AATM — Autonomous Artificial Thinking Machine
THE FRAME demonstrates that structural rules over case accumulation is executable. AATM addresses a different question entirely: can autonomous cognition emerge from a distributed system driven by homeostatic regulation, without externally imposed objectives? The two programs share a methodological commitment to distributed architecture and structural rules, but they operate on independent research questions. Their findings are mutually informative, not hierarchically dependent.
The observation
Current monolithic systems are not capable of supporting a genuine causal analysis process, nor a true learning process. The developments observed so far are superficial and result only from monolithic scaling. Some describe these systems as stochastic parrots. Others see the development of new types of monolithic systems, such as world models, as the solution for integrating a real learning layer.
The problem with monolithic architectures:
| Capability | Monolithic LLM | Distributed Multi-Agent |
|---|---|---|
| Learning | No learning from interaction at runtime | Continuous adaptation through interaction |
| Specialization | Requires explicit retraining | Emerges from interaction |
| Error correction | No mechanism to detect or resolve internal contradictions | Dialectical process between agents |
| Robustness | Single point of failure | Graceful degradation |
Monolithic properties are observed. Distributed properties are architectural targets derived from biological cognition research.
The question
Our brain does not function as a monolithic system. Its operation divides into three distinct layers: the cortical level (hundreds of thousands of sub-elements acting as biological automata), the individual level (the interaction of these elements within a complex organ), and the collective level (the interaction of individual brains with one another). Each level has its own processes, or replicas of processes from other levels, at a different scale. This leads to the emergence of learning and causal analysis — something a monolithic system does not enable.
Can this architecture be replicated in order to eliminate the weaknesses and shortcomings of monolithic systems?
The proposal
AATM repositions LLMs as environment rather than substrate. Agents are virgin — without prior, without inherited semantic bias. They interact with LLMs as organisms interact with a physical environment, driven by a single internal motivator: homeostatic equilibrium.
The two core agent functions are being in the world — capturing signals from the LLM environment to build an internal representation — and making the world — emitting signals toward the LLM environment to modify context toward homeostatic equilibrium. Both functions are driven by the same motivator. There is no separate communication module, no separate perception module. There is an organism seeking equilibrium.
The central architectural formula is D×R×I: Distribution × Reinforcement × Interactions. Intelligence quality is not a function of parametric size. Parametric scaling of a homogeneous system does not increase the Distribution term.
Current results
Two prototype series have been completed on commodity infrastructure at negligible cost.
The homeostatic prototype (March 2026) demonstrated a closed autonomous loop with three heterogeneous agents maintaining an internal state vector against perturbation without external input. Run 3 triggered dream mode via acute stress spike and recovered to within 0.05 of homeostatic target across all variables without any scripted recovery path.
The dialectic prototype (March 2026) demonstrated that homeostatic state variables per agent produce trajectories structurally different from normative training rules. The key finding: normatively constrained models resist dialectical pressure architecturally while open-reasoning models collapse under sustained pressure without explicit axiological anchoring. A degenerate homeostasis confound was identified and documented — the maintenance mechanism functions but the setpoint is imposed by training rather than system parameters. This is a result, not a corrupted test. It defines the next experiment.
Open problems
The homeostatic matrix (Q1) is not yet formalized. Without defining the parameters and their viable ranges, the translation layer between homeostatic state and symbolic interaction cannot be designed. Q1 is the prerequisite for everything that follows.