模型越强，你说不清楚想要什么的代价越大。从地图与疆域讲到四种空白，拆一套我自己验证过的和模型打交道的方法，最后讲怎么让每个坑只交一次学费。

AI 把讲授的边际成本打到零，教育史上第一次，瓶颈从'谁来教'搬到了'人为什么想学'。顺着一个 AI 社群里关于'孩子为什么不肯学'的讨论，我把动机科学、学习科学和 alignment 的证据串了一遍：需求不是找出来的，是点着的；'以后用得上'输给的是一条折现曲线；动机满格的学生加一个直接给答案的 AI，考试成绩反而比不用的还低 17%。教育要解的是双重对齐——先让人想学，再保住那份不可外包的认知劳动。

AI has driven the marginal cost of instruction to zero, and for the first time in the history of education the bottleneck has moved from 'who will teach' to 'why would anyone want to learn.' Starting from a community thread about why kids refuse to study, I traced the evidence across motivation science, learning science, and AI alignment: demand isn't found, it's ignited; 'you'll need it someday' loses to a discount curve; and a fully motivated student with an answer-giving AI scores 17% worse on exams than one with no AI at all. Education in the AI era is a double alignment problem — first get people to want to learn, then protect the cognitive labor that can't be outsourced.

这是《前三次技术革命，红利最后都落到普通人头上》背后的完整框架。正文是给人读的一篇文章，把判断讲清楚就够了；这篇是给框架本身的——把'技术怎么洗牌财富'拆成第一性原理、一个财富捕获公式、六个传导机制、AI 的四层结构、一套九问研究法和几条可验证假说，方便你拿去套到下一个技术上，自己跑一遍。

The full framework behind Railroads, Electricity, the Internet Each Built a Middle Class. AI Might Not. The essay is the readable version; this is the machine — how technology reshuffles wealth, broken into a first principle, a wealth-capture formula, six transmission mechanisms, AI's four layers, a nine-question research method, and a few falsifiable hypotheses — so you can run it on the next technology yourself.

听《十分吸引》串台《听懂涨声》聊 AI 财富再分配，最有意思的是他们对历史的推演——铁路、电力、互联网三轮洗牌的机制。我顺着挖了一层，也想得更硬了一层：洗牌从来不是随机的，它一直按同一套逻辑走——技术重写经济网络，旧瓶颈失效，新瓶颈定价，财富跟着议价权走。铁路重写物流网，电力重写能源网，互联网重写信息网，AI 正在重写任务网。前三轮先极度集中，最后又长出新的中产，因为那些新生产方式胃口很大、需要海量普通人。AI 是重资产、离散型的，可能只走完集中那一半，跳过做大中产那一半。

I listened to an episode on AI and wealth redistribution — the finance podcast 十分吸引 crossed with 听懂涨声 — and the best part was the history: how railroads, electricity, and the internet each reshuffled wealth. I dug a layer under their account and made it harder: the reshuffle runs one logic every time — technology rewires the economic network, the old bottleneck fails, a new one gets priced, and wealth follows the bargaining power. Railroads rewired the goods network, electricity the energy network, the internet the information network, AI the task network. The first three eventually grew a new middle class, because those production modes needed armies of ordinary people. AI is heavy-asset and dispersive, so it may run only the concentration half and skip broaden-the-middle.

Agent 干错事，大多数人的反射是往 CLAUDE.md 加一句话。它错在两个方向：指令文件是上下文不是强制，规则在压力下根本不绑得住；而文件越长，越污染那个让 agent 可靠的上下文本身。承重的规则该进 hook、lint、CI，散文应该越写越短。

The reflex when an agent misbehaves is to add a sentence to CLAUDE.md. It's wrong in two directions: instruction files are context, not enforcement, so the rule won't bind under pressure — and the longer the file gets, the more it pollutes the context that makes the agent reliable. Load-bearing rules belong in hooks, lint, and CI. The prose should get shorter.

Agent 能完美遵守一条它看得见的契约，也能毫无知觉地破坏一条它看不见的。能熬过 agent 时代的边界，是那些写进仓库、有类型、有测试、能被 import 的；会被弄坏的，是那些只活在团队记忆里的裸字符串约定。把每条边界都变成契约，是手里杠杆最高的一招。

An agent is excellent at honoring a contract it can see and terrible at preserving one it can't. The boundaries that survive the agentic era are the ones written into the repo as typed, tested, importable things — and the ones that break are the stringly-typed promises that lived only in a team's shared memory.

Agent 最危险的产出不是烂代码，是"我搞定了"这几个字。纯核心、一条命令的验证、一个独立的判官、一份证据账本，全部加起来就为了一件事：把"做完"的裁决权从作者手里拿走，让它变成被证明出来的，而不是被感觉出来的。这是前三篇一直在铺垫的底座。

An agent's most dangerous output isn't bad code — it's the word "done." The whole point of a pure core, a one-command verify, a separate verifier, and an evidence ledger is to take the verdict out of the author's hands and make "done" a thing that's proven rather than felt. This is the substrate the rest of the series was building toward.

落进我仓库的改动，现在大部分是 agent 敲的。真正变了的不是效率，是代码的读者换了人——一个每次会话都失忆、只能从文件里重新拼出理解的东西。一旦接受这点，仓库就不再是文档，它变成了 agent 编程时要对着写的接口：目录是导航 API，契约是约束 API，验证是校验 API。

The maintainer of your repository quietly changed from a human who remembers to an agent that re-derives everything from the files each session. Once you accept that, the repository stops being documentation and becomes an interface. Directory layout is the navigation API, contracts are the constraint API, and verification is the validation API.

朋友给我一份用拉丁学名的 omakase 认鱼课，可板前点菜时，菜单写的是 Hamachi、是 Sea Bream。我想要一个标美式菜单常用名的版本，于是指挥一队 agent 搭了出来：九个并行调研 agent 上网核实菜单名、gpt-image-2 出约 150 张写实图、一段确定性代码拼成 PDF。73 种鱼、92 页，外带两个踩坑故事和一份免费下载。

I wanted a sushi-counter field guide that used the names you actually order by in a US restaurant — not Latin binomials. So I pointed a team of agents at it: parallel research workflows verifying English menu names with web search, gpt-image-2 for ~150 photoreal neta shots, and a deterministic HTML→PDF assembler. 73 fish, 92 pages, two debugging war stories, and a free download.

一个朋友半夜问我，要不要也做个工具去跟 X 上很火的 Serenity 买进卖出。我的第一反应是没意义：一个策略真能稳定赚钱，最理性的做法是闷声加仓，不会拿出来开直播。但聊到后面我改了主意。值得蒸馏的东西确实有，藏在他提问的顺序里，跟他的持仓没关系。我们把这套研究流程做成了一个开源工具。

A friend pinged me at midnight asking whether we should build a bot to follow Serenity, the supply-chain researcher who blew up on X, in and out of his trades. My first reaction was that it was pointless: a strategy that actually compounds gets quietly levered up, not broadcast. But by the end of the night I'd changed my mind. There is something worth distilling here — it just isn't in his holdings, it's in the order he asks his questions. So we turned that process into an open-source skill.

群里聊升职，起点是鸭哥那篇文章：用 AI 把效率拉爆、成了模范老黄牛，就能加薪升职吗？现实恰恰相反——你越快、越好用，越容易被组织重新定位成一个超高吞吐的执行节点。我把升职拆成一个乘法模型：能力 × 叙事 × sponsor 动机 × 组织坑位 × 政治/时机，任何一项接近 0，整体都垮。AI 主要增强的是能力和速度，却可能同时伤害你的叙事和定位。最怕的不是你不够快，而是你太快了，别人懒得理解你的脑子。