键入会复利，点击不会

一道读者已经不再看见其分层的梯度#

世界上每一台计算机都有命令行。大多数使用计算机的人一辈子也不会看见它。他们不会知道它在那里。今天计算的形状，就是一道读写能力梯度的形状——大多数用户处在“家电”那一端，看不见位于他们之上的东西；而有读写能力的少数人，则太习惯于自己的位置，以至于已经不再注意到这道梯度的存在。

这并不新鲜。同一个模式塑造了从文字发明到文字普及之间的那几个世纪。一名书吏一小时能产出的东西，一个不识字的专业人士一年也产不出来。而那位不识字的专业人士并不知道这一点。他从来没见过书写。

这道梯度，若要精确地命名，共分四层。​家电用户能做按钮允许他做的事。​实用读者能把家电用户所执行的东西编排成句。​脚本作者能把实用读者一步一步组合起来的东西自动化。​系统作者能架构脚本作者无从推理的东西。

这些不是智力的层级。它们是一道有结构的梯度上的位置，每一层对其下方各层都部分不可见——原因和前现代的商人看不见书吏在做什么，是同一个结构性原因：占据上一层所需要的知识，从下一层望过去没有可见的形态。

这道梯度的经验规模不难确立，尽管各层本身抗拒直接测量。梯度的顶端——系统作者层，也就是职业软件开发者——全球约有 2700 万到 4700 万人，具体数字取决于统计口径。¹ 而放在如今已超过 60 亿的全球计算机用户总数上看，这是所有用过计算机的人里的 0.5% 到 0.9%。¹

但系统作者层并不等于整个“有读写能力的阶层”。在它之下、在家电用户这块地板之上，坐着一个大得多的人群。Scaffidi、Shaw 和 Myers 于 2005 年发表的一项 IEEE 研究——2014 年被评为 IEEE 视觉语言与以人为中心计算研讨会的最具影响力论文——预测到 2012 年，美国约有 5500 万劳动者会以构成编程活动的方式使用电子表格或数据库，其中约 1300 万人会自称程序员。相对于不到 300 万的职业开发者，这个比例大约是每一位职业开发者对应四位终端用户程序员。² 论文作者之一 Brad Myers 称这些数字是“惊人的数字”，它们“使这篇论文被高度引用，也极大地推动了更多研究者去关注这一类程序员”。² 二十年后，Gartner 预测，公民开发者——那些使用低代码、宏和自动化工具创建应用能力的员工——在大型企业中的数量将至少以四比一超过职业开发者。³ 同一个比例，维持了二十年。梯度的第二层不是边缘地带。按产量计，它是计算生产的主导模式。

劳动力市场同样精确地显示出家电用户层。对 4300 万条美国招聘启事的分析发现，92% 的岗位至少要求一项数字技能；与此同时，三分之一的美国劳动者不具备当今岗位所要求的基础数字技能。⁴ 在包括德国、英国、美国、加拿大和澳大利亚在内的 15 个国家中，需求量最大的数字技能是 Microsoft Office。⁴ 国际成人能力调查（PIAAC）发现，23% 到 24% 的美国成年人在“技术丰富环境中的问题解决”上低于一级水平，也就是说，他们只能在熟悉的数字环境中完成最基本的单步任务。⁵ 家电用户层不是一个修辞上的构造。它是劳动力中一个可测量的特征。

同行评议的综述确认了这些调查所提示的东西。终端用户软件工程这一领域的综述在 2011 年开篇，说的正是这个观察：“今天大多数程序不是由职业软件开发者写的，而是由那些在其他领域拥有专长、为了需要计算支持的目标而工作的人写的。”⁶ 梯度是计算的主导经验状况，而不是强加在它之上的一个概念。

梯度的各层，从它上面的每一个位置看过去，可见度并不相同。那些花了整个职业生涯编排脚本、流水线和系统的工程师，多年来一直生活在一道梯度上有读写能力的那一侧，而他们周围的大多数人从自己所站的位置看不见这道梯度。这不是指责。这是一个结构性的观察。书吏懂拉丁文不该被责备；商人不懂拉丁文也不算有缺陷。他们在同一个知识经济中占据着结构上不同的位置——而商人从自己的位置，看不见书吏的位置。

接下来要做的，是把这个结构命名得足够精确、从而足够有用——不是作为一份判决书，而是作为一件诊断工具。

GUI 遮住了什么；CLI 又暴露了什么#

GUI 应用与命令行界面之间的差别，首先并不是审美偏好或人机效率上的差别。它是“界面宣称这台机器能做什么”上的差别。

一个 GUI 应用呈现的是一个已完工的表面。文件菜单、功能区、对话框——这些是这个界面关于机器可能性的隐含主张。在那个表面之内操作的用户，是在一个“机器是干什么用的”的主张之内操作。这个表面并不虚假，但它是有界的：GUI 关于机器能力的主张，是机器被打包好的能力，而不是机器真正的生成力。

CLI 则直接呈现那道接缝。每一条命令都是对一个程序的调用；每一个能编排命令的用户，都能编排程序。这个界面不去枚举机器的可能性——它暴露的是可能性被组合起来的机制。这正是这道梯度赖以建立的结构性差别：不是一个界面更快或更高效，而是一个界面把机器呈现为可编程的表面，另一个把它呈现为功能固定的家电。

会写 shell 脚本的人群，即便在能拿到的技术倾向最强的样本里——Stack Overflow 开发者调查，来自 185 个国家的六万五千多名受访者——使用 Bash 或 Shell 进行大量开发工作的也只占全部受访者的 33.9%。⁷ 即便在这个偏向开发者的人群中，能读写 shell 的那一层也远未普及。把这个数字读作 CLI 使用率的上限、而不是一个人口估计，它正好讲出了这道梯度的故事：即便在那些拥有一切理由去了解 CLI 的人当中，仍有三分之二不用 shell 来编排。

针对这套梯度框架最有力的反驳，必须在这里被点名并吸收：每一个主流 GUI 生态内部，都含有自己的一道梯度。Microsoft Excel 从 1993 年起就有 VBA。Apple 同年起就有 AppleScript。Microsoft Office 从 2016 年起有 Power Automate。Mac 在 2005 年加入了 Automator，2018 年加入了快捷指令。Google Workspace 从 2012 年起有 Apps Script。Scaffidi、Shaw 和 Myers 所记录、并在二十年后被 Gartner 确认的那个 4∶1 比例，很大程度上正是那些已经迈进这些可组合层的人。²³ 写 VBA 的人不在家电用户层。写 Excel 公式的人也不在家电用户层。他们在实用读者层，生产的是机器随后会执行的、持久的指令。

这个结构性主张比“GUI 隐藏了可编程性”更尖锐。更尖锐的说法是：​GUI 的表面遮住了它上方的那道梯度，而大多数用户从来找不到向上的第一级台阶。 电子表格本身就是一份清单——在 Jack Goody 那个精确的技术意义上，是一种口头话语无法生成的认知形式。⁸ VBA 宏是一段写下来的程序。已经跨到可组合层上的实用读者，其梯度位置与家电用户确实不同。但家电用户看不见那个位置。GUI 的默认表面不呈现任何需要向上迈一步的东西，而那第一级台阶从外面是看不见的。

家电用户那一端之上的每一层都有这个性质：从上一层看得见，从下一层看不见。脚本作者能看见写公式的人看不见的东西。系统作者能看见脚本作者看不见的东西。而在这个结构里生活和工作的工程师，已经不再看见它是一个结构了——因为它就是他呼吸的空气。

为什么在家电层之上生成的知识会累积，而在家电层生成的知识会随着持有它的从业者一起消失——这个问题需要一个机制。机制来自一套四十年前的词汇。

书写把认知者与被认知者分开：机制#

那个结构性的观察——梯度的各层从下往上看不见，在 CLI 中生成的知识会累积而在 GUI 中生成的知识会随从业者一同消失——不是一个比喻。它有机制。这个机制已经被三个学术体系以不同的语域命名过。每一个都命名了从业者词汇无法命名的东西。

Ong 命名了那次分离。 Walter Ong 1982 年的《Orality and Literacy》确立了“书写把认知者与被认知者分开”这一点——“书写把认知者与被认知者分开，由此为‘客观性’设立了条件，这里的客观性指的是个人的抽离或拉开距离。”⁹ 相比之下，口语文化“让思维贴近人的生活世界，把事物和议题人格化，把知识储存在故事里”。⁹ 存在于口语文化中的知识，存在于人身上、存在于表演之中；要取用它，需要在场。存在于书写文化中的知识，存在于人工制品之中；要取用它，只需要那件制品。

用到计算上，这不是一个类比。它是一个机制。shell 脚本把过程外化成一件可检视的制品。点击会话则把过程保留在从业者的表演之中。脚本会累积；点击会话不会。工程师走了，脚本还在。GUI 从业者走了，他所携带的知识——点了哪个界面、按什么顺序、在什么情况下——就跟着他一起走了。Ong 的框架在形式而非偏好的层面上，命名了这件事为什么会发生。

Goody 命名了那个形式。 Jack Goody 1977 年的《The Domestication of the Savage Mind》确立了：系统性的清单——“离散的、有序的、在物理上被安放的、可以从多个方向阅读的”——是一种口头话语无法生成的认知形式。⁸“特别是，清单很少出现在口头话语中。”⁸ git 提交日志是清单。Terraform 配置是清单。运维手册是清单。shell 管道是清单。一段控制台点击流不是。这不是复杂度或意图的问题；这是形式的问题。清单可以被再次阅读、被查询、被与先前状态比对、被交给一个在它被写下时并不在场的人。表演不能。

这也解答了那个反驳，即现代 GUI 应用是有审计日志的——CloudTrail 记录了在 AWS 控制台里点了什么，Figma 保留了版本历史。日志记录的是一次表演。脚本编排的是一次表演。它们是不同种类的制品：一份“点了什么”的日志无法被当作新配置重新执行；一个 Terraform 文件可以。Goody 关于“清单作为认知形式”与“口头话语作为表演”的区分，是在结构而非技术偏好的层面上完成这项工作的。

Clanchy 命名了那个模式。 CLI／IaC 实践与书写型知识架构之间的结构对应，已经在行政史上被研究得最细致的转变之一中被记录下来，记录者是 Michael Clanchy 的《From Memory to Written Record: England 1066–1307》。

Clanchy 记录了口头行政——经见证的口头誓言、记忆、在场者的证词——如何在英格兰治理的两个半世纪里被书面文书取代的具体机制。那次转变与计算当下的 CLI／IaC 转变之间的对应是结构性的，而不是装饰性的：

口头誓言要求所有各方在场；没有他们，它无法被重演。点击会话要求那个点击过的人；没有他的记忆，它无法被复现。书面特许状则是持久的、可传递的、无需原起草人也能被检视的。Clanchy 记录到，到爱德华一世治下，“国王或领主的令状抵达了英格兰的每一位执行吏和每一个村庄，使书写在整个乡间变得寻常。”¹⁰ shell 脚本是持久的、受版本控制的、无需写它的那位工程师也能被审计的。

《末日审判书》——威廉一世 1086 年对英格兰地产所作的书面清查——让产权无需原见证人也能被检视。基础设施即代码让云端状态无需当初开通它的工程师也能被检视。书面记录在这两种情形下都不只是一次誊录；它是一种治理机制。

爱德华一世治下的 quo warranto​（权源令状）诉讼，把书面行政的实践在此前一个世纪里已在操作上确立的东西形式化了：“记忆，无论是个人的还是集体的，若没有清楚的书面证据支持，都不予采信。”¹⁰ 一份关于自己持有何种特权的口头陈述——没有书面凭据——在法律上是不充分的。由此确立的原则是：作为“做过什么”的权威记载，书面记录优先于口头记忆。¹⁰ 今天的合规审计要求 IaC 状态，理由完全相同：一份关于开通了哪些基础设施的口头说明，在缺少书面记录的情况下，无论在法律上还是在运营上都不充分。

而 Clanchy 对计算这个案例最精确的一项发现是：“世俗读写能力是从官僚制里长出来的，而不是出于任何对教育或文学的抽象渴望。”¹⁰ CLI 的普及走的是同一个机制：靠的不是大规模培训计划，而是不断扩张的、要求产出书面制品的行政需求——要求基础设施受版本控制的 DevOps 实践，要求状态可检视的云治理，以及把点击会话视为不充分的合规框架。

Clanchy 还以一种可以直接映射到计算之上的精度，记录了中世纪英格兰这道梯度的内部结构。中世纪的书写世界不是二元的——一边是书吏，另一边是文盲。它是三分的：能产出文书的职业书写者书吏，使用文书却不产出文书的实用读者，以及虽然完全不识字、却通过代理人参与书写文化的人。“那些使用书写的人参与了书写文化，即便他们并未掌握书记员的技艺。”¹¹ 写 Excel 公式的人参与了有读写能力的计算世界，即便他并未掌握 shell。写 VBA 宏的人是一位实用读者。这道梯度从来都有多个被命名的层——而每一层，对它下面的各层，也从来都是部分不可见的。

结构模板是同一个：需要产出书面制品的人群不断扩张，其驱动力不是教育，而是行政的实际要求。Eisenstein 论证过，印刷机促成了“印刷固定性”——一种抄写文化无法提供的、供知识系统性累积与纠错的机制。¹² Clanchy 这个锚点并不依赖那个论点能否经受住 Johns 的批评：Clanchy 的论证针对的是行政实践，而不是机械固定性理论。

这是一次复现，不是一次范式转移#

计算正在经历英格兰在 1066 到 1307 年间经历过的事情。这不是一次范式转移。这是一个此前跑过的模式的结构性复现——那些不得不产出书面制品的人群，在一场缓慢的、充满争议的、由需求驱动的扩张中不断变大，而梯度的扩张靠的不是普及教育，而是行政上的必要。

“复现”这个框架之所以重要，是因为它约束了主张的边界。在活着的人记忆之内，已经跑过三次生产端的民主化，每一次都既展示了这个模式，也展示了它的界限。

1900 年，柯达以一美元推出了布朗尼相机。此前摄影需要笨重的器材、玻璃干板、有毒的化学显影和专门训练。布朗尼把专家的活儿吸收进了这台家电：柯达在自己的工厂里冲洗胶卷；顾客只管按快门。第一年就有超过 15 万台布朗尼相机出货。¹³ 快照走向了大众。职业摄影并没有消失。电影摄影仍然稀缺。一项有限技能的门槛降到了大众市场的价位。而那个新增人群用这项技能所能做到的上限，仍然是有限的。

1985 年，Aldus PageMaker 于 1 月 23 日在苹果股东大会上发布，并在当年 7 月发出第一个版本。¹⁴ 在 PageMaker 之前，专业品质的文档需要拥有专门设备和训练的排版工。在 PageMaker 之后，一个拥有 Mac 和 LaserWriter 的办公室职员就能做出排版级质量的版面。Paul Brainerd 造出了“桌面出版”这个词。Steve Jobs 后来说，桌面出版救了 Macintosh。¹⁴ 排版业在此后十年间大幅萎缩。产出格式化文档的门槛降到了任何一个 Mac 用户。而那个人群用 PageMaker 所能做到的上限——版面，可以；排版软件，不行——仍然是有限的。

2004 年，苹果推出 GarageBand，随每一台 Mac 免费附赠，明确宣称要“让音乐创作民主化”。到 2023 年，非大厂与独立发行商——这个类别包含占 40.8% 的独立厂牌和占 5.9% 的艺人直发发行商，而不只是独立艺人个人——按所有权口径占据了全球录制音乐市场的 46.7%，收入达 143 亿美元。¹⁵ 家庭录音室的门槛对 Mac 用户降到了零。而在 Logic Pro 那个层级上创作的门槛并没有降。

每一次民主化都有相同的结构。一项此前需要专门知识的有限技能，被吸收进一台家电。非专家获得了这项技能一个特定的、有限的上限。而专家层级的实践——电影摄影、排版软件开发、最高段位的职业音乐制作——仍然稀缺。

Eisenstein 论证过，印刷机的运作方式与此不同——它不是生产端的民主化，而是消费端的民主化。¹² 印刷降低的是阅读的门槛，不是书写的门槛。它让阅读变得值得去学，因为如今有了买得起的、值得一读的东西。印刷之后书写仍然稀缺；即便是 1600 年的一位已出版作者，也很可能是他所在教区里能起草一封信的少数几个人之一。印刷机没有教文盲写字。它终结的是书吏对阅读的垄断，而不是对书写的垄断。

CLI／GUI 这道梯度，在结构上与上述任何一个先例都不同。CLI 不是一件有固定上限的、用途有限的工具。它是通向一台通用机器的操作者界面——这正是为什么这道梯度有如此多的层，为什么这些层还在不断向上延伸，以及为什么“什么算是顶端”这个问题至今真正悬而未决。布朗尼的上限是电影摄影。CLI 的上限，是系统作者此刻正在脚本作者之上建造的任何东西。

梯度何时要紧：组织层面的样貌与适用条件#

这个结构性论证有一副组织层面的面孔，而那些处在能读写 CLI 位置上的工程师，一直生活在它里面，只是不太有词汇去说出它的结构性成因。

那些主要由家电端用户构成的团队——通过 AWS 控制台管理云基础设施的工程师，机构知识活在从业者表演中而非可检视制品中的组织——是在当代规模上运行着口语文化的知识架构。这些病症在从业者话语里都有名字：ClickOps、部落知识、孤儿资源、那个离职时把运维手册一起带走的工程师。从业者词汇没有提供的，是这些失败的结构性成因。

结构性成因在于：口头实践中生成的知识是稳态的。Ong 观察到，“口语社会很大程度上活在一个当下之中，这个当下通过甩掉那些不再具有当下相关性的记忆，来保持自身的均衡或稳态。”¹⁶ 把它用到同期群的尺度上——团队成员大约每 3 到 5 年轮换一轮——这个机制描述的正是那些运营知识只存在于从业者表演中的组织会发生什么。当下没有人在排练的知识，会在持有它的那一批人离开时失传。而在个体离职这个尺度上，起作用的是另一个机制：这是隐性知识传统所描述的那种失败模式，从 Michael Polanyi 的《The Tacit Dimension》到 Nonaka 和 Takeuchi 的《The Knowledge-Creating Company》——那种因为从未被外化成可检视制品、因而无法在个体离开后存续的具身知识。两种机制在不同尺度上运作，产生的却是同一类组织后果：因为从未被写下来，所以再也取不回来的知识。

组织层面的调查证据与这个结构性诊断是一致的。2025 年的行业数据估计，企业云基础设施支出中有 21%——相当于对照 Gartner 当年云支出预测的约 445 亿美元——被浪费在利用率不足、闲置或成为孤儿的资源上。¹⁷ 在同一项调查中，报告拥有实时数据的组织不到一半：闲置资源为 43%，未使用或孤儿资源为 39%，超额配置的工作负载为 33%。¹⁷ 只有 32% 的开发者报告拥有完全自动化的成本节约实践。¹⁷

《末日审判书》让产权变得可检视；基础设施即代码在同样的结构意义上让云端状态变得可检视——书面记录使得查阅无需创建者在场。¹⁰ terraform plan 命令和漂移检测之所以存在，恰恰是因为 IaC 社群已经命名了“地图与疆域”的问题（书面配置可能与已部署状态不符），并建造了机制把这个差距浮到台面上。

但这道梯度是一件诊断工具，不是一份普遍判决。适用条件很要紧。

当同一项任务要做很多遍时，梯度要紧——也就是自动化有正回报的时候。当工作需要被检视、被审计、或被移交给一个在它被执行时并不在场的人时，梯度要紧。当劳动者的处境会改变时，梯度要紧——因为读写型技能是可迁移的，而表演不是。当规模超出操作者持续注意力所及时，梯度要紧——也就是“人能手工追踪的”与“系统所要求的”之间的差距，已经大过活人所能弥合的程度。

而当任务确实是一次性的时候，梯度就没那么要紧。当劳动者的领域专长比自动化节省下来的更值钱时，梯度没那么要紧——会计的时间用来做会计，比用来写 VBA 更值钱。当理性的应对是委托给专家时，梯度也没那么要紧——这就是商人／书吏模型，它在经济上是高效的，而不是一次读写志向上的失败。那位口述契约给书吏的中世纪商人是在专业化，不是在失败。书吏也在专业化。分工创造财富；这道梯度不是一份针对专业化的判决书。

而在梯度确实要紧的那些情形里——重复、审计、移交、规模——口语文化的知识架构会产生可预测的组织性失败，而这些失败有一个结构性成因，从业者词汇为它命了名，却没有解释它。口语／书写这套框架所提供的，正是那个解释。

悬而未决的问题：像布朗尼，像印刷机，还是真正的新事物#

这道梯度挺过了此前每一次门槛降低事件。摄影的民主化没有消解快照与电影摄影实践之间的区别。桌面出版的民主化没有消解一份排好版的通讯稿与一个排版程序之间的区别。家庭录音的民主化没有消解一段 GarageBand 音轨与一次 Logic Pro 母带处理之间的区别。

如今已经出现了两种截然不同的 LLM 智能体架构，而它们之间的区别，直接映射到“累积的制品”与“转瞬即逝的表演”这组口语／书写的区别上。

驱动浏览器的智能体​——Browser-Use（那个 Python 库，截至 2026 年 4 月约有 7.8 万个 GitHub star）和 Skyvern（Y Combinator 投资，“计算机视觉加 LLM”）——通过截图和点击与界面交互。¹⁸ 它们在智能体这一层保留了口语式的点击模式。智能体浏览 GUI 的方式，和从业者浏览 GUI 的方式一样：它看见一屏画面，识别出可点击的元素，发出点击，然后往前走。智能体所执行的那个过程，不会作为一件可检视的制品累积下来。它是一次表演。

生成脚本的智能体​——Claude Code（Anthropic 基于终端的智能体编码工具，能规划并执行一系列动作，包括创建文件、跨文件重构、运行测试和 git 操作）和 GitHub Copilot CLI（把 AI 驱动的编码辅助直接带到命令行）——生成的是可执行、可检视的脚本。¹⁸ 智能体产出的那个过程是一件书面制品：可纳入版本控制，可审计，重新执行时无需智能体在场也能复现。

Skyvern 自身的演进很有教益。这家公司 2025 年 10 月的博客记录了这条推理：“如果 Skyvern 能把自己的推理编译成代码并运行它，而不是让一个 LLM 留在循环里，自动化就会变得更快、更便宜、也更可靠。”¹⁸ 当可复现性和可检视性变得要紧时——当口语模式的智能体的输出需要被可靠地重跑时——智能体想要写一段脚本。即便在一个以浏览器驱动起家的产品内部，这个结构性区分依然存在。

这道梯度挺过了 LLM 转型。口语／书写的区分正在智能体这一层被复制出来，带着同样那些结构属性——可重放、可检视、无需原作者也能传递——正是这些属性，在 Clanchy 笔下的英格兰把书面特许状与口头誓言区分开来。

LLM 这一刻究竟像布朗尼，还是真正的新事物，仍是悬而未决的问题。

如果 LLM 这一刻像布朗尼，那它民主化的是计算实践中一个有限的切片。不会写代码的人写出小小的一次性脚本——就像布朗尼用户拍快照那样——而系统作者那一层，仍旧像今天一样遥远、一样人烟稀少。梯度多出一层：提示词工程师作为新的实用读者，能编排家电用户编排不了的东西，却架构不了系统作者能架构的东西。这次复现，又跑了一遍。

如果 LLM 这一刻是真正的新事物——是历史上第一次，通用可编程指令的生产门槛，为一个此前从未跨过它的人群降了下来——那么这道梯度就正在被重造。不是被填进一个新层。是被重造。

在这一整串事件里，LLM 这一刻是唯一一个可能消解一道梯度、而不只是把它填满的候选者——因为它降低的是生产通用指令的门槛，而不是消费一项有限上限技能的门槛。它到底会不会，目前还看不出来。智能体这一层的分岔是先行指标：每一个新发布的智能体，要么保住了这道梯度，要么延长了它。浏览器驱动者保住了口语模式。脚本生成者保住了书写模式。每一次架构选择，都是这个问题走向答案过程中的一个数据点。

印刷机终结了书吏对阅读的垄断；书写依然稀缺。布朗尼终结了暗房对快照的垄断；电影摄影依然稀缺。LLM 这一刻究竟会终结书吏对通用机器的垄断——还是只不过给一道旧梯度添了一个新层——目前还看不出来。而有读写能力的那一层，正是最有条件看着答案浮现出来的那一层。

延伸阅读#

• Neal Stephenson. "In the Beginning Was the Command Line." 1999. 可在 garote.bdmonkeys.net/commandline/index.html 获取。——关于 CLI 与 GUI 哲学的经典先行文本；写给已经在用 CLI 的读者；它没有问本文所问的人群层面的问题，而这恰恰使它既是最有用的同伴，也是最有用的对照。
• Michael Polanyi. The Tacit Dimension​. Anchor Books, 1966.——关于“无法在个体离职后存续的具身知识”的奠基文本；提供了第 5 节所援引的团队尺度隐性知识机制（区别于 Ong 的同期群尺度稳态机制）。这两个框架在不同尺度上运作，处理的是组织知识流失的不同侧面。
• Ikujiro Nonaka and Hirotaka Takeuchi. The Knowledge-Creating Company: How Japanese Companies Create the Dynamics of Innovation​. Oxford University Press, 1995, 第 3–5 章。——把 Polanyi 的隐性知识传统扩展到组织知识创造；SECI 模型（社会化、外化、组合、内化）提供了一个框架，用以理解隐性知识如何成为显性的制度知识——以及为什么跳过“外化”这一步的组织，会失去其从业者曾经知道的东西。
• Eric A. Havelock. Preface to Plato​. Harvard University Press, 1963.——Ong 那套框架的先驱；考察古希腊由口语文化转向书写文化的认知后果；为“书写重构思想，而不只是记录思想”这一主张提供了最深的根基。
• Alan Kay. "A Personal Computer for Children of All Ages." Proceedings of the ACM Annual Conference​, 1972.——通用个人计算机愿景的奠基文献；阐明了为什么“用途有限的家电”与“通用机器”之间的区别，是本文的结构性主题，而不只是它的由头。

Type Compounds; Click Doesn't

A Gradient with Bands the Reader Has Stopped Seeing#

Every computer in the world has a command line. Most people who use computers will never see it. They will never know it is there. The shape of computing today is the shape of a literacy gradient — with the majority of users positioned at the appliance end, unable to see what sits above them, and with the literate minority so accustomed to their position that they have stopped noticing the gradient exists.

This is not new. The same pattern shaped the centuries between writing's invention and its democratization. A scribe could produce in an hour what an illiterate professional could not produce in a year. And the illiterate professional did not know this. They had never seen a script.

The gradient, named precisely, runs in four bands. The appliance user can do what the buttons allow. The pragmatic reader can compose what the appliance user performs. The script writer can automate what the pragmatic reader composes one step at a time. The systems author can architect what the script writer cannot reason about.

These are not levels of intelligence. They are positions on a structured gradient, each partly invisible to the bands below it — for the same structural reason the pre-modern merchant could not see what the scribe did: the knowledge required to occupy the upper band has no visible form from the lower band.

The empirical scale of this gradient is easy to establish, even if the bands themselves resist direct measurement. The top of the gradient — the systems-author band, professional software developers — numbers approximately 27 to 47 million globally, depending on the methodology.¹ Set against a total global population of computer users that now exceeds 6 billion, that is 0.5 to 0.9 percent of all people who use a computer at all.¹

But the systems-author band is not the literate class as a whole. Below it, and above the appliance-user floor, sits a much larger population. A 2005 IEEE study by Scaffidi, Shaw, and Myers — named the most influential paper at the IEEE Symposium on Visual Languages and Human-Centric Computing in 2014 — projected that by 2012, approximately 55 million US workers would use spreadsheets or databases in ways that constitute programming activity, and roughly 13 million would describe themselves as programmers. Against fewer than 3 million professional developers, that is a ratio of approximately four end-user programmers for every professional developer.² Brad Myers, one of the paper's authors, called the figures "astonishing numbers" that "resulted in this paper being highly cited, and highly influential in getting more researchers to focus on this class of programmers."² Two decades later, Gartner predicted that citizen developers — employees who create application capabilities using low-code, macro, and automation tools — would outnumber professional developers by at least four to one in large enterprises.³ The same ratio, held for twenty years. The gradient's second band is not a fringe. It is the dominant mode of computing production, by volume.

The labor market shows the appliance-user band with equal precision. Analysis of 43 million US job postings found that 92 percent require at least one digital skill; simultaneously, one-third of US workers lack the foundational digital skills today's jobs require.⁴ The most in-demand digital skill across 15 countries — including Germany, the UK, the US, Canada, and Australia — is Microsoft Office.⁴ The international survey of adult skills (PIAAC) found that 23 to 24 percent of US adults perform below Level 1 in problem-solving in technology-rich environments, meaning they can complete only the most basic single-step tasks in familiar digital environments.⁵ The appliance-user band is not a rhetorical construction. It is a measurable feature of the workforce.

Peer-reviewed synthesis confirms what these surveys suggest. The field-level survey of end-user software engineering opened, in 2011, on exactly this observation: "Most programs today are written not by professional software developers, but by people with expertise in other domains working towards goals for which they need computational support."⁶ The gradient is the dominant empirical condition of computing, not a conceptual imposition on it.

The gradient's bands are not equally visible from every position on it. Engineers who have spent a career composing scripts, pipelines, and systems have lived for years on the literate side of a gradient most of the people around them cannot see from where they stand. This is not an accusation. It is a structural observation. The scribe was not blameworthy for knowing Latin; the merchant was not deficient for not knowing it. They occupied structurally different positions in the same knowledge economy — and the merchant could not see the scribe's position from the merchant's own.

What follows is an attempt to name that structure precisely enough to be useful — not as a verdict, but as a diagnostic instrument.

What the GUI Conceals; What the CLI Reveals#

The difference between a GUI application and a command-line interface is not, primarily, a difference in aesthetic preference or ergonomic efficiency. It is a difference in what the interface claims the machine can do.

A GUI application presents a finished surface. The file menu, the ribbon, the dialog box — these are the interface's implicit argument about the machine's possibilities. The user who operates within that surface is operating within a claim about what the machine is for. The surface is not false, but it is bounded: the GUI's claim about the machine's capabilities is the machine's packaged capabilities, not the machine's actual generativity.

The CLI presents the seam directly. Every command is a call to a program; every user who can compose commands can compose programs. The interface does not enumerate the machine's possibilities — it exposes the mechanism by which possibilities are composed. This is the structural difference the gradient is built on: not that one interface is faster or more efficient, but that one interface presents the machine as a programmable surface and the other presents it as a fixed-function appliance.

The shell-scripting population, even among the most technically inclined sample available — the Stack Overflow Developer Survey, with more than 65,000 respondents across 185 countries — stands at 33.9 percent of all respondents using Bash or Shell for extensive development work.⁷ Even within this developer-biased population, the literate-shell band is well below universal. Read as a ceiling on CLI use, not as a population estimate, this figure tells the gradient's story precisely: even among people who have every incentive to know the CLI, two-thirds do not compose in shell.

The hardest counter to the gradient framing must be named and absorbed here: every major GUI ecosystem contains its own internal gradient. Microsoft Excel has had VBA since 1993. Apple has had AppleScript since the same year. Microsoft Office has had Power Automate since 2016. Mac added Automator in 2005 and Shortcuts in 2018. Google Workspace has had Apps Script since 2012. The 4:1 ratio documented by Scaffidi, Shaw, and Myers — and confirmed by Gartner two decades later — is largely the population of people who have moved into these composability layers.²³ The VBA writer is not at the appliance-user band. The Excel formula writer is not at the appliance-user band. They are at the pragmatic-reader band, producing persistent instructions the machine then executes.

The structural claim is sharper than "the GUI hides programmability." The sharper claim is this: the GUI's surface conceals the gradient above it, and most users never find the first step up. The spreadsheet is itself a list — in Jack Goody's precise technical sense, a cognitive form that oral discourse cannot generate.⁸ The VBA macro is a written program. The pragmatic reader who has crossed onto the composability layer is genuinely at a different gradient position than the appliance user. But the appliance user has no view of that position. The default surface of the GUI presents nothing that requires stepping up, and the first step is invisible from the outside.

Every band above the appliance-user end has this property: visible from the band above, invisible from the band below. The script writer can see what the formula writer cannot see. The systems author can see what the script writer cannot see. And the engineer who lives and works in this structure has stopped seeing that it is a structure — because it is the air they breathe.

Why knowledge generated above the appliance band accumulates while knowledge generated at the appliance band disappears with the practitioner who held it — that question requires a mechanism. The mechanism comes from a vocabulary forty years old.

Writing Separates the Knower from the Known: The Mechanism#

The structural observation — that the gradient's bands are invisible from below, that knowledge generated in CLI accumulates while knowledge generated in GUI disappears with the practitioner — is not a metaphor. It has a mechanism. The mechanism has been named, in different registers, by three bodies of scholarship. Each names something the practitioner vocabulary cannot.

Ong names the separation. Walter Ong's 1982 Orality and Literacy established that writing separates the knower from the known — "Writing separates the knower from the known and thus sets up conditions for 'objectivity', in the sense of personal disengagement or distancing."⁹ Oral cultures, by contrast, "keep their thinking close to the human life-world, personalizing things and issues, and storing knowledge in stories."⁹ The knowledge that exists in an oral culture exists in persons and in performances; to retrieve it requires presence. The knowledge that exists in a literate culture exists in artifacts; to retrieve it requires only the artifact.

Applied to computing, this is not an analogy. It is a mechanism. The shell script externalizes the procedure into an inspectable artifact. The click session keeps the procedure embodied in the practitioner's performance. The script accumulates; the click session does not. When the engineer leaves, the script remains. When the GUI practitioner leaves, the knowledge they carried — which interface they clicked, in which order, under which circumstances — leaves with them. Ong's framework names why this happens at the level of form, not preference.

Goody names the form. Jack Goody's 1977 The Domestication of the Savage Mind established that the systematic list — "discontinuous, ordered, physically placed, readable in multiple directions" — is a cognitive form oral discourse cannot generate.⁸ "In particular, the list rarely appears in oral discourse."⁸ The git commit log is a list. The Terraform configuration is a list. The runbook is a list. The shell pipeline is a list. A console clickstream is not. This is not a question of complexity or intent; it is a question of form. The list can be read again, queried, compared against a prior state, handed to someone who was not present when it was composed. The performance cannot.

This is also what resolves the counter that modern GUI applications have audit logs — that CloudTrail records what was clicked in the AWS Console, that Figma keeps version histories. A log records a performance. A script composes a performance. They are different kinds of artifacts: a log of what was clicked cannot be re-executed as a new configuration; a Terraform file can be. Goody's distinction between the list as cognitive form and the oral utterance as performance does this work at the level of structure, not technology preference.

Clanchy names the pattern. The structural mapping between CLI/IaC practice and literate knowledge architecture has been documented, in one of the most meticulously studied transitions in administrative history, by Michael Clanchy's From Memory to Written Record: England 1066–1307​.

Clanchy documents the specific mechanisms by which oral administration — witnessed oral oaths, memory, the testimony of those present — was superseded by written instruments over two and a half centuries of English governance. The correspondence between that transition and computing's current CLI/IaC transition is structural, not decorative:

The oral oath required all parties to be present; it was unreplayable without them. The click session requires the person who clicked; it cannot be reproduced without their memory. The written charter was persistent, transmissible, inspectable without the original drafter. By Edward I's reign, Clanchy documents, "royal or seignorial writs reached every bailiff and village in England, making writing familiar throughout the countryside."¹⁰ The shell script is persistent, version-controlled, auditable without the engineer who wrote it.

The Domesday Book — William the Conqueror's 1086 written survey of England's landholdings — made property rights inspectable without requiring the original witnesses. Infrastructure-as-code makes cloud state inspectable without requiring the engineer who provisioned it. The written record, in both cases, is not a mere transcription; it is a governance mechanism.

The quo warranto proceedings of Edward I's reign formalized what the practice of written administration had been making operationally true for a century: "Memory, whether individual or collective, if unsupported by clear written evidence, was ruled out of court."¹⁰ An oral account of what privileges one held — without a written warrant — was legally insufficient. The principle established was that written record supersedes oral memory as the authoritative account of what was done.¹⁰ Today's compliance audit demands IaC state for the same reason: an oral account of which infrastructure was provisioned is legally and operationally insufficient absent a written record of it.

And Clanchy's most precise finding for computing's case: "Lay literacy grew out of bureaucracy, rather than from any abstract desire for education or literature."¹⁰ CLI adoption has followed the same mechanism: not mass training programs, but the expanding administrative requirement to produce written artifacts — DevOps practice that demands version-controlled infrastructure, cloud governance that requires inspectable state, compliance frameworks that treat click sessions as insufficient.

Clanchy also documents the gradient's internal structure in medieval England with a precision that maps directly onto computing's. The medieval literate world was not binary — scribes on one side, illiterates on the other. It was tripartite: the professionally literate scribe who could produce documents, the pragmatic reader who used documents without producing them, and the fully illiterate who nonetheless participated in written culture through surrogates. "Those who used writing participated in literacy, even if they had not mastered the skills of a clerk."¹¹ The Excel formula writer participates in the literate computing world even if they have not mastered the shell. The VBA macro author is a pragmatic reader. The gradient has always had multiple named bands — and each band has always been partly invisible to the bands below it.

The structural template is the same: the expansion of the population that needs to produce written artifacts, driven not by education but by the practical requirement of administration. Eisenstein argued that the printing press enabled "typographical fixity" — a mechanism for the systematic accumulation and correction of knowledge that scribal culture could not provide.¹² The Clanchy anchor does not depend on whether that thesis survives the Johns critique: Clanchy's argument is about administrative practice, not mechanical fixity theory.

A Recurrence, Not a Paradigm Shift#

Computing is going through what England went through between 1066 and 1307. This is not a paradigm shift. It is a structural recurrence of a pattern that has run before — a slow, contested, need-driven expansion of the population that has to produce written artifacts, with the gradient expanding not through universal instruction but through administrative necessity.

The recurrence frame matters because it disciplines the claim. Three production-side democratizations have run within living memory, and each illustrates the pattern and its limits.

In 1900, Kodak released the Brownie camera at one dollar. Photography had required bulky equipment, glass plates, toxic chemical development, and specialist training. The Brownie absorbed the specialist's work into the appliance: Kodak processed the film at its own facilities; the customer pressed the shutter. More than 150,000 Brownie cameras shipped in the first year.¹³ The snapshot came to the masses. Professional photography did not disappear. Cinematography remained scarce. The threshold for one finite skill dropped to mass-market price. The ceiling of what the new population could do with that skill remained finite.

In 1985, Aldus PageMaker was announced at the Apple stockholders meeting on January 23 and shipped its first version in July of that year.¹⁴ Before PageMaker, professional-quality documents required typesetters with specialized equipment and training. After PageMaker, an office worker with a Mac and a LaserWriter could produce typeset-quality layouts. Paul Brainerd coined the term "desktop publishing." Steve Jobs later said desktop publishing had saved the Macintosh.¹⁴ The typesetting industry contracted significantly over the following decade. The threshold for producing formatted documents dropped to any Mac user. The ceiling of what that population could do with PageMaker — layouts, yes; typesetting software, no — remained finite.

In 2004, Apple launched GarageBand, free with every Mac, explicitly to "democratize music-making." By 2023, non-major labels and independent distributors — a category that includes independent labels at 40.8 percent and artist-direct distributors at 5.9 percent, not solely individual independent artists — commanded 46.7 percent of the global recorded music market by ownership basis, with revenues of $14.3 billion.¹⁵ The home studio threshold dropped to zero for Mac users. The threshold for composing at the level of Logic Pro did not.

Each democratization has the same structure. A finite skill, previously requiring specialist knowledge, is absorbed into an appliance. The non-specialist gains access to a specific finite ceiling of that skill. The specialist-level practice — cinematography, typesetting software development, professional music production at the highest register — remains scarce.

Eisenstein argued that the printing press operated differently — not as a production-side democratization but as a consumption-side one.¹² Print lowered the threshold for reading​, not for writing. It made reading worth learning because there were now things worth reading at affordable prices. Writing remained scarce after print; even a published author in 1600 was likely one of a handful in their parish who could draft a letter. The press did not teach illiterates to write. It ended the scribal monopoly on reading, not on writing.

The CLI/GUI gradient is structurally different from any of these precedents. The CLI is not a finite-purpose tool with a fixed ceiling. It is the operator's interface to a general-purpose machine — which is why the gradient has so many bands, why those bands keep extending upward, and why the question of what counts as the "top" remains genuinely open. The Brownie's ceiling was cinematography. The CLI's ceiling is whatever the systems author is currently building above the script writer.

When the Gradient Matters: The Organizational Face and the Scope Conditions#

The structural argument has an organizational face that engineers in CLI-literate positions have been living inside, without quite having the vocabulary for its structural cause.

Teams staffed predominantly by appliance-end users — engineers managing cloud infrastructure through the AWS Console, organizations whose institutional knowledge lives in practitioner performance rather than in inspectable artifacts — are running oral-culture knowledge architectures at contemporary scale. The pathologies have names in practitioner discourse: ClickOps, tribal knowledge, orphaned resources, the engineer who left and took the runbook with them. What the practitioner vocabulary does not supply is the structural cause of these failures.

The structural cause is that knowledge generated in oral practice is homeostatic. Ong observed that "oral societies live very much in a present which keeps itself in equilibrium or homeostasis by sloughing off memories which no longer have present relevance."¹⁶ Applied at cohort scale — the 3-to-5-year cycles through which team membership turns over — this mechanism describes what happens to organizations whose operational knowledge exists only in practitioner performance. The knowledge that nobody currently rehearses is lost when the cohort who held it moves on. At the individual departure scale, a different mechanism applies: this is the failure mode that the tacit-knowledge tradition, from Michael Polanyi's The Tacit Dimension to Nonaka and Takeuchi's The Knowledge-Creating Company​, describes — embodied knowledge that does not survive individual departure because it was never externalized into an inspectable artifact. Both mechanisms, operating at different scales, produce the same kind of organizational outcome: knowledge that cannot be retrieved because it was never written down.

The organizational survey evidence is consistent with this structural diagnosis. Industry data from 2025 estimates that 21 percent of enterprise cloud infrastructure spend — equivalent to approximately $44.5 billion against Gartner's cloud spending forecast for the year — is wasted on underutilized, idle, or orphaned resources.¹⁷ In the same survey, fewer than half of organizations reported having real-time data on idle resources (43 percent), unused or orphaned resources (39 percent), or over-provisioned workloads (33 percent).¹⁷ Only 32 percent of developers reported having fully automated cost-saving practices.¹⁷

The Domesday Book made property rights inspectable; Infrastructure-as-Code makes cloud state inspectable in the same structural sense — the written record enables consultation without requiring the presence of whoever created it.¹⁰ The terraform plan command and drift detection exist precisely because the IaC community has named the map-versus-territory problem (the written configuration may not match the deployed state) and built mechanisms to surface the gap.

But the gradient is a diagnostic instrument, not a universal verdict. The scope conditions matter.

The gradient matters when the same task is done many times — when automation has positive return on investment. It matters when the work needs to be inspected, audited, or transferred to someone who was not present when it was performed. It matters when the worker's situation will change — when the literate skill is transferable in a way the performance is not. It matters when the scale exceeds the operator's sustained attention — when the gap between what the human can track manually and what the system requires has grown past what liveware can bridge.

The gradient matters much less when the task is genuinely one-off. It matters less when the worker's domain specialization pays better than the automation savings — when the accountant's time is worth more as accounting than as VBA. It matters less when the rational response is delegation to a specialist — the merchant/scribe model, which was economically efficient rather than a failure of literacy aspiration. The medieval merchant who dictated contracts to a scribe was specializing, not failing. The scribe was specializing too. Division of labor creates wealth; the gradient is not a verdict against specialization.

In the situations where the gradient matters — repetition, audit, transfer, scale — the oral-culture knowledge architecture produces predictable organizational failures, and those failures have a structural cause the practitioner vocabulary names but does not explain. What the oral/literate framework supplies is the explanation.

The Open Question: Brownie-Like, Press-Like, or Genuinely New#

The gradient has survived every prior threshold-lowering event. Photography's democratization did not dissolve the distinction between a snapshot and a cinematography practice. Desktop publishing's democratization did not dissolve the distinction between a formatted newsletter and a typesetting program. Home recording's democratization did not dissolve the distinction between a GarageBand track and a Logic Pro mastering session.

Two distinct LLM agent architectures have now emerged, and the distinction between them maps directly onto the oral/literate distinction between accumulated artifact and ephemeral performance.

Browser-driving agents — Browser-Use (the Python library, with approximately 78,000 GitHub stars as of April 2026) and Skyvern (Y Combinator-backed, "computer vision plus LLMs") — interact with interfaces via screenshots and clicks.¹⁸ They preserve the oral-clicky pattern at the agent layer. The agent navigates the GUI the way a practitioner navigates the GUI: it sees a screen, identifies clickable elements, issues clicks, and moves forward. The procedure the agent executes does not accumulate as an inspectable artifact. It is a performance.

Script-generating agents — Claude Code (Anthropic's terminal-based agentic coding tool, which plans and executes sequences of actions including file creation, multi-file refactoring, test execution, and git operations) and GitHub Copilot CLI (which brings AI-powered coding assistance directly to the command line) — generate executable, inspectable scripts.¹⁸ The procedure the agent produces is a written artifact: version-controllable, auditable, reproducible without the agent's presence for re-execution.

Skyvern's own evolution is instructive. The company's blog from October 2025 documents the reasoning: "If Skyvern could compile its reasoning into code and run that instead of keeping an LLM in the loop, automations would become faster, cheaper, and more reliable."¹⁸ When reproducibility and inspectability matter — when the oral-pattern agent's output needs to be rerun reliably — the agent wants to write a script. The structural distinction persists even inside a product that began as a browser-driver.

The gradient has survived the LLM transition. The oral/literate distinction is being reproduced at the agent layer with the same structural properties — replayability, inspectability, transmissibility without the original author — that distinguished the written charter from the oral oath in Clanchy's England.

Whether the LLM moment is Brownie-like or genuinely new remains the open question.

If the LLM moment is Brownie-like, it democratizes a finite slice of computing practice. Non-coders write small one-off scripts — the way the Brownie user took snapshots — and the systems-author band remains as distant and scarcely populated as it is today. The gradient gains a band: the prompt engineer as the new pragmatic reader, able to compose what the appliance user cannot, unable to architect what the systems author can. The recurrence runs again.

If the LLM moment is genuinely new — the first time in history that the production threshold for general-purpose programmable instruction has been lowered for a population that did not previously cross it — then the gradient is being remade. Not populated with a new band. Remade.

The LLM moment is the only candidate in this sequence that might dissolve a gradient rather than merely populate it — because it lowers a threshold for producing general-purpose instruction, not a threshold for consuming a finite-ceiling skill. Whether it does is not yet visible. The bifurcation at the agent layer is the leading indicator: each new agent that ships either preserves the gradient or extends it. The browser-driver preserves the oral pattern. The script-generator preserves the literate pattern. Each architectural choice is a data point in the question's resolution.

The press ended the scribal monopoly on reading; writing stayed scarce. The Brownie ended the darkroom's monopoly on snapshots; cinematography stayed scarce. Whether the LLM moment ends the scribal monopoly on the general-purpose machine — or merely adds a new band to an old gradient — is not yet visible. The literate band is the band best positioned to watch the answer emerge.

References#

1. DataReportal / ITU, Internet Use Statistics 2024–2025​; ITU Facts and Figures 2025 (6 billion users, 74% of global population, confirmed via Xinhua/EurekAlert). Evans Data Corporation, Worldwide Developer Population Report 2025 (27 million). SlashData, Global Developer Population and Demographic Study 2025 (47.2 million). ↩︎ ↩︎

2. Christopher Scaffidi, Mary Shaw, and Brad A. Myers, "Estimating the Numbers of End Users and End User Programmers," IEEE Symposium on Visual Languages and Human-Centric Computing (VL/HCC), Dallas, TX, September 21–24, 2005, pp. 207–214. DOI: 10.1109/VLHCC.2005.16. Most Influential Paper award 2014 confirmed via Oregon State University EECS News (blogs.oregonstate.edu/eecsnews/2014/08/04/chris-scaffidis-first-paper-named-influential/). ↩︎ ↩︎ ↩︎

3. Gartner, "Citizen developers will soon outnumber professional coders 4 to 1," reported in VentureBeat, October 22, 2021, attributing prediction to Jason Wong, Distinguished VP Analyst, at Gartner IT Symposium 2021. Updated 4:1 timeline through 2026 confirmed via Gartner-citing secondary sources. ↩︎ ↩︎

4. National Skills Coalition / Federal Reserve Bank of Atlanta, Closing the Digital Skill Divide​, February 2023. Lightcast, Digital Skills Outlook 2024 (Microsoft Office #1 demanded digital skill across 15 countries). ↩︎ ↩︎

5. National Center for Education Statistics (NCES), Fast Facts: Adult skills in an international context (PIAAC), 2012–15 and 2017 survey cycles. nces.ed.gov/fastfacts/display.asp?id=683. ↩︎

6. Andrew J. Ko, Robin Abraham, Laura Beckwith, Alan Blackwell, Margaret Burnett, Martin Erwig, Christopher Scaffidi, Joseph Lawrance, Henry Lieberman, Brad A. Myers, Mary Beth Rosson, Gregg Rothermel, Mary Shaw, and Susan Wiedenbeck, "The State of the Art in End-User Software Engineering," ACM Computing Surveys 43, no. 3 (April 2011), Article 21. DOI: 10.1145/1922649.1922658. ↩︎

7. Stack Overflow, 2024 Developer Survey​, Technology section (65,437 total respondents, 185 countries). survey.stackoverflow.co/2024/technology. Bash/Shell: 33.9% of all respondents; 34.2% of professional developers. ↩︎

8. Jack Goody, The Domestication of the Savage Mind (Cambridge: Cambridge University Press, 1977), pp. 80–81. Quote: "In particular, the list rarely appears in oral discourse" (p. 80); list properties (p. 81). ↩︎ ↩︎ ↩︎

9. Walter J. Ong, Orality and Literacy: The Technologizing of the Word (London: Methuen, 1982), Chapter 3 ("Some Psychodynamics of Orality"), pp. 37–49. ↩︎ ↩︎

10. Michael T. Clanchy, From Memory to Written Record: England 1066–1307​, 3rd ed. (Wiley-Blackwell, 2013). Passages cited: "royal or seignorial writs reached every bailiff and village in England"; "Memory, whether individual or collective, if unsupported by clear written evidence, was ruled out of court"; "Lay literacy grew out of bureaucracy, rather than from any abstract desire for education or literature"; "The demands of the royal Exchequer and courts of law compelled knights in the shires and burgesses in the towns to create lesser bureaucracies of their own." Quo warranto note: cases were suspended in the 1290s with acknowledgment that "tenure from time out of mind" remained admissible; cited here as the principle being established, not oral memory finally and fully displaced. ↩︎ ↩︎ ↩︎ ↩︎ ↩︎

11. Clanchy, From Memory to Written Record​, 3rd ed. Quote: "Those who used writing participated in literacy, even if they had not mastered the skills of a clerk." ↩︎

12. Elizabeth L. Eisenstein, The Printing Press as an Agent of Change: Communications and Cultural Transformations in Early-Modern Europe (Cambridge: Cambridge University Press, 1980), 2 vols. Passages: "Typographical fixity is a basic prerequisite for the rapid advancement of learning" (p. 113); "scribal culture was so thin that heavy reliance was placed on oral transmission" (p. 111). Contested by: Adrian Johns, The Nature of the Book: Print and Knowledge in the Making (University of Chicago Press, 1998); and the AHR Forum: Elizabeth L. Eisenstein, "An Unacknowledged Revolution Revisited," American Historical Review 107, no. 1 (February 2002): 87–105; Adrian Johns, "How to Acknowledge a Revolution," American Historical Review 107, no. 1 (February 2002): 106–125. The "typographical fixity" thesis is presented here as Eisenstein's argued position; the core claim that print enabled greater standardization than scribal culture is not falsified by Johns, but the specific mechanism is contested. ↩︎ ↩︎

13. Artisan Obscura (2024); JSTOR Daily, "How the Brownie Camera Made Everyone a Photographer"; Franklin Institute Kodak Brownie collection record; EBSCO Research Starters (history, Kodak Brownie). Launch: February 1900; price: $1; first-year sales: 150,000+. Note: the Kodak slogan "You press the button, we do the rest" was George Eastman's 1888 tagline for the original Kodak No. 1 camera, not the 1900 Brownie; the Brownie embodied the same principle at a mass-market price point. ↩︎

14. GeekWire, "PageMaker pioneer Paul Brainerd, 1947–2026," February 2026; Wikipedia, "Adobe PageMaker"; Computer History Museum, "Desktop Publishing — CHM Revolution"; IEEE Annals of the History of Computing​. Announcement date: January 23, 1985 (Apple stockholders meeting); ship date: PageMaker 1.0, July 1985. Steve Jobs quote confirmed via GeekWire Brainerd obituary. ↩︎ ↩︎

15. Rolling Stone, "Inside GarageBand, the Little App Ruling the Sound of Modern Music" (2019); MIDiA Research, October 2024 (via Music Ally, Billboard, Hypebot). Launch: January 6, 2004, Macworld. The 46.7% figure refers to non-major labels (40.8%) and artist-direct distributors such as TuneCore and Ditto Music (5.9%) on ownership basis, as reported by MIDiA Research; $14.3 billion in revenues (2023). ↩︎

16. Walter J. Ong, Orality and Literacy (1982), Chapter 3, approximately p. 46. ↩︎

17. Harness, "FinOps in Focus 2025," press release via PR Newswire, February 2025. Survey: 700 respondents (350 developers, 350 engineering leaders), US and UK, organizations with 1,000+ employees, conducted by Coleman Parkes Research, November–December 2024. $44.5B = 21% waste rate applied to Gartner's 2025 worldwide public cloud end-user spending forecast ($211.86B); this is a model output, not a measured figure. ↩︎ ↩︎ ↩︎

18. browser-use GitHub repository (browser-use/browser-use), approximately 78,000 GitHub stars as of April 2026; Skyvern Y Combinator listing (YC S23); Claude Code: Anthropic primary documentation (anthropic.com, github.com/anthropics/claude-code); GitHub Copilot CLI: docs.github.com. Skyvern evolution quote: "If Skyvern could compile its reasoning into code and run that instead of keeping an LLM in the loop, automations would become faster, cheaper, and more reliable" — Skyvern blog, "Asking AI to build scrapers should be easy right?", October 17, 2025. ↩︎ ↩︎ ↩︎

Further Reading#

• Neal Stephenson. "In the Beginning Was the Command Line." 1999. Available at garote.bdmonkeys.net/commandline/index.html. — The canonical prior text on CLI versus GUI philosophy; addressed to the reader who already uses the CLI; does not ask the population-level question this article asks, which makes it the most useful companion and the most useful contrast.
• Michael Polanyi. The Tacit Dimension​. Anchor Books, 1966. — The foundational text on embodied knowledge that does not survive individual departure; supplies the team-scale tacit-knowledge mechanism referenced in §5 (as distinct from Ong's cohort-scale homeostatic mechanism). The two frameworks operate at different scales and address different aspects of organizational knowledge loss.
• Ikujiro Nonaka and Hirotaka Takeuchi. The Knowledge-Creating Company: How Japanese Companies Create the Dynamics of Innovation​. Oxford University Press, 1995, Chapters 3–5. — Extends Polanyi's tacit-knowledge tradition to organizational knowledge creation; the SECI model (Socialization, Externalization, Combination, Internalization) provides the framework for understanding how tacit knowledge becomes explicit institutional knowledge — and why organizations that skip the Externalization step lose what their practitioners knew.
• Eric A. Havelock. Preface to Plato​. Harvard University Press, 1963. — The precursor to Ong's framework; examines the cognitive consequences of the shift from oral to written culture in ancient Greece; provides the deepest grounding for the claim that writing restructures thought rather than merely recording it.
• Alan Kay. "A Personal Computer for Children of All Ages." Proceedings of the ACM Annual Conference​, 1972. — The founding document of the general-purpose personal computer vision; articulates why the distinction between finite-purpose appliance and general-purpose machine is the article's structural subject, not merely its occasion.