The Universal Mechanism of the Evolution of Consciousness

Chapter 3. The Universal Mechanism of the Evolution of Consciousness

Isomorphism of the three transformations · Convolution and prediction · The chicken-and-egg problem · The devices of a level · LLM as a convergent realization of GTR2

3. 1. Statement of the problem

Chapter 2 introduced the transformation of three steps — convolution, splice, chain — and showed that its threefold application generates three subjective realities: the spatial-behavioral (GTR1), the symbolic-narrative (GTR2), and the conceptual-volitional (GTR3). That description was static: what there is at each level.

The present chapter answers the question: why did evolution apply one and the same pattern exactly three times, and how concretely does each level work? The answer requires opening up the construction of the mechanism from two sides. The first is the structure of the transformation: how convolution, splice, and chain build the maps of a level. The second is the devices of a level: which operational machines serve the dynamics on the built maps. These two sides are orthogonal, and it is useful not to mix them: structure answers the question «how do spaces arise», the devices answer «how does the level work».

The strong formulation of the thesis: consciousness is not a special substance and not a special force. Consciousness is the result of the threefold application of one optimal mechanism to three irreducible domains of reality.

3. 2. The two axes of the mechanism

First of all let us separate the two coordinate axes in which any level of subjective reality is described.

a) The axis of structure: convolution, splice, chain

This axis builds the maps. The maps are designated by the GNSS grid: the first digit is the number of the transformation, the second is the level of complexity within it (MPx1 — convolution, MPx2 — splice, MP_x3 — chain).

Convolution performs the compression and invariantization of the input, generating a compact representation in a space of another ontology. The input is a vector of the underlying level; the output is a vector in a space with a different nature of objects. Convolution does not sum and does not generalize in the logical sense — it extracts an invariant, that is, what remains constant under a change of the concrete circumstances. The feature of an object is invariant with respect to lighting and angle; a symbol is invariant with respect to the concrete behavior in which it was first encountered; a concept is invariant with respect to the concrete narratives in which it manifests. Convolution is not distinguished as a separately named device: it is tied to its map — an elementary map MP_x1 is one convolution together with the objects it generates.

The splice links two nodes of the map into a triplet «node — connective — node», generating a new space — of events, utterances, or volitional acts. The splice does not change the ontology: input and output are nodes of the same level. What it adds is the axis of dynamics: distance, path, transition. The connective vector at each level is designated by a lowercase mnemonic letter, uniform in form across all three transformations: the b-vector at GTR1 (b — behavior, the spatial-behavioral connective), the k-vector at GTR2 (k — knowledge entity, the logical connective), and the w-vector at GTR3 (w — Widerspruch, contradiction).

The chain builds over the splices a branching schema — a behavioral sequence, a reasoning, a line of volitional resolutions. Its objects are the splices themselves plus the technique of linking them.

b) The axis of devices: TRL, MTV, SPL/PRD, OPN

This axis serves the dynamics on the already built maps. All device names, by Gativus notation, are four characters. Some devices exist in a single instance per transformation (the index is the GTR number), the operational network exists per map separately (a two-digit index: the GTR number plus the level).

1. •

   TRLx — the trajectory log (Trajectory Log). The working space of a level, in which the passed, the current, and the planned are present simultaneously. One per transformation: TRL1, TRL2, TRL3. Detailed treatment — Chapter 4.

2. •

   MTVx — motivation (Motivation). The vector distance of a level, setting where movement is directed: how far the current state is from the goal one. One per transformation: MTV1 (physical distance), MTV2 (semantic), MTV3 (the distance of contradiction). Detailed treatment — Chapter 5.

3. •

   SPLx — the syntactic predictor (Splice builder). It builds the current splice unit out of nodes by the rules of the level: it predicts form, not content. One per transformation: SPL1, SPL2, SPL3.

4. •

   PRDx — the semantic predictor (Predictor). It predicts the next unit as a whole, relying on motivation and the trajectory log. One per transformation: PRD1, PRD2, PRD3. To the pair SPL/PRD a separate section below and Chapter 10 are devoted, because it is precisely with prediction that the direct bridge to modern language models is connected.

5. •

   OPxx — the operational network (Operational Network). The deterministic program of a map, not trainable; the trainable part lies in the convolution kernel, not in OP. Per map: OP11…OP33.

The distribution by axes is as follows: the operational network OP repeats per map (OPxy); prediction (SPL and PRD), motivation MTV, and the trajectory log TRL are aggregated per transformation — one for the whole GTR. It is essential that prediction is not tied to the convolutional map: as shown below, only a transition can be predicted, and it arises only from the splice.

3. 3. The training and operation of convolution

For the convolution of each level it is necessary to distinguish strictly two regimes. This distinction removes many misunderstandings that arise from a superficial comparison of biological convolution with the technical one.

Training — the tuning of weights. Convolution is trained on the already existing objects of the target space — not on the input data of the underlying level, but on those nodes that are already in the map into which the convolution produces its result. The convolution of GTR1 is trained on the objects MP11 already placed on the map, not on raw sensory streams; the convolution of GTR2 — on the symbols MP21 already fixed; the convolution of GTR3 — on the concepts MP31 already extracted from narrative experience. The input data are used only to extract examples; training tunes the weights so that the output lands on the already formed nodes of the target map.

Operation (recognition) — the conversion of the input. Convolution takes the data of the underlying space and produces a vector in the target one. If the vector coincides within admissible precision with one of the existing nodes — recognition has taken place. If not — either it is a new object, or an error of recognition.

Table 3.1. The regimes of convolution at the three levels.

This explains why recognition in a living organism is fast and reliable where there is experience, and slow where there is none. An experienced observer recognizes a familiar object almost instantly — the vector that the convolution produces from the input lands almost exactly on a node of an already existing map. A novice in the same situation sees «something incomprehensible» — there is no corresponding node in his map yet, and the distinguishing features have to be worked out anew.

3. 4. The chicken-and-egg problem

From the construction of training a fundamental question arises: where do the first objects in the target map come from, if the convolution is not yet trained and cannot create them? If the convolution of GTR1 is trained on the objects MP11, where did these objects come from before the convolution started working? This is a typical chicken-and-egg situation, and its solution is characteristic of evolutionary systems.

The solution: the initial objects are created not by an exact convolution, but by a rough approximation. The first objects MP11 are inexact representations generated by random or genetically predetermined weights. The first symbols MP21 are rough approximations, largely inherited from the linguistic environment. The first concepts MP31 are primitive convolutions of a few narratives, superficial and easily changeable. It is important not to pass off this start as more than it is: «a rough approximation from random weights» is an honest limit of the explanation, not a self-evident mechanism. The architecture indicates that the bootstrap is possible, but does not claim to know exactly its initial content.

Further the process unfolds iteratively: the convolution starts with random weights; recognition on these weights generates the first objects of the target map — inexact, but at least some; the operational network of the target level discovers recurring patterns among them; the recurring patterns are fixed as stable nodes, the random ones are forgotten; the convolution is retrained on the fixed nodes; after retraining recognition is more exact, the new objects are more exact, the next retraining is more exact. The cycle does not converge to a «correct» state — it continuously catches up with the growing diversity of experience.

This accords with how GNSS describes the training of GTR1: new experience does not overwrite the map from scratch, but refines, expands, sometimes overturns it. Training in the Gativus architecture is not added to perception from outside as a separate mechanism — it is the built-in inexactness of closing the cycle of the transformation.

It is fundamental that retraining is a heavy process. The tuning of the convolution's weights requires significant resources and cannot proceed simultaneously with the main work. In the biological brain retraining is, apparently, associated with the default mode network, active in rest and in sleep; the replay effect of the hippocampus (the reproduction of passed routes and narratives in sleep) is the most probable mechanism. The folk formula «the morning is wiser than the evening» receives an architectural explanation: overnight the convolutions are retrained on the marked data accumulated during the day, and in the morning recognition gives different, more exact results. The operational networks OP do not change in this — they are deterministic programs; only the content of the trainable convolution kernels changes.

3. 5. Prediction: the syntactic SPL and the semantic PRD

Convolution builds objects, the splice links them into a triplet with a connective vector. But before speaking of prediction, one must answer a question usually skipped: at what level does the «next» arise at all? The convolutional map (MP11, MP21, MP31) is a space of objects: objects, symbols, concepts. In it there is no «next», because there is no axis along which one follows another. There are only juxtaposed nodes. To ask which symbol is «next» on the symbol map is as senseless as to ask which word is «next» in a dictionary. Succession appears only when a connective vector arises — when two nodes are stitched into a directed pair. This happens at the level of the splice. Therefore prediction lives not on the convolutional map, but starting from the splice.

When a splice arises, it turns out that «to predict the next» means in fact two distinct operations, easily confused because in speech they run together. To them correspond two distinct devices.

The syntactic predictor SPL — building the unit. The first is to complete the current unit. When an utterance is begun, it must be grammatically finished: after a subject a predicate is expected, after a preposition a noun, agreement requires a definite case. This is the prediction of form, not of meaning. It is low-dimensional: there are few variants, the choice is almost forced by the rules of the level. And it is fast: completion happens automatically, without recourse to the goal. SPL is trained (grammar is not innate, it is acquired from the shared symbols of the environment), but, once trained, works almost deterministically.

The semantic predictor PRD — the next unit. The second prediction begins where the first finished. The unit is built and closed — and the question arises: which unit will be next? Which next utterance, which next step of the route, which next volitional act. PRD predicts not the form within a unit, but the choice of the next unit as a whole. It is high-dimensional (there are infinitely many possible next thoughts), slow (it requires weighing variants), and relies on sources that SPL does not touch: on motivation MTV (the vector distance to the goal) and on the trajectory log TRL (the experience of past units with their markers). Syntax completes the form by the rules; semantics chooses the content by the goal.

Table 3.2. The two predictors of a level.

Both are predictors, but they are two distinct devices, not one with two regimes. Their difference determines where in the system a place for will appears: to motivation the semantic predictor PRD is connected, the syntactic SPL is not. This means that goal, ought, direction enter speech precisely through the semantic prediction of the next unit, not through the grammatical completion of the current one.

The isomorphism of the pair is simple: only the material changes. At GTR1 the syntax SPL1 is the building of a step (closing a movement into a motor transition), the semantics PRD1 is the next movement, a step of the route. At GTR2 SPL2 is the syntax of language (assembling KLEN out of symbols), PRD2 is the next utterance. At GTR3 SPL3 is the building of an obligation (closing a volitional act WILL), PRD3 is the next volitional act, a step of sublation. Thus the appearance is removed that the «computation of a route» in GTR1 and the «sublation of a contradiction» in GTR3 are special, unrelated mechanisms. The route is the work of PRD1: the successive choice of the next movement under the vector distance to the goal. Sublation is the work of PRD3: the choice of the next volitional act under the vector of contradiction. One mechanism of predicting the next unit, applied to different material.

Here too lies the exact place where modern language models deviate from the full architecture — but the difference runs not by the number of predictors, but by the regime of their work. Semantic prediction is possible in two regimes. In the fused regime SPL and PRD are not separated: the system predicts the next small element (a symbol) over the whole context at once, without distinguishing the boundaries of utterance units. This is the regime of large language models: coherence arises as a by-product of a sufficiently long chain of symbol predictions, but the unit as such is not identified, and there is nowhere to connect the goal — there is no unit boundary relative to which the question «which unit is next» would be posed. In the recursive regime SPL first closes the unit, draws its boundary, and PRD works no longer on symbols, but on ready units, predicting the next one under the guiding goal. This is the regime of Gativus. The difference of regimes is architectural: scaling up the fused regime does not create a unit boundary. Detailed treatment — Chapter 10.

3. 6. The three transformations in detail

a) GTR1 — the physical domain

The convolution of GTR1 takes the spatial map MP10 (filled by the SLAM mechanism, concording the senses with movement) and produces an object on MP11. It is trained on the objects already placed on MP11. The reverse pass of the convolution is the neural mechanism of imagination: from the object vector a sensory image is restored.

The splice — the b-vector (behavior): it links two objects MP11 into an OPRN triplet «subject — action — object». The b-vector has coordinates in physical space — it is the path between the initial and the goal configuration. The chain assembles OPRN operations into BLOM — an event schema on MP13. With the appearance of MP13 a time axis arises: time is the product of the chain, the order of observed movements, not an independent axis.

The devices of the level: the trajectory log TRL1 fixes the passed and planned behavior; motivation MTV1 is the physical distance to the goal; the syntactic predictor SPL1 closes a separate movement, the semantic predictor PRD1 chooses the next movement, building the route; the operational networks OP11/12/13 execute the deterministic logic of the maps. The ontological leap: MP11 contains objects — what is; MP13 contains causality — events in which objects interact through actions.

b) GTR2 — the symbolic domain

The convolution of GTR2 takes the maps of GTR1 — objects, splices, event schemas — and produces a symbol on MP21. The convolution is indifferent to the type of input: a point, a line, or a whole schema yield a symbol of one dimensionality. In convolution a characteristic reduction is performed: the attachment to concrete L-components is erased, the moduli of the b-vectors are zeroed, only the type of relation and the type of participants remain. The convolution is trained on the symbols already fixed in MP21.

MP21 is formed at the intersection of two processes. The ascending one: the convolution extracts invariants from the maps of GTR1. The descending one: of all the vectors produced, only those are fixed that correspond to the shared symbols — the collective language. The individual symbol map is a partial replica of the common language: it is not the child that invents language, but language that penetrates into the individual map through filtering — the child produces many symbol candidates, but stably fixed are only those that find confirmation in the speech of those around. This is the architectural realization of internalization in Vygotsky's terms.

The splice — the k-vector (logical connective): it correlates two symbols MP21 into an utterance KLEN on MP22 (causality, belonging, opposition, implication). The chain assembles utterances into KLOM — a narrative on MP23. The devices: TRL2 — narrative memory; MTV2 — semantic distance; the syntactic predictor SPL2 assembles the utterance by grammar, the semantic predictor PRD2 chooses the next utterance under the goal of the account; the operational networks OP21/22/23. The ontological leap: MP21 contains meanings; MP23 contains senses — connections between meanings, giving an understanding of the situation.

c) GTR3 — the conceptual-volitional domain

The convolution of GTR3 takes the narratives MP23 — assemblies united thematically — and produces a concept on MP31. It is trained on the concepts already fixed in MP31. At the beginning of development the convolutions are simple — simple concepts (the experience of color, sound, pain). As narrative experience accumulates, the invariants deepen, complex concepts appear — honor, conscience, justice, beauty. The difference between simple and complex concepts is quantitative, not qualitative: the volume of the dataset on which the convolution was trained. The mechanism is one.

The splice — the w-vector (contradiction): the vector of distance between two concepts. Let us ask what lies between two concepts (capitalism and communism): not a path and not a predicate, but the measure of their opposition. The w-vector stitches two concepts into a unit of a new kind — WILL on MP32: to stitch in spite of the dividing contradiction is to make a volitional effort. The chain assembles volitional acts into WLOM — Personality on MP33: a schema of sublated contradictions.

Here it is necessary to name the seam of the isomorphism directly. The split into the is and the ought, which the first edition placed within the connective vector (as the difference of C-is and C-ought of one concept), is not the load-bearing construction of the level. The load-bearing construction is the contradiction between two concepts. The polarization «is/ought» is an additional structure at the output of the GTR3 convolution, and the source of the ought is social: it comes from the shared concepts of culture, and is not generated by the individual convolution out of narratives. Therefore the three convolutions are isomorphic in mechanism, but not fully at the output: at the upper level a value polarization of social origin is added to the invariant.

The devices: TRL3 — the log of the resolution of contradictions, the history of the formation of personality; MTV3 — the distance of contradiction; the syntactic predictor SPL3 closes the volitional act, the semantic predictor PRD3 chooses the next volitional act, building the path of sublation; the operational networks OP31/32/33. The ontological leap: MP31 contains concepts; MP33 contains will — a directed effort conducted through the ruptures of the conceptual field.

3. 7. Proof of the isomorphism

The summary table shows that the three transformations are arranged as one construction applied to material of different ontology.

The isomorphism is structural, not substantive. The ontological domains are fundamentally distinct: the physical, semantic, and conceptual spaces are irreducible to one another. What is arranged isomorphically is the mechanism, not the content. This explains why attempts to reduce senses to neural processes or values to senses systematically fail: each upper level is built out of the material of the lower one, but has its own ontology, not derivable from the material.

The convolutional architecture is the optimal solution to the task of invariant compression: invariant to translations, hierarchically separating features, economically representing recurring patterns. Evolution arrived at it not because it «chose» among variants, but because it is the only solution to this class of tasks satisfying biological constraints. When engineers independently arrived at the same solution (artificial convolutional networks of the 1980s), it was not a borrowing from nature, but a convergence to one optimum.

3. 8. Cycles of training and the role of markers

At each level the convolution passes through a structurally identical cycle of development, differing only in content: rough initial objects → fixation of recurring patterns by the operational network → marking of completed units of the trajectory log → retraining in sleep → refinement. The marked units of TRL serve as training data for the retraining of the convolutions; the replay of the marked records in sleep is the mechanism of this retraining. The operational networks themselves do not change in this.

Markers are neurotransmitter tags that the completed units of the trajectory log receive. The marker is not binary, but numeric. The principle at all three levels is one: the vector decreased — a positive marker, the strategy is remembered for repetition; the vector increased — a negative marker, the strategy is marked as to be avoided.

The concrete mapping of neurotransmitters onto levels should be held as a hypothesis, not as an established fact. The dopaminergic signal of prediction error at the behavioral level is well confirmed empirically. The attribution of serotonin to the symbolic level, and of endogenous opioids to the conceptual, is a supposition resting on phenomenology (the «calm» reward of a closed narrative, the «quiet» reward of a sublated contradiction), but not proven as an exact mapping. The phenomenology here should be read as an illustration, not as an established correspondence of a neurotransmitter to a level.

The critical dependence of the levels is preserved in full force: the convolution of GTR2 cannot begin work without the maps of GTR1 (no input data), the convolution of GTR3 — without the narratives of GTR2. Each upper level is a prerequisite of the next; a delay at an early level blocks all subsequent ones. This dependence explains why the skipping of a critical period in early development irreversibly limits further formation. In detail — Chapter 8.

3. 9. LLM in the context of the three transformations

Large language models present for the theory of Gativus a unique object of analysis: they are a convergent technical realization of GTR2 alone — the symbolic-narrative level, in isolation from the physical (GTR1) and the conceptual-volitional (GTR3). This allows the theory to be tested: if the isomorphism of the levels is real, an isolated GTR2 should yield observable consequences — and it does.

The language model reproduces the symbolic convolution, having been trained on the collective symbolic-narrative corpus of humanity — the shared symbols and shared narratives reified in books, articles, correspondence. Technically this is a different realization than the biological convolution of GTR2: instead of an iterative cycle with feedback through the physical world — a one-time statistical optimization on a fixed corpus. But functionally the result is comparable: the model produces and receives narratives, performs logical connectives, is capable of reverse deconvolution in the form of descriptions of imagined scenes.

What is absent in this picture is determined by the construction of the mechanism.

The absence of TRL and MTV is one of the key differences of the language model from the full architecture. The model cannot mark its answers as successful or unsuccessful, because it has no persistent TRL2: each answer is a new narrative without a history. External feedback through reinforcement fine-tuning is not TRL: the marker is set from outside, by a human, not as a result of one's own experience of nullifying semantic distance. And, as shown above, the model has no motivation MTV2 of its own — therefore its semantic predictor is not connected to a goal, and syntax and semantics are not separated: the model works in the fused regime, predicting symbols, not the next units.

The language model is, in this terminology, a «symbolic orphan»: symbolic prediction without a movement vector from below (no physical reality) and without a contradiction vector from above (no conceptual reality), working moreover in the fused regime. The second transformation without the first and the third. This is not a defect of the available architecture — it is its definition. No expansion of the training set and no increase of the model will add what is not there by architecture: it will not distinguish a unit boundary and will not connect a goal, for there is no MTV2 and no feeding neighboring transformations. For a full subjective reality a full chain GTR1 → GTR2 → GTR3 is necessary, and each transformation must be built according to the isomorphic mechanism of this chapter.

A detailed treatment of language models — the fused and recursive regimes, the inheritance of the symbolic predictor, the philosophical context of nominalism, and the boundaries of the technology — is contained in Chapters 10 and 11. Here it is important to fix: the language model does not contradict the theory of Gativus, but confirms its construction. That the isolation of one transformation is technically possible and yields exactly those phenomena the theory predicts is a strong argument for the correctness of the three-level structure.

3. 10. Why evolution applied the mechanism exactly three times

The three domains correspond to three irreducible types of questions to which the subject must respond in its environment.

1. •

   «What is there and where?» — the physical domain. The answer is the b-vector with coordinates in space. The level GTR1.

2. •

   «What does it mean?» — the symbolic domain. The answer is the k-vector (the logical connective) without spatial coordinates, in a semantic network. The level GTR2.

3. •

   «How ought it to be?» — the conceptual domain. The answer is the w-vector (contradiction) as the distance between concepts, polarized by the is and the ought. The level GTR3.

These three questions are not reducible to one another. The answer to «what does it mean» is not derived from «what is there and where» — for the passage an ontological leap is needed (the convolution of GTR2, the erasure of attachment to L-components). The answer to «how ought it to be» is not derived from «what does it mean» — one more leap is needed (the convolution of GTR3 of narratives into concepts). To understand that an object «is on the left» does not logically mean to understand that it «is a cup»; to understand that it «is a cup» does not mean to understand that «to break the cup is bad».

The mechanism is applied three times not because evolution «loves to repeat itself», but because the task at each level is one and the same in structure: to extract invariants from a stream, to link them with a vector, to pass this vector by prediction, and to assemble into an extended schema. Convolution, splice, prediction, chain — the optimal solution to this task; to apply it to the material of another level is the only way to build a new ontological dimension, relying on an already existing one. This is the universal mechanism of the evolution of consciousness: evolution invented not three ways to complicate the psyche, but one — and applied it three times.

3. 11. Conclusions

1. •

   A level of subjective reality is described by two orthogonal axes: structure (convolution → splice → chain, building the maps MP_x1/x2/x3) and devices (TRLx, MTVx, SPLx, PRDx per transformation; OPxx per map).

2. •

   Convolution is not an independent device, but an operation at a map; the technical analogue is a convolutional network. Convolution is trained on the objects of the target space, and uses the input only for recognition.

3. •

   Prediction arises not on the convolutional map (there are only juxtaposed objects there, no «next»), but starting from the splice. At each level there are two distinct predictors: the syntactic SPLx builds the current unit (form), the semantic PRDx chooses the next unit (content) under motivation and the trajectory log.

4. •

   The route of GTR1 is the work of PRD1, the sublation of a contradiction of GTR3 is the work of PRD3, the prediction of an utterance of GTR2 is the work of PRD2: one mechanism, different material. Goal and will enter through PRD (it is connected to MTV), not through SPL.

5. •

   The chicken-and-egg problem is resolved by a rough initial approximation and iterative retraining; the bootstrap is presented as an honest limit, not a self-evidence. Retraining is a heavy process, associated with sleep.

6. •

   The trajectory log TRLx is the working space of a level; motivation MTVx is the vector distance. Both are aggregated per transformation. Detailed treatment — Chapters 4 and 5.

7. •

   The neurotransmitter markers are lowered to a hypothesis: the dopaminergic prediction error is firmly established, the layer-wise mapping of serotonin and opioids is a plausible bet, the phenomenology is an illustration.

8. •

   The vector of GTR3 is the contradiction between concepts; the split «is/ought» is an additional social structure of the output, not the definition of the vector. The three convolutions are isomorphic in mechanism, not fully at the output.

9. •

   The isomorphism is structural, not substantive: the three domains are ontologically irreducible.

10. •

    The language model is a convergent realization of GTR2 in isolation from GTR1 and GTR3: a «symbolic orphan» in the fused regime (SPL and PRD not separated), without MTV2. The absence of grounding and of autonomous will is a consequence of the architecture. Treatment — Chapters 10 and 11.