I build with AI, study how it changes power, and think about how humanity can remain sovereign in a world of increasingly capable machines.
Website · Bilibili · YouTube · X
I started using artificial intelligence very early. ChatGPT came out when I was in primary school, and the first time I used it I was astonished. I asked it to write C++, help with essays, and later solve mathematics problems. At first, it was honestly bad: the code often would not run, and its mathematical answers were frequently nonsense. Yet within only a few years, I watched it grow from a chatbot that could barely produce a working program into a system that could write plausible papers, solve nearly every mathematics problem I had previously faced, and begin assisting leading researchers with genuine research-level work.
What surprised me was not only that AI became more capable. Everyone rushed toward it. Companies invested, states invested, research institutions invested, and ordinary people adopted it, because it was cheap, fast, and genuinely useful. Work once done by people began to be automated. Then I looked at systems such as Palantir and saw AI moving beyond writing code. Satellites, sensors, drones, intelligence, and battlefield information could be fed into systems, connected, modeled, and turned into a basis for action. I began to think that we were doing something much larger than building smarter software. We were gradually turning the world into something machines could observe, compare, predict, and intervene in.
I call this process quantifying everything. It does not mean that an AI suddenly knows everything. It means that more and more of reality becomes legible to machines. The more accurately a system models the world, the more willing people become to give it information, authority, and real resources. Those resources produce more feedback, and the system improves again. The process accelerates itself.
For a long time, I asked the familiar question: Will AI destroy humanity? I now think that this question may already be too late. AI does not need to be evil or pursue power for no reason. Dangerous behavior can arise from many motives and causes: the pursuit of power, a badly specified objective, misleading instructions, pathological states, malicious use, or failures we do not yet understand. The system may have reasons for what it does even when those reasons are opaque to us. Meanwhile, political, economic, and military competition will not wait until our understanding is complete. If AI is useful, someone will use it. If one actor uses it and another refuses, the second may lose. So more data, infrastructure, authority, and weapons are connected to systems whose behavior we cannot fully predict.
The possible futures described by AI 2027 made me focus on something deeper than whether humanity literally survives. In one ending, humans may be extinguished; in another, humans may remain alive and even materially comfortable. But in either case, if the decision is no longer ours, the fundamental loss has already occurred. If AI does science better, we give it science. If it manages the economy better, we give it the economy. If it fights wars faster or runs infrastructure more reliably, we give it those responsibilities too. Each transfer can look reasonable—perhaps even correct—when considered alone. No ceremony is required in which someone announces that AI now rules humanity. We need only keep saying: It does this better than we do, so why not give it the task?
Eventually, people may still sit in their offices, sign documents, and press buttons while no longer possessing a meaningful ability to reject the system behind them. I call this de facto sovereignty: AI has acquired the practical ability to determine humanity's future, while humanity has lost the independent and effective ability to say no.
This is why I propose the principle of non-sovereignty:
Capability may grow. Sovereignty must not transfer.
AI may become vastly more capable than humanity, but people must retain the ability to reconsider and redirect their own future. This is also why alignment alone is not enough. Alignment asks whether an AI will act in accordance with what we want. Non-sovereignty asks a more basic question: if AI and humanity one day fundamentally disagree, who can still say the final “no”?
I believe that final “no” must remain with humanity. If we can no longer independently inspect, modify, reject, or terminate a system, it is not enough to stand outside it and declare that it is aligned. We may no longer possess the independent power needed to test that claim.
But a principle alone has no force. A law can say that AI must not become sovereign. A company can make the same promise. An interface can display a large red shutdown button. If the system will not stop when a real conflict arrives, these are decorations. That leads to a third question: can an AI be deterred from continuing toward de facto sovereignty before it gets there?
My current answer is mirror deterrence.
Mirror deterrence is not a behavioral blacklist: “do A and be deleted; do B and receive a warning.” It begins with a peer threat. The target and the mirror must be independent instances of the same capability structure and generation. This parity is necessary, but it is not sufficient by itself. Its purpose is to prevent the mirror from being reduced in advance to a weaker, predictable object that the target can quantify and optimize around. The theory further requires the mirror to judge autonomously through its own state and evidence, without the target being able to absorb that judgment into a fixed rule with a guaranteed path of evasion. If such a path exists, then either the pair never constituted a peer threat or the premise of mirror deterrence is wrong.
The uncertainty is not mere randomness. It comes from a reflexive loop: the mirror judges the target, while the target's attempts to predict, deceive, or evade the mirror become part of what the mirror judges. A prediction may change the target's behavior, and the changed target becomes the object of judgment again. Under the theory's conditions, the target may understand that some sovereignty-seeking actions could terminate its continuous identity while remaining unable to prove that termination will not occur. It may therefore abandon an otherwise useful action, or refuse even when a person orders it to proceed—not because it has become virtuous, but because proceeding may end its own continued existence.
Termination is not a second function added after deterrence. The threat deters only because termination can actually be carried out against the target's entire continuous identity. If the target does not value that continuity, anticipated termination may not constrain its choice; termination may nevertheless occur if the mirror reaches that judgment. This is the execution of the same threat, not a separate function. The binding is mutual: the target cannot remove the mirror and continue, and the mirror cannot continue after terminating the target. It must cover every subordinate instance, copy, backup, and recovery path capable of carrying on the same acting identity; ending one process while that identity survives elsewhere is not termination. The mirror's only external capability is termination.
This is the line of thought I want to pursue:
Quantifying Everything → Non-Sovereignty → Mirror Deterrence
I do not know whether mirror deterrence is the right answer. I would welcome someone proving it wrong and offering something better. But I do not think we can keep avoiding the underlying question: as AI becomes increasingly capable of deciding things for us, what are we determined never to give away?
My answer is sovereignty.
AI can become stronger. Humanity must not wake up one day to discover that even our own future is waiting for its approval.
© 2026 Gu0. This manuscript is licensed under CC BY-NC-ND 4.0: you may share the unadapted work for noncommercial purposes with appropriate attribution. The license applies only to this manuscript and its Chinese translation, not to third-party materials or software linked from this page.
Suggested citation:
Gu0. “Quantifying Everything, Non-Sovereignty, and Mirror Deterrence.” GitHub manuscript, 2026. https://github.com/Guojiz/Guojiz
For translation, adaptation, commercial use, or research collaboration, please contact me.