一张五千多年前的泥板，上面压着几个符号：一种谷物、一个数量、一个人名。你能把它们一个个认出来，却没法把它读成一句话——它没有语序，谁对谁做了什么都没写。它是一张收据。
这很容易被当成「文字还很原始」。但反过来看才有意思：记账根本不需要句子。要让符号能写出任何一句话，得先发生一件看起来不相关的事——有人把一个画出来的东西，拿去记它的读音。
苏美尔人就这么干了。这一步之后，能写下来的不再限于画得出的东西：凡是能念出来的，都能写。
先看门槛这一边。公认最早的一批文字记录出自两河流域的乌鲁克，约公元前 3300 年，绝大多数是账目：某种货物、多少、给谁、合计。它们通常写不出完整的口语句子——语序、谁做了什么，纸面上没有位置。更早的前身甚至不是符号，是封在泥球里的小陶筹，一枚筹代表一只羊。
跨过门槛的那一下，是符号从「画的是什么」切到「念的是什么」。苏美尔语里「箭」念 ti，「命」也念得差不多，于是那个画着箭的符号被借去写「命」。这个把戏叫 rebus 原理（借音写字，汉语传统叫假借）。英语里玩的是同一个：画一只眼睛，表示 I 这个词。
汉字里这一步留下了特别干净的痕迹。「其」的甲骨文字形画的是一只簸箕——扬谷子用的工具，本义就是簸箕。但它被借去写一个读音相近的代词「其」，借得太顺手，后来只好另造一个「箕」来写簸箕：原主人被赶出了自己的房子。
从零开始做到这一步的，目前只认三到四次：两河约公元前 3300 年、中国商代甲骨文约公元前 1200 年、中美洲约公元前 500 年前后（现存最早的是萨波特克人的石刻）。埃及约公元前 3200 年算不算独立还有争论——离两河太近，传过去的可能是「符号能记话」这个主意本身。证据强度：年代和独立性都是考古推断，新出土的东西会挪动数字。
剩下所有文字都是接手来的，接手的往往不是字形而是「这事能做」。日语就很典型：先直接拿汉字当表音符号用（这批字叫万叶假名），再写草、简掉——「安」成了平假名 あ，「伊」掐掉一角成了片假名 イ。
EN The earliest writing we have — Uruk in Mesopotamia, around 3300 BCE — is almost all accounting: a commodity, a quantity, a recipient, a total. It usually cannot be read as a spoken sentence; word order has no place on the tablet. Its own predecessor was not even signs but small clay tokens sealed in a ball, one token per sheep. The threshold is crossed when a sign switches from what it depicts to what it sounds like. Sumerian for "arrow" was ti, and "life" sounded close enough, so the arrow sign was borrowed to write life. The trick is called the rebus principle; in English it is the eye drawn to mean the word I. Chinese kept an unusually clean trace of the same move: the oracle-bone form of 其 (qí) depicted a winnowing basket, which was its original meaning — then it was borrowed for a near-homophonous pronoun, so successfully that a new character 箕 (jī, "winnowing basket") had to be made for the basket. The original owner was evicted from its own house. Writing has been built from scratch only three or four times: Mesopotamia around 3300 BCE, Shang-dynasty oracle bones in China around 1200 BCE, Mesoamerica around 500 BCE (the oldest surviving body is Zapotec stone carving). Whether Egypt around 3200 BCE counts as independent is argued over — it is close enough in time that what travelled may have been the idea that signs can record speech. On evidence strength: both the dates and the independence claims are archaeological inference, and new finds move the numbers. Every other script inherited the idea, and what gets inherited is usually not the shapes but the knowledge that this can be done. Japanese is the clear case: Chinese characters were first used purely for their sound values, a set called man'yōgana, then cursived and clipped into syllabary signs — 安 (an) cursived into あ (a), 伊 (i) clipped into イ (i). One more pattern: once a script starts recording sound, it only goes further down that road. In Chinese the sound-bearing part is the phonetic component, and phono-semantic characters — half meaning, half sound — run above 20% of Shang oracle-bone characters, above 80% of the seal-script characters in the Han dictionary Shuowen Jiezi, and above 90% of modern characters.
文字之间最大的差别不在好不好看，在它的符号对着哪一层单位。只有三条路：对着一个音位（phoneme，能区别意义的最小声音单位），就是字母；对着一个音节，就是音节文字；对着一个语素（morpheme，最小的有意义单位），就是汉字这条路。
走得通哪条，很大程度上被这门语言的音节数量定下来。标准日语只有约 100 个不同音节，所以平假名、片假名各四十来个基本符号就够。切罗基语（Cherokee，北美原住民语言）的音节绝大多数是「辅音加元音」这种开音节，85 个符号也能罩住。
英语就走不通。它可能的音节有多少，不同算法差得很远，几千到三万的估计都有；无论取哪头，一个音节配一个符号都不可能——strengths 这种一个音节塞八个音的词满地都是。只能往下拆到音位那一层：二十来个字母包下整门语言的声音。普通话带声调的音节不到 1300 个，音节文字其实走得通；汉字选的是语素那条路，代价和红利在「汉语哪里特别」那期算过账。
但没有哪套文字是纯的。汉字九成以上是形声字，声旁就是明摆着的表音零件；只是读音几百年里一直在变而声旁没跟着变，现在能可靠提示读音的声旁只剩约四成。反过来，英文字母也在偷偷记语素：cats 和 dogs 末尾都写 s，可前一个嘶出来是清的、后一个嗡出来是浊的。写法记的是「复数」这个语素，不是那一刻的声音。
所以问一套文字「表音还是表意」，问不出什么东西。该问的是两句话：它的符号主要对着哪一层单位，以及这层单位在这门语言里一共有多少个。
EN The deepest difference between scripts is not how they look but which layer of unit their signs point at, and there are only three choices: a phoneme — the smallest sound unit that distinguishes meaning — which gives you an alphabet; a syllable, which gives you a syllabary; or a morpheme, the smallest meaningful unit, which is the road Chinese characters took. Which road works is largely fixed by how many syllables the language has. Standard Japanese has only about 100 distinct syllables, so forty-odd basic signs in each of the two kana sets cover it, one sign per syllable. Most syllables in Cherokee, a North American Indigenous language, are open — a consonant plus a vowel — so 85 signs cover that too. English cannot go this way. Estimates of how many syllables English permits differ wildly by method, from a few thousand to around thirty thousand; on any of them, one sign per syllable is hopeless, and words like strengths pack eight sounds into a single syllable. So English keeps splitting downward to the phoneme layer, where roughly two dozen letters cover the whole sound system. Mandarin has fewer than 1,300 distinct syllables counting tone, so a syllabary is in fact workable, and the Zhuyin system is roughly that; characters took the morpheme road instead, whose costs and payoffs were counted up in the topic on what makes Chinese unusual. No script is pure, though. Over 90% of Chinese characters are phono-semantic, with the phonetic component sitting there in plain sight — but pronunciation kept shifting for centuries while the components did not, so only around 40% of modern phonetic components reliably cue the reading. Conversely, English letters quietly record morphemes: cats and dogs both end in s, yet one is hissed and the other buzzed like a z. The spelling records the plural morpheme, not the sound of the moment. So asking whether a script is phonetic or meaning-based gets you nowhere. Ask two other things: which layer its signs point at, and how many units that layer has in that particular language.
1821 年，一个叫塞阔雅（Sequoyah）的切罗基人完成了一件很难的事。他不识英文，只知道「纸上能说话」这回事是真的；他花了十几年把切罗基语的音节一个个分出来，给每个配一个符号，最后定在 85 个。效果快得惊人：几年之内会读写的人就很多，1828 年《切罗基凤凰报》用切罗基文和英文双栏印刷。
但这套音节表有两件事不记：切罗基语有声调，也区分元音长短，纸面上两样都不标。结果「水」和「盐」写出来是同一串符号——差的正是元音长短。会说这门话的人靠上下文一眼分开，不会说的人分不开。这不是设计失误：塞阔雅是给已经会说切罗基语的人设计的，母语者自己能补回去的差别没必要占纸面。
同一件事在阿拉伯文、希伯来文上更极端：正规文本根本不写短元音，这类文字叫 abjad（辅音文字）。能行是因为意义装在辅音里。k、t、b 这三个辅音管「写」这摊事：kataba 是「他写了」，kitāb 是「书」，kātib 是「写字的人」，maktab 是「办公室」——元音换的是语法和派生那一层，扔掉它，词根完好无损。英语扔不起：bit、bat、but、bet 抹掉元音全变成 bt。
另一头是把差别记得比音位还细。1443 年朝鲜世宗公布的谚文（Hangul，韩文字母），辅音字母照着发音部位画：ㄱ 画舌根向后抬起堵住喉咽（像英语 go 开头那个音），ㄴ 画舌尖抵在上齿龈上（像 no 开头那个音），ㅁ 是双唇合起来的方框（像 me 开头那个音）。这是在音位之下又记了一层「用哪儿发的音」。而字母这条路从头被开出来的可能只有一次：今天的字母文字基本都溯到约四千年前埃及境内的闪米特人借圣书字字形做出的原始西奈文字，拉丁、希腊、西里尔、希伯来、阿拉伯都在这条线上，谚文是公认的例外（这条谱系靠字形与借用链重建，个别分支还有争论）。
所以文字不是语言的镜子，是一层有损的记录：只留设计者听得出、又认为非留不可的那些差别，剩下的交给读者补——补得回来，是因为读者已经会说这门话。这也正是「阅读是大脑没准备过的技能」那期的落点：说话会自己长出来，认字不会。
EN In 1821 a Cherokee man named Sequoyah finished something very hard. He could not read English; he only knew that talk on paper was real. Over more than a decade he picked the syllables of Cherokee apart and gave each one a sign, settling on 85. The uptake was startlingly fast: within a few years a great many Cherokee could read and write, and in 1828 the Cherokee Phoenix was printed in parallel Cherokee and English columns. But the syllabary records two things not at all: Cherokee has tone, and it distinguishes vowel length, and neither is marked. So the words for water and salt are written identically — what separates them is vowel length. Fluent speakers tell them apart from context instantly; learners cannot. This is not a design flaw. Sequoyah built it for people who already spoke Cherokee, and a distinction a native speaker restores by himself need not take up space on the page. Arabic and Hebrew take the same logic further: standard text omits short vowels altogether, a type of script called an abjad. Why that works: the meaning lives in the consonants. The three consonants k, t and b own the business of writing — kataba "he wrote", kitāb "book", kātib "writer", maktab "office" — and the vowels that come and go are carrying the grammar and the derivation. Dropping the vowels drops exactly that layer, leaving the root intact. English cannot afford the move: strip the vowels from bit, bat, but and bet and all four become bt. The opposite extreme records a layer finer than the phoneme. The Korean script Hangul, promulgated by King Sejong in 1443, draws its consonant letters after the speech organs: ㄱ depicts the back of the tongue rising to block the throat (the sound starting go), ㄴ the tongue tip against the upper gum ridge (the sound starting no), ㅁ the square of the closed lips (the sound starting me). That is a layer below the phoneme — where in the mouth the sound is made. And the alphabet road itself may have been opened only once. Essentially every alphabetic script today traces back to Semitic workers in Egypt around four thousand years ago, who borrowed hieroglyphic shapes to make the Proto-Sinaitic script; Latin, Greek, Cyrillic, Hebrew and Arabic all sit on that line. Hangul is the acknowledged exception, designed from articulatory principles instead. Evidence strength: the family tree is reconstructed from letter shapes and borrowing chains, and particular branches, the Indic one for instance, are still argued about. Put the three cases together and the conclusion is no longer that writing mirrors language. Writing is a lossy record: it keeps only the distinctions its designer could hear and judged worth keeping, and leaves the rest for the reader to restore. The reader manages because he already speaks the language. Which is exactly where the topic on reading being an unprepared skill landed — speech grows by itself, literacy does not, and every child has to learn afresh what this record throws away on his behalf.
塞阔雅的选择说明，给母语者用可以大方省掉一堆差别；但同一套省法会让学这门语言的人（包括从小没机会说的本族孩子）很吃亏。这两拨人的需求是真冲突，没有技术上的最优解，决定权该在使用社群手里。
缺的那一步很具体：它还没走到记声音。一个符号只要只对着东西和情绪，能表达的就只有画得出来的东西。真要变成文字，得有人开始拿它记读音。
意味着「看声旁猜读音」是个有一半时间会骗你的策略，而孩子没法事先知道是哪一半。这也解释了为什么这套系统的入门期更长，过了门槛之后阅读速度却并不吃亏。