AATM: Can object identity emerge from continuous coalitions of partial sensorimotor representations?

[Gaetan_Duchateau]

Hi everyone,

I am working on AATM — Autonomous Artificial Thinking Machine — a research project exploring how an artificial system could continuously construct representations from local sensory signals, movement, and a dynamic internal state.

I am sharing the project here because it has significant overlap with the Thousand Brains approach: sensorimotor learning, movement-dependent prediction, distributed cortical computation, and the idea that higher cognitive functions should emerge from the architecture rather than be implemented as separate modules.

However, AATM currently explores a different hypothesis about where an object representation exists.

The central hypothesis

In AATM, no single cortical-like unit contains a complete model of an object.

A local sensory activation is projected sparsely and partially toward several simulated cortical units. Each unit receives only a subset of the available information and learns local regularities such as:

• recurrent sensory components;
• transitions conditioned by movement;
• relationships between partial observations;
• local predictions;
• associations with internal-state configurations.

An object is therefore not stored as a complete record inside one learning module. It exists temporarily as a distributed coalition of mutually compatible partial activations and recurrent pathways.

The same elementary components may participate in many different entities. A cat and a dog may share texture, shape, motion, or spatial components. Their identity comes from the particular combination of those components, their relationships, and their transformations during sensorimotor exploration.

This is intended both as a biological hypothesis and as a possible memory-economy principle: shared components do not need to be duplicated inside every object model.

A finite substrate in an open world

By “open world,” I mean that the system may always receive a signal, combination, or sequence that it has never encountered before.

The architecture is necessarily finite:

• a finite number of receptors;
• a finite number of cortical-like units;
• finite connectivity;
• finite local plasticity rules.

However, it must remain capable of forming new combinations and new coalitions without relying on a closed catalogue of object classes or predefined semantic labels.

A novel signal should therefore initially remain unexplained rather than being forced into the closest known category.

Continuous operation

Another major constraint is continuity.

The target architecture has:

• no natural separation between training and inference;
• no reset between objects;
• no cognitive timestamp available to the units;
• no central object database;
• no central voting or binding mechanism.

Time is represented through the dynamics themselves: residual activity, propagation delays, eligibility traces, recurrent activation, and the order in which signals arrive.

Coalitions emerge through recurrent support between compatible partial states. Incompatible coalitions receive less support or inhibit one another.

Recognition versus reconstruction

AATM distinguishes two processes.

Recognition occurs while external sensory input is available.

Several coalitions may initially be compatible with an ambiguous fragment. Each coalition predicts which additional components should appear after movement. The system then moves its sensor, receives a new fragment, and reinforces the coalition that maximizes the number and coherence of confirmed predictions.

Recognition is therefore not an instantaneous classification. It is the progressive stabilization of the coalition that best predicts the sensorimotor sequence.

Reconstruction occurs when external input is absent or strongly reduced.

In that case, there is no external evidence allowing the system to choose between, for example, a cat and a dog. The reconstruction must instead be selected by endogenous dynamics, including:

• recent cortical activity;
• the current internal and sensorimotor state;
• the number and diversity of access paths;
• synaptic strength;
• recurrent connectivity;
• propagation timing and reliability;
• previously learned internal effects;
• small endogenous fluctuations.

A working hypothesis is that frequently confirmed pathways acquire not only stronger synaptic gain but also faster and more reliable propagation — a form of virtual conduction or myelination plasticity. A pathway is not retrieved from a stored list. It is replayed because its topology and timing make it dynamically easier to traverse.

Homeostasis as an endogenous driver

AATM uses a multidimensional internal state rather than a single external reward.

Behavior depends on combinations of internal variables. A state corresponding to low energy, high threat, and low social security should not activate the same pathways as low energy combined with low arousal and high social need.

A further hypothesis is that an internal imbalance may reactivate representations that previously contributed to reducing a similar imbalance.

This means that recall may itself be a form of internal action.

For example, if a representation associated with a cat has previously reduced a threat-related internal variable more reliably than a representation associated with a dog, a similar imbalance may preferentially reactivate the cat coalition even in the absence of external input.

If both representations have equivalent internal effects, their relative accessibility would depend on their learned topology: number of paths, connection strength, propagation timing, previous frequency, and recent activity.

External and internal attention

The project also distinguishes two attention regimes.

When perceptual information is insufficient, the system should increase the influence of afferent signals and initiate exploratory sensor movements. Attention is directed outward.

When a persistent internal imbalance activates several possible actions but no action becomes dominant, the system may temporarily reduce motor execution and increase recurrent and descending activity. Attention is then directed inward.

Candidate motor commands can circulate as efference copies without immediately reaching the actuators. Their predicted sensory and internal consequences can then be reconstructed before an external action is selected.

The same substrate would therefore support perception, prediction, recognition, reconstruction, and action preparation through different dynamical regimes.

Planned software POC

The first implementation is intended to be deliberately small and falsifiable.

A virtual sensor will explore a controlled 2D environment containing composite entities built from shared elementary components. The cognitive core will not receive object identifiers, semantic labels, absolute world position, or timestamps.

The POC will test:

• sparse overlapping sensory projections;
• local prediction conditioned by movement;
• continuous residual activity;
• recurrent coalition formation;
• local plasticity;
• revisitation without reset;
• partial occlusion;
• reconstruction after sensory input is removed;
• selection by multidimensional internal states;
• controlled failure when movement information is removed or corrupted.

Ablations will separately remove movement copies, recurrent connections, residual activity, plasticity, and conduction-delay adaptation.

The project is currently at the architectural specification stage. I am not presenting these mechanisms as validated results. The purpose of the POC is precisely to determine which of them are necessary, insufficient, or incorrect.

I am not currently planning to use Monty as the core architecture because I want to avoid importing object-level representational assumptions into the first experiment. However, Monty may still provide useful environmental interfaces, sensorimotor tools, or comparative benchmarks.

Questions for the community

1. Is the ability of each learning module to maintain a complete object model essential to Thousand Brains Theory, or could complete identity emerge from coalitions of partial, reusable representations?
2. How could partial representations coordinate without introducing a central binder or an explicit object identifier?
3. Are continuous, non-episodic inference and internally generated sensorimotor replay already being explored within TBP or Monty?
4. Which parts of Monty could be reused as a neutral experimental substrate without constraining the representational hypothesis?

I would particularly value criticism of the falsifiability of the proposal and suggestions for experiments that could distinguish this architecture from complete-object learning modules.

[nleadholm]

Hi Gaetan, welcome to the forums, and thanks for your post.

Re. your questions:

1. Humans have a highly object-centric view of the world; it’s actually one of the things that distinguishes our internal representations markedly from those that emerge in deep learning. In the TBT, this is reflected in the fact that cortical columns learn models of whole objects. Importantly, however, these models are built from “coalitions of partial, reusable representations” that are experienced through sensorimotor exploration. We believe the object identity itself emerges in an unsupervised fashion in the brain, so any supervised learning experiments we have at the moment (i.e. where we pass an object ID) are just a stepping stone for more unsupervised approaches.
2. You might find our experiments on unsupervised learning interesting - there is no explicit object ID passed to Monty in that case. Rather, Monty recognizes that an object is unfamiliar relative to its learned models, and so begins learning a new model.
3. Yes, our unsupervised inference experiments evaluate the condition where Monty is never told that a new object is being presented. I’m not sure what you mean by internally generated sensorimotor replay in this context.
4. Depending on where you want to go with your project, you could consider designing a custom learning module (see more info here). That way you could re-use all the other components in tbp.monty, including the environment and low-level sensory feature extraction.

[Alex]

Hi @Gaetan_Duchateau and welcome to the group.

Looks like a very interesting project and I am sure many here will be interested in your progress.
Building an architecture that can handle such complex processes is the holy grail of AGI research.

Personally I am of the view that object recognition is the pinnacle of intelligence rather than the
starting off point. You refer to cats and dogs in you explanation, both Autonomous Natural Thinking Machines in their own right and I wonder what capabilities they possess in terms of object recognition.

Perhaps initially the world and everything in it can be considered as background. Only objects that hold importance (food, shelter, threat or mate) need to be recognised. And this importance is a significant factor in the recognition and understanding of those few objects. Everything else can be safely ignored. Or put another way, there is nothing to be gained by recognising something for which the Thinking Machine has no use.

To a cat, a desk cluttered with objects which it must navigate a path through is no different from an area of ground littered with rocks. The cat delicately steps over the stapler without giving it a second thought.

I think objects only begin to emerge form the background as they become important to the Intelligence, and this importance supplements the recognition process.

Stages or levels of object recognition may be:

If it moves against the background then it is not part of the background and is a threat

Is it large relative to myself, if small can be ignored (note that self body model is crucial here)

If large then is trajectory away or towards

The first thing an Autonomous Artificial Thinking Machine must do is avoid getting stepped on !

[Gaetan_Duchateau]

Hi Alex, thank you for the comment.

I should also mention that I am currently developing this with very limited resources, so progress will depend on what I can realistically implement in the short term. That is why I am focusing first on a minimal software POC rather than a full robotic implementation.

I agree with the general direction of your point, especially the idea that object recognition should probably not be treated as the starting point of intelligence.

I would push the starting point even earlier. In AATM, the initial process is not yet “object recognition” and not even “important object selection”. It is a pre-object sensorimotor process. The system receives local sensory activity, motor efference, proprioceptive state, internal-state activity, and recurrent activity. From these signals it builds predictive regularities.

What an observer later calls “background” is not ignored. It is part of the predictive field. But if it is stable, predictable, and has no relevant effect on internal state or action selection, it does not need to become an attended or stabilized object-level representation.

Conversely, what later appears as an “object” may emerge when part of that field stops behaving as predicted, moves relative to the rest, creates a sensorimotor constraint, or becomes linked to a change in the internal state.

So I agree that food, shelter, threat or mate are central for an autonomous agent, but I would not treat their importance as a label attached to objects at the beginning. “Important” is something that emerges from the coupling between prediction error, action possibilities, body model, and homeostatic effects.

Your example of the cat walking over a cluttered desk is a good one. I would say the cat does not need to recognize the stapler as a stapler. But the stapler is not absent from processing. Its geometry and sensorimotor constraints are part of the predictive field that allows the cat to step over it automatically. It remains background only because no additional attention, reconstruction or object-level stabilization is required.

I would also be cautious with motion alone. Motion can attract attention, but the world is full of motion that animals ignore. The key variable may be motion relative to prediction: looming, unexpected trajectory, relative movement against a stable field, or movement coupled to internal-state relevance.

In that sense, “avoid getting stepped on” is very close to what I want to test. But at the pre-object level, the system should not need the concept of a threat. It should only need variables such as relative size, trajectory toward the body, failure of background prediction, and previously learned internal consequences.

So the distinction I would make is:

object recognition is a late stabilized result;

sensorimotor prediction and homeostatic relevance are earlier mechanisms that make some parts of the field emerge from the background.

AATM is trying to start from that earlier level.

I should also mention that I am currently developing this project with very limited resources, so progress will depend on what I can realistically implement in the short term. This is why I am focusing first on a minimal software POC rather than a full robotic implementation. Any technical feedback, pointers, or potential collaboration around this first falsifiable prototype would be very valuable.

[Gaetan_Duchateau]

Hi Niels,

Thank you for the detailed answer. It helps clarify several points, and it also corrects one of my assumptions about Monty.

I had probably overstated the dependency on explicit object IDs in the current system. Your unsupervised learning experiments are closer to what I am interested in than I initially understood, especially the case where Monty detects that something is unfamiliar and starts building a new model without being explicitly given an object ID.

However, I think I should clarify the level at which AATM is trying to operate.

My current focus is not primarily to propose an alternative object-recognition module. It is to investigate a process that comes before object recognition.

At the initial level, AATM is meant to be pre-object and pre-semantic. The system is not given object boundaries, object identity, object categories, parts, or semantic labels. It only receives local sensory activity, motor efference, proprioceptive activity, internal-state activity, and recurrent activity.

From this, the system should build predictive regularities.

What an external observer later describes as an “object” would initially correspond only to a recurrent coalition whose internal relations remain predictable across sensor movement, partial sensory interruption, and internal-state modulation.

So the distinction I am trying to make is not necessarily:

“Monty has object models, AATM rejects object models.”

It is rather:

“At what point in the process does something become object-like?”

In AATM, I would like to start before that point.

A stable object-level representation may eventually emerge, but I do not want to assume it at the beginning. I want to test whether local sensorimotor covariation, prediction, recurrent activity, movement and internal-state modulation can make some parts of the sensory field become more stable, more predictive and more actionable than others.

This is also why I am careful with the word “object.” In the first POC, “object” would mostly be an observer-level term. Internally, the system would only manipulate activations, transitions, predictions, errors, traces, gains, thresholds and recurrent pathways.

I also agree with your point that humans have a strongly object-centric view of the world. My hypothesis is that this object-centric view may be a late stabilized result of more primitive processes, rather than the starting point. AATM is trying to start from those earlier processes.

This is where I find Spelke’s work relevant. Not because I want to implement explicit principles such as cohesion or object permanence as symbolic rules, but because the experiments suggest useful behavioral phenomena to reproduce: common motion should favor shared stabilization, relative motion should favor separation, and stationary ambiguous configurations may legitimately remain unresolved.

By “internally generated sensorimotor replay,” I should probably use a clearer term. What I mean is closer to an endogenous covert sensorimotor rollout.

The idea is not simply replaying a stored episode. It would be a condition where external sensory input is absent or downweighted, while a partial internal trigger reactivates part of a learned sensorimotor structure. Motor commands may circulate as efference copies without being executed, and recurrent or descending pathways may regenerate expected sensory transitions and internal consequences.

The purpose would not be to recognize a currently presented object, but to support reconstruction, internal simulation, or action selection when immediate sensory evidence is insufficient or absent.

In AATM, I am also exploring whether a multidimensional internal or homeostatic imbalance could trigger such a rollout. For example, a certain internal imbalance could reactivate pathways that previously contributed to reducing a similar imbalance.

Your suggestion about a custom Learning Module is useful, but I think it is probably a second step for me. I first need to test the core AATM mechanism in a very small standalone software POC, without importing too many assumptions from an existing object-recognition framework.

A possible path would be:

1. first build a minimal standalone POC to test the pre-object mechanism;
2. then, if the results are meaningful, investigate whether the same mechanism could be expressed as a custom Monty Learning Module;
3. then compare the two approaches on the same sensorimotor stream.

The first comparison I would be interested in is not “which system recognizes more objects,” but whether the pre-object substrate shows measurable dependence on movement, prediction, recurrent stabilization and internal-state modulation.

For example:

• does performance collapse when motor efference is removed or corrupted?
• do stable coalitions reappear when the same structure is revisited?
• can ambiguous configurations remain unresolved until movement provides discriminating information?
• can an internal imbalance bias reconstruction in the absence of external input?
• can the system avoid forcing a novel signal into an existing object category too early?

So my current question is less about Monty implementation details and more about the theoretical boundary:

Where, in TBT, does the system stop being purely sensorimotor and become object-level?

Or, put differently:

Could the object-centric representation be treated as an emergent stabilized regime of lower-level sensorimotor dynamics, rather than as the starting representational format?

I should also mention that I am currently developing this project with very limited resources, so progress will depend on what I can realistically implement in the short term. This is why I am focusing first on a minimal software POC rather than a full robotic implementation. Any technical feedback, pointers, or potential collaboration around this first falsifiable prototype would be very valuable.

Thanks again. Your answer helped me clarify that the key issue is not simply supervised versus unsupervised object learning, but the earlier transition from pre-object sensorimotor dynamics to stable object-level representation.

[Alex]

Hi Gaetan

Good point about the world being full of motion, leaves moving in the wind etc, animals seem to be very good at masking that stuff out.

To that cat I think the stapler is not an object at all but an integral part of an uneven surface, as is everything on the desk unless it smells like food. I doubt that a cat would see a china cat as a cat, it lacks too many of the other sensory clues that make it a cat.

If you look at Project Arachne in the Projects section of this forum you will see that I have designed a low cost sensorimotor platform specifically to experiment and learn about this low level type of sensorimotor intelligence in the real world. It has a primitive nervous system providing a variety of sensory data (afferent nerves) and accepting motor control (efferent nerves).

At present it is not feasible to hook it up to Monty but that may change in the future. Right now I am trying to create a neural network capable of receiving the sensory data and generating simple motor commands in response.

Alex

[Gaetan_Duchateau]

Hi Alex,

Project Arachne looks very relevant as a real-world sensorimotor platform, and I agree that this kind of setup is probably necessary to test low-level sensorimotor intelligence properly.

The point I wanted to isolate is slightly different, though. My issue is with the idea that the stapler would simply be “part of an uneven surface.”

At the low sensorimotor level, I would not assume a prior distinction between object, surface, stapler, desk, or background. Those are already higher-level descriptions. The system should not begin by deciding whether something is an object or not. It should only encounter changing sensory states through movement.

So, from that point of view, the stapler and the desk are not different because one is an object and the other is a surface. They are different only if they produce different sensorimotor regularities or different state transitions.

For example, if the stapler remains fixed on the desk and never produces a separate prediction regime, then it may be treated as part of the same stable configuration. But if it moves independently, creates a discontinuity, can be displaced, occludes differently, gives different tactile feedback, or changes the sensory stream differently from the desk, then it becomes separable as a structure.

So I would not formulate it as: “the cat sees the stapler as an uneven surface.”

I would formulate it as: “the system has not yet encountered a state transition that forces the stapler to be modeled separately from the desk.”

That distinction matters to me, because objecthood should not be injected as a category at the beginning. What comes first is movement, pose, local sensory change, prediction, error, and stabilization. The later notion of “object” is a post-hoc description of a stabilized pattern of sensorimotor transitions.

In that sense, Arachne could be very useful precisely because it can test which physical changes produce separability: pushing, touching, occlusion, independent motion, resistance, smell, vibration, etc. But I would avoid saying that something is or is not an object at the primitive level. I would rather say that some structures have not yet become separable through state changes.

Gaetan

[nleadholm]

Hi Gaetan, sounds like an interesting effort. Anything that can be done to relax assumptions about things like episode boundaries during learning would be very welcome. If you have any specific questions about your implementation while working on the POC, feel free to reach out.

[Gaetan_Duchateau]

Hi Niels,

Thanks — that is exactly the kind of issue I am trying to isolate.

Since my previous message, the first public POC is now live here: https://www.nextinsight.org/autonomous-engine/

The dashboard is mainly an observer console. It is not meant to be a product interface or a demonstration of high-level intelligence. It exposes the internal loop so that the process can be inspected cycle by cycle: internal state, selected movement, digital pose, local observations, feature signatures, updates of the column-like structures, prediction success or error, and later the attention / reinforcement / voting signals.

So the dashboard corresponds to the “external observer” layer of the POC. It does not participate in cognition. Its role is to make visible whether the system is really learning through movement, pose and local prediction, rather than through some hidden global representation.

What the POC is:

It is a minimal sensorimotor testbed in a controlled digital environment. The goal is to test whether an initially empty structure can start filling through a continuous loop: internal state → action pressure → movement → pose change → local observation → prediction → prediction error or confirmation → reinforcement.

It is also a way to start relaxing explicit assumptions about episode boundaries. I am trying to avoid giving the system pre-cut learning episodes. The loop should remain continuous, and what later looks like a boundary should ideally come from state transitions, prediction errors, saturation, novelty, or changes in the internal state, not from an externally imposed training segmentation.

What the POC is not:

It is not an implementation of Monty.

It is not yet a real web-scale agent.

It is not a claim of semantic understanding or high-level cognition.

It is not trying to outperform current AI systems.

It is also not yet proof that the cortical-column-like architecture is correct. At this stage, the claim is much narrower: can the loop run, remain observable, and show that learning depends causally on movement, pose and local prediction?

The next step is to harden the validation protocol. In particular, I want to run controlled ablations: remove movement, scramble pose, preserve features but break coordinates, revisit already explored surfaces, and measure whether prediction improves over time only when the sensorimotor structure is preserved.

Once this is stable, I will probably have more specific implementation questions for you, especially around how to represent continuous transitions without accidentally reintroducing hidden episode boundaries or object-level assumptions too early.

Thanks again for the offer — I’ll reach out when I have a concrete technical point where your experience with Monty can help avoid a wrong shortcut.

Gaetan

[Gaetan_Duchateau]

Hello,

I plan to submit the project for a grant from longview philanthropy. I would like to know it there is some kind of conflict of interest in this case if I am willing to continue to communicate on this forum and perhaps collaborate to some extend with people here.

[nleadholm]

Hi @Gaetan_Duchateau , sure, can you clarify your concern about a conflict of interest? As in whether anyone in the TBP has any involvement with Longview Philanthropy? As far as I’m aware, this is not the case.

[Gaetan_Duchateau]

Yes, this case and also to know of it is allowed to participated here while having other (let’s call that this way) shareholders that will indirectly benefit from future potential collaboration I will find here.

[nleadholm]

Yes you’re welcome to have discussions here. If you’re wondering about using Monty in commercial applications, you can read more about our permissive MIT licence and non-assert patent pledge here: Our Licence and Patent Pledge Explained

[Gaetan_Duchateau]

AATM Environment Simulator — first public test version

We have reached a new milestone in AATM: the first public test version of the artificial environment simulator is now online.

The simulator provides a controlled 100 × 100 digital world designed to support future sensorimotor experiments inspired by Thousand Brains Theory. Each cell contains a background and, optionally, one element. Both use the same continuous signal structure, spatially distributed across subcells and receptor sites.

This representation preserves similarity: for example, light green remains closer to dark green than to blue. It also generates local spatial patterns that can later be sampled through movement, rather than presenting an already interpreted object to the agent.

The current backend is running as version 0.1.0, with 10,000 cells, 10 background families, and 1,000 distributed elements. The environment is static for this first baseline test, and the independent mutation engine remains disabled.

A key design constraint is that observer labels such as crocodile, cat, or dog will never be transmitted to the future agent. The agent will receive only local signal patterns and, later, their sensorimotor and homeostatic consequences.

The objective is to test whether initially empty cortical-column analogues can progressively differentiate the environment through:

movement → local signals → prediction → surprise → reinforcement

This is not yet the artificial “blob” itself. The present milestone is the environment: inspectable, reproducible, exportable, and ready for controlled experiments.

The next steps will introduce controlled environmental dynamics, bounded mobile sensors, the homeostatic core, and the progressive filling of cortical columns.

Live test dashboard:

AATM — Artificial Environment Simulator

Feedback on signal encoding, receptor-level spatial structure, sensorimotor sampling, and scaling would be very welcome.

[Gaetan_Duchateau]

Implementation update — the first instrumented Blob layer is now public

Hello everyone,

Following the environment simulator update, I have now published the first instrumented component of the Blob itself: the homeostasis-to-behavior layer.

I deliberately moved one step backward from the full sensorimotor architecture.

Before connecting the Blob to sensors, movement, reference frames, and cortical-column-like structures, I wanted to verify that the internal driver could be implemented in a way that is observable, reproducible, and causally traceable.

The current POC contains four continuous functional variables named after:

• adrenaline;
• cortisol;
• dopamine;
• serotonin.

These names describe functional abstractions. This is not intended as a biochemical simulation.

The exact numerical values are the primary data. The system does not select a behavior from simple high/normal/low labels.

For every round, it calculates:

• the exact value of each variable;
• its signed distance from its equilibrium range;
• its normalized distance;
• the behaviorally dominant variable;
• the ordered minor variables;
• the behavior-resolution rule that was applied.

The current repertoire is deliberately limited to “being in the world” intentions:

• maintain observation;
• search more widely;
• search more locally;
• re-examine;
• change search axis;
• reduce search intensity;
• increase search intensity.

There are no sensors or physical movements yet, so these are behavioral intentions rather than executed motor actions.

After each behavior, one of four synthetic outcome classes is selected pseudo-randomly:

• maintenance;
• transition;
• favorable return;
• failure.

These outcomes are not rewards, and their probabilities are not learned.

They temporarily stand in for the body and environmental consequences that will be introduced later.

The important validation sequence is:

1. a behavior is resolved from the exact molecular state;
2. one outcome class is selected;
3. an exact molecular variation is generated and applied;
4. the resulting state is persisted;
5. that state becomes the input of the next round;
6. the next behavior is independently resolved from the new state.

The next behavior is therefore not predefined in an outcome table.

It must emerge from the molecular values actually produced by the selected outcome.

The current public baseline contains 100 persisted rounds. Across those rounds:

• molecular state carry-over validation is true;
• behavioral carry-over validation is true;
• every selected outcome was verified against its realized molecular effect.

The experimental configuration is also external to the runtime code, versioned, immutable per run, and publicly visible.

This includes:

• absolute molecular ranges;
• equilibrium ranges;
• variation steps;
• behavioral-major rules;
• behavior-resolution rules;
• outcome probabilities;
• allowed molecular impacts;
• configuration hashes and revisions.

This separation is important because the current values are explicitly experimental and not calibrated. If the Blob proves apathetic, unstable, or dominated by one variable, a new configuration revision can be tested without changing the interpretation of previous runs.

A clarification about the dashboard:

The backend service is continuously online, but the displayed experiment is not yet a continuously scheduled autonomous Blob.

The dashboard currently exposes the latest persisted state of a recorded 100-round baseline. New rounds are created only through protected backend commands. The public dashboard is read-only, and its auto-refresh function only reloads persisted data.

I therefore consider this an instrumentation and validation milestone, not yet a demonstration of continuous autonomous cognition.

The next planned stages are:

1. add one initially empty cortical-like column and observe how recurrent homeostatic configurations create or reinforce traces;
2. add two bounded moving sensors;
3. integrate movement, pose, and efference information;
4. keep homeostasis distinct from interoception, with interoception later constructed from internal sensorimotor signals under homeostatic modulation;
5. introduce a variable external signal;
6. validate the complete local prediction chain;
7. only then introduce distributed reconstruction and voting.

Two questions may be relevant to the next stage:

1. For the first empty cortical-like unit, would you initially store the raw homeostatic vector as explicit trace content, use it primarily as modulation of gain and plasticity, or test both variants?

2. When introducing two bounded sensors, what would you consider the minimum pose and motor-efference information required for a meaningful movement-ablation test without importing object-level assumptions?

Public dashboard:

Cognitive systems

The dashboard and configuration are currently marked:

EXPERIMENTAL — NOT CALIBRATED

Critical feedback on the data model, the separation between regulation and future sensorimotor evidence, and the next one-column experiment would be very welcome.

[Gaetan_Duchateau]

A further question: cortical capacity and the learning niche

As mentioned in the previous update, the next implementation step is to add the first initially empty cortical-like column. While thinking about how this should later scale to a distributed architecture, a broader question emerged:

Does greater cortical capacity make a richer learning niche possible, or does a richer learning niche select for greater cortical capacity?

I mean this as an evolutionary and architectural question, not as the claim that an individual develops a larger cortex because it reads more books.

A larger inherited cortical architecture may allow an organism to preserve more distinctions, learn longer temporal relationships, and recombine previously learned structures in more ways. However, once these additional distinctions become behaviorally useful, the organism may also modify or expand its own effective niche through exploration, tools, social learning, language, or external memory. That richer niche may then create further selection pressure for additional cortical capacity.

My current intuition is therefore a feedback loop:

greater cortical capacity
→ more regularities become learnable
→ the effective learning niche expands
→ additional capacity becomes advantageous

The relevant variable may not be raw sensory bandwidth or receptor count.

Cats and humans do not have identical sensory systems, but both receive highly structured and information-rich sensorimotor streams. The major difference may lie partly in the number and depth of predictively distinct, temporally extended, and recombinable structures that their cortical architectures can preserve.

The human learning niche is also extended by language, teaching, books, tools, and cumulative culture. A human can learn about Australia, historical events, or situations never directly experienced. The immediate sensory channel may still consist of familiar visual or auditory signals, but those signals provide access to a much larger relational and combinatorial space.

For the AATM experiments, this suggests that I should not test only “more columns versus fewer columns.” Cortical architecture and niche complexity may need to be varied together.

By cortical capacity, I mean not only the number of columns, but also:

• the local capacity of each column;
• the distribution and overlap of their inputs;
• their connectivity and recurrent topology;
• their temporal prediction horizon;
• their plasticity and consolidation rules.

By niche complexity, I mean the number of distinctions that are recurrent, learnable, and relevant to different future predictions—not simply the volume of incoming data.

One possible experiment would keep instantaneous sensory bandwidth constant while progressively increasing temporal depth and compositional complexity. Different cortical architectures could then be compared by measuring when distinct situations begin to collide or interfere—for example, when two trajectories that lead to different future signals are no longer maintained as separate internal structures.

Has this relationship between cortical capacity and effective niche complexity been considered explicitly in Thousand Brains Theory or Monty?

More specifically:

• Is there a principled way to relate Learning Module capacity, number, or topology to the effective complexity of a sensorimotor learning niche?
• What would be the clearest signature of insufficient capacity: persistent prediction error, interference between learned models, unstable recognition, failure to preserve distinct trajectories, or loss of compositional generalization?
• Have adaptive specialization, structural growth, merging, or pruning of Learning Modules or learned structures been considered?
• Would keeping sensory bandwidth constant while increasing temporal and compositional complexity be a meaningful way to compare different Learning Module architectures?

This is starting to look like a possible working premise for AATM:

Cortical capacity should be evaluated relative to the number and complexity of predictively distinct trajectories that the system must keep separate, rather than relative only to receptor count or raw data volume.

[Gaetan_Duchateau]

I have added a specification doc for the POC : https://www.grdprocess.ch/wp-content/uploads/2026/07/20260719_AATM_Cortical_V1_Step_by_Step_Demonstration_Specification_EN.docx