# The layoffs will continue till we learn to use AI
**作者**: Arnav Gupta
**日期**: 2026-05-05T23:32:32.000Z
**来源**: [https://x.com/championswimmer/status/2051807284691612099](https://x.com/championswimmer/status/2051807284691612099)
---
Somewhere in the upper echelons of my company is a list of 8000 names. There is a 10% chance I am on it. I will get to know in a few days on 20 May.
Seeing today's "AI layoffs" announcement by Coinbase, made me consider writing this. Especially before 20 May, because I would love to share my thoughts on this, untainted, by the knowledge of whether I am on the list myself. These thoughts do not depend on whether I am on the list, not are they about (only) my workplace - it is what I am hearing from all my friends in various companies of various mid to large sizes.
A lot of ink has been poured debating whether these new wave of layoffs (largely regarded as to have started with Jack laying off 40% of Square) are really because of AI or are they merely "AI-washing". I'll spare you the trouble of peppering my article with links to all those articles, essays, news pieces but you either have read them already or they just a Google search ChatGPT query away.
## The oft-touted 'AI productivity' and its elusive proof
Does AI make us productive. Ah that loaded question! If we flip that for a second and just say "AI didn't change anything". I think no one, not even the biggest sceptics of AI's impact will agree to that. Especially in tech companies - the skyrocketing usage of AI is something you just cannot ignore. Even the most conservative of places which are putting AI spending caps, not giving AI tools to employees or what not - even there, undeniably some amount of work is being done by AI, even if it is as sad as just using Gemini or Copilot inside their Google or Microsoft Office suite to edit their docs.
In the more forward-looking companies that dove head first into the ocean of AI tokens, the Ubers, Shopifys of the world (I am not counting Metas or Microsofts - which is making their own model, or Vercels or Cloudflares - which is actively building AI infra; just the ones who are purely "users"), the usage has been crazy. 90-100% code being AI generated, to PRs/diffs per week going up 2-5x, to hundreds of millions of dollars in AI budget for the year being eaten up within months - we have seen it all.
And yet, ofcourse the Ed Zitrons, Will Manidis, Gary Marcus and the Michael Burys of the world will also counter you with the question - why have these companies not 2-5xed their revenue then? Why are their apps still almost exactly the same as it was 6 months back? If AI is really all that productive, what are they even producing with it? If they are generating 5x the code, and the end user doesn't even notice it, then what is the point of all that code? And that's a fair question.
## Input, Output, Outcome
We have to take a little Business Administration 101 detour. When your fast growing mid sized overfunded company throwing money at everything finally starts drying out of funds, and some elder CEO who you go to for advice, asks you to get somebody from McKinsey come and have a look at your state of affairs, they start their presentation with a bland white slide with 3 words written in the default Arial font. "Input, Output, Outcome"
They explain to you, what everybody knows, but loves to forget.
Code is an input.
Features are an output.
Users spending money on your product is an outcome.
AI (or at least Claude Enterprise) is a B2B SaaS product. You'll notice that SaaS products are priced and marketed in different ways. If the product directly changes the outcome, they simply take a cut from the outcome. Imagine the sales pitch "our tool closes sales leads 36% faster. try it out for the low fee of 5% of your sales value"
This is an instant sell. Most other variables unchanged, if you were closing 100 leads in 100 days, you now close them in 63 days, freeing up 36 more days to close (if my math is right) 57 more leads! So your sales potentially goes up by 57% You'll be happy to pay 5% of your sales commission to get 57% more revenue any day. And if you don't use the product, you are paying them $0 anyway.
As you would have predicted where I am going with this - pricing of Claude Code tokens aren't exactly like that. If your software engineers, who are addicted to Claude Code like crack cocaine (I just realised they're both abbreviated to 'cc'), generate 100M tokens a day, you are spending $100 per day, per engineer, on it.
Even if some of the code they generated was discarded because it didn't work
Even if some more of the code was reverted later because it caused a SEV
Even if yet another share of the code was for internal tools to make dashboards look more cute for the VPs to look at
Because, code is input. And while if the direction is right, more input usually tends to more output, and that tends to more outcome - all of that may or may not hold true when you overnight 5x the input. Your "direction" of input might well suddenly point to random places, and not towards the output or the outcome.
## What is blocking us!
Well, every time the CEO or the PM wanted to do 10 things, the team said they could do only the top 2. There is no time for the rest 8. The explanation? Well coding is not child's play. It takes time to code up complex, working software.
Hmmm.... but code is free now. Why are we not doing those other 8 things?
There are 2 answers, one the CEO & PM will not like, and one the middle management and seniors will not like.
1. All those 8 ideas were not actually.... any good?
Just because the CEO or the PM had 10 brainfarts, doesn't mean they were all actually going to lead to outcomes. Even 10 new features (outputs) does not guarantee users like all 10 of them and use your app more for it (outcome). In fact the friction of not having enough bandwidth to code, made people debate a lot more and kill bad ideas much sooner before they hogged too much resources, so you filtered for the top 2 much better. Now that writing code is fast and cheap and easy, there is no point even trying to debate the ideas. Even if you decide to push back against them, do you think it will stop the CEO or the PM from spinning up Claude themselves? Yeah so don't even bother trying.
2. It is a pain to get everyone "aligned"
We know it is. Getting stakeholders to align first "why" we are doing this, and then separately "what" exactly we are doing it, and then once all over again on "how" we are doing it is pain. The more the number of teams, the more the number of projects that get stuck in alignment hell. Writing code being the slow part was hiding this away. Now once the "what" is aligned, overnight someone builds an MVP, and schedules another meeting the very next day. In the meeting, you find out the other team also made an MVP. Yours and theirs work differently based on different assumptions.
Sure you can sit together and iron that out, and discuss which of whose assumptions are correct.
But let's be serious for a moment. You and your team armed with infinte Claude Code tokens are not going to do that. Nor will the other team. You will go right back into the arms of Claude and ask it to re-implement the other teams' part of the work in the way you think is best, and Claude will say "You're absolutely right" and get right to it!
## What will the layoffs solve?
Well you have been bearing with me telling you mostly obvious things so far. But I know you want me to get to the meat of the story. What will the layoffs achieve? If, as I posit, AI is not literally drop-in replacing 30% of the employees. (I think we can agree on this? Although it is better than an entry level white collar worker in many tasks and worse in others - it is not a drop-in replacement, definitely not 10 or 20 or 30% of your company)
The layoffs immensely help with 2 immediate short-term problems which are clear as day.
1. They offset "AI spending"
I mean, this is just cashflow 101. Surely, you can see that if all your Claude-addicted engineers are blowing up $100 per day on Claude (which is $2500 per month, or $30k per year), that is clearly worth 1 SDE salary in India, worth 0.5 SDE in EU and worth 0.25 SDE in USA.
If you just do the dumbest math possible, assuming every employee is an SDE in a flat org, then you need to remove 50%(India) or 33% (EU) or 20% (USA), to continue to meet the same wage bill, inclusive of token spends.
The very fact that AI usage is growing regardless, and revenue is not yet seeing this uptick, this has to happen otherwise the balance sheet of the company goes in disarray. Your entire unit economics of the SDLC goes for a toss - if you spend 50% more in input, with no or little change in outcome.
If we did learn to use AI though - and we figured out how 50% more input costs translate to 50% more revenue outcome, we would not need this to happen. But since you didn't learn to use AI, some of you need to leave to make space for Anthropic's salary.
2. Cut the 'alignment tax'
There is no arguing that any large company, is bigger than it "needs" to be to just survive. That's the whole point. Larger organisations carry organisational fat. That is by design. It is possible for anyone to leave the company, and systems to still continue to work because someone else knew what they did. In big companies, you can go on 6 months parental leaves and things you worked on still keep running. These are good things to have! But this is also proof that if some percentage of people were removed, things will not shut down immediately. In fact, maybe with few weeks of initial system shock, for the next few months, things will get faster!
Remember how the two teams above didn't agree to each other's approaches? Well if you just layoff one of those teams and asked the other one to pull a few all nighters and do their job instead - they have no one to align with. We don't know what happens long term (or as Keynes said - "in the long term we are all dead"), but in the short term, cutting off 10-20% people in a big org, only makes things faster.
Large organisations, over time, invariably build up slack, build up redundancies, and build up 'org debt' just like tech debt. It is the nature of big organisations, and cutting 10% people today doesn't prevent it from happening again in 2 years time. But when you see everyone saying they are generating 5x more diffs, but unable to ship because they are blocked by other teams, at least the most immediate solution does look like removing people so that there is fewer people to block each other.
## These are AI layoffs, even if AI is not replacing you
Is your employee id being replaced by a new instance of Claude running on VM? We know that is not what is happening.
That said, are there many different workflows in the company that were once done by someone hitting keyboard keys and mouse clicks on tools like VS Code, Figma, Canva, Google Docs, and today is basically someone else (who needed that work from you) just yelling a prompt into an LLM instead of bothering to ask you for it? That is true as well.
Are these layoffs "AI-washing"? By which we mean - are there fundamental problems with the company regardless of AI (overhiring, diminishing profits, competitor pressure, bad business decisions) and AI being used as an 'excuse' to lay people off? Well that's somewhat true too.
And you'll also notice that over a period if you collect all these "layoff emails" from CEOs you'll almost feel they are all in the same Whatsapp group writing these emails together. AI-native pods, managers writing code, more reports for managers, flat hierarchy, managing a team of agents, you'll read those exact same terms in all their emails. Almost as if they all gave GPT the same prompt.
But the truth is that these layoffs, even if they they are not because AI is replacing you you, and even if they are some form of AI-washing. These layoffs are still because of AI. And these layoffs will continue till we learn to use AI. Till we learn to convert AI-tokens into outcomes and not just input. Till we learn to re-align the speed of "alignment" with the new speed of coding. And till we figure out, beyond our 2 good and 8 stupid ideas, 10 more ideas that we can chase with our increased productivity.
Till we figure out how the GDP of the world actually grows because of AI, we have to offset the $70 B (combined OAI/Ant enterprise revenue) of annual token spend by cutting some salaries. And till we figure out how to unblock each other faster, we can always be removed from the org chart itself.
I'll know more about my own fate in 15 days. But either way, I think I know why. And even if I were in the corner office making the decisions, I don't even know if I would have or could have done any better, or just did what everyone else in the CEO WhatsApp group is doing.
## 相关链接
- [Arnav Gupta](https://x.com/championswimmer)
- [@championswimmer](https://x.com/championswimmer)
- [133K](https://x.com/championswimmer/status/2051807284691612099/analytics)
- [$30k](https://x.com/search?q=%2430k&src=cashtag_click)
- [Upgrade to Premium](https://x.com/i/premium_sign_up)
- [7:32 AM · May 6, 2026](https://x.com/championswimmer/status/2051807284691612099)
- [133.2K Views](https://x.com/championswimmer/status/2051807284691612099/analytics)
- [View quotes](https://x.com/championswimmer/status/2051807284691612099/quotes)
---
*导出时间: 2026/5/6 20:51:52*
---
## 中文翻译
# 裁员将持续，直到我们学会使用 AI
**作者**: Arnav Gupta
**日期**: 2026-05-05T23:32:32.000Z
**来源**: [https://x.com/championswimmer/status/2051807284691612099](https://x.com/championswimmer/status/2051807284691612099)
---
在我公司的高层某个地方，放着一份包含 8000 个名字的名单。我有 10% 的概率在上面。我将在几天后的 5 月 20 日得知结果。
看到今天 Coinbase 宣布的“AI 裁员”，让我考虑写下这篇文章。特别是在 5 月 20 日之前，因为我很乐意分享我的想法，而不受我自己是否在名单上这一认知的影响。这些想法并不取决于我是否在名单上，也不仅仅是关于我的工作场所——这是我从各行各业的中大型公司朋友那里听来的。
关于这一波新一轮裁员（普遍认为始于 Jack 解雇 Square 40% 的员工）是否真的因为 AI，还是仅仅是“AI 洗白”（AI-washing），人们已经泼洒了大量墨水进行辩论。我就不费劲给文章里塞满那些文章、散文和新闻报道的链接了，你要么已经读过了，要么只需用 ChatGPT 搜一下就能找到。
## 经常吹捧的“AI 生产力”及其难以捉摸的证明
AI 能让我们更有生产力吗。啊，这个充满诱导性的问题！如果我们反过来想一秒钟，直接说“AI 什么也没改变”。我认为没有人，甚至是 AI 影响的最大怀疑论者，会同意这一点。特别是在科技公司——AI 使用量的激增是你无法忽视的。即使是在那些最保守的地方，设置了 AI 支出上限，不给员工提供 AI 工具等等——即便在那里，不可否认也有一定量的工作是由 AI 完成的，哪怕只是像在 Google 或 Microsoft Office 套件里使用 Gemini 或 Copilot 来编辑文档那样悲哀。
在那些一头扎进 AI 代币海洋的更具前瞻性的公司，比如 Uber、Shopify（我不算 Meta 或微软——它们在开发自己的模型，也不算 Vercel 或 Cloudflare——它们在积极建设 AI 基础设施；只算那些纯粹的“用户”），其使用量已经到了疯狂的地步。90-100% 的代码由 AI 生成，每周的 PR/diffs（拉取请求/差异对比）增加了 2-5 倍，数亿美元的年度 AI 预算在几个月内就被用光——这些我们都见识过了。
当然，Ed Zitron、Will Manidis、Gary Marcus 和 Michael Burry 也会反驳你——既然如此，为什么这些公司的收入没有增长 2-5 倍呢？为什么他们的应用程序几乎和 6 个月前一模一样？如果 AI 真的那么有效率，他们到底用它生产了什么？如果他们生成了 5 倍的代码，而最终用户甚至没有注意到，那么所有这些代码的意义何在？这是一个公平的问题。
## 输入、输出、结果
我们需要稍微绕道去上一堂工商管理入门课（Business Administration 101）。当你这家发展迅速、资金过剩、到处撒钱的中型公司终于开始资金枯竭，你去向某位资历深厚的 CEO 寻求建议，他让你请麦肯锡的人来看看你的状况时，他们开始演示的幻灯片是一张平淡的白色幻灯片，上面用默认的 Arial 字体写着 3 个词。“输入、输出、结果”。
他们会向你解释一个大家都心知肚明，却喜欢忘记的道理。
代码是输入。
功能是输出。
用户在你的产品上花钱是结果。
AI（或者至少是 Claude 企业版）是一个 B2B SaaS 产品。你会注意到，SaaS 产品的定价和营销方式是不同的。如果产品直接改变结果，他们通常会直接从结果中抽成。想象一下这样的推销辞：“我们的工具能将销售线索的成交速度提高 36%。试用一下吧，费用仅为销售价值的 5%。”
这是一笔稳赚不赔的买卖。如果其他大部分变量不变，如果你以前在 100 天内成交 100 个线索，现在你可以在 63 天内完成，从而腾出 36 天来成交（如果我的数学没错）另外 57 个线索！所以你的销售额潜在增长了 57%。你会非常乐意每天支付销售额的 5% 来获得 57% 的收入增长。而且如果你不使用该产品，你无论如何也付给他们 0 美元。
正如你可能预料到的我想表达什么——Claude Code 代币的定价并非如此。如果你的软件工程师像吸食强效可卡因一样沉迷于 Claude Code（我刚意识到这两者都缩写为 'cc'），每天生成 1 亿个代币，那你每天在每个工程师身上就要花费 100 美元。
即使他们生成的部分代码因为不工作而被丢弃
即使更多的代码后来因为导致了 SEV（严重事故）而被回滚
即使还有一部分代码是为了让 VP 们看的内部工具，只是为了让仪表板看起来更可爱
因为，代码是输入。而且虽然如果方向正确，更多的输入通常会导致更多的输出，进而导致更多的结果——但当你在一夜之间将输入增加 5 倍时，所有这些可能都不成立。你的输入“方向”很可能突然指向随机的地方，而不是指向输出或结果。
## 是什么在阻碍我们！
好吧，以前每次 CEO 或 PM 想做 10 件事时，团队说他们只能做前 2 件。剩下的 8 件没时间做。解释是什么？嗯，写代码可不是儿戏。编写复杂、能用的软件需要时间。
嗯……但代码现在是免费的。为什么我们不做那另外 8 件事呢？
有两个答案，一个是 CEO 和 PM 不会喜欢的，一个是中层管理和资深员工不会喜欢的。
1. 这 8 个想法其实并不怎么样……
仅仅因为 CEO 或 PM 有 10 个脑洞，并不意味着它们真的都能带来结果。即使是 10 个新功能（输出）也不能保证用户喜欢所有 10 个并因此更多地使用你的应用（结果）。事实上，以前因为没有足够的带宽写代码而产生的摩擦，迫使人们进行更多的辩论，并在坏主意占用太多资源之前更早地扼杀它们，所以你们能更好地筛选出那前 2 个最好的想法。现在写代码变得快速、廉价且简单，去辩论这些想法甚至没有意义。即使你决定反对它们，你认为这能阻止 CEO 或 PM 自己启动 Claude 吗？所以，甚至别费劲去尝试了。
2. 让每个人都“对齐”简直痛苦至极
我们知道确实如此。让利益相关者先就“为什么”我们要做这件事达成一致，然后再就“做什么”达成一致，然后再就“怎么做”达成一致，这很痛苦。团队数量越多，陷入“对齐地狱”的项目就越多。写代码这个慢速环节以前掩盖了这一点。现在一旦“做什么”对齐了，有人一夜之间就做出了一个 MVP（最小可行性产品），并在第二天安排了另一场会议。在会议上，你发现另一个团队也做了一个 MVP。你们的和他们的基于不同的假设，工作方式不同。
当然你们可以坐下来解决这些问题，讨论谁的假设是正确的。
但我们要严肃一点。你和你拥有无限 Claude Code 代币的团队是不会去那样做的。另一个团队也不会。你们会直接回到 Claude 的怀抱，让它以你认为最好的方式重新实现另一个团队的工作，Claude 会说“你绝对正确”并立即开始工作！
## 裁员能解决什么？
好吧，你们一直耐心地听我说了一些显而易见的事情。但我知道你们想知道故事的核心。裁员能成就什么？如果，正如我假设的那样，AI 并不是字面上直接替换了 30% 的员工。（我想我们可以在这个点上达成一致？虽然在许多任务上它比初级白领员工要好，在其他任务上更差——它不是一个即插即用的替代品，肯定不是你公司 10%、20% 或 30% 的人）
裁员极大地帮助了两个明显的、极其迫切的短期问题。
1. 它们抵消了“AI 支出”
我的意思是，这只是现金流 101。当然，你能看出来，如果你所有沉迷于 Claude 的工程师每天在 Claude 上挥霍 100 美元（也就是每月 2500 美元，或每年 3 万美元），这显然值印度 1 个 SDE（软件工程师）的薪水，值欧盟 0.5 个 SDE，值美国 0.25 个 SDE。
如果你只做最简单的数学计算，假设每个员工都是扁平化组织中的 SDE，那么你需要裁员 50%（印度）或 33%（欧盟）或 20%（美国），才能继续满足同样的工资账单，包括代币支出。
AI 使用量无论如何都在增长，而收入尚未看到这种上升，这一事实表明这必须发生，否则公司的资产负债表就会陷入混乱。你整个 SDLC（软件开发生命周期）的单位经济学就乱套了——如果你在输入上多花了 50%，而结果却没有或几乎没有变化。
如果我们确实学会了使用 AI——并且我们弄清楚了如何将增加的 50% 输入成本转化为增加的 50% 收入结果，我们就不需要这样做了。但既然你们没有学会使用 AI，你们中的一些人必须离开，为 Anthropic 的薪水腾出空间。
2. 削减“对齐税”
毫无疑问，任何大公司的规模都超过了仅仅为了生存而“需要”的规模。这就是重点。大型组织都有组织冗余。这是设计使然。任何人离开公司，系统仍然可以继续运转，因为其他人知道他们所做的工作。在大公司，你可以休 6 个月的产假，而你负责的事情仍然继续运行。这些都是好事！但这同时也证明了，如果有一定比例的人员被裁掉，事情不会立即停摆。事实上，也许在最初几周的系统冲击之后，在接下来的几个月里，事情会变得更快！
还记得上面那两个团队不同意彼此的方法吗？好吧，如果你裁掉其中一个团队，并要求另一个团队加几个通宵班来代替他们做工作——他们就没有人需要对齐了。我们不知道长期会发生什么（或者正如凯恩斯所说——“从长远来看，我们都死了”），但在短期内，在一个大组织中裁掉 10-20% 的人，只会让事情变得更快。
大型组织随着时间的推移，不可避免地会积压松弛、冗余和“组织债务”，就像技术债务一样。这是大组织的本质，今天裁掉 10% 的人并不能防止它在 2 年后再次发生。但是当你看到每个人都说他们生成的 diffs 多了 5 倍，却无法发布，因为他们被其他团队阻碍时，至少最直接的解决方案看起来确实像是裁掉一些人，这样互相阻碍的人就变少了。
## 这些是 AI 裁员，即使 AI 没有取代你
你的员工 ID 是否被运行在 VM（虚拟机）上的新 Claude 实例取代了？我们知道这不是正在发生的事情。
话虽如此，公司里是否有很多不同的工作流程以前是由某人在 VS Code、Figma、Canva、Google Docs 等工具上敲击键盘和点击鼠标完成的，而今天基本上就是其他人（需要你完成那项工作的人）直接对着 LLM（大语言模型）大喊提示词，而不是费心去问你？这也是事实。
这些裁员是“AI 洗白”吗？我们的意思是——无论是否有 AI，公司是否存在根本性问题（过度招聘、利润下降、竞争对手压力、糟糕的商业决策），AI 只是裁员的“借口”？嗯，这在一定程度上也是真的。
你还会注意到，如果一段时间以来你收集所有 CEO 发来的这些“裁员邮件”，你几乎会觉得他们都在同一个 WhatsApp 群里一起写这些邮件。AI 原生的小组、管理者写代码、给管理者更多汇报、扁平化层级、管理代理团队，你会在他们所有人的邮件中读到这些完全相同的术语。几乎就像他们都给了 GPT 同样的提示词。
但事实是，这些裁员，即使不是因为 AI 正在取代你，即使它们是某种形式的 AI 洗白。这些裁员仍然是因为 AI。而且这些裁员将持续，直到我们学会使用 AI。直到我们学会将 AI 代币转化为结果，而不仅仅是输入。直到我们学会将“对齐”的速度与新的编码速度重新调整一致。直到我们在那 2 个好主意和 8 个蠢主意之外，想出 10 个更多可以用我们提高的生产力去追逐的想法。
直到我们弄清楚世界 GDP 到底是如何因为 AI 增长的，我们就必须通过削减一些薪水来抵消每年 700 亿美元（OpenAI/Anthropic 企业收入总和）的代币支出。直到我们弄清楚如何更快地互相解除阻碍，我们随时都可能从组织结构图中被移除。
15 天后我会对自己命运的结局了解更多。但无论如何，我想我知道原因。即使我是坐在角落办公室做决定的人，我也不知道我是不是会或能不能做得更好，还是只会做 CEO WhatsApp 群里其他人都在做的事情。
## 相关链接
- [Arnav Gupta](https://x.com/championswimmer)
- [@championswimmer](https://x.com/championswimmer)
- [133K](https://x.com/championswimmer/status/2051807284691612099/analytics)
- [$30k](https://x.com/search?q=%2430k&src=cashtag_click)
- [升级到 Premium](https://x.com/i/premium_sign_up)
- [7:32 AM · May 6, 2026](https://x.com/championswimmer/status/2051807284691612099)
- [133.2K Views](https://x.com/championswimmer/status/2051807284691612099/analytics)
- [查看引用](https://x.com/championswimmer/status/2051807284691612099/quotes)
---
*导出时间: 2026/5/6 20:51:52*