Concept Volumes, Dependency Cells, and the Direction of AI Research

In a recent interview on the idea of Conceptron, Ron Pisaturo presents a valuable and timely challenge to the dominant way language models represent meaning. Current large language models represent tokens through vectors. Words become points in a high-dimensional embedding space. Words that appear in similar contexts tend to be near one another. This is one of the major breakthroughs that made large language models work: language was mathematized into vector and matrix form, allowing massive parallel computation, gradient-based optimization, and context-sensitive transformation.

Pisaturo’s proposal asks whether this is enough.

His answer is that it is not. A word-vector can place “tree,” “bush,” and “plant” near one another, but it does not naturally express the deeper conceptual relationship among them. “Plant” is not merely beside “tree” and “bush.” It is a broader concept that can contain them. A concept is not just a point. It is a bounded range of possible instances. It has scope, extension, measurement, overlap, hierarchy, and concrete reference.

This is the central move in Conceptron: replace concept-points with concept-volumes.

Instead of representing a concept as one location in embedding space, Conceptron proposes representing a concept as a volume across many dimensions of measurement. “Green” would not be a single point. It would be a range of hue, saturation, and brightness. “Tree” would not be a single vector. It would be a region structured by ranges of height, hardness, trunk formation, leaf structure, growth pattern, and other measurable features. Concrete instances would live inside or near these bounded regions. Related concepts could overlap, contain one another, or separate from one another.

This is an important shift. It moves AI representation away from language as token proximity and toward concepts as structured regions of possible reference.

From the perspective of Recursive Generative Emergence, I think this direction is exactly right.

My own work has been developing a broader architecture around a related question:

How does a system turn possibility into stable meaning?

That question begins even before concept representation. Before a concept can be represented as a volume, the system must produce a valid unit of reference. It must distinguish something from its background. It must form a boundary. It must stabilize identity across variation. It must decide when a cluster of variation counts as one thing, when it should split into multiple things, and when a supposed thing is only a false reification.

This is what I describe as Recursive Individuation.

The primitive sequence is:

potential field

→ distinction

→ boundary

→ persistence

→ identity

→ node

→ relation

→ dependency

In this frame, a Conceptron volume corresponds to one important layer of the process. It describes the semantic body of a concept once a candidate unit has already become bounded enough to be treated as a concept.

But my work asks what happens below and above that layer.

Below the concept-volume, I ask: What makes something eligible to become a node at all?

Above the concept-volume, I ask:What makes that node possible? What dependencies support it? What consequences does it enable? What predictions does it generate? What failures expose an incorrect boundary or missing dependency? How does the system revise itself when the concept fails?

This is where RGE, RID, RDD, SCDD, and the Autognizer extend the discussion.

A useful distinction is this: Conceptron addresses concept geometry. The Autognizer addresses recursive concept use.

A concept-volume says: This is the bounded identity of the concept.

A Dependency Cell says: This concept is possible because of these supports, it enables these consequences, it fails under these conditions, and it revises through this mechanism.

So I would not say that Pisaturo’s volumes are the same as Dependency Cells. They are closer to what a Dependency Cell contains. The concept-volume gives the cell its semantic body. The Dependency Cell adds support structure, consequence structure, prediction, failure detection, and revision.

In compact form:

Conceptron volume: C_X = bounded conceptual region for X

Dependency Cell: D_X = C_X + backward dependencies + forward dependencies + possibility gate + local failure signal + revision rule

This gives a nested hierarchy:

Recursive Coherence Formation

→ Recursive Individuation

→ stable node formation

→ concept-volume formation

→ dependency-cell formation

→ dependency graph construction

→ prediction and action

→ failure detection

→ credit assignment

→ graph revision

→ recursive stabilization

→ Autognizer-level intelligence

This is why the interview is exciting from the my perspective. It shows that serious AI discussion is beginning to move toward questions I believe are foundational: concept formation, grounded reference, bounded abstraction, hierarchy, and the difference between linguistic proximity and real conceptual structure.

The current LLM paradigm has been extraordinarily successful, but its strength is also its limitation. It works by learning statistical structure across language. It can model use, association, probability, and context with tremendous power. But the fact that a model can use a word correctly does not mean it has represented the concept as a bounded, dependency-bearing structure. Usage similarity is not the same thing as conceptual identity. Token proximity is not the same thing as grounded reference. Prediction is not the same thing as recursive coherence.

Pisaturo is pointing toward this gap.

He also identifies the hard technical problem. If scalar embeddings are replaced with ranges or volumes, the rest of the LLM machinery does not automatically carry over. Autoregression, backpropagation, gradient descent, and text generation were all designed around the current vectorized token system. A volume-based system needs new machinery for learning, correction, and generation.

This is precisely where my RGE work becomes relevant.

Self-Correcting Dependency Discovery is my attempt to formalize a correction process for meaning-bearing structures. The system does not merely store a concept. It tests the concept through dependency, prediction, and failure. When a prediction fails, when a contradiction appears, or when a missing dependency is discovered, the system assigns credit or blame to parts of its graph. It then revises the node, the edge, the boundary, or the dependency structure.

This means intelligence is not treated as static representation. It is treated as recursive coherence formation.

A system becomes more intelligent as it becomes better at forming stable units, discovering what supports them, predicting from them, detecting failures, and revising its own structure.

That is the larger process I believe AI research is moving toward.

The interview with Pisaturo is valuable because it gives a clear public articulation of one layer of that transition: concepts should not be treated merely as points in language-space. They should have bounded extension. They should relate to concretes. They should be able to contain, overlap, and differentiate. They should have geometry.

RGE agrees with that move and extends it into a broader architecture.

Concepts need geometry, but they also need dependency.

They need boundary, but they also need revision.

They need concrete reference, but they also need recursive validation.

They need to be represented, but they also need to be tested.

This is where the field appears to be moving: away from flat token association and toward structured, grounded, self-correcting conceptual systems.

I am glad to see that movement.

From the RGE perspective, the next generation of AI will not be defined only by larger models, larger datasets, or more parameters. It will be defined by better recursive structure. The important question will not only be, “How much language has the model seen?”

It will be: Can the system form valid nodes? Can it preserve conceptual boundaries? Can it discover dependencies? Can it detect contradiction? Can it assign failure correctly? Can it revise without collapsing coherence? Can it stabilize meaning over time?

Conceptron offers one possible answer at the level of concept representation.

The Autognizer aims at the larger control process: how concepts are born, tested, connected, repaired, and stabilized into intelligence.

That is why I see Pisaturo’s interview as a meaningful signpost. It does not cover the whole architecture I am working on, but it points in a direction I consider essential. The future of AI research will likely require more than better next-token prediction. It will require a theory of how meaning becomes bounded, how concepts become dependency-bearing, and how systems recursively correct themselves when their own structures fail.

That is the research direction I have been pursuing through Recursive Generative Emergence.

And it is encouraging to see more of the field beginning to ask the same kind of question.