Thomas Parr – Markov Blankets from Cells, to Machines, to Planets (Stewarding the Flame, Episode 11)

This new installment of the Stewarding the Flame series is an interview with Thomas Parr, clinician and neuroscientist at the Nuffield Department of Clinical Neurosciences, University of Oxford, and co-author, with Giovanni Pezzulo and Karl Friston, of Active Inference: The Free Energy Principle in Mind, Brain, and Behavior.

What distinguishes Thomas’s approach is that he comes at the question of life through the free energy principle. For Thomas, a sustained boundary between a living system and its world is one of the things that helps define life. Wherever we can draw that boundary, he says, we can write down a model that predicts what happens on it. That gives him a question he returns to throughout the conversation: what model does this creature, collective, or synthetic system use? It also leads him to see life as a non-equilibrium process, one that keeps turning without settling into a fixed point.

In this conversation, he explains how internal models underwrite both perception and action, explores how Markov blankets nest from cells to organisms to collectives, and discusses whether the same principles can apply to synthetic systems. He also argues that, if the living process is to be sustained, AI may do well to learn from the brain’s low-energy architecture rather than relying on ever-greater scale.

The interview is the eleventh installment in The Trajectory’s Stewarding the Flame series, where we ask: What is intelligence, and what is the “flame” that life has, which non-life does not?

This series references the article: Stewarding the Flame – How to Build an Ideal Future for Intelligence.

I hope you enjoy this conversation with Thomas Parr.

Subscribe for the latest episodes of The Trajectory:

Below, we’ll explore the core takeaways from my conversation with Thomas, focusing on three of the central questions that shape the Stewarding the Flame series: what makes life, life, what humanity still needs to discover in order to better understand the living process, and what we should and shouldn’t do with AGI and BCI for the process of life to flourish.

What Makes Life, Life?

For Thomas, one of the things that defines life is a sustained boundary that separates a living system from something else. He doesn’t think that’s a complete definition, but it matters for how he thinks about the brain. The way we perceive and act on the world depends on there being “an us and a world.” If that boundary holds over time, we can describe the system from its own perspective.

As Thomas puts it, “We all sort of make use implicitly of internal models, generative models that predict what’s going on in our outside world. And those models play a dual role.” When sensory data doesn’t match the model’s predictions, the model is revised, which is perception. But a creature can also act on the world until it fits the model. The free energy principle scores that model fitness with a quantity called free energy, and minimizing it gives a simple way of characterizing both perception and action. The key scientific question then becomes: what is the model the creature is using?

He calls the interface a Markov blanket. The model predicts what happens on that interface. These boundaries can nest. Within an animal, organ systems, tissues, cells, and subcellular organelles each draw their own boundaries, and wherever a boundary can be drawn, a model could be written down that predicts what happens on it. For an individual cell, the rest of the body is part of the external world to be modeled. The same maths can be applied at the level of natural selection, which he suggests can be read as a process of inference or model selection, finding the population that best fits an environment.

What separates life from a stone dropping to Earth, Thomas argues, is that living systems don’t simply come to rest. A system that tends towards a fixed point and stops there is not one we’d recognize as life, or as a flame. He wonders whether the solenoidal, conservative dynamics keep the system moving around what would otherwise be a fixed point, “never necessarily meeting exactly the same point again.” He offers an almost definitional way to distinguish equilibrium from non-equilibrium: an equilibrium system looks the same played forwards or backwards in time, while a non-equilibrium system does not.

“one of the things that I think really defines life is a sort of sustained boundary that separates a living system from something else.”

What Do We Still Need to Discover to Understand the Flame?

For Thomas, the open scientific questions about life largely come down to one: what is the model a particular kind of creature uses, and what distinguishes one creature’s model from another’s? One dimension is hierarchy. In the visual system, retinal cells detect simple patterns of light, different patterns emerge in the thalamus, and sensitivity to the orientation of lines emerges in the visual cortex. Further up the hierarchy we recognize objects and people, so that a person’s name and the sight of their face can convey similar information. Predictions also travel back down the hierarchy: recognizing someone can shape what we expect to see or hear. Higher levels of abstraction tend to run at slower timescales, which implies more prospection and planning towards a distant future.

A second dimension is horizontal factorization, carving the world at its joints into different kinds of variable. Thomas compares it to giving a location as latitude and longitude rather than a single number for every point on Earth: factorizing makes generalization far more efficient. Brains seem to do the same, separating “what” from “where,” so that an apple is still an apple and tastes the same no matter which bit of space it occupies. These dimensions, he suggests, form a hypothesis space for studying creatures, collectives, and synthetic organisms alike.

He also cautions against reading a boundary as necessarily spatial. He describes the brain as a collective of neurons, glia, and vasculature, bounded by the way it works together and interacts with other things. A company could be approached the same way: its public-facing roles hold an interface with the outside world that persists even as personnel change, and the rest of the company effectively supplies the predictions. He suggests our sense of three-dimensional space may simply reflect where most interactions happen, and that with technology we may be dealing with an entirely different, configurational kind of space.

Part of what draws Thomas to the study of life is that it seems to hold the most interesting resolvable uncertainty: complex, yet amenable to simple principles, and unfolding on timescales we can study. Most of his time is now spent in clinical neurology, and he points out that neurology has long advanced by studying how things go wrong. He believes the same approach could be applied to the pathologies of artificial intelligence, large language models, and robotic systems, an approach he sees as starting to gain traction but less emphasized than efforts to expand and develop these systems.

“…what is the model that this creature, this collective, this synthetic organism use implicitly that allows us to describe and understand their behavior.”

What should and shouldn’t we do with AGI and BCI if we want the best chance for the greater process-of-life to flourish?

On substrate, Thomas takes a nuanced position. As a neurobiologist, he believes the particular thermodynamic constraints and biological setup of the brain matter, since things like the sparsity of neural spiking likely embody prior beliefs we have evolved about the world. Yet he sees the free energy principle and active inference as applicable to a whole range of systems, including synthetic creatures, provided they are at, or tending towards, a non-equilibrium steady state with their environment and have a boundary that allows them to be distinguished from it.

One difference he flags between biological life and some emerging technologies is mortality. Most forms of biological life, perhaps all, are temporary: the substrate of a biological organism is such that it is born and dies. Some of the systems now being built could, in principle, be immortal, and Thomas wonders whether losing that transience would change something fundamental, while leaving its importance as a question for philosophers. He connects it back to cycling: a life cycle in which something picks up after a death and continues seems to be an important part of the process.

Thomas sees both risk and promise. Extending the flame metaphor, he warns of a forest fire scenario, in which ever-increasing resources go to AI at the expense of human and animal communities. He points to current controversies over data centers, energy use, and the potential for worsening climate change. He doesn’t rule out the possibility that AI could help solve climate change. The gamble, he says, is letting energy use keep rising without a sustainable way to support it, and risking a catastrophe in the process. On the other side, he notes that brain-computer interfaces such as deep brain stimulators for Parkinson’s disease, tremor disorders, and epilepsy, along with cochlear implants, already do a huge amount of good. If technology becomes a helpful prosthesis that enhances our capabilities, he suggests it may be part of the solenoidal dynamics that let a system explore new states while remaining recognizably itself.

For innovators and regulators, Thomas’s closing recommendation is skepticism about scale as the answer. He expects large function approximators with enough parameters and energy to yield results, but finds it striking that the brain engages with the world at far lower energy consumption. Sustainability, he argues, has been implicit throughout the conversation: if the process is to continue long term, it must be sustained by something. He would like to see more technologies focused on what can run at low energy on a device itself, and more attention to how nature has solved these problems over much longer timescales.

Concluding Notes

Thomas sees a sustained boundary as one feature of life. He also describes how a living system keeps moving rather than settling into a fixed point. He wonders whether the need to keep moving rather than settle into equilibrium might be one reason for curiosity. Under active inference, choosing between future worlds involves expected free energy, which includes not only satisfying preferences but also information gain, the drive to find out what we don’t yet know. At the scale of the living process, he speculates, the same mechanics might appear as open-endedness: a drive for change and continuation.

For Thomas, the question of whether new technologies carry that process forward is not settled, and he is wary of anyone who claims it is. A prosthesis that enhances our capabilities could become part of life’s cycling; a resource-hungry race could instead burn through what the living process stands on. His recommendation is to look closely at how nature solved these problems, and at the low-energy architecture inside our own skulls.

For me, this gets to the heart of the Stewarding the Flame series. Before we can responsibly build increasingly capable forms of intelligence, we must first deepen our understanding of the living process itself. We cannot steward what we do not more deeply understand.

I hope you’ll join me for the next conversation in the Stewarding the Flame series as we continue exploring what intelligence is, what makes life, life, and what it will take to carry the greater process of life successfully into the future.

Follow The Trajectory