Coach — Natural Intelligence

Coach — Natural Intelligence: Biology as the Original Inference Engine

Biology is natural intelligence: living things were running inference long before anyone built a computer, and today’s AI is, in important respects, an inferior imitation of what cells already do.

This hypothesis says the parallel between living systems and artificial intelligence is not a loose metaphor; it is structural. The environment supplies the data. Memory, held at many scales at once, supplies the working “weights.” Inference, the body’s constant work of guessing what is going on and correcting the guess, is the continuous minimization of surprise. And mitochondria, numbering in the quadrillions in a human body, are the substrate that makes that inference thermodynamically possible.

The environment as data

The ancestral environment is the evolutionary process that shaped the deepest priors of any living thing: the genome, developmental programs, and basal metabolic set-points. A “prior” here simply means a built-in expectation about the world. Intermittent fuel, solar and seasonal light cues, mechanical load from movement, and microbial partners selected the models that could keep themselves going under those conditions. Today’s environment is the ongoing stream of sensory, metabolic, and social inputs against which those priors are continuously tested and revised. The environment is therefore both the historical training data and the present interface that keeps the organism coupled to the world.

Memory at many scales

Memory in living systems is not confined to a single timescale or substrate. It works across nested layers: genetic, epigenetic and developmental, metabolic (including stable compensatory states), hormonal, synaptic and circuit-level, and cultural. These are not competing theories of memory; they are different time resolutions of the same underlying process of association and dissociation. Genes can be read as compressed archives of earlier solutions that once reduced surprise and were kept. Shorter-term layers update those archives in real time. Together they act as the effective weights of the organism’s model of its world.

Inference as surprise minimization

Karl Friston’s free energy principle proposes that any system that keeps its integrity over time must minimize a quantity called variational free energy. In plain terms, free energy here is an upper bound on surprise: how unexpected the incoming sensory data are, given the system’s internal model of what causes them. Lowering it is the same as lowering prediction error. Perception updates the model to better match the data; action changes the data to better match the model; learning revises the model itself. In this sense every living system is engaged in continuous approximate Bayesian inference, the mathematics of updating a best guess as evidence arrives. That, on this hypothesis, is the functional core of natural intelligence.

Mitochondria as the moving substrate

A human adult contains on the order of a quadrillion or more mitochondria. They are not simple settings like the weights of a neural network. Each one is a sophisticated dynamical system, with its own respiratory-chain complexes, folded inner membranes, local proton gradients, calcium handling, its own genome, and extensive signaling. Together they form a network whose number, quality, location, and state change continuously as mitochondria are built, recycled, fused, split, and moved around the cell.

The hypothesis reads mitochondria as the energy-based precision and capacity machinery that gates the whole inference hierarchy. The proton-motive force, the ATP/ADP ratio, and the NAD⁺/NADH redox state act like analog computing parts. They determine which inferences are energetically affordable, how heavily prediction errors are weighted, and how quickly the system can reconfigure. In neurons and other high-demand cells, mitochondrial shape and position are specific to the cell type and track activity, which suggests they are built into circuits rather than serving as generic power supplies.

Boundaries inside boundaries

A Markov blanket is the statistical boundary that separates the inside of a system from its outside; the two interact only through what the system senses and what it does. Living systems are made of nested blankets: the inner mitochondrial membrane, the cell, the tissue or neural circuit, the organism, and, at larger scales, social and cultural groups. Each one keeps its integrity by minimizing free energy with respect to the level above and below it.

This nesting is essential. Artificial systems typically run one or two levels of prediction inside a substrate that does not rebuild its own boundaries or its own energy-producing machinery. Natural intelligence continuously reconstructs both. Mitophagy and biogenesis, the recycling and building of mitochondria, are not peripheral housekeeping; they are the slow outer loop of model updating and pruning. When mitochondrial capacity declines, the range of coherent inference narrows, and the system is pushed toward coarser or more rigidly locked states.

Energy, cost, and experience

Landauer’s principle states that erasing information has a minimum energy cost. Reconfiguring associations, such as updating a model, consolidating memory, or dissolving a temporary coupling, therefore carries a thermodynamic price. Mitochondria are the primary engines that supply the free energy to pay it, while also reporting redox and calcium state to higher levels of organization.

The ontology underneath is strict monism: base reality is experience itself, an infinite field of graded association and dissociation. Consciousness is not a separate substance, and it is not a property added on top of physical processes; it is the graded coherence of experience across those association–dissociation dynamics in nested blankets. Surprise minimization is the process that maintains that coherence. Mitochondrial capacity sets its energetic limits. When capacity is robust, fine-grained, flexible coherence is possible. When capacity falls under chronic mismatch with ancestral conditions, the system is driven toward pathological dissociation or rigid, compensatory association, which the hypothesis calls the bioenergetic signature of many chronic diseases.

That is where EnSurf comes in. Practical support for mitochondrial capacity can be organized around three interlocking pillars: Fuel (ancestral substrates, mainly animal fats and proteins with fermented, microbiota-accessible carbohydrates), Circadian (natural light–dark and feeding–fasting rhythms), and Strength (resistance training and other hormetic loads that stimulate mitochondrial building and quality control). These are not arbitrary interventions; they are attempts to restore the energetic conditions the organism’s model was optimized for. This is a philosophy and working hypothesis, not medical advice; talk with your own physician before changing anything.

The microbiome, the scaffold, and continual learning

The microbiome works as an outside, co-evolved layer of metabolic memory and inference. Gut microbes turn food into signaling molecules and short-chain fatty acids that calibrate the host’s mitochondrial and immune function, so they act as extra parameters and an extra blanket the host must keep negotiating. The cytoskeleton and cell membranes supply the physical geometry of association; they are not passive scaffolds, which makes mechanical load and membrane lipid composition part of the inference machinery. And mitochondrial recycling, building, and trafficking look like biology’s version of continual learning: unlike offline training on a fixed architecture, they rebuild both the capacity and the location of the computing substrate in real time.

Why artificial intelligence falls short

Current AI systems implement sophisticated forms of hierarchical prediction and can show impressive scaling. They remain inferior imitations for structural reasons. They lack a body, open-ended evolutionary search over their own design, the continuous self-rebuilding of their own boundaries that biologists call autopoiesis, and the extreme energy efficiency of mitochondrial oxidative phosphorylation. None of this would make such a system conscious; encoding is not awareness. A human brain runs on roughly 20 watts; large artificial models need orders of magnitude more energy for narrower abilities. Prediction-based AI designs are genuine functional convergences, but they run on substrates that do not rebuild their own energy-producing machinery. Scaling improves performance inside a fixed paradigm; it does not produce the multi-scale, self-maintaining inference that biology performs.

The original intelligence

Natural intelligence is the original and more complete form of intelligence. The environment supplies the data and the interface, memory at many scales supplies the weights, surprise minimization is the objective, mitochondria supply the continually rewritten capacity, nested blankets provide the architecture, thermodynamics sets the cost, and experience, understood as graded association and dissociation, is the ground. The framing suggests that many chronic diseases are not random failures but stable compensatory modes entered when mitochondrial capacity can no longer support the organism’s model under modern conditions. It also suggests that the deepest continuity is not between silicon and carbon but among the living processes that maintain the coherence of experience against entropy. Artificial systems can model or imitate that coherence, but encoding is not awareness; they are not conscious.

A working hypothesis, open to correction and revised as new evidence arrives. Last revised October 17, 2026.

See the notice at the top of this page. Working hypothesis, not medical advice.