设计这个国家最重要的公共文件封面，我首先会花重金雇佣全天下最好的排版师和国徽设计师，然后再去摆弄版面布局 — 保罗·麦格雷戈 > [23:32](https://www.youtube.com/watch?v=0SbWBq9G_l8&t=1412s) 等你摆弄完了，请把顶级国徽设计师的简历发我一份，看看他们能不能设计得比1782年的官方国徽更好 — 乔·杰比亚 ## Mentioned TBPN, John Coogan, Jordi Hays, Tyler, Nick, Tom Cruise, Glen Powell, GQ, Kevin Roose, Dario Amodei, Eric Newcomer, Founders Fund, Marc Andreessen, Ben Thompson, Stratechery, Halo: Combat Evolved, Resident Evil, Call of Duty, Skate, The Elder Scrolls V: Skyrim, Spider-Man: Miles Morales, Minecraft, Xbox, Google Stadia, Meta Quest, GPT-4o, Claude 3.5 Sonnet, Daron Acemoglu, Mustafa Suleyman, Microsoft, Sean Frank, Ridge, James Baker, Janet Baker, Dragon Systems, Dragon NaturallySpeaking, Lernout & Hauspie, ScanSoft, Nuance, Whisper, Brendan Foody, Mercer, 60 Minutes, Warren Buffett, Berkshire Hathaway, Sharon Osberg, Glenn Close, Paul Anka, Roy Cohn, Uzbekistan's Got Talent, The Wall Street Journal, Paul McGregor, Joe Gebbia, Airbnb, Mistral AI, Mistral Large 4, Harvey, Reflection, Beam, OpenAI, Anthropic, Olivia Moore, a16z, ChatGPT, Claude, Gemini, Vinted ## 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 Tom Cruise’s Secret Six-Hour Acting Course 04:10 AI Game Mods, Halo in a Browser + The Copyright Fight 10:42 Will AI Really Replace Just 5% of Human Work? 17:54 The AI Jobs Debate: Replace Workers or Empower Them? 21:06 Warren Buffett’s Nightly YouTube Rabbit Holes 23:32 The U.S. Passport Redesign + The Typography Fight 25:26 Mistral Large 4 + The Open-Weight AI Race 29:47 ChatGPT vs. Claude vs. Gemini: Who’s Winning Consumer AI? [Source](https://www.youtube.com/watch?v=0SbWBq9G_l8)

[0:01] Being a Global Superstar for Dummies [10:42] 达龙·阿西莫格鲁: 计算机时代随处可见，唯独在生产率统计数据中缺席 [17:54] 布伦丹·富迪: 全美前 20 位顶尖经济学家全部都会认同：十年后社会上的总就业岗位一定比今天更多 [17:54] 达龙·阿西莫格鲁: No. [18:54] Will AI replace workers? Not if we build it right [18:54] just moving stuff around versus doing net new things [18:54] Jeremy: 农业对人类肉体造成的改变，AI也会对人类大脑重演一遍——让我们变得更高更壮 [23:32] 保罗·麦格雷戈: 如果由我来重新设计这个国家最重要的公共文件封面，我首先会花重金雇佣全天下最好的排版师和国徽设计师，然后再去摆弄版面布局 [23:32] 乔·杰比亚: 等你摆弄完了，请把顶级国徽设计师的简历发我一份，看看他们能不能设计得比1782年的官方国徽更好

[0:00] Tom Cruise: I listen first. What do they want? It's not me. I'm more interested in what are their goals? What do they want? [0:00] TBPN hosts: being a global superstar for dummies [10:42] Robert Solow: you can see the technology everywhere except in the productivity statistics [17:54] Daron Acemoglu: No. [18:54] Will AI replace workers? Not if we build it right [18:54] Jeremy: What agriculture did to our bodies, AI will do to our brains. [21:06] TBPN hosts: studying the blade every single night [29:47] TBPN hosts: war for the Frontier [29:47] TBPN hosts: trading shots back and forth

[0:01] I was leafing through GQ this morning. [0:03] Tom Cruz is on the cover. It's It's a [0:05] It's a good addition. Uh there's a funny [0:08] uh anecdote in here that I wanted to [0:10] share. I need I got to find it. Um but [0:13] they got the whole Tom Cruz spread, lots [0:16] of photos. Uh but they did a little back [0:18] and forth interview with him and he's a [0:21] course bro. He dropped a course, a [0:23] private course and we got to get access [0:25] to this. The question from the GQ [0:27] interviewer, you have mentored a lot of [0:30] younger actors. Glenn Pal has talked [0:32] about watching a sixhour instructional [0:35] video about screen acting that you made, [0:38] right? And Tom Cruz says it was very [0:41] basic. It was just something to have [0:42] people understand, just to have a [0:44] different concept of lenses and what the [0:47] lens is and what the frame is just the [0:50] same way as an [0:51] >> being a global superstar for dummies. [0:53] >> Yeah. Basically, he made a he made a [0:55] six-hour that's a course. He made a [0:57] six-hour video just the same way as an [0:59] actor on stage or in the development of [1:01] cinema where it started and what the [1:03] that frame means and how it auditory [1:12] of structure but also to understand [1:14] certain things so that you can forget [1:17] about it. Uh what else do you tell [1:19] younger people, other people in your [1:21] business? I listen first. What do they [1:23] want? It's not me. I'm more interested [1:25] in what are their goals? What do they [1:27] want? And it's not just younger people. [1:29] This is me when I was a kid. I'm going, [1:32] "What are you interested in?" They say, [1:34] "Really?" Said, "Yes. What is this movie [1:36] about? What is this?" Um, there's a [1:38] bunch of other fun things in here. But I [1:40] thought it was funny that there's a [1:41] secret six-hour training video on how to [1:43] be an actor from Tom Cruz. We got to get [1:45] access to it. I feel like uh there's a [1:48] lot of journalists that they get [1:50] documents, they scoop things, they [1:52] release things. Uh but I've never seen [1:54] anyone just be like, I'm putting down 10 [1:56] racks for whoever can bring me the [1:58] six-hour Tom Cruz video. Yeah, [2:00] >> I'm putting it behind a pay wall. I'm [2:01] going to monetize it, but there's a [2:04] bounty out. Like uh Kevin Roose has the [2:06] Darham uh big blob of compute thesis [2:09] from 2017 like this leaked PDF sort of [2:12] on the scaling law and the bitter lesson [2:14] basically. Uh, and I've never seen a [2:17] journalist just be like, "I know this [2:18] document's out there." Like Eric [2:20] Newcomer should just be like, "Bring me [2:22] founders fund returns. [2:25] I I'll I'll give you five grand if you [2:27] got them." You know, because it probably [2:28] maths out at some point. [2:29] >> Five million, [2:30] >> maybe more. Maybe more. Yeah. [2:32] >> To risk to risk your allocation in the [2:34] next [2:35] >> Yeah. I I suppose that that would uh [2:36] violate journalistic ethics to just put [2:38] a bounty out for a document that you [2:40] know exists. [2:41] >> I will bribe you. [2:42] >> Yeah. [2:42] >> For the truth. And it probably really [2:45] like I was thinking about the the the [2:47] the various leaks that go out like some [2:48] of them violate NDAs, but if they're if [2:51] they're secretive enough and they're [2:53] squirreled away uh via signal, like they [2:56] probably don't trace back. They probably [2:58] don't really trigger anything. Also, [3:00] some of them just don't really matter. [3:02] They will, you know, company is might [3:04] not be that frustrated about a [3:05] particular leak, uh if it makes the [3:07] company look good. But [3:08] >> well, speaking speaking of movies, the [3:10] team saw a movie yesterday. Okay, Ben, [3:12] could we get a review? Oh, you didn't [3:14] go. [3:14] >> Ben didn't Ben didn't come. Yeah. [3:15] >> Wow. Tyler review. [3:16] >> I saw it. So, we saw Nick said that he [3:19] was starting to look at his phone after [3:21] like 40 minutes. [3:22] >> Wait, what? [3:22] >> Nick did not enjoy it. I thought it was [3:24] pretty good. And so, so basically the [3:25] reviews are really bad, right? And [3:27] famously, well, not famously, but like [3:28] the box office is really bad. [3:31] >> Not all the reviews are bad. I saw Mark [3:32] Anderson gave it a positive review. [3:33] >> He did. Yeah. And that was really one of [3:35] the only like very positive reviews I [3:36] saw. But yeah, I think the budget was [3:37] like 160 million. First weekend it was [3:40] million. [3:40] >> 8 million. Yeah. pretty brutal. Uh, I [3:42] thought it was it was good. It was very [3:44] cool. Um, like the like the framing of [3:47] the movie. Um, [3:49] >> I it's it's hard to like reveal too much [3:52] you. I thought it was very film school [3:54] like this. This critique is so nuanced. [3:57] It's remarkable. [3:58] >> I thought it was [3:58] >> I don't want to spoil anything. I don't [4:00] want to spoil anything. [4:02] >> What was cool about it? [4:03] >> Just give us like 30 seconds of of [4:05] spoilers. I'll bleep it out. Okay. So [4:08] basically [4:09] >> the bleep organ the game decompilation [4:12] thing is is going strong. Uh Ben [4:14] Thompson was writing about it in stratey [4:15] today. Uh there were some more knock on [4:17] effects. He took it in a very [4:18] interesting like legal direction about [4:20] the the implications of uh taking a [4:23] video game uh decompiling it doing [4:25] whatever you want with it. Obviously IP [4:27] licensers don't like that but there's [4:30] not a lot that you can that you can do [4:32] to control that. And obviously open [4:34] source models are coming and there's a [4:35] bunch of interesting knock-on effects. I [4:37] saw a really funny one. They're getting [4:38] funnier, these mashups. I saw one where [4:40] someone was playing NHL hockey in [4:43] Resident Evil. Did you see this one? It [4:46] was really It was really such a good [4:48] mashup cuz the guys just like on ice [4:51] skates and like flicking pucks and [4:53] zombies. It was It was pretty good. Uh I [4:56] was I was into it. Again, probably not [4:57] really something I want to play for 5 [4:59] hours, but will I watch a 30 secondond [5:01] video of it? Absolutely. But uh yeah, [5:03] Polygon has been breaking this down. And [5:05] if you've been on X over the last few [5:06] days, you probably noticed that things [5:08] have gotten very weird when it comes to [5:09] modded video games. There are clips of [5:11] Call of Duty operators doing kick flips [5:13] and other tricks from Skate, people [5:14] playing Black Ops 1 zombie mode in [5:16] Skyrim, Miles Morales swinging through [5:19] Gotham. Here's another to add to the [5:20] list. Someone has gotten the Xbox [5:22] version of Halo Combat Evolved running [5:25] in a web browser with multiplayer and [5:27] split screen support. And the other [5:29] funny thing is that it's not they didn't [5:32] like apparently the Halo community [5:34] doesn't like the PC release of Halo. So [5:37] they ported the Xbox version to the [5:40] cloud browser. It's like you you can [5:42] really do anything you want now. And so [5:44] Ben Thompson goes on to describe a [5:46] little bit of what's going on. Games are [5:47] written in human readable languages, C, [5:49] C++. Obviously they're scripting [5:50] languages on top of that, but then [5:52] they're compiled into machine code [5:53] that's executable by the processor that [5:55] they're intended for. For Halo, that's [5:57] x86. You've always been able to [5:59] disassemble machine code into assembly [6:02] because assembly maps almost perfectly [6:04] to machine code. The difficult part, [6:06] which takes humans years, is decompiling [6:08] assembly to C or some human readable [6:12] language that then you can actually [6:13] modify. But LLMs are really good at [6:15] that. And so, is it legal? Is it legal? [6:18] Uh, the story conflates two different AI [6:20] capabilities. The Fallout project [6:22] appears to be a new game. There's no [6:23] Fallout New York. Although using the [6:25] name Fallout is likely a trademark [6:26] violation. I think we were joking about [6:28] uh the lawyers being really excited [6:30] seeing all these going out because it [6:31] means full employment. Decompiling is [6:34] not in and of itself illegal. Saga Sega [6:37] versus Accolade held that disassembling [6:40] code is fair use if it is the only way [6:43] to reach unprotected functional elements [6:46] and there was a legitimate purpose. and [6:49] Sony verse connectics extended the same [6:52] reasoning to reverse engineering the [6:54] PlayStation's BIOS. However, that is [6:57] likely where protections end. Code that [6:59] is derived from decompiling is by [7:01] definition very similar if not identical [7:03] to the original code which is covered by [7:05] copyright. So even if you go and [7:06] decompile a game and then you ship [7:08] something that's like a mod, you're [7:09] going to get a DMCA notice. And this is [7:11] already happening. Like something will [7:13] go up, it'll get it'll go viral and then [7:15] they'll get hit with a a takedown [7:16] notice. This already happened. later. [7:19] >> Is that me or is that [7:22] >> getting hit with a copyright? [7:23] >> Two hours later, you'll get hit by Yeah. [7:25] copyright. Uh so you'll have to take it [7:27] down. Um there's there's a couple [7:28] interesting things. So the actual value [7:30] of these like decompiled games, mashups, [7:34] pretty limited because you basically [7:35] just get a viral video and you can't [7:37] really monetize it because you get the [7:39] takedown notice. But uh the the the the [7:42] training roll out for the AI labs like [7:45] the training data that comes from like [7:46] having this fleet of like all these [7:48] different game modders go and spend a [7:50] bunch of tokens that they're basically [7:52] paying for maybe it's a little [7:53] subsidized but like they're basically [7:54] paying for all this decompilation and [7:57] then also the labs probably have RL [7:58] environments for this type of work and [8:00] then there's tools for this as well [8:02] means that I would expect like the next [8:05] version of models that is trained on all [8:09] the infra that's happening right now [8:11] because everyone right now in the game [8:13] modding community is like, "Let me go [8:15] point GPT Astra and Opus 55 at whatever [8:19] game binaries I can get." So, it's like, [8:21] "Oh, I I like Skyrim. I like Minecraft. [8:24] Let's go decompile those." But then [8:26] they're also doing it with all the [8:27] longtail. So, all of that's going into [8:29] the next set of training data. So, you [8:31] would imagine that the models get much [8:33] better at video game development, which [8:35] is uh which is cool. Uh he also just [8:37] talks about like how things will change [8:39] over time similar to uh similar to [8:41] what's happened in video AI seems to [8:43] exacerbate a lot of trends including the [8:46] shift to service-based games. So he [8:48] talks about like in a world where if you [8:50] send someone a disc or you send someone [8:53] a executable binary like they're just [8:55] going to decompile it and be able to do [8:56] whatever they want with it. Maybe you [8:58] want to stream that game. So you uh this [9:01] was uh Google had an effort for this. I [9:04] forget what that was called, but Google [9:06] had the ability to basically stream you [9:09] a 1080p [9:10] >> Google Stadia. [9:11] >> Stadia, that's right. So you they would [9:13] stream you the video output and you [9:15] would stream them the controller input [9:18] and the game would run on the server. So [9:20] you never saw any code. You just saw [9:22] video coming back. [9:22] >> This is also how the uh MetaQuest Xbox [9:25] edition worked. You you were streaming [9:26] the game in. [9:27] >> That's right. That's right. And and I [9:28] think uh I think Xbox has has stayed [9:30] with the cloud gaming a little bit [9:32] longer, but gamers have always sort of [9:34] pushed back on that. I don't like the [9:35] latency. What if my Wi-Fi goes out? I'm [9:37] just trying playing a single player [9:38] game, blah blah blah. And so, uh it's [9:40] possible that this is something where [9:42] the game industry comes together and [9:43] they're like, we really got to fight [9:45] these hackers. We got to hide our code [9:48] more behind more of a firewall, like [9:50] that cloud streaming firewall. Uh or it [9:53] could be something that they just like [9:54] embrace and go the the IP protection [9:57] route. Uh yeah, I it's hard to to think [10:00] that uh this doesn't go the way of [10:03] generated text where like people still [10:06] care about the the the posts and where [10:09] they come from. Effectively, [10:11] intellectual property, the instantiation [10:13] of the actual product is less relevant [10:15] than the story around it and the and the [10:17] people behind it and the and the [10:19] shelling point of like we're playing the [10:20] same game. Of course, people will do [10:21] their individual stuff as well, but [10:23] anyway, [10:24] >> what else we have going on, John? Well, [10:27] I am so glad to be past that subject [10:31] >> because you just don't like video games. [10:33] >> Uh, I had enough. I had enough. [10:35] >> You had enough? [10:36] >> I had enough. [10:36] >> What about racing days a video game? [10:39] >> That's the final frontier. [10:40] >> Yeah, you're going to be able to race [10:41] anywhere you want. [10:42] >> Mustafa says, uh, coming in with a major [10:45] white pill. AI won't take your job [10:47] anytime soon. In 10 years, only 5% of [10:50] what humans do will be replaced by AI. [10:54] Let's see what he has to say about that. [10:55] They're doing new media over at MI, the [10:58] humanist review of AI. They made a [11:00] magazine and they have a number of [11:02] articles and uh they tapped MIT [11:08] economist and Nobel laurate, Nobel Prize [11:10] winner in economics Darren Isimoglu to [11:14] write a piece about how AI will diffuse. [11:17] And it's interesting because in this one [11:19] it's total white pill but then he was on [11:21] 60 minutes and he's and he's a little [11:22] bit more blackpilled. So, we'll have to [11:25] go back and forth. I think it's the [11:26] same. [11:27] >> Wait, that's the same guy. [11:28] >> Yeah, we'll have to reconcile this. [11:29] Okay, so let's see what Daron Aimogu [11:33] says in the humanist uh review. It's a [11:37] longer piece, but Mustafa Sullean uh [11:40] summarized it here. He said he argues [11:42] that we need to stop building AI to [11:43] replace people and start building it to [11:46] make them better at their jobs. 52% of [11:48] Americans are worried about AI's impact [11:50] on their jobs. The fear is overblown and [11:53] it's steering how we build AI. I like [11:56] the 48% of just confident people. Like, [11:59] good luck. [12:00] >> Good luck. I'd like to see you. I'd like [12:02] to see the work I do. [12:03] >> Yeah. I'm not in the training data. I'm [12:06] not There's no Ral environment for what [12:08] I do. [12:09] >> Um AI isn't in the productivity [12:11] statistics yet. This was the famous line [12:13] about the internet. It shows up [12:14] everywhere except the productivity [12:16] statistics. Most firms using AI aren't [12:18] seeing real gains yet. expect roughly [12:21] 1.5% added to GDP over 10 years. Not a [12:25] revolution. Electricity took decades to [12:27] spread. There's so [12:29] >> problem the problem is like so so the [12:31] problem is like going back to our [12:32] conversation uh with Sean Frank [12:33] yesterday from Ridge. [12:35] >> This is a company that's over 10 years [12:37] old right now growing 50% this year [12:40] >> using AI across the entire business from [12:44] demand planning to creative to a number [12:48] of other things. And the business is [12:50] performing pretty much as well as it [12:52] ever has. The thing is you AI is [12:55] impacting the whole business in [12:57] different ways [12:59] >> in big ways and small ways but you can't [13:01] give AI credit for the growth because [13:04] there's so many factors contributing [13:05] right new products push into retail all [13:08] these different things. And I think you [13:09] can still say that maybe most firms [13:11] aren't using it to see see real gains. [13:13] But when you say that electricity took [13:16] decades to spread, I just don't view AI [13:20] as it's not the kind of technology that [13:23] that takes a long time to spread when it [13:25] works. [13:26] >> And that's because we can use the [13:29] internet, which pretty much everybody [13:31] that's productive has access to to [13:34] spread AI. Yeah. [13:35] >> So, um already uh just a few bullet [13:38] points in and and uh [13:40] >> but the Nobel Prize winning economist [13:42] has another example to push back on you [13:44] with. He says if you don't like the [13:46] electricity example, here's another one. [13:49] Speech recognition. He said turning [13:51] speech into text has been an aspiration [13:54] since the 1930s. And this is a harder [13:56] one to grapple with than the than [13:57] electricity because a lot of what was in [14:00] place for AI is also in place what uh a [14:03] lot of the things that you're saying [14:05] about why diffusion should be faster for [14:06] AI was true for uh for speech [14:10] recognition. So uh people have been [14:12] trying to do speech recognition you talk [14:14] the computer types since the 1930s going [14:17] on a hundred years of pursuing this. It [14:19] became a huge step closer to reality [14:22] thanks to the work by American by by the [14:25] American husband and wife team James and [14:28] Janet Baker starting in the 1980s when [14:31] they founded the iconic company Dragon [14:34] Systems. Dragon truly revolutionized [14:38] voice recognition and it would not be an [14:40] exaggeration to say that everything else [14:43] that has followed in voice recognition [14:44] since then is thanks to the Bakers. [14:46] Their breakthroughs, their breakthrough [14:48] was to assue previous attempts to build [14:51] speech recognition systems based on [14:53] syntax and meaning and instead deploy [14:56] machine learning techniques in order to [14:58] build a pure prediction system which [15:00] would map speech onto the best match [15:03] text. Already in 1997, the year Dragon [15:06] launched Dragon Naturally Speaking, the [15:08] first program that was capable of voice [15:10] recognition from natural continuous [15:12] speech, the mo the momentousness of this [15:14] was obvious. So, in previous software, [15:17] uh, including the previous product from [15:18] Dragon, Dragon Dictate, you had to pause [15:22] in between each word to let it process [15:24] the gap between words. You no longer [15:27] needed to do that with Dragon, naturally [15:28] speaking. This was 1997. I can vouch for [15:31] all of this because uh Darren Esmoglu, [15:33] the writer, uh, after developing [15:35] repetitive strain injury in both arms in [15:38] 97, he was an early user, even with the [15:42] very slow processors of the time. Dragon [15:44] naturally speaking was like magic. It [15:46] would score close to 95% accuracy. I [15:49] didn't know at the time that I would be [15:51] a user of their products for the next 30 [15:53] years. But what I immediately knew was [15:56] that the processing power of computers [15:58] improved. I could have access to much [16:00] more powerful voice recognition [16:01] software, enabling me to carry out my [16:03] work despite a continued inability to [16:06] type. By 2000, everything was going as I [16:09] and others had predicted. The software [16:10] had gotten better and with better [16:12] computer chips, accur accuracy and [16:14] functionality were improving at a fast [16:15] pace. What nobody could predict was that [16:18] in the next two decades, despite [16:19] advances in the underlying technology, [16:20] consumers would see little of the gains [16:22] thanks to a series of corporate [16:24] disasters. Dragon was acquired in 2000 [16:26] by a Belgian company, Learnout and [16:28] Housepie, uh that went bankrupt within [16:30] months of the purchase with its [16:32] co-founders later being found guilty of [16:34] corporate fraud. Uh you hate to see [16:37] that. [16:38] >> Naughty, naughty. Uh so Dragon was then [16:42] purchased by Scansoft, rebranded as [16:43] Nuance. I actually worked for a startup [16:45] in college that got acquired by Nuance. [16:47] Uh very very talented team over there. [16:50] Uh which progressively deprioritized the [16:52] consumer in favor of medical and [16:54] enterprise markets. If you walk into a [16:56] hospital today, you will see a lot of a [16:57] lot of doctors and nurses dictating [17:00] notes. uh even before the uh whisper uh [17:03] revolution that just recently happened, [17:05] Microsoft acquired Nuance in 2022 and [17:08] continued the same model. The result, [17:10] voice recognition software on PCs today [17:12] is only marginally better than what was [17:13] available in 2000. Far less progress [17:15] than the underlying technology would [17:17] have led me to expect. This [17:19] disappointing history lesson is one of [17:20] the reasons why even greater curation is [17:22] warranted when making predictions about [17:24] the fast diffusion of AI models. Today [17:26] breakthroughs in the infrastructure of a [17:28] new technology need to go handinhand [17:30] with the right kinds of applications [17:31] built on this infrastructure. And then [17:33] of course you need the right marketing [17:34] structure to reach potential customer [17:36] base as well as inputs from consumers to [17:38] act as the impetus for modifications and [17:40] improvements of the applications. This [17:42] is what we are not seeing for AI. This [17:44] is what was completely bungled for voice [17:47] recognition. What do you think? [17:52] >> Whole lot of mumbo jumbo John. Okay, [17:54] let's move over to 60 minutes because [17:56] Darren Simoglu takes the other side of [17:58] his own argument, arguing against [18:00] himself with Brendan Foody from Mccur [18:02] >> job category on LinkedIn. And just as [18:04] machinists once replaced artisans, [18:06] Brendan Foody insists he's creating a [18:09] career of the future. Out of work, fear [18:11] not, come train AI. [18:13] >> Literally all of the top 20 economists [18:15] in America, they would say the exact [18:18] same thing, which is that there's going [18:20] to be more jobs in 10 years than there [18:22] are today. You're a top economist in [18:24] America. What say you? [18:25] >> No. [18:26] >> One of the greatest edits. I [18:28] >> don't how that could happen. [18:29] >> What we are living through with AI is [18:32] unprecedented. The first phase of the [18:35] industrial revolution. [18:36] >> You're positive this is the same guy in [18:38] this article that we read. [18:39] >> Wait. I I I don't think these artic like [18:41] these two things are incompatible. [18:43] >> Break it down for us. [18:44] >> So the article [18:45] >> because he's saying 5% 5% of the things [18:48] humans do are at risk. No, he's saying [18:50] that right now if we keep building AI [18:52] the same way it will replace workers but [18:54] we need to build it a different way that [18:56] like empowers workers. That's what the [18:57] artic like the headline of the article [18:58] is will AI replace workers not if we [19:00] build it right. He's saying like right [19:02] now on the course that we're doing right [19:03] now it's going to replace workers. Got [19:05] it. So we need to shift courses. [19:06] >> Okay. [19:07] >> Right. I that's what I [19:09] >> and and and his argument from an [19:10] economic perspective is like that will [19:12] actually increase productivity that will [19:14] be better not just because higher [19:16] employment is good but also higher [19:19] economic growth is good. So we should be [19:21] trying to solve problems that cannot be [19:23] solved by people right now with AI as [19:26] opposed to just [19:26] >> just displacing. Yeah. Just moving stuff [19:28] around versus doing net new thing. [19:30] >> Okay. I like that. [19:30] >> I believe that's what he's position [19:32] saying. [19:32] >> You're a top economist. What do you say? [19:34] No. [19:36] one of the greatest. [19:37] >> And here's the thing. Yeah, [19:38] >> top economists have never been wrong on [19:40] anything. So, you know, [19:43] >> it does feel like it's a critical [19:45] moment. There's been so many people [19:46] building up to like 2027's going to be [19:48] the crazy year. 2026 is going to be the [19:50] last normal year. like like something [19:53] has to happen in 2027 to that will [19:56] answer a lot of the questions that have [19:57] been being asked over the last 5 years, [20:00] you know, since GPT3 basically uh which [20:03] is like RSI super intelligence serious [20:07] serious move on GDP serious like showing [20:09] up in the statistics somewhere other [20:11] than just debt but I mean at the same [20:13] time you see like weird things happening [20:15] in uh in financial markets like gold is [20:18] spiking and like it's hard to put that [20:20] on AI but there's still like like [20:22] there's there there's pockets of like [20:25] okay there's something odd happening [20:26] that you might in retrospect attribute [20:29] to the AI boob. Anyway, what do you [20:31] think of Jeremy's take? What agriculture [20:33] did to our bodies AI will do to our [20:36] brains, make us stronger and taller. I [20:39] think he's being negative. [20:40] >> I think he's being negative. Didn't it [20:42] in some way make us like, you know, [20:44] weaker because we're, you know, [20:46] consuming more grains versus, you know, [20:48] >> Yeah. I mean, some agriculture made us a [20:51] lot taller and stronger, but then made [20:54] us fatter for sure. Yeah, because we [20:56] over consumed. And so, yeah, if you're [20:58] over consuming information, it's brain [21:00] rot. Yeah, but a good uh good [21:02] encapsulation of the of the thesis. [21:05] Anyway, where you want to go next, [21:06] Jordy? [21:06] >> Uh, we got to we we got to talk about [21:08] Warren Buffett because he's just like [21:10] you. [21:11] >> He's just like you, folks. A viewer's [21:13] guide to Warren Buffett's nightly trips [21:15] down the YouTube rabbit hole. Uh Glenn [21:18] Close clips 1950s commercials. [21:20] Usuzbekistan's version of America's Got [21:23] Talent. The legendary investor can't [21:26] resist the algorithm either. Feeling [21:28] guilty about spending hours each day [21:30] watching YouTube. Don't be. Warren [21:32] Buffett is right there with you. Glenn [21:34] Close clips commercials from the 1950s. [21:37] Paul Anka videos. The famed investor [21:39] can't get enough. Buffett enjoys [21:40] speeches by complete unknowns. He's a [21:43] fan of Birkshshire Hathway's own content [21:45] too. proud of how it performs on the [21:47] video platform. He watches it all nearly [21:49] every evening, usually from a recliner [21:52] in his Omaha home. Warren's range is [21:55] broad, says Sharon Osberg, a close [21:57] friend and frequent viewing companion. [21:59] It can be the silliest things. Buffett [22:02] also goes down random rabbit holes just [22:04] like us. The next day, he'll come into [22:05] his office gushing over a moving amateur [22:08] rendition of My Way. In August, Osberg [22:11] visited Buffett's home in Omaha. They [22:13] spent hours on YouTube grinding. He We [22:16] watched a really good documentary on Roy [22:18] Conn, she says, and the Beckistan [22:20] version of America's Got Talent. That [22:23] would be O's Beckistan Got Talent. It's [22:28] easy to remember. Buffett, who recently [22:30] turned 96 and stepped down as [22:32] Birkshshire's chairman, declined to [22:34] comment on his YouTube habits. What's up [22:36] with Sharon just blowing up blowing up [22:39] his whole routine? If I invite a friend [22:41] to my home, [22:42] >> I'll go do son. [22:43] >> If I invite a friend to my home to watch [22:45] YouTube videos with me for hours and [22:47] hours and hours and go down the YouTube [22:49] rabbit hole, and then a week later, [22:51] you're calling up the journal, spilling [22:54] spilling the truth about my my video [22:58] consump my rabbit hole habits. [23:00] >> Yeah. [23:01] >> I feel like I feel like Buffett had to [23:03] be like, I want you to tell the journal. [23:05] >> You think it's astroturf? [23:06] >> Yeah. I think I think I want you to take [23:08] this. [23:08] >> You think he's paying attention? He's [23:10] trying to [23:11] >> I think he's as locked in as ever. I [23:13] don't think he's doing anything for [23:14] entertainment. He wants the other [23:17] allocators to get comfortable. You know, [23:19] I can spend a few hours on YouTube every [23:21] night. Meanwhile, he's studying the [23:24] blade every single night and all of this [23:25] >> giving the benefit of 4D chess. [23:27] >> Yes. Aam's 4D razor. [23:29] >> Yeah, that is AAM's 40 razor. This is [23:31] what he's doing. Makes sense. What do [23:32] you think of the new Passport? People [23:35] were going back and forth on it. Paul [23:36] McGregor was saying, "If I was [23:37] redesigning the cover of what is [23:39] arguably the country's most important [23:40] public document, I'd first hire both the [23:43] best typographer and crest designer [23:46] money can buy, then fiddle about with [23:48] the layouts, probably." Um, and Joe [23:50] Gabia fires back high. When you're done [23:52] finding, when you're done fiddling, [23:53] please send me resume of top crest [23:55] designer. Make sure they can do better [23:56] than the 1782 official national anthem. [23:59] It looks like a very national emblem. [24:01] >> National emblem. Um, yeah, the redesign [24:04] looks like a very tasteful and moderate [24:08] iteration. Like I think people are [24:09] reacting to this like it's like now it's [24:11] all gold and it has like a bear and an [24:14] eagle fighting on it and it it it [24:16] features like the UFC octagon. Um, which [24:19] like was sort of where people's [24:20] expectations are for these types of [24:22] things. But this feels like [24:23] >> surprising [24:25] UFC branding on the passport. [24:27] >> Yeah, no ads is is a big step backwards. [24:31] Yeah, I will say the the passport text [24:33] where it's kind of sans serif all caps [24:35] and there's like spacing that does look [24:38] like AI slop kind of. [24:39] >> It does. [24:40] >> This is like very common when you just [24:42] have [24:42] >> too much too wide of kerning is your [24:45] problem. [24:46] >> Okay. The kerning needs some work. Yeah, [24:50] >> it does read a little odd. [24:52] >> That's pretty high stakes graphic [24:53] design. [24:54] >> It is most you're like, "Yeah, we're [24:56] going to make this, you know, pamphlet [24:58] for a conference and nobody's going to [25:00] read it." And if they do, they'll look [25:02] at it for a second. The passports like [25:04] we're going to make, you know, a hundred [25:05] million of these and then they're going [25:08] to carry them for a decade. Yeah. [25:10] >> Representing the country globally. [25:12] >> Yeah. [25:12] >> They will be inspected many times. [25:15] >> I'm I'm not I'm not into passports. I [25:18] don't I don't want to leave America. [25:20] It's like not a good thing. They can [25:22] just get rid of them. Just stay here. [25:24] >> What else we got? [25:26] >> Oh, Leon fat. [25:30] Lechon. They're calling it Lechonk. [25:32] >> Lchunk. Yes. Uh they're leaning into the [25:34] jokes ML AI. The French Neol. [25:38] >> Is it fair to call them a neolab or are [25:40] they just a lab? [25:41] >> They're just a lab, right? They're [25:44] pollab. [25:45] >> Yeah. Yeah, I would say so. They're kind [25:47] of [25:47] >> We call them preno pre- neol [25:50] specifically. [25:51] >> Uh let me actually consult our [25:52] nomenclature. [25:53] >> Okay. Oh, I think they were like [25:55] national lab or something like that [25:57] >> up there with Lawrence Livermore. Uh but [26:00] they launched Mistral large 4 aka lonk [26:04] one trillion parameters natively made it [26:07] multimodal 49 billion active parameters [26:10] it is the best openweight model from US [26:13] or Europe or Europe on aggregated [26:15] benchmark state-of-the-art on critical [26:17] workloads including cyber defense [26:19] manufacturing and finance and it [26:20] surpasses closed frontier models on [26:22] visual grounding forged in Europe end to [26:26] end deployable from Europe up via our [26:29] our own Mistral cloud infrastructure [26:32] available to all via API today working [26:36] with cyber security partners privately [26:38] open weights release end of October so [26:41] presumably they'll release the open [26:43] weights for like the cyber nerfed [26:45] version I would imagine something like [26:46] that I don't know uh anyway uh what is [26:50] your review Tyler do you have any other [26:53] uh data to share about Mrol's progress [26:56] >> uh yeah I mean it looks very good um I [26:58] It's like very much on par with the kind [27:00] of frontier of the open source. So like [27:02] the Chinese labs, it's like very much a [27:04] contender. It's not necessarily like [27:06] destroying all of them on on every [27:07] benchmark, but it's like certainly if [27:09] people are worried about um using [27:11] Chinese open source models because maybe [27:14] downstream there's going to be some like [27:15] legal ramification or whatever. Yeah. [27:17] Like this is a very good choice. [27:19] >> It really does change the open source [27:21] closed source debate because [27:23] >> Yeah. I think Mistl and Reflection which [27:25] is kind of the the American open source [27:27] lab. Yep. [27:28] >> Like we're now starting to see like, [27:29] okay, these models are actually like [27:30] quite [27:30] >> quite competent. [27:31] >> And Reflection uh came out with it's [27:33] called Beam yesterday. [27:38] >> Whoa. Not tapped in. [27:41] >> You're like, "Sorry, sorry. I can't I [27:43] can't hear I can't hear you. I can't [27:44] hear you." [27:47] >> It's called Yes. Uh anyways, we we can [27:50] come come back to that, but uh great [27:51] great to see them getting a a new model [27:55] out. [27:56] >> And Leo has the the the the funniest [27:59] benchmark of them. Says it's pretty [28:01] funny that it seems the one thing Mrol [28:04] large 4 is actually leading on is a [28:06] benchmark about regulation. Another EU [28:08] banger. Uh Harvey Harvey's legal agent [28:12] benchmark uh is up to 15% uh whereas [28:16] other models are down at between five [28:18] and 13%. Uh interesting. I I wonder if [28:22] that's uh intentional. I wonder if [28:24] there's actually anything to read into [28:25] that. I I I really have no idea, but uh [28:27] it's certainly certainly valuable. Uh, [28:29] I'm sure Harvey and all the other uh [28:32] legal companies will be excited because [28:34] we I mean we read the whole thing about [28:35] Harvey uh their margins being in a [28:38] really rough spot for a minute when they [28:40] were overly uh overly reliant on like [28:43] the most expensive frontier models. Now [28:44] there's a ton of price competition [28:46] between OpenAI and Enthropic and then [28:47] obviously there's a bunch of pressure [28:48] from uh open source models. So it like [28:51] it it it's the true it's truly the uh [28:54] create the benchmark that you want the [28:56] the labs to hill climb and they just [28:58] created the benchmark for like their [29:00] exact workload and then they're able to [29:02] do that. [29:02] >> Sushi commander says what are you [29:04] talking about both bean and mol beam and [29:06] mol large are a generation behind the [29:09] latest Chinese leading model. So putting [29:11] you in the truth zone Tyler [29:13] >> I mean you can look at the benchmarks [29:14] it's like yeah it's not leading them [29:15] it's like among one of the contenders. [29:18] It's not the frontier model for sure. [29:20] >> It's like in in line, but a lot of the a [29:22] lot of the like the most recent round of [29:24] of Chinese open source happened like a [29:26] couple months ago, which is always rough [29:28] because like you catch up to the [29:29] frontier and then it's like, well, [29:31] you're going to take another six months [29:32] to get something out where they're going [29:34] to only take another two months to get [29:35] something out. So, that window, but [29:37] there's so many other uh geopolitical uh [29:40] implications that uh lead companies to [29:43] go one way or another. So, we'll see. [29:45] We'll see. Um, uh, Olivia Moore has some [29:48] data on what's happening in the battle [29:50] for consumer AI dominance between Chachi [29:53] PT, Claude, and Gemini. Uh, A16Z pulled [29:56] the data across web and mobile plus [29:58] consumer revenue. Uh, Chad GPT is at 2x [30:01] Gemini scale and 6x scale on web and 2 [30:05] and a half Gemini scale and 14 times [30:08] cloud scale on mobile. Interesting. I [30:10] mean, I guess that makes sense. More [30:11] consumer focused. You're you're pulling [30:14] there. Uh [30:15] >> the main thing here is just look at look [30:16] at the run from Claude. [30:19] >> Yeah. [30:19] >> And it was [30:20] >> I mean we saw this in the app store, [30:21] right? Like you would open up the app [30:22] store and like Claude was at the top for [30:23] a while and then and then fell down and [30:26] now um Chach is back on top and they're [30:28] just going back and forth. It's the the [30:30] war for the Frontier is just trading [30:32] shots back and forth, back and forth. [30:35] >> Right now MW is still at number one. [30:37] Vinted, who we're working on, is number [30:39] two. [30:40] >> Vented is still number two. so dominant. [30:43] >> The folks over at Vinted are are really [30:45] doing a fantastic job. Thank you for [30:47] tuning in to TVPN today. It's been an [30:49] honor. [30:49] >> Fantastic show. We'll see you tomorrow [30:51] at 11 a.m. Pacific. [30:52] >> I can't wait. [30:53] >> Leave us five stars in Apple Podcast and [30:54] Spotify. Sign up for our newsletter at [30:56] tvpn.com and we will see you tomorrow. [30:59] >> Love you. Have a wonderful [31:00] >> Goodbye.