Personal Assistant Age
We believe we are in the early stage of a big shift: the arrival of the personal assistant. Over the next several years, we expect almost everyone will have one. It will order your dinner and buy your groceries, book your doctor's appointments and clear your inbox, plan your trips and keep your calendar straight. 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 in your life — what you like, what you do each day, what you worry about.
Here is the part we think most people haven't fully thought through: the same feature that makes the assistant useful is the one that makes it dangerous. To help with all of your life, it has to see all of your life and be woven deep into it — everything you do, online and off, runs through it. Day after day it accumulates information about you until it holds more than you could ever write down yourself.
So it comes down to one question we need to answer: who owns all of that?
Big Tech Cannot Be Trusted
In our view, it cannot be the big technology companies. Your data is not just useful — it is power. And a personal assistant would concentrate far more of it than anything that exists today. Right now even the largest companies see only a slice of you — Amazon sees what you buy, Uber sees where you go, your bank sees what you spend — and no single one of them sees more than its own fragment. An assistant that runs your life takes all of it in — it knows your every move, reads your temperament and habits, and knows you inside out. Whoever holds it can surveil you, influence you, and steer your choices, and they can turn that power against you the moment what you want conflicts with what they want — blackmailing you with what they hold, cutting you off from your assistant, even threatening your safety. If you hand that power to a few companies, your helper becomes your warden. We believe that is a threat both to your own freedom and to the free and democratic society we share. So this is not a small design choice. It is the whole game.
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, you give the orders, and you own everything your assistant does and everything it learns. No company can see it, and no government can take it.
PhysiClaw
PhysiClaw is our first step toward that principle. To be clear, it is early — rough, unfinished, and far from perfect. But it works. Today it already takes real weight off people's minds by handling the small tasks on a phone that used to eat up hours. More important than what it does, though, is what it proves: that a useful assistant can be built in the open and owned by you, not rented to you and controlled by a company.
We also believe this cannot be a privilege for the few, so we build it and make all of it open — the hardware and the agent both, under the MIT license, yours to use and modify freely. Anyone can build one. The only cost is the parts. We designed the hardware around standard, off-the-shelf parts — cheap, reliable, and easy to find — so almost anyone can afford one. One assistant per person, owned by no monopoly, with no one shut out and no one left holding another person's data.
The Last Gap
Even so, the assistant is not yet entirely yours. The hardware is in your hands. The agent runs on your own computer, and its code is yours to read and change. But the mind it thinks with does not run there. Today PhysiClaw is driven by a large vision-language model, and the strongest of these are too big to fit on the computer in your home. So to use it, your assistant has to send your life to a server you don't own — which means the rule we set, that no company sees your life, is not yet fully kept. This is the last place where big tech still sits inside the loop.
To close that gap, we think we need a different kind of model: small enough to own, strong enough to reason on its own. Not a giant store of everything ever written in human history, but a sharp and simple core — the ability to think, to plan, and to act. This is what's called a world model. It gives up the trivia and endless memorized detail that routine work never needs, and keeps only what matters: the power to reason. Knowledge can be loaded when a task calls for it; the reasoning stays at the center. Strip away the rest and the model gets small enough for everyone to own.
This is not wishful thinking. Consider a bee: it flies with precision, forages with grace, plans its routes, raises its young, and keeps the colony alive — all on a brain of fewer than a million neurons, barely larger than a pinhead. Nature never engineers as wastefully as we do today; it follows Occam's razor, cutting away everything superfluous and keeping only what truly works. Think of the energy that bee spends in a day, then look at the compute today's self-driving cars burn through — clearly, powerful need not mean big.
The Road Ahead
AI models today are like the computers of the 1960s — room-sized, expensive, locked away, and owned by the few. We expect them to travel the same road the computer did: from the mainframe to the personal machine, from the few to everyone. They will be smaller, cheaper, and yours, getting stronger each year. This will not come from scaling them further. We expect a new training paradigm — one that learns to reason, rather than just to predict the next token and compress the entire internet into its weights, the way current models do. That breakthrough may be five to ten years away, or it may arrive overnight. When it does and becomes the norm, people may be surprised at how simple the answer was — and even now, it may be quietly taking shape in a few bright minds.
An AI model small enough to run entirely on a consumer computer — its power consumption low, its weights small enough to fit in your computer's memory — would be more than an economic change. It would help preserve the freedom, democracy, and dignity we have long fought for. A free society depends on ordinary people having enough power to check the few at the top. If a handful of companies own the strongest minds, that balance breaks: they gain great power over everyone who relies on them. If everyone can own one instead, the power stays spread out. A mind held by a few can be used to control people; a mind in everyone's hands keeps that from happening.
When that day comes, the loop is finally whole. Everything the assistant learns, everything it does, and the mind it thinks with — all of it yours, owned by you and answering to you, held by no one else.
个人助理的时代
我们相信，一场巨大的变革才刚刚开始：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属于自己的个人助理。它替你点外卖、买菜，帮你挂号、处理快递，为你规划出行、打理日程。你睡觉时它也在工作，全天候待命。正因为它做这一切，它会比你生命里几乎任何人都更了解你——你喜欢什么，每天在做什么，又在为什么发愁。
但有一点，大多数人并没有认真去思考：让它不可或缺的那一点，恰恰也让它变得危险。要照看你的整个生活，它就得看见你的整个生活，深深融入其中——你所做的一切，线上线下，都要经它之手。日复一日，它不断积累关于你的信息，最终掌握的，比你自己能记下来的还要多。
有一个问题我们需要回答：这一切，归谁所有？
科技巨头不可信
在我们看来，不能是那些科技巨头。你的数据不只是有用——它也是权力。而一个巨头，一旦掌控了所有人的个人助理，它所聚集的权力，将远超今天的一切。如今，哪怕最大的公司也只看到你的一个侧面——淘宝知道你买什么，滴滴知道你去哪，银行知道你花了多少——但谁都看不到自己那一块之外的东西。而一个打理你生活的助理，把这一切看在眼里——它知悉你的一举一动，洞悉你的性情与习惯，对你知根知底。谁掌握了它，谁就能监视你、影响你、左右你的选择；一旦你的意愿与他们相悖，他们随时能利用这些权力来对付你：用手中的信息要挟你，切断你与助理的连接，甚至威胁你的人身安全。若把这份权力交给少数几家公司，你的助理，反倒成了你的主人。我们相信，这既威胁你个人的自由，也威胁我们共处的这个自由民主的社会。所以这不是一个无足轻重的设计选择，而是事关全局的考量。
这引出我们绝不让步的一条原则：你的生活，属于你自己。具体来说，就是：你的数据留在你自己的电脑里，助理只听从于你，它所做的一切、所学到的一切，都归你所有。任何公司都看不到，任何机构都拿不走。
PhysiClaw
PhysiClaw 是我们朝这条原则迈出的第一步。坦白来说：它才刚起步，还很粗糙，远谈不上完美。但它能用。那些过去要耗掉好几个小时的在手机上处理的琐事，如今交给它，实实在在地减轻了人们的心智负担。不过，比它能做什么更重要的，是它证明了什么：一个好用的助理，每个人都可以独立搭建完成、归你所有——而不是技术和专利都被科技巨头垄断，你只能用他们的现成产品。
我们也相信，拥有这样一个助理，不该是少数人的特权，所以我们把它做出来，并且全部开源——硬件和智能体，遵从 MIT 许可协议，你可以任意使用和修改。任何人都能造一个，唯一的成本就是零件。硬件全部围绕标准的、现成的零件来设计——便宜、可靠、随处可买——几乎人人都负担得起。每个人一个助理，不被任何科技巨头掌控，没有人被拒之门外，也没有谁能觊觎你的个人数据。
最后一道缺口
即便如此，这个助理还没有完全属于你。硬件在你手里。智能体跑在你自己的电脑上，代码任你阅读、修改。但它用来思考的那个大脑，并不在那里运行。今天，驱动 PhysiClaw 的是一个庞大的视觉语言模型，而其中最强的那些，无法装载到你家里的电脑。于是为了用它，你的助理不得不把你的数据传输到 AI 服务商的服务器——这意味着我们追求的理念，“任何公司都无法窥探、掌控你的全部数据”，还没有完全做到。这是科技巨头还把持着的最后一环。
要补上这道缺口，我们认为需要一种不一样的模型：小到能在本地运行，强到足以独立推理。它不会把人类至今的所有知识都压缩进权重，而是只保留一个精悍而简单的内核——会思考，会规划，会行动。这就是所谓的世界模型。它舍弃那些日常工作从不需要的琐碎知识和无尽的记忆细节，只留下真正要紧的：推理的能力。知识可以在任务需要时再加载进来；推理，始终居于核心。剥去其余的一切，模型就小到人人都能拥有。
这并非空想。看看一只蜜蜂：它灵巧地飞行，优雅地采蜜，规划路线，哺育后代，繁衍不息——而支撑这一切的，不过是一颗不到一百万个神经元的大脑，比针尖大不了多少。大自然的进化，从不像人类今天的设计那样笨拙；它遵循奥卡姆剃刀，砍去一切多余，只留下真正管用的。想想这只蜜蜂一天所消耗的能量，再看看今天的自动驾驶要烧掉多少算力——可见，强大，未必等于庞大。
前路
今天的 AI 模型，就像上世纪六十年代的计算机——大得占满一整个房间，昂贵，锁在机房里，只属于少数人。我们相信它们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。它们会更小、更便宜、归你所有，而且一年比一年更强。这不会靠把它们造得更大来实现。我们期待一种新的训练范式——它学会推理，而不是像如今的模型那样，只去预测下一个词、把整个互联网压缩进自己的权重里。这一突破，或许还要五到十年，也或许就在一夜之间。等它真正到来、成为主流，人们或许会惊讶：原来答案如此简单。而此刻，它说不定正在某些聪明的头脑里悄然酝酿。
一个强大却小到能完全跑在消费级电脑上的 AI 模型——耗电低，权重小到能装进你电脑的内存——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，靠的是普通人手里有足够的力量，去制衡居于上层的少数。若最强的那些 AI 被少数几家公司垄断，这种平衡就会被打破：他们将对每一个依赖他们的人，握有巨大的权力。而如果人人都能拥有一个强大的 AI 助理，力量就分散在大多数人手中。
等那一天到来，最后一道缺口才算真正补上。个人助理学到的一切，做过的一切，以及它用来思考的那个大脑——全都属于你，归你所有，只听从于你，不被任何公司或机构控制。