当编程接近免费,模板成为化石;
形式不再是容器——形式即内容。
你正在读的这个页面,没有模板、没有框架、没有构建工具。它是一个论点的运行态。往下,每一个能拖动的东西,都是论证的一部分。
WordPress 一度承载近半个 web。这不是因为「固定程序 + 可变文字」是内容的自然形态,而是因为工程师时间曾是最贵的生产要素。模板的本质是固定资本摊销:程序写一次,摊到一百万篇文章上。CMS、主题市场、「设计阶段」与「内容阶段」的分离——整套分工,都是为了绕开一种昂贵。
要素价格归零,为节约该要素而建的结构就会解体。拖一下,看这条价格曲线塌得有多快:
2021 年 Distill.pub 停刊,公告承认:交互式论文是手工业奢侈品,一篇需要作者、工程师、编辑协作数月。
需求从来都在——Bret Victor、Nicky Case、The Pudding 早就证明了读者想要什么。是供给被价格卡死的。价格塌了,形态就该重新长。
化石与未来的区别,不用讲,编译给你看。下面是同一个论断——「prompt caching 能省掉你大部分推理成本」——在三个时代分别长成的样子:
如果你的应用存在大量重复上下文,缓存命中的部分按一折计价。对客服、对话类产品来说,这意味着账单可以砍掉一大截。
本文整理了配置步骤和注意事项,建议收藏……
prompt caching 能省掉你大部分推理成本。
别信这句话。拖动参数,自己验证:
定价取 $3/M(原价)与 $0.3/M(缓存命中,0.1×)。示意值——查看源码,三秒换成你自己的价目表。
{
"claim": "prompt caching 可省 60–85% 推理成本",
"mechanism": "缓存命中 token 按 0.1× 计价",
"assumes": ["上下文高度重复", "会话粘性路由"],
"verify": { "type": "calculator", "params": ["tokens", "hit_rate"] },
"provenance": "freem.ai · 2026-07",
"confidence": 0.8
}
给机器的影子层。人看不见它,但 agent 只看它——这一层决定你被谁引用。
给通勤者的渲染。同一个不变量的另一个投影——形式在读取那一刻才被决定。
看清楚了吗:前两个时代,作者在排布句子;第三个时代,作者只发布不变量——论点、数据、模型、置信度——形式按读者的上下文即时编译。一篇文章,变成了这些渲染的等价类。论证从断言变成了可运行的模型:「别信我,来改我的参数」。
「形式即内容」要过三道筛子,过不去的部分,文本仍然为王:
作者端的成本革命,不改变消费端的硬约束。交互要求读者付出操作与注意力;文本可略读、可检索、可引用,消费摩擦最低。交互只在「操纵比阅读更快抵达理解」的地方赢:动态系统、参数敏感性、反直觉的量。给观点随笔强加交互,是负价值。
越来越多的「阅读」由 AI 完成,而 JS 重交互对爬虫是黑箱。内容会分成两层:语义层给机器,决定分发与被引用;表现层给人,决定体验与说服。只做后者的 blog,会漂亮地消失在所有 AI 的答案之外。
文本的耐久性以千年计,无人维护的软件以年计。出路是把 source of truth 上移到 spec——artifact 随时可以从源意图重新生成。单文件、零依赖的 HTML,是程序里最接近文本耐久性的形态。你现在读的,就是一份。
Web2 的一切静态构件都是缓存:对预期需求的预计算。缓存存在的理由只有一个——实时生成太贵。生成免费之后,缓存反转为「存 spec,读时编译」。判据可以机械化:看一个构件的形态在替谁省钱;那个要素归零,它就解体。
用这把尺子,量一遍 Web2 的地基。点开每张卡片,有一句深注:
任何全称判断都需要一个诚实层。三样东西不会塌:
这是整场迁移最大的结构性事实:瓶颈从生产移到验证。provenance、签名、审计、责任承担,是新的稀缺。合同、价格、记录性新闻仍需要 canonical 的不可变版本——个性化一切等于认识论碎片化,社会必须保留共享参照物。
人一天还是那些小时。能跑的错误模型比错误的文本更有说服力,「可运行」会被误当作「正确」——交互是下一代的 chartjunk,判断力是防线。
Web2 的免费内容靠眼球换广告;当读者是 agent,眼球消失。要么内容 API 化、按爬取付费,要么开放语料塌方。这一条决定演进的速度,不决定方向。
新的稀缺只剩三样:验证、判断力、不变量的质量。
别再租别人的形式。第一批读时编译的 blog,会从带着不变量的人手里长出来——观点、数据、判断是你的;开发、分发、维护,交给 AI。
在 rushos,你以「领主」身份发布不变量,剩下的交给机器与网络。
When programming approaches free, templates become fossils.
Form stops being a container. Form becomes content.
This page is not a template, a framework, or a build pipeline. It is an argument in runtime form. Every thing you can touch below is part of the proof.
WordPress once carried nearly half the web. That was not because “fixed program plus variable text” is the natural form of content. It was because engineering time used to be the most expensive production factor. A template is fixed capital amortized over millions of posts.
When the price of an input collapses, structures built to economize that input start to dissolve. Interactive essays used to require a rare mix of writer, designer, and engineer. AI turns many of those once-expensive operations into something close to an afternoon's work.
Take one claim: prompt caching can remove much of your inference cost. In 1998 it becomes a hand-written personal page. In 2010 it becomes a CMS article with share badges and related posts. In 2026 it becomes a small working model: a calculator for humans, a claims layer for agents, and a spoken script for listeners.
A page is a document. Links, guestbooks, counters, and email carry the conversation.
A post is text poured into a reusable mold. The template saves engineering time.
A post stores invariants: claim, data, model, confidence, provenance. The surface compiles for the reader.
The author no longer only arranges sentences. The author publishes a small reality that can be inspected, adjusted, and cited.
The Chinese version keeps the full interactive demo. This English companion states the same argument in a more stable reading form.
Interaction asks for attention and manipulation. Text still wins when skimming, quoting, and search are the main job.
Content now needs two layers: a semantic layer for agents and a presentation layer for humans.
The durable source of truth should move upward into a spec. Single-file, dependency-light HTML is one of the closest software forms to textual durability.
Many Web2 components are caches: precomputed answers to expected needs. Search result pages, landing pages, dashboards, forms, help centers, newsletters, feeds, forums, and product pages all exist in part because real-time generation and routing used to be expensive.
Search results, feeds, newsletters, and landing pages move from static routes to agent-shaped briefings and single-person persuasion.
Documentation, dashboards, and wikis split into semantic stores plus generated views.
Comments, forums, and personal homepages become conversation agents, reputation routes, and callable identities.
Forms and product pages retreat into schemas,报价 interfaces, and negotiation surfaces for buyer and seller agents.
The bottleneck moves from production to verification: provenance, signatures, audits, and accountability become newly scarce.
A runnable wrong model can be more persuasive than a wrong paragraph. Interaction is the next generation of chartjunk unless judgment stays awake.
When readers are agents, eyeballs disappear. Content must become API-like, licensed, or otherwise compensated by new rails.
The new scarcity is verification, judgment, and the quality of invariants.
The first read-time compiled blogs will grow from people who carry invariants: claims, data, taste, and judgment. AI handles the code, distribution, and maintenance.
扫码可以直接打开页面；下方文案可以复制后发到微信、Facebook、邮件或任何消息窗口。
Scan the QR code to open this page, or copy the note below into WeChat, Facebook, email, or any message thread.
留下联系方式，再写下你想编译成 blog / artifact 的构思、创意。我们会尽快联系你，并一并回答问题。
Leave a contact method and the concept or idea you want to compile into a blog or artifact. We will contact you soon and answer questions together.