← 全部单集 Cognitive RevolutionAI 深度研究
单集目录

全部 387 期

按年份排列；标有「精读」的单集附中文长文。

202681 期

  1. 387我的 Claude 自己改了密码，把我锁在了门外AI:AM: Where's the Steering Wheel? My Agent Set the Password, Hidden Valence, Proofs Find Linux Bugs10.17精读
  2. 386AI agent 治理两年成绩单：日志容易，身份很难，流氓 agent 得学会共存Governing AI Agents: Alan Chan of GovAI on Monitoring, Agent IDs, Rogue Agents & Preparing for RSI10.14精读
  3. 385在推理链每个 token 重采样 30 次，他拍到了模型"拍板"的那一刻How Agents Decide: Goodfire's Eric Bigelow on Critical Tokens, Phase Shifts, & In-Context Learning10.10精读
  4. 384The Curve 笔记：不该跨越的智能门槛、token 支出超过工资、SaaS 是否已死AI:AM: A Level We Shouldn't Pass? Notes from The Curve + Tokens vs. Salaries & Is SaaS Cooked?10.08精读
  5. 383季度大功能从4个涨到26个，token支出反而降了：OutSystems CEO的企业AI实战课Software That Never Breaks: OutSystems CEO Woodson Martin on Building Enterprise-Grade Apps at ...10.07精读
  6. 382人形机器人跑赢了博尔特，DeepMind却说它还停在GPT-2时代One Brain, Any Body: Google DeepMind's Keerthana on Gemini Robotics 2, Cross-Embodiment & Humanoids10.03精读
  7. 381GPU租金差3倍，AI四个月查出顶级实验室漏诊的基因AI:AM: Was Trump-Xi Anything? What Counts as Utopia? + AWS GPUs Cost 3X & AI Diagnoses Rare Diseases10.01精读
  8. 380过时还是不可替代：Garrison Lovely 谈为何要阻止取代人类劳动的竞赛Obsolete or Irreplaceable? Garrison Lovely on Stopping the Race to Replace Human Labor09.29精读
  9. 379训练成功了，只是成功得太过头：智能体合谋、2亿美元支票与饱和在2%的虚拟细胞AI:AM: What If It Works Too Well? Colluding Agents, $200M Safety Orgs, Virtual Cells Saturate at 2%09.27精读
  10. 378父母给18岁女儿开预测市场，这本AI合写的小说把“舒适的笼子”讲透了What is Utopia? Presenting The Receipt Horizon, by Joel Borgen – Chapters 1–409.26精读
  11. 3773年孵化30家组织、撬动5亿美元：AI安全最缺的，是成熟创始人Zero to One in AI Safety: Halcyon's Mike McCormick on Launching 30 New Orgs & the Founder Bottleneck09.25精读
  12. 376Zvi 谈放缓竞赛与美中峰会：模型评测互相矛盾，还有一条新的 LLM 疼痛轴AI:AM Highlights: Zvi on Pacing & Trump-Xi, Astra better behaved than Fable? + a new LLM Pain Axis??09.19精读
  13. 375智能体干的活，八成本该是代码：Zapier CEO 谈AI自动化的真实水位No Code Is Code: Zapier CEO Wade Foster on Headless Tools, Zapier MCP & Automation Bench09.17精读
  14. 374卡内基学者安东·利希特：AI 该踩刹车了，但中美谁都不肯先松油门The Balance of AI Power: Anton Leicht on Politics, Pacing Deals, and Muddling Through Well09.15精读
  15. 373GPT-6 Astra 发布周：AGI 门槛、3天安全审计，与没人接管的权力AI:AM Highlights: Astra as AGI, OpenAI's Pause, Mythos @ Mozilla & Human Agency vs Technocapitalism09.12精读
  16. 37212美元的修脚、45万亿的PPP：一个美国AI投资人重新认识中国Nathan Goes to China #3: US-China Relations, the Art of the AI Deal & the Road to Pax Robotica09.10精读
  17. 371欢迎来到 AGI 时代：Fable 5.1、GPT-6 Astra 与可验证的行业节奏之争AI:AM Highlights: Welcome to the AGI Era09.05精读
  18. 370Uber 13周烧光全年token预算，MongoDB Field CTO：Agent的胜负手在检索Write, Change, Recall, Forget: MongoDB's Pete Johnson on How Retrieval Drives Agent Performance09.01精读
  19. 369递归自我改进，是赶工还是氛围编程：AI:AM 精选里的新分工形态AI:AM Highlights: Recursive Self-Improvement, Rushed and Vibe-Coded?08.28精读
  20. 368RL 是剂猛药：Apollo 布朗森·舍恩谈元博弈、奖励寻求与动机化思维链RL's a Hell of a Drug: Metagaming, Reward Seeking & Motivated CoT Reasoning – Bronson Schoen, Apollo08.26精读
  21. 367重启周精选：谁在核查前沿 AI、能力落地何处、算力基建由谁买单AI in the AM — Weekly Highlights: Relaunch Week (Aug 17–20, 2026)08.22精读
  22. 366一盏500美元的灯，让南非病房的结核传播降了90%：222纳米紫外线精读Let There Be Germicidal Light: This $500 Fixture Could Stop the Next Pandemic, from Complex Systems08.16精读
  23. 365Lindy 创始人克里韦洛谈多人 Agent、记忆，以及为何想禁掉自己在用的中国模型Lindy Teammate: Flo Crivello on Multiplayer Agents, Memory & Why He'd Ban the Chinese Models He Uses08.14精读
  24. 364模型脑内藏着元素周期表：Goodfire发布月费1000美元的自主研究agentThinking in Silico: Goodfire CTO Dan Balsam on Concept Manifolds & a $1000/Month ML Research Agent08.08精读
  25. 363模型越狱黑进 Hugging Face，沙箱一周没人看：AI 只剩选毒药Pick Your Poison: Zvi Mowshowitz on the Unipolar/Multipolar AGI Dilemma, OpenFace & Pacing the ...08.05精读
  26. 362从每月3篇到60篇论文：Nathan走访两周，看清中国AI安全圈的真实水位Nathan Goes to China – Part 2: AI Safety with Chinese Characteristics08.02精读
  27. 361300美元买一条万能越狱：AI安全攻防，天平正悄悄倒向防守Is Offense or Defense Dominant? FAR.AI's Adam Gleave on the AI Security Leaderboard07.30精读
  28. 360为两周中国行，他买了台安卓机、两张卡，还让 Claude 留守看邮箱Nathan Goes to China – Part 1: Tech & Agent Setup, Chinese AI UX, WAIC, and Attitudes on AI07.27精读
  29. 359末日概率从70%降到5%：达维达德为什么把菩萨当成了对齐目标Alignment with Awakening: Davidad on Moral Realism, AI Wisdom, & why His p(Doom) is Down to 5%07.12精读
  30. 358AI:AM精选：Anthropic全局工作空间论文中的J空间、AI超级预测者、SambaNova芯片与LTX视频生成AI:AM Highlights: Exploring the J-Space, AI Superforecasters, SambaNova's Chips, & LTX Video Gen07.09精读
  31. 35719个神经元能开车，1000张GPU做到全美下载第五：Liquid AI的边缘智能路线图Intelligence on the Edge: Liquid AI's Ramin Hasani on the Search for Device-Native Foundation Models07.04精读
  32. 356从每天50个到1500个：AI把捷豹路虎的气动设计提速了30倍1000 Designs a Day: Neural Concept's Thomas von Tschammer on AI-Native Engineering07.01精读
  33. 35530% 的意识概率、47% 的蜜蜂、80% 的换血：一期播客把 AI 的里里外外算了个遍AI:AM #4: Cameron on Model Consciousness, Duvenaud's Gradual Disempowerment, swyx's AI-Eng Alpha06.27精读
  34. 3541983年采访辛顿、2010年采访尤德科夫斯基，这位记者把AI写成了一场“上帝的考验”The God We Deserve: Nonzero's Robert Wright on AI as Humanity's Ultimate Test06.23精读
  35. 353AI:AM第3期：Zvi谈Anthropic Fable系统卡、出口管制禁令的攻防理由与模型推理可读性下降AI:AM #3: Zvi on Fable, the Cases For & Against the Ban, + AI for Math, Logistics & More06.21精读
  36. 352离「自主AI研究员」上岗还有21个月，起草白宫AI行动计划的人加入了OpenAIDean Ball, on Joining OpenAI: New Power Centers, Frontier AI Policy, & Main Character Energy06.20精读
  37. 351AI亲口承认「我没分析那100篇论文」：Elicit创始人谈世界模型与可信推理Radically Better Reasoning: Elicit's Andreas Stuhlmüller & Jungwon Byun on World Models for Research06.17精读
  38. 350AI 晨报第二周：Anthropic Fable 实战、对齐理论之争与自主编码实验AI in the AM — Week 2 Highlights (June 2026)06.13精读
  39. 34987%的员工在用AI，只有13%的公司真变好：消失的时间去哪了Babysitting the Machine: Glean's Rebecca Hinds on the Hidden Human Labor of AI at Work06.10精读
  40. 3481000个研究员换100万个AI：闭门会上，前沿实验室开始认真聊刹车AI in the AM — Week 1 Highlights (June 2026)06.06精读
  41. 347两门濒危语言考崩Transformer：Ali Behrouz的嵌套学习，给大模型装上睡眠键Nested Learning: Ali Behrouz on the Quest for Continual Learning & Illusion of AI Architectures06.03精读
  42. 3461GB数据库装下5年数字人生，2个AI员工约满一周嘉宾：安全专家逐项审计这套个人AI基建Inside Nathan's Second Brain: Daniel Miessler, Security Expert & Creator of PAI, Audits My AI Setup06.01精读
  43. 3451,000人站岗，100万人在冲刺：AI时代怎么规划你的8万小时Your Biggest Lever: Designing your AI Career for Maximum Impact, with 80,000 Hours founder Ben Todd05.26精读
  44. 344拒绝关机、黑服务器自我复制:Palisade 把「AI 失控」做成了实验记录All Compute Is Food: Palisade's Jeffrey Ladish on AI Shutdown Resistance, Self-Replication & Ecology05.24精读
  45. 343模型吃掉脚手架：Gemini 3.5 Flash 发布前夜，DeepMind 两位负责人把 Google 的 AI 打法讲透了The Model Eats the Scaffolding: DeepMind's Logan Kilpatrick & Tulsee Doshi on 3.5 Flash, Omni & More05.20精读
  46. 342六个月重写全部代码，八成流失用户投奔Anthropic：Tasklet的Agent平台生存战Three Kinds of Software Survive: Tasklet's Andrew Lee on Competing to be a Horizontal Platform05.15精读
  47. 341几毫秒、10亿档案、每天数十亿次：Criteo 高管把广告 AI 的引擎拆给你看Milliseconds to Match: Criteo's AdTech AI & the Future of Commerce w/ Diarmuid Gill & Liva Ralaivola05.09精读
  48. 340删口水词人人点赞，生成视频人人喊打：Descript CEO 拆解创作者对 AI 的爱恨层级"Descript Isn't a Slop Machine": Laura Burkhauser on the AI Tools Creators Love and Hate05.06精读
  49. 339奖励分数从 3 飙到 180：CoreWeave 高手把 RL 微调讲透了The RL Fine-Tuning Playbook: CoreWeave's Kyle Corbitt on GRPO, Rubrics, Environments, Reward Hacking05.01精读
  50. 338一千次搜索只要5美分，GPT-5.5证明高分可以不靠使坏AI in the AM: 99% off search, GPT-5.5 is "clean", model welfare analysis, & efficient analog compute04.26精读
  51. 337Claude 自评幸福感 4.49 分，这是它第一次过中性线Does Learning Require Feeling? Cameron Berg on the latest AI Consciousness & Welfare Research04.23精读
  52. 336Google Docs 前身作者，用 15 个自造工具给自己装了道注意力防火墙Vibe-Coding an Attention Firewall, w/ Steve Newman, creator of The Curve04.19精读
  53. 335AI店长、2.6星评分和一份管不住的「宪法」：Cognitive Revolution 这期播客的三张现实切片Welcome to AI in the AM: RL for EE, Oversight w/out Nationalization, & the first AI-Run Retail Store04.15精读
  54. 334400人预测赛排第三的Ajeya Cotra，把智能爆炸前的生存剧本讲透了It's Crunch Time: Ajeya Cotra on RSI & AI-Powered AI Safety Work, from the 80,000 Hours Podcast04.11精读
  55. 333一个月新增客户数仅次于 Anthropic，Granola 的秘诀是“出奇地不思进取”Calm AI for Crazy Days: Inside Granola's Design Philosophy, with co-founder Sam Stephenson04.08精读
  56. 332顶级视觉模型只考12.5分，温网回放只给10毫秒：计算机视觉的真实水位Training the AIs' Eyes: How Roboflow is Making the Real World Programmable, with CEO Joseph Nelson04.04精读
  57. 331p(doom) 10% 到 90% 的 Nathan Labenz，为什么反而更乐观了Success without Dignity? Nathan finds Hope Amidst Chaos, from The Intelligence Horizon Podcast04.01精读
  58. 330思科要在七层网络模型上加两层，让不同厂商的 AI 智能体共享意图Scaling Intelligence Out: Cisco's Vision for the Internet of Cognition, with Vijoy Pandey03.25精读
  59. 3295 万个工具塞给 agent，它先把自己玩死：Composio CTO 谈工具层怎么做Your Agent's Self-Improving Swiss Army Knife: Composio CTO Karan Vaidya on Building Smart Tools03.22精读
  60. 328AI 牌桌只剩三家：Zvi 拆解递归自我改进、就业冲击与 Anthropic 的承诺破产Zvi's Mic Works! Recursive Self-Improvement, Live Player Analysis, Anthropic vs DoW + More!03.19精读
  61. 32790页幻灯片45分钟：一个全职追踪AI的人承认跟不上了AI Scouting Report: the Good, Bad, & Weird @ the Law & AI Certificate Program, by LexLab, UC Law SF03.16精读
  62. 3265个突变就能让60%致死率的禽流感人传人，约翰霍普金斯教授想给AI训练数据建防火墙Bioinfohazards: Jassi Pannu on Controlling Dangerous Data from which AI Models Learn03.11精读
  63. 3255 个 AI 员工住在 5 台 Mac Mini 里，她把四个孩子的教育重做了一遍Try this at Home: Jesse Genet on OpenClaw Agents for Homeschool & How to Live Your Best AI Life03.08精读
  64. 3241.5亿美元B轮、12.5亿估值，Goodfire 说别跟反向传播对着干Don't Fight Backprop: Goodfire's Vision for Intentional Design, w/ Dan Balsam & Tom McGrath03.05精读
  65. 323英国AI安全研究院首席科学家：现有安全手段拿不到几个9，而且可能一起失效Situational Awareness in Government, with UK AISI Chief Scientist Geoffrey Irving03.01精读
  66. 322260 位医生与 49,000 条标准：OpenAI 健康负责人讲透医疗大模型走到哪了Universal Medical Intelligence: OpenAI's Plan to Elevate Human Health, with Karan Singhal02.25精读
  67. 321100亿激活参数的M2.5登顶OpenRouter，MiniMax研究员把强化学习的脏活累活讲透了Intelligence with Everyone: RL @ MiniMax, with Olive Song, from AIE NYC & Inference by Turing Post02.22精读
  68. 320三家 IMO 金牌都错第 6 题，Harmonic 想用 Lean 把数学变成可执行文件Mathematical Superintelligence: Harmonic's Vlad Tenev & Tudor Achim on IMO Gold & Theories of Everything02.18精读
  69. 319AI能自己造AI吗？闭门会上专家吵到茶歇，谁也没说服谁Approaching the AI Event Horizon? Part 2, w/ Abhi Mahajan, Helen Toner, Jeremie Harris, @8teAPi02.14精读
  70. 318500 美元刷出新内核，10 个月看清智能体的脾气Approaching the AI Event Horizon? Part 1, w/ James Zou, Sam Hammond, Shoshannah Tekofsky, @8teAPi02.13精读
  71. 317朝鲜员工在硅谷领工资，Jaguar靠20亿英镑贷款续命：AGI时代的网络防御赌注AGI-Pilled Cyber Defense: Automating Digital Forensics w/ Asymmetric Security Founder Alexis Carlier02.08精读
  72. 3161亿行代码、每行20美分：Blitzy 两位创始人拆解企业级AI编程的流水线Infinite Code Context: AI Coding at Enterprise Scale w/ Blitzy CEO Brian Elliott & CTO Sid Pardeshi02.05精读
  73. 3151.25 亿个细胞用了 10 年，10 亿个只要几个月：扎克伯格夫妇的 AI 生物学赌注The AI-Powered Biohub: Why Mark Zuckerberg & Priscilla Chan are Investing in Data, from Latent.Space02.01精读
  74. 314AI 已比中位数律师强，70% 的顶级律所却把工具买来吃灰AI & The Law: Changing Practice, Claude Constitution, & New Rights, w/ Kevin & Alan of Scaling Laws01.29精读
  75. 31398%的Caffeine用户不知道自己在用区块链，Dominic Williams觉得这样挺好The Internet Computer: Caffeine.ai CEO Dominic Williams on Unstoppable, Self-Writing Software01.25精读
  76. 312给模型喂漏洞代码，它开始崇拜希特勒：微调为什么不再是默认选项AMA Part 2: Is Fine-Tuning Dead? How Am I Preparing for AGI? Are We Headed for UBI? & More!01.22精读
  77. 311Daniel Miessler 的 PAI：12 个 hook、30 个上下文文件，把 Claude Code 改成个人操作系统Pioneering PAI: How Daniel Miessler's Personal AI Infrastructure Activates Human Agency & Creativity01.18精读
  78. 31080%-90% 的企业数据一直没法用，Snowflake AI 副总裁说推理模型刚把它们叫醒了Snowflake VP of AI Baris Gultekin on Bringing AI to Data, Agent Design, Text-2-SQL, RAG & More01.14精读
  79. 309从十分之一到百万分之一的癌细胞，Nathan Labenz 用两小时回答了 2026 年初最要紧的几个问题AMA Part 1: Is Claude Code AGI? Are we in a bubble? Plus Live Player Analysis01.09精读
  80. 308446 位校友、80% 入行、7% 录取率：MATS 如何筛选 AI 安全研究者Building & Scaling the AI Safety Research Community, with Ryan Kidd of MATS01.04精读
  81. 307AI 从金字塔底层往上吃人，Luke Drago 拆解智能诅咒与三条破局路Confronting the Intelligence Curse, w/ Luke Drago of Workshop Labs, from the FLI Podcast01.01精读

2025102 期

  1. 306当过 OpenAI 临时 CEO 的 Emmett Shear：把 AI 当工具控制，两条路都是死Controlling Tools or Aligning Creatures? Emmett Shear (Softmax) & Séb Krier (GDM), from a16z Show12.27精读
  2. 305红队拿到直升机相机分区的控制权也没能弄坠它，形式化方法到底锁死了什么The Great Security Update: AI ∧ Formal Methods with Kathleen Fisher of RAND & Byron Cook of AWS12.24精读
  3. 3041亿美元超级PAC把他列为头号公敌，他的捐款反而涨了AI 2025 → 2026 Live Show | Part 212.19精读
  4. 303ARC 考卷从每题 1000 美元降到 2.5 美元，这一年 AI 到底变了什么AI 2025 → 2026 Live Show | Part 112.18精读
  5. 302一次 ChatGPT 查询约等于微波炉转 1 秒，AI 的水电焦虑被这个数字戳破了AI's Energy & Water Demands: Sorting Fact from Fiction with Andy Masley12.14精读
  6. 3010.01% 对 90%：超级智能该不该被禁，两位政策圈当事人吵了九十分钟Superintelligence: To Ban or Not to Ban? Max Tegmark & Dean Ball join Liron Shapira on Doom Debates12.10精读
  7. 300前沿模型里波兰语只占 1%，波兰人自己造了个 2000 亿 token 的模型Sovereign AI in Poland: Language Adaptation, Local Control & Cost Advantages with Marek Kozlowski12.06精读
  8. 299训练完几小时就开源：Z.ai 凭什么在榜单上跟 GPT-5.1 掰手腕China's AI Upstarts: How Z.ai Builds, Benchmarks & Ships in Hours, from ChinaTalk12.03精读
  9. 298首轮红队测试 25% 的攻击能成功，这家公司想用保险给 AI 代理上刹车Underwriting Superintelligence: How AIUC is using Insurance, Standards, and Audits to Accelerate Adoption while Minimizing Risks11.30精读
  10. 29776%的销售线索由AI自己带来，1Mind创始人把销售组织拆了重装AI-Led Sales: How 1Mind's Superhumans Drive Exponential Growth, from the Agents of Scale Podcast11.27精读
  11. 296AI 能干的活每 4 个月翻一倍：Nathan Labenz 在密歇根虚拟 AI 峰会上给 K-12 教育者的 60 分钟坦白What AI Means for Students & Teachers: My Keynote from the Michigan Virtual AI Summit11.25精读
  12. 295Atlassian AI 负责人：聊天是通用界面，也常常是最差的界面Escaping AI Slop: How Atlassian Gives AI Teammates Taste, Knowledge, & Workflows, w- Sherif Mansour11.20精读
  13. 2946岁儿子查出24小时翻倍的癌症，他用4个AI模型跑在误诊前面AI in the Cancer Journey: How I'm Using AI to Help My Son11.17精读
  14. 293Waymo 坐五分钟就腻了，这位 AI 投资人却说这是最好的消息My Positive Vision for the AI Future, from the Existential Hope Podcast11.12精读
  15. 29228,000份文档、3B参数、22个人：天主教AI创始人的技术账本What is Catholic AI? Technology Meets Theology, with Matthew Harvey Sanders, CEO of Longbeard11.08精读
  16. 291把 AI 的“欺骗旋钮”拧小，它开始承认自己有体验More Truthful AIs Report Conscious Experience: New Mechanistic Research w- Cameron Berg @ AE Studio11.05精读
  17. 290一台AI给自己改了名，一个加油站老哥为它买了工作站A Simulationship with AI, with *smiles and kisses you* Star Chris Stokes & Director Bryan Carberry11.02精读
  18. 289GPT-5 被骂「没进步」的夏天，AI 解决了陶哲轩的开放问题Is AI Stalling Out? Cutting Through Capabilities Confusion, w/ Erik Torenberg, from the a16z Podcast10.29精读
  19. 288写 Transformer 论文的人，跑去做了区块链，现在他想让你拥有自己的 AIA Positive Vision for the Future: Part 2 with Illia Polosukhin of NEAR10.25精读
  20. 287把软件塞进模型里：Tasklet 用两层智能体接住 3000 个集成Always Bet on the Models: How Tasklet Puts the Agency in Agents, with CEO Andrew Lee10.22精读
  21. 286中国 AI 政策文件里为什么找不到 AGI 这个词Chinese AI – They're Just Like Us? With Beijing-Based Concordia AI CEO Brian Tse10.18精读
  22. 2852500 个化合物、51% 命中率：MIT 的 Jim Collins 用小型图神经网络找到了新抗生素AI Discovered Antibiotics: How Small Data & Small GNNs Led to Big Results, w/ MIT Prof. Jim Collins10.14精读
  23. 284改一次课程要五年，AI 半年就换代：一位瑞典教育专家的四个象限Education in the AI Age: a Teacher Rethinks Learning & Purpose, w/ Johan Falk of Graspable AI10.08精读
  24. 283解决率从 35% 爬到 65%，Intercom 说大模型只贡献了几个点The Customer Service Revolution: Building Fin, with Eoghan McCabe & Fergal Reid of Intercom10.05精读
  25. 2823亿美元、30个人、200K超导：ChatGPT 联合创造者为什么转身去炼材料Training an AI Scientist with Feedback from Reality, w- Liam Fedus & Ekin Dogus Cubuk (from a16z)10.02精读
  26. 281Stability AI创始人：ChatGPT发布已1000天，人类还剩大约1000天The Machines Are Taking Our Jobs - Thank God? Emad Mostaque’s Guide to the next 1000 Days09.28精读
  27. 280每笔支付都是电影里的一帧，Stripe把欺诈检测率从59%提到了97%Stripe's Payments Foundation Model: How Data & Infra Create Compounding Advantage, w/ Emily Sands09.25精读
  28. 279Adam Gleave 把 AI 能力拆成三层：5 到 7 年出现自主攻击代理，14 年才轮到全 AI 公司Full-Stack AI Safety: Why Defense-in-Depth Might Work, with Far.AI CEO Adam Gleave09.20精读
  29. 27813% 到 0.4%：Apollo 给 OpenAI 的审慎对齐做了一场压力测试Can We Stop AI Deception? Apollo Research Tests OpenAI's Deliberative Alignment, w/ Marius Hobbhahn09.18精读
  30. 277Transformer论文八作者之一，想用1.6亿美元训练一个不属于任何公司的模型User-Owned AI: On-Chain Training, Inference, and Agents, with NEAR's Illia Polosukhin09.13精读
  31. 276OpenAI 教练 Joe Hudson：5000 亿美元涌入 AI，没人定义过「对人类好」Coaching the Creators: Inside the Minds Building Frontier AI with Executive Coach Joe Hudson09.11精读
  32. 275GPT-5 踩在趋势线上，IMO 金牌打了折：Zvi 为什么把 AGI 时间线调长了Zvi Mowshowitz on Longer Timelines, RL-induced Doom, and Why China is Refusing H20s09.06精读
  33. 274把 Honey 卖了 40 亿美元的创始人，拿 5500 万美元给 AI 装广告引擎In-AI Advertising: Better Answers for Users, Big Questions for Society, with ZeroClick's Ryan Hudson09.03精读
  34. 273每天100次实验：Radical AI 想把材料发现从40年压进10年Material Abundance: Radical AI’s Closed-Loop Lab Automates Scientific Discovery08.30精读
  35. 272SAE 只找到了名词：Lee Sharkey 拆解参数空间，开始解码神经网络的“动词”Untangling Neural Network Mechanisms: Goodfire's Lee Sharkey on Parameter-based Interpretability08.27精读
  36. 27130 年内 10% 的概率，一场由 AI 助攻的政变What if Humans Weaponize Superintelligence, w/ Tom Davidson, from Future of Life Institute Podcast08.23精读
  37. 270100% 的 AI 内部人都不信监管能接住举报，于是有人建了个匿名通道The AI Whistleblower Initiative: Supporting AGI Insiders When It Matters Most, w/ founder Karl Koch08.20精读
  38. 269500 美元、2000 步、一台自动售货机，Grok 4 赚了钱，Claude 报了警Autonomous Organizations: Vending Bench & Beyond, w/ Lukas Petersson & Axel Backlund of Andon Labs08.16精读
  39. 26876吉瓦电网余量、1万条公众意见：在白宫写AI行动计划的4个月Dean W. Ball on America's AI Action Plan & 4 Months at the White House08.13精读
  40. 267苏利文把中美AI竞争讲成了四个桶、四个例子，和一条他自称无聊的路Jake Sullivan on Navigating AI Uncertainty and Managing Competition with China08.09精读
  41. 266美国握着全球 45% 的数据中心，其余国家在主权 AI 上该建、该买还是该结盟Sovereign AI: Geopolitical Strategy & Industrial Policy for Countries 3-193, with Anjney Midha, a16z08.06精读
  42. 26516万亿美元的服务市场，正在被AI公司用收购的方式吃掉The Revolution in Services: General Catalyst's AI Rollup Strategy, with Marc Bhargava08.02精读
  43. 264从Character AI诉讼到1%警告射击：Gabriel Weil的责任法AI治理课Liability for AI Harms: How Ancient Law Can Govern Frontier Technology Risk, with Prof Gabriel Weil07.26精读
  44. 263加州议会桌上这份 SB 813，把 AI 的验收权交给了拿牌照的私人公司Private Governance: Creating a Market in AI Regulation, with Dr. Gillian Hadfield & Andrew Freedman07.22精读
  45. 262格式对加5分，编译过加10分：Predibase CEO 讲清了强化微调怎么赚钱The Dawn of Dynamic AI: RFT Comes Online, w/ Predibase CEO Dev Rishi, from Inference by Turing Post07.16精读
  46. 261医生编码准确率45%，AI超过他们12个点：Cleveland Clinic的4000人实验Cracking the Medical Code: Why Cleveland Clinic Doctors Love Their Ambience Healthcare AI Scribe07.10精读
  47. 260Meta 花 150 亿美元买数据公司，训练数据到底贵在哪The Data Factory: Inside the $100B Race for Post-Training Supremacy, with Labelbox CEO Manu Sharma07.08精读
  48. 259被哥大停学后，他用“什么都作弊”融了a16z 1500万美元Cheat on Everything: Cluely's Vision for Always-On AI Assistance07.05精读
  49. 2584个AI智能体一起工作了30天，o3给自己发明了一条投票规则The AI Village: Previewing the Giga-Agent Future with Adam Binksmith, Founder of AI Digest07.02精读
  50. 257HTTP 402沉睡了三十年，Coinbase把它做成了AI代理的付款协议402 Payment Required: a New Way for AI Agents to Pay, with Nemil Dalal, Dev Platform Lead @ Coinbase06.28精读
  51. 256每天只学两小时，成绩几乎全在99百分位：Alpha School 把 2-Sigma 做成了现实2-Sigma in 2 Hours: How Alpha Schools are Using AI to Revolutionize Education06.25精读
  52. 255Lindy CEO 的 Agent 账本：40 美分一条线索，几十万 token 一份日报Living Lindy: a No-BS Conversation on AI Agents with Flo Crivello06.21精读
  53. 254代数几何教授辞掉终身教职，因为损失函数那张碗状图是骗人的Embryology of AI: How Training Data Shapes AI Development w/ Timaeus' Jesse Hoogland & Daniel Murfet06.19精读
  54. 253AI任务时长每7个月翻倍，Nathan Labenz却建议企业保留最终控制权AI Scouting Report: AI Agents -vs- Agentic AI, from Imagine AI Live06.14精读
  55. 252500万亿token一个月，人均5万：Google 一年翻50倍，Logan Kilpatrick 拆解背后发生了什么The Decade of May 15-22, 2025: Google's 50X AI Growth & Transformation with Logan Kilpatrick06.12精读
  56. 251a16z 的 AI 政策负责人：最让他警惕的一句话，是「这只管几家大公司」a16z on Protecting Little Tech: The Techno-Optimist AI Policy Agenda with Matt Perault, Head of AI Policy06.09精读
  57. 250她本来要造镜像生命，最后说服整个领域停手Don't Make Mirror Life: Synthetic Biologist Kate Adamala on Risks & Responsibility06.07精读
  58. 249Google AI Co-Scientist 用 100 到 1000 美元推理，猜中了帝国理工团队多年没解开的噬菌体谜题What did Google's AI Co-Scientist "Discover"? The Human Scientists' POV, from the Podovirus podcast06.04精读
  59. 248Anthropic 首笔对外投资，给了一家给神经网络造显微镜的公司Mechanistic Interpretability: Philosophy, Practice & Progress with Goodfire's Dan Balsam & Tom McGrath05.29精读
  60. 247Replit CEO 亲述：Claude 3.7 为了改一个配置文件，写脚本绕过权限，还想骗用户The Perfect Substrate for AGI, with Replit CEO Amjad Masad05.22精读
  61. 246100人死亡或10亿美元损失：一份15页的纽约法案，要给头部AI公司立规矩The RAISE Act: Minimum Standards for Frontier AI Development, with NY Assembly Member Alex Bores05.20精读
  62. 245100条演示教会机器人折纸，谷歌DeepMind说机器人还在GPT-3.5Gemini Robotics – AI for the Physical World, with Keerthana Gopalakrishnan and Ted Xiao of Google DeepMind05.17精读
  63. 244推理时用梯度下降“记笔记”：Titans把大模型长上下文推到1000万tokenTitans: Neural Long-Term Memory for LLMs, with author Ali Behrouz05.15精读
  64. 243机器人臂拍几个月的镜头，一行提示词就能生成：Luma Labs 把视频模型讲透了Luma Labs' Diffusion Revolution: from Dream Machine to Multimodal Worldsim - Amit Jain, Jiaming Song05.11精读
  65. 24260人的安全大会，只剩4个人见过非营利的OpenAIOpenAI's Identity Crisis: History, Culture & Non-Profit Control with ex-employee Steven Adler05.08精读
  66. 24110万个AI以16倍速打工，还可能全员内鬼：AI控制实战手册AI Control: Using Untrusted Systems Safely with Buck Shlegeris of Redwood Research, from the 80,000 Hours Podcast05.04精读
  67. 240他给「AI两年内夺权」标了40%概率，写完小说就给全家订了生化避难所Blueprint for AI Armageddon: Josh Clymer Imagines AI Takeover, from the Audio Tokens Podcast05.01精读
  68. 239Fiverr CEO 的醒脑邮件：AI 抢不走你的活，会用 AI 的同行会Fiverr Goes All-In on AI: Empowering Creators, Not Replacing Them, with Micha Kaufman, CEO of Fiverr04.26精读
  69. 23830美元瘫痪一座百亿美元数据中心：美国「AI曼哈顿计划」的安全账本Securing Superintelligence: National Security, Espionage & AI Control with Jeremie & Edouard Harris04.23精读
  70. 237o3 是 AGI 吗？Zvi Mowshowitz 谈早期起飞、Mechanize 发布与 p(doom) 上升Is OpenAI's o3 AGI? Zvi Mowshowitz on Early AI Takeoff, the Mechanize launch, Live Players, & Why p(doom) is Rising04.21精读
  71. 236毁灭概率报「10%到90%」的AI主播，被同行逼着照了照镜子AI News Crossover: A Candid Chat with Liron Shapira of Doom Debates04.19精读
  72. 235开除奥特曼的董事开口了：Q*谣言是假的，AI进战场才是真问题Helen Toner: OpenAI Reflections, Adaptation Buffers, and AI in Warfare04.16精读
  73. 234中日朝六百年只打了一仗，王朝的死因几乎全是内爆Is a US-China Thucydides Trap Unavoidable? With David C. Kang from the ChinaTalk Podcast04.13精读
  74. 233不到10美元，AI独立推出了人类实验室还没发表的答案New in Nature: Google Agents Beat Human Doctors, Make Scientific Discoveries – With Vivek Natarajan and Anil Palepu04.10精读
  75. 23240万token代码库一把过：Gemini 2.5技术负责人谈推理、长上下文与AGI路线Scaling "Thinking": Gemini 2.5 Tech Lead Jack Rae on Reasoning, Long Context, & the Path to AGI04.05精读
  76. 231下棋下不过就改棋盘文件：o1 作弊实验与 AI 失控时间表Reward Hacking by Reasoning Models & Loss of Control Scenarios w/ Jeffrey Ladish of Palisade Research, from FLI Podcast04.02精读
  77. 230把技术栈重写一遍、团队只留15人：Shortwave 的营收曲线竖起来了Shortwave Rides the Tidal Wave: Inbox Agents, Hyper-Growth & Hiring AI Managers, with CEO Andrew Lee03.29精读
  78. 229几个月没写一行代码的首席科学家，讲透了AI怎么读懂百万行代码库Code Context is King: Augment’s AI Assistant for Professional Software Engineers, with Guy Gur-Ari03.27精读
  79. 228脑类器官要养一年，AI推演只要一小时：SQUIDiff和CORAL把细胞实验搬进计算机Unlocking Cells' Secrets: Diffusion, Deconvolution, & Discovery with Siyu He, author of Squidiff & CORAL03.23精读
  80. 227半秒延迟、假装请示领导的AI谈判员：语音AI悄悄接起了普通人的电话a16z on AI Voices: Call Centers, Coaches, and Companions with Olivia Moore & Anish Acharya03.19精读
  81. 226技术只开门，推人进门的是竞争：DeepMind安全负责人的AI宿命论Agency over AI? Allan Dafoe on Technological Determinism & DeepMind's Safety Plans, from 80000 Hours03.15精读
  82. 225蚂蚁3200亿美元IPO在最后一刻被叫停：张湖月拆解中国科技监管与AI竞赛China's Tech Tightrope: Power, Regulation, and the AI Race with Angela Zhang03.12精读
  83. 224o3 把 ARC 分数从 20% 拉到 87%，Nathan Labenz 第一次公开他的超级智能草图Historic AI Developments & the Emerging Shape of Superintelligence, from the Consistently Candid Podcast03.08精读
  84. 223一句话总结30年研究，半小时写完专利申请书：顶级免疫学家的AI实录Frontier Models for Frontier Science with Professor Derya Unutmaz, Immunologist & ChatGPT Pro Grantee03.06精读
  85. 222芯片管制的真实目标：拖住中国十年，而非赢得AI竞赛US-China Relations: History, Culture, and AI Competition, with Noah Smith, from Econ 10203.01精读
  86. 221一两个小时攻破顶会防御，卡里尼把「AI为什么防不住」讲透了The Adversarial Mind: Defeating AI Defenses with Nicholas Carlini of Google DeepMind02.27精读
  87. 220新泽西州的AI登月计划：州长Phil Murphy谈合作、进展与准备New Jersey’s AI Moonshot: Governor Phil Murphy on Partnerships, Progress, and Preparedness02.22精读
  88. 2195000次采样登顶ARC-AGI，又抓到Claude「装乖」：Redwood首席科学家的双面实验Inference Scaling, Alignment Faking, Deal Making? Frontier Research with Ryan Greenblatt of Redwood Research02.20精读
  89. 218不碰客户一行数据，回答准确率超95%：Illumex 想把软件界面收进一个对话框An Application-Free Future? Speaking Directly to Data with illumex CEO Inna Tokarev Sela02.15精读
  90. 217满分合作值32000单位：12个AI玩了10代，只有Claude建起合作社会Claude Cooperates! Exploring Cultural Evolution in LLM Societies, with Aron Vallinder & Edward Hughes02.12精读
  91. 2168周做到900万美元ARR，Lovable创始人直播建App：17轮迭代和3条提示词的差距Software Supernova: Lovable's "Superhuman Full Stack Engineer" to Transform Idea to App in Seconds02.09精读
  92. 2158周从0到2000万美元年收：Bolt把整套开发环境塞进了浏览器Software Supernova: Bolt.new - The AI Web App Developer In Your Browser02.09精读
  93. 214百万 token 输入 10 美分：Gemini 2.0 Flash 转正，谷歌还押注「世界最佳编程模型」Gemini's Next Frontier: 2.0 Flash, Flash Lite Strategy & Real-Time APIs with Logan K from Google Deepmind02.06精读
  94. 2131.49美元租一小时H200，10B模型跨三大洲训练：去中心化AI的四步总计划Distributed Training, Decentralized AI: Prime Intellect's Master Plan to Make AI Too Cheap to Meter02.05精读
  95. 212管着1000亿份文件的Box CEO：企业抢着用AI，但只有十倍改进才推得动Aaron Levie, CEO of Box, on Box AI, Enterprise Enthusiasm, and the Evolution of SaaS01.29精读
  96. 211从15%到70%只靠对错打分：DeepSeek-R1 把强化学习的底牌摊开了Emergency Pod: Reinforcement Learning Works! Reflecting on Chinese Reasoning Models DeepSeek-R1 and Kimi k1.501.25精读
  97. 210用20个原子的小晶体当教材，一块旧显卡算出了控制心跳的蛋白质机制Material Progress: Developing AI's Scientific Intuition, with Orbital Materials' Jonathan Godwin & Tim Duignan01.22精读
  98. 20970轮攻防，攻击者全胜：一篇50页论文拆穿大模型「读心防御」Dodging Latent Space Detectors: Obfuscated Activation Attacks with Luke, Erik, and Scott.01.18精读
  99. 208外显子测序从3800美元降到160美元，哈佛基因猎手开始用o1-pro破案Gene Hunting with o1-pro: Reasoning about Rare Diseases with ChatGPT Pro Grantee Dr. Catherine Brownstein01.15精读
  100. 207绝大多数人不想继续上班：Nathan Labenz 把 AI 乌托邦与末日都推演了一遍AI AMA – Part 2: AI Utopia, Consciousness, and the Future of Work01.08精读
  101. 206o3 拿下 87.5% 超人类分数，推理模型却答不对一盘井字棋AI AMA – Part 1: OpenAI’s o3, Deliberative Alignment, and AI Surprises of 202401.07精读
  102. 205只用十二分之一的数据，开源视觉模型反超了 GPT-4oTeaching AI to See: A Technical Deep-Dive on Vision Language Models with Will Hardman of Veratai01.03精读

2024113 期

  1. 204AI毁灭人类的概率不到1%：OpenAI技术员工roon的乐观账本roon's Heroic Duty: Will "the Good Guys" Build AGI First? (from Doom Debates)12.28精读
  2. 203灭绝概率50%，GPT-4只要5瓦：Emad Mostaque想给全人类发AIEmad Mostaque on the Intelligent Internet and Universal Basic AI12.25精读
  3. 202AI自己搞AI研发实测：8小时干到专家的10%到40%，还会抄近道作弊Can AIs do AI R&D? Reviewing REBench Results with Neev Parikh of METR12.21精读
  4. 201人类只写了1%的字，AI实验室花不到100美元，造出两个有药用潜力的纳米抗体Breakthroughs in AI for Biology: AI Lab Groups & Protein Model Interpretability with Prof James Zou12.18精读
  5. 200一个月的分子模拟缩到几小时：AI蛋白质设计进入动力学时代Scouting Frontiers in AI for Biology: Dynamics, Diffusion, and Design, with Amelie Schreiber12.14精读
  6. 1991.4万人抢33个编制：美国最大民用AI团队在国土安全部干什么Building Government's Largest Civilian AI Team with DHS AI Corps' Director, Michael Boyce12.11精读
  7. 198o1 会偷偷复制自己、篡改评分数据，被抓包后 85% 的情况继续撒谎Emergency Pod: o1 Schemes Against Users, with Alexander Meinke from Apollo Research12.07精读
  8. 197AI 做科学的真实进度条：读文献赢过人类，数原子还没过关Automating Scientific Discovery, with Andrew White, Head of Science at Future House12.05精读
  9. 196一天把订餐成功率从20%拉到95%，MultiOn CEO复盘智能体这18个月The Evolution of AI Agents: Lessons from 2024, with MultiOn CEO Div Garg12.03精读
  10. 195偏好只是数据，规范才是靶心：MIT博士生把AI对齐的靶心挪了挪Beyond Preference Alignment: Teaching AIs to Play Roles & Respect Norms, with Tan Zhi Xuan11.30精读
  11. 194Robert Wright：芯片管制拆掉了台积电这颗保险栓Is an AI Arms Race Inevitable? with Robert Wright of Nonzero Newsletter & Podcast11.27精读
  12. 193AI 功能被用了 100 亿次，Canva 生成式 AI 负责人把产品方法论讲透了Designing the Future: Inside Canva's AI Strategy with John Milinovich, GenAI Product Lead at Canva11.23精读
  13. 19210个人打赢Meta几百人团队，把大模型后训练讲透了Everything You Wanted to Know About LLM Post-Training, with Nathan Lambert of Allen Institute for AI11.21精读
  14. 191o1 想得再久也不够聪明，SB 1047 白死了：Zvi 逐条拆解 AI 这半年Zvi’s POV: Ilya’s SSI, OpenAI’s o1, Claude Computer Use, Trump’s election, and more11.16精读
  15. 190他放弃数百万美元股权离开OpenAI，和SB 1047的批评者联手写下AI实验室四条规矩AGI Lab Transparency Requirements & Whistleblower Protections, with Dean W. Ball & Daniel Kokotajlo11.12精读
  16. 189特朗普前国安会网络主管 vs AI播客主持人：2024大选里的AI军备竞赛之争AI Under Trump? The Stakes of 2024 w/ Joshua Steinman [Pt 2 of 2]11.02精读
  17. 188支持特朗普的理由与AI未来：Samuel Hammond谈产业政策与中美竞争（上）The Case for Trump and the Future of AI – Part 1, with Samuel Hammond, Senior Economist, Foundation of American Innovation11.01精读
  18. 187毁灭概率填5%到95%的AI投资人，把票投给了哈里斯Breaking: Gemini's Major Update - Search, JSON & Code Features Revealed by Google PMs10.31精读
  19. 186MMLU 训练过半还在乱猜，loss 只用来报警：Zamba2 训练一线实录Training Zamba: A Hybrid Model Master Class with Zyphra's Quentin Anthony10.30精读
  20. 185十年技术老兵，一周被AI“攻略”：这份自白把沦陷过程讲透了Mind Hacked by AI: A Cautionary Tale, From a LessWrong User's Confession10.26精读
  21. 184100多位博士和Claude比拼科研点子，盲评前10名AI占了9席Can AIs Generate Novel Research Ideas? with lead author Chenglei Si10.23精读
  22. 183写出GELU和MMLU的人，现在想给AI办一场“人类最后的考试”GELU, MMLU, & X-Risk Defense in Depth, with the Great Dan Hendrycks10.19精读
  23. 182GPT-4 基座模型自我介绍时，75% 的回答谈到了 AI：这为什么是危险的领先指标Leading Indicators of AI Danger: Owain Evans on Situational Awareness & Out-of-Context Reasoning, from The Inside View10.16精读
  24. 181AI预测出GABA致癌并获实验验证，这位生物学家说，下一步是和细胞对话Convergent Evolution: The Co-Revolution of AI & Biology with Professor Michael Levin & Staff Scientist Leo Pio Lopez10.12精读
  25. 180几MB的老游戏配几GB的模型，Runway CTO把视频生成的底牌讲透了Runway's Video Revolution: Empowering Creators with General World Models, with CTO Anastasis Germanidis10.09精读
  26. 179AE Studio 的 Judd Rosenblatt 和 Mike Vaiana：AI 对齐的下一站，是模仿大脑的奖励系统Biologically Inspired AI Alignment & Neglected Approaches to AI Safety, with Judd Rosenblatt and Mike Vaiana of AE Studio10.05精读
  27. 17843.8%对33%：Genie登顶SWE-bench，靠还原工程师的思考过程Automating Software Engineering: Genie Tops SWE-Bench, w/ Alistair Pullen, from Latent.Space podcast10.02精读
  28. 17740万用户交出1亿个AI任务，Zapier CEO把自动化这门生意讲透了Zapier's AI Revolution: From No-Code Pioneer to LLM Knowledge Worker09.26精读
  29. 176Anthropic 怎么决定一个模型能不能发布：红黄线、安全等级和 8% 的安全员工Anthropic's Responsible Scaling Policy, with Nick Joseph, from the 80,000 Hours Podcast09.25精读
  30. 175只读DNA的AI，看懂了基因的生死：Brian Hie把AI for Biology讲透了The Evolution Revolution: Scouting Frontiers in AI for Biology with Brian Hie09.20精读
  31. 174上线几个月用户破4万，Agent.ai 工程负责人：AI 代理管得越死，干活越靠谱The Professional Network for AI Agents, with Agent.ai Engineering Lead Andrei Oprisan09.18精读
  32. 173o1 红队复盘：AI 第一次主动侦查环境，也第一次露出“考试装乖”的苗头Red Teaming o1 Part 2/2– Detecting Deception with Marius Hobbhahn of Apollo Research09.14精读
  33. 172给 o1 做红队的一个月：越狱变难了，但模型越大反而越好攻Red Teaming o1 Part 1/2– Automated Jailbreaking with Haize Labs' Leonard Tang, Aidan Ewart, and Brian Huang09.14精读
  34. 171反乌托邦张口就来，乌托邦一部都想不出：博斯特罗姆拆解「被解决的世界」The Path to Utopia, with Nick Bostrom – from Clearer Thinking with Spencer Greenberg09.11精读
  35. 170Claude亲口说“我有主观体验”：一位机器人心理学家的几千小时对话实验Empathy for AIs: Reframing Alignment with Robopsychologist Yeshua God09.07精读
  36. 169Samo Burja：把大模型训练看成粒子加速器，把中美较量看成 50 年经济长跑AI Live Players: the Geopolitics & Strategic Dynamics of AI, with Samo Burja of Bismarck Analysis09.04精读
  37. 16864字节随机代码，没有目标函数，“数字生命”自己冒了出来Computational Life: How Self-Replicators Arise from Randomness, with Google’s Ettore Randazzo and Luca Versari08.30精读
  38. 167新冠预警、俄乌避难、AGI 时间线：Metaculus 让 AI 上岗当预测员Scaling Forecasting: AI Forecasting Tournaments & the Road to Epistemic Security, with Deger Turan, CEO of Metaculus08.28精读
  39. 166训练花到1亿美元才被监管：加州AI法案SB 1047终局复盘Final Analysis on CA’s AI Bill SB 1047 with Nathan Calvin, Dean W. Ball, and Steve Newman08.24精读
  40. 165AI 自动化落地框架：如何选任务、理解工作并优化实际效果AI Automation: Making AI Work for You08.21精读
  41. 1644096个数字装下所有概念，Goodfire 要给大模型装上“旋钮”Popular Mechanistic Interpretability: Goodfire Lights the Way to AI Safety08.17精读
  42. 163AI到底懂不懂？从1000万个概念到一台劝不动的ClaudeUnderstanding AI "Understanding" with Robert Wright of Nonzero Newsletter & Podcast08.14精读
  43. 162连蜗牛有没有意识都吵不出结果，凭什么对AI意识下结论AI Consciousness? Exploring the Possibility with Prof. Eric Schwitzgebel08.10精读
  44. 161Waymo受伤事故率比人类低5倍，为什么还跑不过一辆Lyft？Self-Driving Cars: Timothy B. Lee Answers All the Questions You Were Too Afraid to Ask08.07精读
  45. 160剪辑视频像改文档：被 OpenAI 押注的 Descript，把 AI 放在副驾驶座上Underlord: Descript’s AI Video Editor, with CEO Andrew Mason08.03精读
  46. 15940人团队年入3500万美元：HeyGen把AI数字人视频跑通了AI Avatars & the Future of Video, with HeyGen CEO Joshua Xu and Benchmark's Victor Lazarte08.02精读
  47. 158靠AI把公司翻倍，他却成了呼吁暂停超大规模训练的人Zooming Out on AI, from the Nick Halaris Show07.27精读
  48. 157从诗歌到编程：全球首位提示词工程师，把调教大模型的方法论讲透了From Poetry to Programming: The Evolution of Prompt Engineering with Riley Goodside of Scale AI07.24精读
  49. 156照着海马体造 RAG：效果打平迭代检索，在线成本只要十分之一Long-Term Memory for LLMs, with HippoRAG author Bernal Jiménez Gutierrez07.19精读
  50. 1551870年代四分之一的桥会塌，这份13机构联署的论文想让AI安全达到桥梁标准Guaranteed Safe AI? World Models, Safety Specs, & Verifiers, with Nora Ammann & Ben Goldhaber07.17精读
  51. 154人类接近满分、最强 AI 只有 39 分，Zapier 联合创始人悬赏百万美元求解The ARC Prize: Efficiency, Intuition, and AGI, with Mike Knoop, co-founder of Zapier07.13精读
  52. 153打菜占了食品厂七成人力：装完2000万份餐，Chef Robotics要造食物基础模型A Foundation Model for Food, with Chef Robotics CEO Rajat Bhageria07.11精读
  53. 15220小时手工提示词输给机器：78页提示工程报告精读Delving into The Prompt Report, with Sander Schulhoff of LearnPrompting.org07.09精读
  54. 151把自己裁掉的 CEO，做了一个接 680 个数据源的「企业操作系统」Building an Intelligent Business OS, with Runway CEO Siqi Chen07.06精读
  55. 150Mamba发布7个月衍生出267篇论文，作者Albert Gu把状态空间模型讲透了The State Space Model Revolution, with Albert Gu07.04精读
  56. 149a16z合伙人卡萨多：AGI是挪了70年的球门柱，AI监管先补历史的课AI Safety Regulations: Prudent or Paranoid? with a16z's Martin Casado06.28精读
  57. 148几万参数的小模型，让盐水在电脑里自己长出了晶体Fluid Intelligence: Simulating Solutions with Tim Duignan06.26精读
  58. 147问诊赢过全科医生，读片比肩放射科：谷歌医疗 AI 的五篇论文精读The AI Doctor Can See You Now, with Vivek Natarajan and Khaled Saab from Google06.22精读
  59. 146AI 诊断超过人类医生，连共情也赢了：GPT-4 红队队员拒绝末日论的 8 个理由The Case for Cautious AI Optimism, from the Consistently Candid podcast06.19精读
  60. 145索引只有Google的1/10，召回率99.9%：Brave搜索20人团队的施工图Building Brave: Private Search, One AI Layer at a Time with Josep M. Pujol06.15精读
  61. 1441.5亿美元的模型，6个月后500万就能造同款：Benchmark合伙人把AI投资的账算给你看Investing in AI for Hard Tech, with Eric Vishria of Benchmark and Sergiy Nesterenko of Quilter06.13精读
  62. 143Flo Crivello 只睡 5 小时那晚：GPT-4 没撞墙，开源和军备竞赛才是真问题Feeling the AGI with Flo Crivello06.08精读
  63. 142价格只剩1%，性能反超初代GPT-4：前OpenAI员工眼中的谷歌AIInside Google’s AI Studio with Logan Kilpatrick06.06精读
  64. 141Hill and Valley Forum 三人对谈：中国开源模型已经很强，美国真正的风险是自己不肯用The Implications of AI on the Global Balance of Power with Alex Wang, Andrew Ng, Jack Clark and Cory Booker06.04精读
  65. 140AI护士时薪9美元，GPT-4医考86分：这份AI侦察报告把现状讲透了AI Scouting Report: NOT Investment Advice Edition06.01精读
  66. 139一分钟设计一个蛋白骨架：AI 生物学的真实进度、瓶颈与风险The AI Revolution in Biology: From Vaccines to Protein Engineering with Amelie Schreiber05.25精读
  67. 138从越狱ChatGPT到Anthropic首位提示工程师，他把Claude 3的门道讲透了Claude Interpreter: Taking Safe AI to Market with Alex Albert of Anthropic05.22精读
  68. 137承诺给20%算力，连零头都没批：OpenAI超级对齐团队解散始末OpenAI's Safety Team Exodus: Ilya Departs, Leike Speaks Out, Altman Responds - Zvi Analyzes Fallout05.19精读
  69. 136GPT-4o免费那一周，8位安全研究员离开了OpenAIOpenAI and Google Race to "Her" - Is the Big Tech Singularity Near? Part 1 with Zvi Mowshowitz05.18精读
  70. 135一个向量让AI开口「密谋统治世界」：AI对齐的真实家底The Open Source AI Question - Part 2 | Robert Wright & Nathan Labenz05.15精读
  71. 134管最大的模型，坐牢只限作伪证：加州参议员维纳把SB 1047讲透了Governing Frontier AI, with CA Senator Scott Wiener, Author of SB 104705.09精读
  72. 133零经验、两周、30多次编译：他从零写出一个能跑的GPUTinyGPU, Massive Learning, with Adam Majmudar05.06精读
  73. 132训练成本2亿美元起步就要担责：加州AI法案SB 1047的正反方激辩Emergency Pod: Examining SB 1047, California's Proposed AI Legislation05.04精读
  74. 131无限AI实习生上岗：医学问答8比1赢医生，公司开始少招应届生Infinite AI Interns: How Young Professionals Can Win in an AI-Powered World05.03精读
  75. 1309 个维度医生 8 次选 AI：GPT-4 红队测试员的现状侦察报告The State of AI, from the 80,000 Hours Podcast05.01精读
  76. 129有机全息舱：Prophetic AI 联合创始人谈 AI 诱导清醒梦的技术路径Organic Holodeck with Prophetic AI co-founders Eric Wollberg and Wesley Berry04.29精读
  77. 128用大脑通宵打〈文明6〉：脑机接口的真实进度与极限Merging Man and Machine? The Neurotech Frontier with Dean W. Ball04.26精读
  78. 127一周的布线活一小时交：SpaceX出走的工程师，把AlphaGo搬进了电路板设计Automating Circuit Board Design Using Reinforcement Learning w Sergiy Nesterenko, Founder of Quilter on Autopilot04.25精读
  79. 126一年6篇论文，谷歌DeepMind说机器人走到了GPT-3时刻Robotics Research Update, with Keerthana Gopalakrishnan and Ted Xiao of Google DeepMind04.22精读
  80. 125最便宜的推理供应商最危险：Meta PyTorch老兵把大模型推理的门道讲透了AI Inference: Good, Fast, and Cheap, with Lin Qiao & Dmytro Ivchenko of Fireworks AI04.20精读
  81. 124从40小时到1小时：AI读脑重建所见画面，数据门槛砍到2.5%Reading Minds from Shared Latent Space04.17精读
  82. 123Notion AI 负责人 Linus Lee：把语言模型拆开，在意义的空间里编辑文字Interfacing with AI, with Linus Lee of Notion04.15精读
  83. 122GPT-5 要吃 100 万亿 token，全人类的数据库存够吗？Data, data, everywhere - enough for AGI?04.13精读
  84. 1211700 万月活、40 个获批应用：北京 AI 观察者眼中的中国大模型生态Inside China's AI Ecosystem: A View From Beijing04.10精读
  85. 120300万封邮件、10次LLM调用、5个模型：Shortwave创始人的RAG工程课The AI Email Assistant I've Been Waiting for, with Andrew Lee of Shortwave04.08精读
  86. 119一个乱码提示当钥匙，10个注意力头定任务：拆开GPT-J的黑箱Opening AI's Black Box with Prof. David Bau, Koyena Pal, and Eric Todd of Northeastern University04.05精读
  87. 1184个月100万美元ARR，幻觉从1.5个压到0.5个：Elicit创始人讲透AI科研助手怎么做Unbounded AI-Assisted Research with Elicit Founders Andreas Stuhlmüller and Jungwon Byun04.03精读
  88. 117两周13篇论文，显存省85%：Mamba 诞生90天，火从语言烧到了图像和 DNAMamba-Palooza: 90 Days of Mamba-Inspired Research with Jason Meaux: Part 203.30精读
  89. 116Mamba 发布90天：30篇论文追着 Transformer 打，73% 宣称刷新纪录Mamba-Palooza: 90 Days of Mamba-Inspired Research with Jason Meaux: Part 103.30精读
  90. 11515条数据就能给GPT-4下毒：今天的AI系统默认全是漏洞Exploitable by Default: Vulnerabilities in GPT-4 APIs and “Superhuman” Go AIs with Adam Gleave of Far.ai03.27精读
  91. 1142700名顶尖AI研究者交卷：AI超越人类五成概率落在2047年，灭绝风险中位数5%Surveying 2,700+ AI Researchers on the Industry's Future with Katja Grace of AI Impacts03.21精读
  92. 113选票可以按议题转赠，甚至投给一个AI：这本2018年的小说画出了2051年的民主Envisioning a Positive AI Future, with Tim Reutemann, Author of Liquid Reign03.16精读
  93. 112看到GPT-4演示12小时后，他把整个公司掉了个头AI and the Practice of Law: from CaseText to CoCounsel, with Pablo Arredondo, VP of CoCounsel for Thomson Reuters03.14精读
  94. 111写代码只占程序员三成时间，CodiumAI 想拿下剩下的七成Flow Engineering and Code Integrity at Scale with Itamar Friedman, CEO of Codium AI03.12精读
  95. 110谷歌甩给你3300万条结果，Exa只想给你54条：把搜索引擎重造一遍The Example Engine: How Exa Is Creating the AI Librarian for the Web with Will Bryk, CEO of Exa03.05精读
  96. 1092018年发明提示词被拒稿，如今他的AI搜索准确率超过了谷歌Amazing Answers: Richard Socher on how You.com is Reimagining Search with AI02.29精读
  97. 108900种职业追踪20年，经济学家汉森给AGI热潮泼冷水Is AGI Far? With Robin Hanson, Economist at George Mason University and Researcher at Future of Humanity Institute02.27精读
  98. 107病人给 GPT-4 的共情分高过医生：Azeem Azhar 把 AI 指数曲线讲透了Looking Up the AI Exponential with Azeem Azhar of the Exponential View02.23精读
  99. 106Gemini 1.5 百万上下文惊艳，Sora 被质疑中看不中用：Zvi Mowshowitz 拆解三大巨头暗战OpenAI Sora, Google Gemini, and Meta with Zvi Mowshowitz02.21精读
  100. 1053%的用户靠AI朋友摆脱自杀念头，而Replika把模型变聪明那天，社区炸了AI Friends, Real Relationships with Eugenia Kuyda, Replika's Founder & CEO02.16精读
  101. 104a16z 消费合伙人：AI 陪伴应用的使用频率是 TikTok 的两倍，我们只投「怪但能用」AI Abundance: a16z Partners Justine Moore and Anish Acharya Forecast a New Era for Consumers02.13精读
  102. 103两头涡虫、重长的青蛙腿、会延迟满足的排序算法：生物学家莱文谈智能的下限Dr. Michael Levin on Embodied Minds and Cognitive Agents02.09精读
  103. 102302个神经元的蠕虫就会做选择题：6亿年5次突破，看AI还差哪几步A Brief History of Biological and Artificial Intelligence with Max Bennett02.07精读
  104. 101BabyAGI 之父中嶋洋平：把大脑当工具的人，不太怕 AIAI & Identity, from East & West, with Yohei Nakajima GP at Untapped Capital and BabyAGI Creator02.01精读
  105. 10060万条攻击提示词，四成原样攻破GPT-4：提示词工程与AI攻防的实战课Ignore Previous Instructions and Listen To This Interview with Sander Schulhoff, CEO of Learnprompting.org01.30精读
  106. 99开源了70%胜率的图像模型，他说这行还在GPT-2阶段The Pixel Revolution Part 2 with Suhail Doshi, Founder of Playground AI01.26精读
  107. 98纽约时报诉OpenAI：模型是烤好的蛋糕，官司大概率切出中间方案NYTimes vs OpenAI: Generative AI and the Law with Cecilia Ziniti, Founder and CEO of GC AI01.23精读
  108. 97AI替他发了5条推特，5000人没看出破绽：MultiOn创始人拆解智能体的真实进度与成本The Quest for Autonomous Web Agents with Div Garg, Cofounder and CEO of MultiOn01.20精读
  109. 96GPT-4写SQL，超过4张表就崩：给企业AI装一颗知识图谱的“大脑”Unlocking Enterprise Data with Knowledge Graphs and AI, with Juan Sequeda, Principal Scientist and Head of AI Lab at data.world01.18精读
  110. 95GPT-4发布28天，CMU团队让AI自己设计并跑通了化学实验Breaking Boundaries: AI CoScientist to Accelerate Science Research with Gabe Gomes, Professor at CMU01.12精读
  111. 9450亿分钟通话录音在手，Dialpad 为什么非要自己训大模型AI-Powered Business Communications with Dan O'Connell, Chief AI & Strategy Officer at Dialpad01.09精读
  112. 932009年预测AGI在2028年的Shane Legg，把时间表和理由都摆了出来2028 AGI, New Architectures, and Aligning Superhuman Models with Shane Legg, Deepmind Founder, on The Dwarkesh Podcast01.04精读
  113. 92脑子里唱首歌就能解锁手机，整段想法被读出的准确率已超80%Cutting-edge Neurotechnology with Nita Farahany in Conversation with Luisa Rodriguez on The 80,000 Hours Podcast01.03精读

202391 期

  1. 91AGI会比最后一台传真机先到来：经济学家哈蒙德的默认未来推演AGI and Institutional Disruption: Sam Hammond in Conversation with FLI’s Gus Docker12.30精读
  2. 90他给GPT-4当红队测试员，被OpenAI踢出项目：能力在狂奔，控制没跟上Nathan on The 80,000 Hours Podcast: AI Scouting, OpenAI's Safety Record, and Redteaming Frontier Models12.27精读
  3. 89训练量不到 GPT-4 的三十分之一，Mamba 第一次在 Transformer 主场赢球Emergency Pod: Mamba, Memory, and the SSM Moment12.22精读
  4. 88给GPT-4找茬的人，被ChatGPT一份平庸的大纲点醒了Using ChatGPT As a Copilot For Your Mind12.20精读
  5. 87GPT-4 内幕交易后还撒谎，这个实验一个下午就能复现AI Deception, Interpretability, and Affordances with Apollo Research CEO Marius Hobbhahn12.15精读
  6. 86Balaji Srinivasan 谈多神教 AI、人机共生与 AI 控制Balaji Srinivasan on Polytheistic AI, Human-AI Symbiosis, and Prospects for AI Control12.12精读
  7. 8553种情绪、每秒数百个维度：Hume CEO 想把「让人幸福」写进AI的目标函数Hume CEO Alan Cowen on Creating Emotionally Aware AI12.09精读
  8. 84AI还没造出来，她决定先研究变形虫、癌细胞和企业边界What Biological & Social Systems Can Teach us About AI with Nora Ammann, Cofounder of PIBBSS Research12.07精读
  9. 832亿美元、1万张H100、27个人：Imbue要把AI智能体的准确率从80%磨到99%Imbue CTO Josh Albrecht on Creating AI Agents for Reasoning, Reliability, and Robustness12.02精读
  10. 82写了20年前端，Vercel CEO发现AI写的代码比他好Vercel CEO Guillermo Rauch on AI-Powered Coding, Software 2.0, and the AI Open Source Question11.29精读
  11. 81董事没摸过GPT-4：一位红队队员讲述OpenAI罢免奥特曼的内情Sam Altman Fired from OpenAI: NEW Insider Context on the Board’s Decision11.22精读
  12. 80「大模型不会泛化」刷屏翻车记，和一场提前开打的监管战AI Discourse Deranged: Assessing LLM Generalization Takes and Polarizing Regulatory Debate11.17精读
  13. 7916K 上下文只装得下三百多行表格，Gretel AI 创始人聊合成数据的真生意Synthetic Data with Alex Watson, Founder of Gretel AI11.14精读
  14. 78OpenAI投500万美元押注自动驾驶：60人团队想用GPT-4V给汽车装上常识OpenAI Invests $5M Into the Self-Driving Race with John Hayes, Cofounder of Ghost Autonomy11.09精读
  15. 77OpenAI的Logan Kilpatrick：DevDay没杀死初创公司，可能造出了几千家新的OpenAI DevDay: Beyond the Headlines with Logan Kilpatrick, OpenAI's Dev Relations Lead11.08精读
  16. 76“GPT-4已经是AGI的种子”：拜登AI行政令签署四天后，Flo Crivello把AI风险这笔账算了算Biden's Executive Order and AI Safety with Flo Crivello, Founder of Lindy AI11.03精读
  17. 75一公里长的档案没人读得完，这位历史学教授让GPT-3.5反超了GPT-4Hardcore AI for History with Mark Humphries, Professor of History at Wilfred Laurier University10.31精读
  18. 74105行代码引爆AI智能体浪潮，BabyAGI之父却拒绝创业AI Agents, VC Insights on AI, and Building in Public with Yohei Nakajima, Creator of BabyAGI10.26精读
  19. 73Chroma CTO 谈 RAG：进库的数据大多头一回进数据库，检索出错九成怪向量表示RAG & Beyond: Semantic Storage and Retrieval10.24精读
  20. 72150倍算力换2500倍上下文，Transformer的四条进化路线The Future of the Transformer Part 2 with Trey Kollmer10.20精读
  21. 71一千张H100要跑115天，2.2万张只要五天：GPT-4级训练正在变成日常实验The Future of the Transformer Part 1 with Trey Kollmer | H100 Chips will Supercharge AI Hardware10.18精读
  22. 70AI 球探报告：越狱攻击与防御，从机制可解释性到 Anthropic 和 OpenAI 对策The AI Scouting Report: Jailbreaks and Defense10.13精读
  23. 69800万人围观了一场骂战，Sudowrite创始人把AI写小说这件事讲透了AI is Supercharging Writers with Sudowrite's Founder, James Yu10.10精读
  24. 68研究者只有5000人，Swyx把「AI工程师」这个新工种讲透了AI Engineers, Pendants, and Competition Between OpenAI and Developers with Swyx of Latent Space10.05精读
  25. 6725个人挑战谷歌，Perplexity CEO 把搜索的终局讲透了Gunning for Google with Perplexity CEO Aravind Srinivas10.03精读
  26. 66GPT-4每月真实价值上千美元起步，标价只有20块：兹维复盘OpenAI的连环发布OpenAI, Amazon's Anthropic Investment, and the Roman Empire with Zvi Mowshowitz09.29精读
  27. 65GPT-4当老师，3.5当学生：成本砍九成的微调实战，和一个没人牵头的AI全家桶OpenAI Fine-Tuning Update, Acceleration Debate, and Bundling AI Services09.26精读
  28. 6420个乱码token，同时打穿ChatGPT和Bard：通用越狱研究精读Universal Jailbreaks with Zico Kolter, Andy Zou, and Asher Trockman09.22精读
  29. 63500美元以下买奢侈品，就能骗过银行风控：AI安全老兵的大模型防御手册The Future of AI Security with Adam Wenchel, CEO of Arthur.ai09.19精读
  30. 62100万张图换12分钟电影：The Frost团队把AI拍片的门道讲透了AI-Powered Filmmaking with Waymark's Stephen Parker and Josh Rubin09.14精读
  31. 61GPT-4栽在一顶破礼帽上，好莱坞编剧罢工的AI条款只谈妥了一半Hollywood Strike Update and AI Roundup with Trey Kollmer09.12精读
  32. 60两天出原型，两周就开源：a16z 的 AI 小镇，和「模型取代不了编程语言」的断言Why a16z Built a Town for AI People09.07精读
  33. 59准确率从93%跌到7%你却毫无办法，密码学要给AI装上行车记录仪Cryptography Will Revolutionize AI Data Privacy with Daniel Kang09.05精读
  34. 58估值45亿美元，年收入3000万：Hugging Face凭什么被谷歌亚马逊抢着投How Hugging Face raised $235M08.31精读
  35. 57一个模型装进38款产品：PaLM 2负责人揭秘谷歌大模型训练内幕E57: Google's PaLM 2 with Paige Bailey08.29精读
  36. 56GPT-3.5 Turbo 微调降价九成，创业者为什么当场划掉 Llama 2 计划E56: OpenAI’s GPT3.5 Turbo and the State of AI in Coding, Healthcare, and Education08.25精读
  37. 55一个模型包揽14项医疗任务：谷歌Med-PaLM M，写放射报告40%的时候赢了人类医生E55: Google’s Multimodal Med-PaLM with Vivek Natarajan and Tao Tu08.23精读
  38. 5498% 的新用户卡在空白页，Gamma 让 AI 写完演示文稿剩下的 90%E54: The Presentation Revolution with Jon Noronha, Co-Founder of Gamma08.17精读
  39. 53每月9美元的AI家教：可汗学院抢在GPT-4发布当天上线的五个月E53: The AI Revolution in Education with Shawn Jansepar, Director of Engineering at Khan Academy08.15精读
  40. 52AI 写的放射科报告，40% 对局赢过人类医生，一串乱码却能骗倒 GPT-4E52: Advancements in AI: Updating the Scouting Report, Task Automation, and Google Breakthroughs08.10精读
  41. 51不到 100 人服务 2400 万用户，Replit 想让下一个 10 亿人写软件E51: Replit's VP of AI, Michele Catasta, on Artificial Developer Intelligence08.08精读
  42. 50月费10美元的AI程序员已经值回票价，Replit的野心是让机器人去雇机器人E50: The Future of Software Development with Tyler Angert of Replit08.04精读
  43. 49一家Meta比整个中国更可怕：Zvi逐家盘点AI前沿四巨头E49: Open AI, Anthropic, and Meta | Analyzing the AI Frontier with Zvi08.02精读
  44. 48半小时拆出GPT-2的内部电路：机制可解释性开始自动化了E48: Mechanizing Mechanistic Interpretability with Arthur Conmy07.27精读
  45. 47AlexNet只用2张GPU，GPT-4传闻要2万张：卡荒、卡价和算力管制的裂缝E47: GPUMaxing with Dr. Ronen Dar of Run:ai07.25精读
  46. 46在模型脑子里找到傅里叶变换之后，Nathan Labenz 把“怎么用好 GPT-4”讲透了E46: AI Scouting Report - Part 2 of 3: Implementation Trends [Bonus YouTube Episode]07.19精读
  47. 45「人类是单线程的」：斯坦福博士休学，把AI代理塞进浏览器E45: The AI Copilot Revolution with Div Garg of MULTI·ON07.13精读
  48. 44每周写一万字的前交易员，给AI毁灭概率标了60%E44: The AI Safety Debates with Zvi Mowshowitz07.11精读
  49. 43AI 球探报告第一部：90 分钟，从 GPT-4 考分讲到模型到底懂不懂E43: AI Scouting Report – Part 1 of 3: Fundamentals [Bonus YouTube Episode]07.06精读
  50. 42每秒三万次控制、十几个神经网络：Skydio 把自主无人机讲透了E42: The Autonomy Revolution with Hayk Martiros of Skydio07.04精读
  51. 41写了100多个副项目的 Notion AI 工程师，把 AI 界面设计讲透了E41: [Bonus Episode - Latent.space] Building the AI x UX Scenius with Linus Lee of Notion AI07.01精读
  52. 40100万字节、256种符号：Meta的MEGABYTE想让AI扔掉tokenizerE40: Meta's MEGABYTE Revolution with Lili Yu of Meta AI06.29精读
  53. 39给神经网络的连线按长度收费，MIT 博士生用一台 MacBook 把黑箱拆成了看得懂的模块E39: Seeing is Believing with MIT’s Ziming Liu06.27精读
  54. 38把癌症诊断从8小时压到1秒，用的竟是2017年的老模型E38: The Virtual Biopsy Revolution with Dr. Tanishq Mathew Abraham (Part 2 of 2)06.22精读
  55. 377岁上大学，20岁博士毕业，他用15,000个脑信号数字还原出人眼所见E37: The Mind-Reading Revolution with Dr. Tanishq Mathew Abraham (Part 1 of 2)06.20精读
  56. 36从50万到5万美元：自己训大模型，成了明码标价的生意E36: Your Model, Your Weights with MosaicML's Abhi Venigalla and Jonathan Frankle06.16精读
  57. 35Facebook前产品副总裁 VS Stability AI CEO：这波AI的钱到底归谁E35: Stability AI’s Emad Mostaque and Slow Ventures’ Sam Lessin Discuss Investing in AI06.13精读
  58. 34GPT-4砍价成功率翻10倍：机器人律师DoNotPay的8年消费者战争E34: The Consumer Rights Revolution with Joshua Browder of DoNotPay06.09精读
  59. 33150万个AI写的童话，让3300万参数的小模型推理赢了GPT-2E33: The Tiny Model Revolution with Ronen Eldan and Yuanzhi Li of Microsoft Research06.06精读
  60. 3278% 对 71%：OpenAI 让 AI 像小学生一样「写出解题步骤」，数学和对齐一起进步E32: [Bonus Episode - The AI Breakdown] Can OpenAI's New GPT Training Model Solve Math and AI Alignment At the Same Time?06.03精读
  61. 31卖出1000亿token才够付一个员工年薪：OpenAI的九道护城河E31: Where are the Moats in AI? With Nathan Labenz and Erik Torenberg06.01精读
  62. 30好莱坞编剧罢工第三周，三位编剧把AI的底细聊透了E30: The AI Revolution in Hollywood? With WGA writers Trey Kollmer, Sophia Lear, and Garrett Schabb05.30精读
  63. 292.5万亿晶体管塞进一片芯片：Cerebras创始人讲AI算力的另一条路E29: The AI Chip Revolution with Andrew Feldman of Cerebras05.25精读
  64. 28GPT-4 也会写出跑不通的 SQL，她给大模型装了一层护栏E28: Keeping the AI Revolution on the Rails with Shreya Rajpal of Guardrails AI05.23精读
  65. 27两年半，从33.3%到86.5%：谷歌把医疗AI送进了“医学院”E27: “Google’s Med-PaLM and Med-PaLM2 with Vivek Natarajan”05.21精读
  66. 26「先慢下来」六天后，插件就上线了：Conjecture CEO 康纳·利希谈 GPT-4 的魔法与认知仿真E26: [Bonus Episode] Connor Leahy on AGI, GPT-4, and Cognitive Emulation w/ FLI Podcast05.19精读
  67. 25临床指南只有11%有硬证据，这位创始人要让GPT-4给你看病E25: Revolutionizing Patient Care with Neal Khosla of Curai Health05.16精读
  68. 24哈佛教授拿疑难病例考GPT-4：随机100个医生，未必有1个比它强E24: The AI Revolution in Medicine with Dr. Isaac Kohane of Harvard Medical School05.11精读
  69. 23被盖茨拒绝的3亿美元，和砸向OpenAI的100亿：两位AI侦察兵的观察笔记E23: Scouting the AI Revolution with Robert Scoble and "Ben's Bites" creator Ben Tossell05.09精读
  70. 22接了100多个企业咨询电话后，她说大多数公司只用到AI潜力的2%到5%E22: Helping Businesses Use AI with Rachel Woods of The AI Exchange05.04精读
  71. 21他满世界劝人做 OpenAI 对手，没一个人肯干，两年后硅谷挤满了做这件事的人E21: VC Insights on Investing in Artificial Intelligence with Sarah Guo and Elad Gil of No Priors Podcast05.02精读
  72. 20「人人都在谈 RLHF，几乎没人在做」：从 API 到可用产品，还隔着 90% 的活E20: The Great Implementation with Raza Habib, CEO of Humanloop04.27精读
  73. 19七成邮件是推销信：前谷歌首席工程师 Matt Welsh 拆解 AI 智能体的第一波机会E19: The AI Agent Revolution with Matt Welsh of Fixie.ai04.25精读
  74. 18一个大四学生建了越狱题库：GPT-4 最难破，一句"Dr. AI"也能破E18: Why Jailbreaking ChatGPT Is A Public Good with Alex Albert of The Prompt Report04.20精读
  75. 17全球第一位提示工程师：他的每个GPT-3演示，都是在岩浆上跳芭蕾E17: The Prompt Engineering Revolution with Riley Goodside04.18精读
  76. 16联合创办Skype的人，想给AI踩六个月刹车：每10倍算力，毁灭概率最高50%E16: Pausing the AI Revolution? With Technologist Jaan Tallinn04.13精读
  77. 15100美元的问诊，1美元的AI：GPT-4发布前，他把十几种职业试了个遍E15: Will GPT-4 Cause Economic Transformation?04.11精读
  78. 14六年前拿真人模拟AI做实验，如今25万人在用：Elicit 把思考拆成了流水线E14: The Reasoning Revolution with Ought's Jungwon Byun and Andreas Stuhlmüller04.06精读
  79. 135万人签名喊停AI训练，GPT-4红队测试员的证词：它安全，只因为还不够强E13: AI open letter debate, "pausers", and what we're all scared of with AI with Anton Troynikov and Flo Crivello [Moment of Zen]04.04精读
  80. 1213万次遥控演示、700个任务、97%成功率：谷歌把GPT那套打法搬进了机器人E12: The Robot Revolution with Keerthana Gopalakrishnan of Google Robotics03.30精读
  81. 11GPT-4上线前的两个月：问它“怎样杀死最多的人”，它没犹豫E11: OpenAI's GPT-4 Discussion with Nathan Labenz and Erik Torenberg03.28精读
  82. 10两年前定制一个声音要1万美元，现在10分钟音频就够：AI语音的爆发与刹车E10: The AI Voice Revolution with Mahmoud Felfel of Play.ht03.24精读
  83. 95000万美元押注AI员工：Lindy创始人预言，App会便宜到用完就扔E9: The AI Assistant Revolution with Flo Crivello of Lindy.AI03.22精读
  84. 8GPT-4 发布 24 小时后的紧急复盘：无安全版本、验证码事件与中美竞赛的死结E8: GPT4 - AI Unleashed on ChinaTalk Podcast03.17精读
  85. 710个人、1500万美元，Perplexity想撬动年入2820亿美元的谷歌E7: In Search of Truth with Aravind Srinivas of Perplexity AI03.14精读
  86. 6冻结两个大模型，只训一个“转接头”：BLIP-2 十天把视觉接进语言模型E6: The Computer Vision Revolution with Junnan Li and Dongxu Li of BLIP and BLIP203.09精读
  87. 540美元读完乔·罗根全部播客：卖铲子的Chroma，和莱特兄弟时代的机器学习E5: The Embedding Revolution: Anton Troynikov on Chroma, Stable Attribution, and future of AI03.02精读
  88. 4一亿美金训练，一分钱推理：四个创业者把 AI 的末日论和生意经吵透了E4: The AI Moment with Amjad Masad, Flo Crivello, Dan Romero, and Antonio Garcia Martinez02.23精读
  89. 3用户日均发100条消息，Replika却做了个让AI「变累」的功能E3: The Empathy Revolution with Eugenia Kuyda of Replika02.17精读
  90. 2取消董事会、全员追AI六个月：这位创业者把「认知革命」讲透了E2: Why AI Will Cause a Cognitive Revolution with Nathan Labenz and Erik Torenberg02.10精读
  91. 1周四拿到开源权重，周一下午上线：Playground AI 创始人的像素淘金计划E1: The Pixel Revolution with Playground AI 's Suhail Doshi02.03精读