Concise Summary简洁概述
AI didn't replace jobs one-to-one, but it inflated token costs and code output faster than revenue or coordination could keep up — making layoffs the fastest lever to fix the resulting unit economics.
The author frames this wave of layoffs as offsetting AI spend and cutting an 'alignment tax' between teams, both short-term fixes that don't require AI to have actually taken anyone's job.
AI 并没有一对一替代岗位，但它让 Token 成本和代码产出的增速远远超过收入与组织对齐速度，于是裁员成了修复单位经济效益最快的杠杆。
作者认为这轮裁员本质是在抵消 AI 支出、削减团队间的「对齐税」——两者都是短期止血手段，都不需要 AI 真的抢走谁的饭碗。
Infographic信息图
Token spend as fixed cost
Token 支出变成硬成本
Unlike outcome-based SaaS pricing (a % of sales lift), Claude's usage-based pricing charges per token regardless of whether the generated code ships, breaks, or gets thrown away — so a 5x jump in usage doesn't guarantee any revenue gain, just a guaranteed cost spike.
与按业绩抽成的 SaaS 定价不同，Claude 按 Token 消耗计费——无论生成的代码是否上线、报废还是引发事故都要付费。用量暴涨 5 倍不代表收入也涨 5 倍，却一定意味着成本先涨了 5 倍。
Input/Output/Outcome ladder
投入/产出/成果三层模型
Borrowing a McKinsey-style framework, the author separates code (input) from features (output) from paying users (outcome) — AI massively cheapens the first rung without moving the third, exposing the gap between activity and value.
作者借用类似麦肯锡的三层框架——代码是投入，功能是产出，用户愿意付费才是成果。AI 大幅拉低了第一层的成本，却没有拉动第三层，暴露出「活跃」与「价值」之间的落差。
Friction used to be a filter
摩擦力曾是过滤器
When coding was slow and expensive, teams were forced to debate and kill weak ideas before wasting resources; cheap code removes that forcing function, so more code gets written without more good ideas actually being selected.
过去写代码又慢又贵，这种「摩擦」逼着团队提前争论、淘汰烂点子；如今代码近乎免费，这层过滤消失了，结果是代码量暴增，却没有更多真正靠谱的想法被筛选出来。
Layoffs as an alignment shortcut
裁员是对齐问题的捷径
Because large orgs accumulate redundant headcount that isn't strictly needed for survival, cutting 10-20% removes duplicate teams and cross-team blocking almost overnight — a faster fix than actually solving coordination.
大型组织本就积累了远超「生存所需」的冗余人力，裁掉 10%-20% 能迅速消灭重复团队和相互卡脖子的局面——比真正解决组织对齐问题快得多，尽管治标不治本。
Detailed Summary详细解读
The piece opens from a place of personal risk — an engineer with a 10% chance of being on an 8000-person layoff list — but deliberately strips out the emotional angle to make a structural argument. Rather than litigating whether AI 'caused' the Square-style layoff wave or whether companies are simply using AI as cover for unrelated failures, the author sidesteps the citation war entirely and instead builds a mechanism-level case for why both explanations can be true at once.
The core evidence is a mismatch: token usage, code volume, and PR counts have exploded 2-5x at heavy-adoption companies like Uber and Shopify, yet revenue and shipped user-facing features haven't grown proportionally. Critics like Ed Zitron and Gary Marcus read this as proof AI isn't productive. The author instead reaches for a McKinsey-style input/output/outcome ladder: code is input, features are output, and only paying users constitute outcome — so a spike in the first rung says nothing about the third.
He contrasts Claude's usage-based token pricing with outcome-based SaaS pricing (e.g., a cut of sales lift). Outcome pricing aligns vendor incentives with results and costs nothing if it fails; token pricing charges in full for code that's discarded, rolled back after a SEV, or built purely for internal cosmetics. Scaling input 5x overnight doesn't reliably scale output or outcome the way it used to at smaller increments — the input-to-outcome relationship breaks down at extreme volume.
The essay then asks why cheap code hasn't unlocked the 8 backlog ideas companies used to shelve for lack of engineering time. Two answers emerge: most of those ideas were never good — friction from slow coding used to force teams to argue and kill weak proposals before they consumed resources, a filter that's now gone; and cross-team alignment has become unbearable because any team can silently build a rival MVP overnight instead of negotiating.
From there the argument lands on its central thesis: layoffs solve two concrete short-term problems even without AI replacing anyone. First, offsetting ballooning token spend against flat revenue — a $30k/year Claude habit per engineer equals a full India SDE salary, half a European one, or a quarter of a US one, forcing headcount cuts of 20-50% just to hold payroll flat. Second, cutting the 'alignment tax': large orgs accumulate redundant headcount as organizational debt, and removing 10-20% of staff eliminates the duplicate teams blocking each other, speeding execution in the short run even though it doesn't fix the underlying bloat long-term.
The piece closes by acknowledging the 'AI-washing' critique has merit — layoff emails across companies reuse suspiciously identical buzzwords like 'AI-native,' 'managers who code,' and 'flatter org' — while insisting the causal chain still runs through AI: token spend must be offset, and coordination speed must catch up to code-generation speed. Until firms learn to convert AI input into revenue outcome, layoffs remain the only lever, closing on the author's own unresolved 15-day wait.
文章从作者本人的处境切入——一名有 10% 概率上榜、身处 8000 人裁员名单中的工程师——但刻意剥离情绪，转向结构性论证。他没有纠缠于「AI 是否真的引发了这波裁员潮」或「企业只是拿 AI 当遮羞布」这类各执一词的争论，而是绕开引用大战，直接从机制层面论证：这两种解释其实可以同时成立。
核心证据是一种错配：在 Uber、Shopify 这类重度使用 AI 的公司，Token 用量、代码量和 PR 数量都暴涨了 2-5 倍，但收入和用户可感知的新功能并没有同比例增长。Ed Zitron、Gary Marcus 等评论者据此认为 AI 并不真正提高生产力。作者则借用类似麦肯锡的「投入-产出-成果」三层框架反驳：代码是投入，功能是产出，只有愿意付费的用户才算成果——第一层暴涨，不代表第三层也会跟着涨。
他对比了 Claude 的按量计费与「成果抽成」式 SaaS 定价（例如按销售额提升抽成）。成果定价让供应商利益与结果绑定，失败了客户不用付钱；而 Token 定价对被废弃、因严重事故而回滚、或只是为了美化内部仪表盘的代码，照样全额收费。当投入在一夜之间放大 5 倍时，投入-产出-成果这条因果链在过去小幅增长时还成立，到了这种极端量级就会失灵。
文章接着追问：既然代码几乎免费，为什么过去因人手不足被搁置的那 8 个想法依然没被做出来？答案有二：其一，那些想法本来就大多不靠谱——过去写代码慢，这种摩擦逼着团队提前争论、淘汰烂点子，而这层过滤如今消失了；其二，跨团队对齐变得难以忍受，因为任何团队都能连夜悄悄做出一个对立版本的 MVP，而不是坐下来谈判达成一致假设。
论证由此推向核心论点：即便 AI 没有真正取代任何人，裁员依然能立竿见影地解决两个短期问题。第一，抵消不断膨胀却未带来收入增长的 Token 支出——每位工程师每年约 3 万美元的 Claude 花费，相当于一名印度 SDE 的全部薪资、半名欧洲 SDE 或四分之一名美国 SDE，为维持工资支出不变，就必须裁掉 20%-50% 的人。第二，削减「对齐税」：大型组织本就积累了大量组织性冗余，裁掉 10%-20% 的人能迅速消灭互相卡脖子的重复团队，短期内提速，尽管长期并未真正解决冗余积累的根本问题。
结尾承认「AI 洗白」的批评有其道理——各家公司的裁员邮件反复出现「AI 原生」「写代码的管理者」「扁平化架构」等高度雷同的措辞，像是抄同一份提示词——但仍坚持因果链条终究绕不开 AI：Token 支出必须被抵消，组织对齐速度必须追上代码生成速度。在企业学会把 AI 投入转化为收入成果之前，裁员将是唯一的解药，文章最终收束回作者本人尚未揭晓的 15 天倒计时。
FAQ常见问答
Is the author claiming AI directly replaced 30% of jobs?作者是否认为 AI 直接替代了 30% 的岗位？
No — he explicitly rejects one-to-one replacement, noting AI beats junior white-collar work on some tasks but underperforms on others, making it a poor drop-in substitute for a fixed share of headcount.
没有——他明确否认一对一替代的说法，指出 AI 在部分任务上强于初级白领，但在另一些任务上不如人类，并非能直接插拔替换固定比例员工的零件。
Why doesn't outcome-based pricing solve the token-cost problem?为什么按成果计费不能解决 Token 成本问题？
Because Claude and similar tools are sold as B2B SaaS priced on usage, not on the revenue they generate — so companies pay full price for discarded, rolled-back, or cosmetic code regardless of business impact.
因为 Claude 这类工具是按用量计费的 B2B SaaS，而非按其带来的收入抽成——无论代码是否被废弃、回滚或只是装饰性改动，企业都要照单全付。
What's the difference between 'AI layoffs' and 'AI-washing' in this piece?文中「AI 裁员」和「AI 洗白」有何区别？
AI-washing means using AI as PR cover for layoffs really driven by overhiring or weak business decisions; the author argues even those cases are still AI-caused at root, since token spend and alignment speed are the real triggers.
AI 洗白指企业拿 AI 当借口，掩盖真正由过度招聘或经营不善导致的裁员；作者认为即便如此，根源仍是 AI——真正的诱因是 Token 支出与对齐速度问题。
Why would firing people actually speed teams up rather than slow them down?裁员为何反而会让团队变快，而不是拖慢？
Because large orgs carry redundant headcount ('organizational fat'); removing overlapping teams eliminates the cross-team negotiation and duplicate MVP-building that was slowing everyone down, at least in the short run.
因为大型组织本就带有冗余人力（「组织脂肪」）；裁掉重叠团队能消除拖慢进度的跨团队谈判和重复 MVP 开发，至少短期内如此。
Does the piece offer evidence beyond anecdote for its cost figures?文章的成本数据是否有超出个人轶事的依据？
Mostly no — the $100/day/engineer token spend and the $70B industry revenue figure are used illustratively, not sourced from audited data, so treat the specific ratios as directional rather than precise.
基本没有——每位工程师每日 100 美元的 Token 支出、行业 700 亿美元营收等数字均为示意性引用，并非经审计的数据，具体比例应视为方向性参考而非精确统计。
In-depth Analysis · Pros & Cons深入解读 · 优缺点
A software engineer facing an 8000-person layoff list explains why AI didn't need to replace anyone's job to cause this wave of cuts. The piece reframes AI layoffs as a cash-flow and coordination-speed problem, not a one-to-one automation event.
一位可能身处 8000 人裁员名单中的软件工程师，解释了 AI 为何无需真正替代任何岗位，就足以引发这轮裁员潮。文章把「AI 裁员」重新定义为现金流问题与组织对齐速度问题，而非一对一的自动化替代事件。
- Clear causal mechanism因果机制清晰Instead of vague 'AI took jobs' rhetoric, the piece names two concrete levers — token cost offset and alignment-tax cutting — that explain layoffs without requiring literal automation.不同于笼统的「AI 抢走工作」说法，文章点明两个具体杠杆——抵消 Token 成本、削减对齐税——在无需真正自动化替代的前提下解释了裁员逻辑。
- Useful business framework引入实用商业框架The input/output/outcome ladder gives readers a reusable lens for evaluating any AI productivity claim, distinguishing raw activity from actual business value.投入-产出-成果三层框架为读者提供了评估任何 AI 生产力主张的通用视角，区分了单纯的活跃度与真正的商业价值。
- Honest about ambiguity坦承不确定性The author doesn't force a single explanation, explicitly allowing that AI-washing and genuine AI-driven cost pressure can coexist, which is more intellectually honest than partisan takes on either side.作者没有强行给出单一解释，明确承认「AI 洗白」与真实的 AI 成本压力可以并存，这比任何一方的极端立场都更诚实。
- Grounded in lived experience来自真实一线体验Observations about PR volume spikes, MVP turf wars, and per-engineer token spend read as credible insider detail rather than secondhand commentary, giving the argument texture beyond punditry.关于 PR 数量暴涨、MVP 地盘之争、人均 Token 支出的观察带有可信的一线细节，而非二手评论，为论证增添了超越空谈的质感。
- No hard data, only anecdote缺乏硬数据，只有轶事Figures like $100/engineer/day or 2-5x PR growth are illustrative, not sourced from disclosed company financials, so the quantitative backbone of the argument can't be independently verified.每位工程师每日 100 美元、PR 增长 2-5 倍等数字均为示意，并非来自公开披露的财报，论证的量化基础无法被独立核实。
- Assumes uniform SDE-flat org假设扁平化 SDE 组织The 20-50% headcount-cut math assumes a company where everyone is an SDE on similar pay — real orgs mix roles and geographies, so the arithmetic oversimplifies actual layoff sizing.20%-50% 裁员比例的算法假设公司里人人都是薪资相近的 SDE，但真实组织角色和地域构成复杂得多，这一算术明显简化了实际裁员规模的测算。
- Doesn't address AI quality/reliability risk未涉及 AI 质量与可靠性风险The piece treats code-cost inflation as the central problem but glosses over the deeper risk that AI-generated code accumulating faster than review capacity could degrade system reliability over time.文章把代码成本膨胀当作核心问题，却轻描淡写了更深层的风险——AI 生成代码的速度若长期超过评审能力，系统可靠性可能随时间恶化。
- Short-term frame, no long-term resolution只谈短期，未解长期问题The author admits layoffs don't fix organizational debt and predicts it recurs in two years, but doesn't explore what a durable solution to the alignment-tax problem would actually look like.作者承认裁员无法真正解决组织债务问题，并预测两年后老毛病复发，但并未深入探讨对齐税问题的长期解决方案究竟应是什么样。
Recommended for managers and engineers trying to make sense of AI-linked layoffs beyond headline panic — its input/output/outcome framework is genuinely useful, though its cost figures are illustrative anecdote, not audited data, so use it for the mental model, not the math.
适合管理者和工程师用来跳出「AI 恐慌式」标题党，理解裁员潮背后的真实逻辑——投入/产出/成果框架相当实用，但文中成本数字属于示意性轶事而非审计数据，读它取其思维模型，不必较真具体算术。
Excerpt原文节选
This is a short excerpt, not the full piece — the complete essay belongs to its original author; please read it in full at the link above.
以下仅为节选，并非全文——完整文章版权归原作者所有，请点击上方链接阅读全文。
The English text on this side is an AI translation provided for convenience; the authoritative version is the source in the other language.
Somewhere in our company's executive offices lies a layoff list with as many as 8,000 names on it. There's a 10% chance I'm on that list. In a few days, on May 20th, I'll know my fate.
Seeing Coinbase's "AI layoffs" announcement today, I decided to write this piece. I made a point of writing it before May 20th, because I want to share some of the most honest thoughts I have, without any personal emotion about whether I'll stay or go. These thoughts have nothing to do with whether I personally get laid off, and they're not limited to just my own company. They come from the genuine voices of my friends who work at various mid-size and large companies.
Right now there are plenty of articles arguing over whether this new wave of layoffs (generally believed to have started when Jack Dorsey cut 40% of Square's staff) is really caused by AI, or whether it's just "AI-washing" (companies using the…
[…the source continues — read the rest at the link above]
[……原文更长，完整内容请点击上方链接阅读]
Fifteen days left, and I'll know my fate. But regardless of the outcome, I think I already know the reason.
在我们公司的高层办公室里，某处正躺着一份多达 8000 人的裁员名单。我有 10% 的概率在这份名单上。再过几天，也就是 5 月 20 日，我就能知道自己的命运了。
看到今天 Coinbase 宣布的“AI 裁员”消息，我决定写下这篇文章。我特意赶在 5 月 20 日之前动笔，因为我想分享一些最真实的看法，不带任何“我是走是留”的个人情绪。这些想法不仅与我是否被裁无关，也不仅仅局限于我所在的公司。它们来自我那些在各大中型企业工作的朋友们的真实心声。
现在有大量的文章在争论：这新一波的裁员潮（大家普遍认为是从杰克·多西裁掉 Square 40% 员工开始的）到底是因为 AI 导致的，还是仅仅在搞“AI 洗白 (AI-washing)” （指企业借着拥抱 AI 的名义，来掩盖其他商业失败或裁员的真实目的） 。我不想在文章里塞满各种新闻和论文的链接来折磨你，这些内容你可能早就看过了，或者只需在谷歌搜一下、问问 ChatGPT 就能找到。
备受吹捧的“AI 生产力”与难以捉摸的证据
AI 真的让我们更高效了吗？这真是一个充满争议的重磅问题！如果我们反向思考一下，断言“AI 什么都没改变”，我想哪怕是那些最怀疑 AI 价值的人，也不会同意这种说法。尤其是在科技公司里，AI 使用量的火箭式飙升是摆在眼前的事实。即便是那些最保守、给 AI 预算设限、不给员工配备 AI 工具的公司，也同样不可否认有一部分工作实质上是 AI 完成的——哪怕员工只是苦哈哈地在谷歌或微软办公套件里，偷偷用 Gemini 或 Copilot 来编辑文档。
至于那些更有远见、一头扎进 AI token（Token）（AI 模型处理文本的基本单位，企业使用大语言模型时通常按消耗的 token 数量计费）海洋的公司，比如优步（Uber）或 Shopify（我这里不包括像 Meta 或微软这种自己开发大语言模型的公司，也不包括 Vercel 或 Cloudflare 这种积极搭建 AI 基础设施的公司；只说纯粹的“使用者”），他们的 AI 用量简直陷入了疯狂。
[…the source continues — read the rest at the link above]
[……原文更长，完整内容请点击上方链接阅读]
还有 15 天，我就能知晓自己的命运了。但不管结果如何，我想我已经知道了原因。哪怕当时坐在角落那间宽敞的 CEO 办公室里做决定的人是我，我也不知道自己能不能做得更好，说不定我也只会和其他拉群的 CEO 们一样，做出如出一辙的选择。