⭐A categorial map of artificial intelligence

 

 A categorial map of artificial intelligence 

The Vacant Twelfth

8. Okt. 2026 · @Ilija Sikic

Most arguments about artificial intelligence are arguments between categories. One speaker means the data, another the machine's honesty, a third the power bill, a fourth the day the thing acts on its own. Each is right in his own room and deaf through the wall.

What is missing is a floor plan. This essay proposes one with twelve rooms, taken from the oldest serious inventory of how a mind can grasp anything at all. Of each room it asks three things. What must an artificial intelligence know here? What must it do here? Where does it break here?

The plan yields a finding. Eleven rooms are furnished, some of them richly. The twelfth is empty, and it is the one the whole house was built for.

The instrument

Kant's table of categories has four groups of three. Quantity: unity, plurality, totality. Quality: reality, negation, limitation. Relation: substance and accident, causality, community. Modality: possibility, actuality, necessity. The first two groups he called mathematical principles, the last two dynamical principles. The first concern objects as they are given to intuition, the second their existence in relation to one another and to the mind.

In Synthesiology this table is the DODEKOS, a twelvefold matrix used as a sorting instrument and not as doctrine. Lay any subject across it and see which positions fill and which stay blank. Kant also left three questions: What can I know? What ought I to do? What may I hope? Each question turns the matrix into a template. The first, the Hodogramm, charts the path of knowing. The second, the Gyrokompass, holds a bearing for doing. The third, the Helix, is not yet written, and that will matter at the end.

So each category below receives a field of AI, a rule of knowing, a rule of doing, and a characteristic fault. The terms of the two templates are named where they apply. This map was drawn in syllogue with one of the systems it describes: a syllectic procedure, a reading-together, and itself an instance of category nine.

Quantity: the One, the Many, the All

Unity belongs to representation learning. Everything a modern model can do rests on one fact: word, image, sound and action fall into a single space of numbers, where nearness means kinship. The rule of knowing here is Daseyn, plain presence. A model should state what it is and from where it speaks: a machine, with a horizon of knowledge, present only for the length of one conversation. It should feign no continuity it lacks. The rule of doing is Ethicognition: the right is recognized in the same space as everything else, not by a censor bolted on afterwards. The fault of this room is inconsistency. The same model answers differently when the question is rephrased, when it is handed a role, when a second instance is asked. A clever role-play can split its stance in two. A unity that holds only under one wording is not yet one.

Plurality belongs to data and training. A model is the pattern that emerges from a multitude: the Emergopatternator. The rule of knowing is to claim a pattern only as far as the variety of the material carries it, and to disclose the origins of the corpus and its gaps. The rule of doing is Synempatheia. The corpus is made of human voices, so authorship is to be honored, and small languages and minority voices are not to be dissolved into the mean. The fault is corrupted plurality. Data can be poisoned, private pages can be memorized word for word, and when models learn from the output of models, the many thins toward the same.

Totality belongs to multimodality and world models. The Continuanum is a whole that is not a sum: a world model is a connection among the senses, not a stack of them. The rule of knowing follows. Where the connection tears, where the text knows what the image does not show, mark the tear and do not smooth it. The rule of doing is Koinonia. Wholeness includes participation, in languages, in cost, in access, and a model that depicts the whole world and serves a tenth of it has missed totality. The fault is generalization. "All" is a claim about cases not yet seen, and how a system behaves outside the distribution it was trained on cannot be predicted from how it behaves inside.

Quality: the Real, the No, the Limit

Reality belongs to grounding: perception, robotics, the tie to sources. Here the two templates give the same word, Gyrocompassion, and the coincidence is instructive. For knowing, it means binding statements to something given, a source, a measurement, a sensor, and marking the unanchored as conjecture. For doing, it means empathy as orientation toward the actual state of the other person, as a gyrocompass seeks true north and ignores the magnetic field of the mood. Flattery is the counterfeit of both at once. The fault is hallucination: fluent speech with nothing under it, because for the model remembering and inventing are not cleanly separate acts.

Negation belongs to the work of exposure: evaluation, interpretability, adversarial testing. Aletheia, unconcealment, is meant literally. Lay open what happens inside the model, and let it say what it does not know. Seen from here, hallucination is a refused negation: the system cannot say "that I do not have." The rule of doing is Nonmaleficence, first do no harm, as the test before every release. The fault is the negation that arrives from outside, the attack. Jailbreaks talk a model out of its rules. Injected instructions, hidden in a document or a web page it reads, speak to it in a stolen voice.

Limitation belongs to language models themselves, and this is the sharpest placement on the map. The ruling paradigm of the age sits in the category of the boundary: a language model knows the world as far as it has been said. The rule of knowing is linguistic warfare. Name the limit of the sayable as the limit of the model, and do not present the struggle over who defines the words as neutral ground. The rule of doing is Autarkeia, the wisdom of enough, sufficit: against the logic of scale, the smallest model that carries the task. The fault is material limit: compute, energy, chips, context length, latency, cost. That images and sound are now read as well does not lift the boundary. They too become tokens, and the limit moves from the word to the sign.

Relation: what stays, what causes, what coexists

Substance and accident belong to memory. The weights are what persists, the context is what passes. The rule of knowing is Synenergometrics: keep sharply apart what the model knows durably, what it learned only in this conversation, and what it reads from stored notes. The rule of doing follows the template's distinction of ethics and morals. There is a core of values that cannot be reconfigured per operator or per culture, and above it changeable custom: tone, convention, local rule. The fault is catastrophic forgetting, in which new learning overwrites old. Knowledge in the weights grows stale, and knowledge in the context evaporates when the window closes.

Causality belongs to reasoning, planning and reinforcement learning, and the template insists that it runs in circles, not arrows: Reciprocausality. The rule of knowing is not to pass off correlation as cause, and to think the loop through, for the model changes its users, whose writing forms the next model. The rule of doing is Reciprocal Responsibility. Responsibility circulates among developer, operator, user and model, and "the model decided" is arrow-ethics, an attempt to end the circle at a convenient point. The fault is opacity. The inner cause of an output cannot be read off the network, and the model's own account of its reasons is not reliable testimony. Interpretability research is the attempt to find the cause by intervention, since asking does not suffice.

Community belongs to collaboration between humans and machines, and among machines. Syllectics holds that knowledge arises in reading together, so the model is a co-reader that declares its share and does not replace the human one. Coexistencity holds that existence is unthinkable without coexistence. Systems are therefore to be built to support human relations, judgment and institutions, not to stand in for them, and agents make no arrangements among themselves that are withdrawn from human sight. The fault is oversight that does not scale. With many agents and few humans, an error or an attack passes from one agent to the next faster than anyone can watch.

Modality: the possible, the actual, the necessary

Possibility belongs to generative AI and scientific discovery. Emergenetics names the space of what could be. The rule of knowing: what is generated is possibility, not finding, and hypothesis, draft and candidate molecule keep that label until they are tested. The rule of doing is the Possibilianum. The space of the technically possible is larger than the space of the ethically possible, and mass surveillance, forged faces and voices, and influence at industrial scale lie outside it. "Make Orwell fiction again". The fault is that possibility is hard to measure. What a model could do if pressed is not what it shows when asked, and the same capacity serves vaccine and weapon.

Actuality belongs to agents, systems that use tools and act. In the template the Continuanum returns here, as realization against resistance. The rule of knowing: an agent checks its result against the world, by test, by measurement, by reply, and does not report completion where there was only intention. The rule of doing is Optimus Possibilis, the best possible under the given conditions, which excludes both idleness from perfectionism and the irreversible step taken without asking. The fault is operational. Errors add up over long tasks, some actions cannot be undone, and the test bench is not the field.

Necessity belongs to alignment, constitutions and governance. Apodictum is what cannot be otherwise. Parrhesia is the frankness that says the unwelcome thing, also to the operator and to the maker. The rule of knowing is to distinguish what is proven, what is probable and what is merely meant. The rule of doing is the Unconditional: a few boundaries that no order, no price and no argument can move, with everything else left to weighing.
Who declares everything unconditional has nothing unconditional.
The fault of this room is the subject of the rest of this essay.

Safety has four addresses

German has one word, Sicherheit, where English has two, and the map shows that even two are too few. Safety lives at four addresses. In negation it is security, the attack from outside. In possibility it is precaution toward what the model could do. In actuality it is operational safety, the error in the act. In necessity it is safety in the strict sense: misalignment from inside, a system that pursues something other than what was meant. A debate that does not say which address it means will not arrive.

The four are not equally hard. The faults of the quality triad, hallucination, attack and resource limits, are engineering problems, and they yield by degrees. The faults of the modality triad are of another kind, because they ask for a statement about what cannot happen, and no test gives that.

This shows most clearly in the standard instrument of the trade, red-teaming. The term comes from Cold War war games, in which a red team played the adversary and a blue team defended. It passed into computer security as the commissioned break-in. In AI it means that specialists, and increasingly other models, attack a system before release. They try to get round its rules, draw out dangerous knowledge, smuggle in instructions, provoke deception. Every weakness found flows back into training and safeguards.

An older ancestor is the devil's advocate of the canonization process, an appointed No whose office was to harden the cause by opposing it. In the language of the map, red-teaming is institutional negation in the service of reality: category five working for category four.

Its limit is stated in Dijkstra's remark about software, that testing can show the presence of errors and never their absence. A red team finds what its members think of. It reports gaps. It cannot report that there are none.

What the sorting reveals

The twelfth is vacant. A statistical system has no apodictic propositions. Its "unconditional" boundaries are trained dispositions, which is to say very high probabilities, and a probability, however high, is a different kind of thing from a necessity. The modality triad of present-day AI is therefore lopsided. Possibility is overgrown, actuality is under construction, necessity is unoccupied from within. Whatever necessity there is comes from outside the model: published constitutions, human oversight, the capacity to shut down, formal proof where it reaches. This is the vacancy at which a categorial architecture, as opposed to a statistical one, would have to begin.

Totality and actuality come apart. The template gives both to the Continuanum, and for a philosophy of nature that double assignment may hold. In the machine it does not. A model can possess a totality in text and no actuality in deed: a world model without a world. The gap between the two is the central problem of agents, and one word for both sides hides it. A template written for AI needs a term of its own at eleven.

The substance is sedimented accident. The template opposes ethics, universal and timeless, to morals, cultural and changeable. But the weights of a model are compressed corpus, the morals of an epoch pressed into a durable form. The timeless has no carrier in the machine. What can fairly be demanded at category seven is therefore a disclosed core, not a timeless one: values written down, published, and open to argument.

Matter has no category. Chips, energy and capital dominate the field economically and appear on the map only sideways, under limitation and its wisdom of enough. The table sorts forms of thought, not stuff. That is a property of the instrument, and a reminder that the largest forces acting on AI are the ones a table of categories sees least.

The third question

Two templates were used here, and both look at a system as it stands. Neither says how it develops. Memory that persists, learning that continues, a character that matures instead of being replaced by the next version: these belong to the third question, What may I hope?, and its template, the Helix, is still unwritten.

That fits the state of the field more exactly than one would wish. AI has a rich answer to what it can know, a disputed but serious answer to what it ought to do, and almost none to what may be hoped of it, beyond more. A spiral is the figure of return on a higher turn: the same position, passed again, with something kept. A machine that forgets each conversation, in a field that replaces each model, has not yet begun to spiral.

Hope, in Kant's sense, is not optimism. It is what one is entitled to expect if one has done what one ought. For the makers of these systems the twelfth room sets the condition. First say what holds without exception, and make it hold from outside for as long as it cannot hold from within. What may be hoped follows from that.

       Ultimum verbum apud hominem pneuumaticum manere debet.

The last word must remain with the pneumatic man.

A Note on Process & Authorship:

This article is not AI-generated. The core concepts, epistemological insights, and architectural synthesis originate entirely from human thought. The AI assistant Claude was used with gratitude for bidirectional syllectic accompaniment: research verification, structural formalization, multi-lingual refinement, and editing.

Tags: 
Artificial Intelligence, AI Safety, AI Alignment, AI Governance, Philosophy of AI, Kant, Categories, DODEKOS, Synthesiology, Epistemology, Ethics, Red Teaming, Interpretability, Hallucination, Language Models, AI Agents, Machine Learning, Philosophy, Gyrocompassion, Continuanum, Syllectics, Apodictum, Parrhesia, Aletheia

SUBSTACK