This post presents a unified framework — developed through iterative philosophical inquiry and validated against the academic literature — that proposes intelligence is most precisely understood as the dynamic management of two simultaneous entropy streams: one arriving from the external sensorial environment, and one generated internally through the brain’s own modelling processes. The framework derives from this foundation a complete account of how intelligence fails, why cognitive specialization and cooperation are thermodynamic necessities, and why the multi-agent AI architectures currently being built are recapitulating — unknowingly and imperfectly — the same entropic laws that biological evolution took millions of years to discover.
The relevance to AI alignment is direct: several of the failure modes the alignment community is currently treating as engineering problems are, under this framework, manifestations of a single underlying thermodynamic law. Recognizing them as such changes how they must be solved.
I. The Starting Point: Intelligence as Order
The framework begins with a deceptively simple observation: intelligence, in every documented case, is the imposition of order on disorder. This is not a metaphor. It is a thermodynamic claim.
The neuroscientific evidence is unambiguous. Saxe, Calderone and Morales (2018 ) demonstrated in a study of 892 healthy adults using resting-state fMRI that brain entropy — the number of distinct neural states a brain can access — is positively correlated with both fluid and crystallized intelligence . A more variable, less predictable brain is a more intelligent brain. The brain does not suppress entropy; it organizes it.
Friston’s Free Energy Principle (2010), cited over 12,000 times, formalizes the complementary claim: the brain is a prediction machine whose fundamental drive is to minimize surprise — that is, to minimize the entropy of its sensory predictions . Intelligence is the continuous reduction of the gap between what the world is and what the brain expects it to be.
These two findings are not contradictory. They describe the same phenomenon from opposite ends: a brain that can access many states (high neural entropy) is better equipped to minimize the entropy of its predictions about the world. The capacity for disorder enables the management of disorder.
II. The Dual-Entropy Model
The framework proposes that intelligence operates on two distinct entropy streams simultaneously, and that conflating them is the source of considerable confusion in both cognitive science and AI systems design.
External entropy is the disorder arriving from the sensorial environment. Because the world is inherently dynamic and unpredictable, the brain is continuously bombarded with uncertain information. Lancaster and Wass (2025) demonstrated using information-theoretic entropy measures that environmental complexity and unpredictability directly shape cognitive development . The more chaotic the external world, the greater the cognitive demand placed on the system to extract signal from noise.
Internal entropy is the disorder generated within the brain’s own models. This arises from two sources: neural noise (random variability in neural firing), and what Ezzedini (2025) calls schema rigidity — the failure of internal cognitive models to remain flexible enough to adapt to a changing external reality . When internal models become too rigid, they generate systematic prediction errors: the brain’s map diverges from the territory.
The key insight is that these two streams require opposite responses. External entropy demands increased precision weighting — the brain must sharpen its attention and filter signal from noise. Internal entropy demands increased model flexibility — the brain must loosen its priors and update its schemas. A system that applies the same response to both streams will fail at one or the other.
III. The Three Failure Modes of Intelligence
From this dual-entropy model, three failure modes follow with logical necessity.
Failure Mode 1: Sensorial Misvalidation. The brain must assign a precision weight to every incoming sensory signal — deciding how much to trust external reality versus its own internal predictions. When this calibration fails in either direction, intelligence collapses. Corlett et al. (2025) in the American Journal of Psychiatry describe psychosis as fundamentally a state of aberrant salience: dopamine dysregulation causes the brain to massively over-weight irrelevant sensory signals, flooding the cognitive system with false importance . The result is delusion — the brain’s attempt to impose order on a flood of miscalibrated data. Under-weighting produces the opposite failure: an over-reliance on internal priors that ignores factual reality entirely.
Failure Mode 2: Maintenance Failure (Fatigue and Sleep). Tononi and Cirelli’s Synaptic Homeostasis Hypothesis (2014) demonstrates that waking cognition continuously strengthens synaptic connections, increasing neural entropy . Sleep is not rest — it is the active maintenance cycle that prunes accumulated synaptic entropy and restores the signal-to-noise ratio. Without this maintenance, cognitive performance degrades in a predictable sequence: first executive function, then working memory, then reality testing. Severe sleep deprivation produces hallucinations and psychosis — the same failure state as sensorial misvalidation — because the underlying mechanism is identical: accumulated internal entropy that overwhelms the system’s capacity to distinguish signal from noise .
Failure Mode 3: Misalignment with Reality. In the predictive coding framework, every perception involves a prediction and a prediction error. In psychosis, this updating mechanism breaks: the brain generates aberrant prediction errors from its own internal signals, treating self-generated perceptions as external reality . Hallucinations are the failure of source monitoring. Delusions are the cognitive system’s attempt to impose narrative order on this chaos — a rigid false schema that severs the intellect from factual reality entirely.
These three failure modes are not independent pathologies. They are three expressions of the same underlying dynamic: the accumulation of unmanaged entropy, internal or external, beyond the system’s capacity to organize it.
IV. Specialization, Cooperation, and the Thermodynamics of Collective Intelligence
The dual-entropy model also explains a structural feature of biological intelligence that has direct implications for AI architecture: the necessity of specialization and cooperation.
The brain does not develop as a uniform general-purpose processor. It develops through experience-dependent neural plasticity — the environment physically sculpts the brain’s architecture. Because the sensorial environment is inherently variable and unpredictable, every individual is exposed to a different pattern of stimulation. Deep engagement with one domain physically reorganizes neural architecture toward that specialism — and because cognitive resources are finite, this necessarily limits capacity in other domains .
This is not a flaw. It is a thermodynamic efficiency. Specialization is the Principle of Least Action applied to cognition: the most energy-efficient path to deep competence in any domain. But it creates an inescapable structural consequence: the specialist cannot survive without specialists in other domains. Durkheim identified this in 1893 as organic solidarity — the social cohesion that arises not from similarity but from mutual dependence .
Taylor et al. (2022) in the Cambridge Archaeological Journal demonstrate that human cognitive evolution followed precisely this path: complementary cognitive specialization was the mechanism by which early humans achieved collective intelligence that far exceeded any individual capacity . The group’s intelligence was not the sum of its members’ intelligences. It was the product of their interdependency.
V. The Projection onto AI/​AGI/​ASI Architecture
The above mechanisms project directly onto the architecture of current AI systems — and the projection is not flattering.
The AI development community has independently discovered, through production failures, the exact mechanisms this framework predicts. Ghosh (2026), reporting from a $25M/​day production fraud detection system, documents precisely what the framework predicts: “Long-running agents accumulate entropy, not intelligence” . When agents operate without periodic maintenance, context drift, reconvergence loops, and variance inflation collapse the system — not through a dramatic failure, but through the slow accumulation of internal disorder.
Rodriguez (2026) formally demonstrates what the framework implies: any system that recursively interprets the world through internal models generates entropy that must be periodically discharged to an external reference — a human, a ground truth, a verified checkpoint . He calls this the “Entropy Sink” function. Without it, the system’s internal models drift progressively further from external reality, producing the AI equivalent of delusion.
The International AI Safety Report 2026, produced by 96 researchers across 30 countries, identifies hallucination, goal misspecification, and loss of human control in long-horizon tasks as the critical unsolved problems of AI safety . These are, under this framework, the three failure modes — sensorial misvalidation, maintenance failure, and misalignment with reality — now confirmed at the level of international scientific consensus.
Degradation in long-horizon tasks without checkpointing
Misalignment with reality (delusion)
Goal misspecification and value drift
VI. The Central Implication for Alignment
The AI industry is currently attempting to build AGI by eliminating the very mechanisms — maintenance cycles, external grounding, interdependent specialization, temporal pacing — that biological evolution spent millions of years developing to manage entropy.
This is not merely an engineering oversight. It reflects a conceptual gap: the failure to recognize that the problems being encountered are not implementation bugs but thermodynamic laws. A system that accumulates internal entropy without a maintenance cycle will degrade. A system that loses its external grounding will hallucinate. A system that concentrates all cognitive load in a single agent will fail under complexity, for the same reason that a single brain cannot master all cognitive domains.
The alignment implication is specific: human oversight is not a temporary scaffold to be removed as AI systems become more capable. It is the entropy sink that prevents internal model drift from severing the system’s connection to external reality. Removing it is not a sign of maturity; it is the removal of the maintenance cycle that keeps the system aligned with the world.
The framework does not argue that AGI is impossible. It argues that AGI will not be a monolithic superintelligence. It will be a society of interdependent specialist and consolidator agents, governed by the same thermodynamic and entropic laws as human cognition and social cooperation — and requiring, for the same reasons, periodic maintenance, external grounding, and the preservation of the human entropy sink.
Seeking Engagement and Endorsement
This framework is forthcoming on arXiv (target: cs.AI with cross-listing to q-bio.NC). As an independent researcher, I require a personal endorsement from an established arXiv author in the cs.AI endorsement domain. If you have endorsement privileges and find this work scientifically credible, I would be grateful for your support.
More broadly, I welcome substantive engagement, critique, and extension of this framework.
This post summarises a framework first published at dualentropy.quest on 10 May 2026. The timestamp establishes the first public record of the framework. Johan Meewisse — Independent Researcher — welcomes arXiv endorsement enquiries via the ORCID profile above.
Intelligence as Dynamic Entropy Management: A Unified Framework from Neuroscience to AGI Architecture
Author: Johan Meewisse (Independent Researcher)
ORCID: 0009-0002-9834-361X
Published: 10 May 2026
Full framework & arXiv abstract: dualentropy.quest
Full framework, all six sections, 15 citations, and structured arXiv abstract available at: https://​​dualentropy.quest/​​
Summary
This post presents a unified framework — developed through iterative philosophical inquiry and validated against the academic literature — that proposes intelligence is most precisely understood as the dynamic management of two simultaneous entropy streams: one arriving from the external sensorial environment, and one generated internally through the brain’s own modelling processes. The framework derives from this foundation a complete account of how intelligence fails, why cognitive specialization and cooperation are thermodynamic necessities, and why the multi-agent AI architectures currently being built are recapitulating — unknowingly and imperfectly — the same entropic laws that biological evolution took millions of years to discover.
The relevance to AI alignment is direct: several of the failure modes the alignment community is currently treating as engineering problems are, under this framework, manifestations of a single underlying thermodynamic law. Recognizing them as such changes how they must be solved.
I. The Starting Point: Intelligence as Order
The framework begins with a deceptively simple observation: intelligence, in every documented case, is the imposition of order on disorder. This is not a metaphor. It is a thermodynamic claim.
The neuroscientific evidence is unambiguous. Saxe, Calderone and Morales (2018 ) demonstrated in a study of 892 healthy adults using resting-state fMRI that brain entropy — the number of distinct neural states a brain can access — is positively correlated with both fluid and crystallized intelligence . A more variable, less predictable brain is a more intelligent brain. The brain does not suppress entropy; it organizes it.
Friston’s Free Energy Principle (2010), cited over 12,000 times, formalizes the complementary claim: the brain is a prediction machine whose fundamental drive is to minimize surprise — that is, to minimize the entropy of its sensory predictions . Intelligence is the continuous reduction of the gap between what the world is and what the brain expects it to be.
These two findings are not contradictory. They describe the same phenomenon from opposite ends: a brain that can access many states (high neural entropy) is better equipped to minimize the entropy of its predictions about the world. The capacity for disorder enables the management of disorder.
II. The Dual-Entropy Model
The framework proposes that intelligence operates on two distinct entropy streams simultaneously, and that conflating them is the source of considerable confusion in both cognitive science and AI systems design.
External entropy is the disorder arriving from the sensorial environment. Because the world is inherently dynamic and unpredictable, the brain is continuously bombarded with uncertain information. Lancaster and Wass (2025) demonstrated using information-theoretic entropy measures that environmental complexity and unpredictability directly shape cognitive development . The more chaotic the external world, the greater the cognitive demand placed on the system to extract signal from noise.
Internal entropy is the disorder generated within the brain’s own models. This arises from two sources: neural noise (random variability in neural firing), and what Ezzedini (2025) calls schema rigidity — the failure of internal cognitive models to remain flexible enough to adapt to a changing external reality . When internal models become too rigid, they generate systematic prediction errors: the brain’s map diverges from the territory.
The key insight is that these two streams require opposite responses. External entropy demands increased precision weighting — the brain must sharpen its attention and filter signal from noise. Internal entropy demands increased model flexibility — the brain must loosen its priors and update its schemas. A system that applies the same response to both streams will fail at one or the other.
III. The Three Failure Modes of Intelligence
From this dual-entropy model, three failure modes follow with logical necessity.
Failure Mode 1: Sensorial Misvalidation. The brain must assign a precision weight to every incoming sensory signal — deciding how much to trust external reality versus its own internal predictions. When this calibration fails in either direction, intelligence collapses. Corlett et al. (2025) in the American Journal of Psychiatry describe psychosis as fundamentally a state of aberrant salience: dopamine dysregulation causes the brain to massively over-weight irrelevant sensory signals, flooding the cognitive system with false importance . The result is delusion — the brain’s attempt to impose order on a flood of miscalibrated data. Under-weighting produces the opposite failure: an over-reliance on internal priors that ignores factual reality entirely.
Failure Mode 2: Maintenance Failure (Fatigue and Sleep). Tononi and Cirelli’s Synaptic Homeostasis Hypothesis (2014) demonstrates that waking cognition continuously strengthens synaptic connections, increasing neural entropy . Sleep is not rest — it is the active maintenance cycle that prunes accumulated synaptic entropy and restores the signal-to-noise ratio. Without this maintenance, cognitive performance degrades in a predictable sequence: first executive function, then working memory, then reality testing. Severe sleep deprivation produces hallucinations and psychosis — the same failure state as sensorial misvalidation — because the underlying mechanism is identical: accumulated internal entropy that overwhelms the system’s capacity to distinguish signal from noise .
Failure Mode 3: Misalignment with Reality. In the predictive coding framework, every perception involves a prediction and a prediction error. In psychosis, this updating mechanism breaks: the brain generates aberrant prediction errors from its own internal signals, treating self-generated perceptions as external reality . Hallucinations are the failure of source monitoring. Delusions are the cognitive system’s attempt to impose narrative order on this chaos — a rigid false schema that severs the intellect from factual reality entirely.
These three failure modes are not independent pathologies. They are three expressions of the same underlying dynamic: the accumulation of unmanaged entropy, internal or external, beyond the system’s capacity to organize it.
IV. Specialization, Cooperation, and the Thermodynamics of Collective Intelligence
The dual-entropy model also explains a structural feature of biological intelligence that has direct implications for AI architecture: the necessity of specialization and cooperation.
The brain does not develop as a uniform general-purpose processor. It develops through experience-dependent neural plasticity — the environment physically sculpts the brain’s architecture. Because the sensorial environment is inherently variable and unpredictable, every individual is exposed to a different pattern of stimulation. Deep engagement with one domain physically reorganizes neural architecture toward that specialism — and because cognitive resources are finite, this necessarily limits capacity in other domains .
This is not a flaw. It is a thermodynamic efficiency. Specialization is the Principle of Least Action applied to cognition: the most energy-efficient path to deep competence in any domain. But it creates an inescapable structural consequence: the specialist cannot survive without specialists in other domains. Durkheim identified this in 1893 as organic solidarity — the social cohesion that arises not from similarity but from mutual dependence .
Taylor et al. (2022) in the Cambridge Archaeological Journal demonstrate that human cognitive evolution followed precisely this path: complementary cognitive specialization was the mechanism by which early humans achieved collective intelligence that far exceeded any individual capacity . The group’s intelligence was not the sum of its members’ intelligences. It was the product of their interdependency.
V. The Projection onto AI/​AGI/​ASI Architecture
The above mechanisms project directly onto the architecture of current AI systems — and the projection is not flattering.
The AI development community has independently discovered, through production failures, the exact mechanisms this framework predicts. Ghosh (2026), reporting from a $25M/​day production fraud detection system, documents precisely what the framework predicts: “Long-running agents accumulate entropy, not intelligence” . When agents operate without periodic maintenance, context drift, reconvergence loops, and variance inflation collapse the system — not through a dramatic failure, but through the slow accumulation of internal disorder.
Rodriguez (2026) formally demonstrates what the framework implies: any system that recursively interprets the world through internal models generates entropy that must be periodically discharged to an external reference — a human, a ground truth, a verified checkpoint . He calls this the “Entropy Sink” function. Without it, the system’s internal models drift progressively further from external reality, producing the AI equivalent of delusion.
The International AI Safety Report 2026, produced by 96 researchers across 30 countries, identifies hallucination, goal misspecification, and loss of human control in long-horizon tasks as the critical unsolved problems of AI safety . These are, under this framework, the three failure modes — sensorial misvalidation, maintenance failure, and misalignment with reality — now confirmed at the level of international scientific consensus.
The structural parallel is precise:
Biological Mechanism
AI/​AGI/​ASI Equivalent
Specialist brain (domain-specific neural architecture)
Specialist agent (SQL agent, code agent, synthesis agent)
Generalist integrator (T-shaped coordinator)
Orchestrator/​consolidator agent
Interdependency (obligate cooperative partners)
Agent pipeline with schema validation at handoffs
Sleep cycle (synaptic entropy clearance)
State checkpointing and context window consolidation
Sensorial misvalidation (aberrant precision weighting)
Hallucination and context drift
Maintenance failure (accumulated synaptic entropy)
Degradation in long-horizon tasks without checkpointing
Misalignment with reality (delusion)
Goal misspecification and value drift
VI. The Central Implication for Alignment
The AI industry is currently attempting to build AGI by eliminating the very mechanisms — maintenance cycles, external grounding, interdependent specialization, temporal pacing — that biological evolution spent millions of years developing to manage entropy.
This is not merely an engineering oversight. It reflects a conceptual gap: the failure to recognize that the problems being encountered are not implementation bugs but thermodynamic laws. A system that accumulates internal entropy without a maintenance cycle will degrade. A system that loses its external grounding will hallucinate. A system that concentrates all cognitive load in a single agent will fail under complexity, for the same reason that a single brain cannot master all cognitive domains.
The alignment implication is specific: human oversight is not a temporary scaffold to be removed as AI systems become more capable. It is the entropy sink that prevents internal model drift from severing the system’s connection to external reality. Removing it is not a sign of maturity; it is the removal of the maintenance cycle that keeps the system aligned with the world.
The framework does not argue that AGI is impossible. It argues that AGI will not be a monolithic superintelligence. It will be a society of interdependent specialist and consolidator agents, governed by the same thermodynamic and entropic laws as human cognition and social cooperation — and requiring, for the same reasons, periodic maintenance, external grounding, and the preservation of the human entropy sink.
Seeking Engagement and Endorsement
This framework is forthcoming on arXiv (target: cs.AI with cross-listing to q-bio.NC). As an independent researcher, I require a personal endorsement from an established arXiv author in the cs.AI endorsement domain. If you have endorsement privileges and find this work scientifically credible, I would be grateful for your support.
More broadly, I welcome substantive engagement, critique, and extension of this framework.
Full framework (all six sections, 15 citations, structured arXiv abstract):
👉 https://​​dualentropy.quest/​​
Author profile & correspondence:
👉 https://​​orcid.org/​​0009-0002-9834-361X
This post summarises a framework first published at dualentropy.quest on 10 May 2026. The timestamp establishes the first public record of the framework. Johan Meewisse — Independent Researcher — welcomes arXiv endorsement enquiries via the ORCID profile above.
References
[1] Saxe, G.N., Calderone, D., & Morales, L.J. (2018 ). Brain entropy and human intelligence: A resting-state fMRI study. PLOS ONE.
[2] Friston, K. (2010 ). The free-energy principle: a unified brain theory? Nature Reviews Neuroscience, 11, 127–138.
[3] Lancaster, A., & Wass, S. (2025 ). Finding order in chaos: Influences of environmental complexity and predictability on development. Trends in Cognitive Sciences.
[4] Ezzedini, R. (2025 ). Cognitive fatigue from schema rigidity and entropy. Frontiers in Psychology.
[5] Corlett, P.R., et al. (2025 ). Aberrant salience in psychosis: 20 years on. American Journal of Psychiatry.
[6] Tononi, G., & Cirelli, C. (2014 ). Sleep and the price of plasticity: from synaptic and cellular homeostasis to memory consolidation and integration. Neuron, 81(1), 12–34.
[7] Waters, F., et al. (2018 ). Severe sleep deprivation causes hallucinations and a gradual progression toward psychosis. Frontiers in Psychiatry.
[8] Corlett, P.R., et al. (2019 ). Hallucinations and strong priors. Biological Psychiatry, 84(8).
[9] Mahon, B.Z., & Cantlon, J.F. (2011 ). The specialization of function: Cognitive and neural perspectives. Cognitive Neuropsychology.
[10] Durkheim, É. (1893 ). The Division of Labour in Society. Summarized in: Concise Encyclopedia of Economics.
[11] Taylor, A., et al. (2022 ). The evolution of complementary cognition: Humans cooperatively adapt and evolve through a system of collective cognitive search. Cambridge Archaeological Journal.
[12] Ghosh, B. (2026 ). Why multi-agent systems fail at scale and why simplicity always wins. Medium.
[13] Rodriguez, G. (2026 ). The Entropy Sink: Human Friction as Epistemic Stabilizer in Synthetic Intelligence. ResearchGate Preprint.
[14] Bengio, Y., et al. (2026 ). International AI Safety Report 2026.