fullmetal_decade

The observation starts from both parts and function: what the components are, how they are connected, what dynamics they generate, and what those dynamics do. In the brain that means cell types, connectivity and microcircuitry, but also ensemble activity, state, context and computation. In cells and tissues it means channels, coupling, fields and patterns, but also control, memory and repair.

The central question has stayed remarkably stubborn:

How does a physical system acquire, retain, transform, and use information in order to control what happens next?

The path I keep taking through that question is:

The last arrow matters as much as the first. I am not only trying to understand biological computation. I want the theory to become constructive enough to guide the design of new computational architectures — including ones that can be built as physical devices — and then use those constructed systems to ask better questions of biology.

That loop is the decade.

Start with the parts — and what they do

My way into this problem was the cortex.

The early question was not “is the brain a computer?” It was much less forgiving: what actually organizes the activity of large populations of neurons, and what biological machinery produces that organization?

That means refusing to separate anatomy from function too early. A cortical theory has to care about which cells are excitatory or inhibitory, what kinds of inhibitory cells they are, how populations are connected, how thalamic and cortical loops interact, and what happens to ensemble dynamics across wakefulness, sleep and pathology.

The most direct version of that problem begins with the parts themselves. Morphoelectric properties of inhibitory neurons shift gradually and regardless of cell type along the depth of the cerebral cortex separates two sources of inhibitory-neuron diversity: a molecular cell-type-specific structure that is largely depth-independent, and a gradual depth-dependent modulation shared across inhibitory types. The point is not simply to classify interneurons, but to understand which properties belong to type and which belong to the cortical environment in which that type is embedded.

From there, the question moves from parts to collective function. In Dynamic Balance of Excitation and Inhibition in Human and Monkey Neocortex, excitatory and inhibitory populations co-fluctuate across many temporal scales and across the wake–sleep cycle, while seizure disrupts that coordination. In Ensemble inhibition and excitation in the human cortex, the same system is asked a different question: how much of the collective statistics can a pairwise maximum-entropy model actually capture, and what goes wrong when inhibition is ignored? A Computational Perspective of the Role of the Thalamus in Cognition pushes beyond a relay picture and asks how thalamocortical loops can reconfigure cortical computation according to context.

This is what I mean by neurophysics grounded in biology.

I am interested in statistical physics, nonlinear dynamics, criticality and scaling of neural activity, but I do not want critical, balanced, collective, or emergent to become free-floating descriptions. I am not interested in placing a beautiful physics frame around a biological system and stopping there. The tools are useful because they let us dig beneath those words: identify the degrees of freedom, find invariances, locate the order parameters, ask what breaks the symmetry, and then connect the effective description back to the cells, circuits and interactions that generate it.

The physics has to remain attached to the machinery.

The ruler is part of the theory

A second idea became impossible for me to avoid: the answer depends on the scale at which the question is asked.

That does not mean that anything goes. It means that scale itself has to become an experimental and theoretical variable.

The 2018 multiscale theory of neural control argued that an intervention should be matched to the computational level one actually intends to control. Causal Scale of Rotors in a Cardiac System asked at what spatiotemporal scale a macroscopic rotor genuinely has causal explanatory power. Symmetry’s Edge in Cortical Dynamics asks which features of excitation/inhibition coordination survive coarse-graining and which broken symmetries distinguish physiological and pathological states.

The newer work pushes the same issue further. How long is the coast of a neuron? asks how apparent interaction order and computational capacity change with the ruler used to describe a neuron. When is there a manifold? treats temporal integration scale as a coordinate of neural population geometry rather than a hidden preprocessing choice.

The question is not “which scale is the correct one?” It is:

What survives when the ruler changes? What reorganizes? What disappears?

And then: at which scale does a new variable become sufficiently stable, predictive and causally useful to deserve its own effective theory?

Information processing is not yet computation

At some point, calling everything an information-processing system stopped being satisfying. The phrase is too cheap unless it makes exclusions.

A wave carries information. A tissue can preserve a pattern. A dynamical system can remember its past. A rock has state transitions. None of those statements, by itself, tells us that the system is computing.

This is where the cellular work and the work on foundations of physical computation meet. Systematizing Cellular Complexity asks what kinds of problems a cell must actually solve — control, reconfiguration, robustness, adaptation, noise management — and treats the formalization of the right problems as part of the scientific task. Physical Computing asks for a stricter bridge between physical processes and abstract computation, with compositional structure replacing an arbitrary labeling of physical states.

The hierarchy that has emerged from the newer work is deliberately demanding:

Physical memory is not yet physical computation, and physical computation is not yet autonomous computation.

A physical system can store a trace without computing with it. A system can implement a robust coarse-grained transition without internally choosing which transition comes next. Autonomous computation needs something stronger: an internal physical readout must participate in selecting the next physical operation.

That is closure. And increasingly, closure has become the word around which the rest of the program organizes itself.

Recurrence, recurrence, recurrence

Long before I had the word closure for it, I kept finding the same structure.

The thalamus does not simply relay cortical information; it participates in a loop that changes the effective computational regime of cortex according to context. Excitation and inhibition stabilize one another dynamically. A reservoir transforms an input because its present state contains its past. A bioelectric tissue persists because local state changes coupling, and coupling changes the future of local state. A neuron writes fields that act back on the neuron. A recurrent network carries state forward rather than reconstructing the past from scratch.

The common architecture is recurrence. Not merely a loop drawn on a circuit diagram, but a physical mechanism by which history becomes state, state becomes context, context changes the next transformation, and the consequences return to alter the future state.

That is why I keep coming back to it:

The forthcoming Computational Irreducibility of Time paper makes one part of this quantitative: for multiplicative temporal structure, persistent recurrent state can realize computations whose fixed-depth flattened implementations pay a growing resource cost. The point is not that every temporal computation requires recurrence. The point is to identify exactly what recurrence buys, and when it buys it.

Cells and tissues are not a detour

The brain is an extraordinary laboratory, but it is not always the easiest place to isolate first principles. It contains too many interacting cell types, scales, feedback loops and effective layers to make every conceptual distinction clean.

Cells and tissues provide a complementary laboratory. They let us ask stripped-down questions about memory, pattern, control and computation in systems where the components can sometimes be written explicitly enough to build mathematical frames around them. A cell has internal state, sensing, actuation and homeostatic control. A tissue adds coupling, collective states, spatial pattern and repair. These systems can expose the minimal ingredients of computation more cleanly than a whole brain, while still remaining genuinely biological.

That is why the bioelectric work matters to the same program. Self-insulating domains asks what physically holds a tissue-scale pattern memory. The Middle Way asks when a membrane-voltage pattern becomes a relatively autonomous mesoscopic computational layer rather than merely a coarse description of channels underneath it. Field closure, ice neurons, and when a dendrite is a motif asks what separates a branching object that looks neuronal from a biological neuron whose internal state and self-generated field participate in a dynamical loop.

Ice can grow a dendritic silhouette. That is precisely why the distinction is useful. A shape is not a computation. Excitability is not yet autonomous computation. The interesting object appears when morphology, state, field, readout and feedback become part of the same causal architecture.

Cells and tissues are therefore not side projects away from neuroscience. They are simpler arenas in which to expose the basis of patterns, memory and computation — and then carry those principles back into the brain.

Prediction, internal models, agency

Control is future-oriented.

To control a changing system, an organism needs internal states that carry information about what is likely to happen next and about the consequences of possible actions. That is the sense in which I use internal model: not necessarily a symbolic representation, and not a little scientist hidden inside the organism, but a physical state whose structure is useful because it predicts enough of the world to guide control.

Prediction and internal model therefore go hand in hand. A cell uses them in the service of viability. A tissue uses them to maintain and repair a pattern. A nervous system extends the same logic over richer spaces — body, environment, other agents, possible actions, future consequences.

My view is that agency and intelligence grow out of this coupling between prediction and control. I would go further: I think consciousness belongs on the same physical continuum. I do not mean that as a spooky extra ingredient added above the biophysics. I mean that whatever effective layer we call consciousness has to sit on, and eventually be pinned back to, physical components, dynamics and causal organization. An effective theory is useful only if we can say what it is effective over and how it couples back to the machinery beneath it.

That is the wager running through this work: increasingly rich internal models, embedded in recurrent control loops, produce increasingly rich forms of agency.

Turn the arrow around

Then comes the part I care about just as much: build with the principles.

If biology has discovered useful computational organizations, the test is not whether we can describe them beautifully. The test is whether we can extract the operative principle, formalize it, and make it do work in another substrate.

That is the point of Bio-inspired AI. Biological intelligence is not interesting because neurons look like neural-network units. It is interesting because living systems are multiscale, context-dependent, recurrent, adaptive and organized around control.

The constructive work follows that line. In Cortical Blueprint RNNs, measured cortical geometry and functional organization become inductive biases for recurrent networks, and the biologically grounded models outperform weaker controls. In Adaptive Reservoir Computing, the recurrent substrate itself is placed under evolutionary selection for prediction of spatiotemporal chaos, exposing structural constraints rather than treating the reservoir as an arbitrary random matrix. All you need is recurrence! asks whether learning operations can live in recurrent state rather than in changing synaptic weights.

This is what I mean by true bio-inspired AI:

extract an organizing principle → formalize it → build with it → test it → bring the constructed system back to biology as a sharper probe.

The research program itself becomes recurrent: understand → build → use what was built to understand better → build again.

Circuits have to compose

There is still another missing level. Knowing what a primitive does does not tell us what happens when primitives are coupled.

Canonical neural motifs are usually assigned functional labels: this circuit normalizes, this one selects, this one gates. But a recurrent circuit exposed to sequences of inputs is not merely one function. It generates a repertoire of transformations.

“More Is Different” in Neural Circuits uses transition monoids, Krohn–Rhodes theory and holonomy decomposition to distinguish what is already present in a primitive update from what appears only under composition, and what belongs to one motif from what genuinely lives on the joint state space of coupled motifs.

This matters in both directions. For neuroscience, it gives us a way to move from saying that cortex contains motifs to asking what computations those motifs can generate together. For engineering, it points toward something closer to programming recurrent matter: choose primitives, interfaces and schedules so that the generated algebra has the repertoire you want.

The interface becomes part of the program. And that is also where effective theories become constructive rather than merely descriptive. A higher-level computational structure can belong to the composite even when no isolated part contains it.

More can genuinely be different.

The loop

The decade is therefore not a march from neuroscience toward AI. It is a loop.

1. observe parts + function
2. identify the relevant physical degrees of freedom
3. ask what survives across scale
4. find the effective state and what it predicts
5. separate information, memory and computation
6. find recurrence, feedback and closure
7. formalize transformations and composition
8. design new recurrent architectures and physical machines
9. use them to ask sharper questions of biology
10. ↺back to the parts

The cortex, a cell collective, a wave-memory system, a reservoir computer and a recurrent circuit motif are obviously not the same object.

The point is to develop a language strong enough to say how they differ, while still asking the same physical question of each.