When Friston Met Levin: The Free Energy Principle Goes Cellular

Series: Basal Cognition | Part: 3 of 11

In 2021, Karl Friston and Michael Levin published a paper together that should have broken the internet. It didn't—because most people still don't know what the Free Energy Principle means, and even fewer understand that cells might be doing it.

Here's what happened: The neuroscientist who thinks brains minimize variational free energy met the biologist who thinks cells minimize bioelectric gradients to maintain morphological coherence. They realized they were describing the same thing at different scales.

This isn't metaphor. This is the mathematical structure of how living systems persist.

What Friston formalized as active inference—systems maintaining their integrity by predicting and correcting deviations from expected states—Levin had been observing in cellular collectives rebuilding salamander eyes and planarian heads. The cells weren't just following genetic blueprints. They were minimizing prediction error about what the tissue should become.

The convergence matters because it reveals something fundamental: the same computational principles that explain how your brain predicts the next word in this sentence also explain how your cells predicted the shape of your face during embryonic development.

Both are systems minimizing surprise. Both are implementing Bayesian inference in biological substrate. Both are active inference all the way down.

What the Free Energy Principle Actually Says

Before we go cellular, let's get the core claim clear.

Karl Friston's Free Energy Principle states that any system that maintains its organization over time must minimize variational free energy—a quantity that bounds the surprise a system experiences when it encounters sensory states.

Translation: If a thing persists, it must resist entropy. To resist entropy, it must predict its environment well enough to avoid states that would dissolve it. To predict well, it must build and update internal models of the world.

This applies to bacteria avoiding toxins, brains predicting visual inputs, cells coordinating tissue repair, and immune systems distinguishing self from non-self. The principle is scale-free. It doesn't care whether you're a neuron or a cell collective or an organism. If you persist, you're minimizing free energy.

The math looks like this:

F = E[log p(s|m)] - H[q(m|s)]

Where F is free energy, s is sensory states, m is the model (hidden states), and the system tries to minimize the gap between what it predicts and what it receives.

In active inference, systems don't just passively update beliefs—they act on the world to bring sensory states into alignment with predictions. You don't just perceive the coffee cup. You reach for it, reducing the prediction error between "I expect to feel warmth" and "my hand is empty."

Now take that principle down to cells.

Cells as Bayesian Agents

Michael Levin's radical claim: Cells are not passive automatons executing genetic programs. They are autonomous agents making predictions about their local and global context, updating beliefs based on bioelectric signals, and acting to minimize the gap between expected and actual morphology.

This is active inference at the cellular scale.

Consider regeneration in planaria—flatworms that can rebuild entire heads or tails from almost any fragment. The cells don't have a blueprint for "head" stored locally. Instead, they read bioelectric gradients across tissue, compare them to an expected pattern, and adjust gene expression and migration to match the target morphology.

The bioelectric field acts as a prior—a probabilistic map of what the tissue should look like. Cells sample this field through gap junctions and voltage-gated ion channels, integrate the information, and take action (differentiate, migrate, proliferate) to reduce discrepancy.

This is variational inference. The cells are approximating the posterior distribution over "what kind of structure should I be part of?" given the sensory evidence (voltage gradients, chemical signals, mechanical forces).

When Friston and Levin published "Life as we know it" (2021), they formalized this convergence. They showed that morphogenesis—the process by which organisms take shape—can be understood as collective active inference where cellular ensembles minimize variational free energy over anatomical configurations.

Cells predict the target anatomy. They sense deviations. They act to correct them. This isn't just growth. It's goal-directed inference.

Markov Blankets at the Tissue Scale

One of the most powerful concepts linking Friston and Levin is the Markov blanket—the statistical boundary that defines where one system ends and another begins.

A Markov blanket partitions a system into:

• Internal states (hidden from the environment)
• External states (the environment)
• Sensory states (how internal states sample external ones)
• Active states (how internal states influence external ones)

At the cellular level, this maps cleanly onto biological structure:

• Internal states = intracellular biochemistry, gene regulatory networks
• External states = extracellular environment, neighboring cells
• Sensory states = ion channels, gap junctions, receptors
• Active states = secreted signals, bioelectric manipulation, migration

The Markov blanket isn't metaphorical. It's a precise statistical formulation of how cells remain distinct from their environment while still coupling to it.

But here's where it gets interesting: Markov blankets can nest.

A cell has a blanket. A tissue has a blanket composed of cells. An organ has a blanket composed of tissues. An organism has a blanket composed of organs.

At each scale, the system minimizes free energy relative to its sensory inputs. The heart doesn't need to know about geopolitics. The cell doesn't need to know about the heart. Each level minimizes surprise relative to its Markov blanket.

Levin's work shows that morphogenetic fields—the bioelectric patterns guiding development—act as higher-order priors that constrain cellular inference. The cells don't individually "know" the target anatomy, but the field encodes a low-dimensional representation that they collectively sample and enact.

This is hierarchical active inference: cells minimize free energy at the local level, while the tissue-level field minimizes free energy at the global level. The field emerges from cellular activity but also constrains it—a circular causality that Friston calls generalized synchrony.

Morphogenetic Computation as Predictive Processing

Let's make this concrete.

When a salamander loses a limb, the wound site exhibits a characteristic bioelectric signature—a depolarization gradient that signals "something is missing." Cells near the wound sense this deviation from the expected voltage pattern (the prior for "intact limb") and respond.

Stem cells migrate to the wound. Differentiation programs activate. Proliferation accelerates. The blastema—the regenerative structure—begins to form.

What's guiding this? Not a genetic blueprint that says "rebuild the limb in 14 steps." Instead, the cells are performing error correction. They're comparing the current bioelectric state to the target state and updating their behavior to minimize the discrepancy.

This is predictive processing: the system has a generative model of what the limb should look like, receives sensory evidence (voltage patterns, chemical cues), computes prediction error, and updates both beliefs (gene expression) and actions (migration, differentiation) to minimize free energy.

Levin has shown you can hack this process. Apply the right bioelectric stimulus, and you can induce planaria to grow heads where tails should be—or tadpoles to grow eyes on their tails. You're not directly programming morphology. You're reprogramming the prior—changing the target distribution the cells are trying to match.

This is the same computational architecture that governs perception and action in brains. When you see a face in the clouds, your visual cortex is matching sensory input to a learned prior. When cells build a face during embryogenesis, they're matching bioelectric input to an evolved prior.

Same algorithm. Different substrate.

Why Cells Use Bioelectricity (And Not Just Chemicals)

A reasonable question: Why bioelectricity? Why not just chemical gradients?

The answer: speed and coherence.

Chemical diffusion is slow—on the order of seconds to minutes across tissue-scale distances. Bioelectric signals propagate at millisecond timescales. For a cellular collective to coordinate morphogenetic decisions in real time, electrical coupling through gap junctions provides the bandwidth.

But there's a deeper reason. Bioelectric fields are integrable.

In information geometry, an integrable field is one where local measurements can reconstruct the global state. The voltage pattern across a tissue encodes a low-dimensional representation of the morphological target—a compressed prior that cells can sample and use for inference.

Chemical gradients encode positional information ("you are 300 microns from the anterior end"). Bioelectric fields encode topological information ("you are part of a structure with these symmetries and boundaries").

This is why Levin's work connects so cleanly to coherence geometry. The bioelectric field isn't just signaling. It's a geometric prior over morphological state space. Cells navigate that space by minimizing free energy, following geodesics toward target configurations.

In AToM terms: M = C / T. Meaning equals coherence over time (or tension). The cells maintain morphological coherence by minimizing variational free energy—the tension between predicted and actual states. The bioelectric field is the manifold they're traversing.

The Friston-Levin Convergence: What It Means

When Friston and Levin converged on active inference as the unifying principle, they weren't just drawing analogies. They were formalizing a claim: Life is computational.

Not in the sense of digital computation (bits and logic gates). In the sense of Bayesian computation—systems building models, making predictions, updating beliefs, and acting to minimize surprise.

This reframes biology. Instead of asking "what genes control development?" we ask "what priors guide morphogenetic inference?" Instead of "what signals trigger regeneration?" we ask "what prediction errors drive tissue repair?"

The difference isn't semantic. It shifts the causal locus from molecular mechanisms to information geometry.

Genes matter—but they matter as parameters in a generative model. Signals matter—but they matter as sensory evidence updating beliefs. The morphology isn't encoded in DNA. It's encoded in the low-free-energy configurations toward which cellular collectives converge.

This has practical implications:

1. Regenerative medicine: Instead of micromanaging gene expression, we can manipulate bioelectric priors to guide cells toward target anatomies. Levin's lab has already shown proof-of-concept—using ion channel drugs to induce tadpoles to regenerate eyes without genetic modification.

2. Cancer treatment: If cancer is coherence collapse—cells losing their morphological priors and reverting to disordered proliferation—then the treatment isn't just killing rogue cells. It's restoring the field that constrains them.

3. Synthetic biology: Instead of engineering organisms gene by gene, we can specify bioelectric targets and let cells infer the pathway. This is morphogenetic programming—top-down design via bottom-up inference.

The Friston-Levin framework doesn't replace molecular biology. It subsumes it into a larger computational picture.

Active Inference as the Deep Pattern

What makes active inference powerful isn't that it describes one phenomenon. It's that it describes all phenomena where systems persist.

Consider the range:

• A bacterium tumbling toward glucose is minimizing free energy over nutrient states.
• A brain hallucinating reality from sparse sensory data is minimizing free energy over perceptual hypotheses.
• A cellular collective building an eye is minimizing free energy over morphological configurations.
• An immune system distinguishing self from pathogen is minimizing free energy over identity classifications.

These aren't separate processes that happen to share a metaphor. They're implementations of the same variational principle in different substrates.

The math is scale-invariant. The mechanism is substrate-neutral. The principle holds from molecules to minds.

This is why Friston's work connects so productively with Levin's. It's not that cells are "like brains." It's that brains and cells are both self-organizing systems minimizing variational free energy under their respective Markov blankets.

The formalism doesn't care about neurons versus fibroblasts. It cares about systems that maintain organization by predicting states and acting to confirm predictions.

When those systems are hierarchically nested—cells within tissues within organisms—you get emergent coherence at each scale. The organism doesn't need to micromanage cellular behavior. It sets higher-order priors, and cells infer the rest.

This is collective intelligence without central control. It's swarm cognition. It's what happens when Bayesian agents couple through shared fields.

Why This Matters Beyond Biology

The Friston-Levin convergence has implications that extend beyond developmental biology and neuroscience.

If active inference is the principle governing all self-organizing systems, then we should expect to find it everywhere structure persists against entropy:

• Ecosystems maintaining homeostasis through predator-prey dynamics (minimizing free energy over population states)
• Economies stabilizing through price signals (minimizing free energy over resource allocations)
• Cultures transmitting norms through ritual and story (minimizing free energy over collective identity)

Each of these is a system with Markov blankets, sensory states, active states, and internal models. Each persists by predicting and acting to minimize surprise.

This isn't reductionism. It's pattern recognition. The same computational architecture appears at radically different scales because it solves the same fundamental problem: how to maintain organization in a dissipative universe.

The cells in your body are doing it. The neurons in your brain are doing it. The cultural practices in your society are doing it.

All the way down, all the way up: systems minimizing free energy, predicting their environments, acting to confirm those predictions.

The question isn't whether you're doing active inference. The question is what priors are guiding your predictions, and whether those priors still serve the coherence you're trying to maintain.

Where This Series Goes Next

We've established the bridge: cells are Bayesian agents minimizing variational free energy through bioelectric inference.

Next, we'll explore morphogenetic fields as Markov blankets—the statistical boundaries that partition developmental space and constrain cellular inference.

Then we'll examine what happens when that inference breaks down: cancer as coherence collapse, where cells lose their priors and revert to disordered proliferation.

And we'll look at systems that demonstrate just how flexible these priors can be: xenobots—cellular collectives repurposed from frog embryos into novel morphologies, still minimizing free energy, still enacting coherence, just toward radically different targets.

The pattern holds. The principle scales. Life is inference, all the way down.

Further Reading

Primary Sources:

• Friston, K., & Levin, M. (2021). "Life as we know it." Journal of the Royal Society Interface. (The paper that formalized the convergence)
• Friston, K. (2010). "The free-energy principle: a unified brain theory?" Nature Reviews Neuroscience. (The foundational FEP paper)
• Levin, M. (2021). "Bioelectric networks: the cognitive glue enabling evolutionary scaling from physiology to mind." Animal Cognition. (Levin's synthesis on bioelectric cognition)
• Fields, C., & Levin, M. (2020). "Scale-Free Biology: Integrating Evolutionary and Developmental Thinking." BioEssays. (On hierarchical Markov blankets and multi-scale coherence)

Accessible Introductions:

• Parr, T., Pezzulo, G., & Friston, K. (2022). Active Inference: The Free Energy Principle in Mind, Brain, and Behavior. MIT Press.
• Levin, M. (2022). "The Computational Boundary of a Self: Developmental Bioelectricity Drives Multicellularity and Scale-Free Cognition." Frontiers in Psychology.

For AToM Integration:

• The Free Energy Principle series (for deeper FEP mechanics)
• Coherence: The Geometry of Systems That Work (for AToM foundations)
• Active Inference: When Perception Becomes Action (for active inference specifics)

This is Part 3 of the Basal Cognition series, exploring how cellular collectives think, remember, and build coherent forms without central control.

Previous: Bioelectric Fields as Coherence Manifolds
Next: Morphogenetic Fields as Markov Blankets