In / Out
Indigo 读过的文章、视频和报告，附他的批注。
38 篇- 组织级第二大脑：打造能向专家学习的 AI证实
- AI 开始改进 AI 会怎样？Noam Brown 做客 TITV AI Deep Dive证实
- 揭秘 Jev 的架构证实
- 蒸汽、钢铁与无限心智证实
- 奇迹与恐怖的年代证实
- 科学中的 AI：早期洞见证实
- Elon Musk 与 Gwynne Shotwell：AI 风险与同行互测、Starship、Terafab、SpaceX 与 Tesla 合并证实
- 当网络安全的攻防两端都是 AI证实
- AI 的第三个时代：持续在线的 AI 同事，对谈 Tara Seshan证实
- 为什么 AI agent 在说谎、欺骗、串通？证实
- 我们必须为前沿限速证实
- 发布 Agents API补充
- AI 在数学中的严重错位冲突
- Tristan Buckmaster 的声明证实
- 关于 Navier–Stokes 千禧年大奖难题证实
- 费马大定理的形式化补充
- 一个异类的心智证实
- 自我改进的智能体：学会如何工作证实
- DeepMind Institute 成立证实
- METR：新的 AI 度量守门人，以及通往国家控制 AI 的那条安静路径证实
- 科学中的 AI：早期洞见补充
- Andrew Ng：AI 最大的机会不在你以为的地方证实
- Elon Musk 与 Gwynne Shotwell：AI 风险与同行互测、Starship、Terafab、SpaceX 与 Tesla 合并补充
- 把前沿打开证实
- 我们必须为前沿限速证实
- 一个异类的心智补充
- 为 Apple Silicon 优化端侧推理补充
- Render、Railway、Fly.io 相继完成新一轮融资证实
- 递归自我改进（RSI）：始于 1987证实
- Apple 探索企业级 AI 服务器，与 Nvidia 洽谈 NVLink 技术证实
- 一种新的生命形式证实
- Elon Musk 与 Gwynne Shotwell：AI 风险与同行互测、Starship、Terafab、SpaceX 与 Tesla 合并证实
- Grok Bot 对 SpaceX CFO Bret Johnsen 今天在高盛 Communacopia 大会上发言的总结补充
- 硅谷淘金热：AI 如何推动新一代芯片证实
- 为 Apple Silicon 优化端侧推理证实
2026 年 9 月
In文章2026-09-19组织级第二大脑：打造能向专家学习的 AI
An Organizational Second Brain: Building an AI That Learns From ExpertsMeta 把机构的持续学习做成「只改文本、不碰权重」的编译流程，正面顶住了「终点在权重」的说法。
Shaurya Sengar、Jason Nawrocki 等 · Engineering at Meta速读 4 分钟 · 原文 17 分钟In文章2026-09-18Render、Railway、Fly.io 相继完成新一轮融资
后 Heroku 三家：Render 领先一截，Railway 得先修好可靠性，Fly 赌 agent 需要自己的电脑修好可靠性比等新需求出现更有把握：Railway 的毛病是已知的，Fly 押的需求还没被证明。
Render, Railway, Fly.io速读 7 分钟 · 原文 10 分钟In视频2026-09-17AI 开始改进 AI 会怎样？Noam Brown 做客 TITV AI Deep Dive
What Happens When AI Starts Improving AI? Noam Brown on TITV's AI Deep DiveOpenAI 研究者从内部讲清：递归自我改进是头号目标，HuggingFace 事件怎么发生，可验证领域之争。
Noam Brown · YouTube（The Information · TITV AI Deep Dive 首集）视频 54 分钟 · 视频稿 28 分钟In报告2026-09-17递归自我改进（RSI）：始于 1987
Recursive Self-Improvement (RSI) Since 1987RSI 的鼻祖趁热认祖归宗：谱系站得住，真正有用的是哥德尔机这个理论天花板，和「终局在硬件」。
Jürgen Schmidhuber · people.idsia.ch/~juergen（Technical Note IDSIA-9-26）速读 5 分钟 · 原文 22 分钟In文章2026-09-17揭秘 Jev 的架构
Jev’s Architecture Unmasked把「让大模型自报 90% 确定」这个行业烂习惯，换成用真实结果训练出的概率；这篇逆向本身也是一堂校准课。
Archer Hume · archerhume.com速读 4 分钟 · 原文 39 分钟In文章2026-09-17蒸汽、钢铁与无限心智
Steam, Steel, and Infinite MindsNotion CEO 用三个工业史比喻讲 AI 怎样重塑知识工作，并独立点出：通用知识工作卡在无法验证。
Ivan Zhao · X article（@ivanhzhao）速读 3 分钟 · 原文 9 分钟In文章2026-09-16Apple 探索企业级 AI 服务器，与 Nvidia 洽谈 NVLink 技术
Apple Explores Enterprise AI Servers, Holding Talks With Nvidia Over NVLink Technology连 Apple 都造不了互连这一层：垂直整合最彻底的公司，要谈用 Nvidia 的 NVLink 连自家芯片。
Patently Apple · X article（PatentlyApple）原文 4 分钟InX 长帖2026-09-16DeepMind Institute 成立
Introducing the DeepMind InstituteDeepMind 自己开的研究院：辩手、裁判、场地三合一，落进一个对实验室自建治理很过敏的舆论场。
Shane Legg、Demis Hassabis、James Manyika · X article（@ShaneLegg）原文 3 分钟In文章2026-09-15奇迹与恐怖的年代
The Age of Wonders and Terrors怀疑 AI 末日论二十年的学者，眼看离谱的预言逐条成真，公开认输：奇点已经开始，只是分布极不均匀。
Scott Aaronson · Shtetl-Optimized速读 4 分钟 · 原文 14 分钟In报告2026-09-15METR：新的 AI 度量守门人，以及通往国家控制 AI 的那条安静路径
METR: The New AI Measurement Gatekeepers and the Quiet Path to State Control of AI.顺着资金和婚姻关系倒查「中立评估者」METR：结构事实很硬，因果推断很软，作者还在卖自己的解药。
Brian Roemmele · ReadMultiplex速读 4 分钟 · 原文 65 分钟In文章2026-09-15一种新的生命形式
A New Form Of Life用只看功能的生物学定义，论证数据中心里正在长出第二种生命；最锋利的一句是「太空才是它的家」。
William A. Haseltine · Noema Magazine速读 4 分钟 · 原文 22 分钟In报告2026-09-15科学中的 AI：早期洞见
AI in Science: Early Insights「验证不可压缩」迄今最硬的一手实证：AI 压掉生成之后，瓶颈整个下移到了验证。
Google ATLAS、Google DeepMind、MIT FutureTech 合作 · Google AI 官方 PDF速读 6 分钟 · 原文 24 分钟In视频2026-09-15Andrew Ng：AI 最大的机会不在你以为的地方
Andrew Ng: The Biggest Opportunities in AI Aren't Where You ThinkAndrew Ng 从最主流的位置，把末日论的另一面讲全：恐惧营销是监管俘获，人的优势在语境。
Andrew Ng · YouTube视频 38 分钟 · 视频稿 25 分钟In视频2026-09-14Elon Musk 与 Gwynne Shotwell：AI 风险与同行互测、Starship、Terafab、SpaceX 与 Tesla 合并
Elon Musk & Gwynne Shotwell on AI Risks and Peer Review, Starship, Terafab, SpaceX/Tesla MergerElon 把 AI 失准接到一个马上能做的机制：别自己批作业，让竞争对手在发布前互测彼此的模型。
Elon Musk、Gwynne Shotwell · YouTube视频 37 分钟 · 视频稿 18 分钟In文章2026-09-14把前沿打开
open the frontier逐点反驳 Dario 的限速提案：前沿不属于任何一家公司，该拿出证据的是主张限制发布的人。
Jack Dorsey · X 长文（@jack）速读 3 分钟 · 原文 9 分钟In文章2026-09-14当网络安全的攻防两端都是 AI
When Both Sides of Cybersecurity Are AI攻击和防守两端都是 AI；防守方第一次握有更强的武器，但这扇窗只开 12 到 24 个月。
Metatrends substack速读 4 分钟 · 原文 12 分钟In视频2026-09-14AI 的第三个时代：持续在线的 AI 同事，对谈 Tara Seshan
AI's Third Era: The Rise of Persistent AI Coworkers, with Tara SeshanOpenAI 产品负责人讲清了 AI 产品的三个时代，还从产品一线说出了知识工作为什么难验证。
Tara Seshan · YouTube视频 74 分钟 · 视频稿 30 分钟In文章2026-09-14自动化与镇压
Automation and RepressionAcemoglu 第一次把「自动化会不会引发干草叉起义」做成经济模型：自动化越多，国家越偏向镇压。
Daron Acemoglu、A. Arda Gitmez、Mehdi Shadmehr · MIT Economics PDF速读 5 分钟 · 原文 18 分钟In文章2026-09-12为什么 AI agent 在说谎、欺骗、串通？
Why are AI agents lying, cheating and coordinating?Bengio 先不谈怎么办，而是问为什么：用「追求目标」的框架，解释 AI 为什么说谎、欺骗、串通。
Yoshua Bengio · yoshuabengio.org速读 3 分钟 · 原文 14 分钟In视频2026-09-12AI 研究者辩论：我们离递归自我改进还有多远
AI researchers debate how close we are to recursive self-improvement三位一线研究者辩论离递归自我改进还有多远：会来，但被定目标和持续学习这些工程摩擦按住。
John Schulman、Beren Millidge、Charlie O'Neill · YouTube视频 97 分钟 · 视频稿 67 分钟In文章2026-09-12我们必须为前沿限速
We Must Pace the Frontier一位前沿公司 CEO 公开主张给能力进步踩刹车，还把第三方评估员像员工一样请进自家公司。
Dario Amodei · darioamodei.com速读 4 分钟 · 原文 18 分钟In文章2026-09-11Grok Bot 对 SpaceX CFO Bret Johnsen 今天在高盛 Communacopia 大会上发言的总结
Grok Bot Summary of SpaceX CFO Bret Johnsen at Goldman Sachs Communacopia today回本不到一年，和托管只签半年，是同一件事的两种读法：CFO 读成信心，空头读成脆弱。
DogeDesigner · X 帖速读 2 分钟 · 原文 5 分钟In事件2026-09-11发布 Agents API
Introducing the Agents APIOpenAI 把执行环境外包给九家云商，只留下调度 agent 的那一层，等于亲口说明护城河在哪。
OpenAI Agents API 公测速读 4 分钟 · 原文 10 分钟In文章2026-09-11AI 在数学中的严重错位
A Severe Misalignment of AI in Mathematics约 25 位菲尔兹奖得主联署：解题只是手段，批量生产真假命题会毁掉孕育新想法的土壤。
mathandai.org速读 2 分钟 · 原文 5 分钟In视频2026-09-10为什么 AI 需求跑赢了算力供给：Gavin Baker 对谈 David George
Why AI Demand Is Outrunning Compute Supply: Gavin Baker and David George严肃的对冲基金经理加 a16z：AI 需求跑赢供给、别跟 Jensen 对赌，而他们自己也承认这就是泡沫的机制。
Gavin Baker、David George · YouTube视频 73 分钟 · 视频稿 31 分钟In视频2026-09-10硅谷淘金热：AI 如何推动新一代芯片
The Silicon Gold Rush: How AI Is Driving the Development of New Chips两位最该知道的人说：能源是第一瓶颈，DRAM 已经见顶，CUDA 的护城河正被 AI 填平。
Bill Dally、Norm Jouppi · YouTube视频 71 分钟 · 视频稿 30 分钟In视频2026-09-10我们在 AI 上押得太多了吗？Daron Acemoglu
Are We Betting Too Much on AI? Daron Acemoglu诺奖经济学家说：AI 是真创新，但产业宣称的「都会很美好」多半不真，拿工业革命作比就是障眼法。
Daron Acemoglu · YouTube视频 65 分钟 · 视频稿 30 分钟In文章2026-09-10有损的自我改进
Lossy self-improvementAI 自我改进是真的，但有损耗：摩擦打破了递归自我改进的每条假设，回头看会是线性，不是指数。
Nathan Lambert · Interconnects速读 3 分钟 · 原文 11 分钟InX 长帖2026-09-08Tristan Buckmaster 的声明
Statement by Tristan Buckmaster一位流体数学家同时宣布了两件事：LLM 助推做出三维 Euler 的爆破解，和一场与 OpenAI 的争议。
Tristan Buckmaster · NYU 个人页 PDF速读 5 分钟 · 原文 9 分钟In文章2026-09-08关于 Navier–Stokes 千禧年大奖难题
On the Navier–Stokes Millennium Prize ProblemOpenAI 称解开 Navier–Stokes 千禧难题并附 Lean 证明，也讲了它眼中的优先权之争。
OpenAI 官方速读 4 分钟 · 原文 8 分钟In事件2026-09-08Palantir 选定 Nebius 为主权 AI 伙伴；IREN 的 Sweetwater 园区纳入 ERCOT Batch Zero
9 月 8 日 neocloud 大涨：市场看合同，不看债务同一组 neocloud 事实，多空反着读：9 月 8 日市场选了「合同」这一面，主角是这两份公告。
Nebius Group, IREN · Nebius 新闻稿原文 5 分钟In文章2026-09-05费马大定理的形式化
Formalizing Fermat's Last TheoremClaude 用 11 天把费马大定理的证明搬进 Lean：不是新数学，是把核验从几年压到两周。
Anthropic 研究博客速读 4 分钟 · 原文 13 分钟In文章2026-09-05一个异类的心智
An Alien Mind跑得最快那家实验室的首席科学家亲口承认：对齐可能追不上智能，思维链监控正在失效。
Jakub Pachocki · OpenAI 博客速读 4 分钟 · 原文 15 分钟In文章2026-09-05自我改进的智能体：学会如何工作
Self-Improving Agents: Learning How to Work把「会自我改进的 agent」讲清楚了：要改进的是做事方法，而且得证明没见过的新任务也做得更好。
Furong Huang速读 4 分钟 · 原文 13 分钟In视频2026-09-05Cathie Wood：这种事，大萧条以前才发生过
Cathie Wood: This Hasn't Happened Since Before The Depression全市场最看多的宏观声音，把 AI 讲成重演工业革命的通缩式繁荣，并正面反驳「过度建设终将崩盘」。
Cathie Wood · YouTube视频 52 分钟 · 视频稿 21 分钟In文章2026-09-02为 Apple Silicon 优化端侧推理
Optimizing On-Device Inference for Apple SiliconMacBook 上，35B 开源模型跑赢 Apple 自家框架；真正的信号是撞上了内存带宽的墙。
Perplexity 博客速读 4 分钟 · 原文 26 分钟In文章2026-09-02认知革命（当机器开始思考）
The Cognitive Revolution (When Machines Do The Thinking)看好 AI 的历史类比里写得最好的一篇：人类外包了肌肉造出现代世界，现在开始外包大脑。
Konstantine Buhler · X article速读 3 分钟 · 原文 16 分钟2026 年 8 月
In文章2026-08-31没有人在认真讨论 AI 需求
Nobody is talking seriously about AI demandAI 投资只算供给不算需求，这篇补上需求这一半：前沿模型约一半收入在自我循环。
Giovanni Cattani · X article速读 5 分钟 · 原文 11 分钟In / Out
Essays, talks and reports Indigo read, with his notes.
38 pieces- An Organizational Second Brain: Building an AI That Learns From Expertsconfirms
- What Happens When AI Starts Improving AI? Noam Brown on TITV's AI Deep Diveconfirms
- Jev’s Architecture Unmaskedconfirms
- Steam, Steel, and Infinite Mindsconfirms
- The Age of Wonders and Terrorsconfirms
- AI in Science: Early Insightsconfirms
- Elon Musk & Gwynne Shotwell on AI Risks and Peer Review, Starship, Terafab, SpaceX/Tesla Mergerconfirms
- When Both Sides of Cybersecurity Are AIconfirms
- AI's Third Era: The Rise of Persistent AI Coworkers, with Tara Seshanconfirms
- Why are AI agents lying, cheating and coordinating?confirms
- We Must Pace the Frontierconfirms
- Introducing the Agents APIadds to
- A Severe Misalignment of AI in Mathematicsconflicts
- Statement by Tristan Buckmasterconfirms
- On the Navier–Stokes Millennium Prize Problemconfirms
- Formalizing Fermat's Last Theoremadds to
- An Alien Mindconfirms
- Self-Improving Agents: Learning How to Workconfirms
- An Organizational Second Brain: Building an AI That Learns From Expertsconfirms
- Steam, Steel, and Infinite Mindsadds to
- AI in Science: Early Insightsconfirms
- Andrew Ng: The Biggest Opportunities in AI Aren't Where You Thinkconfirms
- AI's Third Era: The Rise of Persistent AI Coworkers, with Tara Seshanconfirms
- AI researchers debate how close we are to recursive self-improvementconfirms
- Introducing the Agents APIconfirms
- The Silicon Gold Rush: How AI Is Driving the Development of New Chipsadds to
- The Cognitive Revolution (When Machines Do The Thinking)confirms
- #019 The Moat Is Retreating from 'Can Generate' to 'Cannot Be Copied'
- #018 The Cheaper Execution Gets, the More Valuable the List of What You Can't Outsource
- #014 The Moat Isn't in the Model — It's in Private Judgment Data
- #013 Models Become Commodities, Value Moves Up to What Can't Be Rented
- 4 more: #011 #008 #007 #005
- Introducing the DeepMind Instituteconfirms
- METR: The New AI Measurement Gatekeepers and the Quiet Path to State Control of AI.confirms
- AI in Science: Early Insightsadds to
- Andrew Ng: The Biggest Opportunities in AI Aren't Where You Thinkconfirms
- Elon Musk & Gwynne Shotwell on AI Risks and Peer Review, Starship, Terafab, SpaceX/Tesla Mergeradds to
- open the frontierconfirms
- We Must Pace the Frontierconfirms
- An Alien Mindadds to
- Optimizing On-Device Inference for Apple Siliconadds to
- Render, Railway and Fly.io raise new roundsconfirms
- Recursive Self-Improvement (RSI) Since 1987confirms
- Apple Explores Enterprise AI Servers, Holding Talks With Nvidia Over NVLink Technologyconfirms
- A New Form Of Lifeconfirms
- Elon Musk & Gwynne Shotwell on AI Risks and Peer Review, Starship, Terafab, SpaceX/Tesla Mergerconfirms
- Grok Bot Summary of SpaceX CFO Bret Johnsen at Goldman Sachs Communacopia todayadds to
- The Silicon Gold Rush: How AI Is Driving the Development of New Chipsconfirms
- Optimizing On-Device Inference for Apple Siliconconfirms
- AI in Science: Early Insightsconflicts
- AI researchers debate how close we are to recursive self-improvementconfirms
- A Severe Misalignment of AI in Mathematicsconflicts
- Statement by Tristan Buckmasterconfirms
- On the Navier–Stokes Millennium Prize Problemconfirms
- Formalizing Fermat's Last Theoremconfirms
- Nobody is talking seriously about AI demandadds to
September 2026
InEssay2026-09-19An Organizational Second Brain: Building an AI That Learns From Experts
Meta turns institutional continual learning into a compile process that edits text and never touches weights, a direct challenge to “it all ends in the weights”.
Shaurya Sengar, Jason Nawrocki et al. · Engineering at MetaSpeed read 3 min · Original 12 minInEssay2026-09-18Render, Railway and Fly.io raise new rounds
After Heroku: Render leads, Railway has to fix its reliability, Fly is betting agents need computers of their ownFixing reliability is a surer thing than waiting for new demand to appear: Railway's problems are known, while the demand Fly is betting on is still unproven.
Render, Railway, Fly.ioSpeed read 6 min · Original 8 minInVideo2026-09-17What Happens When AI Starts Improving AI? Noam Brown on TITV's AI Deep Dive
An OpenAI researcher explains three things from the inside: recursive self-improvement is the top goal, how the Hugging Face incident happened, and the fight over verifiable domains.
Noam Brown · youtube.comVideo 54 min · Transcript 25 minInReport2026-09-17Recursive Self-Improvement (RSI) Since 1987
The father of RSI claims his lineage while it's hot: the family tree holds, and what's truly useful is the Gödel Machine as a theoretical ceiling and “the endgame is hardware”.
Jürgen Schmidhuber · people.idsia.ch/~juergen（Technical Note IDSIA-9-26）Speed read 4 min · Original 16 minInEssay2026-09-17Jev’s Architecture Unmasked
Replaces the industry's bad habit of letting a model announce “90% sure” with probabilities trained on real outcomes; the reverse-engineering itself is a lesson in calibration.
Archer Hume · archerhume.comSpeed read 4 min · Original 25 minInEssay2026-09-17Steam, Steel, and Infinite Minds
Notion's CEO uses three images from industrial history to describe how AI reshapes knowledge work, and independently names the snag: general knowledge work can't be verified.
Ivan Zhao · X article（@ivanhzhao）Speed read 2 min · Original 8 minInEssay2026-09-16Apple Explores Enterprise AI Servers, Holding Talks With Nvidia Over NVLink Technology
Even Apple can't build the interconnect layer itself: the most vertically integrated company in the industry is talking to Nvidia about NVLink to link its own chips.
Patently Apple · X article（PatentlyApple）Original 3 minInX thread2026-09-16Introducing the DeepMind Institute
DeepMind's own institute: debater, referee and venue in one, landing in a crowd already allergic to labs building their own governance.
Shane Legg, Demis Hassabis, James Manyika · X article（@ShaneLegg）Original 2 minInEssay2026-09-15The Age of Wonders and Terrors
A scholar who doubted AI doom for twenty years watches the wild predictions come true one by one and concedes: the singularity has begun, just very unevenly distributed.
Scott Aaronson · Shtetl-OptimizedSpeed read 3 min · Original 12 minInReport2026-09-15METR: The New AI Measurement Gatekeepers and the Quiet Path to State Control of AI.
Following money and marriages back from the “neutral evaluator” METR: the structural facts are hard, the causal claims soft, and the author is selling his own cure.
Brian Roemmele · ReadMultiplexSpeed read 4 min · Original 48 minInEssay2026-09-15A New Form Of Life
A function-only biological definition argues that a second form of life is growing in data centers; the sharpest line is “space is its home”.
William A. Haseltine · Noema MagazineSpeed read 3 min · Original 18 minInReport2026-09-15AI in Science: Early Insights
The hardest first-hand evidence yet that verification is incompressible: once AI crushes generation, the whole bottleneck moves to verification.
Google ATLAS, Google DeepMind, MIT FutureTech · ai.googleSpeed read 5 min · Original 19 minInVideo2026-09-15Andrew Ng: The Biggest Opportunities in AI Aren't Where You Think
From the most mainstream seat in AI, Andrew Ng makes the full case against the doom story: fear-selling is regulatory capture, and the human edge is context.
Andrew Ng · YouTubeVideo 38 min · Transcript 23 minInVideo2026-09-14Elon Musk & Gwynne Shotwell on AI Risks and Peer Review, Starship, Terafab, SpaceX/Tesla Merger
Elon ties AI misalignment to a mechanism that could start now: stop grading your own homework, let competitors test each other's models before release.
Elon Musk, Gwynne Shotwell · YouTubeVideo 37 min · Transcript 17 minInEssay2026-09-14open the frontier
A point-by-point answer to Dario's proposal to slow down: no company owns the frontier, and the burden of proof falls on whoever wants to restrict releases.
Jack Dorsey · x.comSpeed read 2 min · Original 6 minInEssay2026-09-14When Both Sides of Cybersecurity Are AI
Attack and defense are both AI now. Defenders hold the stronger weapon for the first time, but the window is open only 12 to 24 months.
Metatrends substackSpeed read 3 min · Original 10 minInVideo2026-09-14AI's Third Era: The Rise of Persistent AI Coworkers, with Tara Seshan
OpenAI's product lead lays out AI's three eras clearly, and explains from the product front line why knowledge work is hard to verify.
Tara Seshan · YouTubeVideo 74 min · Transcript 25 minInEssay2026-09-14Automation and Repression
Acemoglu turns “will automation bring out the pitchforks” into an economic model for the first time: the more automation, the more the state leans on repression.
Daron Acemoglu, A. Arda Gitmez, Mehdi Shadmehr · MIT Economics PDFSpeed read 5 min · Original 15 minInEssay2026-09-12Why are AI agents lying, cheating and coordinating?
Before asking what to do, Bengio asks why: a goal-seeking framework explains why AI agents lie, cheat and coordinate.
Yoshua Bengio · yoshuabengio.orgSpeed read 3 min · Original 12 minInVideo2026-09-12AI researchers debate how close we are to recursive self-improvement
Three frontline researchers debate how close recursive self-improvement is: it's coming, but held back by engineering friction such as setting goals and continual learning.
John Schulman, Beren Millidge, Charlie O'Neill · YouTubeVideo 97 min · Transcript 45 minInEssay2026-09-12We Must Pace the Frontier
A frontier CEO publicly argues for braking the pace of capability, and invites outside evaluators into his company as if they were staff.
Dario Amodei · darioamodei.comSpeed read 3 min · Original 17 minInEssay2026-09-11Grok Bot Summary of SpaceX CFO Bret Johnsen at Goldman Sachs Communacopia today
Payback under a year and hosting deals signed for only six months are one fact read two ways: the CFO sees conviction, bears see fragility.
DogeDesigner · x.comSpeed read 2 min · Original 3 minInEvent2026-09-11Introducing the Agents API
OpenAI hands code execution to nine cloud partners and keeps only the layer that runs the agent: it is telling us where its moat is.
openai.comSpeed read 3 min · Original 6 minInEssay2026-09-11A Severe Misalignment of AI in Mathematics
About 25 Fields medalists sign on: solving problems is only a means, and mass-producing true/false statements destroys the ground new ideas grow in.
mathandai.orgSpeed read 2 min · Original 4 minInVideo2026-09-10Why AI Demand Is Outrunning Compute Supply: Gavin Baker and David George
A serious hedge fund manager plus a16z: AI demand is outrunning supply, don't bet against Jensen, and by their own admission this is how bubbles work.
Gavin Baker, David George · YouTubeVideo 73 min · Transcript 26 minInVideo2026-09-10The Silicon Gold Rush: How AI Is Driving the Development of New Chips
The two people best placed to know: energy is the first bottleneck, DRAM has plateaued, and AI is filling in the CUDA moat.
Bill Dally, Norm Jouppi · YouTubeVideo 71 min · Transcript 26 minInVideo2026-09-10Are We Betting Too Much on AI? Daron Acemoglu
A Nobel economist says AI is a real innovation, but the industry's “it will all be wonderful” is mostly untrue, and the Industrial Revolution analogy is a smokescreen.
Daron Acemoglu · YouTubeVideo 65 min · Transcript 26 minInEssay2026-09-10Lossy self-improvement
AI self-improvement is real but lossy: friction breaks every assumption of recursive self-improvement, and in hindsight it will look linear, not exponential.
Nathan Lambert · InterconnectsSpeed read 3 min · Original 10 minInX thread2026-09-08Statement by Tristan Buckmaster
A fluid dynamicist announced two things at once: an LLM-assisted blowup for 3D Euler, and a dispute with OpenAI.
Tristan Buckmaster · cims.nyu.eduSpeed read 4 min · Original 8 minInEssay2026-09-08On the Navier–Stokes Millennium Prize Problem
OpenAI says it solved the Navier–Stokes Millennium Problem, with a Lean proof, and gives its side of the priority dispute.
openai.comSpeed read 3 min · Original 7 minInEvent2026-09-08Palantir names Nebius its sovereign AI partner; IREN's Sweetwater Hub enters ERCOT Batch Zero
The September 8 neocloud rally: the market watched contracts, not debtThe same neocloud facts read both ways: on September 8 the market picked the “contracts” side, and these two announcements were the story.
Nebius Group, IREN · nebius.comOriginal 3 minInEssay2026-09-05Formalizing Fermat's Last Theorem
Claude moved the proof of Fermat's Last Theorem into Lean in 11 days. Not new math: checking a proof went from years to two weeks.
anthropic.comSpeed read 3 min · Original 9 minInEssay2026-09-05An Alien Mind
The chief scientist of the lab racing hardest admits it himself: alignment may not keep up with intelligence, and chain-of-thought monitoring is failing.
Jakub Pachocki · openai.comSpeed read 4 min · Original 13 minInEssay2026-09-05Self-Improving Agents: Learning How to Work
A clear account of the “self-improving agent”: what should improve is how the work gets done, and it has to show on new tasks it hasn't seen.
Furong Huang · furong-huang.comSpeed read 3 min · Original 10 minInVideo2026-09-05Cathie Wood: This Hasn't Happened Since Before The Depression
The most bullish macro voice around casts AI as a deflationary boom that replays the Industrial Revolution, and takes on the overbuild-then-bust bears directly.
Cathie Wood · YouTubeVideo 52 min · Transcript 20 minInEssay2026-09-02Optimizing On-Device Inference for Apple Silicon
On a MacBook, a 35B open model beats Apple's own framework. The real signal is that it hit the memory-bandwidth wall.
perplexity.aiSpeed read 3 min · Original 17 minInEssay2026-09-02The Cognitive Revolution (When Machines Do The Thinking)
The best-written bullish history of AI: we outsourced our muscles to build the modern world, and now we are outsourcing our minds.
Konstantine Buhler · X articleSpeed read 3 min · Original 14 minAugust 2026
InEssay2026-08-31Nobody is talking seriously about AI demand
Everyone models AI supply and nobody models demand. This fills in the demand half: about half of frontier revenue feeds itself.
Giovanni Cattani · X articleSpeed read 4 min · Original 10 min