Where memory flows foggy forest dark...
Beijing's order this week to unwind Meta's $2bn purchase of Manus has produced predictable gloom about Chinese AI exits. The wrong story. The trade is not who buys Chinese software. It is who builds the boxes that all software runs on. Forecasts of US data centre capex run to $650bn for 2026 alone. About half of that nominal pipeline will not be built on time. The binding constraint is not capital…
Why prefab grey-space is the most underrated industrial story of the AI build-out The dominant narrative about AI infrastructure in 2026 goes something like this: a handful of American hyperscalers — Microsoft, Google, Amazon, Meta — are pouring six hundred and fifty billion dollars into gigawatt-scale campuses in Texas, Virginia, Arizona. NVIDIA sells them GPUs. TSMC fabs the silicon. The story…
For a long time, science followed a familiar script. A human scientist noticed something strange, formed a hypothesis, designed an experiment, gathered evidence, and slowly refined a theory. Computers helped with calculation, but the core act of discovery still belonged to people: intuition first, verification later. That is no longer quite true. AI is not just making science faster. It is…
"Vibe coding" is a new term that emerged with the popularization of large language models (and tools like Cursor, GitHub Copilot, etc.). It refers to a paradigm where programmers no longer meticulously grind through syntax line by line. Instead, they communicate high-level "intentions," "feelings," or "logical frameworks" (i.e., the vibe ) to the AI using natural language, leaving the AI to handle…
If you've spent any time building tools for your own workflow, you've probably discovered it: there's something deeply satisfying about crafting a good command-line interface tool that fits exactly into how you work. Looking at three recent CLI projects we've built — Cubox CLI, Flomo CLI, and Get笔记 CLI — the pattern is clear: making CLIs is vibe coding at its finest. What Is Vibe Coding Anyway?…
Most people assume yes. The reality is stranger and more interesting. There's a question I keep coming back to when thinking about large language models: do they actually understand time, or are they just very good at pretending to? The answer matters more than it seems. If an LLM genuinely reasons about time, it can track causality, handle evolving information, and give you answers grounded in…
Control Claude Code on your Mac from your iPhone anywhere using Termius, Tailscale, tmux, and SSH—no VPS needed. Includes whoami tip.
How I Use Obsidian, Git, and Claude Code Web to Manage My Notes TL;DR: Three tools, one coherent system — Obsidian as the local knowledge substrate, Git as the version-controlled backbone, and Claude Code Web as the AI orchestrator that ties them together. This is a personal sovereignty setup that compounds over time. The Problem With Most Note-Taking Systems Most people's note-taking setup is one…
三个月前，我在想：既然一个AI就能写东西，那如果让10个AI各管一摊，会不会写出更好的东西？ 于是我做了这个实验——一个多代理小说写作系统。现在系统写完了第一篇小说，我想分享一下这段经历。 缘起：从两篇文章到一个系统 这个系统其实是两篇文章的结合。第一篇讲的是5个代理写小说，第二篇讲的是7个代理写博客。我把它们揉在一起，变成了现在的10代理系统。 这10个代理是： 协调者 ：项目经理，先搞清楚我们要写什么 研究者 ：查资料的，一切细节要有来源 世界构建者 ：设定故事发生的地方 角色设计者 ：创造人物，连他们最怕什么都要想清楚 情节设计者 ：搭结构、写场景、埋线索 作家 ：真正写字的人 连续性检查者 ：抓bug的——这个人第一章叫张三第三章怎么叫李四了？ 敏感性读者 ：检查有没有冒犯性的刻板印象 编辑 ：质量守门员，12项指标一项一项过 出版者 ：最后打包交付 最有意思的三个设计…
There's a famous experiment that keeps haunting me. In a California grocery store, researchers set up two tasting booths on alternating days. One day, they offered 24 varieties of jam. The next, just 6. The 24-jam booth drew bigger crowds — more people stopped, sampled, lingered. But when it came time to actually buy? The 6-jam booth won by a landslide. Nearly 24% of people who stopped made a…