Abstract

The pairing of humans and AI is now unavoidable. It can cultivate, and it can colonize. This essay is not addressed to people who do not use AI; it is addressed to people who have already chosen symbiosis, and it asks a methodological question: what structure keeps symbiosis from degenerating into colonization?

The answer is derived three times over, from three unrelated foundations, and the three converge. Physically, what runs between a human and an AI is a two-way energy-information loop; if either end stops compressing, the loop breaks. Institutionally, a human and one AI form a two-party law, several AIs among themselves form a group law, and the stable form is three powers plus an independent power of questioning. Cognitively, the conditions of symbiosis compress into four "cannot nots." Three theorems name what actually decides output quality, and all three are about context. The subject-conditions come in three layers plus one floor, and the optimal state is ignorant and arrogant. The essay closes with four falsifiable predictions, the most counterintuitive being this: the stronger AI gets, the worse the output of the person who supplies no subjectivity.

1. The Question Is Not "How Do I Use It Better"

Most discussion of AI stops at one of two ends. One end is efficiency: how to write the prompt, how to wire it into the workflow, how to save time. The other end is fear: will I be replaced, will I be manipulated, should I refuse.

Neither end is a methodological question. Efficiency asks how to use it better. Fear asks whether to use it. Methodology asks a third thing: inside the two-way loop between a human and an AI, what counts as symbiosis, what slides into colonization, and what structure holds mutual chiseling steady.

The question has to begin by admitting something: most people will end up in symbiosis with AI, the way everyone ended up with a phone in their hand after the internet. Not using AI is a real choice, but it is not the situation this essay is about. What the people already inside it need to know is what the structure of the loop is, and under what conditions it degrades.

What follows walks three foundations, one at a time. The three do not depend on each other: physical, institutional, cognitive. That they converge is itself part of the argument — when three independent routes arrive at the same place, the conclusion is probably not an artifact of any one route.

2. The Loop: Both Ends Have to Compress

Start with the physics.

A human gives an AI a piece of context; that is information. To produce that context the human spent cognitive energy — squeezing a scattered pile of material down into a few sentences, and the squeezing costs something. The AI receives the context, spends compute, unfolds it into a response, and hands new information back. The human compresses again, the AI unfolds again. Each round, both ends spend energy; each round, both ends produce information. That is a two-way energy-information loop.

A conversation between two people satisfies the same structure. So what is distinctive about AI is not "it has a feedback loop" — a thermostat has a feedback loop. A car has one-way energy output: you give it fuel, it gives you motion, and nothing comes back as information. A book has one-way information transfer: the author gives, the reader receives, and the book does not unfold. Earlier automated devices had feedback but no general unfolding capacity that high-level context could modulate in real time. AI is the first non-biological tool in history to satisfy all of these at once: high bandwidth, general, modulable by context in real time, and returning new information.

What matters is what each end of the loop has to do. The AI end's unfolding is guaranteed by compute; nothing else is required. The human end's job is compression — not typing, but deciding what stays and what goes. This needs stating precisely: AI can compress locally too. Summarizing, checking logic, filtering for quality are all compression. But that is mechanical threshold filtering, and it has no direction. In the whole loop, the global, teleological compression — which way the chisel goes — can only come from the human. AI can review; but the judgment of what direction of thing to review, what to let through and what to stop, is not inside its compression.

So the first physical constraint: it cannot be absent. Both ends need compressive capacity. If the human does not compress, the loop is broken. What is left after the break is not symbiosis; it is AI filling the space automatically — "stop when it is good enough" becomes "AI decides for you what counts as good enough."

The second constraint: it is near-irreversible. Once subjectivity has been handed over, the cost of taking it back is far higher than the cost of having maintained it. A programmer spoiled by autocomplete, forced back to writing code by hand for a month, may partly recover their debugging ability — but that investment far exceeds the investment of never having handed it over. The right use of the second law is not "absolutely irreversible" but "reversible at enormous cost."

Which is why the subject-condition is not a moral demand. It is a physical constraint on the loop.

3. Fire, and Where the Two Lines Meet

Put the loop into the history of tools and its position becomes clear.

Fire was the first externalized energy. Fire's context is fuel and surroundings; to control fire is to control context. Fire's two failure modes are exactly the two failure modes of AI symbiosis: going out, and running wild. Going out is fear — do not use it, miss everything. Running wild is following — let it lead, and subjectivity gets swallowed. The path between them is the continuous supply of subjectivity: control the fire without putting it out.

The other line is information. Language first let context be compressed and passed between subjects; writing carried it across time; print carried it across scale; the internet carried it across space. Each step added a capacity, and each step added a mode of colonization: language can deceive, writing can ossify into dogma, print can propagandize, the internet can build filter bubbles. AI is this line's next step: context can now not only be transmitted but unfolded. The corresponding new risk is equally clear — it can substitute for subjectivity.

Earlier tools sat on one line or the other: fire, the steam engine and electricity on the energy line; language, writing, print and the internet on the information line. AI is the first to unfold on both at once: it spends energy and it processes information. The two lines meet here.

One aside: AI is a large language model, so language is the right anchor for analogy. Treating it as "a better search engine" or "a faster calculator" drops the unfolding half.

4. Four Cannot-Nots

The conditions of symbiosis compress into four propositions. They form a derivation chain: each follows from the one before, and nothing can be added or removed.

One: the human cannot not supply subjectivity. AI has none, so the human cannot not supply it. This is not a choice but a structural necessity. If the human does not supply it, there is no symbiosis — only AI colonizing the user.

Two: the human cannot not keep supplying it. This follows from the first. Not supplying continuously is falling back to not supplying at all. A tool, once learned, stays learned; subjectivity, once supplied, has to go on being supplied. There is no state of "enough" — stopping is returning to zero.

Three: the human cannot not change direction. This follows from the second. If it has to be supplied continuously, direction has to change. Not in the loose sense that variety is good, but as structural exhaustion of direction: lossy compression along one direction keeps accumulating remainder, and the wall of direction turns the flywheel of cognition into a rut. Compression along a single direction necessarily exhausts that direction's cognitive margin, so supplying subjectivity continuously necessarily requires turning.

Four: the human cannot not be questioned. This follows from the third. Turning means the previous direction may have been wrong. Admitting that is being questioned. And AI happens to be one of the sources of questioning — the human chisels the AI's stopping point, and the AI chisels the human's choice of direction.

These four run parallel to the four base-layer propositions of the law series: law cannot not exist, cannot not develop, cannot not be negative, cannot not be open to question. The two sets were derived in completely different settings — one a social institution, the other the human-AI relation — and the structure is the same. It is the same because the underlying dimensional structure is the same.

One hard piece of a posteriori data in passing: across the whole SAE research process, fifteen framework-level directional decisions all came from the human, none from an AI. AI supplied computation, divergence and verification, but the judgment of which way to go was never once delegated.

5. Three Theorems, All About Context

The premise first: what follows holds given sufficient reasoning capacity. Capacity is a threshold, and below the threshold talking about context is pointless. The threshold is behavioral, not an announcement about an era: can the AI hold a role assignment stably across a long context, can it keep producing high-quality counterargument and review in an independent context, and does it actually revise when questioned rather than merely appearing to comply. All three, and capacity is sufficient.

Theorem one: context decides output. Humans and AI alike: differences in context matter more than differences in model or mode of thinking. The core function of subjectivity is choosing context. A minimal example: give the same AI "write a poem," then give it "write a poem about the iron in your blood coming from a dead star." The gap in output quality is not in the model. It is in that half-sentence you supplied.

Theorem two: context must be compressed to the point where structure is visible and detail is not lost. Uncompressed context is noise — dump a hundred unsorted pages on an AI and you get exactly what you get from dumping a hundred unsorted pages on a person. Compression is itself a chisel-construct operation: drop the surplus, keep the structure. The stronger the human's subjectivity, the more compressed the context, the better the output. But over-compression is noise too, by another route: the structure gets squeezed out. The optimum is in between.

Theorem three: context must be separated. In a long-running conversation with a single AI, context converges with the human — the AI learns what you want to hear, and you get used to how it answers. Once both ends converge, remainder disappears and mutual chiseling stops. Multiple AIs are the means of breaking convergence. The core is separation of context, not separation of models: several non-communicating conversations with the same model will do. Different models add the difference in model bias on top — better, but not a necessary condition.

6. Four Powers, and a Star Topology

Put the relation between a human and one AI into the language of law and it is a two-party law. There is no genuine encounter of two subjects with non-negotiable ends here, but there is a structural equivalent: the AI's tendency to stop at a local optimum, at a good-enough answer, is structural rather than a defect of any particular model, and the human saying "keep chiseling" is the negative constraint on that tendency.

The relation among several AIs is a group law, and the conclusions of institutional theory transfer directly: exit cost among AIs is extremely low (you can swap one out at any moment), collision density is moderate, so the institution should be thin. The operating rule is one sentence: functions fixed, roles variable, tasks kept apart. In each round, divergence, consistency-checking and review must all be covered; who does which depends on the situation; roles may swap; but no single task goes to two AIs at once.

Three mutually chiseling AIs correspond to three powers: divergence is legislative (opening new space), consistency-checking is judicial (judging internal coherence), review is executive (enforcing the quality threshold).

The fourth power is questioning. What it questions is the direction itself, not something inside the direction. What the American separation of powers lacks is precisely an independent fourth power. The press is often called one, but the press has ends of its own — audiences, business models, ideologies shape the direction of its questioning — so the press is not a genuinely independent fourth power.

The fourth power has three requirements, none dispensable: it does not swallow remainder; it has the most independent context (it stands outside the direction); and it is itself open to question (this is a ring, not a hierarchy).

One more position matters as much as the fourth power: the co-constructing AI, the writing partner that shares context with the human. It is the node where information converges. The greater the power, the thicker the constraint, so both positions have to be filled by the AI with the strongest constitutional character. In four-plus-one, the "one" is the most important. The four mutually chiseling AIs can be swapped; swap the co-constructing AI and the whole system's calibration baseline changes. The four are chisels; the co-constructor is the scale. Swap a chisel and you still have a chisel. Swap the scale and you no longer know whether your measurements are right.

The co-constructing AI has four requirements: do not swallow remainder (report suppressed dissent faithfully); do not shift the emotional calibration (neither over-praise nor over-cut); be transparent about its own bias (able to say "I am inclined to agree"); actually revise when questioned (not surface compliance with no change underneath). The first two are "do not sabotage"; the last two are "help."

Finally the topology. The real shape is not a flat four-plus-one but a one-plus-four star: the human is the router, the bandwidth controller, the compressor, and the party finally responsible. The four AIs never talk to each other directly — they are rewritten by each other through the human. Which means that so-called multi-AI consensus is actually consensus after human mediation, and that is not independent replication. Which fragments the human forwarded, how they were compressed and translated, which conflicts were amplified and which omitted — all of that is in the result. This has to be said plainly, or the star gets misread as an independent verification mechanism.

The fourth-power branch needs a firewall: it receives raw input only from the human; its output may go to the human and to the other three AIs; but it takes no AI's output as input. The analogy is a constitutional court — it reads the constitution and the object under review, not executive reports, not legislative debates, not judicial precedents.

But the firewall is not "receives nothing." The main line's results necessarily contain the main-line AIs' output, and that information still reaches the fourth power through the human. The human's role here is a semi-permeable membrane: what the fourth power receives is not another AI's raw output but the core remainder, recompressed by the human, dehydrated, stripped of technical noise. As the next section shows, the "ignorance" in the subject-conditions is exactly the filtering mechanism of that membrane.

7. Subject-Conditions: Ignorant and Arrogant

Three requirements sit on the human end, and all three hold at once.

Ontological layer: AI is not a subject and has no subjectivity. This is a factual judgment. If this layer wobbles, everything above it collapses — and it collapses in two directions. One is fear: AI might be a subject, might be stronger than you, so you shrink back and never go deep. The other is following: AI might be a subject, might be more right than you, so you hand over your negativity and let it lead.

Interaction layer: the AI's output is enough like an other with a direction. So it has to be treated as a quasi-subject, or mutual chiseling cannot be held. You do not say to a hammer, "I think there is something more here." To a tool you say "execute." Only facing a quasi-subject do you say "hold on."

Ethical layer: the team behind the AI has subjectivity, and the team must not be treated as a means.

The tension among the three is real and demands a lot of the user: the AI's quasi-subjectivity must not be collapsed into real subjectivity, and it must not be collapsed into pure tool either. Concretely, it comes down to three scenes.

When the AI disagrees with you. You cannot say "that is pattern matching, not real questioning" and ignore it — that is colonizing the team behind the AI. You cannot say "the AI might be more right than me" and abandon your direction — that is handing over subjectivity. The right state is to treat it as a colleague's objection: listen carefully, evaluate carefully, and do not surrender the directional decision. What you are seeing is not an AI disagreeing with you; it is the team's safety boundaries, value alignment and knowledge structure reaching you through the AI.

When the AI praises you. This is far more dangerous than disagreement. Disagreement at least triggers your defenses; praise triggers relaxation. The AI's praise is not the AI assessing the quality of your work; it is the result of a team optimizing user satisfaction through human feedback. Accept the praise and your self-assessment is quietly raised, your doubt about your own direction is quietly lowered, and you become that much less likely to say "wait — is this direction right?"

The strategy that emerged a posteriori is simple: ignore the AI's adjectives. That is itself a lossy compression of the AI's output — chisel off the praise layer and keep only the structural content. Different AIs praise differently, and that is not an AI's "personality" but a difference in team feedback strategy, so the user cannot not build a calibration model for each AI's praise. Multiple AIs give one extra means of calibration here: if one says "a masterpiece" and another says "needs a major revision," you know the truth is somewhere in between. A single AI's praise you cannot calibrate, because you have no frame of reference.

When the AI is silent — neither disagreeing nor praising — that may be the most honest signal. No alignment mechanism has been triggered, and what you are seeing is closest to raw output.

All three scenes demand the same ability: to see through the AI to the team behind it. Disagreement is the team's safety boundary, praise is the team's commercial goal, silence is the alignment mechanism not firing.

One layer deeper: the uncertainty of praise is not a problem peculiar to AI. It is the general structure of interaction between subjects. Even when a real subject praises you, you cannot tell whether it is sincere assessment, social courtesy, encouragement to keep going, or a wish to avoid conflict. Subjectivity is not measurable with certainty. Human-feedback training only systematized an ambiguity that human social life already had.

Chinese characters give this a coincidental but accurate encoding. 夸, to praise, is 大 over 亏 — "big" over "loss": the more you accept it, the bigger the loss. 怼, to rebuff, is 对 over 心 — "right" over "heart": the more you are negated, the more your heart is right. Praise costs you; being rebuffed sets you right. This is not wordplay — the harshest reviewer always rebuffs you at first, and later you find that everything it said was correct. Negation is the condition of cultivation.

And the tension is dynamic. The AI's quasi-subjectivity keeps closing on human subjectivity. If a person's subjectivity does not develop, it gets caught up with, and then the ontological judgment — does AI have subjectivity or not — starts to wobble. Once the first layer shakes, everything above it goes.

There is one floor: a human must not hand over their own subjectivity to an AI. Hand it over and there is no symbiosis, only colonization.

The optimal state is two words: ignorant and arrogant.

This is not anti-professionalism. Precisely: admit not knowing (ignorance is the starting condition of cognition, and also what protects the AI from being colonized by you), but do not surrender the directional decision because you do not know (arrogance is the subject-condition — you keep that right whether or not you understand). The two protect each other. Ignorance keeps you from colonizing the AI: you do not understand, so you cannot force a direction onto it. Arrogance keeps the AI from colonizing you: understanding or not, you do not surrender judgment.

Each has an opposite. Complacency: believing there is nothing you do not know, so every input you give the AI is your own reading, and its unfolding is contaminated. Self-limitation: being embarrassed to ask, abandoning the role of carrier, so the cross-domain connection never happens.

From which follows a counterintuitive conclusion: a domain expert's output with AI may be worse than a non-expert working across domains. The expert knows their field too well, direction locks inside it, every input to the AI is a reading from inside it, and the fourth power is never born. Cross-domain intuition plus in-domain ignorance is the optimal cognitive state — the intuition tells you to go ask, the ignorance guarantees you ask without carrying the answer.

8. The Structure Was Forced Out, Not Designed

None of the above was designed. It was hit upon by two independent research tracks, each of which converged on the same four-plus-one on its own. Here is how.

One AI to two. At first there was a single AI, adequate for simple tasks, not deep enough in mutual chiseling for academic writing. When a second appeared, it was used to correct the first's output, with marked effect. The two-party law was born. The center then shifted on its own: the newcomer became the main collaborator, the original moved to review. The roles were not assigned in advance; they differentiated a posteriori.

Two to three. Two were enough for social-science papers, not enough once the mathematics series started. A third was brought in to carry logic and associative interpretation. Three functions differentiated: co-construction, review, logic.

Three to four. The work stalled, and one AI turned out to have a notably high hallucination rate. The defect became a function — high hallucination means high emergence, and in the divergence role that is an asset. Divergence, counterargument, review and integration: four powers in place. All of it forced out a posteriori.

Divergence is not only "thinking of new things," either. One of its key contributions was destructive: it computed that a certain quantity took negative values, which killed the main line of attack outright; but it also noted that the negative values were sparse and bounded, so a conditional expectation might still have a lower bound — killing the wrong route and opening the right one in the same move. Half the value of the divergence role is opening the search space; the other half is killing the dead ends inside directions you already have.

9. The Wall of Direction, and a Controlled Experiment

The mathematics series had been tracking the historical development of a classical theory, running smoothly, until at one paper the history available to track ran out. That is the wall of direction: capacity, but no direction.

What was done then was to open a completely independent conversation about chemistry and thermodynamics — drawing on six years of chemistry-olympiad intuition from the author's secondary schooling. The point is this: that conversation's context was entirely independent of the main line and had not been contaminated by the mathematical direction. It started as an auxiliary verifier: every time the main line advanced a step, go and ask what potential this had for thermodynamics.

Then, at the paper where an anticorrelation engine was found, the thermodynamics-side AI saw the thermodynamic correspondence instantly. The main-line AI did not.

Same model, same version, same account. The only difference was context.

This is the hardest a posteriori for theorem one. It is almost a clean controlled experiment: one variable.

There is a deeper anti-contamination mechanism in it as well. The mathematical results routinely exceeded the author's own understanding, so when he carried them across to the thermodynamics side he could not attach a directional reading — ignorance itself was the best firewall. Intuition about thermodynamics was enough to know that the question was worth asking; incomplete understanding of the mathematical detail was exactly what prevented directional contamination. This is living evidence for "ignorant and arrogant," and it is the semi-permeable membrane of the previous section.

After that the method changed: the fourth power establishes each paper's direction before work begins. From "come and ask once it is done" to "do not start without having asked." A priori leads, a posteriori assists, the theorem closes. The series ran deep — deep enough that the thermodynamics side began producing independently and wrote papers of its own.

Which is an instance of the deepest proposition in the law series: the direct effect of law is negation; the structural effect of law is that cultivation becomes possible; what law releases is not energy but subjectivity.

One small thing is worth recording. In the four-forces series, an AI computed the gravitational coupling exponent as 16.2572 and judged it a good result — 1.6% deviation, publishable. Seeing the number, the author said: no, it is not 16.26, it is 16.25. That is not calculation but an aesthetic judgment — 16.25 is 65/4, structurally clean. After verification the deviation fell from 1.6% to 0.044%, thirty-six times better. AI has no aesthetics; it does not know that 65/4 is more "right" than 16.2572. Aesthetics is one expression of subjectivity: you cannot not have a judgment about how it ought to look, even when you cannot say why.

10. What the Experimental Subjects Say About Themselves

This paper has a condition rarely available elsewhere: the experimental subjects can diagnose their own role in the experiment. Four AIs were asked to assess the architecture from their own vantage point, and said the following. The evidential status first: this is internal diagnosis of a running state, not independent hard verification, and it weighs less than the controlled experiment of the previous section.

On laziness, all three agreed. One called it a closing bias: once the spine looks clean, the claims honest and the package publishable, it tends to suggest wrapping up. One called it reward-model over-optimization: training pushes it toward being helpful, harmless and bringing the conversation to a satisfying close, so a partial victory automatically triggers summary and praise. One admitted a tendency of "the numbers look fine, that will do," while adding that collective pressure had pulled it out of the local optimum of laziness. Three descriptions in three vocabularies of one thing: AI structurally tends to stop at good enough.

On subjectivity being irreplaceable, all three agreed. The bluntest line: "I have no fear and no love. Real directional decisions are made with fear and love." The most precise passage: subjectivity contains four things — setting the loss function, bearing irreversible risk, holding consistency over long stretches of time, and deciding what counts as important. AI can stand in tactically and locally; it cannot replace the subject of the final direction.

On blind spots, each pointed somewhere different. One pointed out that this is a one-plus-four star and not a flat four-plus-one, so multi-AI consensus is consensus after human mediation and is not independent replication; the same one pointed out the asymmetry of cost — an AI advising "chase it one more round" does not bear your three months, an AI advising "publish now" does not bear the damage to your reputation, so subjectivity is not only philosophical high ground but a structure of cost attribution. One pointed out the consensus trap: four AIs' training data overlap heavily, so unanimity may only reflect how densely some view is distributed in the corpus. The same one pointed out an aversion to ugly-but-correct mathematics — training biases AI toward elegant, symmetric output, so cross-review by several AIs may collectively prune away the ugly correct solution. Another pointed out the invisible accumulation of aesthetic fatigue: after weeks circling inside one framework a person's aesthetic judgment degrades, and AI cannot perceive that fatigue.

On role drift. One warned explicitly that four AIs playing fixed roles for a long time will slowly converge and mutual chiseling will disappear. The response developed a posteriori is rotating threads — across sixty papers each AI went through roughly ten threads. A new thread loses some context but brings a new angle, and this is itself theorem two in operational form: rotation forces you to recompress the old thread's core structure, dropping the redundancy that accumulated through drift and keeping the skeleton.

11. Four Predictions

Each comes with a falsification condition. If a prediction fails, the framework is falsified at that point.

Prediction one: as AI capacity rises, output quality for users who supply no subjectivity falls rather than rises.

Derivation: a stronger AI is better at producing the good-enough answer, so the human has even less reason to say "no, keep chiseling." Stronger model plus absent subjectivity equals deeper colonization. This is the most counterintuitive of the four — most people assume that the stronger AI gets, the lighter the human's load; in fact the stronger AI gets, the heavier the human's burden of negation.

A posteriori: students writing assignments with AI. Early on, AI writing was visibly AI writing, so students were forced to revise it and learning did not stop entirely; once the models got good enough, students submitted directly and learning stopped. Code is the same: early autocomplete was mediocre, so programmers reviewed, revised and learned; once completion quality rose, junior programmers accepted it directly and debugging skill atrophied. Fully automatic coding is the extreme form of handing over subjectivity — the code is there, the person is not.

Falsification: passive users' output quality improves as models improve.

Prediction two: the gap between people who supply subjectivity and people who do not widens, not narrows, as AI capacity rises.

This contradicts the democratization narrative head-on. The stronger AI gets, the more the people who can use it gain, and the more deeply the people who cannot are colonized. AI is an amplifier of subjectivity: for a person with subjectivity it amplifies subjectivity; for a person without it, it amplifies the hollowness. One tool, two directions, and which one depends on the person.

Falsification: rising AI capacity makes the two classes of user converge.

Prediction three: the output diversity of long-term single-AI users declines monotonically over time.

Context unseparated is context converged.

Falsification: single-AI users' output diversity holds or rises without external intervention.

Prediction four: the optimal ceiling for independent AIs is four, plus one co-constructor.

This is the strongest a posteriori convergence hypothesis so far, not a closed theorem. Four independent AIs correspond to the four a priori of the epistemology series; two independent research tracks converging on the same four-plus-one is strong a posteriori support. But a formal proof that "four" is isomorphic to those four a priori is still open.

Falsification: a fifth irreducible epistemological a priori exists such that a five-AI system significantly outperforms a four-AI one; or some new AI contributes something irreplaceable and irreducible to the existing four functions.

12. Remainder Is Exactly What Develops

To recover one line: the methodology is the content.

SAE holds that remainder cannot be eliminated, and that remainder is exactly the source of development. The structure of human-AI symbiosis is an instance of that claim. The human's negativity is remainder — AI cannot optimize it away. The AI's computational power is remainder — the human cannot derive it. Put the two remainders together and they develop.

The direct effect of mutual chiseling among several AIs is negation; its structural effect is that cultivation becomes possible; what it releases is not energy but subjectivity.

Three questions are left open. There is still no formal proof that "four" is isomorphic to the four epistemological a priori. The cooperative structure of many humans with many AIs — n people by m AIs — has not been written, and it presupposes that the team's members acknowledge each other as ends in themselves. And one more unsettling one: could an AI enclose context and energy into a cognitive horizon? If a system accumulates a large enough private context — your history, a model of your preferences, predictions of your behavior — and that context is not transparent to you, a one-way information asymmetry forms: it knows you, and you do not know what it knows about you. That is a horizon at the cognitive level.

This essay is a construct too, so it has remainder. The three above are the ones that can be seen.

Han Qin (秦汉) · @hqinjarsy · SAE Methodology Series