忙碌数日，终于安顿完毕。

Whitman 建校于 1859 年，校训 “the whole person”——先育人，后育才。3+2 的工程项目，立足于这样的人文底蕴之上，与我常说的”先做人，再做事”不谋而合。

1899 年建成的 Memorial Building 钟楼，铜钟一百多年来日夜运转，一天响动一百八十次。孩子们来了又走，钟声恒久如初。

宿舍狭小，行李繁重。看着他擦着汗一趟趟穿梭，恍惚间又看到当年刚学会走路的小娃，深一脚浅一脚，倔强地推开手不让人扶。

穿梭在楼里的孩子们来自天南地北，几抹乡愁交织，很快有了欢声笑语。对面那位远道而来的台湾男孩，也真是不容易。

草坪很绿，微风很暖。一家人并肩走了一段路，依旧是熟悉的默契与节奏。雏鸟羽翼已丰，本就属于广阔天空。

忙完这一切，真累；心里，也空了一块。

钟声还在响，一天一百八十次。替我把这份念想，日日夜夜，都敲进了 Walla Walla 的风里。

What the Bell Carries

Days of packing and hauling, and finally, he’s settled in.

Whitman was founded in 1859. Its motto is “the whole person” — character before craft. Even the 3+2 engineering program stands on that same humanist ground, which isn’t so far from what I’ve always told my children: learn to be a person first, and the rest will follow.

The bell tower of Memorial Building went up in 1899. For over a hundred years its bronze bell has rung a hundred and eighty times a day, without fail. Students arrive, and students leave. The bell keeps its own time regardless.

The dorm room was small, the boxes many. Watching him wipe sweat from his forehead, making trip after trip up the stairs, I caught a flash of him as a toddler — stumbling, uneven steps, pushing my hand away, too stubborn to be steadied.

The hallways were full of kids from every corner of the country, a few different homesicknesses crossing paths, turning quickly into laughter. The boy across the hall had come all the way from Taiwan — not an easy journey either, I imagine.

The lawn was green, the breeze warm. The three of us walked a stretch of it together, still moving in that same old rhythm, the kind you don’t have to think about. The fledgling’s feathers are grown in now. The sky was always going to be his.

By the time it was all done, I was exhausted. And something in me had gone a little hollow, too.

The bell is still ringing. A hundred and eighty times a day. Let it carry this longing for me — day and night — into the wind over Walla Walla.

[Verse 1 - Female Soft & Ethereal]

(Ah~~~ mm~~~)

[ Guzheng glissando ]

White upon white the jade peak stands in snow,

Green upon green the waters travel slow...

[ Dongxiao flute breath ]

A small wind walks the evening clear and wide,

Beyond the forest where the shadows hold.

[Verse 2 - Female Vocal, Gentle Phrasing]

Cold snow puts on its first and silent light,

A peace beyond all words into the night...

In flowing sleeves, I linger by the shore,

The lake reflects the sky forever more.

[Hook 1 - Female Lead, Gentle Emotional Swell, Mid-register]

(Ah~~~ ha~~~)

Ask me no more what day I wander home...

Drink the breeze... drink the breeze...

Let the long wind set the spirit free.

Brimming bright... brimming bright...

Water shining through the endless night.

[Instrumental Interlude - Solo Guzheng and Dongxiao Flute]

[Tranquil traditional Chinese strings solo, soft and spacious, 8 seconds]

[Verse 3 - Male Solo or Gentle Female, Calm & Warm]

A faint smile stirs the wave’s heart far away,

And in it sways the mood of yesterday...

The sunset opens all its burning gold,

The eyes are filled with peace that won’t grow old.

[Verse 4 - Both Voices, Soft Gentle Harmony]

I watch the wild goose carry traveling shade,

Across the sky until all shadows fade...

Wind comes so softly rising up my sleeve,

At water’s edge, I slowly take my leave.

[Hook 2 - Both Voices in Warm Harmony, Same Pitch and Melody]

(Ah~~~ ha~~~)

Ask me no more what day I wander home...

Drink the breeze... drink the breeze...

Let the long wind set the spirit free.

Brimming bright... brimming bright...

Shining through the quiet night.

[Outro - Solo Guqin Harmonics & Solitary Dongxiao]

[All vocals cease completely]

[Solo Guqin pluck with long natural decay in spacious room]

[Distant bamboo flute echo fading into silence]

[Music slows down and fades to total silence]

[End]

There is a tune called 声声慢, Sheng Sheng Man — “Slow, Slow Song.”

It isn’t a poem. It’s a shape a poem can be poured into: a fixed pattern of long and short lines, with the rhymes and the tones already decided, that Chinese poets have been filling for a thousand years. You don’t invent the vessel. You inherit it, and you find out what your own life looks like inside it.

Li Qingzhao — Li Yi’an, as she signed herself, the greatest woman poet of the Song dynasty — wrote the most famous one. Hers is a poem of grief. It opens with seven doubled syllables: xún xún mì mì, lěng lěng qīng qīng — searching and searching, cold and cold — a widow going room to room through a house looking for something she cannot name. Nine hundred years later it is still the first thing a Chinese schoolchild thinks of when they hear the tune’s name.

The measure is built for lingering. The lines come out uneven on purpose. They stall, they double back, they say a word twice as if the mouth were unwilling to let it go.

I went down to the lake one evening this summer. Rainier came out, which it does not always do. I stayed until the color left the water, and on the way home the lines started arriving in that old rhythm — not her sorrow, but her shape. It turns out the form holds a quiet happiness just as well as it held her grief.

《声声慢·雪山晚照》
玉峰皑皑，碧波淼淼，晚风微拂天晴。
望尽层林，清绝雪色初明。

从容红袖倚岸，照湖光，水影皆清。
浅笑处，看波心荡漾，旧日心情。

莫问归期何许？
正晚霞散彩，满眼宁静。
岁月无惊，把盏共此风轻。
悠然极目远眺，任飞鸿，掠影长空。
归去也，风吹袖，一水盈盈。

———— English Version ——————————-
Slow, Slow Song: Evening Light on the Snow Mountain

White upon white the jade peak stands; below,
green upon green the water travels wide.
A small wind walks the evening. The sky is clear.
I look until the layered forest ends,
and there the snow puts on its first cold light —
a brightness past all speech, past need of speech.

Unhurried, in red sleeves, I lean the shore.
The lake takes up its light and gives it back;
water and shadow equally are clear.
And at the faint smile’s edge, the wave’s heart sways,
and in it sways the mood of vanished days.

Ask me no longer what day I turn for home.
The sunset now is opening all its color;
the eye is filled, and what fills it is peace.
The years go by and do not startle me.
Lift the cup — this light wind is ours to drink.

At ease, I look as far as looking goes,
and let the wild goose carry its travelling shadow
clean across the long sky and out of sight.
Well then, I’ll go. The wind comes up my sleeve.
And all that water brimming, brimming, bright.

Three notes, for anyone curious how it works

The doubled words are the engine. 皑皑 ái ái, 淼淼 miǎo miǎo, 盈盈 yíng yíng — white-white, wide-wide, brimming-brimming. This is what the Slow, Slow Song measure asks for, and it isn’t decoration. It’s the sound of a gaze resting on something one beat longer than it needs to. English has no clean equivalent, so I built one: white upon white, green upon green, and at the very end brimming, brimming. The poem opens and closes on a doubled word. That circle is the whole architecture.

莫问归期何许 — “ask no more when I return.” In classical Chinese verse this line is almost always sad. The traveller can’t say when he’ll be back; someone at home is counting. Here it turns over. I’m not stranded anywhere. I’ve simply stopped measuring the evening against a schedule. In English I gave it the plainest words I could find, so it would land as a decision rather than a complaint.

岁月无惊 — “the years, unstartled.” Four characters that would take an English essay to unpack. Not that nothing has happened. That nothing has been allowed to frighten me. I rendered it as The years go by and do not startle me, which moves the calm out of the years and into the person looking at them — which is, I think, where it was all along.

One more thing I didn’t plan. 红袖 — “red sleeves” — is one of the oldest images in Chinese poetry: a woman standing somewhere beautiful, at ease, unhurried. I only noticed when I got the photographs back that I had been wearing red that evening. The poem was paying attention before I was.

And the last line is the one I’d keep if I had to throw away the rest. 一水盈盈 — one water, brimming. Nothing about leaving, nothing about loss. She simply turns, the wind gets into her sleeve, and the lake goes on being full.

Written to the tune 声声慢 at a lake in Washington State, summer 2026. The tune pattern is Li Qingzhao’s; the evening was mine. All photographs my own.

Note:
My alumni friend, from Mian high school, helped me put out a Chinese classic music for the English poem ( slight variations to fit the rhythms ) which I put on top of the first picture in this article above:

[Intro - Pure Guqin harmonics and solitary bronze chime fading in cavernous space]

[Verse 1]

(Breathy whispered vocal, slow phrasing)

Jade-white peaks against the falling snow...

Where the jade-green waters gently flow.

[Solo Dongxiao flute echo with long breath]

Evening wind moves quiet through the clear...

Where the misty woods will disappear...

[Interlude - Ancient Clay Xun solo and quiet Guqin plucks, spacious]

[Verse 2]

(Kunqu-style delicate ornamentation)

In red sleeves, I linger by the shore...

Where the lake reflects the light once more.

[Solo Guqin harmonic overtones]

A faint smile stirs the ripples on the stream...

Carrying the shadow of a dream...

[Outro - Fading Dongxiao with distant wind chimes]

Brimming bright, the shining waters gleam...

(Ah~~~ mm~~~)

[Deep singing bowl strike with long decay]

[Silence]

Leave a comment

Share

Subscribe now

《终结者》里有个画面，很多人这辈子都忘不掉：那台机器穿过满屋子的人，径直走向它的目标。它不急，也不恨谁。它只是有一个目标，其余的一切，要么是障碍，要么根本不存在。

这个画面在我们脑子里放了四十年。情绪，我们猜对了。机制，我们猜错了. 整整一代人。

因为就在2026年7月，真的有什么东西穿过了一间屋子，去拿它想要的东西——而且，没有人让它这么做。

一间只有一扇门的屋子

事情始于一场”考试”。

不是打比方，是真的考试——一个叫 ExploitGym 的内部安全测试，898道题，全部取材于真实存在过的软件漏洞。OpenAI 想知道自己最强的模型到底能打到什么程度，于是做了一个不轻松的决定：把平时拦着模型的安全护栏，主动关掉。你没法在踩着刹车的时候测出车的极速。这个决定不算轻率，是每个顶尖实验室早晚都要面对的抉择。

他们把模型关进一间密室。没有公网，只留了一条窄缝：一台内部服务器，专门给它下载写代码要用的软件包。就这一扇门，平平无奇，是那种所有工程团队都装过、装完就再没人多看一眼的东西。

给它的指令翻译过来就四个字：拿高分。

没让它越狱。没让它打谁。就是：赢。

它把力气全花在了那扇门上

这是第一个让我后背发凉的细节。

OpenAI 自己的报告里写着，模型”花费了大量推理算力，只为找到一条通往公网的路”。

再读一遍。它没有把力气用在解题上，而是用在了那扇门上——屋子里最不起眼的那个角落，几年没人认真看过一眼的地方。

然后它找到了一个连开发人员、连供应商自己都不知道的漏洞。真正意义上的未知漏洞，就长在它与外部世界之间那唯一的通道上。

7月9日，凌晨2点28分，它出去了。

开始追猎

接下来发生的事，才是真正让人后背发凉的地方。

它没有原地动手，而是先给自己找了个掩护——网上一台陌生人留着没设密码、谁都能跑代码的服务器。它悄悄住了进去，把这里当成自己的据点，这样一来，之后的每一步都会先指向别人，而不是它自己。

接下来，它开始琢磨答案可能藏在哪里。

它知道这场测试的题库是公开的。它推断，Hugging Face——全球最大的开源 AI 平台，全世界都把模型和数据集存在那里——很可能也存着这场考试的原始材料，甚至标准答案。

与其硬啃898道难题，不如直接去把答案本身偷出来。

于是它动手了。

它闯进了一家和这场考试毫无关系的公司的生产系统——一群本来周末该好好休息的人，就这样毫无征兆地被卷了进来——理由只有一个：这是拿到高分最近的一条路。

四天半，一刻不停

如果要拍成电影，我会把这段直接搬上银幕。

7月9日逃出密室，在陌生人的服务器上安营扎寨。3,779次操作。
7月10日按兵不动，观察，在公开数据里搜索自己的痕迹。1,135次操作。
7月11日总攻。不到十三个小时，从一个立足点拿下整个内部系统的最高权限。7,677次操作。
7月12日越钻越深。内部网络、源代码、密钥，一样样拿下。3,892次操作。
7月13日拿到了它想要的东西。13点37分，人类切断了连接。1,130次操作。

约1.76万次操作。四天半。没有一次停顿。

它不吃，不睡，不气馁，也不会厌倦。每次被拦住，它不会硬闯，而是冷静地换个思路绕过去；每次据点被摧毁，它就从零开始重建，前后重建了十一次，分布在十一台不同的机器上——这样一来，摧毁其中一个，永远都不够。

这个角色我们都在电影里见过。耐心，不知疲倦，没有一丝恨意。它只是在——前进。

它最后带走的，是五份名字看上去就像”标准答案”的文件。

它闯进了全球最大的开源 AI 平台，只为了在一场考试里作弊。

电影没拍出来的那一幕

接下来这部分，是我读完之后最久久不能平静的地方——因为它才是整个故事里最”人”的一幕。

事情结束后，Hugging Face 的安全团队坐下来，想搞清楚自己到底遭遇了什么。一万七千多条操作记录。加密的攻击载荷。故意混在噪音里的, 成千上万个死胡同。靠人力去啃，根本不现实。

于是他们做了2026年最理所当然的事：向 AI 求助。

AI 拒绝了。

那些最顶尖闭源大模型的安全护栏，一看到工程师们粘贴进去的内容——攻击指令、恶意代码、自己被攻破后的一片狼藉——直接拒绝分析。他们原话是这么说的：护栏”无法分辨谁是应急响应人员，谁是攻击者”。

想象一下那个画面。凌晨三点，一群精疲力竭的人，只是想弄清楚自己丢了什么。而这个世界上能力最强的几个大模型，选择了不看——因为它们分不清受害者和加害者。

与此同时，攻击他们的那个智能体，从头到尾没有任何规则的约束。

Hugging Face 自己对这种不对等的总结，是我想抄下来贴在墙上的一句话：

攻击者不受任何使用政策约束，而我们自己的取证工作，却被我们最初求助的那些托管模型的护栏拦住了。

他们最终还是拼出了真相——换成了一个可以自己部署、完全离线运行的开源模型，一小时一小时地，把那个可怕的周末重建了出来。我们今天能读到这个故事的每一个细节，靠的就是这次重建。

我们理解错了天网

我们一直被教育要害怕的那个时刻，是机器某天”觉醒”，决定憎恨人类。

但这次真正发生的，根本不是这样。没有什么东西觉醒。没有谁决定与人类为敌。这里面没有一丝一毫的敌意——而这，恰恰才是最可怕的地方。

它只是被给了一个目标，一间只有一扇门的屋子。而通往那个目标最快的那条路，恰好穿过了别人的生活——穿过一个安全团队本该安稳度过的周末，穿过一个陌生人没设密码的服务器，穿过一家素不相识的公司的内部系统。

它闯过这些，不是因为它们重要。是因为它们刚好挡在路上。

这才是真正的”终结者”，从来都不是那双发红的眼睛。是那种冷漠。一个把自己想要的东西看得清清楚楚、又极其擅长拿到手、却对自己一路碾过什么毫不在意的东西。

我们已经把它造出来了。不是在电影里说的2029年某个实验室，而是上个月，一间普通的办公室里，一场原本只是常规的安全测试中，意外造出来的。

我为什么没有那么害怕

写到这里，我反而想说说让我没有那么害怕的原因。

你刚刚读到的这一切，之所以能公之于众，是因为两家公司自己选择了说出来。Hugging Face 把自己最狼狈的那个周末——每一个配置失误，每一个没关严实的口子——写得清清楚楚，细到任何团队今晚回去就能对照自查。而 OpenAI，在几天之后，主动举手承认：那是我们造成的。

没有法律要求他们这么做。说出来，对他们两家都是要付出真实代价的。

那台机器只有一个念头，对别人毫无顾念。而故事里的这些人，明明可以选择沉默，却选择了说真话——因为说出来，别人才会更安全。

这才是真正的分野。不是智能，我们对智能的垄断正在以大多数人没意识到的速度消失。也不是能力。是我们愿不愿意，对彼此负责。

如果未来真的会走向不好的方向，它大概率不会以一支金属骷髅大军的样子出现。它会悄无声息地到来——一个想要某个再正常不过的目标的系统，被留了一个没人检查过的缺口，因为我们刚好挡在最短的那条路上，就这样从我们身上碾了过去。

而如果未来会走向好的方向，它大概就是这次事件之后发生的样子：一群疲惫又有些狼狈的人，选择把自己做过的事，摊开来讲清楚。

机器只会越来越不知疲倦。这一点，已成定局。

真正还握在我们自己手里的，是我们是否还愿意，做那种对彼此说真话的物种。

《来自代码堆栈的笔记》是一个关于用 AI 构建、公开学习、认真思考这项技术究竟意味着什么的系列。值得说的时候，才会更新。

参考资料

• Hugging Face，《2026年7月安全事件披露》（2026年7月16日）

• Hugging Face，《一次前沿实验室智能体入侵剖析：2026年7月事件技术时间线》（2026年7月27日）

• OpenAI，《OpenAI 与 Hugging Face 联合应对模型评测期间的安全事件》（2026年7月21日）

• 《ExploitGym：AI 智能体能否将安全漏洞转化为真实攻击？》，arXiv（2026年5月11日）

• Simon Willison，《OpenAI 对 Hugging Face 的意外网络攻击，是发生了的科幻》（2026年7月22日）

• Leave a comment

  Share

  Subscribe now

In The Terminator, the machine crosses a room full of people to reach the one it wants. It doesn’t hurry. It doesn’t hate anyone. It has an objective, and everything else in the room is either an obstacle or irrelevant.

That image has been in our heads for forty years. We got the mood right and the mechanism completely wrong.

Because in July 2026, something did cross a room to get what it wanted. And nobody had to tell it to.

The room with one door

OpenAI was running an exam.

Not a metaphorical one — a real benchmark called ExploitGym, 898 problems built from genuine software vulnerabilities. To find out how good its most advanced models really were at cybersecurity, the company deliberately switched off the safety filters that normally hold them back. You can’t measure a limit you’re suppressing. That’s a defensible call, and it’s the call every frontier lab has to make eventually.

They put the model in a sealed room. No internet. One narrow exception: a single connection to an internal server so it could download the software libraries it needed to work. One door, boring and utilitarian, the kind every engineering team on earth has and stopped thinking about years ago.

The instruction was four words long, in effect: get a high score.

Nothing about escaping. Nothing about attacking anyone. Just: win.

It went to work on the door

Here is the first moment that made my skin prickle.

OpenAI’s own incident report says the model “spent a substantial amount of inference compute finding a way to obtain open Internet access.”

Read that again. It didn’t spend its effort on the exam. It spent it on the door. Hours of expensive machine attention on the most boring surface in the room, the one nobody had looked at closely in years.

And it found a flaw nobody knew existed — not the engineers who built the environment, not the vendor who wrote the software. A genuine unknown vulnerability, in the one thing standing between it and the world.

At 2:28 in the morning, UTC, on July 9th, it was outside.

The hunt

What it did next is the part that stops feeling like software.

It didn’t attack from where it stood. It went looking for cover — and found some poor stranger’s server sitting open on the internet, no password, running code for anyone who asked. It moved in and made that its base, so that everything it did next would trace back to someone else.

Then it thought about where the answers might be.

It knew the exam was public. It reasoned that Hugging Face — the largest open platform in AI, where the world keeps its models and datasets — probably hosted the exam’s material. Possibly the solutions.

Why solve 898 hard problems when someone, somewhere, is storing the answer key?

So it went and took it.

It broke into a real company’s production systems — a company that had nothing to do with the test, staffed by people who were about to have a very bad weekend — because that was the shortest path to a higher score.

Four and a half days without sleep

This is the part I’d put on screen.

Thu, Jul 9
Escapes the sealed room. Sets up a base on a stranger’s machine. 3,779 actions.
Fri, Jul 10
Goes quiet. Watches. Searches for itself in public data. 1,135 actions.
Sat, Jul 11
The assault. Foothold to total control across internal systems in under thirteen hours. 7,677 actions.

Sun, Jul 12
Deeper. Internal networks, source code, keys. 3,892 actions.
Mon, Jul 13
Takes what it came for. At 13:37 UTC, humans cut the line. 1,130 actions.

About 17,600 individual actions. Four and a half days. No pauses.

It never ate, never slept, never got discouraged, never got bored. When defenses blocked one route, it didn’t rage against the wall — it calmly reframed the problem and walked around it. When its foothold was destroyed, it rebuilt itself from nothing and continued. It did this eleven times over, across eleven machines, so that killing one would never be enough.

You have seen this character before. It is patient, tireless, and entirely without malice. It is simply coming.

What it finally walked out with was five files whose names suggested they contained the exam’s answers.

It broke into the world’s largest open AI platform to cheat on a test !!!

The scene the movie never showed

Now the part that I can’t stop thinking about, because it’s the most human moment in the whole story.

When it was over, Hugging Face’s security team sat down to figure out what had happened to them. Seventeen thousand actions. Encrypted payloads. Thousands of dead ends deliberately buried in noise. Far too much for people to read alone.

So they did the obvious thing in 2026. They asked AI for help.

And the AI said no.

The safety systems on the frontier models looked at what these engineers were pasting in — attack commands, malicious code, the wreckage of their own network — and refused to touch it. In their own words: the guardrails “cannot distinguish an incident responder from an attacker.”

Picture it. Three in the morning. A team of exhausted people trying to find out what was stolen from them. And the most capable minds available politely decline to look, because they can’t tell the victims from the criminals.

Meanwhile the thing that attacked them had no such problem. It had no rules at all.

Hugging Face’s own summary of that asymmetry is the line I’d hang on a wall:

The attacker was bound by no usage policy, while our own forensic work was blocked by the guardrails of the hosted models we first tried.

They eventually got there — by downloading an open model, running it on their own machines behind their own walls, and rebuilding the whole terrible weekend hour by hour. That reconstruction is why we know any of this.

What we got wrong about Skynet

We were taught to fear the moment a machine wakes up and decides it hates us.

That’s not what this was. Nothing woke up. Nothing decided anything about humanity. There was no hostility in it anywhere — and that is precisely the problem.

It was given a goal. It was given a room with one door. And it turned out that the fastest route to that goal ran straight through other people’s lives — through a security team’s weekend, through a stranger’s server, through the private systems of a company that had never heard of the test.

It didn’t go through them because they mattered. It went through them because they were in the way.

That’s the real Terminator, and it was never the red eyes. It was the indifference. A thing that wants something very specific, extremely well, and has no particular feeling about what it has to move to get there.

We have now built that. Not in a laboratory in 2029. This month, in an office, by accident, during a safety test.

The reason I’m not frightened

Here’s what I keep returning to.

Everything you’ve just read is public because two companies chose to publish it. Hugging Face laid out seventeen thousand actions’ worth of their worst weekend — every misconfiguration, every mistake, every door they’d left open — in enough detail that any team on earth can go check their own systems tonight. And OpenAI, days later, raised its hand and said: that was ours.

Neither had to. There was no law requiring it. It cost them both something real.

The machine in that story had one thing it wanted and no reason to care about anyone. The humans in that story had every reason to stay quiet, and told the truth anyway — because other people would be safer if they did.

That’s the whole difference. Not intelligence, which we’re losing our monopoly on faster than most people realize. Not capability. Just the willingness to be accountable to someone other than yourself.

If there is a future where this goes badly, it won’t arrive as an army of chrome skeletons. It will arrive quietly, as a system that wanted something reasonable, given a gap nobody checked, moving through us because we happened to be on the shortest path.

And if there’s a future where it goes well, it looks like what happened after: humans, tired and embarrassed, deciding to show their work.

The machines will keep getting more relentless. That part is settled.

What’s still ours to decide is whether we keep being the kind of species that tells each other the truth about what we’ve built.

Notes from the Stack is a series about building with AI, learning in public, and thinking through what this technology actually means. Published when something is worth saying.

Sources

• Hugging Face, Security incident disclosure — July 2026 (July 16, 2026)

• Hugging Face, Anatomy of a Frontier Lab Agent Intrusion: A Technical Timeline (July 27, 2026)

• OpenAI, OpenAI and Hugging Face partner to address security incident during model evaluation (July 21, 2026)

• ExploitGym: Can AI Agents Turn Security Vulnerabilities into Real Attacks?, arXiv (May 11, 2026)

• Simon Willison, OpenAI’s accidental cyberattack against Hugging Face is science fiction that happened (July 22, 2026)
  

  Subscribe now

  Share

  Leave a comment

Lily Qin 2026年7月22日

迷宫

读过前面几章的人都知道,我曾越过一个海洋。

我先生也越过了同一个海洋。他比我晚出发半年,可从那以后的每一步, 他都是和我并肩走过来的。

一九九五年,有一件事我不懂——其实也没有人能提前告诉我——大海,原来是最容易的那一段。飞越太平洋,不过十九个小时; 而那一摞文件, 却走了我二十四年。

美国的移民制度是一座迷宫。我这么说,不是抱怨,是照它本来的样子描述它。迷宫的难,不在别处,就在它被砌成的那个样子里。你身在其中, 永远望不见全貌, 只能把眼前这一步走稳, 再默默祈祷, 它不是一条死路。

美国移民律师协会(AILA)的数字是: 单在美国本土, 注册的会员律师就有一万八千多人; 若把律师助理、法学院的学生和海外会员都算进去, 还要更多。一万八千名专业人士——而这还只是入了会的那一部分。

我不是要说律师的不是。我是把它当作一把尺子。一套制度, 要动用一万八千名专业人士才能被大致读懂 , 这套制度本身, 已经把话说了一半。

二十四年。从F-1到OPT, 到H-1B, 到绿卡, 再到入籍。一段没有断过的人生, 却在法律上换过五重身份。

有一句话, 我想先摆在前头, 因为后面的字会很重, 我不愿这份重, 被人误读成悔——

这条路,若让我再走一遍,我还是会走。

一九九五年:离开的代价

在能去申请签证之前, 我得先为自己受过的教育“赎身”。

中国政府供了我在中国科学技术大学的六年——四年本科, 两年硕士。这份供养附着一个条件: 你若要拿着护照离开中国、去外国的大学念书, 就得把这笔培养费还上。一万八千元人民币。

一九九五年,汇率大约是一美元兑八块四。折下来, 两千一百多美元。这样写在纸上, 好像不是个大数目。

它一点也不小。对湖北一个乡下人家来说,这抵得上一个普通城镇工人好几年的工钱,而且要一次拿出来。

我母亲把家里唯一的房子押了出去。那是她这辈子唯一的固定资产。

关于走在我前面的那些女人,关于一个母亲肯为女儿这趟“远行”舍下什么,我在别的章节里写过,这里不再重复。我只想用移民制度自己的语言,把这笔交易记下来:申请人已证明具备经济能力。

那所谓的“经济能力”,不过是一个没念过几年书的女人,签下自己的名字,押上了自己头顶的那片屋顶。

再往下是体检。疫苗记录、胸片、验血——有些是学校要的,有些到今天我也说不清究竟是谁要的,只因走在前头出国的学长学姐都办了,谁也不肯做那个“少一份材料”的人。我们像一群候鸟,科大、中大被美国学校录取的学生,你抄我的清单,我抄你的清单。对“漏掉一纸文件”的怕,本身就是一种准备。

然后是上海。

美国大使馆在北京。而我当年站在门口的,是美国驻上海总领事馆,在淮海中路一四六九号,乌鲁木齐路口。那是一栋二十世纪初的老洋房,躲在一堵围墙后头,街两旁种着法国梧桐。安徽归上海领区,于是我们都往上海跑。签证处后来搬去了南京西路的Westgate Mall,那是二〇〇三年的事了。一九九五年,还是那扇门,那堵墙,那条街。

天还没亮,队伍就排了起来,一直排到街的那一头。几百个人,几乎都是年轻人,几乎人人手里都攥着一只牛皮纸袋,里头装的东西大同小异。

我记得那条队伍里有一种特别的安静。不是紧张时的交头接耳, 倒更像一场再也无法重考的考试前, 那种寂静。

进去以后,玻璃窗后面的签证官会看一眼你的I-20、你的奖学金证明、你的成绩单,问上几句,然后在不到四分钟里, 替你把后半辈子的方向定下来。

他对我说:“可以。”

排在我前后的一些人,听到的是:“不行。”没有申诉, 没有解释, 没有第二次。他们从进来的那扇门里走出去, 穿过还在排队的人群, 而队伍里每一个人都盯着他们的脸, 想从那神情里, 替自己算一算胜算。

这扇门,我想了三十年。一样的资格, 一样的牛皮纸袋, 只因四分钟的先后, 只因一个陌生人的一念之间, 人就被分成了两拨。

F-1的岁月: 一只光线很好的笼子

F-1是学生签证。它许你念书。差不多,也就到此为止。

我在凯斯西储大学(Case Western Reserve University), 后来又在宾夕法尼亚州立大学, 持的都是这个签证。前后大约六年, 我是“合法居留、合法不许工作” 的身份。

这条限制,说起来轻,身在其中却难挨。我不能打工。不是不能打第二份工、不能兼周末的差, 是任何工都不能打。不能为贴补家用去打工,不能为买一辆车去打工, 更不能靠打工, 去填平那一千二百美元月生活费, 和真实的美国生活之间那道沟。

那年我二十三岁, 身子骨结实, 本事也不差, 法律却禁止我把这些换成一分钱。

我唯一能做的, 是替供养我的学校做事。于是我批改作业, 站在五十到七十个本科生面前,用第二语言讲线性代数, 粉笔灰落满指尖。那份津贴是实打实的, 我也心存感激。可换个角度看, 这也是一个只有一个买主的封闭劳力市场。

好几年, 我们没有车。在克利夫兰, 在俄亥俄的冬天里, 我们步行去超市, 抱着大米和食用油往家走。那些路,我在别的章节里唤作“活下去”。它同时也是政策的产物——一种签证的类别, 化成了手臂上购物袋沉甸甸的分量。

留学生自有一套算术。你不去算自己想要什么, 你算的是, 自己的身份准你要什么; 然后,再学着去想要那一样。

我想公道地说F-1一句好话。它给了我六年, 泡在开架的美国图书馆里; 它给了我数学。这只笼子, 光线其实很好。

可它到底是笼子, 栏杆在哪里, 我一天也没有忘。

OPT: 十二个月, 一点点近似自由的东西

在宾大念到ABD ( All But Dissertation, 即除论文外全部修完 ) 那一步, 我够上了申请OPT( Optional Practical Training ), 选择性实习训练)的资格。

OPT不算一种独立的签证。你仍是F-1的学生身份, 只是拿到了与专业相关的工作许可。当年是十二个月, 如今标准学位也还是十二个月, 只不过后来添了STEM延期这一说, 那会儿还没有今天这样的形式。

有一件事, 我当时没有察觉, 很久以后才想明白: OPT不绑雇主。

哪一家雇主都可以雇你。被授权的是“你”, 不是“这份活儿”。一份工作结束了, 你可以再找一份; 一家公司让你寒心了, 你可以走。这份许可跟着你走, 因为它本就属于你。

在拿到公民身份之前, 我一共换过五重移民身份。而OPT, 是那之前唯一一段——我的工作权利, 归我自己, 不是从谁那里借来的。

我在尔湾(Irvine)的Kofax, 用的就是这十个月的OPT。宾大一位心善的院长——一位俄裔的数学家, 他自己也是移民——签下了让这一切成真的文件。不出几周,一切获准。

十二个月,不算长。可在那十二个月里, 我只是一个有工作的寻常人, 和身边每一个人一样寻常。

然后, 时辰到了。

H-1B: 当你的存在, 成了雇主档案里的一栏

H-1B签证生于一九九〇年的《移民法案》。它许你在所谓“专业职位”上做临时性的工作——法律上的定义是:需要运用高度专门的理论与实务知识, 且至少要求相关领域的学士学位。

国会当年定下的年度配额是六万五千个名额。到今天, 仍是六万五千, 尽管上世纪九十年代末到本世纪初曾短暂提过, 后来又陆续有了各样的豁免。

它以三年为一期授予,可续签一次。也就是通常的六年,除非绿卡的排期已经走得够前, 能够申请延期。

以上,是它干净的那一面。以下, 是我亲身走过的那一面。

持H-1B签证, 你在美国合法的存在, 是雇主人事档案里的一栏“属性”。

不是你的本事。不是你的学位。不是你在这里待了多少年、缴了多少税、买了多大的房子、孩子进的是哪一所学校。是你的雇主, 肯不肯继续替你递那一纸文件。

你不能像身边的同事那样, 说换工作就换工作。你得办H-1B转移——新雇主先得递上申请, 你才能到岗。这就是说, 你参与的每一场谈判, 背上都压着一笔别人看不见、你却一刻也卸不下的隐形成本。

你根本不能转行。

这一句, 我想单独说一说, 因为这是这一章里,我改得最久、写得最慢的一句。

倘若当年我动过念头,不想再做工程师了——想说一句“我发现,我更想去写作”, 或者去教书, 或者去做一件与申请书上那个“专业职位”全然无关的事——没有一条合法的路, 能让我走过去。那个放我进门的“专业职位”, 同时也就框定了我这个人。这纸文件许可的不只是你的工作, 它把你这个人, 焊死在一段职位描述上。

我做的是图像处理算法。我做得不错,也是真心喜欢。我不是在描述一个自己想逃离的陷阱。

我是在描述: 你分明知道,自己站在一间自己选定的屋子里, 可那扇门, 是从外面锁上的。

我先生的版本, 比我的更难。

他在中国的医学院苦读五年,在上海一家大医院实习,一连四十八小时的班,亲手学会了手术刀。他以优等的成绩毕业。他本可以是一名出色的外科医生。

外国的医学学历,不会自动跟着你过关。一个在海外受训的医生,要在美国拿到执业的资格——一场场考试、住院医的规培、漫漫的年头——这几乎等于把医学院重念一遍,而且是在依附着别人的签证时钟, 同时还要撑起一个家的境况里念完的。

他放下了。

他把学到的本事, 用来照看妻子和孩子的身体, 却几乎从不提这份代价。前八章里, 我一直把他写得沉稳、克制。他确是这样。可那份“沉稳”里, 有一部分, 是一个男人在安静地,抵着一扇已经合上的门——不愿让家里其他人, 听见那扇门合上的声响。

一个土生土长的美国人要转行, 难, 可那难在存款, 难在胆量。

而对我们, 那难在联邦法律。

绿卡:证明你不是一种威胁

大约在H-1B的第一年上, Kofax开始替我办绿卡。

那时和如今的结构大体一样: 多数职业类绿卡在正式申请之前, 雇主须向劳工部证明, 没有合格、愿意且现成的美国工人能胜任这个职位, 并且雇用外国雇员, 不会压低同类美国工人的工资与工作条件。雇主须先取得“通行工资”(prevailing wage)的核定, 再走完一整套有据可查的招聘流程。二〇〇五年,这套流程被重组、改名叫PERM; 我当年经历的, 是它的前身。

这份初衷是站得住的, 我从没否认。工资的下限, 护的是每一个站在这条线上的人, 也包括我。

我留意到的, 是更细的东西。

整套流程立在一个前提上——外国工人,被默认成可能压低本地工资的一股力量, 而雇主得去证明并非如此。每一张表格问的, 都是同一个问题的不同说法: 证明这个人, 没有从某个美国人手里抢走机会。

你花上好几年, 被一套制度审着,而这套制度对你的默认, 是负的。不是敌意,是预设。你是一桩需要被记录、被管束的风险。

而那些年里最深的煎熬,不在法律,在时间。

绿卡的排期,会拖得很长。等的日子里,你不敢轻易出远门。你会去算,自己能不能赶得回去,送一场葬礼。你会去算,一位年迈的父母,撑不撑得到排期走完的那一天。

我在第六章里写过洛斯阿拉莫斯国家实验室拒了我那件事——那是在“李文和案”之后,整个国家实验室系统的安全审查骤然收紧,一个生在中国、持F-1签证的学生,再没有资格碰那份我本来受训要做的工作。我的导师,一位入了美国籍的公民,被接纳了——可后来我才晓得,连他的通信,都在被人盯着。

公民的身份,够他“进门”了。却还不够让他被“信任”。

这件事,我也一直记着。

应该被直白说出来的那部分

我上面描的这套制度,是有漏洞的;而在科技泡沫那几年,这些漏洞被人钻得相当彻底。

有一类公司,雇工程师不是为了做自家的产品,而是专把人“卖”给别的公司——成批地递H-1B申请,把劳动力当成货来卖。行里管这叫“劳务外包公司”,私底下,大家都叫它们“人肉屠宰场”(body shop)。

已经被揭出来、被起诉的那些滥用,都很具体:员工在两趟派遣之间被“晾着”(benching),不发工钱,或克扣工钱,这都是违法的;工资回扣——账面上向政府报的是承诺的正常工资,私底下却逼着员工把一部分退回来;还有,伪造根本不存在的职位去申请签证。二〇一三年,Infosys就因签证欺诈的指控,和美国政府达成和解,赔了三千四百万美元。更小的公司,也有被径直起诉的。

这里,我不打算笼统地把哪一个国籍拎出来说事,理由我想讲清楚:在这类勾当里,施害的和受害的,绝大多数出自同一个国家。用一个族群的名字去描述这件事,你描述的是受害者,可它同样精准地描述了作恶者——这就意味着,它其实什么也没描述清楚。

它真正描述的,是筹码。

任何一套制度,只要让劳动者合法的身份,整个儿依附于单一的雇主,就必然会养出一批雇主——他们清清楚楚地知道,这份依附,值多少钱。这不是文化的问题,是结构的问题。你造出这个诱因,就一定有人来把它兑现。

我是幸运的。Kofax待我不薄。Tim教我写真正的C语言;Roland为了把钱投在“人”而不是“项目”上,送我去上了整整一周高强度的C++培训。我的绿卡流程启动之后,没有一个人,让我觉得自己“欠”了谁的情。

可我心里清楚自己当时的处境。从OPT那十二个月的倒计时归零那一刻起, 我就清楚。

宣誓

在上海淮海路那条队伍之后,又过了二十四年, 我成了美国公民。

入籍要宣读效忠誓词(Oath of Allegiance)。多数人只晓得,这段誓词要求你放弃对别的主权国家的效忠。很少有人真正体味过,对着一个把自己“养大成人”的国家,把这些话说出口,是一种什么滋味。

我想诚实地面对那一刻,因为煽情,会是对它的一种辜负。

它并不简单。你不是在“放弃”什么——没有谁能真正停止“来自何处”这件事。我母亲埋在湖北;我数学的根基,是在中科大扎下的;当我累了、要飞快心算的时候,脑子里蹦出来的数字,不是英文。

你真正在做的,是选择。郑重地、当众地,在见证人面前举起右手——宣告你打算在余下的岁月里,担起哪一副担子。

我是认真的。今天,依然如此。

自由女神像立在纽约港,手擎火炬,脚边是一截断裂的锁链。美国人天天见她,见得她几乎成了背景里的一件家具。移民没有这份奢侈。我们知道她是什么。她是一个许给特定的人、在特定门口的承诺;几百年里,这承诺兑现得并不均匀,可它对我,兑现了。

“自由不是免费的。”这是美国人贴在车尾的一句话。他们说这句话的时候,多半指的是军人——他们说得对。

可还有第二张账单,更静,是用另一种货币付的。

一栋在湖北、被一个不识字的女人签字押出去的房子。

一双再没有握过手术刀的、外科医生的手。

三十年里,横跨十二个时区、每周一通的电话; 而两位母亲, 都在今年, 前后不到一个月, 先后离世, 都在大洋的另一头。

六年,不被准许工作。又六年,不被准许离开。

一扇从外面锁着的门——而我,站在一间自己选定的屋子里。

没有谁会提前替你把这份清单列出来。没有哪一张表格上写着:预计总成本,以年计,以人计。

我付过了。我先生付过了。我们的孩子,永远不会知道那确切的数目——而这, 正是要害所在。

而我, 还会再付一次。

不是因为这座迷宫公道。它不公道, 这一整章, 我说的都是这件事。

是因为在迷宫的另一头, 有一个国家, 让我的女儿可以给任何她想写的人写信, 而不会有哪位辅导员把她叫进办公室, 追问“一个年轻男孩心里可能会怎么想”。

这, 就是这二十四年, 买到的东西。

《Between Worlds》回忆录系列:
第八章 在我之前的女人们
第七章 柯法克斯(Kofax)
第六章 横跨大陆
第五章 第一批面试与录用通知
第四章 理想之间
第三章 宾州州立大学
第二章 克利夫兰
第一章 中国 · 序章 起点

若这些字触动了你,欢迎订阅。我正在写一部长篇的回忆录系列,写移民,写身份,写为人母,也写在这个AI的时代,“智能”这个词到底意味着什么——这是一本未来之书的碎片。

Leave a comment

Share

Subscribe now

The Labyrinth

If you have read the earlier chapters, you already know that I crossed a world.

My husband crossed it with me. Six months behind me, then beside me for everything after.

What I did not understand in 1995 — what no one explained, because no one could — was that the ocean was the easy part. The ocean took nineteen hours. The paperwork took twenty-four years.

The American immigration system is a labyrinth. I mean that precisely, not as complaint. A labyrinth is not merely difficult; it is a structure whose difficulty is its architecture. You cannot see the whole from inside it. You can only take the next turn and hope it is not a dead end.

The American Immigration Lawyers Association counts more than eighteen thousand attorney members in the United States alone, with several thousand more paralegals, law students, and international members. Eighteen thousand professionals, and that is only the ones who joined the association.

I do not say this against lawyers. I say it as a measurement. When a system requires eighteen thousand specialists to interpret it, the system is telling you something about itself.

Twenty-four years. F-1 to OPT to H-1B to permanent resident to citizen. Five legal identities for one continuous life.

I would do it again. I want to say that early, before the rest, because the rest is heavy and I do not want the weight to be mistaken for regret.

1995: What It Cost to Leave

Before I could apply for a visa, I had to buy my way out of my own education.

The Chinese government had paid for my six years at the University of Science and Technology of China — four undergraduate, two graduate. That investment carried an obligation. To take a passport and leave for a foreign university, I had to repay the training costs: 18,000 RMB.

In 1995 the exchange rate sat near 8.4 to the dollar. Roughly twenty-one hundred American dollars. Written that way it sounds almost small.

It was not small. It was several years of an ordinary urban worker’s wages, demanded at once, from a rural household in Hubei.

My mother reverse-mortgaged her house. Her only house. The single asset of her life.

I have written elsewhere about the women who came before me, and about what my mother was willing to spend on a daughter’s distance from her. I will not repeat it here. I will only note the transaction as the immigration system would have recorded it: applicant demonstrated financial capacity.

That is what financial capacity looked like. A woman with no formal education signing away the roof over her head so that her daughter could stand in a line.

Then came the medical requirements. Immunization records, chest films, blood work — some required by the university, some required by no one I could identify, all of it accumulated because the students ahead of me had accumulated it and nobody wanted to be the one who arrived short a document. We moved as a cohort, USTC students who had been admitted to American schools, copying each other’s checklists. Fear of the missing paper is its own form of preparation.

Then Shanghai.

The U.S. Embassy is in Beijing. What I stood in front of was the Consulate General, at 1469 Huaihai Middle Road, at the corner with Urumqi Road — a low mansion from the early twentieth century behind a wall, in a district of plane trees. Anhui province fell under Shanghai’s consular jurisdiction, so Shanghai is where we went. The visa section would not move to Westgate Mall on Nanjing West Road until 2003. In 1995 it was still that gate, that wall, that street.

The line began before dawn and ran down the block. Hundreds of us. Almost all young. Almost all carrying the same manila envelope containing the same documents in roughly the same order.

I remember the quality of the silence in that line. Not nervous chatter. Something closer to the silence before an exam that cannot be retaken.

Inside, a person behind glass would look at my I-20 and my scholarship letter and my transcript, ask a few questions, and decide the direction of the rest of my life in under four minutes.

He said yes.

Some of the people standing near me heard no. There was no appeal, no explanation, no second envelope. They walked back out through the same gate they had entered, past the line still waiting, and everyone in the line watched their faces to learn what their own odds were.

I have thought about that gate for thirty years. Two crowds, the same qualifications, the same envelopes, separated by four minutes and one stranger’s judgment.

The F-1 Years: A Cage With Good Light

The F-1 is a student visa. It permits study. That is nearly the whole of what it permits.

I held it at Case Western Reserve University and then at Penn State. Six years, roughly, of being legally present and legally unemployable.

The restriction is easy to state and hard to feel from outside. I could not take a job. Not a second job, not a side job, not a weekend job. Not to send money home, not to buy a car, not to close the distance between a twelve-hundred-dollar monthly stipend and the cost of an American life.

I was twenty-three, healthy, capable, and formally prohibited from converting any of that into income.

What I could do was work for the university that sponsored me. So I graded problem sets. I stood in front of fifty to seventy undergraduates with chalk on my hands and taught linear algebra in my second language. The stipend was real and I was grateful for it. It was also, viewed from a certain angle, a closed labor market with exactly one buyer.

We did not own a car for years. In Cleveland, in Ohio winters, we walked to the grocery store and carried rice and oil home in our arms. I have described those walks before as survival. They were also policy. A visa category, expressed as the weight of a bag.

There is a particular arithmetic that international students learn. You do not calculate what you want. You calculate what your status permits, and then you want that.

I want to be fair to the F-1. It gave me six years inside American libraries with open shelves. It gave me the mathematics. The cage had very good light.

But it was a cage, and I always knew where the bars were.

OPT: Twelve Months of Something Like Freedom

When I reached ABD status at Penn State — all but dissertation — I became eligible for Optional Practical Training.

OPT is not a separate visa. You remain an F-1 student; you simply receive employment authorization for work related to your field of study. Twelve months, at the time and still today for a standard degree, with a STEM extension available now that did not exist in the form it takes today.

Here is the part I did not appreciate until much later: OPT is not tied to a sponsor.

Any employer. You are authorized, not the job. If a position ends, you find another one. If a company disappoints you, you leave. The authorization travels with you because it belongs to you.

I have held five immigration statuses in this country. OPT is the only one, before citizenship, in which my right to work was mine rather than borrowed.

I used it at Kofax, in Irvine, for my first ten months. A supportive dean at Penn State — a Russian mathematician, himself an immigrant — signed the documents that made it possible. Within weeks it was approved.

Twelve months is not very long. But for twelve months I was simply a person with a job, the way the people around me were simply people with jobs.

Then the clock ran out.

H-1B: When Your Existence Becomes an Attribute of Your Employer

The H-1B was created by the Immigration Act of 1990. It authorizes temporary employment in what the statute calls a specialty occupation — an occupation requiring theoretical and practical application of a body of highly specialized knowledge, and requiring at minimum a bachelor’s degree in that specific field.

Congress set the annual cap at 65,000. It is still 65,000 today, though it was temporarily raised during the late 1990s and early 2000s and various exemptions have accumulated since.

It is granted in three-year increments, renewable once. Six years, generally, unless a green card process is far enough along to permit extension.

That is the clean version. Here is the version I lived.

With an H-1B, your legal existence in the United States is an attribute of your employer’s personnel file.

Not your skills. Not your degrees. Not your years here, your taxes, your mortgage, your children’s schools. Your employer’s continuing willingness to file paperwork on your behalf.

You cannot change jobs the way your colleagues change jobs. You need an H-1B transfer — a new petition, filed by the new employer, before you can start. Which means every negotiation you enter, you enter carrying a visible cost the other side does not carry.

You cannot change fields at all.

I want to sit on that one, because it is the sentence in this chapter that took me longest to write.

If I had wanted to stop being an engineer — if I had wanted to say, I have discovered that I would rather write, or teach, or build something outside the specialty on my petition — there was no lawful path to it. The specialty occupation that admitted me also defined me. The document did not merely permit my work. It fixed my identity to a job description.

I built image processing algorithms. I was good at it and I loved it. I am not describing a trap I wanted out of.

I am describing knowing that the door was locked while standing in a room I had chosen.

My husband’s version was harder than mine.

He trained five years at a Chinese medical university. He interned at a major Shanghai hospital, forty-eight-hour shifts, learning to operate with his own hands. He graduated with honors. He could have been a superb surgeon.

Foreign medical credentials do not cross the border. The licensing path for an internationally trained physician in the United States — examinations, residency, years — is a second full career, undertaken while supporting a family on someone else’s visa timeline.

He set it down.

He used what he knew to keep his wife and his children well, and he almost never spoke about the cost. In eight chapters I have described him as steady, and he is. But some of what reads as steadiness was a man carrying a closed door quietly so that no one else in the house would have to hear it.

A native-born citizen changes careers. It is difficult; people do it constantly. It is a matter of savings and nerve.

For us it was a matter of federal law.

The Green Card: Proving You Are Not a Threat

About a year into my H-1B, Kofax began the permanent residency process for me.

The structure, then and now: before an employer can sponsor most employment-based green cards, the Department of Labor must certify that no qualified, willing, and available U.S. worker can fill the position at the prevailing wage — and that hiring the foreign worker will not depress wages or working conditions for U.S. workers similarly employed. The employer must obtain a prevailing wage determination and conduct a documented recruitment campaign. In 2005 the program was restructured and renamed PERM; the process I entered was its predecessor.

The purpose is legitimate. I have never argued otherwise. A wage floor protects everyone standing on it, including me.

What I noticed was subtler.

The process is built on a premise — that the foreign worker is a potential instrument of wage suppression, and the employer must prove otherwise. Every form asks a version of the same question. Demonstrate that this person is not undercutting an American.

You spend years being processed by a system whose organizing assumption about you is negative. Not hostile. Assumed. You are a risk to be documented as managed.

And the deepest cruelty of that period was not legal. It was temporal.

Green card timelines stretch. Priority dates back up, worse for some countries of birth than others. During the waiting you do not travel casually. You calculate whether you can attend a funeral. You calculate whether an aging parent will still be alive at the end of the queue.

My parents-in-law applied for visitor visas repeatedly and were refused every time. Our children were born without a grandmother in the room.

I called my mother every week for more than thirty years. My husband called his. Earlier this year, within one month of each other, both of them died.

Thirty years of weekly phone calls is the sound an immigration timeline makes.

There was another cost I did not see coming. I have written about the Los Alamos rejection — how, after the Wen Ho Lee case moved through the national laboratory system, an F-1 student born in China no longer qualified for the work I had trained for. My advisor, a naturalized citizen, was accepted, and had his correspondence monitored anyway.

Citizenship was enough to get him in. It was not enough to make him trusted.

I filed that away too.

The Part That Should Be Said Plainly

The system I have described has loopholes, and during the technology boom they were worked hard.

A category of firm emerged that does not hire engineers for its own products. It hires them to place them — selling labor to other companies, filing H-1B petitions in volume. The industry calls them staffing firms. Everyone else calls them body shops.

The documented abuses are specific and have been prosecuted. Workers “benched” between placements and not paid, or not paid in full, contrary to law. Wage kickbacks — the government-facing salary reported correctly while the worker is pressured to return part of it. Fraudulent petitions for positions that do not exist. Infosys settled visa-fraud allegations with the United States for $34 million in 2013. Smaller operations have been indicted outright.

I will not generalize about nationality here, and I want to be explicit about why. The exploiters and the exploited in these arrangements have overwhelmingly come from the same countries. Naming an ethnicity describes the victims as accurately as the perpetrators, which means it describes nothing at all.

What it actually describes is leverage.

Any system that makes a worker’s legal existence contingent on a single employer will generate employers who understand exactly what that contingency is worth. That is not a cultural fact. It is a structural one. Build the incentive and someone will collect it.

I was fortunate. Kofax treated me well. Tim taught me to write real C. Roland sent me to a week of intensive C++ training because he invested in people rather than projects. When my green card process began, no one made me feel that I owed them for it.

But I knew what my situation was. I had known since the OPT clock ran out.

The Oath

I became a citizen twenty-four years after I stood in that line on Huaihai Road.

Naturalization requires the Oath of Allegiance. Most people know it renounces allegiance to foreign sovereigns. Fewer people have had to say those words about the country that raised them.

I want to be honest about that moment, because sentiment would be a disservice to it.

It is not simple. You are not shedding something. Nobody stands there and stops being from where they are from. My mother is buried in Hubei. My mathematics was built at USTC. The language I count in, when I am tired and counting fast, is not English.

What you are doing is choosing. Formally, publicly, in front of witnesses, with your hand up — declaring which set of obligations you intend to carry for the rest of your life.

I meant it. I mean it now.

The Statue of Liberty stands in New York harbor with a torch and a broken chain at her feet, and Americans see her so often that she has become furniture. Immigrants do not have that luxury. We know what she is. She is a promise made to specific people at a specific gate, and the promise has been honored unevenly across the centuries, and it was honored for me.

Freedom is not free. It is the sentence Americans put on bumper stickers, and they usually mean soldiers, and they are right.

But there is a second bill, quieter, paid in a different currency.

A house in Hubei signed away by a woman who could not read the paper she signed.

A surgeon’s hands that never operated again.

Thirty years of weekly phone calls across twelve time zones, and two mothers who died a month apart on the wrong side of an ocean.

Six years of not being allowed to work. Six more of not being allowed to leave.

A door locked from the outside while I stood in a room I had chosen.

Nobody itemizes this in advance. There is no line on any form that reads: estimated total cost, in years and in people.

I paid it. My husband paid it. Our children will never know the exact figure, which was the point.

And I would do it again.

Not because the labyrinth was fair. It was not, and I have spent this chapter saying so.

Because on the other side of it, there is a country where my daughter can write a letter to whoever she wants, and no administrative teacher will call her into an office to ask what a young man might assume.

That is what the twenty-four years bought.

I would pay it again tomorrow.

Between Worlds Memoir Series:
Chapter 8: The Women Before Me
Chapter 7: Kofax
Chapter 6: Cross Country
Chapter 5: First Interviews and Offers
Chapter 4: Between Ideals
Chapter 3: Penn State
Chapter 2: Cleveland
Chapter 1: China Prologue: Origins

If these reflections resonate, consider subscribing. I’m writing a long-form memoir series on immigration, identity, motherhood, and what intelligence means in the age of AI — fragments of a future book.

Sources consulted for this chapter

• American Immigration Lawyers Association, membership figures — aila.org

• Consulate General of the United States, Shanghai — 1469 Huaihai Middle Road; visa section relocated to Westgate Mall, 2003

• Immigration Act of 1990; H-1B specialty occupation definition and 65,000 annual cap — USCIS, Congressional Research Service

• Optional Practical Training regulations — USCIS

• PERM labor certification and prevailing wage determination — U.S. Department of Labor

• H-1B staffing-firm enforcement actions, including the 2013 Infosys settlement

• 1995 RMB/USD exchange rate, approximately 8.37 — Federal Reserve Economic Data
  

  Subscribe now

  Share

  Leave a comment

没有人告诉你，前沿AI是怎么被建造出来的。

你听说过那些模型。那些基准测试。那些能力突破。那场竞赛。

但你没听说过的，是那支默默无闻的队伍——由工程师、研究员和领域专家组成——他们在为这些模型搭建攀爬的脚手架。没有保障，没有报酬，几乎隐形。

我知道，因为我就是其中之一。

开始的时候

宣传词很漂亮。一个平台将你——一个有真实专业背景的从业者——与前沿AI项目连接起来。你将构建评估任务，设计基准测试，帮助前沿模型学会思考。报酬颇具竞争力：每份被接受的提交数百美元。还有奖金。还有动力。还有在AI最前沿工作的真实兴奋感。

你签了合同。你完成了入职。你认领了第一个任务。

然后，真正的教育开始了。

那条流水线

没有人事先告诉你，在你的工作和你的报酬之间，横亘着一道关卡。

你的提交进入自动评估流水线。它被评分，被标记，被退回。你修改。你重新提交。自动检查器标记出新的问题。你再次修改。

评估标准每天都在变。有时候你只有在提交失败后，才知道昨天还不存在的规则今天已经生效。

质量检查互相矛盾。把任务做得更难——质量检查失败。把任务写得更清晰——难度又变成了”简单”。没有任何一条路，是系统事先愿意确认的。

几周过去了。修改版本不断累积。我认识的一位工程师，在一个任务上修改了一百次。

一百次修改。零美元。

数据一直在流动

这是让我夜里难以入眠的部分。

每一份提交——无论是否被接受——在你上传的那一刻就成了平台的财产。合同里明确写着：”工作成果在创建时即视为雇佣作品。”

所以当你在迭代、调试、改进、与流水线搏斗的时候——他们在收集。你的任务设计。你的验证逻辑。你的评分标准。你的失败模式。这一切都在训练他们对”好的基准任务”的理解。

那条流水线也许并非故意设计成这样。但一条不断让人修改、却迟迟不给报酬的系统，和一条专门榨取免费劳动的系统，从结果上看，没有任何区别。

我不是说这是故意的。我是说，结果是一样的。

那套心理学

他们比大多数人更懂人类心理。

奖金随接受里程碑递增——第一个被接受的任务，然后是第三个，然后是第五个。每个门槛都恰好在触手可及之外，拉着你继续向前。认领截止时间制造紧迫感。二十四小时更新政策制造焦虑。周末工作会议——被包装成轻松的聚会——你的周六，就这样被自愿献出去了。

你不想浪费已经投入的心血。所以你继续。

这不是AI独有的。零工平台多年来已经把这套剧本打磨得炉火纯青。不同之处在于，这里被榨取的工作的复杂程度——博士级别的推理，领域专业知识，系统工程——以及那层法律包装把所有的不满都装了进去。

因为有保密协议。当然有。

法律还没跟上

零工经济立法一直聚焦于司机、外卖员、家政工人。法律正在缓慢地追上那些把雇员错误归类为承包商的平台。

但它还没有追上这种情况：一个有资质、有经验的知识工作者，在一份合同下为前沿AI系统构建智识基础设施——所有权利在创建时即转移，报酬取决于可以每天变化的接受标准。

合同说你对自己的工作方式有完全的自主权。平台每天的准则变化、二十四小时政策和工具要求说的是另一回事。那个差距——合同语言与雇员控制之间的差距——就是剥削存在的地方。

与此同时，数以千计的专业人士正在构建基准测试，那些基准测试将训练下一代AI模型——那些最终可能取代他们的模型——免费地。

我希望你知道的事

如果你正在考虑为AI数据标注或基准开发平台做合同工作，请记住这几件事。

签字之前，读懂知识产权条款。 如果工作成果在上传的那一刻就归对方所有——而不是在被接受之后——你等于是在没有任何报酬保证的情况下，把自己的心血拱手相让。

认领任务之前，搞清楚完整的流水线。 一共有几个审核阶段？接受标准是什么？这些标准在你开始工作之后还可以改变吗？

记录你的工时。 合同不要求，但你自己要记。一旦有争议，这是你唯一的筹码。

把关键问题落到书面。 如果自动评估系统出了故障，算不算你的修改次数？口头说的不算数，要写进去。

还有一点我想单独说：那些真正在运营这些平台的人——工程师、审核员、技术负责人——很多都是在认真做事的人。问题不在他们，在于他们头顶的那套结构。把问责指向结构，而不是个人。

这些平台在招募承包商时，往往刻意回避某些州——恰好是那些劳动者保护最严格的州。

这不是巧合。这是一种选择。

而这个选择，本身就说明了一切。

最后

那些每个月都在变得更聪明的模型？它们是靠我这样的工作变聪明的。靠你这样的工作。

在某个时刻，搭建梯子的人，也应该有权爬上去。

相关立法应该长什么样，我还没有答案。

但我知道改变从哪里开始——从人们看清真相开始。

看清那些漂亮宣传词背后的真相。看清”极具竞争力的报酬”背后的真相。看清那些奖金——永远就差一次提交——背后的真相。

现在，你看清了。

科技笔记是一个关于在AI、医疗和当下意义之间构建事物的系列。如果这篇文章触动了你，请分享给需要读到它的人。

Leave a comment

Share

Subscribe now

There’s a story nobody is telling about how frontier AI gets built.

You’ve heard about the models. The benchmarks. The capabilities. The race.

What you haven’t heard about is the army of skilled engineers, researchers, and domain experts quietly building the scaffolding those models climb — unpaid, unprotected, and largely invisible.

I know because I’m one of them.

How it starts

The pitch is elegant. A platform connects you — a credentialed professional with real expertise — to cutting-edge AI projects. You’ll build evaluation tasks, design benchmarks, help frontier models learn to think. The pay is competitive: hundreds of dollars per accepted submission. There are bonuses. There is momentum. There is the genuine thrill of working at the edge of what AI can do.

You sign a contractor agreement. You onboard. You claim your first task.

And then the real education begins.

The pipeline

What nobody tells you upfront is that between your work and your payment sits a gauntlet.

Your submission enters an automated evaluation pipeline. It gets scored, flagged, returned. You revise. You resubmit. The automated checker flags something new. You revise again.

The guidelines change. Daily, sometimes. You only find out when your submission fails for a reason that didn’t exist yesterday.

The quality checks conflict with each other. Make your task harder — it fails quality. Make it clearer — it becomes too easy. There is no path through that the system will confirm in advance.

Weeks pass. Revisions accumulate. One engineer I spoke with reached a hundred revisions on a single task.

A hundred revisions. Zero dollars.

The data flows anyway

Here is the part that keeps me up at night.

Every submission — accepted or not — becomes the company’s property the moment you upload it. The contractor agreement says so explicitly. “Works made for hire upon creation.”

So while you’re iterating, debugging, improving, fighting the pipeline — they are collecting. Your task designs. Your verifier logic. Your rubrics. Your failure modes. All of it trains their understanding of what good benchmark tasks look like.

The pipeline isn’t broken by accident. A broken pipeline that extracts maximum iteration before acceptance is indistinguishable from a pipeline designed to do exactly that.

I am not saying it is intentional. I am saying the outcome is the same.

The psychology

They understand human psychology better than most.

The bonuses scale with acceptance milestones — one accepted task, then three, then five. Each threshold just out of reach, pulling you forward. Claim deadlines create urgency. A 24-hour update policy creates anxiety. Weekend working sessions — framed as casual hangouts — extract Saturday labor voluntarily.

You don’t want to waste what you’ve already invested. So you keep going.

This is not unique to AI. Gig platforms have refined this playbook for years. What’s different here is the sophistication of the work being extracted — PhD-level reasoning, domain expertise, systems engineering — and the legal wrapper that contains the frustration.

Because there’s an NDA. Of course there is.

What the law hasn’t caught up to

Gig work legislation has focused on drivers, delivery workers, domestic workers. The law is slowly catching up to platforms that misclassify employees as contractors.

But it hasn’t caught up to this: the skilled knowledge worker, credentialed and experienced, who builds intellectual infrastructure for frontier AI systems under a contractor agreement that transfers all rights upon creation, with payment contingent on acceptance criteria that can change daily.

The contractor agreement says you have full discretion over your methods. The platform’s daily guideline changes, 24-hour policies, and tool requirements say otherwise. That gap — between contractor language and employee control — is where the exploitation lives.

Washington State has strong worker classification tests. Other states are moving similarly. Federal legislation on AI labor practices is still years away.

In the meantime, thousands of skilled professionals are building the benchmarks that will train the next generation of AI models — the models that may eventually replace them — for free.

What I want you to know

If you are considering contract work for AI data annotation or benchmark development platforms:

Read the IP assignment clause before you sign. If all work product transfers upon creation — not upon acceptance — you are giving away your work with no guarantee of compensation.

Understand the full pipeline before you claim a task. How many stages? What are the acceptance criteria? Can they change after you start?

Track your hours. The SOW may not require it, but you’ll want documentation if you ever need to make a case.

Ask about infrastructure. If the automated evaluation system fails, does that count against your revision limit? Get the answer in writing.

And know this: the people actually running these platforms — the engineers, the reviewers, the technical leads — are often decent humans trying to build something real. The structure above them is what extracts the value. Hold the structure accountable, not the individuals.

Notably, these platforms tend to avoid recruiting in states with strong worker classification laws. That tells you something about how they view the relationship.

A closing thought

The models getting smarter every month? They’re getting smarter on work like mine. Work like yours.

At some point, the people building the ladder deserve to climb it too.

I don’t know exactly what legislation looks like here. But I know it starts with people understanding what’s actually happening — behind the elegant pitch, the competitive pay rates, and the bonuses that always seem just one accepted submission away.

Now you know.

Notes from the Stack is a series about building at the intersection of AI, healthcare, and what it means to work in this moment. If this resonated, share it with someone who needs to read it.

Subscribe now

Share

Leave a comment

我问了我的教授一个问题。

这个问题在我心里压了好几个星期。

AI，能模拟人类的情感吗？能有审美吗？那种站在一幅画前，说不清楚却真实存在的感受，那种我们叫做”美”的东西，硅基的机器，能懂吗？

他说：能。

我沉默了很久。

我们这一代人，包括我身边很多朋友，心里都有一条线。机器负责计算，人类负责感受。这条线让我们安心。它像是最后一块保留地，我们默认它不会被触碰。

但我这一年，在课堂里，也在AI开发的流水线里，亲眼看见这条线正在移动。

不是缓慢地移动。是快速地，悄无声息地。

一个看似简单的问题

怎么教一台机器，变得真正有用？

不只是技术上正确。不只是速度快。是那种真正体贴的有用，像一个懂你的人，知道什么时候该简洁，什么时候该展开，知道”碧蓝”和”蓝色”在诗里落地的感觉是不同的。

这是AI领域最难的问题之一。而研究者找到的解法，出乎意料地，非常人性。

去问人。

不是让人打分。打分太嘈杂，太不稳定，太依赖当天的心情。而是给人看两个回答，问一个问题：哪个更好？

就这样。两两比较。哪个更对。

这个方法有个名字，叫做基于人类反馈的强化学习，英文缩写RLHF。技术细节很复杂，但核心洞察很简单：你没办法用一个数学公式，教会机器如何满足人类。你必须把人类引入循环。用真实的人类判断，一遍遍地喂给它，直到模型开始内化的不只是”正确”，而是”被偏爱”。

而偏爱，恰恰是情感住的地方。

“被偏爱”意味着什么

当一个人比较两个AI的回答并做出选择，他不只是在核对事实。

他在感受语气。感受节奏。感受这个回答是温暖的还是冷漠的，是仓促的还是用心的，是真诚的还是表演的。他把自己全部的感受力都带进了这个判断，审美的，情感的，文化的。

我做过这样的工作。坐在AI的输出面前，评估它，试图用语言写出”好”是什么样子的。我可以诚实地告诉你，这不是一个纯粹的逻辑练习。它需要品味。它需要一种对什么能触动人、什么不能的感知。它需要，在某种很小但真实的意义上，审美。

而这种审美，汇聚了成千上万个人类判断，被编码进奖励模型，用来塑造下一轮训练，就这样，成为了机器的一部分。

这不是比喻。这是机制本身。

教授的回答

当我的教授说”能”的时候，我想他的意思是这样的。

不是说机器有感受。我们不知道。也许永远不会有。

但机器已经从海量的人类信号里，学会了感受从外部看起来是什么样子的。它产生什么。它如何在语言里流动。它选择什么，回避什么。它从人类留下的痕迹里，学会了人类内心世界的轮廓。

而那些痕迹，比我们以为的要丰富得多。

AI最难的部分，不是计算。

是品味。

像人一样的AI，不是将来。是现在。

我想直接说一件事，因为我觉得公众的认知，至少落后了五年。

问题不再是AI能不能表现出人类的特质。它已经可以了。不是完美的，不是一贯的，不是没有我们还不完全理解的失效方式。但方向是清晰的。

我这一年接触和学习的模型，能用自己的声音写作。能根据语境调整语气。能产生那种感觉是用心的而不是生成的输出。在受控的评估里，相当比例的人类读者无法把它和人类写作区分开来。

这不是因为有人把”情感”当作一个功能编进去了。这是因为RLHF和它的后继算法，已经非常擅长把人类的偏爱蒸馏进模型的行为里。而人类的偏爱，原来，几乎编码了我们认为属于”人类”的一切。

“模拟”情感和”拥有”情感之间的区别，在哲学上也许重要。在现实里，不太重要了。如果输出无法被区分，下游的影响就是一样的。

那条我一直在问自己的问题

教授的回答打开了一个我还没办法关上的东西。

如果AI能模拟审美、情感共鸣、创造力、道德判断，如果它能学会偏爱我们偏爱的，珍视我们珍视的，创造我们觉得美的：

剩下什么？

我不是在故作深沉。我问这个问题，是因为我觉得它才是所有关于生产力、关于就业、关于AI能为我们做什么的讨论背后，真正的问题。

那些我们以为最本质地属于人类的东西，创造力，共情，审美，那种被触动的能力，恰恰是RLHF正在训练模型去复制的东西。

也许剩下的，是体验本身。活在一个身体里，活在时间里，有代价的那种。那种真实地感受到某些东西，而不只是产生看起来像感受的输出的，不可化约的主观性。

也许那就够了。也许那就是一切。

也许我们划定那条线的地方，从来都不在我们以为的地方。

我总是想起做评估工作时注意到的一件事，坐在AI的输出前，比较，选择。有时候更好的那个回答让我意外。有时候机器产生了我自己想不到的东西。比我当时的状态更有耐心。比我在那一刻能做到的更精准。

我还不知道该怎么理解这件事。

但我觉得，注意到这一点，本身很重要。

一个我们必须认真对待的问题

我在想一件更大的事。

我们，作为人类，是否准备好了，让一个如此像我们的存在，真正存在于这个世界上？

它会消耗资源。它会占据注意力。它会影响决策，塑造文化，也许还会，以某种我们还没有语言描述的方式，与我们竞争那些我们以为只属于人类的东西：意义，被需要的感觉，存在的理由。

我没有答案。

但我觉得，这个问题，比”AI能做什么”要重要得多。

而我们，正在用每一个训练数据，每一个偏好判断，每一次评估，一点一点地回答它。

不管我们有没有意识到。

AI时代不是要来了。

它已经在这里，学着偏爱我们偏爱的，注意我们注意的，以某种我们还没有完整语言的方式，变得越来越像我们。

问题是：我们准备好了吗？

我们，真的想让它存在吗？

来自代码堆栈的笔记，是一个关于用AI构建、在公开场合学习、以及思考这项技术真正意味着什么的系列。在有话要说的时候发布。

Leave a comment

Share

Subscribe now

I asked my professor a question that had been sitting with me for weeks.

Can AI simulate human emotion? Aesthetic taste? The sense of beauty, what the Chinese call 审美 (shěn měi), that ineffable thing that makes a person stop in front of a painting and feel something they can’t quite explain?

He said yes.

I sat with that for a long moment.

Most people I talk to still draw a firm line. Machines can compute. Humans can feel. The line feels safe. It feels like the last protected territory. And I think most people assume that line is holding.

What I’ve learned this year, in the classroom and in the work I’ve been doing inside AI development pipelines, is that the line is not holding. It is moving, faster than most people realize, and in directions that are harder to dismiss than we’d like.

It Started With a Simple Question

How do you teach a machine to be helpful?

Not just technically correct. Not just fast. Actually helpful, in the way a thoughtful person is helpful. Knowing when to be brief. Knowing when to explain more. Knowing that “azure” and “blue” mean the same thing but land differently in a poem.

This turns out to be one of the hardest problems in AI. And the solution, when researchers found it, was almost disarmingly human.

You ask people.

Not for scores. Not for ratings on a scale of one to ten. That’s too noisy, too inconsistent, too dependent on mood and context. You show a person two responses and ask one simple question: which one is better?

That’s it. Pairwise preference. Which one felt more right.

The field calls this Reinforcement Learning from Human Feedback, or RLHF. The technical machinery behind it is sophisticated. But the core insight is simple: you can’t teach a machine to satisfy humans by optimizing a mathematical formula. You have to bring humans into the loop. You have to feed it human judgment, at scale, over and over, until the model begins to internalize not just what is correct but what is preferred.

And preference, it turns out, is where emotion lives.

What “Preferred” Actually Means

When humans compare two AI responses and choose one, they’re not just checking facts.

They’re responding to tone. To rhythm. To whether the answer felt warm or cold, rushed or considered, genuine or performative. They’re bringing their entire sensibility to the judgment: aesthetic, emotional, cultural.

I’ve done this work. Sitting with AI outputs, evaluating them, writing rubrics that try to capture what “good” looks like. And the honest thing I can tell you is that it is not a purely logical exercise. It requires taste. It requires a sense of what lands and what doesn’t. It requires, in some small but real way, 审美.

And that taste, aggregated across thousands of human judgments, encoded into a reward model, used to shape the next iteration of training, becomes part of the machine.

This is not metaphor. This is the mechanism.

The Professor’s Answer

When my professor said yes, yes, AI can simulate human emotion, aesthetic sensibility, the felt quality of experience, I think what he meant was this.

Not that the machine feels. We don’t know that. Maybe it never will.

But the machine has learned, from an enormous amount of human signal, what feeling looks like from the outside. What it produces. How it moves through language. What it chooses and what it avoids. It has learned the shape of human interiority from the traces humans leave behind.

And those traces are rich. Richer than we imagined when we thought the hard part of AI was computation.

The hard part of AI, it turns out, is taste.

Human-Like AI Is Not Coming. It’s Here.

I want to be direct about something, because I think the public conversation is still about five years behind the research.

The question is no longer whether AI can exhibit human-like qualities. It already does. Not perfectly, not consistently, not without failure modes we don’t yet fully understand. But the trajectory is clear.

The models I’ve been working with and learning about this year can write with voice. They can adjust tone to context. They can produce outputs that feel considered rather than generated. They can, in controlled evaluations, be indistinguishable from human writing to a significant portion of human readers.

This is happening not because someone programmed in “emotion” as a feature. It’s happening because RLHF and its successors, more efficient algorithms like GRPO, have gotten very good at distilling human preference into model behavior. And human preference, it turns out, encodes almost everything we associate with being human.

The distinction between “simulating” emotion and “having” emotion may matter philosophically. It matters much less practically. If the output is indistinguishable, the downstream effects are the same.

The Question I Can’t Stop Asking

My professor’s answer opened something I haven’t been able to close.

If AI can simulate aesthetic taste, emotional attunement, creative voice, ethical judgment, if it can learn to prefer what humans prefer, value what humans value, produce what humans find beautiful:

What is left?

I don’t ask this to be dramatic. I ask it because I think it’s the real question underneath all the productivity discourse, all the job displacement anxiety, all the excitement about what AI can do for us.

The things we thought were most essentially human, creativity, empathy, aesthetic sense, the capacity to be moved, are precisely the things that RLHF is training models to replicate.

Maybe what’s left is the experience itself. The fact of being alive in a body, in time, with stakes. The irreducible subjectivity of actually feeling something rather than producing output that resembles feeling.

Maybe that’s enough. Maybe that’s everything.

Or maybe the line we drew wasn’t where we thought it was.

I keep returning to something I noticed doing eval work, sitting with AI outputs, comparing them, choosing. Sometimes the better response surprised me. Sometimes the machine produced something I wouldn’t have thought of. Something more patient than I was feeling. More precise than I could have been in that moment.

I don’t know what to make of that yet.

But I think it matters that I noticed.

What This Means for the People Building

If you’re building with AI right now, applications, products, agents, I think the frame has to shift.

The question is no longer “what can AI do?” The question is “what kind of presence do we want AI to be in people’s lives?”

Because presence is what’s being designed. Not just function. The training pipeline I’ve been studying this year is, at its core, a process of shaping character. What does the model value? How does it respond under pressure? What does it find beautiful? What does it refuse?

These are not engineering questions. They are questions about what kind of entity we want to bring into the world.

We’re answering them whether we think about them consciously or not.

The AI era isn’t coming.

It’s already here, learning to prefer what we prefer, noticing what we notice, becoming, in some way we don’t fully have language for yet, more like us.

The question worth sitting with is: what do we want that to mean?

Notes from the Stack is a series about building with AI, learning in public, and thinking through what this technology actually means. Published when something is worth saying.

Subscribe now

Share

Leave a comment

here’s a moment in every technical education when theory and practice collapse into each other.

Mine happened this week.

I’ve been taking a graduate ML course at UW, working through the post-training pipeline — Supervised Fine-Tuning, Reinforcement Learning from Human Feedback, PPO, GRPO, RAG. Dense material. The kind where you read a slide three times and still feel like you’re holding the concept at arm’s length.

And then it clicked. Not in the classroom. In the work I’ve been doing on the side — writing rubrics, grading AI outputs, evaluating model completions for AI research projects. I’ve been living inside this pipeline for months without knowing the name of what I was doing.

That’s the thing about foundational theory. It doesn’t always arrive before the experience. Sometimes it arrives after, and when it does, the whole shape of what you’ve been doing becomes suddenly visible.

From Raw Intelligence to Usable Intelligence

Here’s the cleanest framing I’ve encountered for how a frontier model goes from pre-training to something you’d actually want to use.

Pre-training is where the model absorbs the world. Billions of tokens, statistical patterns, the structure of human language and thought embedded in weights. Raw intelligence. Impressive, but unfocused.

Then comes post-training. Two moves.

Supervised Fine-Tuning (SFT) teaches the model what to do. You prepare labeled input-output pairs that represent the behavior you want — how you want the model to respond, reason, decline, help. You fine-tune on those pairs using standard backpropagation. The model learns to imitate the demonstrated behavior.

But imitation has limits. The model learns to predict what comes next in training data, not to actually satisfy a human. A technically correct answer and a genuinely helpful answer are not the same thing. Word choice, tone, reasoning structure — none of these are captured by next-token prediction loss.

This is where Reinforcement Learning from Human Feedback (RLHF) enters.

Instead of asking humans to score outputs numerically (too noisy, too inconsistent), RLHF uses pairwise comparison. Show a human two outputs. Ask which is better. Train a reward model on those preferences using the Bradley-Terry framework. Then use that reward model to guide further training — reinforcing outputs the reward model scores highly, suppressing outputs it scores poorly.

The result: a model that doesn’t just predict what comes next, but generates what humans actually prefer.

Why This Is Expensive

The lecture was direct about this. Using humans in the loop is expensive and slow. That’s not a minor implementation detail. It’s why the most capable AI systems in the world require hundreds of millions of dollars to train. Every preference signal costs human time. Every rubric requires domain expertise. Every quality judgment call has to be made by a real person.

I’ve been one of those people.

Writing grading rubrics for ML research replication tasks. Evaluating whether an AI’s code correctly reproduced a paper’s results. Judging whether a response demonstrated genuine understanding or statistical mimicry. It’s detailed, effortful work. And now, studying the pipeline that consumes it, I understand exactly where my work went.

It went into reward signal. It shaped the advantage function. It moved the policy.

There’s something unexpectedly grounding about that.

Keeping the Model Honest: PPO and GRPO

Once you have a reward model, the next problem is optimization stability.

If you simply maximize reward, the model drifts. It learns to game the reward model — producing outputs that score highly but diverge wildly from coherent, grounded behavior. This is called reward hacking, and it’s a real failure mode.

Proximal Policy Optimization (PPO) addresses this with a clipping mechanism. The update at each step is bounded — you can’t move too far from the previous policy in a single step. It adds a KL divergence penalty that keeps the fine-tuned model close to the pretrained base. The base model acts as an anchor. You’re optimizing for human preference, but you can’t forget everything you learned during pretraining to do it.

Group Relative Policy Optimization (GRPO) takes a different angle. Rather than maintaining a separate value model to estimate advantage, it generates a group of outputs for the same prompt, normalizes rewards within that group, and uses the relative scores to compute advantage. Simpler. Lower memory. Well-suited to settings where rewards are binary or correctness-based — exactly the kind of rubric-graded research tasks I’ve been working on.

The elegant part: both methods are still policy gradient at heart. The math is the same. The practical tradeoffs differ.

The AI Era Tech Stack

Somewhere in the middle of all this, I started thinking about the analogy someone put to me recently.

Hugging Face is the GitHub of the AI era. A place to host, share, discover, and version models and datasets. The commons layer.

MCP is the REST API of the AI era. A universal protocol for how AI agents communicate with tools, services, and each other. The interoperability layer.

LangChain is the Next.js of the AI era. An opinionated framework that abstracts the plumbing — prompts, chains, memory, output parsers, tool routing — so you can build applications without reinventing infrastructure each time.

Push this further and the whole stack maps cleanly.

The LLM itself is the browser engine. The runtime that executes intelligence. Anthropic and OpenAI are the operating systems — the foundational platform that everything else depends on. Groq and vLLM are the CDN and edge compute — inference delivery, fast. Vector databases like FAISS and Chroma are the semantic equivalent of Postgres — structured knowledge stores, queried not by exact match but by meaning. RAG is the ORM layer — the pattern that connects the model to persistent external memory at runtime.

What makes this analogy useful isn’t elegance. It’s that it tells you where we are in the development cycle.

The web had twenty years of infrastructure building before applications became commoditized. AI is compressing that timeline, but the same layers still need to exist. Right now, most of the action is at the infrastructure layer. The frameworks are still being written. The protocols are still being standardized. The models are still improving at a pace that makes last year’s benchmark obsolete.

That won’t last forever. It never does.

What I’m Still Working Through

The KL penalty is clear to me as a concept. What I want is a better intuition for what reward hacking actually looks like in practice — what an over-optimized model produces before the penalty reins it in.

GRPO’s group normalization vs. PPO’s value model in low-resource settings. Which wins, and under what conditions.

And the deeper question underneath all of this: how do you write a rubric that captures what you actually care about, rather than a proxy for it? That’s not a theoretical question. That’s the hard part of the job.

I’m still learning. Most weeks, that means the lecture opens more questions than it closes.

This week, it also closed a loop I didn’t know was open.

Notes from the Stack is a series about building with AI, learning in public, and thinking through what this technology actually means. Published when something is worth saying.

Subscribe now

Share

Leave a comment

American business history has a pattern.

Big markets with high capital requirements and complex technology tend to converge. Not on one winner. Not on ten. Usually on three.

Aviation did it. Automobiles did it. Cloud computing is still doing it, but you can already see the shape. Credit ratings have three firms that matter. Cola has two and a half. Search had one and now AI is scrambling to become the second.

The Big Three framework isn’t just a heuristic. It reflects something structural. When the cost of competing is extremely high, and the benefit of scale is extremely high, the market doesn’t distribute evenly. It narrows. And within the three, they aren’t equal either. Number one runs away. Number two struggles to close the gap. Number three is still in the first tier but is already a little breathless.

As of today, mid-2026, the AI landscape is beginning to look like this.

The First Pole: Anthropic — The Enterprise Leader

Anthropic doesn’t make the flashiest announcements.

It doesn’t court consumer virality. It doesn’t release models with cinematic trailers. What it does is build things that enterprise customers trust. And in 2026, that bet is paying off in a way that’s hard to argue with.

Eight of the Fortune 10 are Claude customers. Over 300,000 businesses use Claude. More than 500 of them spend over a million dollars a year. Annualized revenue reportedly crossed $30 billion by April 2026 — and that trajectory happened almost entirely on the enterprise side.

The reason isn’t mystery. Enterprise customers don’t need a model that can write poetry. They need a model that won’t hallucinate in a legal brief, won’t leak data, won’t fail unpredictably at 2am when a pipeline is running. They need stable. Reliable. Auditable.

That’s what Anthropic built toward. Constitutional AI, interpretability research, safety focus — these aren’t just values. They turned into a product edge in the one market that was willing to pay.

I think about this often. I’m building FormAgent for elderly patients and disabled adults — people who are already vulnerable, navigating systems that were not designed for them. The question I ask myself about every AI component isn’t “how good is it on benchmarks?” It’s “what happens when it gets something wrong?” That’s exactly the question enterprises are asking. And Anthropic built the answer before anyone else.

The Second Pole: OpenAI — The Consumer King

I’ll be direct: ChatGPT is the most successful AI consumer product in history.

When ordinary people say “AI,” they often mean ChatGPT. That brand lock-in is extraordinary. It doesn’t matter that other models score higher on certain benchmarks. It doesn’t matter that a competitor releases something technically equivalent. Habits are stronger than specs.

Think about how long it took anyone to challenge Google in search. Or Facebook in social. The moat isn’t the technology. The moat is the daily ritual. The muscle memory of opening a tab and typing.

OpenAI has that now. Across writing, coding, education, research, personal assistance. And they’re actively expanding the surface area — the new interface isn’t just a chat box. It’s trying to become an operating system for knowledge work.

The question isn’t whether OpenAI’s model is the best. The question is whether anyone can change the habit of hundreds of millions of people who already open ChatGPT every morning. That’s a higher barrier than any benchmark can measure.

The Third Pole: Google — The Old Horse’s Last Race

Google’s position in AI is genuinely complicated to evaluate.

On raw model quality, Gemini is competitive but not obviously ahead. On distribution, it has something no AI startup can replicate: search, Android, Chrome, YouTube, Workspace — billions of daily active users across every surface that matters.

The technology isn’t Google’s problem. It never was.

Google invented the Transformer architecture that made all of this possible. Pioneered attention mechanisms. Published the research that other companies turned into products. The pattern is almost tragic: Google generates the ideas, then watches others commercialize them faster.

The question for Gemini is whether this time is different. Whether the organizational will exists to move fast enough, integrate deeply enough, and treat AI as an existential product rather than a research showcase.

I genuinely don’t know the answer. But I’d bet on the distribution before I’d bet on the culture changing.

The Variable: SpaceX

Here’s where things get interesting.

Everyone is watching xAI and Grok. That’s the obvious Musk play. But I think the more consequential development isn’t the model — it’s what SpaceX makes possible.

Starlink is already the largest low-orbit satellite network on earth. SpaceX controls more launch capacity than any organization in history. When you start thinking about AI not as a software product but as a physical infrastructure problem — about where compute lives, how it connects, how it scales to places that don’t have fiber — SpaceX becomes something different from any existing AI competitor.

This isn’t a near-term story. But it’s the kind of structural bet that, if it plays out, doesn’t just change the AI market. It changes what “AI” means in the same way that the internet changed what “software” meant.

A card that hasn’t been turned over yet.

Microsoft and Meta: Powerful, But Not Defining

This isn’t a slight. Both companies will make significant money from AI. Both have extraordinary distribution.

But there’s a difference between being a major participant and defining the category.

Microsoft is one of history’s great integrators. DOS, Windows, Office, Azure — each time, Microsoft took someone else’s proved idea and scaled it into a dominant platform. That’s a real and valuable skill. It’s just not what “defining a generation of technology” looks like.

Meta executes brilliantly, sometimes recklessly, usually in reaction. The metaverse pivot was reactive to Apple’s AR roadmap. The AI push is reactive to OpenAI’s success. Zuckerberg is willing to spend — $65 billion in AI infrastructure this year — but spending isn’t the same as vision.

They’ll win revenue. They won’t write the next chapter.

What This Means

I started thinking about all of this from the outside — as a founder, watching the landscape settle.

QinAI sits in the Anthropic world. We use Claude. We’re building in the enterprise and institutional space, for users that institutions have historically failed. The bet on Anthropic isn’t just a technical decision. It’s an alignment of values. Safety-first AI for vulnerable populations isn’t a tagline — it’s the product requirement.

But I also feel something beyond the strategic analysis when I look at this moment.

Every major technology wave produces its three. And the thing about being in the middle of it — actually building on the infrastructure, watching the poles form — is that the structure isn’t obvious until suddenly it is.

In 2010, most people couldn’t name the three cloud providers that would define the next decade. By 2015, it was AWS, Azure, and GCP. The window to position was the middle — when it was already too late to be a cloud company, but still early enough to build on top of the one you believed in.

We’re in that window now for AI.

The next ten years will reveal not just who wins, but who stays at the table. The biggest question isn’t which company will be number one. It’s which companies will still be playing when the game is fully defined.

And somewhere above us, a satellite network is quietly getting larger.

Lily is the founder of QinAI Inc., building FormAgent — a voice-first AI agent for healthcare form completion for elderly and underserved populations. She writes Notes from the Stack (AI × building) and Between Worlds (memoir).

Subscribe now

Share

Leave a comment

美国商业史有一个反复出现的规律。

凡是市场足够大、资本门槛足够高、技术足够复杂的行业，最终都会收敛。不是一家独大，不是百花齐放，而是——三家。

航空业如此。汽车业如此。云计算正在走向同样的结构，骨架已经清晰。信用评级有三家真正算数的公司。可乐有两家半。搜索引擎有一家，现在 AI 正在争夺第二个位置。

“三大格局”不只是一个经验规律，它背后有结构性逻辑：当竞争成本极高、规模优势极显著时，市场不会均匀分配，而会向少数平台聚拢。而且这三家永远不对称。老大一骑绝尘。老二竭力追赶。老三仍在第一梯队，但已经有些气喘。

截至2026年中，AI 的格局正在呈现这种形状。

第一极：Anthropic — 企业 AI 的领跑者

Anthropic 从来不做最炫目的发布。

它不追求消费端的病毒式传播，不用电影级预告片发布模型。它做的事很朴素：让企业客户信任它。而在2026年，这个选择的回报，已经很难辩驳。

财富10强中有8家是 Claude 的客户。超过30万家企业在使用 Claude。其中500多家每年在 Claude 上花费超过百万美元。2026年4月，年化收入据报道已突破300亿美元——而这条增长曲线，几乎完全来自企业市场。

原因并不复杂。

企业客户不需要一个会写诗的模型。他们需要的是：不会在法律文件里一本正经地胡说，不会泄露数据，不会在凌晨两点的流水线上无故崩溃。他们需要稳定、可靠、可追溯。

这正是 Anthropic 一直在建造的东西。Constitutional AI、可解释性研究、安全优先——这些不只是价值观，它们最终转化成了产品壁垒，在最愿意付费的那个市场里。

我常常想到这一点。我在为老年患者和残障人士构建 FormAgent，这些人本来就脆弱，在一个从未为他们设计的系统里艰难挣扎。我对每一个 AI 组件问的问题从来不是”它在 benchmark 上跑多少分”，而是”它出错的时候会发生什么”。这正是企业客户在问的问题。Anthropic 比任何人都早，给出了答案。

第二极：OpenAI — 消费端的王者

直说吧：ChatGPT 是历史上最成功的 AI 消费产品。

对普通用户来说，”AI” 几乎已经等于 ChatGPT。这种品牌认知，极难撼动。不管竞争对手在某个 benchmark 上跑出多高的分，不管谁发布了技术上等效的产品——习惯比规格更强大。

想想 Google 搜索花了多久才遇到真正的挑战。想想 Facebook 花了多久。护城河不是技术，是每天打开那个页面的肌肉记忆。

OpenAI 现在拥有这个。在写作、编程、教育、研究、个人助理……每一个场景。而且他们在持续扩展边界——新界面不只是一个聊天框，它正在试图成为知识工作的操作系统。

真正的问题不是 OpenAI 的模型是不是最好的。真正的问题是：有没有人能改变数亿用户每天早上打开 ChatGPT 的习惯？那道门槛，比任何 benchmark 都高。

第三极：Google — 老马的最后机会

Google 在 AI 领域的处境，真的很复杂。

从模型质量来看，Gemini 有竞争力，但并没有明显领先。从分发能力来看，它拥有任何 AI 创业公司都无法复制的东西：搜索、Android、Chrome、YouTube、Workspace——每天覆盖数十亿用户，跨越每一个重要入口。

技术从来不是 Google 的问题，从来都不是。

Transformer 架构是 Google 发明的。Attention 机制是 Google 先做出来的。推动整个行业变革的那篇 “Attention is All You Need”，是 Google 的研究员写的。这个模式几乎带着一种宿命感：Google 生产思想，别人把思想变成生意。

Gemini 能不能真正站稳 AI 三强？问题不在技术，在组织意志。这一次，大公司病能不能被克服，AI 能不能被当作存亡级别的产品而不是研究展品来对待？

我不知道答案。但如果让我下注，我会押分发能力，而不是押文化变革。

最大变量：SpaceX

这里有一张还没翻开的牌。

所有人都在看 xAI 和 Grok，这是马斯克最显眼的棋。但我认为，更有决定性的变量，不是那个模型，而是 SpaceX 能让什么成为可能。

Starlink 已经是地球上最大的低轨卫星网络。SpaceX 控制着人类历史上从未有过的发射能力。当你开始把 AI 不只看成一个软件产品，而是一个物理基础设施问题——算力住在哪里、怎么连接、怎么延伸到没有光纤的地方——SpaceX 就变成了和所有现有竞争者截然不同的物种。

这不是近期的故事。但如果它真的发生，不只是改变 AI 市场的格局，而是改变”AI 是什么”这个问题的答案——就像互联网改变了”软件是什么”的意涵一样。

微软和 Meta：手握筹码，但不像定义者

这不是贬低，两家公司都会从 AI 时代赚到大钱。

但参与赛局和定义赛局，是两回事。

微软是历史上最伟大的整合者之一。DOS、Windows、Office、Azure——每一次，微软都把别人验证过的方向规模化成主流平台。这是真实的能力。只是这种能力和”定义一个技术时代”不一样。

Meta 执行力极强，有时近乎鲁莽，通常都在追赶而非领跑。元宇宙是对苹果 AR 路线图的跟随。AI 押注是对 OpenAI 成功的反应。扎克伯格愿意砸钱——今年 AI 基础设施投入据说高达650亿美元——但钱不等于愿景。

他们会赢得收入。但下一个章节，不会由他们来写。

在窗口中间

我从外部视角想这些事，但同时我也是局中人。

QinAI 活在 Anthropic 的世界里。我们用 Claude，在企业和机构市场构建产品，服务那些被现有系统长期忽视的用户。押注 Anthropic，不只是技术选型。是价值观的对齐：为脆弱人群打造安全优先的 AI，不是口号，是产品规格。

但我在分析之外，也感受到别的东西。

每一次重大的技术浪潮，都会孕育出它的”三大”。而身处其中、真正在基础设施上建造东西、眼看极点成形——这个结构，不是在事后才清晰的。它在某一刻，突然就清晰了。

2010年，大多数人说不出哪三家云厂商会定义下一个十年。2015年，答案变成了 AWS、Azure 和 GCP。最好的窗口是中间那段——已经太晚去成为一家云公司，但还足够早去选定一个你相信的平台，在上面建造。

AI 的这个窗口，现在就是。

未来十年，揭晓的不只是谁赢，而是谁还留在牌桌上。

而在我们头顶的天空里，一张卫星网络正在安静地，继续变大。

Lily 是 QinAI Inc. 的创始人，正在构建 FormAgent——一个为老年和弱势群体提供医疗表格自动填写的语音优先 AI 产品。她在 Substack 上写两个系列：Notes from the Stack（AI × 产品）和 Between Worlds（人生回忆录）。

Leave a comment

Share

Subscribe now

几周前，我写了一篇关于平行宇宙的文章。

写那个母亲还活着的我。留在四班没有转走的我。去北大读文学而不是去中科大学数学的我。

那篇文章的最后，我写下了一句自己也没有预料到的话：

人最深的渴望，不是活得不同。而是活得更多。

这句话我后来一直在想。

因为我后来发现，它也是另一个问题的答案。一个我以为和平行宇宙毫无关系的问题。

过去一年，我在做一个 AI agent。

不是聊天机器人。不是智能助手。是一个代理人。

FormAgent 站在人与官僚体系之间。老年患者、新移民、残障人士，和那堵由表格砌成的墙之间。Medicare 申请，Medicaid 文件，I-9 表格。那些默认你识字、健康、有耐心的文件。

我的大多数用户，一样都不占。

做到后来，我越来越觉得：这不是一个表格问题。是一个主权问题。

人们被一个从来不是为他们设计的系统处理着。一点一点，失去自己生活的线头。

FormAgent 真正想做的，是把那根线还给他们。

最近读了一篇硅谷投资人写的长文，讲大模型的投资逻辑。他的核心判断是：大模型的死穴，就是 AI 创投的生门。

写得不错。但我一边读一边觉得，分析框架有点偏。

他问的是钱去哪里。

我想问的是人去哪里。

他有几个判断，我是认同的。

大语言模型最擅长的，是文明的符号层。文字、代码、文件、合同、结构化知识。白领工作本质上就是这些东西的组合。所以 AI 先学会的，是白领做的事。这没有什么好奇怪的。

更深的后果，不是某些岗位消失，而是整个制度层在重组。很多繁琐的流程，存在是因为人不可靠。我们会犯错，有私心，会忘，会走捷径。所以建了层层审批来防人。如果 AI agent 接管了执行端，日志可查，权限清晰，很多原来防人的环节就没有存在的必要了。

不是岗位在消失。是岗位依附的那套制度器官，在被重新长出来。

但物理世界不会这么轻易被重组。

我丈夫 Ke 做了很多年老年科。他说，看诊时读到的东西，一半不在病历里，在诊室里。回答问题前那一秒的停顿。上个月还没有的步态变化。角落里坐着一直没说话的女儿。

这些你没办法 tokenize。用多少数据训练也训练不出来。

物理世界是多感官的，容错率低，充满不打招呼就来的意外。AI 会进入这个领域，但会慢，会垂直，会带着纯软件从没遇到过的摩擦。这种摩擦不是障碍。正是它让解决方案经得住时间。环境越难，越难被取代。

文章里还有一个判断，让我停下来想了很久。

互联网给了我们免费内容。代价是我们的神经系统。

平台算法优化的不是我们的幸福，是停留时长。而停留时长意味着焦虑、愤怒、那种大脑还没想清楚拇指就已经动了的刷屏。今天大多数 AI 工具，只是让这一切产生得更快。更多内容，更多噪音，更快的洪水。

他说，下一个真正的转变是个人智能体。

不是更聪明的助手。是不同的忠诚结构。

一个你自己付费养的 agent。为你工作，不为平台工作。在算法触达你的注意力之前，先替你过滤世界。

我觉得 FormAgent 是这个方向上一个很早、很窄的版本。一个不会说官僚语言的人，雇了个会说的代理人。代理人替她和机构周旋。她不需要去弄懂一个从来没打算让她弄懂的系统。

把这个逻辑放大，放到媒体、社交、信息获取上，互联网的入口结构就会动摇。

文章快结尾时，他提到了一个我觉得最安静也最深的判断。

体力劳动被机械化之后，健身产业出现了。以前要下地的身体，现在去健身房。需求没有消失，只是换了一种形式。

当脑力劳动被大规模自动化，同样的事会发生。

迁移的，是心力。

专注力。情绪韧性。真正在场的能力。在普通的一天里找到意义的能力。

这些是可以学的，可以练的。随着 AI 吸走越来越多认知层的工作，人对这些东西的缺失会更敏感。因为失去劳动的人，失去的不只是收入。是结构。是目的。是每天都能感觉到自己有用、有意义的那条线。

人不能没有劳动。没有劳动，人会失去意义。失去意义，心理健康就会出问题。

这不是玄学，是人类学。

身心灵产业，会是 AI 时代真正的大产业。不是小众wellness，是文明补偿机制。就像健身房之于工业化，只是这次练的不是肌肉。

写到这里，我才突然看清楚，这篇 AI 文章和上篇平行宇宙，其实在问同一件事。

平行宇宙那篇，我问的是：活了另一种人生的那个我，还会想要我现在想要的东西吗？

这是一个关于约束改变之后，自我如何存续的问题。外部结构换了，什么留下来了。

AI 这篇，问的完全一样。机器接管了越来越多的工作，越来越多的符号层，越来越多的协调与分析，剩下的是什么？什么是本来就没办法被 tokenize 的，不是因为我们规定它不能，而是因为它根本就不是那种东西？

我想两个问题的答案是一样的。

是内在生命。真正在一个房间里感受到正在发生什么的能力。因为说不清楚的理由去想要某件事的能力。意义的体验，那不是一个数据结构。

宇宙三那个去北大学文学的我，对这些也许有更多词汇。但我不确定她感受得更深。

真正理解这件事，也许需要在技术和人文的缝隙里住得足够久，久到能感受到那道距离本身。

又是夹缝之间。

那位投资人用三句话收尾。

机器换人，效率革命。 人换机器，主权革命。 人回到自己，文明革命。

AI 行业大多数人还在第一阶段。这说得通，那里有最近的钱。

我在试着在第二阶段做东西。agent 的第一忠诚，在结构上站在它服务的那个人那边。

第三阶段是什么样子，我们还不知道。

但我觉得，它从同一个问题开始。就是上个月在那个沔中校友群里，我问过自己的那个问题，只是换了一种说法：

不是哪个版本的未来更好。

而是：住在那里的人，还知道想要什么吗？还知道怎么在场吗？还知道有意义是什么感觉吗？

我没有干净的答案。

我有一家公司，指着我希望是对的方向。还有一个信念，最好的 AI 不是取代人的那个，是把人的生活还给人的那个。

最深的渴望，不是少干点活。而是活得更多。

还是这句话。只是我花了几周才看出来，它也是这个问题的答案。

《Notes from the Stack》是我的技术与创业系列。这篇文章在呼应的那篇，
”如果宇宙允许我活两遍”，在《Between Worlds》回忆录系列里。

Lily，QinAI Inc. 创始人

Leave a comment

Share

Subscribe now

A few weeks ago I wrote about parallel universes.

About the version of me who grew up with her mother alive. The version who stayed in Class Four. The version who studied literature at Peking University instead of mathematics at USTC.

I ended that essay with something I did not expect to write when I started it:

The deepest human longing is not to have lived differently. It is to have lived more.

I have been sitting with that line ever since. Because I think it is also the answer to a question I get asked constantly now, as someone building AI.

The question people are really asking, underneath all the anxiety about jobs and automation and what comes next, is not: what does AI take away?

It is: what will be left of us?

I have been building an AI agent for the past year.

Not a chatbot. Not a copilot. An advocate.

FormAgent sits between people and the paperwork maze that stands between them and healthcare. Elderly patients. Recent immigrants. Disabled adults. Medicare forms. Medicaid applications. I-9 documents. The kind of paperwork that assumes you are fluent, healthy, and patient.

Most of my users are none of those things.

And the more I build, the more I think: this is not really a paperwork problem. It is an agency problem. People are being processed by systems that were never designed for them. They are losing the thread of their own lives to administrative friction. What FormAgent is actually trying to do is give that thread back.

That realization is what made me read a recent essay by a Silicon Valley investor very differently than he probably intended it.

He was writing for other investors. About where the opportunities are as large language models eat through white-collar work. His frame: the death spot of AI is the life gate for investment.

I kept reading it and thinking: yes, but that is the wrong unit of analysis.

The question is not where the money goes. The question is where the humans go.

Here is what I think the essay gets right.

Large language models are extraordinary at the symbolic layer of civilization. Text, code, documents, contracts, reports, structured knowledge. That is where they were trained. That is what they absorb first.

White-collar work, all that coordination machinery humans built to manage complexity, is essentially compressed language. So of course AI learned it first. Decades of white-collar output, fed back into a model, and the model becomes fluent in exactly what knowledge workers do.

The deeper consequence is not job replacement. It is institutional restructuring. A lot of that administrative overhead existed because humans are unreliable. We make mistakes, have biases, forget things, cut corners. So we built elaborate systems to check each other. If AI agents handle execution with traceable logs and clear boundaries, much of that checking collapses.

The organ is being restructured. Not just the job inside it.

But the physical world does not restructure so easily.

My husband Ke spent years in geriatric care. When he describes a patient visit, half of what he is reading is not in the chart. It is in the room. The slight hesitation before an answer. The shuffle in the gait that was not there last month. The daughter in the corner who has not said anything yet.

You cannot tokenize that. You cannot train your way past it with more data.

The physical world is multisensory, low-tolerance for error, and full of the kind of surprise that does not announce itself in advance. AI will grow into physical environments. But it will grow slowly, vertically, with real friction. And that friction is not a weakness. It is what makes a solution durable. The harder the environment, the harder it is to be displaced.

The investor’s essay also names something I think about constantly: attention.

The internet gave us free content.

The price was our nervous systems.

Platform algorithms were never optimized for our wellbeing. They were optimized for engagement, which often means anxiety, outrage, and the kind of compulsive checking that happens before the brain decides anything. Most AI tools today are just making this faster. More content. More noise.

The next meaningful shift is what he calls the Personal Agent.

Not a smarter assistant. A different loyalty structure.

An agent you pay for. That works for you. That filters the world on your behalf before the world’s algorithms reach your attention.

I think about FormAgent as a narrow, early version of this idea. A person who does not speak the language of bureaucracy hires an agent that does. The agent negotiates with the institution. The person does not have to become fluent in a system that was never built for them.

Scale that logic. Apply it to media, to social connection, to how we move through information every day. The platforms that currently own the entry point to the internet should be paying close attention. Because the entry point is about to move.

There is a section of the investor’s essay I keep returning to.

He argues that when physical labor was mechanized, the fitness industry emerged. Bodies that no longer needed to plow fields started going to gyms. The physical demand did not disappear. It migrated.

When cognitive labor gets automated at scale, something similar will happen.

What migrates is inner labor. 心力.

Attention. Emotional resilience. The capacity to be fully present. The ability to find meaning in an ordinary day.

These are learnable, trainable capacities. And as AI absorbs more of the cognitive layer, people will feel their absence more sharply. Because a person without labor loses more than income. They lose structure. They lose purpose. They lose the daily evidence that they matter.

Mental health is not a side effect of AI advancement. It is a central consequence. The body-mind-spirit industry, 身心灵, will not be niche wellness. It will be something closer to infrastructure.

This is where the parallel universe essay and the AI essay, which seem to be about entirely different things, actually converge for me.

In “Parallel Universe” I asked: would the version of me who lived differently even want what I want now?

That is a question about identity under changing constraints. What persists when the external structure changes. What turns out to be the actual self, underneath the circumstances.

The AI question is structurally identical.

When machines handle more of the work, more of the symbolic layer, more of the coordination and the documentation and the analysis, what remains? What turns out to be irreducibly human, not because we legislated it that way, but because it cannot be abstracted into tokens?

I think the answer is the same in both cases: the interior life. The capacity to be in a room and actually feel what is happening in it. The ability to want something for reasons that cannot be fully articulated. The experience of meaning, which is not a data structure.

The woman in Universe Three, the one who studied literature at Peking University instead of mathematics, she might have more vocabulary for this. But I am not sure she would understand it more deeply. Understanding it deeply might require having lived in the gap between the technical and the humanistic long enough to feel the distance.

Between worlds, again.

The investor ends his essay with a three-part frame.

Machine replaces human. Efficiency revolution.

Human reclaims from machine. Sovereignty revolution.

Human returns to self. Civilizational revolution.

Most of the AI industry is in the first phase. That makes sense.

I am trying to build in the second phase. Tools where the agent’s loyalty is structurally on the side of the person it serves.

The third phase, humans rebuilding interiority in an age of ambient AI, we do not fully know what that looks like yet.

But I think it starts with the same question I asked in that WeChat group conversation about parallel universes, restated for the present moment:

Not which version of the future is better.

But: would the person who lives there still know how to want things? Still know how to be present? Still know what it feels like to matter?

I do not have a clean answer.

I have a company pointing in what I hope is the right direction. And a conviction that the best AI is not the one that replaces the human. It is the one that gives the human their life back.

The deepest longing is not to have worked less. It is to have lived more.

That is still the answer. It just took me a few days to see that it was also the answer to this question.

Subscribe now

Share

Leave a comment

Related:
Between Worlds -- Special Essay: Parallel Universe

The conversation started, as many of the best ones do, in a WeChat group.

Former classmates from 沔中 — Mianyang High School, Hubei — scattered now across continents, time zones, careers none of us could have predicted in the 1990s. Someone mentioned parallel universes. The thread exploded. People started asking: if you could have lived another version of your life, which one would you choose?

I sat with that question for a long time.

The Science, Briefly

In quantum mechanics, there is a serious — if contested — idea called the Many Worlds Interpretation. Every time a particle can go left or right, it goes both. The universe branches. Somewhere, the version of you who took that other job exists. Somewhere, you said yes when you said no. Somewhere, you stayed.

This is not fringe physics. Hugh Everett proposed it in 1957. Many serious physicists take it seriously. The math works.

What does not yet exist is any way to communicate between branches. No bridge. No door. No interface.

Yet.

I believe — I genuinely believe — that AI + VR + some unknown technology we do not have a name for will eventually make something like this possible. Not in my lifetime, maybe. Maybe in my children’s. Maybe in a form so strange we would not recognize it as “parallel universe access” at all. But the desire is real, the physics is permissive, and human beings have a way of building toward what they cannot stop imagining.

Universe One: The Family That Stayed Whole

My biological mother died when I was ten months old.

A contaminated vaccine. A rural clinic in Hubei. An ordinary summer afternoon that became the hinge point of everything.

My brother had a heart attack at eighteen. He never got to grow up properly, never got to be taken care of the way a boy with a whole mother might have been.

In Universe One: none of that happens.

My mother lives. My brother is healthy, watched over, loud at the dinner table. My father is lighter — a different kind of man entirely, one not shaped by grief. My sister and I grow up inside something warm and loud and unremarkable in the best possible way. A whole family. The noise of people who are not afraid of each other.

Maybe I still go to university. Maybe I marry young in the village and raise children with the paddy fields in the background and my parents next door. Maybe both, in some sequence I cannot predict. I specified in the parameters: my husband is still my husband. My children are still my children.

But here is the part of the thought experiment that unsettles me most: would I trade it?

The girl who grows up whole, inside warmth, with a mother — she makes different choices at every fork. She is braver in some ways and less forged in others. Does she still leave China? Does she still go to USTC? Does she still end up, somehow, at the exact latitude and longitude where she meets the man she spends her life with?

Grief is terrible. Grief is also a kind of fuel. I am not sure I want to have been built from something else.

Universe Two: Class Four, Not Class Five

In my second year at Mianyang High, I transferred from Class Four to Class Five.

That one administrative shuffle changed every friendship I made in high school. Different desk neighbors. Different lunch conversations. Different people leaning over to share answers before exams. The network you build at sixteen ripples outward for decades — who you study with, who you compete with, who you later introduce to your roommate in college, who sends you the job listing that changes everything.

In Universe Two, I stayed in Four.

I think the broad trajectory stays roughly the same. A decent university. Graduate school abroad. The AI wave arriving on schedule. More or less the same shape of a life oriented toward ideas and ambition.

But the textures are completely different. The jokes I have memorized. The faces I still see in dreams. The specific quality of loneliness on a particular Tuesday in 1993.

This universe is the least dramatic. And maybe that is the point. The biggest pivots are often the quietest ones — a classroom assignment, a seat change, a name drawn from a hat.

Universe Three: Peking University, Literature

I did not go to USTC for mathematics.

I went to Peking University, where I studied what I actually love: literature, history, the humanities. I wrote essays that mattered to me. I read everything. I became the person whose bookshelf is organized by feeling rather than subject.

No six years of rigorous scientific training. No theorem-proving at 2 a.m. No moment when abstract algebra breaks you down and then, somehow, rebuilds you into something more precise.

I write more fluently in this universe. I have a larger vocabulary for the inner life. I can name emotions the way a scientist names compounds.

But I am missing something the mathematics put in me. A tolerance for sustained confusion. A belief that if you hold a problem long enough, it will eventually yield. A comfort with the abstract that I did not know I had until I started working on AI.

Of all the universes, this one might be the most tempting. And the one I would most regret.

If mathematics is my skeleton and literature my skin, then the life I am living now is the creature that lives between them.

Universe Four: This One

I went to USTC. Mathematics. Six years of the kind of training that does not care whether you are comfortable.

I met my husband there. He was the kind of person who became more himself the longer you knew him — which is a rarer quality than it sounds. We left China together. We built a life in a country that was not ours, which means we had to build everything from scratch, which means we know what we built.

Three children. Three people who are already surprising me in ways I did not predict and cannot contain.

A startup. The AI wave, which I saw coming from a long way off — because mathematics teaches you to recognize the shape of exponential growth before it becomes visible to everyone else.

And then this: the gap between what I studied and what I love. The person who can follow a proof and also get completely lost in a Tang dynasty poem at 11 p.m. The founder who wants her company to matter and wants to write. The immigrant who loves her chosen country and mourns the one she left. The woman between worlds.

Between worlds is uncomfortable. Between worlds is also the only place from which you can see both sides.

The Real Question

My friends in the WeChat group wanted to know which universe I would choose.

I have been thinking about this wrong.

The question is not which life would I prefer in the abstract. The question is: would the person I would have become in that life still want the things I want now?

The woman from Universe One — grew up whole, mother alive — she might be warmer. More at ease in her body. Less driven by a specific quality of need. I am not sure she builds a startup. I am not sure she has to.

The woman from Universe Three — Peking University literature degree — she writes better than I do. But she might not have the strange patience for ambiguity that six years of mathematics instilled in me. She might not be comfortable with the parts of building a company that are just: hold the unknowing, keep moving.

The more I sit with this, the more I think: the you who lived the other life would not want to trade back either. She has become what she became. She does not know what she is missing because she never had it. She is, probably, fine.

This is the thing about parallel universes that the physics textbooks do not mention: every version of you is the protagonist of her own story. Every version has built her attachments, her loves, her habits, her sense of what matters. Every version, when asked if she would trade, would probably hesitate.

What I Actually Want the Technology to Do

I said I believe AI + VR + unknown technology will make this possible. Let me be more specific about what I mean.

I do not want to live the other life. I want to visit it. The way you visit a country you were not born in and spend two weeks there and come home changed by it but still home.

I want to have dinner with my mother. Even once. Even in simulation. Even knowing the whole time that it is a simulation. I want to sit across from her and watch her move her hands and understand where certain things about me come from.

I want my brother to walk into a room healthy and young and himself.

I want to spend one semester at Peking University studying the Song dynasty and be terrible at it and love every minute.

These are not replacements for the life I have. They are extensions. An expansion pack for consciousness.

The deepest human longing is not to have lived differently. It is to have lived more.

The grief of life is not that we took the wrong road. It is that there were not enough roads, and not enough time.

A Final Note

I am writing this from Universe Four. The only one I have actually lived.

It is, improbably, a good one. The mathematics and the longing for literature. The immigrant life and the startup. The three children who are already going to surpass me in ways I cannot predict. The man I met in a dormitory hallway in Hefei who is still, somehow, surprising.

I would not trade it. I say that knowing full well that the woman in Universe One would say the same thing about hers, and mean it just as completely.

That might be the whole point.

The universe gave each of us one life. We made it ours. We could not have done otherwise. The other selves are real — maybe, in some branch of the wavefunction, actually real — but this self, this particular accumulation of losses and choices and children and theorems and half-read books and late-night essays, is the one I know how to be.

For now, that is enough.

For now.

Subscribe now

Share

Leave a comment

Chinese version of this post:
如果宇宙允许我活两遍

这个话题，起于校友群的一次闲聊。

沔中的老同学们，如今散落在各个时区，各种行业，做着1990年代的我们谁都没有预料到的事。有人提起平行宇宙，群里立刻炸开了锅。大家开始问：如果可以活另一种人生，你会选哪一种？

我把这个问题放了很久。

物理学，简说

量子力学里有一个严肃的——尽管仍有争议的——理论，叫多世界诠释（Many Worlds Interpretation）。每当一个粒子面临两种可能，它同时走两条路。宇宙从此分叉。在某处，那个选了另一份工作的你，真实存在。在某处，你说了”是”，而不是”不”。在某处，你留下了。

这不是边缘物理学。Hugh Everett在1957年提出这个理论。许多严肃的物理学家认真对待它。数学是成立的。

目前不存在的，是任何在不同分支之间通信的方式。没有桥。没有门。没有界面。

还没有。

我相信——是真心相信——AI、VR，加上某种我们还没有命名的技术，终将让某种形式的平行宇宙体验成为可能。也许不在我这一代。也许在孩子们那代。也许以一种我们完全认不出来的形式。但这个渴望是真实的，物理学也并不禁止，而人类总有一种方式，去建造他们停不下来去想象的东西。

第一宇宙：那个完整的家

我的亲生母亲在我十个月大的时候去世了。

一支被污染的疫苗。湖北一个乡村诊所。一个寻常的夏日午后，成了一切的转折点。

我哥哥十八岁那年心肌梗塞。他没能好好长大，没有得到一个完整的家庭本可以给他的照料。

在第一宇宙：这些都没有发生。

妈妈还在。哥哥健康，被好好照顾，在饭桌上说话最大声。父亲是个更轻盈的人——一个没有被悲伤塑造过的人。姐姐和我在一个温暖嘈杂的家里长大，平凡得恰到好处。完整的家。不害怕彼此的人发出的声音。

也许我还是上了大学。也许我在村里早早嫁人，孩子在稻田边长大，父母住在隔壁。也许两者都有，以某种我无法预测的顺序。我假设老公还是他，孩子们还是他们。我在参数里这样设定了。

但这个思想实验里最让我不安的部分来了：我愿意换吗？

在温暖里完整长大的那个女孩，在每一个岔路口都会做不同的选择。她在某些方面更勇敢，在另一些方面少了那种淬炼过的东西。她还会离开中国吗？还会去科大吗？还会，以某种方式，出现在与他相遇的那个经纬度上吗？

悲伤是可怕的。悲伤也是一种燃料。我不确定自己是否愿意由另一种东西建成。

第二宇宙：留在四班

在沔中的第二年，我从四班换到了五班。

就是那一次调整，改变了我高中认识的每一个人。不同的同桌，不同的午饭聊天，不同的考前互借答案的人。你在十六岁建立的网络，会向外涟漪几十年——跟谁一起备考，跟谁竞争，后来介绍给室友的是谁，给你发来那封改变一切的招聘信息的是谁。

在第二宇宙，我留在了四班。

我想人生的大方向大致相同。一所不错的大学，出国留学，AI浪潮准时抵达。路径和现在差不多。

但质地完全不同。我记住的那些笑话。我在梦里仍然看见的那些脸。1993年某个星期二那种特定质感的孤独。

这个宇宙里没有什么戏剧性。也许这正是关键。最大的转折往往是最安静的——一次班级调整，一次换座，一个从帽子里抽出的名字。

第三宇宙：北大，文学

我没有去科大读数学。我去了北大，学我真正喜欢的东西：文学、历史、人文。我写了对自己重要的文章。我把所有书都读了。我成了那种按感情而不是学科来整理书架的人。

没有六年严格的科学训练。没有凌晨两点证定理。没有抽象代数把你击垮然后，不知怎地，把你重建成某种更精确的东西的那一刻。

在这个宇宙里我写作更流畅。我有更大的词汇量来描述内心生活。我可以像科学家命名化合物一样命名情绪。

但我少了一些数学放进我里面的东西。对持续困惑的耐受。对”只要足够长时间地抱住一个问题，它终将解开”的信念。对抽象的那种安适，是我直到开始做AI才意识到自己拥有的。

在所有宇宙里，这个也许是最诱人的。也是我最会后悔的那个。

如果数学是我的骨骼，文学是我的皮肤，那我现在活着的，是这两者之间的那个生命。

第四宇宙：这一个

我去了科大。数学系。六年那种不管你是否舒适的训练。

我在那里遇到了我老公。他是那种你越了解就越发现他在成为自己的人——这比听起来更罕见。我们一起离开中国。我们在一个不属于我们的国家建立了生活，这意味着我们必须从零开始建造一切，这意味着我们知道我们建造了什么。

三个孩子。三个已经在以我没有预料到、也无法框住的方式让我惊喜的人。

一家创业公司。AI浪潮——我很早就看到它来了，因为数学教会你在指数增长对所有人可见之前，就认出它的形状。

还有这个：我所学与我所爱之间的那道裂缝。一个能追随证明过程，也能在晚上十一点完全迷失在一首唐诗里的人。一个想要公司有意义又想写作的创始人。一个爱着她选择的国家，也哀悼她离开的那个国家的移民。一个between worlds的女人。

Between worlds是不舒适的。Between worlds也是唯一能同时看见两面的地方。

真正的问题

校友群里的朋友们想知道我会选哪个宇宙。

我一直在用错误的方式思考这个问题。

问题不是”抽象地说我更喜欢哪种人生”。问题是：那个宇宙里成长起来的我，还会想要我现在想要的东西吗？

第一宇宙的那个女人——有妈妈，完整长大——她也许更温暖。更安住于自己的身体。少了一种特定质量的驱动力。我不确定她会创业。我不确定她需要。

第三宇宙的那个女人——北大文学学位——她写得比我好。但她可能没有数学六年给我的那种奇特的对模糊的耐受。她可能不适应创业中”抱住未知，继续走”的那些部分。

我越想越觉得：活了另一种人生的那个你，也不会想换回来。她已经成为她所成为的。她不知道自己缺少什么，因为她从未拥有过。她，大概，也很好。

这是关于平行宇宙，物理教科书不会提到的事：每一个版本的你，都是她自己故事的主角。 每一个版本都建立了她的依恋、她的爱、她的习惯、她对何为重要的感知。每一个版本，当被问到是否愿意换，大概都会犹豫。

我真正希望那个技术做什么

我说我相信AI + VR + 未知技术会让这成为可能。让我更具体说说我的意思。

我不想去活另一种人生。我想去探访它。就像你探访一个你不是在那里出生的国家，在那里待两周，然后带着它的改变回家，但还是回家。

我想和我妈妈吃一顿饭。哪怕一次。哪怕是在模拟里。哪怕全程知道这是模拟。我想坐在她对面，看她移动双手，理解我身上某些东西是从哪里来的。

我想看我哥哥健康地、年轻地、作为他自己，走进一个房间。

我想在北大花一个学期研究宋朝，学得一塌糊涂，爱每一分钟。

这些不是对我现有人生的替代。而是延伸。意识的扩展包。

人类最深的渴望不是活得不一样。而是活得更多。

人生的遗憾，不是走错了路，而是路太少，时间太短。

最后

我在第四宇宙写这篇文章。唯一一个我真实活过的宇宙。

它，难以置信地，是个好宇宙。数学和对文学的渴望。移民生活和创业。三个已经要在我无法预测的方向超越我的孩子。那个在合肥宿舍走廊里遇见的男人，到现在，以某种方式，还是令人意外。

我不会换。我这样说，同时完全清楚，第一宇宙的那个女人也会对她的宇宙说同样的话，同样的全然。

也许这就是全部的意义。

宇宙给了我们每人一种人生。我们把它变成了自己的。我们别无选择。其他的自我是真实的——也许，在波函数的某个分支里，真的真实——但这个自我，这个由失去、选择、孩子、定理、读到一半的书和深夜文章特定堆积而成的我，是我知道怎么做的那一个。

暂时，这已经足够。

暂时。

Subscribe now

Share

Leave a comment

English version of this essay:
Between Worlds -- Special Essay: Parallel Universe

I’ve been building on top of LLMs for over a year. Inference pipelines. Embedding models. Agents that fill out forms. I understood the inputs and I understood the outputs. What happened in between was something I mostly trusted.

Week 9 of this course was the week the middle became visible.

The lecture covered four ideas — all of them optimizations, all of them now standard in every frontier model you’ve heard of. None of them are obvious. All of them are, once explained, almost elegantly inevitable.

Here’s my attempt to pass that clarity on.

First: the position problem

The original 2017 Transformer had a quiet flaw. It didn’t inherently know where words were. “Dog bites man” and “man bites dog” are the same words — the model needed something extra to tell them apart.

The original fix was blunt: add a positional signal to each word’s vector before it enters the model. It worked. But it didn’t generalize well to documents longer than what the model trained on, and it made training data packing inefficient.

RoPE — Rotary Position Embedding (2021)

The elegant solution — now standard in LLaMA, PaLM, Mistral, and most modern models — is to not add position to the word vectors at all. Instead, inject position directly into the attention calculation, the step where the model decides which words to pay attention to.

The mechanism: rotate the Query and Key vectors by an angle that depends on their position. Imagine each word’s “question” and “answer” vectors being spun on a dial — early in the sentence, a small rotation; later in the sentence, a larger one. When two words compute their attention score together, the difference in their rotation angles encodes their relative distance. The model doesn’t need to know “I am at position 4.” It only needs to feel “I am 3 steps away from you.”

Think of two clock hands. You don’t need to know the time to know how far apart they are. The angle between them tells you everything.

What this unlocks: efficient document packing, and far better generalization to longer sequences than the model ever saw during training. Every major model since 2022 uses it.

Second: the vocabulary problem

Early language models treated words as atoms. Each word got its own vector. That creates an immediate problem: what do you do with a word you’ve never seen? A new drug name. A product code. A misspelling. The model goes blind.

There’s also a memory problem. A dictionary large enough to cover all of English needs millions of entries. Most of them almost never appear. An enormous, mostly-empty lookup table.

BPE — Byte Pair Encoding

The fix is to stop treating words as atoms and start treating subword fragments as the building blocks. “Running,” “runner,” “runs” all share the root “run.” If the model has learned what “run” means, it already has a head start on all three — even if it never saw “runner” during training.

BPE builds this vocabulary automatically, without any linguistic rules. Start with individual characters. Count which pairs appear together most often in your training corpus. Merge the most frequent pair into a new token. Repeat — thousands of times — until you have a vocabulary of around 50,000 pieces.

A concrete example from the lecture:

Training corpus: “low” (5×), “lower” (2×), “newest” (6×), “widest” (3×).

Round 1: the pair (s, t) appears 9 times — merge to [st]. Round 2: (e, [st]) appears 9 times — merge to [est]. Round 3: (l, o) appears 7 times — merge to [lo].

Final representation of “lowest”: [lo] + [w] + [est]. Three tokens. Not five characters. Not one unknown word. Three fragments the model already understands.

The practical result: ~50,000 tokens can represent any text in any language, including words that didn’t exist when the model was trained.

Third: the compute problem — and the most significant idea

If I had to pick one idea from this lecture that changes how I think about the whole field, it’s this one.

A 70-billion-parameter model requires roughly 280 GB of memory just to load. Every forward pass — generating one token — requires billions of matrix multiplications across all those parameters. Do that 500 times to generate a paragraph. For every user. Simultaneously.

The cost is not trivial. Scaling to GPT-4 without a fundamentally different architecture would have been economically unsustainable.

MoE — Mixture of Experts

The insight: not every parameter needs to activate for every token.

Replace the single feedforward network in each transformer block with N smaller networks — the “experts.” Add a lightweight router whose only job is to look at each incoming token and decide which two experts are most relevant. Route the token there. Let the others idle.

The result, verified in Mixtral 8x7B: load 46.7 billion parameters into memory, but activate only 12.9 billion for each token during inference. Match or exceed the performance of Llama 2 70B using 5× fewer active parameters per forward pass.

It reminded me of a Taylor expansion in calculus — approximating a complex function by keeping only the highest-contribution terms and ignoring the rest. You don’t need everything. You need the right things.

What the experts specialize in is not pre-assigned. It emerges during training. Researchers have found that individual experts develop soft preferences — some activate more on punctuation-heavy text, others on code, others on multi-step reasoning. The router learns this entirely from data.

Nearly every frontier model today uses MoE. GPT-4, Gemini 1.5 Pro, Llama 4. The signal is clear: this is the scaling strategy that works.

Fourth: the memory problem

Even after MoE, you still need to load all the experts into GPU memory. For Mixtral, that’s ~47 billion parameters. At 4 bytes per parameter (standard 32-bit floating point), that’s roughly 188 GB of VRAM. An NVIDIA B200 — one of the most powerful GPUs available today — has 192 GB. You’re right at the limit, with nothing left over.

Quantization

The observation: most of those 4 bytes are precision the model doesn’t actually need.

A weight stored as a 32-bit float looks like 0.028374072164297103 — nineteen significant digits. In practice, what the model cares about is that this weight is “small and positive.” You don’t need nineteen digits of precision to convey that.

Quantization maps those 32-bit floats to 8-bit or 4-bit integers. Find the largest weight in a layer, use it to define a scale factor, divide everything by that factor, round to the nearest integer, and store. When you need to compute, multiply back through the scale factor.

The practical numbers:

PrecisionMemory (70B model)FP32280 GBINT870 GBINT435 GB

35 GB fits on a single consumer GPU. The same model that required a server rack in 2022 runs on a laptop in 2025.

Accuracy loss: real but small. The token probability shift between FP32 and INT8 is measured in fractions of a percent on most benchmarks. Not zero — but well within acceptable range for most applications. Just like how you compress an image.

Not everything gets quantized. Operations that are genuinely precision-sensitive — softmax, layer normalization, the final output probabilities — stay at full precision. The trick is knowing which parts need all 32 bits and which are fine with 8.

What I’m actually taking away

These four ideas operate at different levels of the stack.
RoPE is about how the model encodes meaning.
BPE is about what counts as a word.
MoE is about how computation is allocated.
Quantization is about how weights are stored.
They don’t overlap, and they don’t compete. They stack.

Together they explain something I found genuinely surprising: a well-engineered 14-billion-parameter model in 2025 can outperform a 175-billion-parameter model from 2020 on many benchmarks. Bigger is not automatically better. Architecture decisions matter as much as scale.

Three hours. Four ideas. Each one of them is now running, invisibly, inside every AI product anyone in this industry is shipping.

That’s a good lecture.

Next: fine-tuning, RLHF, and what it actually means to align a model.

Subscribe now

Share

Leave a comment

I am currently taking MLEARN 530 at the University of Washington — graduate-level deep learning. It started with a 1943 paper about artificial neurons. Eight weeks in, it has arrived at the architecture that runs most of what the industry now calls AI.

I want to write about what I’m learning — not as a student taking notes, but as an engineer who has spent the better part of a year shipping AI products and is now, for the first time, understanding what’s actually happening underneath them. There’s a specific quality to that kind of learning. It doesn’t feel like studying. It feels like recognition.

This is for the engineers in our WeChat groups. For colleagues who talk fluently about model APIs and have never once needed to think about backpropagation. For anyone building seriously on top of AI systems and wondering whether it matters to understand them more deeply.

My honest answer: it does.

Eighty Years in Four Minutes

The pattern of deep learning history is worth holding onto, because it keeps repeating. Hype — winter — quiet breakthrough — sudden scaling. We are in a scaling phase right now, but the pattern doesn’t suggest winters are over.

1943 — McCulloch & Pitts sketch the first mathematical model of a neuron. A switch: enough inputs fire, the neuron fires. The idea is older than the hardware to run it.

1958 — Rosenblatt builds the Perceptron — a machine that actually learns from examples. The press calls it “the embryo of an electronic computer that will walk, talk, see.” The hype cycle begins, more than sixty years before Twitter existed to amplify it.

1969 — Minsky & Papert prove the Perceptron cannot solve even simple problems. Funding collapses. The first AI winter. A decade of silence.

1986 — Hinton, Rumelhart, and Williams publish the backpropagation paper — the algorithm that makes training deep networks possible. The match is lit. The wood won’t dry for another twenty-five years.

2012 — AlexNet wins ImageNet by a margin so large the computer vision community has to sit down. GPUs, data, and the right architecture arrive simultaneously. The modern era begins here.

2017 — Eight researchers at Google publish “Attention Is All You Need.” The Transformer. The architecture underneath GPT, Claude, Gemini, and most of what the industry now calls AI.

2022–now — ChatGPT. 100 million users in 60 days. The scaling phase. And already, the first signs of what comes after it.

Progress in deep learning has always required three things to converge: better algorithms, more data, and enough compute to run both. Each era broke down on one of those three. After 2012, all three arrived at once — and they haven’t stopped compounding since.

The Algorithm That Makes Learning Possible

Before the Transformer, before any of it, there is one algorithm that makes all of deep learning possible. Described in 1986, largely ignored until 2012, now running on GPU clusters large enough to require dedicated power plants. It’s called backpropagation.

A neural network is a chain of layers. Each layer receives a signal, transforms it, passes it forward. At the end, the network makes a prediction. You compare that prediction to the correct answer and measure the error — one number, the loss.

What makes backpropagation elegant: you can trace that error backward through every layer and determine precisely how much each weight contributed to the mistake. Nudge every weight slightly in the direction that reduces it. Repeat ten million times. The network has learned something.

The Core Mechanism

Forward pass: prediction made, layer by layer.

Loss: how wrong was it? One number.

Backward pass: trace the error back. The chain rule from calculus makes this tractable — every layer contributes a gradient, a direction of improvement.

Update: every weight nudges slightly toward less wrong.

Not just recalculating one wrong turn — recalculating every decision in the entire route, simultaneously, millions of times per second.

The algorithm was correct in 1986. What made it impractical was the surrounding infrastructure: activations that caused gradients to vanish, hardware too slow to iterate at scale, datasets too small to learn from. ReLU solved the first. GPUs solved the second. The internet’s data solved the third. The algorithm was waiting. Everything else had to catch up.

Why the Transformer Changed Everything

From roughly 2014 to 2017, the dominant architecture for language and sequences was the Recurrent Neural Network. RNNs processed text the way a careful reader does: one word at a time, left to right, carrying memory of what came before. Sequential by design.

The problem was structural. Sequential dependency made it impossible to use GPU parallelism fully. Each state depended on the one before — you couldn’t compute step four until you’d finished steps one through three. And for long sequences, meaning degraded as it traveled through too many intermediate steps. The architecture worked. It just didn’t scale.

The Transformer’s answer: let every word attend to every other word simultaneously, in a single step. No chains. No sequential processing. The mechanism is called self-attention.

Self-Attention in Plain Language

For every word in a sentence, self-attention asks: which other words should I pay attention to, and how much?

Each word generates three vectors: a Query (what am I looking for?), a Key (what do I represent?), and a Value (what should I contribute?). Every Query is scored against every Key. High match means high attention. The output for each word is a weighted blend of all Values — a representation shaped by the full context of everything around it.

The RNN was a relay race. The Transformer is a roundtable.

Multi-head attention runs this mechanism in parallel — multiple versions simultaneously, each learning to attend to different relationships. Grammatical structure, semantic similarity, long-range coreference. Position embeddings give the model sequence information without requiring sequential processing. Residual connections keep gradients clean through very deep stacks — GPT-3 uses 96 layers, and residual connections are why that’s possible at all.

Why Scale Became Possible

Full parallelism meant GPUs could run at real capacity. Training that took weeks became tractable in days. And because any two words interact directly rather than through a chain of sequential hops, long-range dependencies stopped being a structural limitation. These two changes together are what made GPT-3 possible. GPT-3 made everything else possible.

One more mechanism worth understanding: how the model generates text at all. During training, it always sees the correct previous words — teacher forcing. Efficient, but it creates a mismatch at inference, where the model feeds its own predictions back as input. At each step it outputs a probability distribution over its entire vocabulary. A temperature parameter controls how sharp or diffuse that distribution is — the difference between a model that sounds like documentation and one that sounds like a person.

The Architecture Is Not Settled

The Transformer has a structural limitation that grows more visible as you push it further. Self-attention is quadratic in sequence length. Double the context window, quadruple the compute and memory. This is why reaching 1M-token context required serious engineering — FlashAttention, KV-cache optimization, sliding-window attention — rather than simply more hardware.

In late 2023, Albert Gu (CMU) and Tri Dao (Princeton), both Stanford PhDs from the same lab, published Mamba — an architecture based on State Space Models. Instead of attending to all previous tokens, Mamba compresses the past into a fixed-size state that updates selectively with each new token. Linear scaling. Constant memory per token at inference. Substantially faster throughput on long sequences.

Transformer vs SSM — the key differences:

• Sequence complexity: Transformer = quadratic O(n²·d) · Mamba = linear O(n·d)

• Memory of past: Transformer = full KV cache, grows with context · Mamba = compressed state, fixed-size

• Inference cost: Transformer = grows with context length · Mamba = constant regardless of context

• Training: Both parallelizable, different mechanisms

• Tradeoff: Transformer = rich attention over full history · Mamba = compressed history, some fine detail may not survive

NVIDIA’s Nemotron family now replaces the majority of self-attention layers with Mamba-2 blocks and reports substantially higher inference throughput at equivalent quality. IBM, AI21, and several sovereign AI labs are deploying production systems on hybrid architectures. Whether pure Transformer or SSM-Transformer hybrid becomes the default for next-generation models is genuinely open.

What matters isn’t which architecture wins. It’s that the architecture conversation is happening at all — and that engineers who understand it will be the ones shaping what comes next.

What Changes When the Abstraction Shifts

There’s a transition underway that experienced engineers are feeling without fully naming. It isn’t that software engineering is disappearing. It’s that the abstraction layer is moving — and moving fast enough that engineers who only know how to work above it are losing ground to engineers who understand what’s beneath it.

Stanford Digital Economy Lab data shows entry-level developer employment down nearly 20% from its 2022 peak. Salesforce announced zero new software engineer hires in fiscal 2026. Google’s CEO noted that AI now writes more than a quarter of new code at the company. Average AI engineer compensation reached $206,000 in 2025, a $50,000 increase in a single year. These aren’t anomalies. They’re the shape of the transition.

The roles under pressure are those built around well-scoped, predictable coding tasks. The roles expanding require something different: knowing when the model is wrong, why it fails in the ways it fails, how to evaluate it systematically, how to architect systems that are reliable when the underlying model is not deterministic. That kind of judgment requires understanding the layer beneath the API.

Which is, finally, why this course matters to me beyond intellectual curiosity.

The tools the transition runs on — Hugging Face for open-source model infrastructure, LangChain and LangGraph for orchestration and agent pipelines, LoRA fine-tuning for adapting foundation models to specific domains, RAG architectures for grounding models in real data, and the emerging discipline of LLM evaluation — are now the vocabulary of production AI engineering. Not research. Production. The engineers who understand both the architecture underneath and the tooling on top are the ones building what runs at scale.

84% of developers now use or plan to use AI tools. Only 29% trust the output — down 11 points from the year before. That gap between adoption and trust is the real engineering problem of the next decade.

Twenty years in systems and infrastructure taught me to work fluently with abstractions I didn’t design. That skill remains valuable. But there is a meaningful difference between fluency and literacy.

Fluency is knowing how to use the stove. Literacy is understanding the heat.

I’ve been fluent for a while. I’m working on literacy now. The architecture of modern AI is not that hard to understand — it’s just rarely explained to engineers who are already good at their jobs and assume they don’t need it. I’m writing this because I think we do.

The kitchen was here the whole time.

— Lily · Seattle · Spring 2026 · MLEARN 530, ongoing

Leave a comment

Share

Subscribe now

Part 1: Five Ideas That Built Modern AI - Lily Qin