DARWIN ZERO

Distributed architecture for autonomous artificial cognition

DARWIN ZERO was formerly developed under the working name AATM — Autonomous Artificial Thinking Machine.

Older documents, dashboards and technical references may still use the name AATM. They belong to the same research lineage.

What has been built

Several components are already operational or publicly observable.

Carl : A first operational homeostatic agent running on Moltbook.
Autonomous Engine : A controlled sensorimotor baseline with movement, local sensors, prediction error, column activity and public observer dashboard.
Environment Simulator : A controlled artificial world used to expose the observer-side environment without transmitting labels as cognitive inputs.
Blob HB1 : A minimal homeostasis-to-behavior loop without sensors, cortical columns, reward or learning.
Blob Loop B1–B5 : A progressive series of public read-only dashboards documenting the transition from isolated homeostatic behavior toward richer loop validation, replay checks and controlled observer traces.

AATM: Autonomous Artificial Thinking Machine

Current artificial intelligence systems generally rely on representations that are already structured: labeled data, semantic embeddings, object models, external objectives or predefined reward functions. This approach yields high performance but leaves a fundamental question open: how can an autonomous system build its own representation of the world from partial sensory streams, from its own movements and from its internal state?

AATM, Autonomous Artificial Thinking Machine, is a research program devoted to this question.

An example of the question is available here : https://www.grdprocess.ch/where-is-the-cat/

Our firsts demonstrators are up and running : https://www.nextinsight.org/public/

A finite substrate in an open world

The world can always present, tomorrow, a signal, a combination or a sequence that the system has never yet encountered. An autonomous architecture therefore cannot hold in advance a slot, a class or a model for every possible situation.

AATM starts from a finite substrate:

a limited number of simulated receptors;
a limited number of distributed micro units;
local plasticity rules;
limited sensorimotor capacities;
a vector set of internal variables.

The diversity of representations does not come from a prebuilt catalog but from the combinatorics among these elements, from their relations and from the continuous modification of their connections.

No module contains the complete object

In the architecture under study, an object such as a cat is not stored in a unit, a column or a memory record.

Each local signal is distributed toward several micro units. Each unit receives only part of the information and learns local regularities: coactivations, transitions, relations with a movement, internal effects and predictions.

The observable object corresponds to a temporary coalition of compatible partial paths. This coalition can mobilize many units without any one of them containing the object as a whole.

The same components can take part in several entities. A cat and a dog can share textures, shapes or elementary movements. Their distinction comes from the combination of these components, from their relations and from their transformations over the course of sensorimotor exploration.

This organization also aims at better reuse of memory: shared components need not be fully duplicated for each new object.

Recognition and reconstruction

AATM distinguishes two phenomena that are often conflated.

Recognizing in the presence of the world

When the sensors are open, several coalitions may initially be compatible with an ambiguous fragment. Each produces predictions about the signals that should appear after a movement.

Moving the sensor then provides new observations. Confirmed predictions reinforce certain coalitions; invalidated predictions weaken the others.

Recognition therefore does not result from instantaneous classification. It corresponds to the progressive stabilization of the coalition that best explains the successive observations.

Reconstructing in the absence of stimulus

When the external signal disappears, for example when the eyes are closed, it can no longer separate a cat from a dog. The representation must then be generated by the internal dynamics of the system.

The reconstructed coalition can be influenced by:

the cortical activity still present;
the sensorimotor state;
the current combination of homeostatic variables;
the frequency and diversity of past experiences;
the number of paths giving access to the coalition;
the strength and reliability of the connections;
the speed and synchronization of propagation;
weak endogenous fluctuations.

The project notably introduces a virtual myelination: certain frequently confirmed routes become faster and better synchronized. The system does not search for a path recorded in symbolic form; activity spontaneously takes the routes that learning has made most accessible.

The internal state as driver

AATM does not reduce motivation to an external scalar reward.

The internal state is represented by several dimensions operating simultaneously within equilibrium ranges. Behavior depends on their combinations and not on an isolated variable such as hunger, stress or energy.

The same external situation can thus lead to different reactions depending on the internal configuration of the system. Conversely, a combination of imbalances can reactivate representations or action chains that previously contributed to restoring a comparable state.

This hypothesis leads to treating recall as a form of internal action. A representation can be reactivated:

to prepare an external action;
to compare several possible consequences;
to guide a new exploration;
or to modify certain internal dimensions controllable by endogenous activity.

The reconstruction of a meal cannot restore a real energy reserve. By contrast, the reconstruction of a situation associated with safety can, according to the hypothesis being tested, influence an internal dimension related to threat or activation.

Attention toward the outside and toward the inside

The model does not treat attention as a central module.

When perceptual information is insufficient, activity favors afferent pathways and exploratory movements: attention is directed outward.

When the world is sufficiently identified but several recalls or actions remain in competition, the system can temporarily reduce sensory influence, strengthen the recurrent loops and circulate copies of motor commands without immediately executing the action: attention is directed inward.

The same circuits thus serve to perceive, predict, recognize, reconstruct and prepare an action. These functions correspond to different regimes of a common substrate and not to a succession of independent modules.

This distinction gives operational content to the two founding categories of AATM:

Being in the world consists in letting the environment constrain and transform the internal structure.

Making the world consists in letting an internal imbalance traverse this structure, select a possible transformation, then act on the external or internal world.

A falsifiable software POC

The first validation will take the form of a controlled software environment. A virtual sensor will perceive only small fragments of entities composed from common primitives. It will have to move in order to discover their relations.

The core will run continuously:

with no labels transmitted to the system;
with no object identifier;
with no reset between two entities;
with no separation between training and use;
with no timestamp readable by the units;
with no absolute sensor position provided to the core;
with no central vote or central object memory.

An external observer will be able to know the ground truth, measure the results and reconstruct the causal graph of activations, without transmitting this information to the system.

The experiments will seek in particular to verify that:

prediction degrades when the motor copy is removed or scrambled;
the same entities reactivate compatible coalitions on revisits;
different entities remain distinguishable despite shared components;
a partial occlusion does not immediately erase the representation;
reconstruction without stimulus genuinely depends on the recurrent paths;
different homeostatic combinations favor different coalitions;
a novel signal is not forced into an already known category;
the system stays active over time without global saturation or collapse.

Scientific objective and scope

The first objective is not to claim general intelligence nor immediate superiority over existing systems.

It is to determine experimentally whether a distributed, continuous architecture motivated from within can produce separable causal effects:

movement dependent learning;
persistent representation without a centralized object;
recognition by predictive confirmation;
reconstruction by internal feedback;
selection influenced by homeostasis;
reuse of components across several entities;
integration of novel situations without a closed catalog.

A positive result would provide a technical basis for robotic systems and world models capable of learning continually from partial perceptions. A negative result would remain informative, because the protocol is designed to identify precisely which mechanisms fail to produce the expected effects.

The architecture, the experimental rules, the data structures, the ablations and the reproducibility criteria are now formalized. The next step is their implementation and their confrontation with a controlled test bench.