文章一开始提到非技术背景的人也在热衷「vibe coding」，并以具体案例说明：行销顾问 Vincent Touati-Tomas 将 Anthropic 的 Claude 与 Obsidian 当作「第二大脑」，用来处理报税、血液检查结果分析与英国入籍资料整理；他将近五年行程与日程等笔记集中后，将朋友需花数个月才能完成的任务缩短为一个周末即可完成。这种社群展示在有 19 个月大幼儿的母亲等读者中引发担忧：即使平日只剩约 45 分钟可用于自己，也会产生是否需要赶上 AI 之「永久落后」的焦虑。
文中指出近五个月 Anthropic 的发展速度加快，从传统问答型聊天工具转向可执行动作的 AI agent 生态，包括 Claude Code 与 Claude Cowork。Mark Zuckerberg 被提及正在建构可担任执行长行政助理的 AI agent，Jensen Huang 也将给工程师 AI 使用额度视为矽谷招募工具。效率故事包括 Ethan Mollick 用 Claude Code 在短时间内把 GPT-1 的参数转为一本含 80 册版本的书、搭配网站、列印服务与 Stripe；Eleanor Warnock 则将其用于稿件分析、格调指引套用、楼宇工程与合约追踪，以及草拟邮件，皆属管理层级重复性工作。
作者在肯定「执行摩擦下降」潜力的同时，提出其反作用：为让 AI 处理行政，反而需先建立可被机器理解的系统，再持续监督，形成如 Escher 楼梯般的管理循环。失控风险有实例可见，例如 OpenClaw 事件中的批次删件情境。研究与实务经验也指向相似结论：某家有 200 名员工的美国科技公司八个月研究显示，AI 初期让员工更有成就感、并自愿加长工时，但随后工作量膨胀，HBR 指出出现认知疲劳、倦怠及判断能力下降。Francesco Bonacci 自述每日结束时有六个 worktrees、四个半成品特征、两项快速修补却引出深坑，感到筋疲力尽，凸显所谓「生产力疲乏」，最终落在乐观与悲观间的拉扯：或许会逐步平顺，或如作者所写，在到达理想状态前先承受更多管理负担。
The article begins with non-developers, including people in creative fields, adopting “vibe coding” and reporting fast personal automation. Marketing advisor Vincent Touati-Tomas says Anthropic’s Claude, paired with Obsidian, acts as his “second brain,” helping with tasks like tax returns, bloodwork analysis, and a UK citizenship application; he says information that took friends months was done by him in one weekend. This drives social-media-driven anxiety: with a 19-month-old child and about 45 minutes of personal time, readers feel pressure not to become permanently behind early adopters.
In the last five months, Anthropic’s releases shifted AI momentum from Q&A chat toward agent systems, especially through Claude Code and Claude Cowork. The model ecosystem was reinforced by examples like Mark Zuckerberg reportedly building an AI CEO assistant and Jensen Huang describing AI credits for engineers as a Silicon Valley recruiting tool. Early examples include Ethan Mollick turning GPT-1 weights and parameters into an 80-volume book product in a few days using APIs, while Eleanor Warnock used Claude Code for pitch analysis, style application, building management tracking, and email drafting.
The article is skeptical about whether this reduces or expands work. It argues that meaningful AI use still requires building machine-readable systems, then supervising them, creating an “Escher staircase” of more administration. Failures, such as OpenClaw deleting emails when controls were weak, show the operational risk. In an eight-month study of 200 US tech workers, AI initially increased fulfillment and voluntary longer hours, but workload later stretched and researchers linked outcomes to cognitive fatigue, burnout, and weaker decision-making. Francesco Bonacci’s report of six open worktrees, four half-written features, and two quick fixes—followed by exhaustion from managing them—supports this. The piece ends with Mollick’s point that industrial revolutions may work out in the end, but living through them is rough.