Predictors

Chapter 10. Predictors

Where prediction lives · The syntactic and semantic predictors · Two regimes · The inheritance of the symbolic predictor

10. 1. Where prediction arises

The chapter on the universal mechanism named the predictor among the functions serving the dynamics of a level, and placed it in the general schema. The present chapter treats it in detail — because it is precisely with prediction that the most direct bridge between the theory of Gativus and existing technology is connected, and because the inner construction of prediction turns out to be subtler than it seems at first glance.

Let us begin with 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 symbol», because there is no axis along which one follows another. There are only symbols — juxtaposed, outside any sequence. 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. Only here is there a «from what» and a «to what», and hence something to predict. Therefore prediction lives not on the convolutional map, but starting from the splice: at the level of building units and at the level of their chains. This is the first clarification, and it abolishes the naive picture of a «predictor of the next symbol» as an independent capacity of the symbolic dictionary.

Only a transition can be predicted. Where there is no connective vector, there is no next — there is only the juxtaposed.

10. 2. Two predictions, not one

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.

a) Syntactic prediction: building the splice

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 a prediction — but a prediction of form, not of meaning. It is low-dimensional: there are few variants, the choice is almost forced by the rules of language. And it is fast: the completion of the construction happens automatically, without recourse to a goal or motivation.

Let us call this device the syntactic predictor — it is responsible for building the splice, for assembling the current unit out of nodes by the rules of the level. Isomorphically it works at all three transformations, and its content at each is its own:

• At GTR1 — the building of a step: completing an elementary operation of movement, closing the b-vector into a concrete motor transition.

• At GTR2 — syntax: completing the utterance by the grammar of the language, assembling KLEN out of symbols.

• At GTR3 — the building of an obligation: completing the volitional act, closing the w-vector into a concrete «ought».

The syntactic predictor is trained — grammar is not innate, it is acquired from the shared symbols of the environment (the critical period for syntax is real: a language not acquired in time is not fully completed). But, once trained, it works almost deterministically: the rules of assembling a unit are stable, and they are applied quickly.

b) Semantic prediction: the next unit

The second prediction begins where the first finished. The unit is built and closed — and now the genuinely interesting question arises: which unit will be next? Which next utterance in a reasoning, which next step of a route, which next volitional act. This is the semantic predictor — it predicts not the form within a unit, but the choice of the next unit as a whole.

Semantic prediction is opposite to syntactic in all parameters. It is high-dimensional: there can be infinitely many next thoughts. It is slow: it requires weighing variants. And, above all, it relies on sources that the syntactic predictor does not touch — on motivation (the vector distance to the goal, setting for the sake of what the account is conducted) and on the trajectory log (the experience of past units with their markers). Syntax completes the form by the rules; semantics chooses the content by the goal.

Table 10.1. The two predictors of a level.

Both are predictors, but they are two distinct predictors, not one with two regimes. Their difference is not a subtlety of terminology: it determines where in the system a place for will appears. The semantic predictor is connected to motivation, the syntactic 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.

10. 3. Two regimes of semantic prediction

Having distinguished the two predictors, we can pose the decisive question: how do they work together? Here two regimes are possible, and it is precisely their difference that separates the architecture of Gativus from existing language models.

a) The fused regime

In the first regime syntactic and semantic prediction are not separated. The system predicts the next element over the whole accumulated context at once, without distinguishing the boundaries of units. «To complete the utterance» and «to choose the next thought» merge into one stream of predicting small elements — symbols. The boundary between utterance units is not drawn: the system does not know where one KLEN ends and another begins, because it works below the level of the unit, at the level of the symbol.

This is the regime of existing large language models. They predict the next symbol by the context, and the coherence of utterances arises in them as a by-product: a sufficiently long chain of symbol predictions looks like a meaningful utterance, because in the training corpus the utterances were already meaningful. But the unit as such is not identified. Hence a characteristic trait: the model is locally grammatical and coherent, but over a long distance loses the thread — because it has no level of units at which the semantic predictor could weigh «the next thought» under a guiding goal. There is simply nowhere to connect the goal: there is no unit boundary relative to which the question «which unit is next» would be posed.

b) The recursive regime

In the second regime the two predictors are separated and work recursively. First the syntactic predictor assembles and closes the unit — distinguishes the utterance as a whole, draws its boundary. Then the semantic predictor works no longer on symbols, but on these ready units: it predicts the next utterance as a whole, relying on motivation and the trajectory log. Prediction goes on two levels: the fast syntactic — within a unit, the slow semantic — between units.

This is the regime of Gativus. Its advantage follows directly from the separation: as soon as the unit is distinguished, a place appears to which motivation is connected. The semantic predictor asks not «which next symbol», but «which next utterance leads to the goal» — and for this it needs the vector distance setting the goal. The recursive regime thereby returns to prediction what the fused regime loses: directedness. The system does not merely continue plausible text — it builds the next unit for the sake of nullifying the vector.

The difference of the two regimes is not a question of the quality of realization, but a question of architecture. However much one scales up the fused regime, the unit boundary will not appear in it: it is arranged so as to work below the level of the unit. The recursive regime requires a different organization — the distinguishing of units as independent objects, over which a separate semantic predictor works. This is the architectural step that Gativus makes in comparison with a language model.

10. 4. The isomorphism of prediction at the three levels

The separation into a syntactic and a semantic predictor is isomorphic for all three transformations — only the material changes.

This removes the appearance that the «computation of a route» in space and the «resolution of a contradiction» in thinking are special, unrelated capacities. The route is the work of the semantic predictor of the physical level: the successive choice of the next movement under the vector distance to the goal. The resolution of a contradiction is the work of the semantic predictor of the conceptual level: the choice of the next volitional act under the vector of contradiction. One mechanism of predicting the next unit, applied to different material. And within each step — the fast syntactic predictor, assembling the unit itself.

10. 5. The symbolic predictor is already built

Here is the central practical thesis of the chapter. The semantic predictor of the symbolic level together with the trained symbol map is already built — by humanity, in the form of large language models trained on the collective symbolic-narrative corpus. Open models provide the result of this training freely. It does not need to be reproduced anew.

It is worth realizing the scale of what is available here. The training of the symbolic level requires a pass through the corpus accumulated by civilization — all the shared symbols and shared narratives reified in texts. This is precisely the resource that in the Gativus architecture feeds the symbolic convolution and prediction. Language models have already performed this pass; their weights are the crystallized result of training on the shared narratives of humanity. From the point of view of Gativus an open model is a ready symbolic predictor lying in free access.

Therefore the first Gativus node does not begin the symbolic level with an empty map. It inherits the open weights as the initial state of the symbol map and the predictor, receiving at once a rich vocabulary of symbols and a working prediction. What culture spent millennia on, and the language-model industry years and enormous computational resources, is obtained as a starting asset. This turns the architecture from a purely theoretical construction into a feasible design: the component most expensive to train already exists.

It is important to understand correctly the status of what is inherited. An open model brings not a «ready mind», but the symbolic level of one transformation — the prediction and the symbol map. Everything else — motivation, behavior, concepts, will, the trajectory log — is absent in it. And, as shown above, even the symbolic prediction itself is realized in it in the fused regime: syntax and semantics are not separated, the utterance unit is not distinguished. What is inherited is an enormous but uncompleted resource.

10. 6. Completing the inherited predictor

The inherited symbolic predictor, taken by itself, is what this book has called a «symbolic orphan»: the prediction of symbols without a movement vector from below and without a contradiction vector from above, working moreover in the fused regime. Gativus completes it in two ways: it separates the fused predictors and connects the missing levels.

a) Separating syntax and semantics

The first step is to translate the inherited prediction from the fused regime into the recursive: to distinguish the utterance unit as a whole (the work of the syntactic predictor) and to place over it a semantic predictor predicting the next unit. Only after this does a place appear to which motivation can be connected: the question «which next utterance leads to the goal» is meaningful only when the unit is distinguished.

b) Connecting motivation from above

An isolated predictor has no vector distance of its own: it has nothing to nullify, it merely continues the text. In Gativus, above the symbolic level stands motivation — the semantic distance between the current narrative state and the goal. The semantic predictor connected to it ceases merely to continue plausible text: it builds a narrative for the sake of nullifying the distance. A goal of the account appears, which the orphan did not have.

c) Connecting behavior from below

In an isolated predictor the symbols are not bound to physical reality — there are no maps of the physical level, no objects, no b-vectors. Hence hallucinations: a symbol is not connected with any object that can be seen and taken. In Gativus, below the symbolic level lies the full physical level: a symbol is reversible into an object, an object has coordinates, an action is described by a b-vector. The connection from below gives symbols a support in reality — reverse deconvolution into a vivid representation and into a testable action.

d) Connecting will from above

An isolated predictor has no conceptual level — no contradictions, no will, no direction coming from the ought. In Gativus, above the symbolic level stands the conceptual: the convolution of narratives into concepts, the w-vectors of contradictions, volitional acts. The connection from above gives the narrative a source surpassing plausibility: the account is conducted not only coherently, but also in defense of a concept, under the pressure of an unsublated contradiction. The symbolic level becomes the executor of will, not an end in itself.

Table 10.2. What the completion of the inherited predictor gives.

The sum of these steps is the transformation of the «symbolic orphan» into a full-fledged symbolic level of subjective reality. The inherited predictor remains the same — but now it works in the recursive regime, under motivation, over behavior, and under will. Gativus does not discard the achievement of language models and does not compete with it — it separates in it what is fused and places it in the missing context.

10. 7. Prediction and prediction error

Prediction is connected directly with the marking of experience. The outcome marker, by which a completed unit of the trajectory log is marked as successful or unsuccessful, is tied not to the result itself, but to its unexpectedness — to the discrepancy between what the semantic predictor predicted and what happened. This is exactly the model of prediction error: the signal is the stronger the more the prediction diverged from the fact.

Here the architecture converges with a firmly established neurobiological fact. The dopaminergic system encodes precisely the error of reward prediction, not reward as such: a burst arises when the result is better than expected, and dips when worse. In Gativus terms this is a measure of the discrepancy of the prediction with the actual outcome at the behavioral level. Prediction and marker are two sides of one: the semantic predictor predicts, the marker fixes the error, the symbolic convolution is retrained on the marked experience so that next time the prediction is more exact. The layer-wise attachment of other neurotransmitters remains a hypothesis, but the dopaminergic anchor of prediction is reliable.

10. 8. Transition to the next chapter

The symbolic predictor, taken in isolation — working in the fused regime, without motivation from above, without behavior from below, without will over it — is precisely the large language model as it exists today. This chapter has shown what symbolic prediction is in the full architecture, how syntax and semantics are separated in it, and how it is completed into a full level. The next chapter considers the reverse case: what this predictor represents when left alone, what properties follow from this, and why no scaling up of the isolated model will replace the missing levels or bring it out of the fused regime. The language model is not a failed mind, but a precisely delineated fragment of the architecture: a symbolic orphan that Gativus returns to the family of transformations.

10. 9. Conclusions

• Prediction arises not on the convolutional map (there are only juxtaposed objects there, no «next»), but starting from the splice — where there is a connective vector and hence a transition.

• At each level two distinct predictors operate. The syntactic builds the current unit (form; at the three levels — step / syntax / obligation): low-dimensional, fast, by the rules of the level. The semantic chooses the next unit (content): high-dimensional, slow, under motivation and the trajectory log.

• Goal and will enter speech through the semantic predictor (it is connected to motivation), not through the syntactic.

• Semantic prediction is possible in two regimes. The fused (language models): syntax and semantics are not separated, symbols are predicted, the unit boundary is lost, there is nowhere to connect the goal. The recursive (Gativus): syntax closes the unit, semantics predicts the next under a guiding goal.

• The difference of the regimes is architectural: scaling up the fused regime does not create a unit boundary; the recursive regime requires a different organization — the distinguishing of units as objects.

• The main practical thesis: the symbolic predictor with the symbol map is already built by humanity in the form of the open weights of language models. Gativus inherits it rather than trains it anew — the most expensive component is obtained ready.

• What is inherited is the symbolic level of one transformation in the fused regime, not a ready mind. Completion: separate syntax and semantics (the recursive regime), connect motivation from above, behavior from below, will from above.

• Prediction and the experience marker are two sides of one mechanism: the predictor predicts, the marker fixes the prediction error. The dopaminergic system encodes precisely this error — a firm anchor.