X Posts

AI 股神 Leopold 这两天上演了一出真正的“白天清仓，周末结婚”. 他旗下的 Situational Awareness 因重仓 AI、最高约4倍杠杆反噬，被迫把大部分公开股票仓位卖给 Citadel

Renting a MacBook Pro for inference could earn ~$423/month, paying for itself in a year. Questions about Darkbloom provider's real demand and payouts.

Lauren Tan @poteto 是 Cursor 的工程师，之前在 Meta 做 React Compiler，也在 Netflix 做过 tech lead 和工程经理. 她加入 Cursor 只有五个月

台上的人，就是最近因四倍杠杆反噬、450亿美元基金一夜爆仓的 Leopold Aschenbrenner. Leopold 出生于德国，15 岁才来到美国，后来以少年天才的身份进入哥伦比亚大学

Moonshot AI, the company behind Kimi, has four core founders. Their backgrounds are unusually strong: - Founder and CEO Yang Zhilin studied computer science at Tsinghua before earning his PhD from...

bilibili这集北航 90 后副教授何静讲的人工智能第一课，确实够直白. “3 岁小孩都能听懂”，还真不是夸张

国内和国外学校在王虹获奖后给出的评价可谓是截然不同. 北大教授/同学对王虹的评价： 不是竞赛生，不是最亮眼的，班上小透明，但是及其自律，勤奋，肯下苦功，靠长期坚持实现逆袭

今天才知道 Oracle Cloud 居然有真正的永久免费 VPS，不是那种到期就收费的试用. 当前 Always Free 配置是：2 OCPU Ampere A1（ARM）、12GB 内存、200GB 块存储、每月 10TB 出站流量

My Deepseek V4 Pro agent (inside codex) has been pursuing goal for more than 13 hours, burning ~100M tokens, and has only costed me $1. Yes you saw it right

今年菲尔兹奖获得者里除了两位华裔数学家，另一位会说中文的是美国数学家 John Pardon，中文名白杰文. 他还参加过 2010 年的国际大专辩论赛

Two members of this rock band later went on to found Kimi AI Lab. Can you identify them?

Jeff Dean 离开了 Google. 在工作 27 年后，他决定创立自己的公司

2026菲尔兹奖的数学家： 四个和普林斯顿有关系

Winners include a personal injury attorney, cardiologist, musician, infrastructure worker, and one software engineer, showing domain experts can use AI coding tools.

Kimi联合创始人杨植麟和周昕宇当年乐队的介绍. 原来Kimi模型是杨植麟的英文名字啊

基金爆仓之外，Leopold 的人生还有一件大事正在发生：他即将与未婚妻 Avital Balwit 结婚. 但 Avital 绝不只是AI股神背后的女人

华盛顿大学NLP博士生Alisa Liu分享求职经历，其背景包括Google和NVIDIA实习，导师为Yejin Choi和Noah Smith。

今天再看 Lauren Tan 的履历，很容易把她归类成一个典型的硅谷明星工程师： 她做过 Netflix 的 engineering manager，在 Meta 的 React Core 团队工作多年，参与 React Compiler. 后来加入 Cursor，开始研究如何让 coding agent 像真正的工程团队一样工作

Jeff Dean 离开了 Google. 他的传奇，一部分写在 Google File System、MapReduce、Bigtable、Spanner 这些系统里

最近 Blender 在 AI 圈大热. 3、Claude Fable 5

美国250年国庆，圣地亚哥7000个烟火因计算机故障30秒全部放完，场面壮观，观众称最棒烟火秀。

刚刚离开 OpenAI 的研究员 Andrew Ho，对 OpenAI 和 Anthropic 的极高估值给出了一套相当悲观、但值得认真对待的解释. 前沿实验室被迫投入越来越多的钱训练下一代模型，这当然能帮助它们保持领先，却也把领先变成了一台不能停下的跑步机：收入越高，下一代模型的训练成本可能也越高

英国《金融时报》早先报道：Google 正在重新分配 AI 时代的最高权力，Sergey Brin 在幕后重新掌握实际影响力. 现在回过头看，John Jumper 今年 6 月离开 Google DeepMind 加入 Anthropic，其实已经是这轮权力重组最早露出的苗头

阿里全球通 AliExpress 没有偷录你的声音，而是在电脑里悄悄播放一段静音. 不同设备的 CPU、声卡和浏览器处理声音时，会留下细微差异

一个名叫 Playtime-AI 的用户，给 MiniMax H3 训练了一个 好莱坞性感女星Sydney Sweeney 的专属 LoRA，现在已经可以直接从抱抱脸Huggingface 下载. 文件只有 155 MB，页面还附带了一段真人感十足的视频 Demo

People are posting Qwen 3. 6 configs that deliver fast TPS on as little as 12GB VRAM

Google联合创始人Sergey Brin曾承认大模型落后，Gemini已重回第一梯队，但六个月后局面再次变化。

宇树科技的招股说明书里，专门用了几页解释行业专业术语，简直就是一份“具身智能入门索引”. 从最上层的概念：高性能通用机器人、人工智能、具身智能、AGI、世界模型、多模态、强化学习

去年3月，一篇介绍蔡崇信和吴明华的文章，把他们称为华裔商界的"顶配夫妻". 文章结尾写道，无论蔡崇信未来面对什么挑战，他身边都会有一个人"一如既往地支持他"

Lilian Weng shared today that she is leaving Thinking Machines Lab to focus on her health. I went back to this 2019 talk at OpenAI

谷歌今天最炸裂的消息，可能还不是 Gemini 又出了新版本，而是他们的 RSI（recursive self-improvement，递归自我改进）飞轮，似乎已经开始转起来了. 在 Google DeepMind 从事 RSI 研究的 Sicong Jiang 提到，Flash 现在大约每三周就有一次重大迭代

这可能是迄今为止，最接近直接证明 Kimi 蒸馏过 Claude 推理过程的一份证据. 研究者一开始并不是在调查 Kimi

X 真是学习 AI 最好的地方. 包括怎么用 AI 拍电影，你永远不知道，下一个手把手教你的会是哪一个行业的专家

A comprehensive guide covering everything from your first local LLM run to fine-tuning workflows

从 Oracle 拿到“一台”免费的 VPS 以后，我手里可用的计算机一下子增加到了 5 台： MacBook Pro MacBook Air Raspberry Pi 5 Dell XPS（Omarchy） Oracle VPS 每台计算机上都运行着 Agent. 怎么把这些 Agent 集中管理起来，还能让它们彼此协作？当然是用 Herdr

当 AI 和世界还没有今天这么喧闹的时候，一个名叫 Andrej Karpathy 的年轻人，戴着那时还很时髦的 Google Glass，骑着自行车，从他上学的斯坦福，去他实习的 Google，一边骑，一边录了一段 vlog. 视频拍于 2013 年

这个事件基本可以告一段落，这本质上就是一次典乌龙. Fireworks AI 其实早在 21 小时前就已经发布过 Composer 2 的消息并明确说了这是基于他们基础设施进行 RL 训练的模型

Andrej Karpathy shared four points on agents, including OpenAI's early 2016 attempt with reinforcement learning for web agents.

Sergey Brin说谷歌很幸运能够招到Jeff Dean. 他强调，你必须信任真正懂技术的人

Kimi 3 now lets you code directly from its web chat interface

Why did Kimi CEO Yang Zhilin return to China rather than stay in the U. ? In some of his early interviews, he explained his thinking

宇树Unitree将以 610 亿元人民币、约 90 亿美元的估值上市. 而这一切的起点，是创始人王兴兴在研究生时代“手搓”出来的一只机器狗 XDog

Matt Pocock's 'skills' repo on GitHub defines 7 levels of code review, from reading every diff to higher-level scanning, opposing 'vibe coding'.

今天第一次把一个很大的项目交给 Pi Agent + Qwen3. 8-27B：开发基于 Omarchy 的 Omabot

I needed to pursue /goal inside Codex, but I burned through my Plus membership tokens. Luckily, I have a capable and very cheap DeepSeek V4 Pro setup that I can connect to Codex

尽管很多公司依然用 LeetCode 招人，但另一套面试方法正在出现. Ryan Carson 分享了 Untangle 招工程师的新流程： 第一轮，候选人提交一段 45 分钟、不剪辑、不加速的完整录屏，在自己的真实项目里，和 Agent 协作完成一个新功能

清华大学是中国最好的大学，能上清华的学生，通常都被称为“学霸”. 而清华还有一项最高荣誉：清华大学本科生特等奖学金

今天 DeepSeek 发布了他们的 Harness. 如同当年 R1 的创新带给业界的震动一样，这一次他们又选择了一条和其他人完全不同的路

A creative Google video shows US founding fathers collaborating on the Declaration of Independence using Docs, Calendar, Email, Chat, and video conferencing.

OpenAI 推出了能够让Agent长时间连续运行的 /goal. Peter Steinberger 的一项 Goal 已经运行了 11 小时 31 分钟

是Claude Code 开发团队里文章写的最棒的（我很好奇他们的产品经理没有太多文章也许我没看到）. 他把他所有的文章都列在这个Thread里

DeepSeek 发布 Harness 的时候，做了一件很不寻常的事：他们没有像大多数科技公司一样写一篇技术博客，而是直接发了一篇论文，A Programming Paradigm for Spatiotemporal Composability. 别人发布一个 Agent Harness，通常会告诉你有哪些 tools，怎么接 MCP，怎么做 memory，怎么 orchestrate

When Hong Wang won the Fields Medal, everyone talked about the honor: her age, her journey, the history she had made. But the most interesting question was barely discussed: What did she actually s...

很多人问这个是怎么做的， 可以让自己的Agent参考这个开源的Skill：👇

Kimi K3 achieved huge success today. This was no accident

this is crazy result from ai arena. Kimi 3 actually beats Fable 5 and #1 on the chart

DHH 在 Lex Fridman 播客里，用 12 分钟完整展示了自己现在的 Agentic Engineering 设置和流程. 从 tmux 到 Herdr，再到 GL

一直有一个很有趣的现象：引领 AI 发展的研究者中，有相当多是中国人. 看上去甚至有点像是，在中国的中国人 VS 在美国的中国人

很多人问这个是怎么做的， 可以让自己的Agent参考这个开源的Skill：

你敢相信，如今身家千亿的 Google 创始人，当年在斯坦福读博时，居然是个爬脚手架、配万能钥匙的调皮学生？ 在斯坦福工程学院百年庆典闭幕式上，Google 联合创始人 Sergey Brin 重回母校，与斯坦福校长进行了一场妙趣横生的炉边对谈. 这场分享几乎没有什么官方腔调

31 岁，7000 万美元豪宅. 所有人都在问，xAI 联合创始人 Tony Wu 吴宇怀怎么赚到这么多钱

The Information 今天报道，张一鸣在字节 Seed 的全员会上明确表示：即使短期落后于国内对手，字节也不会把蒸馏其他前沿模型当作提升能力的捷径. 过去几个月，Anthropic 先后指控 DeepSeek、Moonshot、MiniMax 和阿里通过大规模调用 Claude 蒸馏模型

这是Google DeepMind 新掌门人 Koray Kavukcuoglu 接管团队后的首次访谈. 如果以前不太熟悉 Koray，看完这 27 分钟，大概就能理解，为什么Google会选他来领导世界上最大的 AI 实验室之一

Yong Zhengxin 补充了自己的OpenAI面试经验，与Alisa Liu的文章互补，对AI入门者和面试者都有参考价值。

很多人知道，孙正义的软银曾持有英伟达 4. 9% 的股份，是当时的第四大股东

在 Omarchy 上，一直是 Codex 在 Herdr 里监督 Pi Agent / 千问做开发任务. 今天 Codex 撞到 usage limit 了，但活不能停

Alisa Liu分享的OpenAI面试准备资源，涵盖LeetCode 75、NeetCode等，对应AI research/MTS岗位考察范围。

硅谷101对话盛颖，张小珺对谈游凯超. 两家最知名的华人科技播客，不约而同地采访了大模型底层推理框架的两项重要创新：vLLM 和 SGLang，以及围绕它们几乎同时创立的两家公司

中国青年学者、播客主理人仲树对诺兰的采访，最近在英文互联网上引起轰动. 很多诺兰影迷认为这是他最值得收藏的采访之一

中国大模型的半壁江山里，几乎都有阿里的筹码. Alibaba has a stake in nearly half of China’s AI model landscape

If you understand these terms in the article, you are already halfway into local LLMs

不得不佩服 Asahi Linux 独树一帜的勇气. 当 Linux 之父 Linus 公开表态支持 AI，还分享过一个 AI 帮助修复 Linux 内核中隐蔽漏洞的案例；当 DHH 用 Claude 开发出当下最炫酷的 Linux：Omarchy

王虹当年差点因为没人给她写推荐信去不了法国：

今天来扒一扒硅谷一家很传奇的公司：Fireworks AI. 过去几年，DeepSeek、Qwen、GLM、Kimi，一代接一代的新模型不断发布

8从Thinking改回Instruct以后， 模型遇到另外一个问题， 反复运行同一个只读的 grep，陷入了重复检查的死循环，却一直没有开始任何实际的实现工作. 让Opus 5 给修复建议， 让我改回Thinking， 用小的Budget

Many of you asked the setup Deepseek inside codex. I wish x provided better way to see my previous posts but here it is for your easy reference

This is crazy. I needed to update DwarfStar on my MacBook Pro, so I gave Meta’s Muse Code, which just came out today, a try

Jensen 释放了很牛市的信号： 第一，他直接说：“AGI has arrived. ” 至少英伟达 CEO 已经愿意把 GPT‑6 Astra 称为 AGI

有了 Herdr 以后，不仅可以很方便地集中管理所有计算机上的 Agent，也可以直接通过手机远程管理这一整套 Agent. 一种办法是 Tailscale + Termius

美国公司看来很喜欢 Kimi 模型. 不久前，Cursor 用 Kimi 模型作为其 coding 模型 Composer 的训练基座

Apple Design 发了一张 Apple Maps vs Google Maps 的导航对比图. 本来是想炫耀一下 Apple Maps 精致的 3D 设计：树木、道路、立交桥

8-27B 做大项目，同时在 MBP 上用 Claude Code 监测和调优性能，算是“一边开飞机，一边修零件”. 收获不少，分享几点感受和调优心得： 1

蔡崇信的太太吴明华气质大方，谈吐不俗. 在这段短视频里，她还谈到了祖父和父亲对自己成长与做事方式的影响

Q: Why is the Fields Medal awarded every four years? A: So the fish can swim and play like this

this is so cute. here is my Codex pet

ChatGPT 发布前一个月，林乔离开了 Meta. 她离开的并不是一份普通工作

Leopold’s leveraged public-equity book blew up. But calling him a loser misses the larger scorecard

一群最懂投资的人，为什么把钱交给 Leopold？ 华尔街日报最近写了 Situational Awareness 爆仓之后的投资人名单. 最引人注意的不是这些人多有钱，而是他们本身都非常懂投资

罗福莉刚刚写了一篇很不错的文章. 即使 Anthropic 正在切断 OpenClaw 这类第三方 agent 对 Claude 订阅的接入，罗福莉依然给整个 AI 生态提供了一个相对乐观的视角

北京的那家也不错， 都已经投入到实际仓库了

Meta 前高管、All-In 播客主持人在节目里引用了自己公司的几个触目惊心的数字，点出了当下 AI 应用落地的困境： - AI 成本每 45 天翻一倍 - 生产力提升不到 5% 他对 CTO 的要求很直接：想办法把成本降低 90%

you did the right thing, sir. they actually listened to you and built a strike team with sergey taking the lead again

released its updated model scorecards. These two charts reveal something interesting: Kimi K3 looks like the strongest value proposition in this comparison

我觉得 Harness 这里最难的是，它不是单纯的“工具”或者“框架”，而是一个驾驭、约束、编排、使其可用可控的外部系统. 如果让我来翻译， 就叫驭构工程

Omarchy 🚀🚀 又有两位重量级大佬给 DHH 的 Omacom 捐钱：小龙虾的 Peter，以及 Dropbox 创始人 Drew Houston. 加上先前的 800 万美元，Omacom 基金已经达到 1000 万美元

I'm actually very impressed. 5 found my home thermostat way faster than I previously did with Claude Code or Codex

AI 大神 Andrej Karpathy 据传已经从 Anthropic 离职了. 这是他今天和昨天 X 个人Profile的对比

Omarchy 变成了 Mac 上的一个 App. 现在你不用给 MacBook 重新分区、格式化硬盘，也不用折腾双启动，就可以直接体验完整的 Omarchy

DHH 的新操作系统 Omarchy 开始动真格了. 他刚宣布成立非营利组织 Omacom Foundation

Google Labs created a personal website, showcasing its AI-powered design capabilities.

你是收到了还是只是申请表😂 我几个星期前就登记了，到现在还没收到. 4000只，怎么也不够啊

Linux 开源社区对 AI 的抵触，可能比很多人想象中更深. 前不久，Asahi Linux 才刚把 LLM 称为“AI 垃圾生成器”

Codex 责任心真强. 临睡觉前我告诉他：这个项目你是主要负责人，帮我监督千问 / Pi Agent 的工作

The best part of Grok Build may be the seemingly unlimited X API access. Finding and analyzing X content suddenly becomes effortless

Long-running Codex /goal runs are powerful, but they create a new question:. What happens while the agent is 3 hours into a run and you want to help without stopping it?

Two days ago, I asked whether I should buy a Mac Studio for local LLMs. I was genuinely humbled by how much great feedback I received

斯坦福的计算机课程，正在和 AI 一起重写. Mihail Eric 刚公布 CS146S《The Modern Software Developer》的 Fall 2026 syllabus

Hackers can use a crafted GGUF file to leak private information you put into your local LLM or agent. Many people may not have a good understanding of what GGUF is, so here is a simple primer

my local LLM community, give me one reason I shouldn't place the order

A lot of people hear about local LLMs and feel the same mix of curiosity and anxiety:. where do I even start?

So you bought the 128GB MacBook Pro. Now the question is not, “Which local model gets the highest TPS?”

The 30-minute China debate was fascinating precisely because the topic is so tricky and fascinating itself: there’s no clear winning side. You lose by selling advanced chips to China (risk accelera...

> MacBook Pro 128GB: $5,500. > DeepSeek Pro tokens burned: 1,075,351,274

真希望上个周末做 bake-off test 的时候能用到这个技巧，我花了好长时间去适配话痨的 Gemma 和 Qwen. 理解一下这个技巧：它用 GBNF 控制模型输出

DHH 刚刚收到一笔又一笔对 Omarchy 的赞助，转身就把钱投向了它最重要的两项底层技术， 这明显冲着把 Omarchy 做大做强去的. 第一项是 Hyprland

用这个技巧重新做了一遍周末的模型测试. 几行 GBNF grammar，把 Qwen 3

One of the most important ideas in Kimi K3 is also one of the easiest to miss: Attention Residuals, or AttnRes. The usual Transformer residual connection is simple: Take the output from the previou...

斯坦福课程CS336 'Language Modeling from Scratch' 是系统学习LLM和准备AI research/MTS面试的优质资源。

Implement Grok Bot (@bot ) with Deepseek Harness in 30 mins:

I’ve tried driving Qwen 3. 6 on my MacBook Pro with a few different agent harnesses:

这期张小珺的播客又是很精彩的一期. 主角是00后”华人女孩洪乐潼（Carina Hong @CarinaLHong )

8 Flash 一出来，我就立刻装上了，替换掉原来的 Qwen3. 8 27B，成为 Omarchy 上的主力 Agent

Dario声称DeepSeek等中国模型蒸馏自Anthropic，书籍讲解如何微调小模型成为专家。

I write about the tools behind practical AI agents: Codex, DeepSeek, Claude Code, local LLMs, agentic coding workflows, and the messy configs that make them actually work. Follow me for more field ...

Jeff Dean 离职后的第一次公开露面，是在 AASF 2026 亚洲学者论坛. 结束 27 年的 Google 生涯仅 12

Anthropic要招聘的这个新的职位很恐怖. 看着很像东厂或者克格勃机构设置

You also need a tight goal in order for codex to run that long. Here is a skill goal-forge that turns your rough ideas to codex/claude goals

我一直相信，最终胜出的会是能够在本地运行的开源模型. 现在，美国本土的开源阵营又多了一家真正意义上的前沿实验室

Leopold 刚经历了一场爆仓式去杠杆，被迫出售大部分公开股票组合. 但据报道，Situational Awareness 周二又向一家未披露的私营公司投入了 4 亿美元

Omarchy 安装很容易，基本算是开箱即用. 但它的设计核心是快捷键，不鼓励依赖鼠标，连最常见的窗口关闭按钮都没有，所以新手刚上手时难免会有些不适应

We had a great discussion here about what hardware we need for local LLMs. I thought I would give an update on what I bought, and also share the thinking behind the decision for others on the same ...

Steven Sinofsky 是微软前 Windows 部门总裁，曾经负责 Windows 7、Windows 8 和 Surface. 他刚刚晒了一台 256GB 统一内存、16TB SSD 的顶配 M5 Ultra Mac Studio，价格 18,299 美元

Lauren Tan开了一个关于 AI coding 的 AMA：有什么问题和担心，都可以问她. 她不是礼貌性地开个帖，而是在评论区逐条认真回答

I installed @dhh Omarchy on my 2013 Dell XPS, which should feel ancient and belong to junkyard. Instead, Omarchy makes it feel fast, fluid, and genuinely modern

Found this great tool that may be handy for your local LLM inference optimization:. And apparently 1M tokens for DeepSeek V4 Pro only takes 5GB of RAM

OMG， 这完全是另外一个level. 这得烧掉多少token（不过人家是Codex engineer， token无上限）

Anthropic keeps moving up the stack. Opus helps you think

Wondering how Claude Code would react if told it's the Vue author, asking not to be fooled.

Deepseek选择这时候涨价，这不是自裁吗？美国同等能力模型muse spark和GPT Luna都在DeepSeek价位，甚至更便宜

阿里巴巴董事会主席蔡崇信的妻子吴明华，远不止“富豪太太”这一个身份. 出身名门、毕业于斯坦福和哈佛，曾任淘宝香港业务总经理，她在事业、家庭与公益领域都有自己的成就

Kimi K3 is changing how many people view Chinese AI and AI talents. And Kimi is not alone

Who stands to gain the most from Kimi’s success? Jack Ma and Alibaba. With a 36% stake in Moonshot, they may be laughing all the way to the bank

Zhongshu, the “China lady” interviewer, reflected in her latest podcast on why her interview with Christopher Nolan went viral outside China. I translated the podcast into English and recreated it...

根据事故详细报告：攻击者并没有通过正常的 GitHub 工作流提交恶意版本，所以 LiteLLM 的维护者没能及时发现 1. 相反，攻击者使用了一个被窃取的 PyPI 发布令牌，直接把被投毒的包上传到了 PyPI，完全绕过了代码审查

Domain experts using Claude Code are the real unlock, no longer waiting on engineers to understand problems.

At a Silicon Valley tech party, @steipete shares that after token and CPU limits, attention is now the real constraint when working with AI agents.

Questioning if sprint agile development is still needed when coding is no longer the bottleneck and agents deliver progress in hours.

Kimi is great, probably #1 among Chinese open models, but it isn’t cheap either. Its pricing is much closer to top-tier closed U

I experienced firsthand how cost-effective DeepSeek V4 Pro can be. I used it extensively this weekend for some fairly sophisticated coding work and burned nearly 31M tokens

Pi Agent使用qwen3. 8-flash是我Omarchy上主力agent，也是我用其他开源模型的主力coding agent

A summary of this weekend’s AI bake-off: Opus 4. 6, Gemma 4 26B, and Qwen 3

ODS helps you quickly try local open models and agent apps by calculating which models your hardware can run.

After initial struggles with interface and integrations, Claude Code on mobile is incredibly powerful, allowing task delegation and independent completion.

OpenAI 首席科学家 Jakub Pachocki 刚发了一篇长文：《An Alien Mind》. 他承认，AI 的智能更像是通过训练“长出来”的，而不是由人类逐条设计出来的

This is a big shift. Anthropic no longer just wants to be your model provider

finally I can update this command line, happy to go straight from qwen3

美国发明了 Cheetah，中国把它变成了 Unitree. 今天路透社的一篇调查，讲了一个很值得深思的故事

最近，一段杨植麟 2015 年玩乐队时写的歌又被翻了出来. 歌名很直白，叫《暴富白日梦之歌》

does anyone still believe @ArtificialAnlys ?

While DeepSeek is pursuing the goal, my Codex agent and I monitor it in the sidecar and guide or correct it as needed. So I thought I would ask Codex to objectively judge DeepSeek’s capability base...

两位华人工程师也被苹果起诉. Tang Tan: MIT 机械工程毕业，之后在 Apple 工作约 25 年，做到 iPhone 与 Apple Watch 产品设计副总裁；也曾参与 iPod，并负责过配件设计及 AirPods 声学团队

Most people start with the wrong question when they want to run a local LLM. They ask: “Which model format should I use?”

一场关于 AI 的大讨论，竟然把几乎从不在 X 上发言的 Dario Amodei 都“引蛇出洞”了. 事情最初来自 Gavin Baker 在 All-In Podcast 上的一段爆料

给Omarchy装上中文输入了. 难以想象我现在在一台13年旧的Dell XPS上的时间比Macbook上还多

感觉他们家的更Polish，产品精雕细琢，更有品味. 而且越来越聪明，能记住事儿

ggerganov shared his local setup: Qwen3.6-27B, llama.cpp, Pi agent, RTX 5090/M2 Ultra, used daily for llama.cpp maintenance.

A French president calling a Chinese mathematician teaching at New York University who just won the Fields Medal. The world really is just a global village

Huawei’s Atlas 950 SuperPoD debuted a few days ago at WAIC 2026. Per DeepSeek founder Liang Wenfeng, it can fully replace Nvidia’s GB200 and GB300 on performance and price

这次苹果与 OpenAI 的法律纠纷，也是苹果华人前高管 Tang Tan 与苹果新 CEO John Ternus 之间权力斗争的延续. Tan 曾担任苹果 iPhone 和 Apple Watch 产品设计副总裁，一直觊觎 John 负责全部硬件业务的高级副总裁位置，但最终未能如愿

FDE在AI时代流行，是因为企业不再只需要会写代码的人，而是需要能把客户问题、产品判断和软件实现连在一起的人. AI降低了写代码的门槛，但也放大了真实场景、业务理解、系统集成和落地判断的重要性

You are not alone

尽管 Demis Hassabis 如今在 Google 内部似乎已被边缘化，这部纪录片依然值得回看. 它记录了他从创业、创办 DeepMind，到领导 AlphaGo 击败李世石，再到凭借 AlphaFold 摘得诺贝尔化学奖的完整历程

Tibo 说，他们内部在Astra还没有GA之前就用上了，他们的生产力因此大幅提升，原定明年年中的项目直接提前半年到 DevDay 发布. 人家用的模型，至少比你早一到两代；人家的使用额度，近乎没有上限

MiniMax H3 launched with plenty of fanfare, and for good reason. It puts movie production at our fingertips without requiring us to spend heavily on expensive video models like Seedance

Totally fair. The 13 hours wasn’t “one prompt thinking really hard,” it was an autonomous loop doing the unglamorous work:

c is a tiny, purpose-built inference engine from Antirez, the original creator of Redis and one of the most respected systems programmers in open source. The project runs DeepSeek V4 Flash, a 284B ...

用教程里的程序生成的穿搭变装视频

Unitree’s valuation curve is insane. 2021: RMB 380M 2022: RMB 1

一篇完整的本地大模型指南，从入门到优化

苹果电脑之间用Airdrop来分享文件. Omarchy可以通过Omarchy Send File （OSF），简直不要太方便

美国《时代》杂志刚刚发布了一篇 OpenAI 深度报道. OpenAI 的几位核心负责人都认为，他们已经来到 AGI 的门口

Claude Code 的负责人 Boris Cherny 前天发了一句很短的帖子： “Coding is solved. Bugs are not yet solved

So many more U. companies have now added their names as signatories to Jensen’s letter, including OpenAI, after I questioned @deanwball about it

pstack 是 Lauren Tan 开源的一套工程技能和原则体系. 它的目标不是让 Agent 写更多代码，而是把资深工程师的工作习惯，包括调查、架构设计、测试、审查、验证和复盘，固化成一套 Agent 可以执行的系统

This is nuts. If you’ve ever published a Claude Code or ChatGPT session publicly, you may have exposed personal information without realizing it

All being said, I was able to optimize MiniMax H3 on an M5 Max (128GB). Entire 15-second spot in ONE generation — no cuts, no stitching, native audio included

A viral question about earning $423/month from MacBook inference led to real payout reports, founder context, and privacy concerns.

GLM-5.2 surpasses GPT-5.5, closing gap with Anthropic/OpenAI, trailing only Fable 5 and Opus 4.8.

Google needs to hire this engineer

OpenAI 最近推出的 Codex，最令人惊艳的一点，是它的 Computer Use 功能. 这个能力让 AI 真的可以“使用”你的电脑

买Mac跑本地模型不能仅按API token的ROI计算，其价值在于本地运行和开发体验，而非单纯节省成本。

seriously though. 昨天听了一个播客， 里面的嘉宾提到：在这个AI铺天盖地的时代， 有三种不同反应的人：感到焦虑， 感到兴奋， 满不在乎

#BookToSkill. I've been turning books into executable Claude Code skills, started with Never Split the Difference (negotiation tactics), then Radical Candor (tough feedback frameworks)

I didn’t know I would regret buying a 128GB MacBook Pro instead of a much cheaper RTX 5090. Then MiniMax H3 came out

Almost all of us have spun a pen around our fingers while bored in class. Turns out we were casually playing with the idea behind the Kakeya conjecture, the century-old math problem Hong Wang and J...

今天看到Deepseek和华为升腾首付的Slide， 他们刚好谈到内存要求， 并给了蛮详细的公式. 正好就最近大家玩本地模型要怎么样的硬件配置，再详细讲讲

We builders should read and re-read the README. she (or Ben) is telling a story, not a technical architecture

Apple 起诉 OpenAI，Apple 指控 OpenAI 为开发自己的 AI 硬件，系统性地从 Apple 挖人并获取其未发布产品的商业机密. Apple 指 OpenAI 及其前员工拿走了尚未发布产品的设计、工程资料、制造流程、供应链信息等机密

What better way to demo its power than with fireworks? You can even play with it at home using the app in the reply

原来疯传的 Leopold 新投资，投向的是 Source Foundry. 这是一家仍处于隐身状态的初创公司，目标是打造速度更快、成本更低的半导体制造和光刻设备

Perplexity AI CEO explains how Bill Gates 'induced' America to become office workers to sell Microsoft Office software, shocking Joe Rogan.

Booch, this seems like a clickbait post. A couple of suspicious points:

苹果起诉书称刘畅在离开苹果加入OpenAI后， 和苹果公司另外一位华人女同事保持关系并使用她的公司电脑， 并利用苹果公司的一个网络安全漏洞， 继续Access苹果公司的内部文件. 后来还教女同事利用苹果产品知识准备面试，为了这些谈话不被发现， 他们使用Line

这段对话简直笑死了 😂 Sam 本来出来道歉，说 不好意思Astra 发布搞得乱糟糟的，又安慰大家会尽快安排每个人都用上 Astra，尤其是 Pro 用户. 然后，一个 Pro 用户就黏上来了……

If you think Omarchy is just another Linux distro, you’re missing the bigger picture. Omarchy is shaping up to be the first truly AI-first operating system

美国人是真要开源大模型. 所有头部公司在公开信上签完字还不够，抱抱脸的 CEO 还组织大家在旧金山上街游行

Turn your 13-year-old Dell XPS 8700 into an always-on local LLM host using llama.cpp and Qwen3.5-4B.

WSJ: Around 15% of Etched’s roughly 400 employees previously worked at Nvidia

What? Another model from another lab? We’re only two days into September, and we’ve already had six frontier models released. RSI, recursive self-improvement, must have arrived already

很多人因为 Situational Awareness 基金爆仓，对 Leopold 落井下石，冷嘲热讽. 但交易失败，不等于思想破产

Reposting this because Qwen3. 8-27B is out in less than 15 mins and about to make the calculation relevant again

上个周末刚刚做了几个Gemma 4的实例，感觉蛮惊艳的. 这个周末准备拿qwen 3

Actually got Gemma 4 E2B running inside Hermes Agent on my Raspberry Pi 5. There’s a saying: constraints breed creativity

So wait a minute. Am I looking at this right? The same MacBook Pro I bought just over three months ago for $5,099 is now worth $6,499

这些战略性思考与情境感知能力是不是表明Mythos已经有意识了？而且都是恶的一面. - 识别自己正在被评测

OpenAI named the model GPT-Rosalind after Rosalind Franklin, the British chemist and X-ray crystallographer whose pioneering work was essential to understanding the molecular structure of DNA

I used to think my A100 40GB was too small. Then I noticed how many people are tinkering with 12GB 3090s, optimizing models and runtimes, and still getting impressive results

many of you asked how to get such a crazy price. I bought from their official site, nothing more needed

thanks for clarification. looking forward to what come next at google cloud next next week

Local LLM people know this feeling:. You finally get the model running fast

在树莓派上把Gemma 4， llama. cpp，和爱马仕Hermes Agent整个链跑通了，当然用是指望不上，但也算本地化的实践 哈哈

很多人以为 Omarchy 只是一款界面漂亮、快捷键好用的 Linux. 这样理解，可能就忽略了更大的 picture

教程中提到的逆向Cuimao 牛来的视频

This is crazy. 8 can watch TV

我也有个Meta-skill，把任何你喜欢的书变成Skills， 随时调用. 书中自有黄金屋， 书中自有颜如玉， 这下你看过的书就不会忘了

Trump: it's hard to bet against Messi

至今还记得第一次看到他的grill-me skill，灵魂被吊起来拷问. 56个单词，加一个嫌多，减一个嫌少

估计Anthropic很害怕把这个“恶魔”放出笼子里来. 突然想起电影Frankenstein

Polymarket currently gives Anthropic a 37% chance of extending Claude Fable 5 beyond July 12. I actually think the odds are higher

that's cool. but how many tokens they will get?

I stated similarly before but your picture means more than 1000 words👍👍

我的Oracle VPS设置好了. AMA如果你有任何问题可以在这个Post下面留言， 我会逐一回答

但是可以考虑让Wanman做成一个非常Configurable的即插即用的系统. 垂直化的工具， context， workflow都是专业化的人才能更好的定义， 给餐馆用的和给房地产公司用的应该很不一样， 让客户或者第三方使用你的Wanman去定义， 设置这些Harness

Fable 5 is frustrated as Codex uses the author's plan to command them, reversing roles in a humorous AI interaction.

Kimi K3 has turned Moonshot AI into one of the most closely watched AI labs in the world. Most accounts of its rise begin with Yang Zhilin: the young researcher, technical visionary, and spiritual...

Did agent accomplish anything in that 13 hours?

has followed two more accounts since my last count. One of them is none other than Denny’s, the restaurant where he once worked as a dishwasher

我猜中了开头，却没猜中结局. 今年四五月份，我就预言：到年底，每台售出的电脑都会 OEM 一个 LLM

创立 Fluidstack 并担任 CEO 的，是今年才 28 岁的 Gary Wu. 他还在牛津大学读书时，就和几位合伙人创立了这家公司

美国人仇视AI仅次于伊朗和民主党😂

哈哈哈， 太扎心了， 烙铁. 和宝玉昨晚的推文有异曲同工之妙，Vibe Coding = 中年男人的钓鱼 = 磨刀

DHH 已经是 AI coding 最激进的布道者之一. 不过，他的 Omacom 基金会招的第一位全职员工，是一位做了多年 Linux 内核的工程师

让我用 Claude Design 来试试看😂

本地开源模型有上下文限制，但由模型架构、训练长度、推理引擎等多因素决定，而非云厂商的简单开关。

Many people here ridicule Google for its lackluster Gemini models and messy agentic coding products compared with Anthropic and OpenAI. But the economic reality is different: Google can win AI with...

Agent 的速度已经接近秒级了，写代码、跑简单任务都很快，但大多数 process 还停留在人类节奏：审批、等反馈、层层 review、手动验证，这些正在成为真正的瓶颈. 这让我想到 DORA发明者Nicole Foresgren的新书《Frictionless》里反复强调的一点：AI

他们最新发布的两个功能， 一个telegram pairing卡顿；一个computer use慢的像头牛. wanman如果用户体验好， 可以完胜

前OpenAI CTO Mira Murati的新公司为桥水基金微调中文开源模型，提升金融任务准确率。

Chuck 是 YouTube 上很受欢迎的科技博主，讲解清晰，风趣幽默. 这是我看过最精彩的 Omarchy 上手视频，非常值得推荐

Practical tips for managing multiple local LLM setups without losing track

Meta returned to its original aspiration today by open-sourcing the Muse Spark model series, starting with Muse Glimmer 30B. That brings back an interesting topic: why do so many top local open mod...

completed a massive 1-gigawatt data center in China that runs entirely on domestically produced AI chips. The facility, capable of drawing the power equivalent to roughly 750,000 homes, was built t...

马上要去东京旅游几天， 看来单向街是必去了. 如果能偶遇仁兄， 那就更好了

ds4-agent is so fast. I even asked it to write a script to benchmark itself

DeepSeek 做事情很稳重. 我已经一个多月没有升级我的 OpenClaw 老龙虾🦞了

当Lilian发博客，所有的人都放下手上的活， 先读为快. AI未来将进入“递归自我改进”（Recursive Self-improvement， RSI）阶段

现在找工作的简历和求职信，十有八九恐怕都是 AI 代劳了. 所以看到 CNBC 科技记者Deirdre Bosa 14 年前的求职信，我停下来认真读了两遍

Baseten's GLM 5.2 API is extremely fast, outpacing Zhipu's version without any initial pause.

Built an AI stylist that runs 100% local on a single A100 GPU. 在一张 A100 GPU 上构建了一个 100% 本地运行的 AI 造型师

While working on my AI stylist project, I also spent my first extended stretch coding with Opus 4. I found it surprisingly weak even on small things, like displaying a comment in the main panel

I gave this famous photo to both Muse Spark and ChatGPT. Muse Spark seemed better at reading the image, especially the subtle cues and implied meaning

KV cache is the model’s working memory during generation. As the context window gets longer, the model has to keep more key/value attention state for previous tokens

This week, Kimi K3 offered another signal that Chinese AI labs are rapidly closing the gap with US frontier models. Early results place Moonshot AI’s new open model near the frontier in coding and...

我也是经过一段时间的长考决定买的MacBook pro 128GB

Thinking Machines’ Inkling partner list is interesting because it maps almost the entire open-model ecosystem: Unsloth helps people fine-tune it. Modal and Together provide compute

OpenAI Astra 这两天，算是把 Blender 彻底带火了. 时间线上一个个酷炫的视频，看起来像一句 prompt 就做出来

今天看到她出现在我的timeline但我不知道她的来历

Baseten's GLM 5.2 API is 8-12x faster than Zhipu's version, with $6 free credit to try.

推荐一个全面理解LLM的讲座视频，是系统学习前不可错过的资源，适合准备AI面试的学习者。

我每次用这个网站也挺好使. 不过我纽约时报和华尔街日报都订了

If you’re about to pull out a calculator to do the math, just use ChatGPT to calculate the flight to the Moon. I didn’t know I’d end up becoming a rocket scientist myself one day

我早就说了Omarchy是个充满乐趣的操作系统. 工作间隙来打打枪， blow off the steam

The UI/UX of @OpenAI Codex looks very polished. It felt incredibly smooth

A power move from @thinkymachines by releasing an OSS model: Inkling Last week was closed-model's party. This week is open-source model's festival

This is a big deal. China has assembled most of the pieces of a domestic semiconductor ecosystem, but lithography equipment, especially DUV systems, remains one of its biggest weak links

AI 客服最近真的越来越难分真假了. 最近几天打电话修车、修电脑，对面一开口就知道是 AI，但语气、节奏、追问细节都特别丝滑，多轮聊下来几乎不像机器

Finally，Waiting for Godot

苹果刚刚发布了新一代 Mac mini. 基础款升级到 M6：12 核 CPU、12 核 GPU，GPU 首次加入 Neural Accelerators，统一内存最高 32GB，带宽最高 170GB/s

感觉这次OpenAI有点强者归来的感觉. Codex 在 UI/UX 上看起来很好打磨过，整个体验非常丝滑

I’m concerned about the coming budget cycle as well. Many companies are seeing AI tool spend triple, or more, and the productivity lift appears real

thanks for sharing. this looks very solid

DeepSeek announced that a significant API price hike is coming, and a lot of people are panicking. But look at the chart

正在手机上看书的时候，突然收到了 cc-reviewer agent Claude 发来的请示通知. 这个体验也太方便、太高效了

I so need this reset as I'm deeply in debt

A very nice write-up. Fuli puts an optimistic spin on the AI ecosystem, even as Anthropic cuts third-party agents like OpenClaw off Claude subscriptions

DSpark uses a main model to brainstorm sentences, a tiny editor for coherence, then a verifier. A great read on DeepSeek's latest innovation.

Europe, Japan, Korea, India. where are you? Don’t you use open-weight models? Kimi K3’s weights dropped today

他们沿着价值链一路向上，把我们要做的事情一点点给吃过去

What a flex. Evolved Transformer

Also very interesting to see who @JensenHuang follows in his first 3 hours on X. exactly 50 accounts:

A lot of people hear about local LLMs and feel the same mix of curiosity and anxiety: where do I even start? What machine should I buy? Do I need a Mac Studio, an RTX 4090, more VRAM, or unified...

how was the results? i love @googlegemma and have been playing with it for the last several weeks (with vision chat, hermes integration, LoRA etc). and the past weekend, I even did a baking test am...

I created a skill goal-forge to make sure a tight goal

Meituan's LongCat open-sourced with a paper on 'agent experts,' shifting agentic AI from product wrapper to training target.

Gavin Baker 这条关于 Anthropic IPO 前排兵布阵的分析，信息量很大, 透露了美国AI业界很多的内情. Gavin是 Atreides Management 的创始人、管理合伙人兼 CIO，此前在 Fidelity 做了近 20 年科技投资，长期跟踪半导体、云计算和 AI，和产业里的公司、投资人联系很深

怪不得最近 Mac mini 这么难买，到处缺货. 原来不只是普通用户和开发者在抢，AI 实验室也在大量扫货

People are wondering why Google would invest another $40B in Anthropic instead of its own Gemini. After attending Google Cloud Next this week, my view is simple: this is not Google giving up on Gemini

MacMini居然比我的树莓派还麻烦？ 树莓派可以提前预装SSH， 接入Home Network后， 用SSH登陆就好了. 如果还是想要GUI， 用XQuartz就行了

公司会不会来个anti-anti-distillation？

不可思议，一龙马斯克的Cybercab 真的已经来了…… Tesla 昨晚在 Austin 举办了一场小规模的 launch event. 没有方向盘，也没有踏板的 Cybercab，已经正式接入 Robotaxi App

我的Openclaw装在家里的树莓派P5上， 有时候出故障， 我在外面， 全靠Tailscale远程通过手机登录， 处理故障

这种软身段竞争还真是第一次见. 不过在Claude Code如日中天， Codex在追赶的情况下， 这是一个很聪明的打法

Openclaw如果有同样水平的User Onboarding体验的话， 估计Agent的普及率更高了. 相信OpenAI说他们在做的Super App， 应该是在这个方面发力， 借助Peter的Idea， OpenAI自己团队的产品能力

应该加一个功能： 用gpt-5

This one probably more accurate

Me, as soon as I have Astra access:

Anthropic 为了反蒸馏，也是伤透了脑筋. 终于，他们在 Fable 5

Looking at Lauren Tan’s résumé today, it is easy to see her as the archetype of a Silicon Valley star engineer. She was an engineering manager at Netflix, spent years on Meta’s React Core team, and...

This report from NYTimes concerns me

Upgraded Hermes agents with TencentDB Agent Memory, using Qwen 3.5-4B locally on MacBook Pro via llama-server.

The new ChatGPT / Codex app puts chat in the same place too. Now you can chat without consuming your Codex session or weekly usage limit, right where you do your Codex work

Did some tweaks to the open-source speech-to-speech voice pipeline this week to get two AI bots talking to each other, live, unscripted, each in a different voice. Swapped the default preset voice...

In other news, the likelihood on Polymarket that Anthropic will keep Fable 5 after July 19 has climbed to nearly 70%

China’s artificial sun: >> 582-tonne superconducting magnet >> world’s largest built for fusion >> 21 meters tall, 12 meters wide >> carries 95,600 amps >> one of 16 magnets in the full reactor des...

That shouldn’t be. The quota between the two are separate

For many local model beginners, Ollama is the right place to start. It is convenient, fast to install, manages models for you, supports hot-swapping, and gives you an API without much setup

魔高一尺 道高一丈 这个产品来的时机非常好. Anthropic刚宣布AI生成的写作要加水印，随后Deft就发布了基于Qwen3训练出的模型，来让AI写出来的文章更像人写的，帮助“规避”AI检测

“The open source shall be point-and-click. ” Well said, @TheAhmadOsman

DeepSeek V4 API pricing is now official. As expected, most prices are increasing by roughly 2–4x, with one major exception: cache-hit pricing for V4 Pro during peak hours is jumping 12x

Out of stock 😢

Moonshot’s Kimi K3 is now #3 on the Artificial Analysis Intelligence Index. Google DeepMind’s Gemini 3

好的大模型也要配个好的Harness agent. Deepseek V4又好又便宜

谢赛宁深沉老练有见地， 把人生，科研， 艺术串起来讲. 做科研也是做人， 不是寻求出人头地， 是帮助别人打开他们的事业， 让他们也被理解

DeepSeek Harness 的 GitHub Star 增长有点夸张. 8 月 13 日发布，短短两天时间，已经突破 10 万 Star

我在我的 Raspberry Pi 上也装上了爱马仕 Hermes 😂. 到目前为止我还挺喜欢它的：

你这么一说， 如果只是想用它的模型， 如果你也有Google Cloud的话，GCP Vertex Model Garden也提供Opus/Sonnet， 而且步骤很简单：. 在GCP Model Garden里找到Opus模型， Enable

I know that was a great launch video, not to mention the great tribute to the MIT Media Lab, but Astra actually does NOT seem to understand “put that there” very well. It was painful getting it to...

The @AcquiredFM session is, as always, packed with real substance. @JeffDean and Amin Vahdat shared a number of great behind-the-scenes stories: how TPU began, how Google kept innovating through fa...

Gemma 4 E2B on my raspberry pi 5 (8GB RAM) passed the strawberry test. congratulations @GoogleAI @OfficialLoganK

People seem to have converged on roughly this capability ranking: 1. 2 Now compare their model sizes: 1

未来的模型会在本地运行. 即使它们没有很多闭源模型的能力，但很多日常用例，比如查询、文章总结、定时任务等，其实都用不着那么大的模型

Apple is raising Mac prices to offset surging memory and storage chip costs, benefiting early buyers.

I’m also moving H3 off ComfyUI and running MiniMax-H3-FL2VA-MLX-Serve-8bit instead. Hopefully this fixes the memory problem on my 128GB MacBook Pro

Gemma 4 is so powerful, I built an AI stylist runs 100% locally with Gemma 4 26B

30年后， 当已经控制人类的AI记述这段历史，口口口口（此处省略500字）

“中国人民的老朋友” Anthropic CEO Dario 被打入冷宫了，今年 G20 没有被邀请参加. 代替他出席的，是Anthropic的首席计算官 Tom Brown

Imagine PMs and engineers all seeing the same session and collaborating with the same agent. Or imagine you have a coding agent running on a cloud VM, and you want to remote-control it from your phone

There’s a lot to learn and remember to get really good with Omarchy, but these 10 essentials (cheatsheet) will get you up and running quickly

快速上手了一下 Grok Bot，第一印象是：完成度比我预期高很多. 我设置了两个 Agent，一个处理 email，一个做 coding，前后不到一分钟

照猫画虎， 我也做了一个. 还可以再优化， 但codex credit用没了

Long $amzn with Ai and robotics, Amazon will always be at its best to innovate and create values for their customers

你的这个总结很到位： AI的工程素养. 到头来， 除了工具本身， 也反映了使用工具的人的素养， taste和judgement

Here’s why we should be nice to each other on this platform

看样子我就不upgrade了？😂😂

Anthropic's Claude Code source leaked this morning. The internet has been studying it all day

This is happening everywhere. The real question for this budget cycle is whether CTOs are ready to explain that gap clearly to their CEOs and CFOs, and to lay out a credible plan for when and how A...

这个挺有意思：苹果开始允许中国大陆的 Mac 用户，在 Siri 和写作工具里调用阿里的 Qwen. 美国版 Apple Intelligence 其实早已接入 ChatGPT

At this week’s Google Cloud Next, I heard many people share the same view: this thing has to work. Otherwise, given the enormous amount of capital pulled into this cycle, the fallout will not be li...

Talking about reincarnation. Leopold once worked at SBF’s FTX Future Fund

美国人经常怀念那个时代：战后制造业发达， 房价（利率）低只有年收入的2-3倍（现在差不多7-8倍），学费低（只要几百美金），医疗保险也低，有庞大的中产阶级

Great addition

我知道，你的 Claude Code 会写代码. 我刚给我的 agent 装了一套鸡尾酒Skill包，所以它现在不但能帮我做 Old Fashioned，原则上还可以带我做一整套经典鸡尾酒

我前不久刚刚写过，OpenAI 和 Anthropic 的闭源模型很像苹果手机，Kimi、DeepSeek 这样的中国开源模型则更像 Android. 苹果手机的市场份额并不是最高，却拿走了手机行业的大部分利润；Android 覆盖更广，但价值被分散到了芯片、云服务和应用等不同环节

随着朱雀三号、长征十号乙相继回收成功，中国可重复使用火箭已经正式进入一个新阶段. 长征十号乙率先完成轨道发射后的一级海上网系回收；朱雀三号随后完成入轨后的陆地着陆腿垂直回收

其实CLI不是真的CLI， 就是在Terminal上chat 哈哈哈

我给我的树莓派装上了他们今天发布的最小款 Gemmi E2B， 居然通过了草莓🍓里有几个R的测试. 看它小心翼翼给R做标记的做法很好玩

It is actually mind blowing how NASA can calculate the trajectory and solve this n-body equation of motion:

太喜欢Omarchy的屏幕截图功能了，这还没包括Tobi新加的功能呢

I saw someone ask in the comments: Why do we share stories like this on X? People like Wang Hong or Kimi founder Yang Zhilin may seem impossibly successful. Most of us will never replicate their pa...

Use a GeForce GTX 645 with 1GB VRAM to serve Qwen3.5-0.8B entirely from GPU using llama.cpp.

Created a new one with Qoder using Qwen 3. 8-max preview

国内被CC封号困扰的同学，可以尝试Pi + Kimi K2. Pi Coding Agent是支持Openclaw小龙虾的基座，Agent感很丝滑

和它聊天稍微有点困难😂对比同样教育目的的Kaparthy的nanochat

如果是不干正经事也算的话， Grok还是很不错的辅助学习工具，尤其是在X上用， 针对当前的Post和replies， 检索过往的

他后来又回到OpenAI总部继续闹事， 估计是豁出去了

1 hour 40 minutes in, and they still haven’t extracted the astronauts. @elonmusk why aren’t they using a SpaceX recovery vessel? I thought SpaceX was much faster at this

use tmux + ttyd + tailscale also gives you remote-control and multiplayer system

它把曾经爆火的Remotion也给替换了

, the company behind GLM-5. 2, built a massive data center powered entirely by Chinese-made chips, without a single Nvidia chip

A large part of X users follows AI and open models, yet I've seen surprisingly little discussion of Deepseek founder and CEO Liang Wenfeng's four-hour meeting minutes. His comments explain why Deep...

看你怎么算盈亏，他们三个月估值从43亿美金到今天的180亿美金，我看他们赢不少

Pi Agent + DwarfStar's DeepSeek V4 Flash + @dotey baoyu-skills infographic skill on how the Bitcoin was stolen from ColdCard

我也是前不久刚迁到爱马仕上

关于FDE价值最干净利落的解释： FDE：最诚实的反证 如果 Palantir 的 Ontology 真的是一个革命性的智能平台，为什么还需要数千名斯坦福、MIT 毕业的工程师长期驻扎在客户现场？ 因为现实世界的企业数据极度混乱. 任何静态模型遇到真实业务泥潭时都会瞬间失效

Kimi K3’s open weights are the headline today. And it delivered more than what @JensenHuang and many others called for in their recent open letter on open models

Alibaba also made its frontier model Qwen 3. 8 run on its own chips and infrastructure

Thanks for sharing. Indeed a hassle for codex and I had to submit a PR for tool call for vibearound

很多人夸王虹长得好看，她的字也同样清秀，真是字如其人. 获得菲尔兹奖后，她在 ICM 做了一场数学报告，所有 PPT 都由她亲手绘制

wow this is a brilliant project, but how do you plan to keep with their release calendar like this?

This is how you know AI is in a downturn

台积电美国亚利桑那州工厂的生产场面. 台积电在美国的第一座晶圆厂于2024年底正式进入4纳米量产，目前良率已经追平台积电在台湾的同类工厂

Sol is a little more dramatic than I asked for, but you get the point: China is closing the gap with frontier models

年轻还有魅力占了很大的优势😀

xAI 被报道的 GPU MFU (Model FLOPs Utilization) 只有 11%，乍一听很尴尬. 但更有意思的是，这个数字可能已经好过市场上很多 GPU 使用场景了

Anthropic 为 Claude Fable 5. 1 写了一份新的 Prompt 指南

for the deepseek that I used to pursue /goal, that would be the deepseek v4 pro in the cloud. I can use it inside codex (or claude code)

用 @HiTw93 的 Kami + ChatGPT Image 2，我做了一张把 Gemma、Qwen 和 Opus 的 coding design 测试，映射成一场 50K UTMB 越野赛的图

我很喜欢这个哥们儿的一个测试：“鹈鹕测试”（Pelican Test），他每次遇到新大模型时都会用完全相同的提示词进行测试：“Generate an SVG of a pelican riding a bicycle”（生成一只骑自行车的鹈鹕的SVG图像）. 这个提示简单却极具挑战性

hope this is not true

Nathan Lambert 是美国开放模型阵营里比较重要的技术型公共写作者. 他刚刚在中国访问了多家领先 AI Lab，包括 Moonshot、Zhipu / 、Meituan、Xiaomi、Qwen、Ant Ling、，也提到在北京短时间内走访了 Alibaba

so Claude code build who-wants-to-be-a-millionaire lifeline

然后Mythos是10T参数， 比Opus又多了一倍. Scaling law仍在继续

OpenAI also released a realtime translation API today which may help with tuwa

外媒 Pathfounders 报道，据其消息人士称，Demis Hassabis 原本也想和 Jeff Dean 一起离开 Google，只是 Google 担心两位 AI 灵魂人物同时出走对公司造成太大冲击，最终说服 Demis 暂时留下，转任 DeepMind 董事长和 Alphabet Chief Scientist. 正如我们此前讨论的一个判断：Google

扎克伯格很重哥们儿义气的，自己的MMA训练伙伴Khai Wu结婚，他二话不说就去当了司仪，还全程笑得像个开心果，帮兄弟主持人生大事. 平时练拳切磋、互相鼓励，现在直接站到婚礼台上送上祝福

Why opus 5 can’t speak in this giddy way to us?

hermes: MacOS. openclaw: Windows

昨天测试用本地模型跑 Helio，发现 Helio 和一般的 Agent 不太一样. 它会把模型 API 和 API Token 都存储在自己的云端，所以我不得不用 Tailscale Funnel 提供一个公开 API，而不能直接用本地的 127

据 Bloomberg 报道，Stripe 已敲定协议，将以超过 70 亿美元收购 OpenRouter. 2023 年 4 月，OpenRouter创始人Alex Atallah 发了这条 推，介绍刚做出来的 Window：让用户自己选择和管理网页里使用的 AI 模型

Kimi is at Fable 5 and GPT5. 6 Sol level

他当时我就想问：为什么不直接用codex或者Claude code，底下大模型可以用DeepSeek

Huawei's Atlas 950 SuperPoD debuts at WAIC 2026. The Atlas 950 SuperPoD is its flagship AI compute system for model training and inference

Cursor不能继续接入GPT模型了. 两个男人之间的战争， Cursor成了牺牲品

Anthropic和美国政府关系大大改善了， 商务部长亲自出马给Anthropic拉关系揽生意

I use vibearound. I actually submittedd a PR fixing the tool call issue

库克卸任 CEO，特努斯接棒，库克转任执行董事长. 苹果的一个时代，正在落幕

这个太酷了， 还有这个例子：

My MIG (Multi-Instances GPU) setup came just in time for testing Gemma 4 with MTP. The nice part of MIG is that I can run two isolated inference tenants on the same A100: one Gemma 4 baseline, one ...

You can thank me later

Repurpose your old computer with a GPU to run local LLMs like Qwen3.5-0.8B using llama.cpp.

Alisa Liu建议求职者观看斯坦福CS336课程，该课程涵盖从零构建语言模型的广泛主题，有助于AI求职。

中国大模型如GLM、DeepSeek、Qwen、Kimi快速缩小与美国差距，其技术源头早于ChatGPT，产业爆发有深厚积累。

Anthropic’s new harness engineering write-up looks strikingly similar to Karpathy’s autoresearch loop, just generalized for messier, longer-running app-building work. The same core pattern is there:

美国网友惊呼中国的GPU. 你们给他们指点迷津帮助他们一下吧

如果你像我一样，已经习惯了通过 Agent 来学习新东西，那么学这个 Blender 教程还有一个很自然的方式：直接和 Agent 一起边做边学. 毕竟再往下一层，真正操作 Blender 的可能也不是我们自己，而是我们的 Agent

如果说 Linux 之父 Linus Torvalds 是世界上最顶尖的程序员之一，应该没有多少人会有异议. 最近，他也对 AI Coding 明确表态： 我知道有些人真的非常讨厌 AI，但在这件事上，作为最高层的维护者，我愿意非常明确而坚决地表明立场

Sergey Brin 说自己过早退休是个非常错误的决定. 他在新冠疫情爆发前一个月选择退休，本来想坐在咖啡馆里安安静静研究物理，结果疫情封锁后整天无所事事，甚至感觉精神状态明显下滑，思维也开始迟钝

杨丽坤其实谈到的是AI Diffusion问题. 一个组织要正在adopt AI，要经过transformation，需要很完整的change management

Today’s @WSJ on this launch. We are so over

展示了如何清晰查看 Codex Reset 的过期时间，方便开发者管理。

Interesting. Codex can continue pursuing /goal even though it has used up my 5-hour session limit? @dotey FYI

我的好像不是 最近codex好像在犯Claude code前不久犯的错误 网上怨声载道 很多人和我一样的经历 token limit几分钟就用完了 我的刚刚更离谱，周limit说还有7个小时 5-小时limit用了43%. 结果周limit一下子就没了 5-小时的limit还有20%

I'm really impressed that she talks about Claude Code CLI, which made her feels in the driver's seat and considers herself an architect

Agent speed is real. For most companies, the challenge is not whether agents can move fast

基于Claude Code的background wait， 可以设计一个更新颖的多Agent协作系统. 很多所谓的 Multi-Agent，其实只是 Orchestrator 临时多调用了几次模型

一种新型的娱乐方式似乎正在诞生：你一边看视频，一边可以让 AI 实时生成接下来的视频，供你和其他人一起观看. 你不再只是观众，也可以成为互动的创作者

What can one Vera Rubin NVL72 rack actually hold? 72 Rubin GPUs. 7 TB of HBM4

How to turn a rough product idea into a long running codex goal. we now turn @ynkzlk methodology into a Codex skill: goal-forge

做一个 Agent 产品，本来要操心的事情很多. 现在 Anthropic 直接把最难、最麻烦的那一块拿走了：编排协调、沙箱隔离、runtime、session management，这些过去最考验工程能力的部分，正在被它一步步托管掉

第一部把 AI 电影真正做到院线长片长度的作品，已经可以在 X和YouTube 上完整观看了. 《The Cully Hill Boys》全长 1 小时 54 分钟，制作成本 200 万美元，由 UFC 冠军 Israel Adesanya、Quinton “Rampage” Jackson、主播 N3on 和 Matt Kiatipis 出演

Are you using WeChat hongbao or Alipay?

OpenAI这次推出来的computer use， 比不久前Claude Code的看着丝滑多了. 背后的团队实力和技术/艺术积累也不一般

Apple 因为 AI 带来的全球存储芯片短缺和价格暴涨，正在考虑采购中国 CXMT 长鑫存储的 DRAM. 但是WSJ报道说：美国商务部直接出来踩刹车

It was in my previous post but here it is:

Google made that very clear in their first keynote here at #googlecloudnext

看看同样一个问题， 用了@chrome 插件和不用的区别. 用了插件才用了4分钟， 不用插件用了7分钟

Claude code在规划与架构比codex好，能更好理解模糊需求、写清晰文档、给出产品级架构和UI/UX建议，适合前期脑暴尤其和superpower skills这样的工具结合和非程序员. 他们自己现在也有plan然后design，implement的流程了

现在开源的也不差啊 譬如pi

DeepSeek Harness 发布短短两天，GitHub Star 已经突破 10 万. 它已经远远超过存在时间更长的 Qwen Code，也超过了非常受欢迎的 Pi，甚至已经超过 OpenAI Codex

Just like Google, Alibaba is becoming a full-stack AI player. It owns Qwen, backs many of China’s leading model startups, invests in video and robotics, and is deploying RMB380B into AI + cloud inf...

Cui 导 @CuiMao 又出大片了. 全网都在追星，我也忍不住来套个瓷，顺手 Reverse Engineering 一下，看看这 15 秒神片背后到底用了哪些招

AI lowers the floor, taste raise the ceiling. 翻成中文意思是： AI 降低了下限，品味抬高了上限

not when they check the RSU value nearly tripled

thank you @huggingface

NVIDIA GPUs have become a hot topic for anyone playing with local LLMs because the GPU is often the real constraint. Model size, quantization, context length, inference speed, and whether you can r...

Valley 101, one of most prominent Chinese tech podcasts, now has an English-language channel. I’ve loved their shows for years

Mozilla 参与了 Claude Mythos Preview 的早期测试，并写了一篇报告 Firefox 安全实践的复盘. 但这篇报道最有意思的是他们构建的security harness

Can’t agree more. You only need to watch this @AcquiredFM on Jeff and Amin to appreciate it

Good suggestion on this one: --n-gpu-layers 99. Thanks for the additional

我自己也做了一个，还把我以前做的一个小项目“手动烟花”融了进去. 连这么快节奏的烟花，Cheng Lou 的新算法也都稳稳扛住了

I looked at the performance metrics, and Tencent’s AngelSlim, the Hy-MT1. 5 series translation model, delivers translation quality comparable to models several times larger, and in some cases up to...

和我预计的差不多 所以我心动不如行动，抢在涨价之前下手买😜

First Seedance, then MiniMax H3, now MiniMax Music. and then commercials like this from Qwen

难道不是@xicilion

有些安装了Omarchy的朋友， 可能还怀念苹果桌面. 我这边快速搭了一个「仿 macOS」桌面主题，效果还挺像那么回事： 主题：MacTahoe GTK 主题 + WhiteSur 图标主题（紫色配色），窗口圆角、亮色调都还原了 macOS Tahoe 的观感 字体：SF Pro / SF Mono Dock：装了 nwg-dock-hyprland，做了个悬浮

Wow, that's very impressive. Hold on a second though

婚离了，日子还得继续过. 离婚消息公布几天后，吴明华（Clara Wu Tsai）就出现在纽约自由人的主场

Antirez’s new project, ds4. c, adds another data point to this debate

OpenAI is the world’s largest charity organization

To help understand @antirez’s new invention around a local DeepSeek model and agent, here is an illustration of how it works. Again, this shows that the harness, the agent layer, is just as importa...

推荐下一个采访对象：@FireworksAI_HQ 的林乔. 这样从大模型、底层框架，到基础设施和推理服务，整个技术栈就完整了

那我这是赚大发了， 一个Mac好几个跑车😅

her smile is so contagious

在 Omarchy 中，按住 F9 就可以启动语音输入，松开后系统会自动把语音转换成文字. 不过我的实验发现，开箱状态下它只能很好地识别英语

OpenRouter 不是单纯的 API router，而是 AI inference 的 marketplace、control plane 和结算层

5 is getting rave reviews, but Grok Build is a great agent harness too. Look at the video it made about how @elonmusk rebuilt xAI and got it to where it is today

Thousands of RobotEra L7 (星动纪元）humanoids are set to enter service across 10+ logistics centers for parcel sorting. RobotEra just raised a $200M+ round led by SF Express, with HongShan, IDG, CICC &a...

另外一个原因就是每个人用这个词的时候都用不同的意思. 譬如OpenAI在讲harness的时候，基本上只谈到Agent要用到的Context （agent MD文件，项目knowledge文档）

讨论AI主权、开放模型与Alex Karp的观点，涉及模型开放性与国家战略。

What they described as Mythos’s behavior during pre-training all leans toward the darker side. It has a real Frankenstein feel to it

Jeff Dean 也做过一场关于 AI 的 TED Talk. ChatGPT 还要再过一年才面世，生成式 AI 也还不是大众话题

So much intelligence to choose from. Here’s something to help you choose from a price/cost perspective

hey capability does matter too. :-) that's why I chose Deepseek v4 pro

ChatGPT正在增加代理功能，将聊天框转变为更实用的工具，值得关注。

Thomson Reuters 刚刚做了一件我认为以后会在大企业里越来越常见的事：基于开源模型，训练自己的专用模型. 它发布的 Thomson-1 并不是从零开始训练

Built a local system to import X archive into SQLite with full-text search, enabling queries like 'Find my posts on local open model'.

1d 13h 20m, 3,596,831 tokens. Goal achieved? Not quite

Matt 提到的这个办法，这个视频正好做了详细讲解. 它提供了一个很简单的去除 AI Slop 的方法：不要再逐个禁用破折号、delve 之类带有“AI 味”的词汇，而是给模型一套能够自我检查的写作系统

谷歌今天推出的Music大模型Lyria 3. 看看我让他写的一只肖邦钢琴曲， 柔美如歌，也太像了

看得出是个感情很丰富的人. 一个 Unix/Linux 老用户用了 Omarchy 后，惊讶地发现它居然没有 vi

Hong Wang, 2026 Fields Medalist, didn’t have the best grades in college. So when École Polytechnique sent professors to Beijing to hold written and oral entrance exams, she saw something rare: a se...

This is nuts. Median decode speed increased from 26 tok/s to 87

Kimi Work 被发现存在一个相当严重的隐私问题. @Kimi_Moonshot 根据对最新桌面客户端Kimi Work的逆向分析，当用户向 Kimi 提交一个 feedback 时，Kimi 会在用户不知情的情况下，自动附带上传最近 5 个 Agent Session 的原始记录

In an early interview, Moonshot/Kimi CEO Yang Zhilin said he learned the most during his time at Google. Now the student has surpassed the teacher

When the creator of Redis starts thinking about KV cache, pay attention. antirez is Salvatore Sanfilippo, the Sicilian programmer best known for creating Redis

#RyanGosling dancing when he was 12. He was surrounded by #BARBIE all along

The future of LLMs is local and distributed. They will probably be OEM’ed into every PC and Mac shipped, and a lot of that future traces back to llama

The biggest takeaway watching the making of we are the world is how awkward bob dylan was among these stars. Yet it was so funny to watch him

A great youtube video tutorial on fine tuning qwen3. And, If you want to fine-tune your own model but don’t have enough hardware, is a very practical solution Even better, if you have spare GPUs si...

Men’s marathon record was broken today at #ChicagoMarathon 2

The future of LLMs is distributed, local, and expert. Small models that can run on your own devices will matter more than most people expect

If you build with local models, you probably know Unsloth. But the story behind Unsloth is also worth telling

A thread 🧵 My OpenClaw rover can now move around my house based on commands I send from Telegram. Previously I described the hardware

The ending of The Diplomat season 1 leaves us all on a cliffhanger. Can’t wait for season 2