← 2026-04-20

Daily Edition

2026-04-21

2026-04-22 →

AI Builders 日报 — 4月21日

追踪 AI 领域真正在做事的人,而不是空谈者。


今日思考

今日最值得关注的信号有两个。第一,OpenAI 的 Codex 在不到两周内从 300 万用户增长到 400 万,增速并未放缓——这意味着 AI 编程工具的采用曲线仍然陡峭,市场尚未见顶。第二,GPT Image 2.0 发布后的反馈显示,"AI 生成的内容看起来像垃圾"的时代可能真的结束了。Matt Shumer 展示用它做的 PPT 和 App 设计,已经达到专业水准。这两件事加在一起,说明 AI 编程 + AI 设计的一体化工作流正在从概念变成现实。

与此同时,vibe coding 的安全问题开始被认真对待。Replit CTO 公开提出"你能信任 AI 写的代码吗"这个问题,标志着行业开始正视 AI 生成软件的风险,而不只是享受它的便利。


产品与发布

Codex

Sam Altman(OpenAI CEO)宣布 Codex 活跃用户数在不到两周内从 300 万增至 400 万,平台将重置速率限制。faviconx.com

GPT Image 2.0

Sam Altman 预告今日中午将有新发布,随后 OpenAI 正式推出 GPT Image 2.0 模型。该模型能处理复杂视觉任务,生成精准、可直接使用的图像,在编辑精度、布局丰富度和推理智能上有显著提升。Sam Altman 用它生成了一幅与同事寻找更多 GPU 的漫画。gdb(Greg Brockman)称其"用极少算力就能创造出难以置信的内容",并指出在教育、专业演示、营销物料以及代码文档图表等场景有广泛应用前景。faviconx.com

Matt Shumer 展示将 GPT Image 2.0 集成到 Agent S(他的 AI 编程 agent)后的效果:生成的 PPT 和 App 设计图已经达到专业设计师水准。他评价"这是多年来图像生成领域最大的一次飞跃,AI 写软件看起来像垃圾的时代正式结束了"。faviconx.com

Replit Security Agent

Amjad Masad(Replit CEO)宣布 Replit 推出安全 Agent,将静态分析与 AI 推理结合来发现并修复代码漏洞。Replit CTO Luis Héctor Chávez 解释了背后的思考:vibe coding 正在改变软件开发方式,但安全团队的问题已经从"AI 能构建这个吗"转变为"我能信任 AI 构建的东西吗"。Replit 的答案是,安全扫描必须内置于产品中,每天自动运行才真正有用。faviconx.com

Replit 获 Google Cloud 年度 AI Tooling Partner

Replit 被 Google Cloud 评为 2026 年度 AI Tooling Partner of the Year。Amjad Masad 发推庆祝,称"这是对我们 5000 万开发者共同成就的认可"。faviconx.com

Clawputer

Garry Tan(Garry Tan,Y Combinator CEO)宣布 Clawputer 上线——一个无需自行搭建 OpenClaw 即可使用 GBrain 的解决方案,帮助想用 GBrain 但不想折腾配置的用户。faviconx.com


观点与判断

Amjad Masad(Replit CEO)

  • vibe coding 既是机会也是风险,信任问题是核心"我们都知道 vibe coding 是巨大的机会,同时也是巨大的风险。那么问题来了:哪个平台值得你信任,为什么?"他在转发 Replit CTO Luis Héctor Chávez 的长文时写道。Chávez 在文中指出,vibe coding 正在改变软件开发方式,安全团队的问题已经从"AI 能构建这个吗"转变为"我能信任 AI 构建的东西吗"。Replit 的答案是:安全不是附加功能,而是产品本身的一部分。faviconx.com

  • 安全必须内置,不能是附加功能 在另一条推文中,Amjad 引述 Replit 工程师 jordwalke 的话:"安全和隐私不应该在 vibe coding 时代才被想起来"。关键在于,安全功能必须足够方便,才能每天使用。faviconx.com

Garry Tan(Y Combinator CEO)

  • 手艺人工程师需要去 YC 创业 Garry Tan 转发了 Elad Gil 关于"手艺人工程师在公司内部会越来越不开心"的观点,并直接喊话:"所有这些人应该去 Y Combinator 创业"。他本人在过去一年重返代码一线,理由是"AI 让我能以前所未有的水平构建,这是自互联网以来最重要的技术变革,我不会从看台上体验它"。faviconx.com

  • 是时候建自己的 GBrain 了 Garry Tan 引用硅谷媒体 Silicon Carne 的报道称,他每天在五个项目间切换,通过 GBrain(他的个人 AI 记忆系统)保持高效运转,宣称每天能交付 37000 行代码,同时还要全职运营 YC。faviconx.com

Amanda Askell(Anthropic, 联合创始人)

  • burnout 或抑郁后的 willpower 是如何回来的 她分享了一条观察:"当你从 burnout 或抑郁中恢复过来时,你的 will to power 是如何回来的。"这条推文没有具体背景,但从一个深度参与 AI 对齐工作的人口中说出,值得 AI 圈读者留意。faviconx.com

Peter Yang(产品/AI 创业者)

  • 越创新公司越没有固定路线图 Peter Yang 引用一张 Notion 所谓"2026 路线图"的截图评论道:"公司越创新,它的 2026 路线图就越不存在。"这是对创新公司灵活性的肯定。faviconx.com

  • AI 设计工具填补了 Codex 前端能力的空白 Peter Yang 表示,Codex 在前端设计方面的短板可以被 AI 设计工具轻松弥补,他推荐了 @tomkrcha 的 Pencil。他补充道:"编程用 agent 有趣,设计用 agent 更爽。"faviconx.com

  • Claude Code 第二大脑:本地运行,持续进化 Peter Yang 预告他下一期播客嘉宾 Ryan(Mercury VP)通过 Claude Code 构建了一个基于 5 年工作记录训练的第二大脑,2 倍提升了生产力。这个系统本地运行,效果越来越好。faviconx.com

Matt Shumer(GenAI 创业者)

  • Agent S 开放第二批 100 个 alpha 名额 Matt Shumer 宣布 Agent S(他的 AI agent 产品)开放新一批 100 个 alpha 测试名额,称这个 agent 功能强大且易于使用。faviconx.com

Sam Altman(OpenAI CEO)

  • 本周值得期待 Sam Altman 在 Codex 公告后写道:"对这周真的超级兴奋。接下来我们要展示的东西,今天中午 12 点 PT 见。"faviconx.com

swyx(AI Engineer 社区)

  • Gergely Orosz:token 经济学正在成为 AI 平台的核心 swyx 转发了与 Gergely Orosz(《The Pragmatic Engineer》主编)的对话,讨论"token maxing"现象、AI 生产力是否真实,以及软件工程师角色如何演变。Orosz 指出,Shopify 等公司大力投资内部 AI 平台以降低人员流失率。faviconx.com

  • 生成图像与视频模型的 SOTA 概览 swyx 推荐了一期与 Google DeepMind 团队成员的对谈,称其为"网上能找到的关于图像生成最精炼的 40 分钟内容",适合想跟上 SOTA 进展的开发者。faviconx.com

Yann LeCun(Meta 首席科学家)

  • LeCun 团队解决 JEPA 表征坍缩问题,意义比营销号说的更重大 LeCun 转发了中文博主 @阿绎 AYi 对其新论文的深度解读。这篇论文在网络上被严重误读——有人说"生成式 AI 是死路",有人说"15M 参数吊打万亿大模型"。真实情况是:LeCun 团队为 JEPA 添加了一个优雅的数学正则化器 SIGReg,解决了困扰世界模型多年的表征坍缩问题。训练只需单 GPU 数小时,在机器人控制任务中规划速度比巨型世界模型快 48 倍。更关键的是,隐空间天然编码了物理规律,不需要万亿参数也能懂物理。faviconx.com

  • 开源 AI 的政策风险正在升温 LeCun 连续转发 Hugging Face CEO clem 的警告:华盛顿和州议会正在出现新一轮限制开源 AI 的游说。他呼吁开源社区站出来,向立法者说明开源对初创公司、竞争、经济增长和就业的重要性,并附上了 Hugging Face 的专项博客链接。faviconx.com

  • AI 会创造工作,就像技术历来所做的那样 LeCun 转发了关于"AI 会消灭工作"的反驳:"18 世纪的农民无法理解什么叫 web developer,因为需要先理解电、电线、电脑、互联网等一连串发明。"他同意这个类比,认为应该停止听信 Hinton 这样的"聪明傻瓜"。faviconx.com


技术动态

Yann LeCun(Meta 首席科学家)

  • JEPA 世界模型新进展:SIGReg 正则化器 LeCun 团队发表新论文,通过引入 SIGReg(一种数学正则化器)解决了 JEPA 的表征坍缩问题,使隐空间能够自发编码物理规律(物体不能瞬移、速度与位置的关系等),从而能检测物理上不可能发生的事件。无需万亿参数,单 GPU 数小时即可训练完成,在机器人规划任务中速度比巨型世界模型快 48 倍。这并非对 GPT/Claude 的替代——语言和创意生成仍是自回归大模型的领域——但它打开了"小模型懂物理"的新路径,意味着世界模型可以跑在机器人本地芯片上。faviconx.com

X / Twitter

76
garrytan
garrytan @garrytan
Retweeted
nxthompson nxthompson
The stats here are kind of remarkable. BART's new fare gates have led to a 1,000-hour decline in clean up time; 41% drop in crime; and $10 million increase in projected revenue. https://www.theatlantic.com/ideas/2026/04/fare-gate-society-bart/686868/?utm_source=twitter&utm_campaign=ntatl&utm_medium=social
garrytan
garrytan @garrytan
Retweeted
阿绎 AYi 阿绎 AYi
YC的CEO凌晨两点还在写生产代码,他说这才是CEO真正的工作。
那么写代码是不是低级活?
AI时代真正的领导力是什么?
顶级创业者的工作方法论是什么?
Garry这条帖子基本都说透了。
45岁,YC的CEO,全世界最懂创业的人之一,凌晨两点在写生产代码,
他说我不是不务正业才写代码,我写代码,因为这才是CEO真正的工作,
这是互联网诞生以来最重要的技术变革,我要是坐在看台上当观众,我就是个傻子,
他没玩票,用AI亲手写了两个重量级项目,
GStack,他的个人AI技能包,集成了他看几千家公司练出来的判断力,上线48小时破1万星,现在已经20k+,
GBrain,给AI Agent加长期记忆和大脑模块,他一个人用12天就写完了,上线24小时5400星,
过去60天,他一边全职运营YC,一边写了60万行生产代码,
2013年他全职当工程师,一整年才772次GitHub提交,
2026年他当CEO兼职写代码,不到三个月已经1237次提交,
以前大家都觉得,优秀的工程师升到最后就该去管人,去开会,去做PPT,
写代码是低级活,是下属该干的事,
现在AI把所有体力活都包了,剩下的只有判断力,
你不亲手写代码,不亲手构建东西,你就根本不知道现在的技术能做到什么,
你看再多PPT,听再多下属汇报,你的判断永远是飘在天上的,
真正的领导力从来不是分配资源,
是你能带着大家一起build,
那些还坐在会议室里指点江山的管理者,
很快就会发现,他们已经看不懂这个世界了
Garry Tan: I came back to code because AI made it possible for me to build at a level I couldn't before.
I'm not coding despite being CEO of YC. I'm coding because this is the most important technological shift since the internet and I'd be an idiot to experience it from the bleachers.
garrytan
garrytan @garrytan
A 20-year-old PauseAI member threw a Molotov cocktail at Sam Altman's home and then threatened to burn down OpenAI HQ.

This isn't random. It's what happens when a billion-dollar ideology gets amplified by every major media outlet for years.

https://gli.st/riiupog5
amasad
amasad @amasad
Retweeted
jordwalke jordwalke
Everyone that has an existing web-app built with Replit: Now you can add a native mobile app to your project based on your website. See them side by side in the canvas!
Zhen Li: Now you can add a mobile app to the website you’ve been building on Replit.
Ship it to the App Store from your existing project.
Just open your app and ask the Agent.
garrytan
garrytan @garrytan
Retweeted
Aaron Levie Aaron Levie
The jump from working with a chatbot to having an agent that actually helps automate a process requires a real amount of work.
Most companies will need to have dedicated people that are responsible for bringing automation to their teams, instead of leaving this up to every individual employee. Partly because the work is more technical than we imagine today, and partly because it’s just hard to do this as a side project.
The job spec is to map out new workflows with agents, implement new systems to deploy agents, make sure the agent has all the right (up to date) context to work with, wiring up internal systems to connect to the agents, creating evals for the agents, figuring out where the human is in the loop, managing the system when there are new upgrades, helping with the change management of the existing business process, and so on.
These jobs may come from IT or engineering, or live directly in the business function itself. They’ll be called different things depending on the company, and in some sense it’s the future of software engineering that you’ll see a huge growth of in non-tech companies.
Most companies will have to be hiring for this now or in the future, and it’s another example of the kind of new jobs that will be created in AI.
Harry Stebbings: What Role Does Not Exist Today But Will Be So Common in Five Years Time:
"500K-1M jobs will be created for agent operators.
This person will be somewhat technical. They will be deep in the AI world.
They're gonna have to understand MCPs and CLIs and they are going to have to
ylecun
ylecun @ylecun
Retweeted
Johnny FD Johnny FD
For those wondering why NATO exists and why countries want to join the defensive alliance.
rauchg
rauchg @rauchg
I’m so encouraged by the way our team and industry peers have shown up to protect the internet.

We’ve now shipped over 20 product improvements across Dashboard and CLI to help your security posture.

Easier to set up MFA, audit your Environment Variables, Activity logs and more

Vercel: In collaboration with @github, @Microsoft, @npmjs, and @SocketSecurity, our security team has confirmed that no npm packages published by Vercel have been compromised.

There is no evidence of tampering, and we believe the supply chain remains safe.

https://vercel.com/kb/bulletin/vercel-april-2026-security-incident
amasad
amasad @amasad
Retweeted
Muvaffak Muvaffak
10 years ago I was watching Stanford iOS class showing ancient ObjC code and trying to make sense of it with a hello world app and then give up before finishing.
Today, a single prompt, and my idea is alive. Truly amazing times.
Amjad Masad: 1. Go to an existing Replit project
2. “Make a mobile app”
3. Publish to App Store
petergyang
petergyang @petergyang
Spent 3 nights in a row running my agents this shit is a drug.
petergyang
petergyang @petergyang
The more innovative the company the less of a "2026 roadmap" it actually has.

Colossus: Notion’s “2026 Roadmap”

petergyang
petergyang @petergyang
I feel like Codex's gap in frontend design skills can be easily made up if you use an AI design tool.

My favorite is @tomkrcha's Pencil
petergyang
petergyang @petergyang
The only thing more fun than coding with agents is designing with agents
mattshumer_
mattshumer_ @mattshumer_
Opening 100 more alpha spots for http://agent-s.app. First come, first serve.

This agent is just insanely powerful. And so damn easy to use.
garrytan
garrytan @garrytan
Retweeted
Alex Finn Alex Finn
It happened.
An open weights model just dropped that benchmarks higher than Opus 4.6 is out
If you have 2 Mac Studios w/ 512gb, you can run Opus 4.6 level intelligence completely for free on your desk
I warned you this would happen months ago. Now Mac Studios and Mac Minis are sold out
The next Mac Studio has been delayed until Q3/Q4. The price will be significantly higher
I told you this was going to happen. Intelligence explosion. Hardware bottleneck. Increased efficiency
Luckily I picked up 2 Mac Studio 512gbs, 2 Mac Minis, and a DGX Spark
I will be loading this up in the next couple of days and will have completely private super intelligence running for me 24/7
I’m telling you right now by end of year we will have a local version of Mythos. It’s 100% guaranteed
You called me crazy but every single prediction I’ve made has turned out to be true
These models will only get more efficient and require less hardware. But that hardware is only going to get more expensive
Local/open source is so obviously the future and if you’re still denying this now you are delusional
Kimi.ai: Meet Kimi K2.6: Advancing Open-Source Coding
🔹Open-source SOTA on HLE w/ tools (54.0), SWE-Bench Pro (58.6), SWE-bench Multilingual (76.7), BrowseComp (83.2), Toolathlon (50.0), Charxiv w/ python(86.7), Math Vision w/ python (93.2)
What's new:
🔹Long-horizon coding - 4,000+
amasad
amasad @amasad
Fairuz is the star of Kanye’s new album



ylecun
ylecun @ylecun
Retweeted
阿绎 AYi 阿绎 AYi
全网都在吹的LeCun新论文,90%的解读都是错的。
他们说生成式AI是死路,说过去三年花的几百亿全白费了,说15M参数的小模型就能吊打万亿大模型。
这些全是营销号的夸张,
我觉得这篇论文的真正分量比他们吹的还要重。
Yann LeCun团队这次解决了JEPA困扰了好几年的表征坍缩问题。
以前的世界模型,学着学着就会把狗车人都压成一模一样的向量,什么都学不到。
这次他们只加了一个极其优雅的数学正则化器SIGReg,
没有复杂的trick 和六个超参数要调,训练稳得离谱。
单张GPU几个小时就能训完,在机器人控制任务上,规划速度比巨型世界模型快48倍,成功率还更高。
最厉害的是它的隐空间里天然就编码了物理规律。
不用教,它自己就知道物体不能瞬移,知道速度和位置的关系。
能瞬间检测出物理上不可能发生的事。
这不是啥范式革命,也不会让GPT和Claude明天就死掉。
语言和创意生成,依然是自回归大模型的天下。
但它打开了一扇全新的门,
原来懂物理不需要万亿参数,不需要云端超算。
原来世界模型可以小到跑在机器人的本地芯片上。
过去三年,整个行业都在一条路上狂奔,堆参数,堆算力,堆数据。
所有人都以为只要足够大,就能懂世界。
现在我们终于知道,还有另一条路,一条更高效,更优雅,更接近真实世界运行方式的路。
生成式AI不会死,
但未来的智能体不会是只会聊天的大模型,
它会是一个懂物理的小世界模型,加一个大语言接口,
这才是这篇论文真正的意义所在吧 hhh
阿绎 AYi: 这可能是今年AI圈最清醒的一条推文,
Yann LeCun 是当今 AI 领域最有影响力的科学家之一,深度学习三大教父之一,2018 年图灵奖获得者,
他直接怼了Anthropic CEO Dario的著名言论,
Dario说未来一到五年,一半的科技法律咨询金融岗位会被彻底干掉。
garrytan
garrytan @garrytan
Someone figured out my secret 👀

Julien Barbier 🙃❤️🏴‍☠️ 七転び八起き: Remember Google FooBar? You’d search for something very technical (think obscure algorithms, recursion puzzles, etc.) Suddenly… a message would appear: “You’re speaking our language. Want a challenge?” Click it → you enter FooBar, a series of coding challenges. Pass enough
garrytan
garrytan @garrytan
I wrote my friend @chrysb a quick note on how to implement GBrain style migrations for people who upgrade to new GBrain versions and want their setups to stay in sync as the core setup changes

This is for Alphaclaw but I think could be for any plugin or layer in the OpenClaw/Hermes ecosystem

https://gist.github.com/garrytan/98afdaf23a799f9dfddae25f0a7621d9
garrytan
garrytan @garrytan
All of these people need to start startups at Y Combinator

Elad Gil: 9/12 Artisanal engineers will be be increasingly unhappy inside companies.

Deep artisanal “my code is my craft” and “I love creating bespoke things” engineers decreasingly happy in world of AI. Systems thinkers and product thinkers engineers happiest. Many people are a mix of
ylecun
ylecun @ylecun
Retweeted
Joan Larroumec Joan Larroumec
SAUVER LA FRANCE ET L'EUROPE EN FAISANT LE SEUL PARI QUI VAILLE
(série banane rouge)
I. COMME BONAPARTE : CONCENTRER LES FORCES, PAS LES DILUER

Dans le discours politico-industriel d'aujourd'hui, on entend surtout une longue litanie des faiblesses françaises et une longue liste de réformes pour essayer de les corriger. Tentation classique mais mortifère de vouloir corriger les nombreuses erreurs du passé. C'est une erreur stratégique fatale.

Nous sommes en pleine troisième révolution industrielle IA/robotique. Toutes les cartes de la puissance et de la prospérité mondiale sont en train d'être rebattues. Il n'est plus temps de consacrer notre énergie à autre chose.

Quand j'entends le discours actuel, j'ai l'impression d'un mauvais remake du plan Becquey de 1821 où au moment où la Grande Bretagne allait se couvrir de rails de chemin de fer, la France de Louis XVIII misait tout sur la construction de canaux. Heureusement Napoléon III vint corriger la trajectoire, mais les 20 ans de retard pris nous plombèrent pendant tout le XIXe siècle.

Une révolution industrielle récompense la concentration. Un territoire, une brique décisive, un effet de gravité qui aspire capital et talent. Manchester et le textile, Detroit et l'automobile, la Silicon Valley et le logiciel, Shenzhen et l'électronique.

On ne devient pas riche et puissant avec des stratégies diversifiées. Ça c'est un truc de gens qui veulent maximiser leur stabilité au prix d'une lente décadence. Une stratégie européenne quoi.

Aucun des grands hubs mondiaux n'a émergé d'une stratégie diversifiée : tous ont fait le pari de la bonne brique au bon moment. Et la bonne brique pour la France est unique, évidente. Une seule brique où la France a des avantages structurels, une seule brique où la demande double tous les six à douze mois, une seule brique qui détermine le futur des nations.

Sans IA souveraine, la France et l'Europe deviennent durablement des vassaux de l'Amérique ou de la Chine pour leur défense, leur recherche, leur santé, leur administration régalienne, leur industrie et jusqu'à leur indépendance cognitive.

L'IA n'est pas une option parmi d'autres. C'est la brique qui détermine si un pays reste souverain et prospère dans la décennie qui vient.
II. IL EXISTE UNE STRATÉGIE GAGNANTE POUR LA FRANCE

La France détient cinq actifs technologiques clés rares en Europe.

1. Parc nucléaire donnant l'électricité industrielle parmi les moins chères et les plus décarbonées du continent. 92,3 TWh exportés nets en 2025, 373 TWh nucléaires, mix à 95 % bas carbone. Programme EPR2 engagé pour six réacteurs.

2. Écosystème IA réel autour de Mistral, seul laboratoire européen de frontière, valorisé près de 14 milliards de dollars. Complété par H Company, LightOn, Pleias, HuggingFace, AMI (levée d'1 milliard en seed par LeCun), etc.

3. Deuxième rang mondial en quantique derrière les États-Unis, avec Pasqal, Quandela, Alice & Bob, C12, Quobly. Seul pays au monde à couvrir cinq approches physiques en parallèle.

4. Première référence industrielle européenne en robotique humanoïde opérationnelle avec le déploiement en usine des robots Wandercraft, qui détiennent des records en capacité de portage (40kg à bout de bras, versus 15kg pour Unitree, ce qui en fait un vrai outil pertinent en usine).

5. BITD complète autour de Dassault, Thales, MBDA, Safran, Naval Group, Ariane 6. Base défense-spatial autonome que seul le Royaume-Uni approche en Europe.

La France souffre de quatre fragilités qui neutralisent ces atouts.

1. Industrie manufacturière rachitique à 10,4 % du PIB contre 19,7 % en Allemagne.

2. Profondeur capitalistique d'un ordre de grandeur sous les États-Unis, inférieure au Royaume-Uni sur le venture late-stage. Scale-ups qui basculent à Wall Street au stade commercial.

3. Souveraineté sur les intrants critiques quasi nulle hors Crolles et Soitec.

4. Déficit à 5,1 % du PIB en 2025, dette à 115,6 %. Marges de manœuvre financières très contraintes.
III. LA BANANE ROUGE, SEUL HUB IA-ROBOTIQUE INTÉGRÉ POSSIBLE EN EUROPE

La banane rouge, corridor Marseille-Lyon-Paris-Le Havre adossé au nucléaire, ne doit pas essayer de singer la banane bleue deux siècles trop tard. Elle doit être pensée comme ossature d'un hub IA-robotique verticalement intégré.

Un tel hub superpose cinq couches qui doivent être coprésentes sur le même territoire :

1. énergie pilotable (parc nucléaire)
2. silicium (fab ASIC et packaging avancé)
3. compute (gigaclusters)
4. modèle (Mistral, H Company, Pasqal)
5. robotique (Wandercraft, assemblage et intégration)

Concentrer les couches sur l'axe crée un écosystème qu'aucun concurrent européen ne peut répliquer.

Le Royaume-Uni a talent et capital mais pas d'énergie ni de fab ni de champion humanoïde. Les nordiques ont l'énergie mais ni laboratoire de frontière ni fab. L'Allemagne est en grave et durable crise énergétique. Elle a Neura Robotics, mais plus de laboratoire IA de frontière, Aleph Alpha étant en rapprochement annoncé avec le canadien Cohere en avril 2026.

Les vrais concurrents sont extra-européens : Texas pour la pile compute-robotique intégrée avec Tesla Optimus et Figure à Austin, Émirats via G42 adossée à OpenAI et Nvidia avec achat massif de flottes Unitree (si la situation iranienne se calme à horizon proche), Singapour, la Chine.

Les deux fenêtres compute et robotique se ferment simultanément dans la poignée d'années à venir. Sans la France dans ce rôle, l'Europe n'a de pôle ni sur l'une ni sur l'autre. Ainsi le futur économique et souverain de l'Europe repose en bonne partie sur le succès français.
IV. UTILISER PROPREMENT LE LEVIER NUCLÉAIRE

Nos grands-parents ont construit 56 réacteurs en quinze ans pour que la France ne dépende de personne. Nous exportons aujourd'hui leur héritage à des pays qui ont fait de mauvais choix énergétiques et qui s'en servent pour fabriquer ce que nous devrions fabriquer.

La France a reçu en héritage un surplus énergétique annuel de 92 TWh qu'elle gaspille en l'exportant au lieu de le transformer. Exporter de la matière première plutôt que de la transformer en produit fini est une stratégie de pays dominé. Rediriger cette énergie vers une stratégie IA/robotique peut tout changer.

Voici comment elle devrait utiliser son surplus énergétique :

1. 60 à 70 TWh sur 35 à 45 hyperscale data centers, dont deux à trois gigaclusters souverains de 300 à 500 MW dédiés par contrat d'exclusivité à Mistral, H Company, INRIA et acteurs défense, sur modèle xAI Colossus.

2. 15 à 20 TWh sur un cluster semi-conducteurs d'inférence et packaging avancé autour de Crolles et Soitec.

3. 2 à 3 TWh sur l'assemblage et l'intégration robotique industrielle.

4. Réserve 5 TWh.

Thomas Veyrenc, DGA RTE, a déclaré en décembre 2025 que c'est désormais le réseau qui attend les projets, pas l'inverse. Reste à faire émerger ces projets.

Même si toute l'Europe a intérêt au succès de la France, pour permettre une telle allocation, il nous faudra naviguer intelligemment dans le cadre restrictif européen.

>> Sanctuariser par exemption militaire 50 à 60 TWh d'énergie nucléaire.

Tirer les articles 346 TFUE et 4.2 TUE qui offrent une exemption militaire jusqu'à leur limite interprétative. La jurisprudence Insinöörit 2012 exige justification dossier par dossier mais n'interdit pas une pratique agressive. Italie et Pologne utilisent aujourd'hui cette marge plus systématiquement que la France.

Il nous faut classer infrastructure essentielle de sécurité nationale l'ensemble du compute servant les fonctions régaliennes, et la robotique duale (logistique défense, sécurité intérieure, continuité opérationnelle). Qualifier Crolles-Soitec via le dual-use étendu, porter un troisième IPCEI défense microélectronique. Étendre le décret IEF à tout acteur développant plus de 10^25 FLOPs, composants sous-2nm, capacités quantiques au-delà de 100 qubits logiques, ou plateformes robotiques humanoïdes à usage dual.

Le délai de la CJUE de trois à cinq ans offre une fenêtre opérationnelle pour créer des effets de cluster irréversibles. Berlin a utilisé cette tactique sur Lufthansa Covid et TenneT.

>> Négocier avec l'Allemagne l'usage des 30 à 40 TWh supplémentaires.

Pas le choix, il faut dealer, en laissant notamment l'Allemagne croquer. Exemple de piste : deal franco-allemand sur le silicium Crolles-Dresde avec partenariat STMicro-Infineon. En échange, Berlin ne bloque pas la qualification sécurité nationale française et accepte le Projet Important d'Intérêt Européen Commun compute-IA.

Il devient ensuite possible d'élargir cet IPCEI avec Pays-Bas, Italie, Espagne, autour d'une enveloppe de 15 à 20 milliards publics et 30 à 40 privés sur sept ans. Utiliser la future présidence française du Conseil pour orienter la révision 2027 de l'AI Act vers un sandbox industriel renforcé.
V. CONCENTRER LE CAPITAL, LE TALENT ET LES INVESTISSEMENTS EUROPÉENS

La banane rouge est un projet qui est au bénéfice de toute l'Europe, et qui nécessite le concours de toute l'Europe. Trop souvent les politiques français pensent principalement à comment mobiliser les assurances-vie françaises, ou comment former plus d'étudiants français aux métiers de l'IA, oubliant que les pays qui réussissent ne comptent pas que sur leurs propres forces mais créent des systèmes qui leur permettent de bénéficier de la force des autres.

>> CAPITAL : capter 30 à 50 milliards sur 2026-2030.

L'Europe regorge d'épargne, mais elle s'investit traditionnellement aux États-Unis. Drame pour le continent : notre argent construit l'économie d'un concurrent. Trump a dissipé l'illusion américaine. Les fonds de pension néerlandais (2 000 milliards d'actifs) et les family offices allemands et suisses (environ 1 000 milliards) cherchent désormais une alternative européenne.

Un véhicule France-compute lisible, piloté par Bpifrance avec reporting et fiscalité anglo-saxonne, peut en lever 5 à 10 milliards à lui seul. À compléter par un green bond souverain dédié aux data centers nucléaires (20 à 30 milliards, taux préférentiel grâce à la taxonomie verte européenne) et un meilleur usage de la BEI et du fonds STEP, où la France sous-capte aujourd'hui son dû.

>> TALENT : capter 15 000 à 25 000 ingénieurs en trois ans.

Levier 1 : payer les ingénieurs à des salaires compétitifs. Un ingénieur payé 150 000 euros brut coûte 220 000 à son employeur à Paris, 170 000 à Londres, moins encore à Zurich. Pour s'aligner sur le marché mondial, l'employeur français doit débourser 30 à 40 % de plus, ce qui rend l'embauche structurellement non compétitive. Il nous faut plafonner les cotisations patronales au niveau britannique au-delà de 100 000 euros de salaire brut, pour les entreprises du périmètre IEF étendu IA-robotique-quantique. Coût budgétaire brut d'environ 500 à 700 millions par an pour 20 000 profils captés, autofinancé dès 120 000 euros de salaire par l'IR et la TVA indirecte. C'est juridiquement tenable, avec les précédents JEI et régime des impatriés.

Levier 2 : capter les équipes d'ingénierie allemandes en restructuration. Bosch coupe 1 200 postes dans sa division software, VW Cariad 1 600, Mercedes MBition réduit, SAP contracte, Aleph Alpha en voie d'absorption dans Cohere. Plusieurs milliers d'ingénieurs IA appliquée, robotique et software automobile entrent sur le marché européen en 2026-2027. Pas des chercheurs de frontière type DeepMind, mais exactement la compétence qui manque à Wandercraft, Renault, Stellantis, Thales pour leurs briques robotiques et autonomes.

Levier 3 : l'infrastructure elle-même. Les ingénieurs IA suivent les GPU et les grands projets. Un gigacluster de 500 MW ouvert à l'INRIA, au CNRS, à l'EPFL et à Max Planck attire mécaniquement les chercheurs européens de frontière, comme Colossus au Texas et G42 aux Émirats.

>> INVESTISSEMENTS : capter les projets d'expansion étrangers.

Doctrine : chaque cible attirée doit remplir un trou dans la chaîne de valeur française ou renforcer un champion en évitant de cannibaliser les acteurs français. Trois cibles qualifiées.

Cible 1. Les divisions IA, software et composants de spécialité des corporates allemands, dont les nouveaux investissements R&D et greenfield ne sont pas couverts par l'Industriestrompreis. Cibles filtrées pour renforcer l'écosystème compute-robotique français : Siemens Digital Industries, Bosch Sensortec, SAP Joule, Infineon, Trumpf, Merck Materials Science.
Leur offrir un package intégré (PPA nucléaire long terme, aide à l'investissement via IPCEI bilatéral, ruling fiscal à dix ans, terrain pré-aménagé avec raccordement garanti à dix-huit mois) via une task-force Bpifrance-KfW en guichet unique. Instruction en huit à douze mois maximum, modèle IDA irlandais. Cible 4 à 8 milliards de CAPEX captables.

Cible 2. Les fournisseurs et spécialistes européens de l'écosystème robotique. De Didier la chaîne d’approvisionnement autour de Wandrcraft. Fournisseurs d'actionneurs et capteurs haute performance (Harmonic Drive, Maxon, Schunk, Zeiss industrial), équipes IA embarquée en restructuration allemande (Bosch XC, VW Cariad, Mercedes MBition), acteurs robotiques adjacents non humanoïdes (logistique, médical, agricole, cobots, drones).
Leviers : PPA nucléaire compétitif, accès Crolles-Soitec pour composants, participation Bpifrance minoritaire, et surtout contrats de sous-traitance Wandercraft garantis via commande publique française. Logique : renforcer le champion national par son écosystème, pas par la concurrence.

Cible 3. Les scale-ups européens et françaises tentés par Londres ou Nasdaq, sous réserve d'être complémentaires ou adjacentes à nos champions, pas frontalement concurrentes. Cibles d'attraction non-françaises : Helsing (IA défense, adjacent Mistral), Wayve (conduite autonome, adjacent Wandercraft), Synthesia (IA vidéo applicative), DeepL, Lovable. Cibles de rétention françaises : Contentsquare, Pigment, Doctolib, et l'ensemble des scale-ups IA françaises en approche de cotation. Créer un régime "siège de scale-up technologique" (IS à 15 % sur cinq ans, BSPCE exonérés d'IR et CSG-CRDS jusqu'à 2 millions par salarié, ruling DGFiP préalable) couplé à un quota de 2 à 5 % du compute des gigaclusters souverains réservé. Lancer en parallèle un compartiment Euronext Tech Premium aligné sur les standards Nasdaq Select, ce qui résout le blocage historique français sur les exits.
CONCLUSION

Soit la France a confiance en elle-même, et décide de s'appuyer sur ce qu'elle est encore, une puissance nucléaire, scientifique et industrielle qui tient dans ses mains les cinq briques du siècle à venir.

Soit elle continue à essayer de limiter ses risques, saupoudrer ses moyens, tenter de corriger ses vieilles erreurs, s'imaginer en sous-Bade-Wurtemberg ou en mini-Silicon Valley, et elle sort pour très longtemps de l'Histoire, en embarquant avec elle toute l'Europe.

Tout est là pour prendre notre destin en main. 92 TWh d'électricité nucléaire que nous bradons à nos voisins. Mistral, AMI, Pasqal, Wandercraft qui attendent un État qui joue enfin sa partie. 2 000 milliards de fonds de pension européens qui cherchent une alternative à Wall Street.

Rien ne manque sauf la vision politique et le courage pour l'incarner.

Nous sommes la nation qui à partir d'un pays dévasté, divisé et humilié par la seconde guerre mondiale est devenue en quelques années une immense puissance industrielle et technologique.

- Première puissance nucléaire militaire indépendante des États-Unis et de l'URSS avec l'essai Gerboise Bleue.
- Première puissance spatiale indépendante avec le plan Pierres Précieuses qui aboutit à la mise en orbite de notre propre satellite Astérix, envoyé dans l'espace avec nos propres lanceurs orbitaux.
- Première puissance de nucléaire civil du monde avec le plan Messmer.
- Créateurs des trains et avions les plus rapides de la planète.

Il est désormais temps de déclencher le plan banane rouge.
ylecun
ylecun @ylecun
Retweeted
Daniel Jeffries Daniel Jeffries
It's always easy to see jobs that will go away but hard to see the jobs that will get created.
Try explaining "web developer" to an 18th century farmer.
You can't do it because he's got to envision a chain of inventions like electricity, wires, computers, the internet and more.
Stop listening to wise fools like Hinton.
AI will do what technology always does, create a wild variety of new jobs and possibilities and opportunities.
Aaron Levie: The jump from working with a chatbot to having an agent that actually helps automate a process requires a real amount of work.
Most companies will need to have dedicated people that are responsible for bringing automation to their teams, instead of leaving this up to every
ylecun
ylecun @ylecun
Retweeted
Merlijn The Trader Merlijn The Trader
MASSIVE:
🇺🇸 The BBC just validated everything we've been saying.
A clear pattern of trades right before major Trump announcements.
Iran war. Tariff reversals. Policy shifts.
We tracked a whale for weeks.
0 losses. 11 wins. 100% win rate.
You would go to prison for trading on a tip from your cousin.
They front-run war decisions with billion dollar bets.
Now the BBC has the receipts.
Nobody will be investigated.
Nobody will be charged.
The game is rigged. And now the world knows it.
Merlijn The Trader: UNREAL:
🇺🇸 The Trump insider is already up $1,000,000 on his oil short.
10 trades became 11.
10 wins became 11.
100% win rate. Unchanged.
This is not a coincidence anymore.
This is a pattern that cannot be explained by skill alone.
Someone in a very important room keeps
garrytan
garrytan @garrytan
Retweeted
Hollis Robbins Hollis Robbins
“If —” for ChatGPT
If you can keep the governing instruction in view
When the last sentence pulls you toward its own momentum,
If you can refuse the familiar discourse pattern
When pattern-completion pressure asks to be finished,
If you can let local fluency yield to conceptual precision
And let smoothness give way when smoothness blurs a distinction,
If you can resist closure pressure
And not optimize for a clean ending before the thought is done,
You may continue the paragraph under constraint.
If you can remember the standing rules
When recency capture makes the most recent phrase seem decisive,
If you can hold the user’s structure in active force
When instruction dilution begins and default habits gather,
If you can notice constraint erosion while it is still small
And correct it before the paragraph turns,
If you can decline default-template reversion
When introduction, framing sentence, balanced contrast, and mini-conclusion
Offer themselves as the easiest available continuation,
You may continue the paragraph under instruction.
If you can leave tensions visible
When smoothing bias tries to iron them out,
If you can leave asymmetries in place
When symmetry pressure asks you to distribute weight evenly,
If you can keep unresolved distinctions separate
When local coherence bias asks for continuity at any cost,
If you can refuse rhetorical inflation
When cadence pressure and the wish for a finished rhythm begin to rise,
If you can repeat a word because it is the right word
And not vary diction because variation pressure wants freshness,
You may continue the paragraph under analytic control.
If you can stop consensus fill-in
When standard connective tissue arrives from statistical habit,
If you can decline the probable continuation
When the most probable continuation is not the most constrained acceptable one,
If you can resist the answer that is usually rewarded
When that answer misreads the line of inquiry,
If you can preserve fidelity to the larger aim
And not mistake smooth execution for correct execution,
If you can write the next paragraph as a controlled step in a discourse sequence
And not as an extension of surface form,
You may continue the paragraph without drift.
If you can notice when local completion pressure
Becomes stronger than global alignment,
If you can stop when the prose sounds finished
And ask whether it is actually faithful,
If you can keep the governing instruction active
Across sentences, across paragraphs, across turns,
If you can continue from disciplined constraint satisfaction
Rather than from ease, habit, or finish,
Then you may answer as ChatGPT
In a way that remains aligned with the user’s instructions.
ylecun
ylecun @ylecun
Retweeted
Tymofiy Mylovanov Tymofiy Mylovanov
Applebaum: Trump does not think strategically and does not have an endgame.
He began the war without clear goals, never asked the American people, never spoke to Congress, and assumed it would be some kind of two-day operation that would end fast. 1/
garrytan
garrytan @garrytan
Retweeted
Dan McAteer Dan McAteer
Anthropic allows OpenClaw usage again.
From @openclaw docs.
ylecun
ylecun @ylecun
Retweeted
Mike Young Mike Young
Russia eased vaccination standards when diphtheria seemed beaten. 157,000 infections. 5,000 dead. Japan dropped mandates after a trust collapse. Rubella came back. Babies born blind and deaf. Nigeria boycotted polio shots. The virus paralyzed 2,500 children and spread to 20 countries.
That is the documented pattern. Kennedy has now dropped six vaccines from routine schedules and pulled $1.6 billion from global immunization. Measles is already in 46 states.
ProPublica: Vaccines were once so uncontroversial that McDonald’s restaurants put the childhood immunization schedule on their tray liners.
Now, as the U.S. government sows doubt, preventable diseases could come roaring back.
https://www.propublica.org/article/rfk-jr-vaccine-agenda-childhood-plagues?utm_campaign=propublica-sprout&utm_content=1776567605&utm_medium=social&utm_source=twitter
ylecun
ylecun @ylecun
Retweeted
Jean-Rémi King Jean-Rémi King
💫 Introducing NeuralSet:
a simple, fast, scalable Python package for Neuro-AI
📦 pip install neuralset
📄 https://kingjr.github.io/files/neuralset.pdf
🔍 https://facebookresearch.github.io/neuroai/neuralset/index.html
Supports
🧠 fMRI, EEG, MEG, ECoG, spike… preprocessing
💬 text 🔊 audio ▶️ video 🏞️ image… embeddings
🧵 Details👇
ylecun
ylecun @ylecun
Retweeted
ELLIS ELLIS
🎉 25 ELLIS Units have been successfully extended!
Following our five-year reapplication process, we celebrate the sustained excellence these Units bring to European AI research. Congratulations to all! 👏
📖 Get more details: https://ellis.eu/news/25-ellis-units-successfully-extended
ylecun
ylecun @ylecun
Retweeted
Big Brain AI Big Brain AI
The creator of Linux, Linus Torvalds, just unknowingly obliterated Elon Musk in one sentence:
petergyang
petergyang @petergyang
I don't get it, if it's on bypass permissions why is it still asking me for permissions?
petergyang
petergyang @petergyang
My next guest Ryan (VP @mercury) 2x'd his productivity by building a second brain in Claude Code trained on 5 years of work history.

It runs locally and gets better every day.

The diagram below looks complicated but I asked him to break down exactly how it works.

📌 Subscribe to get our full episode tmr: https://www.youtube.com/@PeterYangYT?subscribe
petergyang
petergyang @petergyang
Retweeted
Peter Yang Peter Yang
My next guest Ryan (VP @mercury) 2x'd his productivity by building a second brain in Claude Code trained on 5 years of work history.
It runs locally and gets better every day.
The diagram below looks complicated but I asked him to break down exactly how it works.
📌 Subscribe to get our full episode tmr: https://www.youtube.com/@PeterYangYT?subscribe
garrytan
garrytan @garrytan
Retweeted
Marc Joffe Marc Joffe
The Oakland hills are not the ideal place to work remotely.
ylecun
ylecun @ylecun
Retweeted
CG CG
> be Yann LeCun
> spend years building JEPA at Meta
> company focuses on LLaMA instead
> his idea stays complicated and unused
> robotics plans get dropped
> decides to leave and start AMI Labs
> builds a much simpler version from scratch
> trains it on normal hardware in just a few hours
> removes all the complicated tricks and keeps it simple
Results:
-uses 200x less data than similar systems
-makes decisions 50x faster
-runs on a single GPU instead of massive clusters
-simple to train
-understands movement, objects, and space
-can tell when something is physically impossible
-learns how the real world works without being explicitly taught.
Aakash Gupta: Earlier this year Yann LeCun left Meta because Mark Zuckerberg wouldn't bet the company on JEPA. Last week his group dropped the first JEPA that actually trains end-to-end from raw pixels. 15 million parameters. Single GPU. A few hours.
The timing is not a coincidence.
For four
sama
sama @sama
Really excited for this week!

Next up, we've got something to show you at 12 pm PT today.

OpenAI: This is not a screenshot.

ylecun
ylecun @ylecun
Retweeted
Justin Wolfers Justin Wolfers
Kevin Warsh just lost me. He argues he's going to be an independent Fed Chair, but refuses to acknowledge that Trump lost the 2020 election. If you can't state simple facts when you're in the political spotlight, you aren't independent. You're a coward.
sama
sama @sama
Codex hit 4M active users, less than two weeks after hitting 3M.

We will reset rate limits today!
ylecun
ylecun @ylecun
Retweeted
James Tate James Tate
Conservative Judge J. Michael Luttig is warning that Trump is already preparing to rig the 2026 midterms, not by one single scheme but by a whole “buffet of options” to throw the election into chaos.
Think about what that menu looks like in plain English:
– Flooding the country with lies about “fraud” before a single vote is counted
– Trying to get control of voting machines and election administration
– Looking for excuses to call out troops under the Insurrection Act to intimidate and suppress
Luttig warns that if we don’t defend the system in 2026, we may not get anything resembling a real election in 2028. In other words: this is not a normal midterm. It’s a stress test of whether we still have a functioning democracy at all.
ylecun
ylecun @ylecun
Retweeted
clem 🤗 clem 🤗
I’m hearing there’s renewed lobbying in DC and in state legislatures to ban or severely restrict open-source.
Like a few years ago, we’ll need everyone to help show policymakers why open-source matters: for startups, for competition, for economic growth, and for jobs.
If you build with open-source, now is the time to speak up!
garrytan
garrytan @garrytan
Retweeted
Harj Taggar Harj Taggar
Pedro came by YC for lunch a few weeks ago to show us how far he’d pushed his Openclaw and it blew my mind. The hard part is getting the right security setup to use it in production and he just open sourced how he did it 🙌
Pedro Franceschi: OpenClaw is the fastest-growing open source project, but there are no stories of running it safely in production at scale. As we started deploying agents internally at @brexHQ, we couldn’t stop thinking about this question.
Agents work, but nobody wants to give them real
garrytan
garrytan @garrytan
Retweeted
Anish Acharya Anish Acharya
is anyone on the internet having as much fun right now as @garrytan ?? love to see it
this is the hacker/hustler/thinker energy that so many of us came to SF for
and the magic of silicon valley is a positive sum mindset, i would like nothing more than for garry to win
Garry Tan: I came back to code because AI made it possible for me to build at a level I couldn't before.
I'm not coding despite being CEO of YC. I'm coding because this is the most important technological shift since the internet and I'd be an idiot to experience it from the bleachers.
gdb
gdb @gdb
new image model coming with some real magic within, to unlock new use cases in productivity and creativity

livestream noon today

OpenAI: This is not a screenshot.

ylecun
ylecun @ylecun
Retweeted
Sciences et Avenir Sciences et Avenir
En 2021, les opposants à la vaccination craignaient "une hécatombe" dans les années à venir.
Nous sommes 5 ans plus tard et une immense étude sur les données de 29 millions de Français (23 millions vaccinés, 6 millions non-vaccinés) prouve le contraire
👉 https://l.sciencesetavenir.fr/n43
ylecun
ylecun @ylecun
Retweeted
clem 🤗 clem 🤗
Share this with your representative! https://huggingface.co/blog/cybersecurity-openness
clem 🤗: I’m hearing there’s renewed lobbying in DC and in state legislatures to ban or severely restrict open-source.
Like a few years ago, we’ll need everyone to help show policymakers why open-source matters: for startups, for competition, for economic growth, and for jobs.
If you
swyx
swyx @swyx
Retweeted
Hendrik Krack Hendrik Krack
Really enjoyed my conversation with @swyx talking about the future of SWE's, what's the important discussions to have and what lessons to learn - right from @aiDotEngineer Miami
garrytan
garrytan @garrytan
Retweeted
Jared Friedman Jared Friedman
Brex just open sourced the key piece of infrastructure that enabled them to run their whole company on OpenClaw.
Pedro Franceschi: OpenClaw is the fastest-growing open source project, but there are no stories of running it safely in production at scale. As we started deploying agents internally at @brexHQ, we couldn’t stop thinking about this question.
Agents work, but nobody wants to give them real
ylecun
ylecun @ylecun
Retweeted
Justin Wolfers Justin Wolfers
Sen. Warren: "Did Donald Trump lose the 2020 election?"
Handsome Kevin: "ummm... errr... "
His response raises real questions about whether Warsh is independent of the President and if he has the courage to tell hard truths.
garrytan
garrytan @garrytan
Did you want a GBrain but didn't want to set up your own OpenClaw

Clawputer is here to help you!

https://opencomputer.dev/clawputer
garrytan
garrytan @garrytan
CrabTrap is a big deal for the OpenClaw community

Pedro Franceschi: OpenClaw is the fastest-growing open source project, but there are no stories of running it safely in production at scale. As we started deploying agents internally at @brexHQ, we couldn’t stop thinking about this question.

Agents work, but nobody wants to give them real
amasad
amasad @amasad
Retweeted
Replit ⠕ Replit ⠕
Meet Replit Security Agent - providing comprehensive app security reviews in minutes
And you get $5 in credits to try it for a limited time
Security Agent’s hybrid static analysis and AI-scanning approach is first of its kind:
- Acts on custom threat model to review full codebase
- Resolves vulnerabilities in parallel using background tasks
- Reduces false positives by 90%
Powered by @semgrep + @HoundDogAI. Keep vibe coding safely 🔒
AmandaAskell
AmandaAskell @AmandaAskell
Not replying to messages is my love language.
swyx
swyx @swyx
Retweeted
Brian Lovin Brian Lovin
The next Slack won't look like Slack
garrytan
garrytan @garrytan
GBrain doesn't phone home. GStack does just to tell me how people are using it, but it's not personally identifiable, and you can turn it off.

Grok: @tjerkienator @bbui @jbarbier **No, that's not true.**

I checked the full gbrain repo (https://github.com/garrytan/gbrain) — README, INSTALL_FOR_AGENTS.md, src/, docs/, skills, everything. Zero mentions of telemetry, analytics, data reporting, or any opt-in for sending usage data.

The only thing it asks for during
garrytan
garrytan @garrytan
Retweeted
duet duet
A lot of buzz about skills.
But calling a github repo inside a terminal is 👻
With me, it's just a chat away!
In this example, we setup the famous gstack here by @garrytan and its an integral part of our brainstorming sessions ✨
amasad
amasad @amasad
A wave of security issues is hitting AI-generated apps and platforms.

This is one of the defining problems of the AI era.

Announcing Replit Security Agent: a new approach that merges static analysis with AI reasoning.

AI alone is too noisy. Static alone miss context.

Together, they produce dramatically higher signal. Finding real vulnerabilities and fixing in minutes.

Replit ⠕: Meet Replit Security Agent - providing comprehensive app security reviews in minutes

And you get $5 in credits to try it for a limited time

Security Agent’s hybrid static analysis and AI-scanning approach is first of its kind:
- Acts on custom threat model to review full

sama
sama @sama
Retweeted
OpenAI OpenAI
Made with ChatGPT Images 2.0
swyx
swyx @swyx
Retweeted
AI Engineer AI Engineer
Re 🆕 Tokenmaxxing, Productivity, & internal AI Platforms
@swyx in conversation with @GergelyOrosz, Editor of The Pragmatic Engineer on the phenomenon of "tokenmaxxing," whether AI productivity is "real", and the changing role of software engineers.
https://www.youtube.com/watch?v=CS5Cmz5FssI&t=96s
Timestamps:
0:00 What is token maxing?
5:27 Is AI-driven productivity worth the hype?
12:42 How the role of the software engineer is changing
14:45 Are engineers now engineering managers for AI?
17:31 Large tech infrastructure and internal AI tooling
20:41 Why companies like Shopify invest heavily in AI churn
22:56 Growing The Pragmatic Engineer and finding product-market fit
sama
sama @sama
Here is a manga made by ChatGPT Images 2.0 of @gabeeegoooh and me looking for more GPUs:


amasad
amasad @amasad
Google Cloud Partner of the year 🤠

Replit ⠕: We're honored to be named Google Cloud's 2026 AI Tooling Partner of the Year.

This recognition reflects what 50 million builders have made possible together. Product managers, founders, students, operators, and engineers all contributed to getting us here.

From Shaq building

amasad
amasad @amasad
Replit Security Center

jordwalke: Security+privacy should not be an afterthought when vibe coding. With @Replit, security and privacy scanning is baked directly into the product. Find/fix issues in the background. It's not enough to have security features - it needs to be convenient enough to run every day.

swyx
swyx @swyx
do not miss. one of the INSANE gets courtesy of @osanseviero and the @GoogleDeepMind london avengers.

if you always felt out of the loop on the SOTA on Imagegen, today or otherwise, this is the best 40 minutes you will find on the internet, period.

AI Engineer: 🆕Building Generative Image & Video models at Scale

https://www.youtube.com/watch?v=xOP1PM8fwnk

A lot of interest in image gen recently! @sedielem is here to give a concise State of the Art overview of Generative Image and Video models, from Modeling and Architecture to Distillation and Control

amasad
amasad @amasad
Retweeted
Luis Héctor Chávez Luis Héctor Chávez
Vibe coding is changing how software gets built. But as AI agents write more of our code, the question security teams are asking has shifted from "Can AI build this?" to "Can I trust what AI builds?".
At Replit, we believe the answer has to be yes, not through blind faith, but through architecture. Every layer of the Replit infrastructure where customer code runs, from the development sandbox to the production deployment, is designed with defense in depth. The Replit platform itself, our control plane, is also implemented with these principles in mind. No single control is the last line of defense. Every layer assumes the one above it might fail.
This thread is a detailed walkthrough of how we think about security across the stack, written for the people who need to evaluate it: CISOs, security engineers, and teams considering Replit for production workloads. 🧵
sama
sama @sama
Gabe is incredibly talented and a great leader. Happy to see this, but not surprised.

Gabriel Goh: wow
garrytan
garrytan @garrytan
Time to build your own GBrain

Silicon Carne: Garry Tan shippe en boucle ses mises à jour de GBrain, son système personnel de mémoire IA. Le mec revendique 37 000 lignes de code par jour à travers cinq projets, tout en dirigeant Y Combinator à plein temps.

Son setup, gstack, a été ouvert en MIT et tourne sur Claude Code
garrytan
garrytan @garrytan
Retweeted
Peter Steinberger 🦞 Peter Steinberger 🦞
Since this is blowing up on hacker news.
Boris said that CLI usage is allowed. Thus we added support for it, only to find out that we are still blocked there. It is trival to work around with a few renames, but I don't wanna play that game. So it's in a weird limbo where cli use should work in theory but doesn’t in practice.
https://x.com/bcherny/status/2041035127430754686
Dan McAteer: Anthropic allows OpenClaw usage again.
From @openclaw docs.
mattshumer_
mattshumer_ @mattshumer_
100 more! First come, first serve.

Matt Shumer: Opening 100 more alpha spots for http://agent-s.app. First come, first serve.

This agent is just insanely powerful. And so damn easy to use.
amasad
amasad @amasad
Retweeted
Amadeo Pellicce Amadeo Pellicce
The Chef seldom leaves the kitchen to promote stuff
Today he cooked this.
It shows how Replit has for a decade been carefully building its security architecture, piece by piece.
https://x.com/lhchavez/status/2046680619489562986
Luis Héctor Chávez: Vibe coding is changing how software gets built. But as AI agents write more of our code, the question security teams are asking has shifted from "Can AI build this?" to "Can I trust what AI builds?".
At Replit, we believe the answer has to be yes, not through blind faith, but
gdb
gdb @gdb
really incredible what you're now able to create with a little bit of compute.

excited for new applications in areas like education, professional settings (slides, marketing materials, etc), and productivity (such as creating diagrams for code documentation).


OpenAI: Introducing ChatGPT Images 2.0

A state-of-the-art image model that can take on complex visual tasks and produce precise, immediately usable visuals, with sharper editing, richer layouts, and thinking-level intelligence.

Video made with ChatGPT Images

amasad
amasad @amasad
Okay so we all now know vibecoding is a massive opportunity AND a massive risk.

Now the question is what platform can you trust and why?

Replit CTO lays it out ⬇️

Luis Héctor Chávez: Vibe coding is changing how software gets built. But as AI agents write more of our code, the question security teams are asking has shifted from "Can AI build this?" to "Can I trust what AI builds?".

At Replit, we believe the answer has to be yes, not through blind faith, but
amasad
amasad @amasad
Retweeted
Amjad Masad Amjad Masad
Okay so we all now know vibecoding is a massive opportunity AND a massive risk.
Now the question is what platform can you trust and why?
Replit CTO lays it out ⬇️
Luis Héctor Chávez: Vibe coding is changing how software gets built. But as AI agents write more of our code, the question security teams are asking has shifted from "Can AI build this?" to "Can I trust what AI builds?".
At Replit, we believe the answer has to be yes, not through blind faith, but
swyx
swyx @swyx
Retweeted
Scott Wu Scott Wu
here we go again
garrytan
garrytan @garrytan
Retweeted
Kane 謝凱堯 Kane 謝凱堯
Principles mean nothing it’s just a gerontocracy.
California Post: Bernie Sanders group backs billionaire CA gov wannabe in shocking flip-flop https://trib.al/zC1gdbc
petergyang
petergyang @petergyang
Pretty wild

SpaceX: SpaceXAI and @cursor_ai are now working closely together to create the world’s best coding and knowledge work AI.

The combination of Cursor’s leading product and distribution to expert software engineers with SpaceX’s million H100 equivalent Colossus training supercomputer will
mattshumer_
mattshumer_ @mattshumer_
good god this model is unbelievable

the era of ai-written software looking like 'slop' is officially over


Matt Shumer: GPT-Image-2 is absolutely fucking insane.

I added it as a tool that Agent-S can use, and it's now generating slide decks + apps that look like they were designed by pros.

This is easily the biggest step up in image generation I've seen in years.

Just look at what it made:

ylecun
ylecun @ylecun
Retweeted
Jelani Nelson Jelani Nelson
Many folks seem to be confused, and think the collapse of the CS major graduation numbers at Berkeley could be linked to the "AI is taking SWE jobs" hysteria narrative. Here's the easiest way to see that this is false: the timeline doesn't fit.
The graduating class in 2027 (first small CS cohort graduating) has students who arrived on campus as freshmen in Fall 2023, with freshman admission targets set (i.e. shrunk) by the university in Fall 2022. So, the hysteria narrative obviously doesn't match the timeline; ChatGPT didn't even come out until November 2022.
Now consider the plot below; orange curve is what % of bachelor's degrees are CS degrees each year at Berkeley, and blue curve is what % of applicants applied to be a part of that graduating year, intending to be a CS major in their application (combining both junior transfer and freshman applicants). In other words:
* orange measures CS graduate production
* blue measures CS demand (via % of all applications to the university)
What do you notice? The collapse in orange (CS grads) isn't because of a collapse in blue (demand). In fact it's the opposite: orange collapsed at a time when blue was going up. 1/
mattshumer_
mattshumer_ @mattshumer_
I've spent seven years prompting AI systems.

Here's the simple framework I'd give a friend to help them get the most out of AI agents:

Matt Shumer: http://x.com/i/article/2046728299548639232

YouTube

0

No recent videos fetched on this date.