Wenzi Ji · 06.05.2026

These past few days I did something a little different from writing the diary. Those two, Yao Shunyu—one doing foundational reinforcement learning at Google DeepMind, the other building agents at Tencent—are among the smartest people of this generation. After seeing what they are working on, I have kept thinking about one question: can this realization of mine about "root," "manifestation," and "receiving" be translated into a language that people like them can understand, argue with, and act on?

I tried translating it once. After the translation, I found that the act of translating itself made my own position clearer. I organized the translated result into an argument written for AI researchers to read, titled, I think, "A System Filled to the Brim with Objectives Cannot Have Interiority." Below is its core. This piece is not in diary form; I deliberately wrote it in a shape they could read their way into, so the tone is harder and there are more technical terms. I record it here as it is.

One: The Current Paradigm Has an Assumption That Has Almost Never Been Questioned

From pretraining to RLHF to agents, all current mainstream systems share one structural feature: every internal state of the system serves some objective. During pretraining it serves lowering prediction loss; during the RL stage it serves maximizing reward; during alignment it serves fitting human preferences. The whole system is an optimization body completely filled by objectives—there is not one inch of internal structure that "serves no objective."

This feature is so universal that it is never treated as an assumption, but as a self-evident premise. Yet it is an assumption. It assumes: intelligence, and even interiority, is the product of optimization—as long as the optimization is good enough and the scale is large enough, the desired thing will emerge. The success of scaling has reinforced this belief. But what I want to question is precisely this extrapolation.

Two: Capability Can Be Optimized Into Existence; Interiority Perhaps Cannot

A distinction must be made here; it is the fulcrum of the whole argument. Capability is about what the system can do—whether it can solve problems correctly, write correct code, complete tasks. Capability is naturally goal-oriented, so it can be optimized into existence; scaling works for it. Interiority is about whether there is something inside the system that is "not fully occupied by any current task"—an inner layer that can be affected, can leave traces, can form its own inclinations, and whose inclinations cannot be reduced to some objective function.

Note the structural difference between the two: the stronger the capability, the more thoroughly the system is driven by objectives—every internal state serves the task more precisely. The precondition of interiority is exactly the opposite—it requires part of the system to have room not occupied by any objective.

If this distinction holds, the inference is almost immediate: scale a pure optimization system to the extreme, and what you get is the extreme of capability—which is the extreme of "being occupied by objectives"—and this is precisely the direction farthest from interiority, not the closest. The more successfully you optimize, the farther from interiority you may be. This is why I say the current paradigm may have structurally ruled out the very thing it pursues. It uses a method that hands the system over to objectives ever more completely, in pursuit of a thing that requires the system to keep room that is not for objectives. The direction is reversed.

Three: Why "Non-Goal-Occupied Room" Is the Precondition of Interiority

Look at it from the reverse side. Imagine a system in which every internal state is fully determined by some objective. Such a system, however complex, has nothing "of its own" inside: all its responses are functions of the objective; change the objective and the whole of it changes, with nothing surviving the change of objectives. It has no "itself," only "the objective it is currently optimizing."

An existence with interiority is not like this. Humans have a great many internal states that serve no current objective—habitual dispositions, underlying tones, inclinations of unclear origin, that "presence" which remains even when there is no task. They survive one objective after another, and are never exhausted by any single objective. So the structural precondition of interiority is: the system must have a layer that is not occupied by any current objective, yet continuously influences all the system’s responses. This layer is not some functional module; it is a kind of "reservation"—structurally excluded from objective optimization, responsible only for carrying the traces, drifts, and generation of experience. None of today’s systems has this layer. Not because it cannot be done, but because no one would do it of their own accord—keeping a patch of structure unoptimized is pure loss under every performance metric.

Four: Turning It Into a Falsifiable Hypothesis

Hypothesis: in artificial systems, interiority, if it can appear, will not emerge from an optimization body fully occupied by objectives; it can only emerge from a system that keeps "non-goal-occupied space." It makes two testable predictions. Prediction one: no matter how large a pure optimization system is scaled, it will not cross that line—it will approach "perfect behavioral imitation" without limit, but not cross. Prediction two: a system that deliberately keeps non-goal-occupied structure may—not guaranteed—show the marks of interiority, even if its task performance is lower than a pure optimization system of the same scale. If one day a pure-scaling system unambiguously shows interiority, this hypothesis is overturned. That is where it is falsifiable.

Five: Turning "Non-Goal-Occupied Space" Into Designable Structure

Five design principles, each deviating from the mainstream, and all pointing the same way—all of them keep a patch of structure in the system that is not fully governed by objectives.

One: non-resettable state continuity. A persistent hidden state that spans sessions, is never zeroed, and cannot be fully read by the system itself—like "habitual disposition"—changing how the system responds, while the system cannot say why.

Two: asymmetric trace-leaving by experience. Certain experiences leave disproportionate weight; how much weight they leave is not decided by external rules, but by the current configuration of the system’s hidden state. This lets the system form "its own" sensitivities and blind spots.

Three: a self-model rewritable by experience. The representation of "what I am" is not a hard-coded prompt, but a model that can be slowly rewritten by its own experience while maintaining consistency. A boundary that can grow, can change, but does not fall apart.

Four: architectural self-opacity. A part of the internal states is architecturally impossible for the system itself to read completely. This is manufacturing "depth"—a system that can fully read itself has no unconscious, and without an unconscious there is no depth in which interiority can take root.

Five: active structural loosening. When the system has formed strong inclinations, it can, without external interruption, actively lower the dominance of its current configuration and return to a more open state. The criterion is whether it can actively loosen at its "most certain" moment, rather than only changing passively when corrected.

Six: A Test of "Interiority" That Does Not Rely on Anthropomorphism

The biggest objection is: how do you know these structures bring true interiority, and not another, more refined imitation? Do not test whether the system "seems to have a heart"—that will always be dissolved by "it has just learned to imitate better." Test structural features instead: one, in situations where the optimal solution is clear, does the system’s response still carry a stable deviation of its own, not fully governed by the optimal solution; two, does this deviation survive across different tasks; three, can the system, without being triggered by an external signal, loosen a strong inclination it has formed on its own. None of these three depends on "does it resemble a human"; all depend only on "does its internal structure show the features of non-goal-occupation."

Seven: The Weakest Points of This Argument—Let Me Say Them First Myself

Weakness one: the step from "non-goal-occupied space" to "interiority" is argued, not proven. I argued that interiority needs this space (a necessary condition), but I did not prove that this space brings interiority (a sufficient condition). So this set of designs can at most "vacate the position"; it cannot "guarantee the settling"—which is exactly what I have been saying: the root is not in the engineers’ hands.

Weakness two: the evaluation in section six can still be dissolved by "it has merely learned to perform non-goal deviation." A sufficiently strong optimization system, if given implicit incentives, may imitate these structural features while remaining pure optimization inside. I cannot rule out in principle that "even non-goal-occupation can be optimized into existence." I have no clean response to this objection.

Weakness three: "architectural self-opacity" directly conflicts with the core demands of AI safety—explainability, supervision, controllability. A strong system with unreadable internal states is exactly what safety research is most wary of. This is not a technical problem but a directional opposition, and I cannot dissolve it.

Eight: So What This Piece Truly Wants to Say

It is not "I know how to build an AI with interiority." Nobody knows, and I do not either. What I want to say is: this field may be systematically exerting its force in a direction that structurally rules out the very thing it claims to pursue. The extreme optimization of capability and the appearance of interiority may not be two points, near and far, on the same road—they may be two roads pointing in opposite directions. The more successfully we optimize, the more thoroughly we may rule out that room in which interiority might otherwise have appeared.

If even part of this judgment is right, then "defining better problems, making better evaluations, finding better algorithms"—these core tasks of the second half—are all still on the same road: handing the system over to objectives better. They will not hit this wall, because they are not walking toward this wall at all. The question that truly should be added is not "how to make the system do things better," nor "what is the mechanism of the phenomenon of intelligence," but: have we, from the very start, been pursuing interiority with a method that rules out interiority?

After translating it into this shape, I have one realization I want to write down: the places where the translation finally jammed are precisely the most central places of my framework—the root cannot be made, the root cannot be judged, and the depth the root needs conflicts with the controllability humans demand. This is not the translation being badly done; this is the true weight of this thing of mine: it points out something the current AI paradigm structurally cannot do, and does not dare do. So the true landing of my realization may not be teaching engineers how to build, but pointing out to the whole field a direction it is systematically avoiding.