Memory, Money, and the Limits of Expectation

Original原始 by Chaping Jun, 2026-10-28 00:00, Zhejiang原创 差评君 2026-10-28 00:00 浙江

What什么 should应该 an Agent智能体 that can能够 actually get things done look like?能管事的Agent应该长啥样？

Is the personal Agent智能体 really a false proposition命题?个人 Agent 到底是不是一个伪命题？

Recently, there has been quite a lot of discussion about this.最近一段时间，关于这个事儿的讨论还挺多的。

Optimists乐观者 think that the next step for AI is of course to take over接管 our lives, ideally理想情况下 so that we don't have to worry about anything ourselves. But the other camp is less polite, and has outright直接 labeled贴标签 the personal Agent智能体 a false need.看好的人觉得，AI 下一步当然是承包咱们的生活，最好啥都不用自己操心。但另外一拨人就没那么客气了，愣是给个人 Agent 扣上了伪需求的帽子。

But in Chaping Jun's view观点, rather than arguing争论 over接管 whether it's a false need, it's better to think carefully about what什么 kind of personal Agent智能体 is actually painless无痛 to use.但在差评君看来，与其争论是不是伪需求，不如好好想一想，什么样的个人 Agent 用起来才不烧心？

After all, AI being able to do work is nothing new anymore; it can能够 plan steps on its own, call tools to handle处理 complex复杂 tasks, and operating software软件 isn't hard困难 either.毕竟 AI 能干活已经不是什么新鲜的事儿了，会自己规划步骤、调用工具来处理复杂任务，操作软件也不难。

But what什么 the personal Agent智能体 wants to do goes far beyond超出 that.但个人 Agent 想做的，可不止这些。

Take MuseMuse as an example. Getting it to book a tennis lesson is nothing surprising, but if the tennis lesson is suddenly canceled被取消, it will proactively主动地 remind you and then adjust调整 your schedule日程 based on your arrangements安排, see whether there are other suitable times, and rebook重新预订.以 Muse 为例，让它去预约网球课没啥稀奇，但网球课临时取消了，它会主动提醒你然后根据你的安排调整行程，看看有没有其他合适的时间，再重新预约。

What什么 the personal Agent智能体 needs to solve is how to turn AI from a tool that does work on command指令 into an assistant that can能够 help you manage long-term长期 complex复杂 affairs. The key is being able to remember what什么 was previously先前 assigned已分配, proactively主动地 follow跟进 up, and know what什么 to do when changes come up.个人 Agent 要解决的，是怎么让 AI 从一个听令干活的工具，变成能够帮你打理长期复杂事务的助理。关键是能记住之前交代过的事，会主动跟进，碰上变化也知道该怎么办。

Another example: if you want to buy cheap plane tickets机票, you have to go into the app and check查看 from time to time, and also repeatedly switch切换 between之间 different platforms平台 to compare比较 prices. So can能够 this kind of thing be handed over接管 to an Agent智能体?再比如你想买便宜机票，时不时就得进 App 里刷一刷，还要在不同平台之间反复切换比价，那这种事情能不能交给 Agent 来做？

If you hand it over接管 to a proactive主动的 personal Agent智能体, you only need to say once, "Keep an eye on tickets机票 from Beijing to Chengdu for me," and it will go check查看 on its own, compare比较, proactively主动地 notify通知 you when the price is right, and even甚至 string串联 price comparison比较 and ordering together into one flow.如果交给一个有主动性的个人 Agent，你只需要说一次“帮我盯着北京飞成都的票”，它就会自己去查、去比、价格合适了主动通知你，甚至帮你把比价和下单串成一条线。

Of course, we can't just keep firing off opinions into the air here. Just yesterday, the Lingguang App released发布 version版本 2.0, said to be centered on a personal AI assistant that can能够 actually manage things.当然了，咱们也不能在这儿对着空气一顿输出，正好昨天灵光 App 发布了 2.0 版本，说是主打一个能管事的个人 AI 助理。

The direction of this update更新 happens to coincide相巧合 with the thinking behind the personal Agent智能体: remember you, follow跟进 up with you, and proactively主动地 help push things forward推进 continuously, somewhat like a Chinese version版本 of MuseMuse.这次的更新方向倒是跟个人 Agent 的思路不谋而合，记住你、跟进你、主动帮你把事儿往前一直推，有点中国版 Muse 的意思。

So Chaping Jun got hands-on and tried it for a few days right away. Whether it can能够 actually manage things well, where it really saves effort, and where it still falls变成，成为，陷入 short, let's take a look together.所以差评君第一时间上手体验了几天。它到底能不能把事情管好，哪些地方确实省心，又有哪些地方还不太行，咱们一块儿来看看。

After the upgrade升级, Lingguang 2.0 has a "My Services" section部分, divided into two categories种类: application应用 services and system services.升级之后的灵光 2.0，有一个“我的服务”板块，分成了应用服务和系统服务两类。

In the app services, you can能够 currently目前 see Alipay, QQ Mail, NetEase Mail, and the fitness健身 tracking追踪 app Xunji. On the system services side, it includes Calendar, Reminders, Alarm闹钟, Health, and other built-in phone features功能.应用服务里，目前能看到支付宝、QQ 邮箱、网易邮箱，还有健身记录 App 训记。系统服务这边，则是包括了日历、提醒事项、闹钟和健康等等手机自带的功能。

Differential差异化 Review tried it out: when you converse交谈 n反面说法 a相反的 with Lingguang in the dialog box（法庭的）专席, it can能够 connect to DingTalk, Mijia, Didi Chuxing, Feishu, and many other external外部 tools and services.差评君试了下，你在对话框与灵光对话，可以连接到钉钉、米家、滴滴出行、飞书等众多外部工具和服务。

这些服务有啥用呢？

For example, after connecting to Alipay, you can能够 ask Lingguang what什么 deals交易 are available可获得的，可利用的 on Alipay to take advantage优点， 优势 of.就比如连接支付宝之后，你可以问灵光支付宝有哪些羊毛可以薅。

Behind it, it is also linked to a whole set设置 of commonly used services such as Alipay bills, Huabei, Jiebei, membership会员, and travel. You can能够 have it go through通过 bills, uncover automatic自动 renewals续费, see where spending goes, and keep an eye on flight航班 and hotel deals交易 and travel discounts折扣, and so on.它后面还连着支付宝账单、花呗、借呗、会员、出行等一整套常用服务，你可以让它去翻账单、扒自动续费、看消费花在哪、盯机酒和出行优惠等等。

A lot of information that was originally最初 scattered撒， 散播 across different pages has been linked together and processed过程 as a whole by it.很多原本散在不同页面里的信息，被它串起来一块处理了。

However, Differential Review still wants to put Lingguang under more pressure大气压力: if it is given规定的，特定的 a task that lasts a long time and needs to be followed从事（职业），经营 up every day, can能够 it still manage it well?不过，差评君还是想给灵光上上强度，找个时间长还天天都需要跟进的活儿，它还能管好吗？

We are not forcing强迫 a need either; let's start开动，着手 from real life.咱们也不硬造需求，就从真实的生活下手。

To be honest, Differential Review has a friend who has been clamoring喧闹 for two years about learning Spanish, but after all that studying, he still only knows the pronunciation of a few letters, no different from memorizing记忆 English vocabulary词汇 and always getting stuck at "abandon放弃."说来惭愧，差评君有一个朋友已经嚷嚷了两年要学西班牙语，结果学来学去，还是只会那几个字母发音，跟背英语单词永远停留在 abandon 没啥区别。

So we simply had Lingguang help arrange安排 a three-month三个月 beginner Spanish learning plan.所以干脆让灵光帮忙安排了一份三个月入门西班牙语的学习计划。

And would you believe it, it arranged已安排 things quite meticulously仔细地,谨慎地: for the three-month三个月 course, the first two weeks are for learning pronunciation, and once that is more or less done, it advances to regular and irregular verbs. Along the way方式, it also intersperses dialogues from common scenarios剧情说明书，剧本, and in the end目标，目的 it checks检查 whether an A1 level can能够 be reached.你别说，它安排得还挺细致，三个月的课程，前两周先学发音，学得差不多了再进阶规则动词和不规则动词，这个过程中还会穿插一些常见场景的对话，最后看能不能达到 A1 的水平。

I did not give any specific具体 requirements either, yet it arranged已安排 the timing and content满意的 for me.我也没提什么具体要求，时间点和内容都帮我排好了。

Down to the daily schedule日程, 40 minutes a day, it proactively主动地 reminds me to study at 20:30 in the evening晚上 and 10:30 in the morning. The evening晚上 session时段 covers遮盖，遮蔽 new knowledge, and after a night's sleep, I spend another 5 to 10 minutes reviewing and consolidating巩固.具体到每天的安排，每天 40 分钟，晚上 20:30 和早上 10:30 主动提醒我学习，晚课安排新知识，睡一觉起来再花 5 - 10 分钟复习巩固。

After connecting to the phone alarm闹钟, five minutes before the morning study session时段 every day, Lingguang uses the alarm闹钟 to remind me so I do not forget, and I no longer need to set设置 the alarm闹钟 manually手工的.连接了手机闹钟之后，每天早自习前五分钟灵光都会用闹铃提醒我，防止我忘了，也不需要我再手动去定闹钟。

And after some tinkering做焊锅匠, I also had Lingguang specially make a "Spanish Daily Learning Card" app. What什么 to learn and what什么 to practice做法 each day is all in there, including key explanations, word shadowing, listening discrimination歧视 mini-games, plus post-lesson assessments评估, stringing线， 细绳 the whole learning process过程 into one line台词.并且经过一番折腾，我还让灵光专门做了个“西班牙语每日学习卡”应用，每天要学什么内容、练什么都在里面，包括重点讲解、单词跟读、听辨小游戏，再加上课后考核，把整个学习流程串成了一条线。

For the first lesson's assessment评估，评价, although Differential Review answered符合，适合 all five questions问，询问，讯问 correctly正确地，恰当地, I still honestly wrote one line台词 in the self-evaluation: I cannot不能 remember many example words.第一节课的考核，差评君虽说五道题全答对了，但还是老老实实在自评里写了一句：很多例词记不住。

I originally最初 thought this was just a casual偶然的 piece of feedback, but I did not expect料想，认为 Lingguang to actually take it seriously.本来以为这就是随手填的一句反馈，没想到灵光还真当回事儿了。

That evening晚上, it took the initiative主动性 to do a review, listing my weak points one by one. When I reviewed the next morning, it also proactively主动地 brought out the example words I hadn't remembered firmly the day before and went over接管 them again, because I had misremembered the word for "table" and it caught卷入 me red-handed again.当晚它就主动做了复盘，把我的薄弱项一条条列了出来。第二早复习的时候，它还主动把昨天我没记牢的例词拿出来重新过了一遍，因为我把“桌子”这个单词记错了，又被它逮了个正着。

Put it this way方式: what什么 I've learned, how the assessments评估 went, and what什么 difficulties困难 I've reported报告—every time I chat with Lingguang, it picks up right where we left off.这么说吧，学过什么、考核结果如何以及我反馈过什么难点，每次聊灵光都能接上进度。

It doesn't just remember—based on these learning records, it also arranges安排， 准备 and adjusts适应 vt. 调整，调节 subsequent尔后的 lessons and review priorities优先事项. This ability to adjust调整 the course on its own based on feedback is the biggest difference between之间 it and ordinary AI chat.它不光记得住，还会根据这些学习记录，安排、调整后续的课程和复习要点。这种根据反馈自己调整课程的能力，才是它跟普通 AI 聊天最大的区别。

And these learning records don't just stay in the Lingguang chat window.而且这些学习记录，不只是留在灵光的对话框里。

Every morning before self-study, based on previous之前的 class performance演出, it automatically自动地 compiles汇编 a summary概括的 of key knowledge points and review priorities优先事项 into my Feishu document. I don't have to scroll卷轴 through通过 chat history every day to find what什么 I learned yesterday, nor do I have to take notes all over接管 again myself.每天早自习前，它还会根据之前的上课表现，把一份知识点汇总和复习要点，自动整理到我的飞书文档里。我不用每天翻聊天记录找昨天学了什么，也不用自己再做一遍笔记。

Mark this: I can能够 even甚至 schedule日程 Lingguang to connect to a printer打印机 and print out the organized PDF.划重点，我甚至可以定时让灵光连接打印机，把整理好的 PDF 打印出来。

If you have a Mijia printer打印机, it can能够 connect directly once you install安装， 设置 the API; if your printer打印机 isn't a Mijia one, a little DIY will do the trick骗局，诡计 too.如果你手里有米家的打印机，让它安装 API 就能直接连上，如果打印机不是米家的，稍微 DIY 一下也行。

At least from these past few days of experience, Lingguang doesn't just hand me a study plan—it feels摸起来 like there's a real teacher supervising监督，监视 my learning.至少从这几天的体验来看，灵光不单单是给我一张学习计划表，像是真有个老师在监督我学。

I also discovered that it can能够 string串联 together information from different conversations.我还发现，它能把不同对话里的信息全都串起来。

I had Lingguang arrange安排 a long-term长期 fitness健身 plan, mentioned提到 in passing that I'd fractured破裂 my right ankle脚踝, and in another conversation planned计划的 a four-day trip to Chengdu.让灵光安排过一个长期的健身计划，顺嘴提了自己右脚踝骨折的事，又在另一个对话里规划了成都四日游。

Then one morning, it not only reminded me to keep an eye on my ankle脚踝 but also remembered that I was about to go to Chengdu.结果有天早上，它不光叮嘱我留意脚踝，还记得我马上要去成都。

I hadn't even甚至 mentioned提到 canceling取消 the training, but it went ahead and canceled被取消 all the arrangements安排 for the next few days on its own.我都没提取消训练的事儿，它就自个儿把后面几天的安排全都取消了。

Of course, for things like studying and fitness健身, at least you have a clear idea of what什么 to do each day.当然了，学习、健身这些事儿，至少每天要干什么还是有数的。

But some things are troublesome困难的 precisely精确的， 准确的 because you have no idea when you should应该 deal处理 with them.但有一些事，麻烦就麻烦在你根本不知道该什么时候去管。

During Double Eleven, you wait for price drops—buy early（稻麦等的）穗 and you fear paying too much, buy late去世不久的 and you fear missing the discount折扣 v. 打折扣卖. You can't open shopping apps eight hundred times a day just to save a few dozen yuan, right? What's more, more than one app is running promotions促销.双十一蹲降价，早买怕买贵，晚买又怕错过优惠。总不能为了省几十块钱，一天打开购物 App 八百遍吧？更何况，搞促销的，还不止一个 App。。

So I decided to hand the bargain-hunting task over接管 to Lingguang as well.所以蹲好价这个事儿，我也决定交给灵光来干。

Headphones耳机, perfume香水, cat food, and a humidifier—check查看 the prices on each shopping platform平台 every two hours, and remind me when there's a good deal处理.耳机、香水、猫粮和加湿器，每隔两个小时核查一次各个购物平台的价格，有合适的再提醒我。

As for me, I sleep when it's time to sleep and work when it's time to work—once Lingguang spots点 a good price, I'll just place the order.至于我嘛，该睡觉睡觉，该上班上班，灵光蹲到好价了再去下单就行。

After watching for two days, for things like headphones耳机 and cat food, it advised me to buy now—with subsidies补助 and coupons优惠券 stacked堆，垛 vt. 堆积， 堆放 on Tmall, the headphones耳机 could come to around 680–700 yuan.盯了两天，像耳机和猫粮，它劝我现在就入手，耳机在天猫叠加补贴和优惠券之后，到手价能做到 680～700 元左右。

I didn't believe it, so I checked已核对 the latest bargain讨价还价 n. 特价商品 guides on Xiaohongshu, and sure enough, it really was in that range排列.我不信邪，去小红书上查了最新的好价攻略，结果还真在这个区间。

There's also perfume香水. The reminder threshold门槛 it initially最初，开始 set设置 was 1,114 yuan, but after rechecking historical prices across other channels, it found that this standard was set设置 too high and adjusted适应 vt. 调整，调节 the reminder target目标 v. 把…作为目标 itself to within 680 yuan.还有香水，它最开始设置的提醒线是 1114 元，后来重新核对了其他渠道的历史价格，发现这个标准定得太高，又自己把提醒目标调到了 680 元以内。

Besides prices, it also keeps an eye out for me on what什么 coupons优惠券 and red packets can能够 be claimed声称.除了价格，有什么优惠券、红包可以领，灵光也帮我盯着。

One morning, it proactively主动地 reminded me that the second支持 wave of JD red packets had opened for claiming声称, and along with that told me which products they could be used for, when they would expire过期, and that there would be another boosted提高 round of activity in the evening晚上.有天早上，就主动提醒我京东第二波红包已经开领了，顺带告诉我哪些商品能用、什么时候过期，以及晚上还有一波加码活动。

In addition加， 加法, this time I also wanted to see whether, with the help of external外部 tools, Lingguang could let me open fewer apps.另外，我这次还想看看灵光如果借助外部工具，能不能让我少打开几个 App？

I had it arrange安排 a two-day trip to Qiandao Lake this weekend. Being able to make an itinerary行程 is nothing unusual. But after connecting to the 12306 MCP, Lingguang can能够 check查看 specific具体 high-speed rail trains, ticket机票 prices, and remaining剩下，余留，保持 tickets机票, and with the Amap interface分界面 connected in the background出身背景， 经历, I don't have to switch切换 back and forth among several apps for what什么 to eat or where to stay.让它安排了这周末去千岛湖的两日游，会做攻略不稀奇。但接上 12306 的 MCP 之后，灵光可以查询具体的高铁车次、票价和余票情况，后台连了高德的接口，吃什么、住哪里，不用我在几个 App 里来回切换。

From time to time, it also keeps an eye on Sunday's weather for me, and if it rains too heavily, it will replace替代 the original原始 orange-picking摘橙子 itinerary行程 with indoor室内 activities.它时不时还会帮我盯着周日的天气，如果雨下得太大，就把原先的采橘行程换成室内项目。

But what什么 I found more pleasantly surprising was probably not this itinerary行程.但我觉得比较惊喜的，可能还不是这份攻略。

Because I was thinking of taking another day off next Monday, I let Lingguang try to see if it could help me request the leave, and as a result, my manager received the application应用 instantly立即的.因为想着下周一再请天假，就让灵光试试能不能帮我把假请了，结果就是领导那边秒接到申请。

I didn't have to go into Feishu myself to fill填写 out the form, select the person, and submit提交 it — it handled已处理 it on its own.我不用自己去飞书里填表、选人、提交——它自己搞定了。

A colleague同事， 同僚, after connecting Lingguang to Feishu and QQ Mail, solved the headache of organizing invoices发票 for travel reimbursements返还费用.有同事用灵光连接飞书和 QQ 邮箱以后，把差旅报销让人头疼的发票整理都给解决了。

First, it dug out the invoices发票 from August to October from the mailbox and organized them into a Feishu spreadsheet表格, then checked已核对 against the Alipay bill to see what什么 was missing. It ran back and forth across three applications申请 the whole time without any manual手动 operation操作.先在邮箱里翻出 8 到 10 月的发票，整理进飞书表格之后，再去对一遍支付宝账单看漏了啥。跨三个应用来回跑，全程没让人手动操作。

All five invoices发票 were registered登记， 注册 in the spreadsheet表格, and the itineraries旅程，路线 were also checked已核对 to confirm证实， 肯定 there were no overlaps重叠.5 张发票全登记进了表格，还核对了行程，确认没有重叠。

The spreadsheet表格 also had a "pending invoicing发票，发货清单" page, which was the meal and travel expenses花费， 消费， 费用 Lingguang analyzed from the Alipay bill to ensure no invoices发票 would be missed. This is the kind of thing you wouldn't think of yourself, and it proactively主动地 filled担任（职务） it in for you.表格还有一页“待开票”，是灵光从支付宝账单里分析出的吃饭、出行消费，确保不会漏开发票。这种你自己想不到的事，它都主动给你补上了。

Anyway, after using it for the past few days, Chapingjun found that there really are quite a few places in daily life where Lingguang can能够 be used.反正这几天用下来，差评君发现生活中确实有不少地方能用上灵光。

If something in the fridge is about to expire过期, have it remind me to eat it soon; give it my watchlist自选列表 stocks股票 and have it keep an eye on the market市场 and send me any changes at any time; have it remind me of anniversaries纪念日, buy gifts in advance, and send greetings on time, so I can能够 avoid避免 a scolding...冰箱里有什么东西快过期了，让它提醒我赶紧吃；把自选股丢给它，让它帮忙盯盘，有什么行情变化随时发过来；纪念日让它来提醒，提前买礼物、准时送祝福，可以少挨一顿骂。。。

Of course, there are shortcomings too. For example, with Double 11 price comparisons比较， 对照, it can't check查看 some real-time实时 prices in Douyin livestream直播 rooms or some Pinduoduo prices, the information updates更新 for market市场 watching aren't very timely及时的, and the setup of the Spanish course is not systematic成体系的 enough.当然，短板也是有的，像双十一比价，抖音直播间的实时价格和拼多多的价格有些它查不到，盯盘的信息更新不算很及时，包括西班牙语课程的设置，也不够体系化。

Overall总的, this attempt尝试， 试图， 努力 by Lingguang 2.0 showed us the value尊重，重视 of a personal Agent智能体.总的来说，灵光 2.0 这次的尝试让我们看到了个人 Agent 的价值。

An Agent智能体 that can能够 actually get things done, remembers what什么 happened before and what什么 needs to be done next, and can能够 keep pushing tasks forward推进, is far more substantial大量 than simple简单 conversation. And with the ability to call various各种各样的， 不同的 external外部 services, many things no longer necessarily require you to open an app.一个能管事的 Agent，记得住之前发生了什么、接下来要做什么，还能持续推进任务，要比简单对话实在得多。而且有能力调用各种外部服务，很多事情就不一定非得你来打开 App 了。

This may also be why the big companies are rushing to build personal Agents智能体.这或许，也是大厂们抢着要做个人 Agent 的原因。

After all, the longer an Agent智能体 spends with you and the better it understands your habits and needs, the more likely it is to arrange安排 things properly合乎体统的，正派的.毕竟，Agent 跟你相处得越久，越了解你的习惯和需求，就越有可能把事情安排妥当。

Over接管 time, you may no longer jump back and forth between之间 several OTAs for a plane ticket机票; instead, you first tell the Agent智能体 what什么 you want to do and let it decide which services to call. The time everyone originally最初 spent searching, comparing比较， 对照 prices, and looking up information across different apps can能够 all be used for something else.久而久之，你可能就不会为了一张机票在几个 OTA 反复横跳，先告诉 Agent 想干啥，再让它决定该调用哪些服务。大伙儿原本花在不同 App 里，搜索、比价、查信息的时间，都可以拿来干点别的。

If that step is truly achieved达到，实现, our future internet habits may all have to change.如果真做到了那一步，咱们以后上网的习惯可能都得改改。

Of course, Chaping Jun thinks the ultimate最终的 goal of personal Agents智能体 is not to let us do nothing ourselves in the future.当然，差评君觉得个人 Agent 的终极目标，倒不是让咱们以后什么都不用自己干。

Being able to solve those trivial琐碎的 matters重要 that are not worth the effort, and letting everyone do something they enjoy, isn't that pretty good too?能把那些不值得费心的琐事解决了，让大伙儿干点自己喜欢的，不也挺好？

撰文：西西

编辑：江江

美编：焕妍

图片、资料来源：

灵光App

跳转微信打开

Original原创 by Shichao, 2026-10-28 00:00, Zhejiang原创 世超 2026-10-28 00:00 浙江

Is the largest最大的 IPO in history about to change hands again?史上最大IPO又要易主了？

The largest最大的 IPO in human history is here—or rather, it's here again.人类史上最大的 IPO，它来了，或者说它又来了。

Four months ago, Musk had just taken his SpaceX public, breaking the IPO record by raising筹集 $85.7 billion at a valuation估值 of $1.77 trillion万亿.四个月前，马斯克才带着他的 SpaceX 打破 IPO 记录，以 1.77 万亿美元的估值筹了 857 亿美元。

This time Anthropic is even甚至 more extreme极端. Although the offering发行 details细节 haven't been announced宣布 yet, the rumored传闻的 fundraising融资 target目标 has already hit达到 $100 billion.这回 Anthropic 更夸张，虽然还没公布发行细节，但传出的募资目标已经干到了 1000 亿美元。

啥概念呢？

Put it this way方式: in all of last year, 2025, never mind A-shares, Hong Kong stocks股票, or US stocks股票—the 1,293 IPOs worldwide全球 raised筹集 a combined合计 $171.8 billion. Together they still can't absorb吸收 as much as these two.这么说吧，去年 2025 一整年，也别说啥 A 股、港股、美股了，全球 1293 家 IPO，拢共筹了 1718 亿美元，特么合一块儿都没这哥俩能吸。。。

Of course, even甚至 setting设定 aside the exaggerated夸大的 numbers数字, Anthropic's prospectus招股书 this time is still quite worth reading. It discloses披露 many billing details细节 and operating conditions, and it's packed with substance实质内容.当然， 哪怕撇开夸张的数字，Anthropic 这回的招股书还是挺有看头的，它公开了很多账单细节和经营状况，干货拉满了。

The impression after reading it is that Anthropic is already trying to redefine重新定义 the AI market市场, because what所…的东西 it sells is the new concept概念 of "knowledge labor."看下来的感觉就是，Anthropic 已经在试着重新定义 AI 市场了，因为他们卖的是 " 知识劳动 " 这个新概念。

And the size of this pie is even甚至 more imaginative富有想象 than Musk's was before—it has reached a full $40 trillion万亿.而这张饼的大小，也比之前的马斯克更有想象力，来到了足足 40 万亿美元。。。

First, look at the most eye-catching抢眼的 set一组 of data: in Q1 of 2026, revenue营收 reached $4.731 billion, directly crushing碾压 all of last year.先来看最扎眼的一组数据，2026 年 Q1 季度，营收达到 47.31 亿美元，直接干翻去年一整年。

Q2 was even甚至 more vigorous强劲, with revenue营收 of $11.56 billion, up 1,368% year over超过 year—more than 13 times higher. Q3 wasn't directly disclosed披露, but based on the clues线索 left in the business section部分, a quick calculation计算 puts it at around $16.8 billion, and growth is still continuing.Q2 季度更来劲，115.6 亿的营收，同比增长 1368%，翻了 13 倍多。Q3 没有直接公布，但根据业务部分留下的线索，掐指一算，应该在 168 亿美元左右，增长还在继续。。。

At the same time, the loss for the first half of the year was also outrageously惊人地 large, reaching $80.748 billion.而与此同时，上半年亏损的数额也夸张得一批，直接干到了 807.48 亿美元。

No way方式—even甚至 if Anthropic were burning money for fun, it couldn't lose this much, right?不儿，就算 Anthropic 把钱拿去烧着玩儿了，也不能亏这么多吧？

Well, don't be so sure—there really is a hidden隐藏的 reason原因.诶，你别说，还真有隐情。

If you dig through通过 the financial财务的 report报告, you'll find that of this $80.7 billion, $72.388 billion comes from changes in the fair公允 value价值 of convertible可转换的 bonds and other financial财务的 instruments工具.仔细翻财报你会发现，这 807 亿里面，有 723.88 亿，来自可转债等金融工具的公允价值变动。

In plain直白的 terms, Anthropic previously此前 raised筹集 a large sum of money, and many of the financing融资 instruments工具 can能够 be converted转换 into company stock股票 in the future under certain conditions.说人话，就是 Anthropic 之前融了一大笔钱，其中不少融资工具，可以在将来按条件转换成公司股票。

When the financing融资 was done, the price was still very low, but as the company's valuation估值 climbed all the way方式 up, these things became变得 more and more valuable有价值的.当初融资的时候，价还很低，但随着公司估值一路上涨，这玩意儿就越来越值钱了。

For the investors投资者 holding them, this is of course a good thing—they can能够 make money lying down—but for Anthropic, it is purely a liability负债, recorded on the books as a huge loss.对持有的投资者来说，这当然是好事，躺着就能赚钱，但对 Anthropic 来说，就纯纯负债了，记到账面上就是一笔巨额损失。

That's why there's this absurd荒谬的 picture: the company becomes more and more valuable有价值的, while the income statement报表 looks worse and worse.所以才有这么个抽象画面，公司越来越值钱，利润表反倒越来越难看。

To put it plainly直白地说, though, Anthropic hasn't actually paid out real money from its bank银行 account账户 for this—it just looks bad on paper贴壁纸.不过说白了，其实 Anthropic 也没从银行账户里真金白银的给这掏钱，这部分也就看着难看罢了。

If you strip剔除 these things out, under Anthropic's own disclosed披露 non-GAAP非公认会计 basis基准, it still had net净 profit in the first half after adjustments, and it already reached 2.671 billion USD.如果把这些玩意儿剔除掉，按 Anthropic 自己披露的非 GAAP 口径，上半年调整后还是有净利润的，而且已经达到 26.71 亿美元。

That's right—its cash flow is actually pretty normal...没错，其实它的现金流还挺正常的。。。

And the user data supporting支撑 revenue营收 is also strong. One metric in there, called net净 revenue营收 retention留存 (NRR), excluding不包括 subscribers订阅用户, is around 900%.而且支撑营收的用户数据也很有力气，里面有个叫净收入留存率 （ NRR ） 的指标，剔除订阅用户，数字大概是 900%。

In plain直白的 terms, last year's old customers spent roughly 9 times as much this year as in the same period last year.大白话理解就是去年的老客户，今年的消费额比去年同期大概翻了 9 倍。

It's certainly sticky粘性的 enough—with that level of paying appeal吸引力, how could it worry about not making money?确实足够黏呐，有这种等级的付费吸引力，哪愁赚不着钱呢。

But this part doesn't have much to do with us—after all, you can能够 still occasionally偶尔 spot发现 a few grasshoppers蚱蜢 after autumn, but a Claude account账户 that survives存活 more than a year is truly rare罕见.不过这部分跟咱倒是没太大关系，毕竟秋后的蚂蚱偶尔还能见着几只，活过一年的 Claude 账号那是真的稀有。

The funniest最滑稽的 part is that Anthropic also casually随意地 released发布 a set一组 of data saying that 5%–7% of revenue营收 is identified识别出 as coming from customers using it in violation违规 of the rules规则.最搞的是，Anthropic 这厮还顺手公布了一组数据，说有 5%～7% 的营收，被认定来自违规使用的客户。

Thinking of those popular science short videos视频 made by Opus 5.5 all over超过 the place a few days ago—this is almost naming names...想起前些天 Opus 5.5 制作的满山遍野的科普小视频，这就差指名道姓了。。。

Stacking堆叠 data on data like this, according根据 to their estimate估计, the economic经济的 value价值 corresponding相应的， 符合的 to global全球的 knowledge labor in 2025 exceeds超过 40 trillion万亿 USD.就这么数据叠数据，按照他们的估算，2025 年全球知识劳动所对应的经济价值，就超过了 40 万亿美元。

In other words, if AI really keeps going like this, then what所…的东西 it sees is no longer the small change of traditional software（计算机的）软件—how could it not take a few bites out of the entire knowledge labor market市场?也就是说，假如 AI 真这么一路干下去，那么它眼里的，可就不是传统软件的那仨瓜俩枣了，整个知识劳动的市场哪不能啃上两口啊。

Of course, this 40 trillion万亿 USD is also standard pie-in-the-sky stuff材料，东西, just like Musk saying he wants to colonize拓殖,殖民 Mars火星—please respect and understand it, but for now it's just fighting for the future.当然，这 40 万亿美元也是画饼基操了哈，就跟老马说要殖民火星一样，期待尊重理解，但眼下也只是战未来。

After all, with a pie drawn this big, to bake it, how much batter连续猛击 do you need to mix, and how big a furnace do you need to build?毕竟饼画的这么大，要想烤出来，你这面糊得打多少，炉子得垒多大啊？

Well, Anthropic is also quite honest—it immediately lays out its spending budget预算 for you.诶，Anthropic 也很老实，转头就把支出预算给你摆出来了。

It went through通过 all the compute计算 suppliers供应者, 补充者 on the market市场—Google, Amazon, Microsoft, SpaceX, and so on—and among the eight companies' orders, contracts合同 are scheduled计划表 as far out as 2045.他把市面上的算力供应商撸了个遍，什么谷歌亚马逊，微软 SpaceX 等等，八家的订单，合同最远给排到了 2045 年。

Among these orders, the tens-of-billions level is just an appetizer; only the hundreds-of-billions level gets a seat at the table, and piled together they total more than 800 billion USD.这堆订单里，百亿等级是开胃菜，千亿等级的才能上桌，摞起来足足有八千多亿美元。

Yes, Anthropic's pre-booked fight-for-the-future bill is roughly equal to one Musk—it's genuinely真的 been scared瘢痕 poor by compute计算.是的，Anthropic 预定的战未来账单，约等于一个马斯克，算力这块儿是真给哥们儿穷怕了。

But these contracts合同 can't simply be seen as liabilities责任, because apart from delivery details细节, many of these companies are more than just creditors债权人.不过这些合同倒也不能单纯的看成负债，因为撇开交付细节，就这几家，很多都不止债主这一个身份。

For example, in May this year, Amazon invested投资， 投入 5 billion USD in Anthropic, and Google was even甚至 more generous慷慨的, investing投资 10 billion.比如今年 5 月，亚马逊就给 Anthropic 投了 50 亿美元，谷歌更大气，投了 100 亿。

And if you've used their respective各自的 Agent代理人， 代理商 products, like Amazon's Kiro or Google's Antigravity, you'll find that Claude's models are in there too.而且如果你用过他们各自的 Agent 产品，像亚马逊的 Kiro，或者谷歌的 Antigravity，会发现里面也有 Claude 的模型。

On top of that, after the partnership合伙(关系) they can能够 also deploy部署， 调度 Claude on their own servers服务器 and then sell the API. In the first half of 2026, 48% of Anthropic's revenue营收 already came from sales through通过 these third-party第三方 cloud platforms政纲，党纲，宣言.除此之外，合作后他们还能在自家服务器上部署 Claude，然后售卖 API，2026 年上半年，Anthropic 已经有 48% 的收入来自这些第三方云平台的销售。

That's right, these big companies are even甚至 sharper than we thought: they make a buck反对 n. 美元 from investing投资, another from renting out compute计算, and yet another from helping sell the goods...没错，这些大厂比咱想的还要猴精，投资赚一笔，出租算力赚一笔，帮忙卖货还能再赚一笔。。。

Even甚至 though they are themselves training top-tier顶尖 models and battling激战，战役 each other at the AI frontier边境, that doesn't stop them from sitting down to talk real business.即使他们自己也在训练顶尖模型，在 AI 前沿互相厮杀，但丝毫不影响坐下来谈现实生意。

This may also be the awkward尴尬的 spot发现 for Anthropic. Frankly坦率地, its ideals理念 are still too idealistic理想主义的, 空想的.这可能也是 Anthropic 的尴尬之处，说实话，它的理想还是太理想了，

Although it has always thought highly of itself—it even甚至 chose PBC (public benefit受益 corporation) as its registration注册，登记 type—tokens代币 don't care about any of that: without compute计算, it simply can't spit them out.虽说它一直自视颇高，连公司注册类型都选了 PBC （ 公益公司 ），但 token 不跟你讲这那的，没算力就是吐不出来。

Half your revenue营收 depends on the giants selling for you, and you've already placed over超过 800 billion in orders waiting to recoup收回， 弥补 the cost—and now you still want not to be led by the nose突出部分（如船头等） by commercial商业的， 商务的 interests? That's a lot of head-scratching ahead.你一半的收入都靠巨头带货了，又预定了八千多亿的订单等着回本，这时候还想不被商业利益牵着鼻子走？那可有得头秃了。

After all, even甚至 someone as willful故意的， 特意的 as Musk—complaining抱怨， 诉苦 about Anthropic one moment, then landing a compute计算 rental order the next—can能够 grin咧着嘴笑 and cooperate.毕竟哪怕任性如马斯克，前脚吐槽 Anthropic，后脚搞到算力出租订单，也是能合作笑嘻的。

In such a chaotic混乱的, 无秩序的 capital资金的，资本的 market市场, when AI safety and making money come into conflict冲突, can能够 you, Company A, still hold on to the original原创 aspiration强烈的愿望 you set一组 for yourselves?就这么个混沌的资本市场，当 AI 安全和赚钱起了冲突，你 A 社还能守住自己设立的初心么？

In response回应, they simply set一组 up an extremely极其 complex复杂 equity股权 structure.对此，他们直接搞了套异常复杂的股权架构。

As everyone knows, to protect their say, founders创始人 of listed companies generally separate the right to share profits from the right to vote on management管理层.众所周知，上市公司的创始人为了保障话语权，一般会把分红赚钱和管事投票的权利分开算。

Well, Anthropic did the same, and in even甚至 finer detail琐碎，小事: they split分裂，分化，派系 their stock股票 into five classes—Class A, N, E, T, and F.诶，Anthropic 也这么干了，而且干的更细，他们把股票分成了 5 类，Class A、N、E、T、F。

Among them, Class A shares are normal stock股票—what所…的东西 ordinary investors投资者 buy—offering发行 both profit and voting rights.其中 A 类股是正常的股票，就普通股民买到手的，既能赚钱，也有投票权。

Class N shares are mainly总管道，干线 reserved保留 for strategic partners, like Google and Amazon mentioned提及 earlier早些时候: they can能够 share in the profits but basically cannot不能 participate参与 in company decisions.N 股主要留给战略合作伙伴，像前面的谷歌亚马逊，能分钱，但基本不能参与公司决策。

What's really interesting are the latter three classes.真正有意思的是后面三类。

First, Class F shares, reserved保留 specifically for the seven co-founders: there is only one such share in total, it pays no dividends股息, yet it carries刊登 50.1% of the company's voting rights.先说 F 股，专门留给七位联合创始人，总共就这么一股，不参与分红，却拥有公司 50.1% 的投票权。

In other words, no matter重要 how many more rounds of financing融资 they raise提出，发起，发出 in the future, as long as this mechanism机制 remains剩下，余留，保持, the seven founders创始人 will jointly hold the decisive决定性的 power in shareholder股东 votes.也就是说，哪怕今后融了一轮又一轮资，只要这套机制还在，七位创始人就能共同掌握股东表决的决定权。

But relying依靠， 依赖 on the founders创始人 alone is clearly not enough, so they created Class T shares and handed them to an independent organization called LTBT (Long-Term Benefit受益 Trust).但光靠创始人，显然还不够，于是他们又弄出了 T 股，交给一个叫作 LTBT （ 长期利益信托 ） 的独立组织。

This thing is even甚至 more extreme极端: it takes no dividends股息 and does not participate参与 in ordinary shareholder股东 votes, yet it can能够 decide the selection选择 of the majority大多数 of the company's board董事会 seats.这玩意儿更绝，它既不拿分红，也不参与一般的股东投票，却能决定公司董事会多数席位的人选。

The members of the trust are generally experts熟练的 n. 专家 with backgrounds出身背景， 经历 in AI safety or public policy, kept利益-isolated使隔离 from the company's stock股票 price and profits.而信托里的成员，一般是有 AI 安全或者公共政策背景的专家，与公司股价和利润保持利益隔离。

Put simply, they specifically brought带来 in a group of people who don't profit from Anthropic's stock股票 price to supervise监督 it neutrally中立地.说白了，就是专门找了一帮不靠 Anthropic 股价赚钱的人，来中立的监督它。

The final E shares are employee governance治理 shares, which carry刊登 no dividend分红，红利 rights but can能够 participate参与 in voting when trust rules规则 are amended修订 or certain director candidates候选人， 投考者， 申请求职 are disputed.最后的 E 股就是员工治理股，没有分红权，却能在修改信托规则、部分董事人选出现争议时参与投票。

This way方式, each party has its own power, and no one can能够 use a single class of stock股票 to monopolize垄断 all decisions—adding a few more layers层， 层次 of armor盔甲， 装甲 to their original原创 ideals理念.这么一来，各方有各方的权力，谁也没法靠一类股票包办所有决定，算是给自己的初心多叠了几层甲。

In addition加， 加法, they set一组 up a public interest fund资金，基金 holding 5% of Anthropic's shares, with more to be added as the stock股票 price rises; they say it is dedicated奉献 to the public interest, but how exactly it will be distributed has not been specified.除此之外，他们还搞了个公益基金，里面存了 5% 的 Anthropic 股份，后面股价涨了还会继续存，说是致力于公共利益，但具体咋分配倒还没说法。

Still, I think Musk's suggestion is a good one—just give it all away to everyone outright全部地.不过我看马斯克的建议就不错，索性直接白送给大伙儿就完了。

So even甚至 though we usually grumble抱怨，发牢骚 about Anthropic, when a giant like this goes public and is willing to experiment with a corporate法人团体的， 公司的 governance治理 setup, it's still worth discussing.所以别看咱平时老蛐蛐 Anthropic，但这么个大家伙上市，还愿意折腾一套公司治理的实验，还是值得说道说道的。

After all, these things were previously此前 all locked inside their company, kept secret layer层 by layer层; forget the specific具体的 cost breakdown崩溃，倒塌，失败—we probably couldn't even甚至 see which companies they were secretly eyeing.毕竟这些东西之前全都封在他们内部，层层保密，别说具体的成本组成了，怕是他们跟哪些公司眉目传情了咱都瞅不着。

Once the prospectus招股书 is disclosed披露, at least we can能够 see clearly how their business actually operates and how they burn money.招股书一披露，起码咱能看清他们的业务到底是怎么个情况，烧钱是怎么个烧法。

As for their management管理层 methods and how much of their original原创 ideals理念 they can能够 ultimately最终的 hold onto, that is still for time to verify证实， 查证， 证明.至于他们的管理方法，最后能搂住多少初心，那还是交给时间来验证。

Speaking of which, OpenAI's IPO is not far off either; we're looking forward向前 to that too, to see what所…的东西 new and unseen未见过的 things this other veteran经验丰富的 n. 老手 can能够 pull out...说起来，OpenAI 的上市也不远了，咱也期待一手，看看这另一位老资历又能掏出点什么没见过的新东西。。。

撰文：风华

编辑：江江 & 面线

美编：焕妍

图片、资料来源：

X, Anthropic, Bloomberg, Wikipedia, some images from the internetX、Anthropic、Bloomberg、维基百科、部分图源网络

跳转微信打开