Coherence Dynamics Circa 1890

Ned Block's recent paper "Can only meat machines be conscious?" (Trends in Cognitive Sciences, 2025) poses a difficult dilemma for functionalism: if consciousness is just computation (roles), why does it seem so tied to biological "meat" (realizers)? He highlights a specific evolutionary puzzle: purely electrical nervous systems ("sparks") are fast and reliable, yet evolution heavily favored electrochemical systems ("soup") for complex cognition.

Block asks: What is special about electrochemical computation?

He considers inhibition, learning, and various biological factors, but finds them inconclusive. As he notes, the "War of the Soups and the Sparks" (Valenstein's phrase) was supposedly settled in 1921 when Otto Loewi showed chemical transmission existed---but we still don't know why evolution favored the messy, slow "soup" over the fast, reliable "sparks."

I want to propose a narrower answer: electrochemistry may enlarge the repertoire of states that can be reached, controlled, and stably coupled to an environment. Effective dimensionality is one way to measure part of that repertoire, but it is not the whole explanation.

What I mean by dimensionality: Not the number of neurons or the raw size of a state space. I mean a task- and scale-dependent estimate of how many independent modes a system reliably occupies or controls. Digital and electrochemical systems can both support high-dimensional dynamics. The empirical comparison concerns their reachable states, perturbation responses, energy use, latency, noise, and physical coupling under matched tasks.

The Wrong Metric: Speed Alone

If we view the brain as a digital computer, the evolutionary choice makes little sense. Electrical gap junctions are faster than chemical synapses. Biological systems also need plasticity, context-sensitive control, local memory, and robust coupling across many timescales. Speed is only one constraint.

The Transistor (Digital/AI): Digital hardware can represent high-dimensional vector spaces and simulate attractors. The open question is physical realization: energy use, latency, noise response, memory, and environmental coupling may differ from those of the system being simulated. Matching an abstract state-transition rule does not establish equivalence under every relevant intervention.

The Oscillator (Analog/Biology): An electrochemical system contains coupled oscillatory, diffusive, and biochemical processes. Its material dynamics can settle into phase-locked states or other attractors directly. That may make some control problems cheaper or more robust, but it does not make digital realization impossible in principle.

The testable thesis is narrower: for some adaptive control tasks, electrochemical oscillators may realize a richer intervention repertoire per unit energy, volume, or latency than current transistor architectures.

Chemical transmission exposes many control variables: transmitter and receptor composition, release probability, metabotropic cascades, uptake, plasticity, neuromodulation, and several time constants. Electrical synapses also have richer dynamics than an ideal resistor, including rectification, plasticity, and network-dependent effects. The hypothesis is comparative: chemical modulation may expand context-dependent reconfiguration in some circuits, often with different delays and energetic costs. That reachable-set advantage has to be measured rather than inferred from the word “soup.”

The "Transmission" Hypothesis (William James, 1898)

Start with William James. He famously argued that the brain does not produce consciousness (like a generator produces steam); it transmits it (like a prism transmits light or a lens focuses it).

The 1890 View: The brain is a "reducing valve" (to use Aldous Huxley's later phrase) that filters a larger consciousness down to a trickle.

The 2025 Update: In dynamical terms, the brain can be modelled as low-bandwidth controls steering a much larger physical state. This does not vindicate James's metaphysics, but it gives the transmission analogy an operational question: which distinctions and intervention responses survive the control interface?

The Selection Machine (Henri Bergson, 1896)

This is the strongest historical ally for coherence dynamics. In Matter and Memory, Bergson argued that the brain is not a storehouse of images (pixels), but a mechanism of "motor selection" acting on a universal field. (For an excellent exposition, see Stephen Robbins' video "Exploring Bergson's Holographic Theory".)

The 1890 View: Bergson described the brain's role as diminishing a richer field to a point of action, not generating experience from scratch. The brain selects; it does not create.

The 2025 Update: Bergson's "selection-for-action" picture maps well onto task-relative reduction. Neural activity can constrain a high-dimensional electrochemical state into an actionable trajectory. That modelling analogy does not decide whether the brain produces, transmits, or realizes experience.

Control theory already supplies the cleaner comparison: a low-dimensional input can select among trajectories in a much larger state space when the system is controllable over the relevant region. No quantum analogy is needed.

Treating the brain as a high-dimensional dynamical system sharpens Block's “Meat vs. Silicon” debate. The relevant comparison is not meat versus code by label, but whether the proposed realizations remain equivalent under interventions on phase, noise, memory, contact, and environmental coupling.

The Great Detour (1900--2000)

One reason we lost sight of this: the Neuron Doctrine gave us tractable measurement and intervention.

Neuron-centred measurements and models made spikes, synapses, and circuits experimentally tractable. This was scientifically productive: it revealed enormous amounts about transmission, coding, and control.

Some modelling traditions treated extracellular fields, population oscillations, and biochemical context mainly as background or averaged them away. Other traditions studied them directly. The historical record is plural, not a single detour.

We also built extraordinarily capable digital machines. They can represent high-dimensional spaces and coupled dynamics; the open comparison concerns physical cost, latency, noise, and intervention response, not whether computers are intrinsically "flat."

The Return to Coherence (2025)

The older field-and-flow intuitions are useful again. Coherence Dynamics asks what systems can do when resonance, phase-locking, chemistry, and field coupling participate in the operation rather than being averaged away.

This isn't fringe thinking anymore. In a recent Current Opinion in Behavioral Sciences paper, Miller, Brincat, and Roy argue that "Cognition is an emergent property" and review population-level coordination, subspace structure, and possible field-mediated interactions. Those mechanisms complement spike-based accounts; they do not establish that fields alone generate cognition or consciousness. The useful shift is from isolated units to coupled population dynamics.

Block worries that functionalism (roles) ignores the “meat.” The stronger response is that a functional description is only as complete as its intervention set.

• Current AI usually abstracts away much of the material terrain.
• If that terrain changes relevant intervention responses, the abstraction has omitted part of the mechanism.

Evolution may have retained the “soup” because chemistry provides reconfiguration, memory, amplification, and multiscale control that purely electrical coupling does not provide as readily. Whether an engineered substrate can reproduce those capacities is an empirical question.

The same modelling question arises beyond neurons. Chromatin conformation, protein folding, cell signalling, and bioelectric patterning all involve spatially extended physical interactions in addition to symbolic descriptions. In each case, the scientific task is to identify which variables and interventions are causally required, not to label every distributed field as cognition.

Pre-genetic protocells already coupled membranes, ion gradients, charge distributions, and chemical potentials. It can be useful to model a membrane as a restricted interface controlling a larger physicochemical state. Calling that state "thinking" or "computing" adds a functional claim that must be earned by a task, counterfactuals, and interventions. Genetic codes later stabilized and transmitted some of this organization, but the evolutionary history does not by itself decide a theory of mind.

The Physical-Realization Hypothesis

The software metaphor leaves out part of the story when the physical medium changes what interventions do. Low-bandwidth signals can steer a much larger state without exhaustively describing it.

Concretely, one candidate mechanism is attractor selection: electrochemical dynamics may have stable or metastable states, and controls such as neuromodulator release, phase resets, or tonic input can change which basin the system enters. Spikes, fields, chemistry, and synaptic state can all participate; their causal roles must be separated experimentally.

This reframes the hardware-versus-software dispute as a comparison of causal models. Two realizations count as equivalent only at a stated grain and for a stated repertoire of interventions.

The “meat” is part of the terrain. It may contribute memory, phase relations, noise, and coupling that disappear from an algorithmic summary. If removing those properties changes the system's response, they belong in the mechanism. This is a realization claim, not the claim that a high-dimensional state space is consciousness.

An abstraction that preserves only steering commands has not yet shown that it preserves the vehicle–environment dynamics. Sometimes that omission is harmless; sometimes it changes the intervention response. The comparison has to be made, not assumed.

Dimensionality Is Not a Consciousness Test

Effective dimensionality can measure the spread of activity across modes. It cannot, by itself, measure consciousness.

The useful question is: does the proposed abstraction preserve the intervention responses relevant to the phenomenon we are trying to explain?

Dimensionality may enter that audit, alongside memory, causal integration, noise sensitivity, recurrence, embodiment, and environmental coupling. Silicon and biology can each realize high-dimensional attractors. The difficult question is whether their causal organization matches at the grain that matters.

This leaves the consciousness question open. A rich manifold is neither necessary nor sufficient on present evidence, and no scalar threshold here licenses claims about animals or AI.

What Would Prove Me Wrong?

If physical realization matters, we should see:

• Anesthesia/sedation: Changes in effective dimensionality should be tested alongside perturbational complexity, connectivity, and report, without treating any one measure as consciousness itself.
• Neuromorphic systems: Physical and digital realizations of the same coarse algorithm may show different intervention responses, failure modes, and energy costs. If they do not, the coarse abstraction has survived that test.
• Electrical vs. chemical networks: Purely gap-junction-coupled networks should show measurable limits in context-dependent reconfiguration compared to chemically modulatable ones, controlling for network size.

These experiments could establish when material dynamics matter to a task. They would not, without an independent theory and measurement protocol, establish that the material property is consciousness.

This post is a response to Ned Block's "Can only meat machines be conscious?" in Trends in Cognitive Sciences (2025). The published research context is developed in High-Dimensional Coherence and Coherence Time.