Britannica drew a map of all human knowledge in 1974. An LLM is what it was trying to be. 1974 年，大英百科画了一张人类知识的全图。它想成为的东西，今天叫大语言模型。

In 1974, Encyclopædia Britannica published its most radical edition. The 15th was not a revision but a restructuring, and its architect was not an editor but a philosopher: Mortimer J. Adler, the man behind How to Read a Book and the Syntopicon — the two-volume index of “Great Ideas” he built for Great Books of the Western World.

Adler’s design split the encyclopedia into three tiers:

• The Micropædia: 12 volumes, roughly 65,000 short entries under 750 words each. Quick facts.
• The Macropædia: 17 volumes, 699 long essays by named experts, the longest running 310 pages. Deep understanding.
• The Propædia: a single volume called An Outline of Knowledge, dividing everything humans know into ten great divisions, subdivided into a hierarchy of branches, each section annotated with pointers into the other two parts.

The problem Adler was attacking was not content. It was navigation: the reader who doesn’t know where to look. The Propædia was his answer — a curriculum in book form. Follow the outline downward and you get a designed learning path through any field. It was a knowledge graph in 1974, and it sat on the shelf for four decades nearly unread.

1974 年，《大英百科全书》出了它历史上最激进的一版。第 15 版不是修订，是重建；主持这件事的不是编辑，是一位哲学家——莫蒂默·阿德勒，《如何阅读一本书》的作者，也是《西方世界的伟大著作》那套书里”伟大观念”两卷本索引的搭建者。

阿德勒把百科拆成三层：

• 微编（Micropædia）：12 卷，约 65,000 个短条目，每条不超过 750 词。用来查事实。
• 大编（Macropædia）：17 卷，699 篇署名专家写的长文，最长的一篇 310 页。用来求理解。
• 导编（Propædia）：只有一卷，叫《知识提纲》，把人类已知的一切分成十大部类，往下再层层分枝，每一节都标注了通向另外两编的指针。

阿德勒要解决的不是内容问题，是导航问题：读者不知道该去哪儿找。导编就是他的答案——一本书形态的课程表。顺着提纲往下走，任何领域都有一条设计好的学习路径。这是 1974 年的知识图谱，然后它在书架上躺了四十年，几乎没人翻。

The Propædia did not fail for lack of ambition. It failed because its architecture assumes exactly the knowledge its reader lacks.

You had to classify before you could ask. To find superconductivity you first had to know it lived under Matter and Energy, then physics, then the right branch of condensed matter. A taxonomy is built for the person who already understands the domain; the person with the question is, by definition, the one who doesn’t. The two-volume alphabetical index shipped at the end of the set required no prior knowledge at all, and it’s what everyone actually used.

One tree can’t hold knowledge. Every real subject belongs to several branches — game theory is mathematics and economics and computer science at once. The outline had to pick one home for everything, and the ten divisions carried their era’s biases with them. Worse, a printed tree freezes at press time while the territory it maps keeps moving.

Navigation costs more than search. Cross-references in the Micro- and Macropædia averaged one per page. A reader following pointers across volumes lost the thread within a few jumps. People don’t want a location. They want an answer.

Libraries reported the Propædia was scarcely used; reviewers recommended dropping it from the set entirely. Britannica kept printing it until the final edition — an idea sustained by conviction, not demand.

导编的失败不是因为不够有野心。它失败，是因为它的结构恰好预设了读者手里没有的那样东西。

你得先会分类，才能提问。 想查超导，你得先知道它归在”物质与能量”底下，然后是物理学，然后是凝聚态的哪一枝。分类体系是为已经懂这个领域的人建的；而带着问题来的那个人，按定义就是不懂的那个。整套书末尾那两卷按字母排的索引，不要求任何前置知识——那才是所有人真正在用的东西。

一棵树装不下知识。 任何真实的主题都同时属于好几枝——博弈论既是数学，又是经济学，又是计算机科学。提纲却必须给每样东西选定一个家，而那十大部类还捎带着它那个年代的偏见。更糟的是，印刷出来的树在付印那天就冻住了，而它要描摹的地形还在动。

导航比搜索更贵。 微编和大编里的交叉引用平均每页一条。读者顺着指针在卷册之间跳，跳不了几次线索就断了。人们要的不是一个位置，是一个答案。

图书馆的反馈是导编几乎无人使用；书评人建议干脆从整套书里拿掉。大英百科一直印到最后一版——一个靠信念、而不是靠需求维持下来的想法。

While Adler drew the map from the top down, the internet built the same structure from below.

Google sorted the web with links instead of categories. Wikipedia replaced the tree with a graph: articles that cite and connect to one another, organized by thousands of editors making local decisions over decades, navigable from any entry point. Nobody ever decided which branch of knowledge an article belongs to — the structure is emergent, continuously updated, and free.

The hierarchy inverted. The alphabetical index became the search box. The designed learning path became “you might also like.” And the reader no longer needed to know where anything lived before asking about it.

阿德勒自上而下地画那张图的时候，互联网正在自下而上地把同一个结构搭起来。

谷歌用链接而不是类目给网页排序。维基百科用图取代了树：文章之间互相引用、互相连接，由几千名编辑在几十年里各自做局部决定组织起来，从任何一个入口都能进去。从来没有人决定过某篇文章属于知识的哪一枝——结构是长出来的，持续更新，而且免费。

层级被翻转了。字母索引变成了搜索框。设计好的学习路径变成了”你可能还喜欢”。而读者，在提问之前不再需要知道答案住在哪里。

A large language model is the latest form of the same idea: a map of knowledge, except the map was never drawn.

The model’s embedding space is the Propædia’s outline, learned statistically from hundreds of billions of tokens instead of imposed by design. Every direction in it is continuous and traversable; there are no dead branches and no single home for game theory, which sits near mathematics and economics and computer science at once. You do not classify your question into a branch. You ask it in natural language and get an answer — the interface Adler’s readers never had.

Adler wanted the reader to arrive at knowledge without first knowing where it lived. Fifty years later, a machine does exactly that. But the LLM version deleted two things the paper volumes guaranteed. Every Macropædia essay had a named author; every claim led back to sources you could check. An LLM is the sum of what was written, with the receipts dissolved — it cannot tell you which encyclopedia it is quoting, or whether that source exists at all. The structure survived. The provenance didn’t.

大语言模型是同一个想法的最新形态：一张知识的地图，只不过这张图从来没有人去画。

模型的嵌入空间就是导编的那张提纲，区别在于它是从几千亿 token 里统计地学出来的，而不是设计着强加上去的。里面每一个方向都是连续可走的；没有死枝，博弈论也不必只有一个家——它同时挨着数学、经济学和计算机科学。你不需要先把问题归到某一枝，你用日常语言问，然后得到答案。这正是阿德勒的读者从来没拿到过的那个界面。

阿德勒想让读者不必先知道知识住在哪儿就能抵达它。五十年后，一台机器做到了。但大模型这一版删掉了纸书保证过的两样东西：大编的每一篇长文都有署名作者，每一个论断都能顺着线索回到可以核查的来源。而大模型是所有被写下来的东西之和，凭据被溶解掉了——它没法告诉你它在引用哪一本百科，甚至没法告诉你那个来源是否存在。结构活下来了，出处没有。

The company still exists — and not quietly. After Wikipedia killed the print edition in 2012, Britannica rebuilt itself around licensing its archive to AI systems, then turned around and sued Perplexity in 2025 and OpenAI in March 2026 for scraping content without paying for it. There is a clean irony in the geometry: the machine intelligence that finally delivers what the Propædia outlined is built on corpora that include Britannica’s, and Britannica’s only remaining weapon is copyright.

Its new business in the AI era is selling exactly what the LLM erased — verified, attributed, expert-written answers. The legend of the map, sold to everyone who uses the map.

这家公司还在，而且动静不小。2012 年维基百科杀死了纸质版之后，大英百科把自己重建成了一门生意：把档案授权给 AI 系统。然后它回过身来，2025 年告了 Perplexity，2026 年 3 月告了 OpenAI，理由是未付费抓取内容。这里有一个几何意义上很干净的讽刺：那个终于兑现了导编设想的机器智能，是建立在包含大英百科在内的语料上的，而大英百科手里剩下的唯一武器是版权。

它在 AI 时代的新生意，卖的恰恰是大模型抹掉的那样东西——经过核实、有署名、由专家撰写的答案。地图的图例，卖给每一个正在用这张地图的人。

Adler spent decades trying to organize all knowledge — first the Great Ideas of the Western world into two volumes of topics, then the Propædia into one. Both failed at the interface, not the ambition. The lesson is not that knowledge can’t be organized. It’s that you cannot hand readers a map and expect them to navigate; they will demand to be teleported to the answer. A machine can now do the teleporting. The map turns out never to have been the hard part. The citations are.

阿德勒花了几十年试图把全部知识组织起来——先是把西方世界的伟大观念压成两卷主题索引，然后是把一切压成一卷导编。两次失败都败在界面上，不是败在野心上。这里的教训不是”知识没法被组织”，而是：你不能把一张地图交给读者，然后指望他们自己走；他们要的是被直接传送到答案跟前。 现在机器可以负责传送了。事后看来，难的从来不是那张地图，是那些出处。