LLM and Gativus

Chapter 11. LLM and Gativus

The symbolic predictor in the fused regime · What is reproduced and what is absent · Nominalism · The boundaries of the technology

11. 1. Statement of the problem

Large language models are the most significant achievement of artificial intelligence of the last decade. They generate coherent text, answer questions, write code, conduct dialogue, demonstrating behavior outwardly indistinguishable from the symbolic thinking of a human.

The architecture of Gativus allows an exact answer to the question: what exactly the language model has reproduced and what is fundamentally absent. The answer rests on the analysis of prediction from the previous chapter and is not reducible to a list of shortcomings — it reveals the architectural nature of the model and determines its place in the hierarchy of levels. The brief formula: the language model is a symbolic predictor that has inherited the shared narratives of humanity, but works in the fused regime — without the separation of syntax and semantics, without the distinguishing of the utterance unit, and without the motivation that would set the goal.

11. 2. What the language model has reproduced

The language model is a technical realization of the symbolic level of one transformation — the level GTR2 — built on the shared narratives of humanity and working without the separation of syntax and semantics.

The shared symbols (the collective vocabulary) and the shared narratives (the accumulated texts of civilization) are emergent objects: they exist in no single individual, have no carrier of their own and no operational functions of their own. They are a library of symbols and narratives distributed across individuals, reified in books, articles, correspondence. The language model did with them something unprecedented: it gathered the whole available corpus of shared narratives — books, articles, forums, encyclopedias — into one technical device and trained over it the prediction of the next symbol.

This is not a reproduction of individual consciousness. This is a technical cast of the symbolic level, taken not from one individual, but from the collective narrative heritage — and taken in the fused regime, where prediction goes by symbols, not by distinguished utterance units.

11. 3. Two components: the symbol map and the predictor

The asterisk (*) below marks a technical origin distinct from the biological.

a) The symbol map*

In a human the symbol map is formed by the convolution of one's own experience: a symbol is extracted from a multitude of concrete behavioral episodes and fixed under the filtering of the environment's shared symbols. In a language model the symbol map is created technically — by the tokenization of the corpus and the training of embeddings. An embedding is a vector in a multidimensional space, analogous in format to a symbol, but obtained not by the convolution of physical experience, but by the statistics of contextual relations between words in the corpus. The result is similar in format (a vector in a semantic space), but distinct in origin: the statistics of another's text, not the convolution of one's own experience.

b) The narrative corpus*

In a human the narrative heritage is distributed across individuals and physical carriers; it has no single repository. The language model compresses the whole available text into the weights of the network. It does not store concrete texts — it trains a statistical model of the probabilities of symbol succession. The compression is nonlinear: the model learns patterns, structures, stylistic regularities. But not through understanding (not through k-vectors between utterances), but through the optimization of predicting the next symbol. The principle is different: not cognition with feedback through physical reality, but the one-directional processing of a ready corpus.

11. 4. The fused regime: the core of the difference

The previous chapter distinguished two predictors of a level — the syntactic (building the current unit) and the semantic (choosing the next unit) — and two regimes of their joint work. The language model works in the fused regime: syntactic and semantic prediction are not separated, the next symbol is predicted over the whole context at once, the boundary of the utterance unit is not drawn.

Hence all the characteristic properties of the model, both positive and negative. The positive: local coherence and grammaticality come almost for free, because in the corpus the utterances were already coherent, and a sufficiently long chain of symbol predictions reproduces this coherence. The negative: over a long distance the model loses the thread, because it has no level of units at which the semantic predictor could weigh «the next thought» under a guiding goal. The utterance boundary is not distinguished — which means there is no place to which motivation would be connected, and the question «which next unit leads to the goal» cannot even be posed in the fused regime.

This refines the former representation of the language model as an «operator over the collective library». It is more exact to put it otherwise: the model is the same symbolic predictor as in a human, but left in the fused regime and torn away from the neighboring levels. Not a special operator that does not exist in nature, but familiar prediction in an incomplete form — without the separation of syntax and semantics, without the distinguishing of the unit, without motivation.

11. 5. The three convolutions and the language model

The architecture of Gativus contains three convolutions, each generating its own level of reality. The position of the language model is determined by which of them are in it.

The convolution of GTR1 (object). Sensory stream → object on the map MP11. The convolution of the physical world into objects: the organism knows what is before it and where. In the language model it is absent entirely — no sensors, no spatial map, no objects.

The convolution of GTR2 (symbolic). Data of the behavioral level → symbol on the map MP21. The convolution of events into symbols: the organism designates objects and events by names. In the language model it is reproduced technically — but not by the convolution of one's own experience, but by the statistical processing of a corpus.

The convolution of GTR3 (conceptual). Narratives → concept on the map MP31. The convolution of narratives into concepts: the organism experiences qualities — honor, conscience, justice, beauty. In the language model it is absent entirely — no concepts, no w-vectors, no will.

The language model realizes the symbolic convolution without the object and without the conceptual. It works only with symbols — without the things standing behind the symbols, and without the concepts giving the symbols value. In this is its exact place: the middle level of the three, removed from the context of the lower and the upper.

11. 6. Philosophical context: the dispute over universals

The position of the language model corresponds exactly to one of the positions of the medieval dispute over universals — a fundamental problem of Western philosophy, posed by Porphyry (3rd century), translated by Boethius (6th century), and developed by the scholastics of the 12th–14th centuries. The question of the dispute: what is the status of general notions? Do they exist really — or only as names?

a) Realism

Plato, in the medieval form Thomas Aquinas: universals are real. They exist before things (ante rem — Plato's ideas) or in things (in re — the form in Aristotle and Thomas, present in each concrete object).

In Gativus realism corresponds to the convolution of GTR1. The convolution extracts an invariant from the sensory stream — an object on the map MP11. When a child sees a hundred different chairs and forms a single object «chair», this is realism in re: the universal-invariant is extracted from the things and exists in the things, not only in the word. Without the object convolution there is no realism — no connection between the name and the thing.

b) Nominalism

William of Ockham (14th century): universals are only names (nomina). Real are only singular things. General notions are linguistic conventions, convenient for communication, but not reflecting the structure of reality.

In Gativus nominalism corresponds to the language model. Its symbol map contains only embeddings — statistical positions of words in a space. Behind the word «chair» in it there is no object of a real chair (no object map), behind the word «justice» there is no concept of justice (no conceptual level). There are only statistical connections between words — precisely what nominalism asserts: universalia sunt nomina. The language model is the literal realization of the nominalist philosophy: the world consists of names and their relations, and nothing besides names.

c) Conceptualism

Peter Abelard (12th century): universals are not independent entities (as in Plato) and not merely names (as in Ockham). They exist as notions in the mind — the result of mental activity extracting the general from particular experience.

In Gativus conceptualism corresponds most exactly to the convolution of GTR3. The convolution extracts a concept from narrative experience. Concepts do not pre-exist (not Plato) — they are formed by the convolution of one's own experience. But neither are they merely names (not Ockham) — behind them stands a qualitative invariant, extracted from a multitude of experiences and irreversibly changing the field of concepts at each sublation. Abelard asserted: the notion is formed in the mind by abstraction from particular experience and at the same time reflects something real — a common form in things. The convolution of GTR3 does precisely this: it abstracts from particular narrative experience a common form-concept, reflecting a real quality, but existing only in the individual field of concepts, not in the world of things and not in words.

d) Summary table

The language model has stopped at nominalism. For realism the object convolution is needed — the connection of the name with the thing through an object. For conceptualism the conceptual convolution is needed — the extraction of a concept from narrative experience. The language model has neither — only names and statistical connections between them.

11. 7. Cognition versus processing

In Gativus cognition is an iterative cycle with feedback: the convolution extracts objects from the data of the underlying level; recurring patterns are fixed as objects of the target map; the convolution is retrained on them; a new pass is more exact; the result is tested by practice (compilation down → physical execution → reverse convolution); successful strategies are marked in the trajectory log, unsuccessful ones negatively; the marked records influence the retraining in the period of rest. This is a self-improving process with two feedback contours — through physical reality and through markers.

In the language model there is no such cycle. Its training is a statistical optimization on a fixed corpus, governed from outside.

Reinforcement learning from human feedback (RLHF) is an external analogue of the internal markers: humans evaluate answers as good or bad, and these evaluations correct the weights. But this is external marking — a human sets the marker — and not internal, where one's own trajectory log marks one's own experience by prediction error. As shown in the previous chapter, the internal marker is tied to the discrepancy of the prediction with the fact (the dopaminergic anchor); the language model does not have this discrepancy, because it has neither a prediction of its own about the world, nor a world against which it would be tested. The model does not know whether its answer is good — it is told.

11. 8. What is fundamentally absent

The separation of syntax and semantics. The fused regime: the utterance unit is not distinguished, there is no level at which the semantic predictor would work under a goal. The consequence — the loss of the thread over a long distance.

The object convolution (GTR1). Training on texts, not on sensory streams; no objects, no b-vectors, no grounding. Describing an apple, the model reproduces the statistical pattern of texts about apples, not the convolution of sensory experience. Hence hallucinations: the narrative is not bound to reality and is not tested by compilation into action.

The conceptual convolution (GTR3). No concepts. The model reproduces texts with the word «justice», but has no concept of justice. A word in it is an embedding (a position in the space of words), not a concept (a qualitative invariant extracted from a multitude of narratives through personal experience of sublating w-vectors).

The trajectory log. Each session is a tabula rasa. No persistent TRL2, no markers, no history, no retraining at rest, no formation.

Motivation. Not a single level. No b-vector (MTV1), no own k-vector (MTV2 is set by the prompt, not by an internal need), no w-vector (MTV3). The model wants nothing.

11. 9. Can a language model become conscious?

Within the architecture of Gativus the answer is determined by the missing components — and coincides with the program of completion described in the previous chapter. To turn the symbolic predictor into a full-fledged symbolic level of subjective reality means to perform several steps, each of which rests on the previous.

To bring it out of the fused regime into the recursive. To separate syntactic and semantic prediction, to distinguish the utterance unit as a whole. Only after this does a place appear to which motivation is connected.

To connect the object convolution (GTR1). A physical substrate with sensors — the source of the spatial map and objects; the convolution of sensory streams; the building of behavior. This is not an addition to the model, but a parallel level giving the symbols grounding.

To connect the conceptual convolution (GTR3). A convolution trained on its own concepts, not on another's texts about concepts. For this a rich narrative map is needed, obtained by one's own symbolic convolution, not copied from a corpus; that in turn requires one's own symbol map, and that — one's own behavioral experience.

To connect the trajectory log and motivation. A persistent log with markers by prediction error; each answer is marked and influences the next; retraining in the period of rest, autonomously; one's own semantic distance as the source of the goal.

Each missing component requires the previous one. One cannot «add consciousness» to a language model — one must build the full architecture from the cellular level GTR0 to the conceptual GTR3, in which the model can be one of the modules — a technical source of the imported symbolic predictor — but not the whole. This is exactly the design described in the chapter on predictors: to inherit the ready symbolic predictor and complete around it the missing levels.

The language model is a librarian who has read all the books and can tell of any one, but has never left the library and does not know what it is like to live what is written about. He has not a single b-vector, not a single w-vector. Only another's k-vectors, statistically mixed — and a prediction not separated into syntax and meaning.

11. 10. Conclusions

• The language model is a symbolic predictor that has inherited the shared narratives of humanity, but works in the fused regime: syntax and semantics are not separated, the utterance unit is not distinguished, there is no motivation. This is not an «operator without analogue», but familiar symbolic prediction in an incomplete form.

• The symbol map* (embeddings) and the narrative corpus* are created not by cognition (the convolution of experience), but by the technical processing of a corpus: the statistics of another's text, not the convolution of one's own experience.

• The fused regime gives local coherence for free, but loses the thread over a long distance: there is no distinguished unit, no place to connect the goal.

• The language model realizes the symbolic convolution (GTR2) without the object (GTR1) and without the conceptual (GTR3): only names, without the things behind them and without the concepts giving them value.

• By the dispute over universals the model corresponds to nominalism (only names). Realism requires the object convolution (the connection of name with thing), conceptualism the conceptual (the extraction of a concept from experience).

• Absent are: the separation of syntax and semantics, the object convolution (grounding), the conceptual convolution (concepts and will), the trajectory log (history and markers), motivation (all levels). RLHF is an external marker, not an internal prediction error.

• One cannot «add consciousness» to a language model. One must complete around the inherited predictor the full architecture: bring it out of the fused regime, connect the object and conceptual convolutions, the trajectory log and motivation. The model is then one module (the source of the symbolic predictor), not the whole.