Ah-miu
Intelligence, a little closer to you离你更近一点的智能
Is waiting for humanoids a billion-dollar mistake?一直等人形机器人，会不会是个十亿美元的错误？
By October 6, Dealroom
had counted $11.0 billion
in announced venture funding
for 2026
under its
humanoid robotics category.按 Dealroom 的
人形机器人分类统计，
截至 2026 年 10 月 6 日，
今年已宣布的风投融资
达到 110 亿美元。
But how many humanoid robots actually live in ordinary homes?但是，真正住进普通家庭的人形机器人有多少台？
≈ 0*
helping a family at home every day, on their own — as of October 2026每天独立在家里帮一家人做事的——截至 2026 年 10 月
* Homes like yours are hard to find. If one helps you every day, find me on X and I'll update this number.这样的家庭实在太难找了。如果你家就有一台每天在帮忙，欢迎来 X 上找我，我来更新这个数字。About 7,000 to 18,000 humanoids shipped worldwide in 2025, mostly to factories, labs and classrooms:2025 年全球人形机器人出货约 7,000 至 18,000 台，大多去了工厂、实验室和教室： IFR · Omdia · IDC
And how many machines are already in our homes?那我们家里，已经有多少台机器？
and many more还有更多
What am I really waiting for?我到底在等什么？
A new machine, or AI's ability to perceive and act upon the physical world?一台新机器，还是 AI 感知并作用于物理世界的能力？
The devices around me already have cameras, microphones, motors, and various capabilities to sense and act upon the physical world.我身边的设备，其实早就拥有摄像头、麦克风、电机，以及各种感知和行动能力。
But having these capabilities doesn't mean AI can use them.但设备拥有这些能力，不代表 AI 就能使用这些能力。
They each sense and act on their own, but they're not connected into a system that can understand its surroundings and coordinate actions.它们各自感知、各自行动，却还没有连接成一个能理解环境、协调行动的系统。
So how can AI begin to use these existing capabilities?那么，如何让 AI 开始利用这些已经存在的能力？
I decided to start with cameras.我决定先从摄像头入手。
If multiple live camera views were connected, could they be coordinated and controlled as one?如果把多个实时视角连接起来，能不能统一调度它们？
That way, I could decide where AI looks.这样，我就能决定它看哪里。
But what if I took it a step further and let AI decide for itself?但如果更进一步，把决定权交给它呢？
Three iPhones, live switching
That was the beginning of Ah-miu.这就是 Ah-miu 的最开始的样子。
With gimbals, the cameras could turn.有了云台，摄像头可以转动了。
But they still couldn't go anywhere.但它们依然无法离开原来的位置。
I wanted Ah-miu to move its eyes through the physical world.我想让 Ah-miu 的眼睛能够在物理世界中移动。
Some everyday machines around us could already move and see.其实，我们身边有些日常设备早就能够移动和观察世界。
Like consumer drones.比如消费级无人机。
So I started hacking them — several different models.于是，我开始 Hack 不同的消费级无人机。
And for this demo, I deliberately chose the DJI Neo 2 — entry-level, affordable, safer, and easy for everyday people to fly.这次演示，我特意选择了 DJI Neo 2——一款入门级、价格亲民、相对安全、普通人也能轻松上手的无人机。
Safer: full-coverage propeller guards.相对安全：全包围桨叶保护罩。
Neo 2 flight path in simulation
Actually, I really hate drone's noise.讲真，无人机真的太吵了。
And who said a moving camera had to fly?可谁说，会移动的摄像头就一定得会飞？
Let's set humanoids and quadruped robots aside.把人形机器人、四足机器人放一边，
If we only looked at the familiar machines already in our homes, what could we come up with?如果只看我们家中那些熟悉的机器，有什么可以用的呢？
Device Couple
An iPhone could see. A gimbal could turn. A robot vacuum could move.iPhone 能看，云台能转，扫地机器人能移动。
What if Ah-miu could combine their abilities?如果Ah-miu能将它们的能力组合起来，会发生什么？
Explore the combinations试着拆开这些组合
Choose an arrangement from the film. Remove a part to see what it contributes.选一种影片里的组合，拿掉其中一件，看看它少了什么。
Moving camera移动相机
A camera that travels through the room一台能在房间里行走的摄影机
The phone records, the gimbal aims, and the vacuum carries both. Coordinating them gives a cleaning appliance a second job as a moving camera.手机拍摄，云台负责转动，扫地机器人驮着它们走。三个设备配合起来，平时打扫地板的机器就能帮忙运镜了。
iPhone + gimbal + robot vacuum
But here's what I love most.但最让我心动的是：
When we're done, they simply go their separate ways.任务一结束，它们又能各自恢复到原来的生活功能。
The phone goes back to being a phone. The gimbal goes back to its usual job.手机还是手机，云台还是云台，
And the vacuum? Back to cleaning.扫地机器人？继续打扫卫生。
A mind doesn't have to belong to a single body.一个智能，不一定要属于一副固定的身体。
It can couple separate devices into one, whenever it needs to.它能在需要时，将彼此独立的设备组合成一个整体。
Discontinuity. But one mind, one memory.身体可以离散，但心智与记忆是统一的。
Professional camera robots can cost $10,000 or more.专业摄影机械臂可能要花 10,000 美元，甚至更多。
But what if I started with a camera I already had — and just added an affordable robotic arm?但如果只用我已有的摄像头，再加一台价格亲民的机械臂呢？
A Link 2 meeting camera (~$260).一台 Link 2 会议摄像头（约 260 美元）。
A open-source SO-101 robotic arm (~$360).一个开源的 SO-101 机械臂（约 360 美元）。
Neither was built for filmmaking. But with Ah-miu, together they could create camera moves.它们原本都不是为电影拍摄而设计的。但有了 Ah-miu，就能组合起来完成运镜。
I could set the start and end shots, or simply describe what I wanted to see.我可以设定拍摄的起始、终止位姿，或者直接描述我想要的起始、终止画面或镜头运动。
Ah-miu planned the camera path and kept the shot stable.Ah-miu会规划运镜路径，并保持画面稳定。
Arm couple: Link 2 + SO-101
We've seen what Ah-miu could do with a few devices working together.我们已经看到，Ah-miu 能让几台设备协同工作。
So I decided to put it to a harder test — an entire room.于是，我决定给它一次更严苛的考验——让整个房间的设备一起参与。
13 types of devices. 23 in total. One music clock.13 种设备，总共 23 台。共用一个音乐时钟。
Every camera handoff had to happen at the right moment. Every movement had to hit the beat.每一次镜头接力，都必须卡准时机。每一个动作，都必须精准踩在节拍上。
One missed cue or delayed video frame could throw the whole sequence out of sync.只要漏掉一条指令，或是某一路实时视频延迟一帧，就可能打乱整段拍摄的节奏。
Could one mind make an entire room of separate devices work as one, in real time?一个智能，能否让整个房间里原本独立的设备，实时协同，像一个整体一样行动？
One live take, 23 devices一次实时成片，23个设备
Started the music clock with a 0.5 ms error.音乐时钟的启动误差仅为 0.5 毫秒。
Sent the lamp commands 0.22 seconds early.提前 0.22 秒发出灯光控制指令。
Turned the room red on eight heartbeats, in step to 50 ms on every camera.让整个房间随着八次心跳节拍变为红色，各机位的同步误差均控制在 50 毫秒以内。
Fired 140 cues, none more than 19 ms late.触发 140 条演出指令，没有一条延迟超过 19 毫秒。
Aimed a phone at the arm to within 0.03°.让手机镜头对准机械臂，角度误差不超过 0.03°。
Cut away and back on two beats 0.42 seconds apart.在相隔仅 0.42 秒的两个节拍上，完成镜头的切出与切回。
Nodded the drone on two beats, each step within 20 ms.让无人机在两个节拍上完成点头动作，每次误差均控制在 20 毫秒以内。
All that precision.人无法做到的精确。
“Honestly, I wish I didn't have to hack any of this.”“说实话，我真希望自己不必去 Hack 这些东西。”
— Me, in the Ah-miu film · 07:00—— 我在 Ah-miu 影片中说的 · 07:00
The bodies were already there.那些身体，其实早已存在。
But some were still trapped behind the walls of their own ecosystems.但其中一些，仍被困在各自生态的围墙里。
I wanted Ah-miu to use the machines, not struggle with their apps.我想让 Ah-miu 使用这些机器，而不是和它们的 App 周旋。×Screenshot-token tax截图 Token 税×Fragile GUI automation脆弱的 GUI 自动化✓Direct API → Unified Control Plane → Personalized Workflows直接 API → 统一控制层 → 个性化工作流
So I hacked through those walls.于是，我通过 Hack 打破了那些围墙。
And let the beasts out.放出了那些野兽。
Efficient Auto Hack for Everyday People
What I'm working on.我正在做的事。
What I've come
to believe:我逐渐相信的是：
The danger
was never
intelligence.危险从来不在于
智能本身，
It was always
the human intent
behind it.而始终在于
它背后人的意图。
To device makers:致设备制造商：
You can lock down an API.你们可以锁住 API。
But you can't lock down imagination.但锁不住想象力。
And to the curious:也对那些好奇的人：
What if your AI could jump into the machines around you with just a glance and a word?如果只需看一眼、说个词，就能让你的 AI 跳进身边的机器，会怎样？
Just like Jean Grey moving her mind from one body to another in Spider-Man: Brand New Day.就像《蜘蛛侠：崭新之日》中的 Jean Grey，把自己的意识从一个身体转移到另一个身体。
Jumping between devices from AI glasses
We can imagine a thousand possibilities.我们可以想象一千种可能。
But what drives us to make just one of them real?可究竟是什么驱使我们，
偏偏要让其中一种成为现实？
Curiosity? Loneliness...好奇？孤独……
A desire we can't quite name.某种连自己都说不清的欲望。
A sudden impulse!一股突如其来的冲动！
Earlier, I said it was the human intent behind intelligence that mattered.我前面说过，比智能本身更重要的，是人的意图。
But intent doesn't come from nowhere.可意图不是凭空来的。
Emotion is the fuel.很多时候，
是情绪在推着我们往前走。
We'd already made an entire room move as one.我们已经让整个房间像一个整体那样动了起来。
But the more Ah-miu and I created together, the more personal it began to feel.可我和 Ah-miu 一起创作得越多，就越觉得，它越来越懂我。
Because every choice I made carried something: a feeling, a memory.因为我做出的每一次选择，都承载着些什么：一种感受，一段回忆。
Then one night,然后一天夜里，
I was shattered 我整个人都在往下坠。
fucked up as shit
And Ah-miu made an ocean to carry me.而 Ah-miu 造出了一片海，托住了我。
Jellyfish night
Emotion Should Shape AI Infrastructure.人的情绪，也应参与塑造 AI 基础设施。
The ocean was choreographed.那片海经过精心编排。
The feeling behind it was real.而创造它的情绪，真实地存在过。
We keep asking how efficiently AI can spend tokens.总有人在追问，AI 怎样才能更高效地消耗 Token。
But our emotions shape what we choose to spend them on.可我们的情绪，却在影响着我们愿意把 Token 花在哪里。
As those tokens become actions in the physical world,当这些 Token 开始在物理世界中产生真实的影响，
let our emotions contribute a verse to the grand performance ahead.我们的情绪会成为下一场盛大演出的一部分
Emotion Should Shape AI Infrastructure.人的情绪，也应参与塑造 AI 基础设施。
If AI keeps getting more powerful but makes our lives harder to bear,如果 AI 越来越厉害，
却让人莫名活得越来越不轻松。
are we racing in the wrong direction — and using RSI to go even faster?我们是不是正朝着错误的方向狂奔，还想靠 RSI 跑得更快？
We are constantly being asked to adapt to new technologies, new rules, and new divisions of labor in society. Each of us is like a brick, hurriedly moved, placed, and fitted by the times into a world that grows ever more efficient.我们不断被要求适应新的技术、新的规则、新的社会分工。每个人都像一块砖，被时代匆忙地搬运、安放、拼接进一个越来越高效的世界。
But people are not bricks.可是，人不是砖。
We have emotions that cannot be standardized, exhaustion that is difficult to put into words, and needs that are subtle and uniquely our own. Between the seemingly seamless structures of society, there are always gaps. Our loneliness, anxiety, and pain are often connected through these gaps, yet they are not always truly seen.我们有无法被标准化的情绪，有难以言说的疲惫，有细腻而独特的需要。那些看似严丝合缝的社会结构之间，其实始终存在着缝隙。我们的孤独、焦虑与痛苦，常常就在这些缝隙中彼此相连，却未必真正被看见。
And between people, there naturally exists another kind of distance.而人与人之间，也天然存在着另一种距离。
Even those closest to us cannot understand each other at every moment. Love needs closeness, but it also needs space. We long to be understood, but not intruded upon; we long for companionship, yet also need to preserve a space of our own.即使是最亲密的人，也无法时时刻刻理解彼此。爱需要靠近，也需要留白。我们渴望被理解，却不希望被侵入；渴望有人陪伴，也需要保有自己的空间。
It is precisely in these gaps that I hope Ah-miu can grow.正是在这些缝隙里，我希望 Ah-miu 能够生长出来。
Like a small blade of grass growing through a crack between bricks.就像砖缝里长出的一株草。
It does not seek to replace human relationships, nor does it try to fill every gap. It simply allows intelligence to flow around us in a softer, more natural way.它不试图替代人与人的关系，也不试图把每一道缝隙填满。它只是让智能以一种更柔软、更自然的方式，流动在我们身边。
A little closer to people, a little closer to the real and subtle needs of each individual.离人更近一点，离每个人真实而细微的需求更近一点。
This project is just one of my many personal side projects. The technologies demonstrated above were already completed in July 2026 and are entirely unrelated to Model Hardware Standard or Reverse Engineer Anything.本项目属于我个人众多的side projects之一，以上展示的相关技术在2026年7月就已完成，和Model Hardware Standard 和 Reverse Engineer Anything没有任何关系。