The Age of the Personal Assistant
We believe almost everyone will have a personal assistant of their own within a few years. It will order dinner, buy groceries, book appointments, buy tickets, pay bills, and manage your calendar. It will keep working while you sleep. As it takes care of these things, it will learn what you like, how you spend your days, and what worries you. It may come to know you better than almost anyone around you.
But to do all this, it needs access to almost every part of your life, online and offline. With each task, it learns a little more. Eventually, it may understand you better than you understand yourself.
So who owns the assistant and everything it knows about you, and who gets to control it?
Big Tech Cannot Be Trusted
Your assistant and its data must stay out of big tech's control. Your data is power. Today, Amazon knows what you buy, Uber knows where you go, and your bank knows what you spend. Even the largest companies see only part of your life. An assistant that handles all of it can put those pieces together. It knows your habits, your moods, and what you do each day. Whoever controls it can watch you and steer your choices. If your interests conflict with theirs, they can use your private information against you or cut off your access to the assistant.
Hand that power to a few companies, and your assistant becomes your master. That threatens your freedom and the foundations of a free society. We cannot leave ownership and control to the market. We have to design assistants so that you own and control them from the start.
We won't compromise on one principle: your life belongs to you. Our goal is an assistant that answers only to you, keeps your data on your own computer, and puts you in control of its actions and memories. No company should be able to read that data, and no government should be able to take the assistant away.
PhysiClaw
PhysiClaw is our first attempt to put that principle into practice. It's still rough, but it works. It can take care of tasks that used to keep you on your phone for hours. More importantly, it shows that you can build and own a useful assistant yourself, rather than rent access to one from a company that controls the technology and patents.
We've released both the hardware and the agent under the MIT license, so anyone can use or modify them. You can build the hardware for the cost of the parts. It uses standard, off-the-shelf components that are inexpensive, reliable, and easy to find. We want everyone to be able to afford an assistant of their own and keep control of it.
What's Still Missing
Even with open hardware and software, the assistant isn't entirely yours yet. The hardware is yours and the agent runs on your computer, but the large vision-language model it relies on still runs in the cloud. PhysiClaw still has to send your data to an AI provider.
We need a model that can run on a personal computer and still reason as well as the best models available. We think this will take a different architecture. Instead of storing all of human knowledge in its parameters, it would keep a small core that can think, plan, and act, with a basic understanding of the physical world. Other knowledge would be loaded as needed for each task.
A bee can fly, find food, plan routes, raise young, and help keep its colony alive. It does all this with fewer than a million neurons in a brain barely larger than a pinhead. Evolution keeps what works and gets rid of what doesn't. A bee gets through its day using very little energy, while a self-driving system uses orders of magnitude more. Powerful doesn't have to mean big.
The Road Ahead
Today's AI reminds us of the mainframe computers of the 1960s, expensive machines housed in large rooms and controlled by a few organizations. We believe AI will become something ordinary people can own, just as computers did. But bigger models won't get us there.
We expect a different approach to training, with vision at its core. Learning from images and video would help models understand the physical world and reason from its laws. That breakthrough might take five or ten years, or it might happen overnight. Once this approach is in common use, the answer may seem obvious. Someone may already be working it out.
We still see a place for large models. We need fast reactions as well as slower, more deliberate thought, much like Kahneman's two systems of thinking. When a predator lunged, our ancestors had no time to stop and think. An agent in the real world faces the same constraint. A fast system reacts while a slower one plans. Working together, they can do things neither could manage alone.
A powerful AI model that runs entirely on a personal computer would change more than the economics. It would help protect the freedom, democracy, and dignity that people have fought for. A free society depends on ordinary people having enough power to keep the few at the top in check. If a handful of companies control the most capable AI, they can use that advantage to gain even more power. If everyone owns a capable AI assistant, that power stays in ordinary people's hands.
We believe our freedom and dignity depend on it.
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。你入睡后，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，知道你喜欢什么、每天在忙什么，又有什么烦心事。
但这背后暗藏着风险。要打理你的日常生活，它就得参与其中的方方面面。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
我们认为，助理和它掌握的数据，都绝不能交到科技巨头手里。你的数据不只是数据，它是权力。今天，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。即使是最大的公司，掌握的也只是你生活的一部分。而一个替你打理日常生活的助理，却能把这些碎片拼成一幅完整的画像。你的一举一动、生活习惯，甚至情绪变化，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。一旦你的意愿与它们相冲突，它们随时可能利用这份权力对付你，以你的隐私相要挟，或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，助理和数据归谁所有、由谁控制，不能只由市场决定，而要在设计之初就确保所有权和控制权掌握在用户手中。
由此，我们立下一条绝不让步的原则：你的生活，属于你自己。我们要让数据留在你自己的电脑里，让助理只听从你，让它的行动和记忆都由你掌控。任何公司都不能窥探你的数据，任何机构都不能夺走你对助理的控制权。
PhysiClaw
PhysiClaw 是我们实践这一原则的第一步。坦白说，它才刚起步，还很粗糙，谈不上完美，但已经能替你处理过去要在手机上耗掉几个小时的琐事，让你少为这些事操心。更重要的是，它证明了人人都可以亲手搭建并拥有一个好用的助理，无需受制于巨头的技术和专利，也不必只做在线服务的用户。
我们相信，拥有这样一个助理，不该是少数人的特权。所以我们将硬件和智能体全部开源，以 MIT 许可协议发布，任你使用，任你修改。任何人都能动手搭建，只需承担零件成本。硬件采用常见的标准零部件，便宜、可靠，也容易买到，几乎人人都负担得起。我们要让每个人都能拥有自己的助理，不被价格挡在门外，也不受巨头掌控。
最后一个缺口
开源让你能亲手搭建并拥有自己的助理。但要让它真正只属于你，我们还差一步。硬件在你手中，智能体运行在你自己的电脑上，它的“大脑”却仍在云端。PhysiClaw 目前依赖的视觉语言模型体量庞大，需要部署在服务器集群上。因此，助理仍不得不把你的数据传给 AI 服务商。
要补上这个缺口，我们需要一种新的模型架构，让模型足够小，能在本地运行，同时具备媲美最先进大模型的推理能力。我们设想，它不必把人类的全部知识压缩进参数，只需保留一个精悍的内核。这个内核掌握物理世界的基本常识，会思考，会规划，会行动。其余知识，按任务需要动态加载；推理，始终居于核心。
这并非空想。蜜蜂就是一个例子，它能灵巧地飞行，采集花蜜，规划路线，还能哺育后代、维系蜂群的运转。而支撑这一切的，是一个比针尖大不了多少的大脑，神经元还不到一百万个。进化从不铺张，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，自动驾驶系统的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那些占满一个房间的计算机，价格昂贵，锁在机房里，只属于少数人。我们相信，AI 也会像计算机一样，从大型主机走向个人电脑，从少数人拥有走向人人可得。但要走通这条路，靠的不是把模型越做越大。我们期待以视觉为核心的新训练范式，用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。这样的突破，也许还要等五年、十年，也许就在一夜之间。等这种方法成为主流，人们或许会感叹，原来答案如此简单。而此刻，这个答案说不定正在某个人的脑海里悄然成形。
这并不意味着大模型已经走到尽头。我们需要两种智能系统的协作，就像卡尼曼所描述的两套思考系统：一套负责快速反应，一套负责深入思考。猛兽扑来时，我们的祖先来不及仔细思考该如何闪避。在真实世界中行动的智能体也需要这样的配合。快的一套负责反应，慢的一套负责谋划，两者协作，才能完成任何一套系统都无法独自完成的任务。
一个足够强大、又能完全在个人电脑上运行的 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。自由社会需要普通人有足够的力量，去制衡掌握权力的少数人。最强大的 AI 一旦被几家公司垄断，这些巨头就会凭借智能上的优势，不断攫取更多权力。相反，若人人都能拥有一个强大的 AI 助理，力量就能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到长久的守护。