Original video in Chinese.
Key Takeaway
- The core of sovereign AI is three rights: the right to exit (after a provider bans your account, degrades the model, or raises prices, can you switch seamlessly?), the right of ownership (can your conversations, memories, and history be fully exported and migrated to the next system?), and the right to define (must the AI act according to your goals, workflows, and evaluation standards, rather than being governed by the platform’s default logic?).
- The engineering standards and core assets of sovereign AI: the three engineering standards are replaceable, migratable, and recoverable; the four core assets are data, memory, workflows, and evaluation. It is a personal infrastructure controlled by the individual, capable of quickly restoring core capabilities after external services change.
- The essence of sovereign AI is the awakening of subjectivity: when AI shifts from a tool to an extension of personal capability, you can hand execution over to AI, but you cannot outsource goals, judgment, and value standards along with it. Sovereign AI is not a technical preference, but the infrastructure for rebuilding individual sovereignty in the AI era.
Once a person uses AI deeply enough, they will definitely develop the idea of “sovereign AI.” In plain language, it means:
Is this AI mine?
Let me give a few examples; I’m sure everyone will deeply relate.
At the service level: when the provider bans your account, degrades the model, or raises the price, can I switch to another one and keep doing the original work?
Over the past two years, we’ve experienced this kind of thing far too many times. Just think about the top three—OpenAI, Anthropic, and Google. Which one hasn’t annoyed people repeatedly?
So when we subscribe to these AI services, we don’t own intelligence; we’re only renting intelligence. That’s also why a small group of people has always been doing local deployment.
This is a right to exit, meaning I can refuse to be tied to you. And this right to exit has nothing to do with cost performance.
At the data level: can my conversations, memory, and historical materials be fully exported, independently stored, and handed to the next system for continued use?
Many people build knowledge bases, but that still isn’t enough. Because traditional knowledge bases mainly store relatively static materials.
What about all the large volumes of daily conversations you have with AI?
That’s why I built a memory system for newtype OS, because I wanted to fully control ownership of the data. This memory system includes full preservation of conversations, as well as automatic distillation of conversation content. All of it is stored locally in Markdown form.
Then all my important discussions are completed inside newtype OS. The dirty work, the repetitive work, and the purely executional work are handed over to Codex.
The direct forces and the outsourced team must, of course, be treated differently.
At the execution level: can AI act according to my goals, processes, and evaluation standards, rather than being governed by the platform’s default product logic?
Why were Skills so popular before? Because the question of how exactly a task should be done, and to what standard it should be done, is decided by people. It has nothing to do with how powerful your AI is.
I didn’t just build Skills; I also turned the logic of content production into a multi-Agent orchestration framework. This framework is the core of newtype OS.
At the execution level, how things are done and what counts as well done is the right to define.
The right to exit, the right of ownership, and the right to define—what I just mentioned—I don’t believe you haven’t thought about them.
So the question at the beginning—“Is this AI mine?”—what it is really asking is:
Why is a system that has deeply participated in my thinking, work, and life ultimately controlled by someone else deciding what it remembers, what it can do, and when it stops serving me?
This is actually a demand for “sovereign AI.”
As for the concept of “sovereign AI,” I gave a full explanation in Knowledge Planet:
It is personal infrastructure controlled by the individual, with data, memory, workflows, and evaluation all in the user’s hands, capable of replacing models, migrating systems, and quickly restoring core capabilities after external services change.
If broken down, sovereign AI has three engineering standards—replaceable, migratable, and recoverable—and four core assets—data, memory, workflows, and evaluation.
As for how to implement it, I also posted an article in the Planet. This implementation path consists of five steps in total, and I’ve already completed the first three and started preparing for the last two.
In fact, the specific steps are not the most important thing. What really matters about sovereign AI is what it represents behind the scenes. As I said in the Planet:
When you start asking, “Is this AI mine?”, what you care about is no longer just technical things like data backup or local deployment. In fact, your subjectivity has already begun to awaken.
That is to say, I can hand execution over to AI, but I cannot outsource goals, judgment, and value standards along with it. I can use the system, but I will not accept being defined by the system.
When AI shifts from a tool to an extension of personal capability, sovereign AI is no longer just a technical preference, but a demand for subjectivity.
Sovereign AI is not the end point, but the infrastructure for rebuilding individual sovereignty in the AI era.
OK, that’s all for this episode. If you want to understand AI, want to take back individual sovereignty, and want to find like-minded people, come join our newtype community. See you next time!