The Age of the Personal Assistant
We believe the age of the personal assistant is coming. Within a few years, almost everyone will have one of their own. It will order your dinner, buy your groceries, book your doctor's appointments, do your shopping, get your tickets, pay your bills, and keep your calendar straight — all the little chores of life, taken off your plate. It will work while you sleep, on call around the clock. And because it does all of this, it will come to know you better than almost anyone around you: what you like, what fills your days, what keeps you up at night.
But a risk lurks underneath. To look after your whole life, it has to be woven into your whole life — online and off, big and small, everything passes through its hands. Day by day, it comes to know you better than you know yourself.
So there's no getting around one question: this assistant, and everything it knows about you — who owns it, and who controls it?
Big Tech Cannot Be Trusted
Our answer: neither can be allowed to end up in the hands of big tech. Your data is not just data — it is power. Today, even the largest companies see only one side of you — Amazon sees what you buy, Uber sees where you go, your bank sees what you spend — and none of them sees beyond its own fragment. An assistant that runs your whole life puts those fragments together into a complete picture: every move you make, every habit, every mood — it knows you inside out. Whoever holds it can watch you, influence you, and steer your choices; and the moment you want something they don't, that power can turn against you — your privacy used as leverage, your connection to your assistant cut off. Hand that power to a few companies, and the assistant stops being your helper and becomes your master. That threatens not only your own freedom, but the foundations of a free society. So ownership and control are not questions to leave to the market. They have to be built into the design from day one.
That leads to the one principle we will not bend on: your life belongs to you. In practice, that means your data stays on your own computer, the assistant answers to you alone, and everything it does and everything it learns is yours — no company can see it, no government can take it.
PhysiClaw
PhysiClaw is our first step toward that principle. To be honest, it's early days — rough, far from perfect — but it already works: the chores that used to eat up hours on a phone can now be handed to it, taking real weight off people's minds. More important is what it proves: that a useful assistant is something anyone can build with their own hands and own outright — not something locked up in a giant's technology and patents, available only as an online service.
We believe owning such an assistant should not be a privilege reserved for the few. So we have open-sourced all of it — the hardware and the agent alike, released under the MIT license, yours to use and modify freely. Anyone can build one; the only cost is the parts. The hardware is designed entirely around standard, off-the-shelf components — cheap, reliable, available everywhere — so almost anyone can afford one. The result is an assistant anyone can have — no one shut out, no giant in control, and your data beyond anyone else's reach.
The Last Gap
Even so, the assistant is not yet entirely yours. The hardware is in your hands and the agent runs on your own computer, but the mind it thinks with does not run locally. PhysiClaw is driven by a huge vision-language model, and models like that have to run on server clusters. The assistant has no choice but to send your data to an AI provider. This is the last gap, and the loop is not yet closed.
Closing it will take a different kind of model architecture: small enough to run locally, yet able to reason on par with the most advanced models of the day. Instead of compressing the whole of human knowledge into its parameters, it would keep only a lean core — one that thinks, plans, and acts. Knowledge gets loaded on demand, task by task; reasoning stays at the center. Everything else is stripped away, leaving only a basic common-sense grasp of the physical world.
This is not wishful thinking. Consider the bee: it flies with precision, forages with grace, plans its routes, raises its young, and keeps the colony going. All of that runs on a brain of fewer than a million neurons, barely larger than a pinhead. Evolution never over-engineers; it has always known how to cut to the essentials — trimming away everything redundant, keeping only what truly works. A bee gets through its day on next to no energy; a self-driving car burns through orders of magnitude more. Powerful, clearly, does not have to mean big.
The Road Ahead
AI today is like the computers of the 1960s: room-sized, expensive, locked away, and owned by the few. We believe it will travel the same road the computer did — from the mainframe to the personal machine, from the few to everyone. But we will not get there by scaling models further. We expect a new training paradigm: one built around vision, with images and video as its foundational training data, so the model learns the real physical world firsthand and reasons from the laws of physics. The breakthrough may be five or ten years away, or it may arrive overnight. When it does, and becomes the norm, people may be surprised at how simple the answer turns out to be. And it may already be taking shape, quietly, in some bright mind.
That said, this does not mean large models have hit a dead end. It means we need two kinds of intelligence — the two systems Kahneman described in the mind. One thinks fast and acts instantly; the other thinks deeply but slowly. Our ancestors could not spend seconds deliberating how to move when a predator lunged. Nor can an agent acting in the real world: the fast system reacts, the slow system plans, and together they do what neither can alone.
A powerful AI model small enough to run entirely on a consumer computer would change more than the economics. It would help us defend the freedom, democracy, and dignity humanity has fought so hard to win. A free society depends on ordinary people holding enough power to check the few at the top. If the most powerful AI is monopolized by a handful of companies, those giants will use their edge in intelligence to seize ever more power. If instead everyone owns a capable AI assistant, power stays spread across the many.
We believe that only then can human freedom and dignity be protected — and endure.
个人助理的时代
我们相信，个人助理的时代正在到来。未来几年，几乎每个人都会拥有一个自己的助理。点外卖、挂号、购物、订票、缴费、打理日程——这些琐事都可以交给它。你入睡时它仍在工作，全天候待命。也正因如此，它会比你身边几乎所有人都更了解你：你喜欢什么，每天在忙什么，又在为什么发愁。
但这背后暗藏着风险。要照看你的整个生活，它就得融入你的整个生活——线上线下，大事小事，都经它之手。日积月累，它对你的了解，甚至会超过你自己。
于是，一个问题摆在眼前：这个助理，连同它掌握的关于你的一切，归谁所有，由谁控制？
科技巨头不可信
我们的答案是：无论归属还是控制，都绝不能落在科技巨头手里。你的数据不只是数据，它是权力。今天，再大的公司也只能看到你的一个侧面——淘宝知道你买什么，滴滴知道你去哪里，银行知道你花了多少钱，但谁也看不到自己那一块之外的东西。而一个打理你全部生活的助理，却把这些碎片拼成了完整的你：你的一举一动、性情习惯，它都了然于心。谁掌握了它，谁就能监视你、影响你、左右你的选择；一旦你与他们的意愿相悖，这些权力随时可能转而针对你——以你的隐私相要挟，切断你与助理的连接。把这份权力交到少数几家公司手中，助理就不再是你的助手，而成了你的主人。这不只威胁你个人的自由，也动摇整个自由社会的根基。所以，归谁所有、由谁控制，我们不能听凭市场选择，而要从一开始就做好顶层设计。
由此，我们立下一条绝不让步的原则：你的生活，属于你自己。具体而言：数据留在你自己的电脑里，助理只听从于你，它做过的一切、学到的一切，都归你所有——任何公司看不到，任何机构拿不走。
PhysiClaw
PhysiClaw 是我们朝这条原则迈出的第一步。坦白说，它才刚起步，还很粗糙，谈不上完美，但已经可用：过去要在手机上耗掉几个小时的琐事，如今都可以交给它处理，切实减轻了人们的心智负担。更重要的是，它证明了一件事：一个好用的助理，人人都可以亲手搭建完成、完全归自己所有，而不是受制于巨头对技术和专利的垄断，只能在线使用他们的产品和服务。
我们相信，拥有这样一个助理，不该是少数人的特权。所以我们将它全部开源——硬件和智能体，均以 MIT 许可协议发布，任你使用，任你修改。任何人都能造一个，唯一的成本就是零件。硬件全部围绕标准现成件设计——便宜、可靠、随处可得，几乎人人负担得起。这样一来，每个人都能拥有自己的助理，没有人被拒之门外；它不受任何巨头掌控，也没有谁能染指你的数据。
最后一道缺口
即便如此，这个助理还没有完全属于你。硬件在你手中，智能体运行在你自己的电脑上，但思考的大脑，不在本地运行。驱动 PhysiClaw 的是庞大的视觉语言模型，需要部署在服务器集群上。助理不得不把你的数据传给 AI 服务商，这是最后一道缺口，尚未闭环。
要补上这道缺口，需要一种不同的模型架构：小到能在本地运行，推理能力却足以媲美当前最先进的大模型。它不把人类的全部知识压缩进模型参数，只保留一个精悍的内核——会思考，会规划，会行动。知识，按任务需要动态加载；推理，始终居于核心。剥离其余一切知识，只留下对物理世界的基本常识。
这并非空想。看看蜜蜂：它灵巧地飞行，优雅地采蜜，会规划路线，也会哺育后代、繁衍不息。而支撑这一切的，不过是一颗不足一百万个神经元的大脑，比针尖大不了多少。进化从不铺张，它天然懂得删繁就简：去掉一切冗余，只留真正有用的。一只蜜蜂一天消耗的能量微乎其微，自动驾驶能耗却高出几个数量级。可见，强大，未必等于庞大。
前路
今天的 AI，就像上世纪六十年代的计算机：占满一整个房间，昂贵，锁在机房里，只属于少数人。我们相信它会重走计算机走过的那条路——从大型主机到个人电脑，从少数人到所有人。但要走通这条路，靠的不是把模型越造越大。我们期待一种新的训练范式：以视觉为核心，以图像和视频作为基础训练数据，让模型直接认识真实的物理世界，并基于物理法则进行推理。这个突破也许还要五年、十年，也许就在一夜之间。等它真正到来、成为主流，人们或许会感叹：答案原来如此简单。而此刻，它说不定正在某个聪明的头脑中悄然成形。
话虽如此，这并不是说大模型走到了尽头，而是说我们需要两种智能系统——正如卡尼曼所描述的，人脑中有两套思考系统：一套反应迅捷，即刻行动；一套深思熟虑，行动迟缓。猛兽扑来时，我们的祖先没有时间去思考如何闪避。在真实世界中行动的智能体也是如此：快的一套负责反应，慢的一套负责谋划，两者协作，才能完成任何一个系统都无法独自完成的事。
一个强大却小到能完全跑在消费级电脑上的 AI 模型，带来的将不只是经济上的改变。它会帮助我们捍卫人类奋力争取的自由、民主与尊严。一个自由的社会，靠的是普通人手中有足够的力量，去制衡居于上层的少数。最强大的 AI 一旦被几家公司垄断，这些巨头就会利用智力优势，不断攫取更多的权力。反过来，若人人都拥有一个强大的 AI 助理，力量就分散在大多数人手中。
我们相信，唯有如此，人的自由与尊严，才能得以守护和延续。