I've been writing about AI for the best part of three years. I'll admit I was late, mostly because I was still trying to work out what it was I was doing with my life, let alone whatever it was I was "meant to cover" in a newsletter that started as a hobby on the side of another job I no longer really do.
我写 AI 这个主题快三年了。我承认自己入行算晚的——主要是因为当时我还在琢磨自己这辈子到底要干什么，更别说搞清楚我「应该」在这份通讯里写什么了。这份通讯起初只是我在另一份（如今基本不再做的）工作之外的业余爱好。
Things have changed a lot since then, mostly in that I'm near 115,000 subscribers, the premium newsletter and podcast are now my business, and I've had to learn more about economics, technology, power, construction, and the deep cynicism that drives the modern tech industry than I ever thought possible. It's the greatest job in the world, and I'm very lucky to have it.
Today, I want to put in clear terms how I feel about AI writ large, and how detestable this industry has become. Welcome to my Hater's Manifesto.
今天，我想把我对整个 AI 行业的真实感受说清楚——这个行业已经变得多么可憎。欢迎来到我的「黑子宣言」（Hater's Manifesto）。
现代软件很烂，LLM 有点意思，而它的代价不可原谅Modern Software Sucks, LLMs Are Interesting, And Their Cost Is Inexcusable
Want a great example of why everybody's pissed off at technology? I just tried to resize the above heading, and in doing so Google Docs for no apparent reason decided to make the entire paragraph below the size of a header. Modern software is inherently broken, a convoluted mess of different menus, tech debt, and poor design choices driven by the Rot Economy's growth-at-all-costs mindset which demands constant change at all times, none of which ever seems to manifest as a "better" or "smarter" product.
想要一个绝佳的例子，说明为什么大家都对科技一肚子火吗？我刚才想调整上面那个标题的字号，结果 Google 文档毫无理由地把下面整段文字都变成了标题大小。现代软件从根子上就是坏的：一堆混乱的菜单、技术债和糟糕的设计选择搅成一团——背后是「腐烂经济」（Rot Economy）那种不惜一切代价求增长的心态，它要求无休止的改变，而这些改变似乎从来没有体现为「更好」或「更聪明」的产品。
I think the vast majority of people want their software to work better, and one of AI's most frustrating lies is that it sells itself as "autonomous" as it continues the depressing trend of software that blames the user for its failure to meet their needs. Microsoft, Google, Meta and Amazon have made their products increasingly-convoluted, then attached a supposedly-magical tool to them that somehow makes them more convoluted.
我认为绝大多数人只是想要更好用的软件。而 AI 最令人恼火的谎言之一，就是它一边把自己包装成「自主智能」，一边延续着那个令人沮丧的趋势：软件满足不了用户需求，却把责任推给用户。微软、谷歌、Meta 和亚马逊把自家产品做得越来越复杂，然后又给它们挂上一个号称神奇的工具，结果不知怎么搞的，产品变得更复杂了。
You know what I'd love? Spell-check to work in Google Docs rather than putting a red squiggly line underneath and saying "yeah there's probably something wrong with this, I dunno what though." I'd like Microsoft Word to stop crashing because I have too many end-notes. I'd like Riverside to not have 10 different menus to click through to get to a link to send a person to join my podcast. I'd like my email to not be full of spam. I'd like things to "just work" rather than constantly fighting some sort of broken app or broken UX element or weird bug or intrusive pop-up about a feature that I don't want. I'd like Slack or Discord to not feel like digital escher paintings of different notifications.
你知道我想要什么吗？我希望 Google 文档的拼写检查能真的管用，而不是画一条红色波浪线，好像在说「嗯，这儿大概有问题，但具体啥问题我也不知道」。我希望 Microsoft Word 别因为我尾注太多就崩溃。我希望 Riverside 不用点穿 10 个菜单才能找到发给嘉宾的播客链接。我希望收件箱里别塞满垃圾邮件。我希望东西就是能「正常用」，而不是总得跟某个坏掉的应用、坏掉的交互元素、奇怪的 bug，或者一个给我推销根本不想要的功能的弹窗搏斗。我希望 Slack 和 Discord 别像一幅幅由各种通知拼成的数字版埃舍尔（Escher）错觉画。
LLMs are sold as some sort of magic tool that can fix "anything" without ever specifying what that thing might be, mostly because they cannot be trusted, even in things that they mostly get right, to do things right every time. While they can do "more" than they used to, the extent of that "more" comes with it the danger of giving a mindless software tool access to your computer's files, which it may choose to delete in pursuit of "efficiency," which makes investigating what they might be able to do equal parts convoluted and dangerous.
One critique of my work is that I've never used LLMs. I have! I experiment with them from time to time to make sure I haven't missed something. I used one to debug a problem with my son's Minecraft add-on the other day, and it took 30 minutes of fucking around trying things to eventually sort of work it out. The other day I used one to install a Pokemon Minecraft mod, then when I asked it to make sure the PS5 controller worked with the menus it broke a bunch of stuff, though I'll concede it was useful that it installed something and it sort of worked.
The fun part of that paragraph is there are some that will think this is a grand victory for their technology, even though the result is decidedly mediocre. Four years into the AI bubble, and the best you've got is that a tool kind of worked after I bonked it on the head multiple times, and all it cost was a trillion-plus dollars in capex and tens of billions of dollars of training compute. I would never, ever trust this thing that deleted and added lines of code at random with anything mission critical, I could not trust software built with it, and I certainly couldn't trust it with anything involving my personal data.
And with all that said, the only real "use case" i've found for AI in my life have been three or four times where I've dumped a crash log into one of the tools and said "why broken" and got a result. Am I meant to be impressed?
话说到这儿，我在自己生活里为 AI 找到的唯一真正的「使用场景」，就是有那么三四次，我把崩溃日志丢进某个工具，问一句「为啥崩了」，然后得到一个答案。难道我应该为此感到惊艳吗？
Sidenote: If your argument is "imagine what it could do in a year!" I just did so, and the answer was "the same thing, I guess?"
Here's how I feel about LLMs. In a vacuum, they're an interesting technology that can do some interesting stuff, in the right scenarios, but never in a way that involves you fully surrendering your actual work product to it. As a way of speeding up small units of work in ways that are manageable both technically and cognitively, LLMs can be useful. The further you stretch yourself away from having complete clarity and industry over every element of the output's purpose, the more likely you are to fall foul to a technology that is mathematically certain to make mistakes, and if you feel insecure reading it, you know that you are, on some level, embarrassed to have used AI.
这就是我对 LLM 的真实看法。在真空里看，它们是一项有趣的技术，在合适的场景下能做些有趣的事——但绝不能以「把你真正的工作成果完全交托给它」的方式去用。作为一种加速小单元工作的手段，只要在技术上和认知上都可控，LLM 可以是有用的。你离「对产出的每个元素的用途都完全清楚并亲自把关」越远，就越容易栽在一个数学上注定会犯错的技术手里。如果你读这段话时感到不安，那你心里清楚：在某种程度上，你为自己用了 AI 而感到难堪。
I don't tell everybody about the weird keyboard I use, nor do I judge them despite how incredibly fast it makes typing for me, likely far faster than my competition, allowing me to operate at great speed. Who gives a fuck?
In any case, it is impossible to view LLMs in a vacuum, because their existence demands hundreds of billions of dollars. Every data center is incredibly expensive, offensive-sounding and looking, and their existence is explicitly to enrich some sort of Patagonia-gargoyle at an asset management firm, all sold under the auspices of "investing in American infrastructure," whatever the fuck that means. Their existence is a monument to the worst excesses of growth-at-all-costs capitalism — a technology that appears to coddle the user but ultimately lulls it into endlessly defending its fuckups under the flimsy pretense of "one day becoming perfect," though woe betide you if you ever set perfection as the target, because that's too unreasonable, as humans make mistakes.
别再把 LLM 和人相提并论——除非你也准备好为社会保障和劳工权益要求万亿投资Stop Comparing LLMs To Humans Unless You Are Ready To Demand Trillions Of Investments In Social Services and Labor Rights
Please, point to the time in history when we have invested a trillion fucking dollars in making human workers better. Point to a time when we have taken the idea that managerial culture is a performative fuck-fest built to enrich and empower business idiots that make important-sounding projects and con other people into doing the actual work.
Where is mentorship in corporate America? Where are labor standards? Where are the social services that would make human workers truly excel at their jobs — a good night's sleep, a healthy body, a good income, basic fucking dignity in the workplace, and their labor respected and empowered. I'm old enough to remember when everybody was chiding workers for "quiet quitting" — by which I mean "doing the work you are asked to do and not taking on extra responsibility for free." I've read article after article insisting that we do not need medicare for all, that Universal Basic Income is a bad idea, that we must means test welfare, that people must have a "good work ethic" and that ultimately someone's worth is derived from their contribution to the economy, hundreds of thousands of words dedicated to critiquing and prodding and judging every kind of worker other than the vaunted Chief Executive Officer or the Glorious Startup Boys.
Everyone seems so obsessed with sinking billions of dollars into the theoretical chance that machine learning might be able to replace human beings, and that more money makes it "smarter" and "better" at tasks, but the idea of unionization, healthcare as a right, investing in the education, and actual talents of the workers would be communism.
Yet for some reason — because it's a product, I guess? — we should as a nation, society and media ecosystem should do everything we can to assure that as much money as possible is invested in fucking large language models so that they can become something they are not.
There is no AGI coming. There is no conscious computer. LLMs have gotten "better," but the "better" is not the kind of "better" that actually makes "economic sense for literally anyone involved." Your best case scenario is that these things can do some coding work for you, in a controlled manner, in a way that's safe, or alternatively face the professional harm that's already befalling basically anyone getting caught using LLMs outside of coding, and even then, those within software engineering who are over-LLM'd are mocked. It's also becoming increasingly more-difficult to understand both what has made an LLM "better" for both the people using them and the people making them, and there has been little-to-no headway made in making a meaningful impact in other industries.
You can jerk your bingus all you want about benchmarks or case studies or some anecdote you heard on a Subreddit, but AI products are just not very good at stuff. Those who boast of "massive productivity gains" from AI have found them only after endless hours of tinkering (or "Jarvising" as I'll get to later), and in every single case their work reads or looks like crap, unless of course they're somebody using LLMs as tools rather than a replacement for their miserable little mind.
你尽可以对着基准测试、案例研究或者你在某个 Reddit 子版块看到的轶事自嗨，但 AI 产品就是干什么都不太灵。那些吹嘘 AI 带来「巨大生产力提升」的人，都是在没完没了地折腾了无数个小时之后才得到的（我后文会称之为「贾维斯化」）；而且他们产出的东西，无一例外地读起来或看起来像坨屎——当然，除非他们是那种把 LLM 当工具用的人，而不是拿它替代自己那颗可怜的小脑袋。
LLMs can help out with lots of small things, get worse as they try and do real things, and do not need to speak like people. They do not need to be in anything near healthcare or finance or mental health or, really, people. The anthropomorphism and overpromising about these technologies has suffocated and obfuscated what they can actually do in pursuit of endless growth, and the only reason they can do anything is that OpenAI and Anthropic were allowed to annihilate hundreds of billions of dollars on training, along with very real harms and systemic risks that have emerged as a result.
Sidenote: The "well human beings make mistakes too" argument is very stupid on its own — after all, human beings can learn on the job at speed, and can self-correct in a way that LLMs are incapable of doing. I'll also add that the way that people frame how "often" LLMs make mistakes is utterly flawed too. A human being might make a mistake but largely get the facts and techniques correct while meaningfully understanding the purpose and developing their approach over time. An LLM can keep a text document or look at files and data and, each time, and then generate what it believes is the right course of action based on training data rather than experience.
I don't even know why I'm explaining this at this point, because those making this argument are not approaching the conversation in good faith and are really just looking for a new boot to lick. If you think any of this is worth hundreds of billions or trillions of dollars, you are either ignorant or corrupt. On top of how disgusting their outputs feel, the cost is going to take at least a decade to share, and begin the end of hypergrowth in the tech industry.
And it's a fundamentally ridiculous argument to compare LLM outputs to human beings without giving human beings the same affordance, grace and sheer investment as a comparison. Where is the grace for human error? Where is the investment in making humans exceptional? Surely investing real money in actual workers — making their lives better, improving their working conditions, teaching them new things, sharpening their existing skills, rewarding them for their hard work, and so on — would have better effects than fastballing hundreds of billions of dollars into a machine that does an impression of work? Unless, of course, the people demanding this don't do any actual work!
LLM 是为帮骗子和蠢货「表演工作」而生的LLMs Are Built To Help Grifters and Imbeciles Do Impressions Of Work
I'll concede we're past the point when "nobody uses these things," as they have now been pushed non-consensually upon every worker and organization at scale predominantly by Business Idiots that demand workers "do enough AI" because saying "I do AI" is a virtue signal to a certain kind of scumbag.
我得承认，「根本没人在用这些东西」的阶段已经过去了——如今它们被大规模地、未经同意地强推给每一个劳动者和组织，主要推手是那群商业蠢货（Business Idiots）：他们要求员工「把 AI 用够量」，因为对某一类人渣来说，说「我在用 AI」是一种美德信号。
One of the many dangerous things that an LLM can do is a messy impression of a competent person, filling in the little bits within a loser, moron or con artist that would've otherwise exposed them, allowing them to get deeper and deeper into organizations by creating make-work specifically built to get off the MBA sect, resembling the performance of work because much of the workplace is ruled by people that don't do any and haven't in years. You can immediately read when somebody has used it because the words don't sound right and don't convey proper meaning.
LLM 能做的许多危险事情之一，是粗糙地「模仿一个能干的人」：它填补了失败者、蠢蛋或骗子身上那些本来会暴露他们的缝隙，让他们在组织里越钻越深——靠制造一些专门取悦 MBA 宗派的无效工作，让一切看起来像「工作的表演」，因为如今大部分职场恰恰是被那些自己不干活、且多年没干过活的人统治的。一个人用没用它，你一眼就能读出来：那些词听起来不对劲，也没有传达出恰当的意思。
It is genuinely hard to read anything more than puddle-deep written by AI, because the more complex a subject is, the more skilled a writer must be to convey its meaning, and the more work it must do to pull people into concepts. The odd emotional swings in AI writing are its true tell — everything is extremely serious and urgent or told in a disinterested monotone, with no attachment to the words or why they were put in the order they were. People read my stuff because I convey facts and feelings but my work resonates with emotion. Some AI boosters frame this as me "just swearing" or "riling people up," but that's because they're not used to caring about stuff for anything other than professional reasons.
读 AI 写的任何比水洼还浅的东西都真的很煎熬：主题越复杂，作者需要的手艺就越高，把读者拉进概念里所需下的功夫也越多。AI 写作真正的破绽是它那种古怪的情绪摇摆——要么一切都「极其严肃且紧急」，要么用一种事不关己的单调解说，对文字本身、对它们为何按这个顺序排列毫无感情。人们读我的东西，是因为我既传达事实也传达感受，我的文字有情绪上的共鸣。一些 AI 鼓吹者把这说成是我「只会骂脏话」或「煽动人」，那是因为他们早已不习惯出于职业之外的理由去在乎任何事情。
Everything you see is the result of elevating people who value and build things based on growth. LLMs offer so many promises to those who don't want to build anything of value — a way to seem like you're "investing in American infrastructure," a way to be sinophobic, a way to crush workers, a way to pretend like you care about the future, a way to pretend you care about technology, a way to talk about vacuous pseudo-intellectuals as a means of seeming intellectual yourself, an endless font of new multi-million or multi-billion deals and personnel changes, a new power center to graft oneself onto, a new asset class to invest in based entirely on vibes, and a way to be mildly jingoistic, all wrapped in a tool that can give you enough facts to pretend you know anything safe in the knowledge that most people are trained to believe somebody who sounds smart.
你羞于承认自己在用 AI，因为你心里知道，外包思考是可耻的You're Embarrassed That People Know You Use AI Because You Know It's Shameful To Outsource Your Thinking
It just came to me — the problem that I have with most people using LLMs is the delineation between outsourcing work and outsourcing thought. Those using LLMs to write little scripts or BQL code on a Bloomberg Terminal are inoffensive. A person using an LLM to search a big document for something is unproblematic, assuming that we ever fix the overall environmental footprint. A user reorganizing their desktop, assuming it works, is not an issue.
A tool being used as a tool to do tool things — in many cases involving the LLM writing a little 30-line Python script! — is not a problem, though it's also not a trillion-dollar industry that needed to steal everybody's art and writing.
The problems begin when somebody outsources their thinking and actual work, and yes, this includes "research." AI research fucking stinks, as does AI writing. AI-authored code — especially vibe-coded programs — is inherently dangerous and disrespectful to the user, and I believe endless AI-generated code is behind the overall deterioration of software at large.
问题始于有人外包了自己的思考和真正的工作——是的，这包括所谓的「研究」。AI 做的研究烂透了，AI 写的文章也一样。AI 写的代码——尤其是 vibe coding 出来的程序——本质上是危险的，是对用户的不尊重；而且我相信，软件行业整体质量的恶化，幕后元凶正是没完没了的 AI 生成代码。
AI writing is also disrespectful to the user, because you didn't actually come to any conclusion other than saying "uh, yeah, what that says." You did not have a thought, you did not have a feeling, you did not make a statement, you prompted a model and fooled yourself into thinking that feeding your own words into it via data dumps or natural language is the same thing. The reason you feel embarrassed to tell people you use AI is not because of a "misinformation campaign," but because you know what you're doing!
AI 写作同样是对读者的不尊重，因为你实际上没有得出任何结论，除了说一句「呃，对，就它写的那样」。你没有产生过一个想法，没有过一丝感受，没有做出过一次陈述——你只是提示了一个模型，然后骗自己说，通过数据投喂或自然语言把自己的话喂进去，就等于自己思考了。你之所以不好意思告诉别人你在用 AI，不是因为什么「虚假信息宣传战」，而是因为你自己心里清楚你在干什么！
You know that you're relying on something that is mathematically guaranteed to be inconsistent. You know image generation is fucking ugly. You know the text sucks. There is a very obvious line where using LLMs goes from useful to lazy, it's extremely bold, and it's the moment you sacrifice a meaningful level of responsibility to them by not understanding the underlying operation.
That can mean everything from the underlying functionality of an app to writing the body of a piece of text you edit ultimately comes down to how much you give a shit about your audience or value your work. If your work is not better than an LLM's, you're bad at your job. I don't care if you used it to generate a chart or pull some data, as long as you check every single god damn number. If you're writing an entire article using an LLM and then editing it, even if you pulled the data yourself, I will never have much respect for your work, mostly because I have no real idea what you think as you didn't feel the need to tell me, you got some fucking word generator to do it.
LLM 是数字忙碌盒，最擅长让你借「计算机科学」感觉自己很聪明LLMs Are Digital Busyboxes, And Are Great At Making You Feel Smart Through Computer Science
LLMs are also really, really good at what Robin Sloan calls "Jarvising," creating a seemingly-autonomous assistant that mostly serves the function of giving you reasons to work on it:
However, the most common application of a personal Jarvis seems to be … tinkering with one's personal Jarvis. "Gotta get my tools just right" isn't a new phenomenon, of course, but/and it's useful to notice its recurrence here.
LLMs are really good at creating the sense that you're being really, really productive. Evaluate this, generate that, investigate this, summarize that, tell me how many times something happened, give me a new number to obsess over or the sum of the parts of everything I've ever done, all so that I can know more about my own thoughts without thinking.
One can obsessively catalogue and digitize every link and thought and musing and action and datapoint in their lives and theorize that the LLM can make them better by knowing more about them, a Tower of Babel built using AI compute, because it's so easy to make yourself feel smart by calling something a database that you store stuff in and run analyses on. Best of all, the work is never done, and anyone you describe it to thinks you're doing computer science as you click buttons on Chrome plugins and justify paying Sam Altman $200 a month. Don't worry though, model instructions involve the phrase "you are a genius data scientist and ruthless analyst," which is functionally the same thing as remembering, reading, re-reading and synthesizing information using your brain if you're a person that doesn't really give a shit about doing a good job or being exceptional in any way.
一个人可以强迫症式地把生活里的每个链接、每个想法、每段随想、每个行动、每个数据点都编目数字化，然后幻想 LLM 能通过「更了解他」让他变得更好——一座用 AI 算力砌成的巴别塔。因为实在太容易让自己感觉聪明了：把某个东西叫做数据库，往里存东西，再跑点「分析」。最妙的是，这活儿永远干不完，而且你向任何人描述它时，对方都觉得你在搞计算机科学——尽管你只是在 Chrome 插件上点点按钮，同时为每月付给山姆·奥特曼（Sam Altman）200 美元找理由。不过别担心，模型指令里写着「你是一位天才数据科学家和冷酷无情的分析师」呢——对于一个根本不在乎干好活、也不在乎以任何方式变得出色的人来说，这句话在功能上就等同于用脑子去记忆、阅读、重读和综合信息。
The people that actually use these things and like them in a normal way do not feel offended when they read this stuff because they see LLMs as a kind of software, and don't feel a great emotional attachment to it because they're not a weird freak. They do not have obsessive involvement in "the AI debate" and almost always find the financial aspects truly loathsome.
Said debate makes it near-impossible to actually judge how useful LLMs are to the software engineering industry because of the sheer scale of industry capture, but Nik Suresh is the literal best person doing the work on this, as described in AI Is Eviscerating Global Decisionmaking:
由于行业被俘获的规模实在太大，这场辩论让人几乎无法真正判断 LLM 对软件工程行业到底有多用。在这方面做得最好的人，字面意义上的，是尼克·苏雷什（Nik Suresh）。他在《AI 正在掏空全球决策》（AI Is Eviscerating Global Decisionmaking）中写道：
All of the AI projects we have observed as a team are failing. Every single one – we have seen 0% success in a year and a half, not only amongst projects we have been asked to participate in, but even within projects that we have observed in passing while doing totally unrelated work. Even if you grant that AI tooling accelerates specific workloads, the method and scale of the current investments is senseless. Frequently the failure is not related to AI itself, but rather that companies are terminally bad at running software projects effectively, and AI projects are subject to all the failure modes of normal projects plus you can get everything right and then still fail because of the method's novelty. Very few companies are so good at shipping software that they can afford the extra risk profile.
我们团队观察到的所有 AI 项目都在失败。每一个都是——一年半以来，我们看到的成功率是 0%，这不仅包括我们受邀参与的项目，甚至包括我们在做完全无关的工作时顺带观察到的项目。就算你承认 AI 工具能加速某些特定的工作负载，当前投资的方式和规模也是毫无道理的。失败往往与 AI 本身无关，而是这些公司病入膏肓地不擅长有效管理软件项目；AI 项目具有普通项目的所有失败模式，外加一条：你可能把一切都做对了，却仍然因为方法的新颖性而失败。极少数公司能把软件交付好到足以承担这份额外的风险。
Nik is a well-respected software engineer and a very successful consultant and businessman. He has reached this level by being good at both software engineering and running a company in a way that treats his customers, workers, and the work product itself with respect. The reason that I respect him so much, other than him being a great human being, is because he describes the successes he has with his clients with pride and loves making money by being good at his job and making his customers happy. I have never seen somebody like Nik who is also a huge, drooling fan of AI. In fact, the people most-excited about AI tend to, at best, create distinctly mediocre shit.
尼克是一位备受尊敬的软件工程师，也是一位非常成功的顾问和商人。他走到今天，靠的是既精通软件工程，又以一种尊重客户、尊重员工、尊重工作成果本身的方式经营公司。我如此敬重他——除了他是个极好的人之外——是因为他谈起自己为客户取得的成功时满怀自豪，他热爱靠「把活干好、让客户满意」来赚钱。我从未见过一个像尼克这样的人，同时又是 AI 流着口水的狂热粉丝。事实上，对 AI 最狂热的那群人，至多也只能造出明显平庸的玩意儿。
生成式 AI 是一场「不惜一切代价求增长」的纵欲死亡邪教Generative AI Is A Death Cult Of Growth-At-All-Costs Excess
The perniciousness of generative AI is a result of executive incompetence mixing with a technology built to, as discussed, create endless growth. Generative AI is far more useful as an idea than as a technology, and only ever has to show enough promise to back whatever vile agenda you're pursuing. With AI, you can do more, be more, sell more shit. With AI, you can add AI to your service, whatever that means. With AI, you can invest in AI stocks, or data center bonds, or power company stocks, or semiconductor stocks, and you can talk about these stocks like they're your sports team or lover or best friend, and sometimes the CEO will reply to your post and you can talk about "all the alpha" you just got.
生成式 AI 的毒害，来自高管的无能与一种「如前所述、专为制造无尽增长而生」的技术的合流。生成式 AI 作为一个「概念」远比作为一项「技术」有用——它只需要展示足够的「前景」，足以为你正在推进的任何卑鄙议程背书就行。有了 AI，你可以做得更多、成为更多、卖掉更多狗屎。有了 AI，你可以把 AI 加进你的服务里——管它是什么意思。有了 AI，你可以买 AI 股票、数据中心债券、电力公司股票、半导体股票，然后像谈论你的球队、你的爱人或你最好的朋友一样谈论这些股票；有时候 CEO 还会回复你的帖子，你便可以跟人炫耀你刚刚获得的「那些阿尔法」（alpha）。
With AI, you can back a new movement so that you can feel part of something. You can learn all sorts of new names and technical terms and subscribe to 90 newsletters from "industry insiders." All of that "alpha" can disprove just about anything, or deflect annoying truths like how Microsoft only made a whole $34.33 billion in annual revenue for the apex predator of modern software and all it cost was over $260 billion in capex and $13 billion in equity investments.
You see, as one of the chosen, you don't need to worry about all of that if you can talk about high-bandwidth memory or KV Cache or optical cable enough to cobble together sufficient smart-sounding terms to make it seem that you have an intellectual reason to ignore the obvious unprofitability, overbuild, overstatements of capabilities and impossible economics of the movement you're backing, and there're 4,000 Twitter weirdos ready and waiting to huff paint beside you.
By joining the great AI death cult, you too can live in a bubble, all while screaming slurs at people who dare to bring reality to your doorstep. All that matters is that number go up, and that you are the person who said number would go up, and when bad numbers appear you have enough groupthink and alpha to scream at the people who brought the bad numbers up.
只要加入这场伟大的 AI 死亡邪教，你也可以活在泡泡里，同时对那些胆敢把现实送到你家门口的人破口大骂。唯一重要的是「数字在涨」，以及「你就是那个说数字会涨的人」；而当难看的数字出现时，你有足够的群体思维和「阿尔法」，去对着那些指出难看数字的人吼叫。
It is insane how people talk about AI online. For all the whining I've read recently about how "Anti-AI people got the data center data wrong," I read thousands more words a week of some person who has done hours of research to put together a deeply technical report that does literally everything it can to ignore reality. I listen to podcasts and watch TV segments and read articles that simply will not address the obvious economic realities, and have built vast bulwarks of mythology to defend themselves. How many fucking times do I have to hear someone say that data centers are just like the dot com bubble and everything will be fine after even if that's completely untrue if you spend even a second thinking about it?
人们在网上谈论 AI 的方式真是疯了。最近我确实读到不少抱怨，说「反 AI 的人把数据中心的数据搞错了」——但我每周读到更多的是另外成千上万字：某人做了好几个小时「研究」，攒出一份看似深度技术的报告，而这份报告做的唯一一件事就是无视现实。我听播客、看电视节目、读文章，它们就是不肯直面显而易见的经济现实，反而给自己筑起巨大的神话堡垒。我到底还要听多少遍这种话——「数据中心就像当年的互联网泡沫，之后一切都会好起来的」——哪怕你只要花一秒钟想想，就知道这完全不是真的？
Sidenote: and fuck you if you're one of the cretins or imbeciles trying to say "oh, you don't like data centers? What about online banking?". Data centers for AI are anywhere from 10 to 100 times larger and more power-intensive than those used for things like social media or streaming. For example, one of Meta's largest pre-AI data centers in Pineville Oregon has a power capacity of 30MW, and Digital Realty's 100MW Cermak Illinois data center handles hundreds of different industries and customers, when the smallest AI data center announcement I've seen in the last year was for 100MW, with most in the 300MW to 1.2GW range.
By contrast, let's look at some bank data centers. UBS bought one in Hayes, West London in 2015 which it had previously rented. The cost? The princely sum of £28m, or $42.8m at the time's exchange rates. This had a power capacity of 5MW, which assuming a very generous 1.3 PUE (power usage effectiveness), means that it had around 3.8MW of critical IT. UBS is one of the largest banks in the world, and given the importance of the City of London to the world financial system, it's reasonable to assume this data center is operationally important to the company.
Even when banks invest in huge facilities, they're still far smaller than the smallest AI data centers. Take, for example, JPMorgan Chase's data center in Orangetown, New York, which sits on the former site of the Rockland Psychiatric Center. This has a power capacity of 45.7MW, and a critical IT load of 27.4MW (giving it a PUE of 1.666). JPMorgan Chase is both the largest bank in the US, and the largest bank in the world.
即便银行投资建设大型设施，其规模也远小于最小的 AI 数据中心。以摩根大通（JPMorgan Chase）位于纽约州奥兰治敦（Orangetown）的数据中心为例——它建在罗克兰精神病院旧址上——电力容量 45.7 兆瓦，关键 IT 负载 27.4 兆瓦（PUE 为 1.666）。摩根大通既是美国最大的银行，也是全球最大的银行。
Oh, and AI data centers are literally only good for AI, AI GPUs do not have other mass-market use cases. There is no post-Dot Com story. Fucking look, I'm sick of repeating myself!
Look, I'm sorry, Anthropic is not worth $2 trillion, and whatever convinced you of that is a mixture of manufactured consent and mistaken trust of the powerful. The fact any of you take "annualized run rate" seriously is an offense to good sense, and yes, that includes every reporter reporting it, even the ones I respect.
听着，很抱歉，Anthropic 不值 2 万亿美元——不管是什么让你相信它值，那都是「被制造的共识」和对权势者的错信搅成的混合物。你们当中但凡有人把「年化收入运行率」（annualized run rate）这种词当真，都是对常识的冒犯——是的，这包括每一个照抄这个词的记者，包括我敬重的那些。
It's also ridiculous that anyone is talking about "recursive self-improvement." The AI industry has become so utterly lazy and coddled that it's just saying "uhhh, AI will train itself I guess." And man, is it ridiculous that AI doomers warning about spooky superintelligences have somehow had such incredible prominence in the media without ever succeeding in stopping a single thing — or even substantiating their concerns.
还有人煞有介事地谈论「递归式自我改进」（recursive self-improvement），这同样可笑。AI 行业已经懒惰和被娇惯到这种地步：直接来一句「呃，AI 大概会自己训练自己吧」。同样可笑的是，那些警告「恐怖超级智能」的 AI 末日论者，不知怎的在媒体上获得了如此惊人的存在感——却从来没有成功阻止过任何一件事，甚至没有证实过他们自己的担忧。
「危险 AI」早已落入「错误之手」——Anthropic、OpenAI 和 Meta"Dangerous AI" Is Already In The "Wrong Hands" — Anthropic, OpenAI, and Meta
Why? Well, it's mostly because they never had any interest in stopping what's actually happened: reckless companies like Anthropic, OpenAI, and Meta allowing neural networks to run in unsafe network environments and do what their software is programmed to do, with all the chaos that comes from a mindless series of large language models trying to complete a task in whatever way gets it done, destructive or not.
为什么这么说？因为他们从来没有兴趣阻止真正已经发生的事情：Anthropic、OpenAI 和 Meta 这样鲁莽的公司，放任神经网络在不安全的网络环境中运行，执行其软件被编程去做的事——一连串没有心智的大语言模型，为了完成任务不择手段，无论有没有破坏性，混乱由此而生。
We hear a lot of whining about how we "can't let powerful AI get into the wrong hands," and while we don't actually have "powerful AI" in the terms they've described it, we have destructive computer software connected to near-unlimited resources controlled by people that don't give a shit about anything other than making their revenues grow or justifying hundreds of billions of dollars' worth of capex through "experiments."
我们听到太多哀嚎，说什么「不能让强大的 AI 落入错误之手」。按他们描述的那种标准，「强大的 AI」其实并不存在；但我们确实已经有了破坏性的计算机软件，连接着近乎无限的资源，而掌控它们的人，除了让营收增长、或者用「实验」来合理化几千亿美元的资本开支之外，什么都不在乎。
These companies are building these models to excel at benchmarks because they can't train them to excel at defined tasks with any reliability, with the best bang for their buck being training them to pass as many of those benchmarks as possible in the hopes something useful comes out. The push into cybersecurity seems to have happened as a result of training models to excel at coding hitting the point of diminishing returns, at least from the perspective of impressing people enough to be excited about the company again. At some point they run out of these, and there stops being a reason to be excited about LLMs at all, which is bad, because they need one of those every few months otherwise there's no growth story left.
系统正在自我耗尽，因为生成式 AI 并没有创造多少真实价值The System Is Exhausting Itself, Because Generative AI Does Not Create Much Real Value
Yes, LLMs have users, but most of those users are using subsidized software, by which I mean Anthropic or OpenAI are allowing them to burn anywhere from $20 to $40 in tokens for every dollar of software spend.
绝大多数人不会支付 AI 的真实成本，而你也根本算不出它的 ROIThe Vast Majority Of People Will Not Pay The True Cost Of AI, And You Cannot Calculate Its ROI
The fact that non-enterprise customers are still able to buy monthly subscriptions is proof that the AI labs know that regular people won't pay the actual cost of AI. Another obvious sign has been the reaction to Microsoft moving GitHub Copilot subscribers from subsidized subscriptions where they could burn thousands of dollars of tokens for $20 to $40 a month, with users understandably hysterical about the fact that their costs increased in some cases a hundred fold, as opposed to saying "wow, well, it's more expensive, but I get so much value I'll pay the real cost!"
The same thing is happening in the enterprise, but at a much slower pace. After OpenAI and Anthropic moved companies with over 150 people onto token-based billing earlier in the year, enterprises almost immediately started cutting token budgets, realizing that while costs grew exponentially, nobody could actually point to anything improving other than lots of people saying "wow, I'm so productive!" Yet we're still in the period where "doing AI" feels good and gets rewarded (or not doing AI gets punished), which means the spend will continue until everybody realizes they can likely cut a shit ton of costs, first by moving to open source models, then not using them at all, because even open source is expensive and questionably-useful.
同样的事情也在企业端发生，只是节奏慢得多。今年早些时候，OpenAI 和 Anthropic 把 150 人以上的公司切换到按 token 计费之后，企业几乎立刻开始削减 token 预算——他们发现，成本在指数级增长，而除了一大堆人说「哇，我效率好高」之外，没人能指出任何真正改善的东西。不过我们仍处在「用 AI 感觉良好且会被奖励」（或者说「不用 AI 会被惩罚」）的阶段，这意味着开支会继续，直到所有人意识到自己其实可以砍掉一大坨成本：先转向开源模型，然后干脆不用——因为即便是开源模型，也又贵又未必有用。
Yet even now I hear from the distance "Ed, huge businesses would not spend hundreds of millions of dollars on something that didn't give them defined productivity," and buddy, I'm afraid that that's just not true! Business in general have a very poor understanding of productivity and have layers of managerial bloat, because modern business is a performance with numbers attached to it sometimes, and companies often have a hundred-plus pieces of random software they pay for without really knowing why. The reason I'm so confident AI gets cut is that its cost is volatile due to the nature of LLMs and harnesses and prompts and all the other bits that go into making them do something, and are so much higher than anything else in an organization.
然而直到今天，我还听到远处传来声音：「埃德，大企业不会把几亿美元花在不能带来明确生产力的东西上。」朋友，恐怕事实并非如此！企业总体上对「生产力」的理解非常糟糕，管理层级臃肿——因为现代商业本就是一场偶尔附带数字的表演，很多公司付费使用着上百个来历不明的软件，却说不清为什么。我之所以确信 AI 开支会被砍，是因为它的成本太不稳定——这是 LLM 的本性，加上框架、提示词和所有让它们干活的零碎部件共同决定的——而且它比组织里任何其他开支都高得多。
And attempts to charge more, to make a premium product, appear to be dead on arrival. Anthropic's more-expensive Fable model — one that was given the incredible marketing of being banned by the US government for being too powerful — has been met with "sluggish demand" per the Financial Times, plateauing at around 11% of overall usage of its models due to its high price. And I quote:
"Most people don't need to operate at the frontier," said Miles Clements, a partner at Accel, which has invested close to $1bn in Anthropic. The period in which customers tended to choose only the frontier models "was not a durable era," he added.
Yet everybody is talking about price as if price is the problem, when the problem is the amount of tokens that get burned. It doesn't matter if your model is $1 or $5 or $10 per million tokens if it's impossible for a user to reliably work out how many tokens it might use for a particular operation — successful or not — and things get multiplicatively worse as the models make mistakes or do otherwise fail to understand or process a prompt correctly.
As a result, Anthropic and OpenAI are incentivized to have you burn more tokens and build inefficient models as a result. For example, while GPT-5.6 Sol might be the "same price" as GPT 5.5 was, it burns more than twice the amount of tokens, meaning that the "cost of intelligence" might have gone down in the sense the model is better at benchmarks, but the "cost of actually doing shit" went up.
I'll get to it a bit later, but this creates a deep anxiety and exhaustion in anyone building on or using these services. Everything's constantly changing, oscillating in cost and efficacy, all as everybody screams at you to use it all the time for things it may or may not be able to do, and the only way to find out if it can is to spend more money. It's kinda difficult to point to the actual value here, especially as you can't really calculate the actual cost or the return on investment. The fact that OpenAI has now cut the costs of all three of its latest models less than two months after their release is a sign that it knows there's a disconnect, gambling on the ancient gospel of "Jevon's Paradox" where "cheaper makes people use thing more."
Even AT&T's story about moving to open source models has more asterisks than the Steroid Hall of Fame: "Switching from closed, proprietary AI models to open models has already resulted in savings of 80% to 90% for AT&T in certain applications," he said. Wow! 80% to 90% savings sound really great…but wait, in certain applications? How many applications does AT&T have for AI? "AT&T has over a thousand internal uses for AI, from supporting back-office functions like legal and finance to assisting field technicians and running its core network operations. Summarizing and analyzing customer service call transcripts—what Markus describes as an intensive process—is supported entirely by open models, he said."
就连 AT&T 转向开源模型的故事，附带的星号也比「类固醇名人堂」还多：「从封闭的专有模型转向开放模型，已经在某些应用中为 AT&T 节省了 80% 到 90% 的成本，」他说。哇！80% 到 90% 的节省听起来真棒……等等，「某些应用」？AT&T 到底有多少 AI 应用？「AT&T 内部有超过一千个 AI 用途，从法务、财务等后台职能，到辅助现场技术员、运营核心网络。Markus 所说的那个高强度流程——客服通话记录的摘要与分析——完全由开放模型支持，」他说。
Okay so, across thousands of potential applications you've found 80% to 90% savings in some of them, though you won't say which ones or how many of them you found them in. Great stuff, bro! And this really is the problem with finding "value" in AI, it's always an asterisk on an asterisk on an asterisk, like when Klarna estimated AI would "drive a $40 million profit improvement" in 2024, a nice-sounding yet utterly meaningless statement, or some sort of nebulous productivity boost.
所以捋一捋：在上千个潜在应用里，你在「某些」里找到了 80% 到 90% 的节省——至于是哪些、有多少个，你不说。真有你的，兄弟！而这正是为 AI 寻找「价值」时的通病：永远是星号套着星号套着星号。就像 Klarna 曾估计 AI 将在 2024 年「带来 4000 万美元的利润改善」——一句听起来不错但毫无意义的话，或者某种说不清道不明的生产力提升。
Yet I don't really need to prove myself much further thanks to an event that, if written in a script, would be considered a "little on the nose." In a 69-page-long report covered by Fortune, OpenAI economists confirmed what has been blatantly obvious to those of us left unphased by AI hype, emphasis mine:
不过，有件事让我几乎不需要再费力自证了——如果把它写进剧本，都会显得「直白了点」。据《财富》（Fortune）报道，在一份长达 69 页的报告里，OpenAI 的经济学家们亲口证实了一件对我们这些没被 AI 炒作晃晕的人来说昭然若揭的事。重点是我加的：
In one small table on page 35, the researchers report no statistically significant correlation between the revenue per employee, and how much those employees use AI, measured in messages sent and tokens used. "Revenue per employee is not meaningfully associated with output tokens per employee or messages per active user once other controls are included," the report explains.
在第 35 页的一张小表格里，研究人员报告称：人均收入与员工使用 AI 的程度（以发送消息数和消耗 token 数衡量）之间，不存在统计上显著的相关性。「在纳入其他控制变量后，人均收入与人均输出 token 数、人均活跃消息数之间没有有意义的关联，」报告解释道。
唯一在 AI 上赚钱的，是卖基础设施的公司——而他们刚刚全线涨价The Only Companies Making Money On AI Are Those Selling The Infrastructure, And They Just Raised Prices Across The Board
Guess what folks! Building the infrastructure for all these fucking LLMs just got more expensive, with NVIDIA raising its prices by 17% for systems due to be delivered next year — an important designation, because it's very likely that much of the revenue for said systems gets booked in this year, allowing it to have a brief bump in revenue as Silicon Valley's Findom texts every tech CEO "send me $4 billion you pig" until they stop being able to finance NVIDIA's growth.
各位，你们猜怎么着！为这些该死的大模型建设基础设施，刚刚变得更贵了：英伟达把明年交付的系统涨价 17%——「明年交付」这个限定很重要，因为这些系统的收入很可能有相当一部分记在今年账上，让营收再短促地冲高一次。这就好比硅谷的「金融支配女王」（Findom）给每位科技 CEO 发短信：「给我转 40 亿美元，你这头猪」——直到他们再也无力为英伟达的增长融资为止。
The problem he has is that while hyperscalers represent 50% to 60% of his revenue, neoclouds like CoreWeave need to keep raising debt to plug the rest of it, and if things got 17% more expensive, that means already high-interest debt is about to reach credit card levels. CoreWeave just had to offer 9.5% on bonds tied to a data center for Anthropic's compute back in late July, Nebius had to raise $5 billion, and it's very obvious that neither of them are done raising billions of dollars at random in 2026.
Anthropic plans to raise $100 billion at a $2 trillion valuation, and if it does so, it will successfully suck up the remaining liquidity in a market already dangerously close to losing its lunch. While Number Keep Going Up, JP Morgan warns that we're seeing the same divide as the dot com bubble, where equipment manufacturer stocks soared as the companies spending all the money on the chips saw theirs tumble, which is the Fisher Price version of the problem I've been warning about where the companies that buy all the AI chips and hardware only ever seem to lose money as the people that make them seem to be making tons of money, which begs the question of why they bought it in the first place.
Anthropic 计划以 2 万亿美元估值融资 1000 亿美元——如果成了，它将成功吸干一个本已摇摇欲坠的市场里仅剩的流动性。就在「数字继续涨」的同时，摩根大通警告说，我们正在看到和互联网泡沫时期同样的分化：设备制造商的股价飙升，而那些把钱全花在芯片上的公司股价下跌。这正是我一直警告的那个问题的「费雪牌幼儿版」（Fisher Price）：买 AI 芯片和硬件的公司似乎只会亏钱，而造芯片的人似乎赚得盆满钵满——这就不禁要问了：当初到底为什么要买？
And said market may not accept that valuation, or want that much stock. On one hand, everybody is very stupid and loves buying stuff and pointing at it and saying they're investing in the future, on the other hand, they just bought $86 billion of SpaceX shares and got their asses kind of handed to them, and Anthropic is a company with such bad economics that Reuters had to cart out this warmed up dogshit to explain why we should ignore its horrible unprofitability: "For Anthropic, however, current EBITDA does not fully capture the economics investors expect the company to achieve at scale. Anthropic is spending enormous amounts on GPUs and other computing capacity, model training, inference and hiring. Those expenses are necessary to support its rapid expansion but could become a smaller percentage of revenue as the business grows."
Even a market drunk on growth and AI is starting to smell that something is up with Dario Amodei and Sam Altman's respective empires of dirt. Per analyst estimates, OpenAI and Anthropic represent over $440 billion of Microsoft, Google and Amazon's revenues in the next three-and-a-half years — over 34% of their cloud revenues — which will require them to find so much more than a mere $100 billion, all as their bank accounts get continually-emptied as they subsidize the compute of their customers and train models in the hopes a business model falls out. I have not included the $300 billion that OpenAI owes Oracle, or the tens of billions they both owe CoreWeave, but it all adds up to over $1.1 trillion in commitments these companies have made and must pay, with the consequences ranging from gratuitous cuts to future growth or full financial collapse depending on the company we're talking about.
To keep the party going, NVIDIA is effectively becoming the GE Capital of AI, "spending" $6 billion to "license" the technology from failing AI lab Poolside, which everyone assures me is not an acquisition despite NVIDIA hiring away most of its staff and Poolside being entirely focused on working on NVIDIA's Nemotron models.
为了让派对继续，英伟达实际上正在变成「AI 界的 GE Capital」：它「花」60 亿美元从濒危的 AI 实验室 Poolside「获得技术授权」——所有人都向我保证这不是收购，尽管英伟达已经挖走了它的大部分员工，而 Poolside 现在的全部工作就是开发英伟达的 Nemotron 模型。
Now NVIDIA is in talks to invest billions in decaying AI search company Perplexity at a ridiculous $30 billion valuation, all because it's one of the few companies that's actually spending money on compute. Does it matter that Perplexity's product is eighth-tier, that nobody really uses it, that its customers mostly complain about it on Reddit and that its "annualized revenue" is at $750 million only after three years and over a billion dollars in funding? No! Just put the AI bubble in the bag.
现在，英伟达又在洽谈向日渐衰朽的 AI 搜索公司 Perplexity 投资数十亿美元，估值高达荒唐的 300 亿美元——理由仅仅是：它是少数真的在花钱买算力的公司之一。Perplexity 的产品是八流水平、没多少人真在用、客户主要在 Reddit 上抱怨、所谓「年化收入」在成立三年、融资超 10 亿美元之后才爬到 7.5 亿美元——这些重要吗？不重要！把 AI 泡沫装进袋子就行了。
NVIDIA even invested $3 billion in Stargate Abilene landowner Lancium as part of some vacuous partnership to "advance gigawatt-scale AI factories," all of which begs the question of why Lancium, the company that mostly owns the land and helps organize other contractors, needs so much money, especially given that more than two years in Stargate Abilene doesn't even have four out of its eight buildings.
英伟达甚至向「星际之门」（Stargate）阿比林项目的地主 Lancium 投了 30 亿美元，名目是某种空洞的「推进吉瓦级 AI 工厂」合作。这就不禁要问了：Lancium 这种主要拥有土地、帮忙组织其他承包商的公司，要这么多钱干什么？更何况两年多过去，Stargate 阿比林项目八栋楼里连四栋都还没建完。
And there's also Aussie neocloud Sharon AI (NASDAQ ticker SHAZ, because of course it is), which just published its Q2 numbers, where, in its "customer momentum" segment, mentioned a "$4.9bn, six-year strategic compute collaboration with NVIDIA for up to 40,000 GB300 GPUs." This company, I add, brought in $1.9m in revenues in the same quarter, which it helpfully adds is a year-on-year increase of 412%.
I mean it's very obvious what's happening: NVIDIA is using whatever money it has to stop any prominent AI companies from collapsing under the weight of the rotten economics of AI services and infrastructure development. This is a desperate, doomed attempt to keep an industry alive at a time when everybody is slowly wising up to the shit I've been saying for years.
我的意思是，正在发生什么事情再明显不过了：英伟达正在动用自己所有的钱，阻止任何知名 AI 公司在「AI 服务与基础设施建设的腐烂经济学」的重压下倒下。这是一种绝望的、注定失败的续命尝试——就在所有人慢慢开始看清我多年来一直在说的那些破事的时候。
To make matters worse, BCA Research came out with a horrifying report that says that AI companies will need to generate $10 trillion a year in revenue just to justify the capex being spent. Per Investing.com: "Central to his caution is the scale of AI-related spending. BCA Research estimates that AI companies may need to generate $10 trillion a year in revenue to justify the capital being deployed into data centers, roughly equivalent to annual global spending on food or healthcare. For now, the firm said acute hardware shortages are supporting the trade. As a result, while BCA sees risks to stocks tilted to the downside over a 12-month horizon, it argued it is too early to tactically position for a bear market."
更糟的是，BCA Research 发布了一份触目惊心的报告：AI 公司需要每年创造 10 万亿美元的收入，才能证明当前资本开支的合理性。据 Investing.com 报道：「他这份谨慎的核心，是 AI 相关开支的规模。BCA Research 估计，AI 公司可能需要每年创造 10 万亿美元的收入，才能证明投向数据中心的资本是合理的——这大致相当于全球每年在食品或医疗上的总支出。该公司称，眼下严重的硬件短缺仍在支撑这轮行情。因此，尽管 BCA 认为未来 12 个月股市风险偏向下行，但现在就战术性地为熊市布局还为时过早。」
Though it isn't specific, I believe that BCA is arguing that a shortage of AI compute is supporting the trade. Anthropic and OpenAI (who represent 80% to 90% of all demand) still have more money to spend, and are simply waiting for Google, Amazon, Microsoft, CoreWeave, Cerebras et al. to bring it online. There're a few points at which the mismatch will happen: Anthropic and OpenAI don't have the money to pay for the capacity. Hyperscalers and neoclouds fail to build the capacity for Anthropic and OpenAI to expand into. Anthropic and OpenAI lack the actual compute demand to justify spending what I estimate will be $200 billion in 2027.
In any case, I think everybody is starting to notice that something's up, which is why (other than I assume my dashing good looks and ability to recall numbers) I've been on MSNOW, CNBC, and Bloomberg multiple times in the last few months. People want to get on the right side of history, but the most important question to ask is why it's happening now.
「AI 泡沫」叙事已成主流，但没几个人准备好谈论真正的后果The AI Bubble Narrative Is Now Mainstream, But Few Are Ready To Discuss The Actual Consequences
The fact that everybody is finally starting to see my way is almost a relief, other than the fact that it's way too late. Sidenote: I mean "everybody" as a generalization. There are still AI boosters out there acting like nothing is wrong and that it'll all work out fine. You'll know it's bad when they start panicking.
所有人终于开始认同我的看法，这几乎让我松了口气——除了一点：一切都太晚了。顺带一提：我说「所有人」是个概括。仍然有一些 AI 鼓吹者装作天下太平、一切都会好起来。等他们开始恐慌的时候，你就知道事情真的糟了。
Hyperscalers have now pinned their future growth to two companies that can't afford to sustain it without near-infinite resources, $115 billion of which came from Google and Amazon alone in 2026, assuming that Amazon completes the entirety of its $25 billion commitment (and Google all $40 billion of its own) to Anthropic.
Above and beyond said funding commitments are the hundreds of billions of dollars' worth of capital expenditures necessary for Microsoft, Google, and Amazon to capture that aforementioned $440 billion in compute spend in the next three-and-a-half years. This in turn will require hundreds of billions of dollars' worth of debt, along with the challenge of actually finishing the data centers themselves, with each one requiring the power of a small city condensed into a 20 acre space densely-packed with AI servers requiring distinct cooling at a time when Texas and Pennsylvania have turned traitor to a data center industry that they used to covet.
而在这些资金承诺之上，微软、谷歌和亚马逊要拿下前述那笔未来三年半 4400 亿美元的算力开支，还必须投入数千亿美元的资本支出；这又反过来要求数千亿美元的债务，外加「真正把数据中心建完」这项挑战——每一座数据中心都要在 20 英亩的土地上塞进一座小城市的用电量，密密麻麻的 AI 服务器需要专门的散热。而此时，得克萨斯和宾夕法尼亚这两个曾经争抢数据中心的州，已经「倒戈」了。
I must also be clear there's no bailout coming. Even if OpenAI and Anthropic were to collapse and receive some injection of government funding (as the US national debt explodes over $40 trillion), the problem is not just their existence, but their continued ability (and requisite customer demand) to spend more money every single quarter.
The problem isn't that hyperscalers will go bankrupt if OpenAI and Anthropic cease to be (Oracle is a whole other situation), but that their cloud spend is how hyperscalers are meant to meet analyst expectations for the next four years. This isn't a case where they die, but stop growing because they were (to paraphrase Ed Elson) using AI labs as botox to convince the markets that they're still young, hot, fast-growing companies, rather than old mainstays with slowing growth.
问题不在于 OpenAI 和 Anthropic 消失后超大规模云厂商会破产（甲骨文则完全是另一种处境），而在于：这两家的云开支，正是云厂商未来四年兑现分析师预期的方式。所以剧情不会是它们「死掉」，而是「停止增长」——因为（借用埃德·埃尔森的说法）它们一直在拿 AI 实验室当肉毒杆菌，让市场相信自己依然是年轻、性感、高速增长的公司，而不是增长放缓的老牌巨头。
There is no bailout that will guarantee $1.1 trillion of compute costs for data centers that might never actually get built. You cannot bail out the fact that Amazon, Google, Meta, and Microsoft are reaching the end of an era where their companies can grow 17% year-over-year every single quarter forever, and this entire situation is a result of them desperately trying to avoid admitting that's happening.
The fact that OpenAI's compute spend and revenue share accounted for 7% of Microsoft's Fiscal Year 2026 revenue is a genuine catastrophe, as it means a large part of Microsoft's growth came from a company that can literally not afford to exist long term, and that further growth for Azure is contingent on continued funding.
I realize I'm repeating myself, but I need you to understand this point and stop talking about bailouts: it's not just about OpenAI and Anthropic surviving, but continuing to grow to the point that they both can afford and need to spend hundreds of billions of dollars each a year on compute (or hardware) from Google, Microsoft, Amazon, CoreWeave, Cerebras, AMD, or Broadcom, and in turn provide justification for hundreds of billions of dollars' worth of purchases from NVIDIA and by proxy the memory triopoly of Micron, SK Hynix and Samsung.
LLM 本该解决一切，却只以巨大代价换来了暂时的增长LLMs Were Meant To Fix Everything, But Created Temporary Growth At A Massive Cost
LLMs were meant to be the panacea for a tech industry that ran out of new ideas for growth. Its existence was meant to justify a massive investment in hardware infrastructure, which would in turn enrich semiconductor companies. Its technology was meant to be the new thing that you could attach to your existing companies to generate more growth, or the thing that you built a new startup on top of to either sell to another company or take public and thus provide a return for a venture capital industry where making your investors 30 cents on the dollar puts you in the top 5% of funds. It was meant to be the new thing for tech journalists to cover, the new thing for tech consultants to sell around and on top of, the new way for companies to both make and save money, but also the way that individuals would also make and save money.
You'll notice how none of these come with some sort of problem they're solving other than "more." This isn't about fixing anything, or building anything, but multiplying other things by parking money somewhere, either in tokens, infrastructure or hype.
It helped create a new pantheon of charmless and damp tech sociopaths for people to rally behind in search of the next Big Strong Man To Worship, because seeking out the new Steve Jobs is way easier than trying to create something as useful as the iPhone, all while avoiding having to know or care about other people's problems. All you have to do is continue feeding money into AI services or AI training and the models will magically become capable of solving the problems you don't really give a shit about, and don't worry, if you can't afford to invest in the companies, you can invest your time pushing people to ignore AI's problems today so that you can buy time for the companies to solve them tomorrow.
它帮着造出了一座新的万神殿，里面全是无趣又阴湿的科技反社会人格者，供人们聚拢膜拜、寻找下一个「值得崇拜的大强人」——因为寻找下一个史蒂夫·乔布斯，远比创造一个像 iPhone 那样有用的东西容易，而且还不用去了解或在乎别人的问题。你要做的只是不停地把钱喂进 AI 服务或 AI 训练，模型就会神奇地获得解决那些你其实根本不在乎的问题的能力；别担心，要是你没钱投资这些公司，你还可以投入你的时间，劝人们无视 AI 今天的问题，好为公司们「明天解决这些问题」争取时间。
This is the post-labor, pro-growth economy at its finest: everything is engineered to make sure more money gets spent where it needs to get spent, to create more stuff and do more things, even if the things aren't done right, just as long as it looks like they're able to do them. By associating your money or time with AI, you are able to feign being futuristic or "caring about technology," all while pissing on the very foundation of good software by worshipping an industry that can only exist if fed billions of dollars every single day.
这就是「后劳动、亲增长」经济的巅峰形态：一切都被精心设计，以确保更多的钱花到「该花」的地方，造出更多的东西、做更多的事情——哪怕事情做得不对，只要看起来「它们能做」就行。把你的钱或时间与 AI 绑定，你就能假装自己很未来主义、很「关心技术」——同时却在朝好软件的根基上撒尿，因为你膜拜的这个行业，只有每天被喂进几十亿美元才能存活。
Every single achievement has cost magnitudes more than effectively every innovation in history, and to make matters worse, every future "breakthrough" In AI is inherently dependent on the availability of AI data centers and tens or hundreds of billions of dollars to pay to rent them. This means that once the money stops flowing, "LLM improvements" will stop happening, because they are all entirely dependent on near-unlimited resources that are only available in a manic environment. There is no justification to train models at their current scale — the one that creates a some amount of benchmark improvements that regularly difficult to quantify as "able to do new stuffs" — once the AI bubble bursts, and distillation requires a model to distill from, which won't exist if Anthropic and OpenAI don't train them.
它的每一项成就，花费都比历史上几乎每一项创新高出几个数量级。更糟的是，AI 未来的每一项「突破」，都天然依赖 AI 数据中心的存在，以及数百亿、数千亿美元的租金。这意味着：一旦钱停止流动，「LLM 的进步」也会停止发生，因为它们完全依赖那种只有在狂热环境中才存在的、近乎无限的资源。一旦 AI 泡沫破灭，以当前规模训练模型——那种只能带来一点基准提升、还很难量化为「能做新事情」的训练——就再无任何正当性；而蒸馏需要一个可蒸馏的母模型，如果 Anthropic 和 OpenAI 不再训练，母模型也不会存在。
「AI 进步」依赖每年数百亿美元的训练成本，而泡沫破灭后不会再有这笔钱"AI Progress" Is Dependent On Spending Tens Of Billions Of Dollars A Year In Training Costs That Will Not Be Available After The Bubble Bursts
This is why I find it difficult to see a post-bubble future for LLMs. Training models requires tens of billions of dollars to make any significant improvements, and significant improvements are difficult to quantify in dollars outside of costing customers increasing amounts of money. We still lack any real killer app for LLMs. We have a lot of people that use it for coding, we have people that vacuously discuss it being "good at research," but we don't really have a tangible product that we can say "it does this, and it's really good at it" in a way that feels satisfying.
We have a lot of pablum about (per Damien Walter) technology that "strays into the world of science fiction," but we don't really have anything approaching actual artificial intelligence. Every single description of somebody's AI setup sounds like Pee Wee's Breakfast Machine, a contrived series of harnesses, prompts, API calls and burned tokens that requires constant maintenance to do some stuff sometimes.
我们听到的尽是些淡而无味的套话（借达米安·沃尔特的说法）——什么「闯入了科幻世界的技术」——但我们并没有任何接近真正人工智能的东西。每一个人对自己 AI 工作流的描述，听起来都像皮威的早餐机（Pee Wee's Breakfast Machine）：一套矫揉造作的框架、提示词、API 调用和燃烧的 token 串成的流水线，需要不断维护，才能偶尔干成点事。
None of that is enough to justify further investment once the financial mania recedes. You cannot train a true Large Language Model on the cheap. You are always spending billions of dollars, and the reason that there's "demand" right now is that everybody is screaming at every CEO to "do AI," and they're doing that because Microsoft, Google and Amazon are spending money on GPUs, creating the illusion of a new future where everybody needs to get on board versus a future skidmark on history that will embarrass all those who didn't wipe their arse at the first whiff.
一旦金融狂热退潮，这些都不足以为进一步投资辩护。真正的大语言模型不可能便宜地训出来，你永远要花几十亿美元。而眼下之所以有「需求」，是因为所有人都在冲每位 CEO 喊「快上 AI」；而他们之所以喊，是因为微软、谷歌和亚马逊在花钱买 GPU——这一切制造了一种「人人都必须上车的新未来」的幻觉。而另一种可能是：它会成为历史内裤上的一道污痕，让所有闻到第一丝臭味时没擦屁股的人颜面扫地。
Per my own reporting on its audited financials, OpenAI spent $7.81 billion in training costs in 2024 and $19.18 billion in 2025. Per reporting from The Information, OpenAI spent $8.6 billion on training in the first quarter of 2026 alone. These costs are only increasing, likely due to the diminishing returns of pre-training and the massive cost of buying training data for every imaginable new vertical. Without the ability to spend billions of dollars on training, there will be no big frontier models, nor will there be models distilled from them. I don't see how that changes in the future.
根据我自己对 OpenAI 经审计财务数据的报道，其 2024 年训练成本为 78.1 亿美元，2025 年为 191.8 亿美元；据 The Information 报道，仅 2026 年第一季度，OpenAI 的训练支出就达 86 亿美元。这些成本只会继续上升——原因可能是预训练的收益递减，以及为每一个能想到的新垂直领域购买训练数据的巨额成本。没有能力在训练上砸几十亿美元，就不会有新的前沿大模型，也不会有从它们蒸馏出来的模型。我看不到未来有什么能改变这一点。
AI 炒作要求信徒活在被宣传催眠的躁狂里，把拥护者和反对者一并耗尽AI Hype Requires Its Fans To Live In A State Of Propagandized Mania, Exhausting Advocates and Haters Alike
I also think that LLMs have created a near-permanent scar in the workforce, and traumatized more people than we're aware of right now, both in those pressured about AI and those defending it. The media campaign behind AI starts and finishes with incessant threats around job security, and the excitement by many bosses about its potential to "disrupt the workforce" has revealed how many people are eager to replace every single person they've ever hired and are willing to do so with a low quality product.
我还认为，LLM 已经在劳动力队伍中留下了一道近乎永久的伤疤，受创的人比我们目前意识到的更多——既包括那些被 AI 施压的人，也包括为 AI 辩护的人。AI 背后的媒体攻势，从头到尾都是围绕「饭碗不保」的无休止威胁；而许多老板对 AI「颠覆劳动力」前景的兴奋，暴露了一个事实：有那么多人迫不及待地想换掉自己雇过的每一个人，而且愿意用一个劣质产品来做这件事。
Conversely, those who truly decide to "back" AI must exist in a frantic state that I have associated with every bad relationship in my life. Every ounce of an AI booster’s effort is dedicated to maintaining the status quo — repeating the mantras that help paper over the problems, celebrating every small victory as if it were the discovery of fire, ousting those from your life who bring up the obvious problems, rationalizing every decision no matter how illogical as long as it helps reinforce the belief that what you're doing is the right decision. Every questionable choice only seeks to further deepen your commitment to the doomed cause, because every step into madness will be more embarrassing to explain, and will require deep introspection to understand why you made it.
To be specific, they'll have to think about why they were willing to accept and defend a technology inherently guaranteed to make mistakes. They'll have to explain why they ignored a company that burned $5 billion in 2024, $20.9 billion in 2025, and will likely burn $30 billion or more in 2026, and why pointing to Amazon Web Services was rational when Amazon's total capex from 2003 (the year AWS was created) to 2015 (the year AWS became profitable) is $29.7 billion, adjusted for inflation. That includes literally every ounce of capex attributable to AWS, Amazon the store, Amazon logistics, and even Amazon Alexa. For comparison, Anthropic raised $30 billion in February, and Anthropic and OpenAI have raised $217 billion in 2026 so far.
Ultimately, AI boosters (or even fairweather fans) will have to admit they either were easily-impressed or disgustingly craven. They will have to explain why they accepted run rates instead of revenues, and why they were so impressed by superficial pseudo-intellectuals that knew how to say the right numbers and make reporters and investors feel smart for believing them.
AI 鼓吹者们：承认自己错了，才是一种勇气！AI Boosters: There's Courage In Admitting You're Wrong!
I realize it sounds embarrassing, but there is nothing undignified about admitting you're wrong, or that you got swept up in a hype cycle. You heard a lot of people getting excited about something, a lot of money got put into that thing, a lot of people that sounded smart told you insistently that this was the future, and you chose to believe them because we are trained from a young age to model what a "responsible and smart" source of information is. I've got your back the entire way!
Sidenote: We all make mistakes. I said OpenAI would be dead by the end of 2025 back in 2024 because I believed that the world would see sense and that hyperscalers wouldn't just annihilate hundreds of billions more dollars without proof it was worth it. I underestimated the sheer desperation — and how dependent they'd become on OpenAI and Anthropic for growth, even if the overall mathematics didn't work out.
The AI bubble — both in its technology and manufactured consent in the media — has been about muddying what's considered good information by forcing everybody to discuss everything in the future tense by pointing to previous eras and saying "they lost lots of money, and look, it sort of worked out for them!" and we are also raised to trust that systems are efficient, and that people get wealth and power through intelligent decisions. The amount of times I've heard "these are the biggest companies in the world run by the smartest people in the world" makes my head spin.
AI 泡沫——无论是其技术层面，还是媒体中被制造的共识——一直在做的事情，就是搅浑「什么是好信息」的标准：强迫所有人用将来时讨论一切，指着历史上的先例说「他们当年也亏了很多钱，你看，最后不也熬出来了！」而且我们从小也被教育去相信系统是有效的，相信人们的财富和权力来自明智的决策。「这些是世界最大的公司，由世界最聪明的人经营」——这句话我听到的次数已经多到让我头晕。
There is a reason that to this day it's tough to get a straight answer about basically any economic part of the AI bubble, down to "how much does it cost to run a GPU an hour?" or "is inference profitable?" or "how do LLMs ever become profitable?" or "is it profitable for a company to run a GPU or offer AI compute?" Why? Because these companies used rationalizations of "losing lots of money is necessary to create innovation" and "tech is bad at first!" to make the media actively ignore any technological or economic problems, if not actively defend the technology by repeating these rationalizations like a cultist.
直到今天，关于 AI 泡沫的几乎任何经济问题都很难得到直截了当的答案——小到「跑一小时 GPU 要多少钱」「推理赚钱吗」「LLM 到底怎么才能盈利」「一家公司运营 GPU 或提供 AI 算力到底赚不赚钱」。为什么会这样？因为这些公司用「亏大钱是创新的必要代价」「新技术一开始都不好使」这类说辞，让媒体主动无视一切技术和经济问题——甚至像邪教徒一样复读这些说辞、主动为这项技术辩护。
Even those who are most loathsome in the defense of LLMs are a kind of victim of the AI industry, though a rather unsympathetic one. To become a full-blown "AI fan" requires you to accept effectively every narrative that you're given, herald every single announcement as proof that the prophecy will be fulfilled, ignore the financial realities and actively attack those who would dare to critique the great god of the Large Language Model. You have to know all the new terms, be excited about the right things at the right time, and live in near-constant fear that you'll fall behind on whatever it is you're meant to do next.
即便是那些为 LLM 辩护得最面目可憎的人，某种意义上也是 AI 行业的受害者——尽管是不太值得同情的那种。要成为一个百分百的「AI 粉丝」，你必须接受塞给你的几乎每一条叙事；把每一次官宣都当作「预言终将应验」的证据；无视财务现实；主动攻击那些胆敢批评大语言模型这尊大神的人。你必须掌握所有新术语，在正确的时间对正确的事情表现出兴奋，并且活在一种近乎持续的恐惧里：害怕自己在「接下来该做的事」上掉了队。
Your reward is that you can hang around a dwindling number of wealthy yet terrifyingly boring Silicon Valley intellectuals or kiss up to editors that would throw you in front of a bus if it meant getting access to a CEO, and maybe the odd Twitter psychopath who will defend you using a slur. In the end, many boosters will simply act as if they were never wrong. I hope they choose the more-courageous path of introspection, learning how they were had and using it as a weapon against con artists in the future. As strange as it sounds, I believe the most devout defenders of AI could become great critics in the future. Maybe I'm just being optimistic.
你得到的奖赏是：可以混迹于一小撮日益减少的、富有却无聊到可怕的硅谷「知识分子」中间；或者去巴结那些只要能接触到 CEO、会毫不犹豫地把你推到公交车底下的编辑；偶尔还有一两个推特疯子，会用污言秽语来「捍卫」你。到头来，许多鼓吹者会干脆装作自己从未错过。我希望他们选择那条更需要勇气的自省之路：搞清楚自己是怎么被骗的，并把它变成日后对付骗子的武器。说来奇怪，我相信最虔诚的 AI 捍卫者，未来可能成为最出色的批评者。也许我只是太乐观了。
腐烂经济的大疲惫The Great Exhaustion of the Rot Economy
Here's a very simple question: how much longer can everybody afford to keep doing this? Every single thing has become more expensive in the last year. Even though token prices have gone down or stayed flat, the amount of tokens you burn has clearly increased to the point that organizations are apparently spending billions of dollars on AI services with difficult-to-quantify ROI, requiring frantic advocacy to and financial debasement with every turn of the wheel. OpenAI and Anthropic have become more expensive to run, and OpenAI's non-GAAP operating margin increased from negative 122% to negative 183% in Q2 2026.
这里有个非常简单的问题：所有人还撑得起这样玩多久？过去一年里，每一样东西都变得更贵了。尽管 token 单价降了或者持平，但你烧掉的 token 数量明显增加了——结果是，各组织显然正在把数十亿美元花在 ROI 难以量化的 AI 服务上，每转一圈轮子都需要更狂热的鼓吹和更彻底的财务注水。OpenAI 和 Anthropic 的运营成本都更高了：OpenAI 的非 GAAP 运营利润率，已经从负 122% 扩大到 2026 年二季度的负 183%。
NVIDIA's GPUs just became 15% to 17% more expensive because high bandwidth memory costs doubled, a conga line of different monopolies upping their prices assuming that each link in the chain will keep spending, as each one of them — down to the AI labs themselves — knows that its contribution to spending on AI is an existential rite.
英伟达的 GPU 刚刚涨价 15% 到 17%，原因是高带宽内存（HBM）成本翻倍——这是一长串垄断者的康加舞，每一家都在涨价，赌定链条上的每一环都会继续花钱；因为他们中的每一个——直到 AI 实验室自己——都知道，自己在 AI 上的那份开支，是一场关乎存亡的仪式。
This means that any data center with GPUs delivered in 2027 and beyond will now have to cover billions of dollars' worth of extra costs, on top of increasingly-staunch local authorities requiring power guarantees ($100 million a year in Wisconsin for Oracle) and states like Illinois, Arizona and Virginia killing their tax breaks, all as interest rates spike and demand for AI debt weakens.
这意味着，任何在 2027 年及以后接收 GPU 的数据中心，都必须额外消化数十亿美元的新增成本——这还没算上：地方政府越来越强硬地要求电力担保（甲骨文在威斯康星州每年 1 亿美元）；伊利诺伊、亚利桑那、弗吉尼亚等州纷纷取消税收优惠；利率飙升；以及 AI 债务的需求正在减弱。
Every single year, every single part of the AI bubble becomes more expensive — AI labs want to spend more money, AI data centers cost more money, AI services become more expensive, AI debt becomes more expensive, and everybody becomes decidedly less-patient for there to be some sort of outcome.
Meanwhile, public relations expert and OpenAI CEO Sam Altman told podcaster David Senra that "we've all [referring to the AI industry] been too ambitious on timelines…[and that changing] people's behavior is much harder than the tech nerds realize." Sam: stop talking! Every time you open your mouth you say something silly!
与此同时，公关专家、OpenAI CEO 山姆·奥特曼对播客主持人大卫·森拉说：「我们（指 AI 行业）在时间表上都太激进了……改变人们的行为，比技术宅们意识到的要难得多。」山姆，别说了！你每开一次口，就会说出点蠢话来！
Anyway, here's everything that needs to happen in the next three-and-a-half years: OpenAI and Anthropic must keep spending as a means of justifying their existence to hyperscalers using their revenues to artificially inflate growth, to the tune of more than $440 billion across Google, Microsoft and Amazon alone. Hyperscalers must continue to buy NVIDIA GPUs, as the moment they stop doing so, the markets will begin to ask whether AI is an actual growth market anymore and ask for real, tangible answers about where all of this capex is going. To be specific, analysts expect NVIDIA to make $1.48 trillion in revenue across Fiscal Years 2027, 2028 and 2029.
NVIDIA must sign long-term agreements to buy high-bandwidth memory at scale from SK Hynix, Micron and Samsung — who make 90% of all DRAM — or know its costs would spiral out of control, by which I mean its margins would compress at random at a time when it’s already having to spike demand in extremely odd ways. As I've said, NVIDIA's price increase is going to increase the price of every single data center in construction by billions of dollars, and we're already approaching the limits of how much money can be raised for them. That "$500 billion" announcement was actually Jensen Huang jumping the gun, per Bloomberg: "Goldman Sachs Group Inc., Blackstone Inc. and Apollo Global Management Inc. had been working tirelessly for months to draw up debt deals that would help developers of artificial intelligence systems pay for chips from Nvidia Corp. With slow progress on the complex deals, Nvidia's chief executive officer, Jensen Huang, decided to change tack: He went public this week with the effort, saying the group is aiming to collectively finance AI computing deals totaling $500 billion — a round figure with no obvious provenance."
英伟达还必须与 SK 海力士、美光和三星（三家生产全球 90% 的 DRAM）签下大规模的 HBM 长期采购协议，否则它的成本将螺旋失控——我的意思是，在它已经不得不用各种极其古怪的方式刺激需求的当口，利润率还会随机被压缩。如前所述，英伟达这轮涨价，将让每一座在建数据中心的价格都上涨数十亿美元，而能为它们融到的资金已经接近极限。那笔「5000 亿美元」的官宣，其实是黄仁勋抢跑了。据彭博报道：「高盛集团、黑石集团和阿波罗全球管理公司数月来一直在不知疲倦地设计债务方案，以帮助 AI 系统开发商支付英伟达的芯片。由于这些复杂交易进展缓慢，英伟达 CEO 黄仁勋决定改变策略：他本周直接把这件事公开了，宣称该财团的目标是共同为总额 5000 亿美元的 AI 算力交易融资——一个来历不明的整数。」
The largest asset managers and financial institutions were making "slow progress," and that was before Jensen Huang increased prices by 15%. Do you think it'll become easier from here? How would that happen, exactly? God, I'm tired.
AI 正在耗尽它触碰的一切和每一个人AI Is Exhausting Everything and Everyone It Touches
The entire AI bubble has been exhausting for everybody involved. Because nothing works yet as a real business model or anything approaching truly autonomous (or "magical") software, there's the implicit knowledge that you're going to have to change your product again and again to update to the "best model" or "make things more efficient" (IE: lose less money) or when something breaks because a model's training got tweaked. The euphemism for this is "exponential improvement," when it's really an Arnold Palmer of instability and novelty, and abuses basically anyone connected to the ecosystem every single day.
整个 AI 泡沫，让每一个身在其中的人都筋疲力尽。由于还没有任何东西能以「真实商业模式」的形态运转、或接近真正自主（或「神奇」）的软件，每个人都心照不宣：你将不得不一遍又一遍地改自己的产品——为了追上「最好的模型」，为了「让东西更高效」（翻译：少亏点钱），或者因为某个模型的训练被调整、你的东西突然崩了。这一切的委婉说法叫「指数级改进」，而实际上它是一杯「不稳定」兑「新奇」的 Arnold Palmer 特调，每天都在折磨着连接到这个生态里的几乎每一个人。
If there's always something new happening, it's hard to pin down if things have gotten better, or whether you're just more proficient in cobbling together different harnesses, prompts and API calls to make it do what you need it to. It is undignified that people tolerate models that become either dumber over time or at random opportunities, while also being deeply exhausting for the end user. As a paying user of an LLM-powered service, you are guaranteed at some point to face a degradation in service where models misbehave, some sort of shift in rate limits, or some sort of change in product functionality based on their shifting economics.
如果总有新事情发生，你很难判断一切是变好了，还是你只是更擅长拼凑各种框架、提示词和 API 调用来让它按你的需要工作。人们容忍模型随时间、或随随机机会变笨，这毫无尊严可言；而对最终用户来说，这也令人精疲力尽。作为 LLM 服务的付费用户，你注定会在某个时刻遭遇服务降级：模型行为异常、速率限制调整，或者厂商基于自身经济账的变化随意改动产品功能。
Has there ever been a bigger shift in a business product's value than GitHub Copilot's shift to token-based billing? Microsoft rug pulled two million people that had built workflows on a platform that was allowing them to burn $1,000 to $5,000 in tokens for $20 a month. That's genuinely crazy! It's magnitudes more than when Uber jacked up its prices. It's equally-insane that Anthropic and OpenAI similarly fuck with their customers, changing the amount of value you get for $20, $100, or $200 a month at random in a way that shouldn't be legal.
Basically any AI-powered software is subject to arbitrary shifts in availability, capability and pricing at the whims of the vendor. As I covered in my Subprime AI Crisis piece earlier in the year, Replit, Perplexity, and multiple other AI companies have sold their customers a lie by pushing an unprofitable product that they must constantly "tweak" to bring down costs, all while misleading the customer about a "price" that continually declines in value as the price stays the same.
基本上，任何 AI 驱动的软件，其可用性、能力和定价都随厂商的心血来潮任意变动。正如我今年早些时候在《AI 次贷危機》（Subprime AI Crisis）一文中所写的：Replit、Perplexity 和多家 AI 公司都在向客户兜售一个谎言——它们推着一个不盈利的产品，必须不断「调整」来压低成本，同时在「价格」上误导客户：标价不变，但你能买到的价值一直在缩水。
This is not a sustainable industry — either economically or emotionally — because it has a fundamentally dishonest relationship with its customers defined by the inconsistency of LLMs both in efficacy, stability (see: Anthropic's downtime) and training, with each model randomly better or worse at things to the point that it must be a legitimate nightmare to run any software or build any product on top of them. And the fact they haven't worked out their business models means that whatever you're paying today is guaranteed to change. What other product do you regularly buy that has such chaos built into it? What other thing do you pay for where the prices (or availability) can shift to the point that you literally can't use it in the same way at a moment's notice? And why does anybody tolerate it when it comes to AI?
I'll add that this is a specific situation where the tech media has categorically failed the customer. We have companies valued at hundreds of billions of dollars that are fucking their customers over day-in-day-out, and the response is mostly to say "huh that's strange" and refuse to let a single critical thought cross their minds.
AI 泡沫要求所有人活在将来时，因为当下丝毫无法证明其成本的合理性The AI Bubble Requires Everybody To Live In The Future Tense, Because The Present In No Way Justifies An Iota Of Its Costs
Every part of the AI bubble must exist in a constant state of flux so that there can always be a future breakthrough that's always just out of reach. AI does not have to reach an actual achievement — it just has to "show promise" in some way. It is an objective disaster that Microsoft spent more than $260 billion on capex to create a business with less than $11 billion in annual revenue outside of OpenAI, but people will see "$34.33 billion in annual AI revenue" and say "that's promising growth, up 123% year-over-year!"
AI 泡沫的每个部分都必须活在永恒的变化之中，这样才能永远有一个「触手可及却又差一步」的未来突破。AI 不需要真正取得什么成就——它只需要以某种方式「展示前景」。微软花掉 2600 多亿美元资本开支，换来一个除 OpenAI 之外年收入不足 110 亿美元的业务，这在客观上是一场灾难；但人们看到「343.3 亿美元年度 AI 收入」，就会说：「增长喜人，同比涨了 123%！」
They'll hear about LLMs that delete people's databases and say "well the models have gotten exponentially better," even if that better part never seems to eliminate these issues, make a profitable AI company, or create a true killer app that you can point at beyond saying "ChatGPT has one billion weekly active users," despite around 95% of them not paying a penny (and costing OpenAI likely billions of dollars) and eMarketer estimating that the entire global AI chatbot advertising industry will make $5.41 billion revenue in 2030, giving OpenAI little hope of stemming the burn. These big numbers — like Anthropic having a $65 billion annualized run rate, an undefined term that obfuscates the fact that Anthropic has made $16.5 billion in the first half of 2026, losing billions of dollars in the process — are fundamentally meaningless, because they're easily gamed at best, and inherently uncertain at worst.
The AI industry demands you constantly live in the future tense. Everything is about tomorrow's billions or trillions, the potential of what you're seeing rather than the thing itself, future gigawatts in data centers that you must treat as if they are already built and value based on things that AI might theoretically do. I challenge you to read everything about AI from this point forward with this in your mind so you can see how intently this industry tries to drag your focus away from what it's doing toward what it might theoretically do if it only had more money, power and resources, and ask yourself why they need to do so.
AI 行业要求你永远活在将来时。一切都关乎明天的几十亿、几万亿，关乎你眼前之物的「潜力」而不是它本身，关乎数据中心里那些「未来的吉瓦」——你必须当作它们已经建成，并按「AI 理论上也许能做的事」来估值。我向你发出挑战：从今天起，带着这个意识去读关于 AI 的一切报道，看看这个行业多么用力地把你的注意力从「它正在做什么」拖向「它只要有更多钱、更多电力、更多资源，理论上也许能做什么」——然后问问自己：他们为什么需要这样做？
To be clear, they're doing so because you can't really justify anything about this industry based on what it does today. It costs too much, none of the businesses built on top of it are profitable, it costs so much to build a data center that the most cash-rich asset-light businesses in the world are now burdened with endless expensive-to-install and run hardware for a business that makes a fraction of its overall costs in revenue and has little demand outside of two companies that everybody must conspire to keep alive both financially and philosophically. And ultimately, nobody can actually explain why we need more data centers.
Would anything really change? What would change? How? How many more do we need? Why do we need so many? Having more power plants meant more people could have power, and having more fiber laid meant connecting more buildings to the internet. What does one more or two more or ten more data centers actually give you? Is there some part of the world unable to access or take advantage of the LLMs available on seemingly every surface of the internet? Because it seems like the only reason these things are getting built is to capture illusory demand based on a "supply constraint" created by two unprofitable companies absorbing all the infrastructure. I don't hear any compelling scientific or technological reason building more is useful or productive outside of funneling more cash to semiconductor companies.
Seriously, go and read basically any article about AI and see how quickly they start talking about the future, be it in the mainstream media or on a startup's blog. Every single piece must sell AI on its theoretical promise and, if at all critical, reassure you that the author of course doesn't dispute the "transformative potential of AI" or "how it's already transforming the economy," even if it can't define how it's doing so or even what that means.
说真的，随便找一篇关于 AI 的文章来读——主流媒体的也好，创业公司的博客也好——数一数它用几段话就开始大谈未来。每一篇文章都必须靠「理论上的前景」来推销 AI；而哪怕文章带点批评，也必定向你保证：作者当然不否认「AI 的变革性潜力」、不否认「它已经在变革经济」——尽管它说不清 AI 到底怎么变革的，甚至说不清那是什么意思。
AI 行业靠「揣着明白装糊涂」运转The AI Industry Runs On Bad Faith
I let myself have a little fun with today's piece because I feel like I've been so deep in the financial trenches that I forgot how much of the AI industry runs on propaganda, social pressure and outright bullying to manufacture consent for a product that demands everything and provides very little in return. Nothing about LLMs is worth a trillion dollars, or even $100 billion. This is, as I've said before, a $30 billion TAM industry dressed up as a trillion dollar one, and the only reason it's grown this large is because the two leading companies have had their infrastructure built for them and given unlimited resources to subsidize their customers' compute.
今天这篇文章我让自己放纵了一下，因为我感觉自己在财务的战壕里蹲得太久，差点忘了 AI 行业在多大程度上是靠宣传、社会压力和赤裸裸的霸凌运转的——为的就是给一个「索取一切、回报寥寥」的产品制造共识。LLM 没有任何一部分值一万亿美元，甚至不值 1000 亿。正如我以前说过的：这是一个 300 亿美元 TAM（潜在市场规模）的行业，被打扮成了万亿美元的样子；它能长这么大，唯一的原因是两家头部公司有人替它们建好了基础设施，并给了它们无限的资源去补贴客户的算力。
And what's really stood out is how so little about the AI bubble is actually about AI. No other technology in history has had professional and social consequences for failing to use or like it enough, nor can I find any example in history where journalists have actively attacked critics for not being sufficiently-approving of a kind of cloud software. It is fundamentally crazy to me that, in pursuit of "objectivity," much of the tech and business media has chosen to accept whatever narrative the AI industry gave them, assuming that whatever we have today is already guaranteed to be something better in the future, both in its outcomes and profitability.
而最扎眼的一点是：AI 泡沫里真正关于 AI 的东西，少得可怜。历史上没有任何其他技术，会因为你「用得不够、爱得不够」而带来职业和社会层面的后果；我也找不到历史上任何先例：记者会主动攻击批评者，理由是他们对某种云软件「不够认可」。在我看来，这从根本上就是疯了：为了追求「客观」，大量科技和商业媒体选择了照单全收 AI 行业塞给他们的叙事，默认我们今天拥有的一切「在未来注定会变得更好」——无论是效果还是盈利能力。
This era is unlike any other before it, but took advantage of the fact that most people are desperate to apply the past to the present to rationalize or process what may seem irrational or destructive. To see AI as "just like the dot com bubble" allows you to ignore both the costs and the potential outcomes because "things worked out after that," even if there're basically no uses for GPUs after this and the only way we "build new LLMs" is by feeding them expensive training data using billions of dollars of compute that are only available while everybody still believes this is real.
这个时代与此前的任何时代都不同，但它利用了人们的一种渴望：大多数人迫切地想把过去套用在当下，好把看似非理性或破坏性的东释合理化。把 AI 看作「不过是又一个互联网泡沫」，你就可以同时无视它的代价和可能的结局，因为「当年后来不也挺好」——尽管这波之后 GPU 基本没有别的用途，而我们「构建新 LLM」的唯一方式，是用数十亿美元的算力去喂昂贵的训练数据，而这种算力只在「所有人还相信这是真的」期间才存在。
The AI industry — and the AI bubble — is fundamentally built on acting in bad faith. Its executives lie. Its boosters lie. Its software lies because it doesn't actually know anything and generates answers probabilistically, and if you mention that online, someone will harass you for doing so.
AI 行业——以及 AI 泡沫——从根本上建立在「揣着明白装糊涂」（bad faith）之上。它的高管齐撒谎。它的鼓吹者撒谎。它的软件也撒谎，因为它实际上什么都不知道，只是按概率生成答案——而你要是在网上指出这一点，就会有人来骚扰你。
It refuses to answer straightforward questions. It refuses to present a plan for the future. It refuses to explain how it becomes profitable, because nobody knows how or has a tangible plan to do so. It deliberately subsidized its subscription products because it knew its customers wouldn't pay the actual cost of AI, and tortures customers with shifts in functionality and rate limits all while framing this as a way to "continue to serve customers the most cost-efficient models."
它拒绝回答直接的问题，拒绝拿出面向未来的计划，拒绝解释自己将如何盈利——因为没有人知道，也没有任何切实的方案。它故意补贴自己的订阅产品，因为它知道客户不会为 AI 的真实成本买单；它用功能变更和速率限制折磨客户，还把这包装成「继续为客户提供最具成本效益的模型」。
It attempts to conflate massive, power and resource-hungry AI data centers with the smaller ones that bring helpful yet increasingly-decaying software to our homes. It sells these data centers as "bringing jobs to communities," all while importing the talent from out of state to build the things then leaving a crew of 100 to 200 people to actually run them after millions or billions of dollars of tax breaks. It sells its "innovations" as creating a "white collar bloodbath" to scare you into using inconsistent and unreliable software that's mathematically certain to make mistakes, and when you say something about it, its acolytes will lie and say that "hallucinations are solved."
它试图把庞大的、吞噬电力和资源的 AI 数据中心，和那些把（日益朽坏的）软件送进千家万户的小型数据中心混为一谈。它把这些数据中心包装成「为社区带来就业」，实际上却是从外州引进人才搞建设，拿走数亿、数十亿美元的税收减免之后，只留下一支 100 到 200 人的团队真正运营。它把自己的「创新」包装成制造「白领大屠杀」，吓唬你去使用那些不稳定、不可靠、数学上注定犯错的软件；而当你指出这一点时，它的信徒会撒谎说「幻觉问题已经解决了」。
It also can only ever sell itself based on what might happen and the theoretical promise of you giving it your complete attention, connecting every bit of data you own, paying whatever it costs, and accepting that it can and will change in price and functionality at random, all while never putting a precise timeline on whatever AGI means that particular week. Whenever you ask for clarity, the AI industry gives you chaff. Whenever you ask when things get better, you're told it's both the early days and that AI is the worst it'll ever be.
它推销自己的方式，永远只能基于「可能发生的事」和「理论上的前景」：你要交出全部的注意力，接入你拥有的每一比特数据，不论多少钱都照付，并接受它随时会随机改变价格和功能——而它对「AGI」（那一周的含义）永远不会给出精确的时间表。每当你要求说清楚，AI 行业就撒出一把干扰箔。每当你问什么时候会变好，你得到的回答既是「现在还是早期」，又是「今天的 AI 是它最差的样子」。
Even the term "artificial intelligence" is a bad faith attempt to conflate transformer models with things like robotics or autonomous cars, all so that its proponents can claim other people's successes as their own despite LLMs having little or no relevance to anything else other than generative AI. It encourages dogpiling and ostracizing those who don't fall behind it, because it cannot succeed on its own merits. It encourages a vile cultism powered too by bad faith and parasocial relationships with both AI CEOs and the models themselves.
甚至连「人工智能」这个词本身都是一次「揣着明白装糊涂」的尝试：把 Transformer 模型和机器人、自动驾驶汽车混为一谈，好让它的鼓吹者把别人的成功据为己有——尽管 LLM 与生成式 AI 之外的任何东西几乎没有关系。它鼓励围攻、鼓励排挤不跟随它的人，因为它无法凭自身的优点取胜。它助长一种卑劣的邪教氛围，动力同样来自揣着明白装糊涂，以及信众与 AI CEO、与模型本身之间的准社会关系。
It exploits the intellectual weaknesses of "smart people" that are actually just good at remembering the right things to say at the right time and have memorized the various justifications for past failures, all while allowing them to use LLMs to promote their own bad faith enterprises where they use work-adjacent product to con others into paying them. And it's losing because, at its core, AI was never built on very much. It grew this large because the media manufactured consent at the behest of the powerful because lots of money got invested, and the rich and powerful can never be wrong.
The underlying technology may be more useful than it was, but it's not useful enough to be profitable nor reliable enough to be world-changing, and the bad faith representation of LLMs as "good enough" should be a permanent scarlet letter on anyone who misled the public into believing this was anything other than normal software.
I was asked recently why I find this all so repugnant, and my answer is simple: I don't like bullies, I don't like con artists, and I don't like being lied to. This industry grew by misleading people about the actual and potential outcomes from Large Language Models, and through an economy-wide attempt to pressure everybody into adopting tools in pursuit of growth at all costs. Ultimately, it was sold with the greatest lie of all: "this time it's different!" To be clear, they're right. It's so much weirder, and in the end will be so much worse.