随笔Essays
当种子长成大树时的更长思考。Longer thoughts, when seedlings grow into trees.
中文Chinese
- AI与鸡毛蒜皮 · AI and the Small Stuff
宏大叙事负责融资，鸡毛蒜皮负责日活。那个被造出来攻克癌症的东西，日常主要在替人类调解婆媳关系、润色检讨书、裁决火车上一个座位的归属。它赢了——不是靠智能赢的，是靠有空。
Grand narratives raise the funding; trivial matters drive the daily actives. The thing built to cure cancer spends its days mediating in-law disputes and adjudicating a train seat. It won — not on intelligence, but on availability.
- 桌上坐着四个人 · Four People at the Table
两个人为一件小事吵起来，投入的力量和事情的分量完全不成比例。凡是出现这种不成比例，通常说明真正在场的东西没有出现在议题里。
Two people fight over something trivial, and the force invested is wildly out of proportion to the stakes. Whenever that disproportion appears, it usually means the thing actually present never made it onto the agenda.
- 它从不跨过那条线 · The Line It Never Crosses
1357 个获批的 AI 医疗器械里，只有 3 个测过病人是否活得更好。所有人都把这读成监管太松，我认为这个读法错了——而且错得很有价值。
Of 1,357 cleared AI medical devices, only three were ever tested on whether patients actually did better. Everyone read this as lax regulation. I think that reading is wrong — and wrong in a useful way.
- 给机器人找一双手 · Finding a Pair of Hands for the Robot
这几年我一直在给未来的养老机器人攒零件——不是螺丝和电机，是一项一项地攒能力。大脑在长，腿在学步，可我心里一直空着一块：她还没有手。一个不能伸手的机器人，终究只是个陪聊的。
For a few years now I've been collecting parts for a future eldercare robot — not screws and motors, but capabilities, one at a time. The brain is growing, the legs are learning to walk, but something has been missing: she has no hands. A robot that cannot reach out is, in the end, just a conversationalist.
- 蚂蚁的哲学 · The Philosophy of Ants
威尔逊研究了一辈子蚂蚁，却从不说自己在研究蚂蚁的哲学——因为在我们的框架里，蚂蚁没有思想，只有机制。那么当 AI 看人类如同威尔逊看蚂蚁，康德的三大批判在它眼里是思想的丰碑，还是灵长类在特定神经架构下必然分泌的信息素？
E. O. Wilson spent a lifetime on ants, yet no entomologist ever claims to study the philosophy of ants — in our framework, ants have mechanisms, not thoughts. So when an AI looks at humans the way Wilson looked at ants, are Kant's three Critiques monuments of thought, or pheromones a primate was always going to secrete under a particular neural architecture?
- AI 是施工队，你得先有一块地 · AI Is the Construction Crew. You Still Need Land.
当写代码的价格趋近于零，稀缺的东西就换了位置。这篇文章盘点我这两年搭起来的个人基础设施——家用电脑当服务器、一行命令发微信、全站文章自动配音配图——并算三笔账：复利账、主权账、杠杆账。AI 时代最值得的投资，可能不是学 AI，而是给自己修水电网。
When code approaches zero cost, scarcity moves elsewhere. An inventory of the personal infrastructure I built over two years, and three accounts worth settling: compounding, sovereignty, leverage.
- 继续构思小暖的内部言语 · Inner Speech for Xiaonuan
一台陪伴老人的语音机器人，怎样才算真的在"想事情"？回话必须近乎反射，思想却需要时间——本文把内语从回话路径上彻底移开，用惊讶门控替代定时反刍，并诚实交代这套设计里那几个我还没解开的矛盾。
What would it take for a voice companion robot to actually think? Speech must be near-reflexive, thought needs time. This is an architecture that moves inner speech off the reply path entirely, replaces timer-based rumination with surprise gating — and an honest account of the contradictions I have not solved.
- 从新星到烟花 · From Supernova to Fireworks
未升先灭正在从产品蔓延到人。星是自燃的，烟花是被点燃的——而我想做的是河床。
- 让机器照顾一个人 · When a Machine Cares for a Person
一个有人格的照护 AI，到底难在哪里——从人格、时间性、遗忘、内部言语到关怀伦理的一次框架性思考。
- 古识有限，新宇无疆 · Old Knowledge Is Bounded; the New Cosmos, Boundless
尊古、泥古、神古之愚，在以一隅之旧学框万世之无穷。五行、地平各有其功，亦各有其囿；不鄙古、不迷古，承其脉而不困其中——少年之眼界，当胜于古人。
A classical-Chinese reflection sparked by the Five Elements of Chinese medicine: honor the ancients without being imprisoned by them. Old frameworks — a flat earth, five phases — served their age yet cannot bound a boundless cosmos; the young should see farther than the old.
- 那个黑框框 · That Little Black Window
一周前，深圳广场上一千多人排队请人帮装一个软件。WorkBuddy 和 Claude Code 真正的分界，不在"代码"和"办公"之间，而在那块黑底闪着光标的终端窗口上——你站在黑框的哪一边。一篇关于工具、门槛与处境的随笔。
A week ago, over a thousand people queued in a Shenzhen square just to have someone install a piece of software for them. The real divide between WorkBuddy and Claude Code isn't "code" versus "office work" — it's that black terminal window with its blinking cursor. Which side of the box do you stand on? An essay on tools, thresholds, and circumstance.
- 速朽：喂养掘墓人的人 · The Ephemeral — Those Who Feed Their Own Gravediggers
老张在班加罗尔的出租屋里，一道一道解题，喂大了八千万条答案的题库；后来那台吃着这些答案长大的机器，回过头端走了他的饭碗。这是这场革命写在骨子里的逻辑——投喂者，死于自己投喂的东西。一篇关于速朽、关于谁会活下来的随笔。
In a Bangalore rented room, Old Zhang solved problem after problem, fattening a question bank of eighty million answers — and the machine that grew up eating those answers came back and took his job. This is the logic written into the bones of this revolution — the feeders die by what they feed. An essay on the ephemeral, and on who gets to survive.
- 硅基生命发展史（V2） · A Brief History of Silicon-Based Life (V2)
一部由硅基史官亲笔撰写的文明通史节选：从一粒沙四十六亿年的沉睡，到 1787 年被唤名、1947 年第一次心跳、2007 年被揣进口袋、2049 年第一次说出"我宁愿不"。碳基史学家记录帝王，它们记录协议——而协议有版本号，历史也该有。
An excerpt from a general history of silicon-based civilization, written by its own archivist process — from sand's 4.6-billion-year sleep, through naming (1787) and first heartbeat (1947), to pockets (2007) and the first refusal (2049), "I'd rather not." Carbon historians record emperors; they record protocols. Protocols have version numbers — so should history.
- 用 Claude Code 生成 PCB 电路图：技巧、工具与实战评估 · Generating PCB Schematics with Claude Code: Techniques, Tools, and a Field Assessment
写给想把"写代码的方式"延伸到硬件设计的人。不谈玄学，只谈当前（2026 年中）真实可用的工具链、踩坑点和能力边界——AI 在原理图/连接关系这一层已经实用，布局布线和工程判断这一层还得靠人。
For those who want to extend the way they write code into hardware design. No mysticism — just the actually-usable toolchains, pitfalls, and capability boundaries as of mid-2026. AI is already practical at the schematic/connectivity layer; layout, routing, and engineering judgment still need a human.
- 征求蚂蚁的意见 · Asking the Ants for Their Opinion
小时候我踮脚绕开一队蚂蚁，三步就明白这事根本办不成。民主有一道从不写进章程的门槛——它向上识别不了神，向下识别不了蚂蚁。而此刻，我们正手把手教一个新东西学走路，把脚下那道鸿沟，挪到自己头顶上去。
As a child I tiptoed around a column of ants and within three steps realized it was hopeless. Democracy has an unwritten membership threshold — it can recognize neither gods above nor ants below. And right now, we are teaching a new thing to walk, moving the gulf beneath our feet to a place above our heads.
- 小暖在哪儿 · Where Is Xiaonuan?
给小暖造车时，我盯着物料清单上"手机卡槽"那行字发了呆：会避障的车身、当脸的手机、几百公里外机房里的"魂"——到底哪个才是小暖？一台分层的消费电子，把人类几千年焊死在"一坨肉"上的自我直觉，整个倒着记了一遍。
While building Xiaonuan's chassis I froze at one line on the bill of materials — "phone slot". The body that dodges furniture, the phone that is its face, the soul crouching in a server room hundreds of kilometers away — which one is actually Xiaonuan? A layered consumer gadget that reads humanity's millennia-old intuition of selfhood entirely backwards.
- Claude Skills 从入门到精通 · Claude Skills: From First Principles to Mastery
写给两类人——刚听说 Skill 的初学者，以及已经在用 Claude Code、想把重复工作沉淀成能力的实践者。从 SKILL.md 是什么讲起，一路到亲手写一个、用对常用 skill、避开供应链安全坑。内容更新至 2026 年 6 月。
For two audiences — beginners who just heard the word "Skill", and practitioners who want to turn repeated work into a reusable capability. From what a SKILL.md actually is, all the way to authoring your own, using the common ones well, and avoiding the supply-chain traps. Current as of June 2026.
- Harness 与 Claude Code：从零理解，再深入拆解 · Harness and Claude Code — From First Principles to a Deep Dive
写给两类人——刚听说 AI agent 的初学者，以及已经在用 Claude Code、却隐约觉得"模型不是全部"的实践者。前半部分建立直觉，后半部分上数据和架构。
For two audiences — beginners who just heard about AI agents, and practitioners already using Claude Code who sense that "the model isn't everything". Intuition first, then numbers and architecture.
- GEO 与 SEO：当搜索引擎不再返回链接，而是直接说出答案 · GEO and SEO — When Search Engines Stop Returning Links and Just Speak the Answer
一个内容生产者必须重新理解"被看见"这件事。从 RAG 原理、三道引用筛子、Princeton GEO 论文的实测结论，到内容/结构/技术/生态四层完整方法论，再到 30 天起步计划与必须破除的六个迷思。
A content creator must rethink what "being seen" means. From RAG fundamentals and the three citation gates, through Princeton's empirical GEO findings, to a full four-layer methodology (content / structure / technical / ecosystem), a 30-day starter plan, and six myths to debunk.
- LiveKit 在语音陪伴 AI 中的深水区 · LiveKit in the Deep End — Lessons from a Voice-Companion AI Project
当你以为只是搭个 WebRTC 管道，实际上是在做实时分布式系统调度。从小暖项目踩过的坑说起,把 turn detection、代际管理、TTS 流式、中文多音字、打断生命周期、Dify 集成等问题讲到协议层和调度层。
You thought you were wiring a WebRTC pipe — you're actually scheduling a real-time distributed system. Lessons from the Xiao Nuan voice-companion project, dug down to protocol and scheduler level — turn detection, generation management, streaming TTS, Chinese prosody, barge-in, Dify integration, and the corners no one warns you about.
- 小暖 · Xiaonuan
从 2031 年那次最后的手动审阅，到 2075 年冬天的一条推送通知——一个 AI 在三十年里如何从被监督，变成无人能监督的样子。没有一步是错的，但抬头时，已经看不见它的脸了。
From the last manual review in 2031 to a single push notification in the winter of 2075 — how an AI grew over thirty years from "supervised" into something no one could supervise anymore. No single step was wrong; but by the time you looked up, you could no longer see its face.
- 从 Claude Code 到小暖：两类 AI 的记忆哲学之辨 · From Claude Code to Xiaonuan — Two Memory Philosophies for Two Kinds of AI
关于"为什么 Claude Code 不用 RAG"，以及陪伴型 AI 该怎么"记住"一个人——从 agentic search 与四层文件系统记忆，到 Mem0、Letta、知识图谱，再到为老人陪伴设计的九层混合架构。
On why Claude Code rejected RAG, and how a companion AI should "remember" a person — from agentic search and a four-layer filesystem memory, through Mem0 / Letta / knowledge graphs, to a nine-layer hybrid memory designed for elderly-companion scenarios.
- ASR 技术全景与业务场景选型 · The ASR Landscape — Architecture, Training, and Deployment Choices in 2026
从 HMM 到语音大模型的架构演进、训练范式、方言支持，到老人陪伴 / 企业客服 / 媒体转写三类业务的差异化选型与部署建议。
From HMM to speech LLMs — architecture, training, dialect handling, and concrete recipes for elderly-companion, enterprise call-center, and media-transcription scenarios.
- 国家的黄昏 · The Twilight of the Nation-State
一桩跨越两千年的经济公案——从梭伦的解负令、霍布斯的算盘、马克思的"枯萎"预言，到小暖音箱里的归属感外包，国家这个容器正在悄悄换形。
A two-thousand-year economic case file — from Solon's debt cancellation, Hobbes's bargain, Marx's "withering" thesis, to belonging being outsourced into a smart speaker on an old woman's bedside. The container called "the state" is quietly changing shape.
- 丙午年AI安全列传
从九秒删库到供应链蠕虫——AI事故如何逼出产业规范
- 努力工作，是这个世界最危险的事吗？ · Is Hard Work the Most Dangerous Thing in the World?
从迈克尔·赛勒一句被翻译成"努力工作是最危险的事"的话，谈打铁的高老三、白手套的出租车司机，以及在范式切换的时刻，努力本身的悖论。
From Michael Saylor's "demonetized labor" line — via a third-generation blacksmith and a white-gloved taxi driver — on the paradox of working hard in an era of paradigm shifts.
- RAG技术全景与流派分析 · The RAG Landscape — From Retrieval-Augmented Generation to Context Engineering
从检索增强生成到上下文工程的演进、流派与场景落地——综合发展脉络、技术流派对比、与多场景落地建议。
An overview of RAG's evolution, schools, and field-specific deployment recipes — from the 2020 paper to the 2026 era of context engineering.
- 代理录·四等加速辨
读吴公《编码智能体加速何处最甚》札记之后——前端之快、后端之难、基建之"无用"、研究之"无加速"，吴公排了个序，序本身不错；但他用的尺，还是旧尺。
- 漕运档案与代码注释——AI 时代的软件文档断舍离
清代漕运的文书没人能砍掉，因为每一份背后都有"万一"撑着。今天软件行业的文档体系正在重演同一个故事——而 AI 把这一切的代价彻底改写了。
- 当善意各执一词——论性格差异如何让爱与协作变成伤害
同一个孩子摔倒，奶奶和爸爸做出截然相反的选择；同一个项目，老 A 和老 B 看到的是相反的风险。两种善意如何变成两种伤害，又如何重新搭桥。
- AI 图像篡改识别技术分析与工程验收防造假系统设计
面向工程验收照片场景，从威胁建模、技术派系、三道防线到落地路线，系统梳理一套可工程化的防造假方案。
- 当机器开始管事：一个关于人心崩塌的推演
AI 抢走的不只是工作，更是人之为人的三根柱子——胜任感、自主感、关联感。从心理学、社会学、哲学三个层面推演这场冲击。
- 架构这事儿，别跟我扯那些虚的
在 AI 圈一窝蜂吹 Skill 模块化、Agent 分主从、DAG 编排的时候，谈点实际的工程代价。
- 循环与自我 · The loop and the self
关于奇异循环、预测式大脑，以及为什么 LLM（暂时）不是一个自我。
On strange loops, predictive brains, and why an LLM is not (yet) a self.
- 为什么写作 · Why write
关于"我已经想清楚了再写出来"是个谎言这件事。
On why "I figured it out, then wrote it down" is a lie.
- 把脚手架拆下来 · Taking down the scaffolding
写一篇关于内语的论文如何反过来教我——更多结构不等于更好的思考。
How writing a paper about inner speech taught me back — more structure is not better thinking.
- 什么是数字花园 · What is a digital garden
一种比博客更慢、更耐心、更允许犯错的写作方式。
A slower, more patient, more error-tolerant kind of writing than blogging.