Coherence Isn't Enough

For twenty years, the working model of the cortex has been a hierarchy of specialists. Lower areas detect edges and tones; higher areas assemble them into faces, words, and categories; a given neuron, once you decode it, always means the same thing. It is a clean picture, and it is wrong in a specific, well-documented way: most cortical neurons don't mean one thing. They multiplex. A prefrontal cell might fire for a butterfly seen first in a sequence and stay silent for the same butterfly seen second. Earl Miller and colleagues at MIT's Picower Institute have spent a decade building the case that this "mixed selectivity" isn't noise sitting on top of a cleaner architecture — it's the architecture. Their new paper, "Analog Cognition and Consciousness," asks the obvious follow-up question: if neurons are this promiscuous about what they represent, what tells a given population which job to do at a given moment? Their answer is brain waves. Not as a side effect of spiking, but as a physical control system in their own right — one that also, they argue, does actual computation, and one whose large-scale coherence is where consciousness comes from.

Recursive Social Prediction (RSP) is a different kind of theory: a claim about what phenomenal experience is made of, not primarily about how cortex routes information. But the two projects turn out to overlap in an unexpected place. RSP's mathematical appendixes describe a hierarchy of predictive-coding message-passing and gain control (Appendix A) and formalize the whole system as nested feedback loops with reference signals, gains, and stability conditions (Appendix D) — and never say what physically implements the gating. Miller, Brincat, and Roy hand over a candidate answer. That's worth taking seriously. It's also, on the question that matters most to RSP, not enough.

What Miller, Brincat & Roy Get Right

Mixed selectivity is a real problem for the connectionist picture

The paper's opening move is hard to argue with. If cortical function really worked by assigning each neuron a stable, specialized role, learning any three cognitive functions would exhaust a huge chunk of prefrontal capacity before you'd even gotten interesting behavior out of it. Instead, the same neurons participate in different computations depending on context, and this "nonlinear mixed selectivity... reflect[s] a core computational principle" of cortex, not an exception to it, present in 30–40% of neurons in prefrontal cortex and also found in primary sensory and motor areas. That's a strong, well-cited empirical claim, and it forces a real question: something has to route these overlapping populations into different roles moment to moment, without rewriting the synaptic wiring every time. Digital, discrete-neuron accounts of cognition don't have a good answer to that question. RSP, for what it's worth, doesn't either — Appendix A's precision-weighting mechanism says representations get gated by expected reliability, but it is silent on the physical channel that carries the gate.

Ephaptic coupling is a serious, underused mechanism

The paper's case for brain waves as more than epiphenomenal readout is genuinely persuasive: oscillating extracellular fields modulate spike timing and probability directly, without synapses or gap junctions, and cortical neurons spend much of their time "teetering" near spike threshold, where even weak fields can tip the balance. Field effects propagate at the speed of electromagnetic conduction rather than synaptic transmission, meaning waves can coordinate activity across a cortical region essentially instantaneously compared to the millisecond-scale relay of spike-to-spike signaling. This is not speculative neuroscience; it's a well-supported, if still under-appreciated, physical mechanism, and the paper deploys it carefully.

The mechanism isn't confined to spike-generating cells, either. Astrocytes — glial cells that don't spike, yet reach near parity with neurons in the human cortex (a 1:1.4 ratio, up from 1:3 in mice) — contribute to the same field dynamics from outside the spiking system entirely. The authors push the physical story further still: microtubules, the cytoskeletal scaffolding inside neurons, physically align to external electric fields, which could in principle retune a neuron's own resonance properties by altering dendritic resistance and capacitance. Whether or not that molecular-level claim holds up, it signals how far this account is willing to push "wave" past metaphor and into concrete biophysics — considerably further than anything RSP's appendixes attempt.

The anesthesia data is a clean convergence result

One of the paper's stronger empirical moves is noting that anesthetics with entirely different molecular targets — GABAergic propofol, NMDA-antagonist ketamine, alpha-2-adrenergic dexmedetomidine — all converge on the same systems-level signature: a collapse into high-power, temporally misaligned slow delta activity across frontal, parietal, and sensory cortex. The authors' own prior work (Bardon et al., Bastos et al.) backs this up directly. Their reading is that consciousness "depends less on specific receptors or cell types and more on the integrity of large-scale wave organization" — synaptic circuitry stays largely intact under anesthesia, but the coordinated wave patterns that would normally organize it collapse. That's a genuinely elegant unification of a pharmacologically diverse phenomenon, and it sits comfortably next to RSP's own anesthesia-adjacent predictions (Appendix G) without contradicting them.

The Fault Line: Coherence as a Sufficient Condition

“Consciousness, then, emerges when these dynamic wave patterns bring the cortex into an organized, globally integrated state, one that naturally links and influences widespread activity.”

Read that sentence carefully and it makes a strong claim: sufficiently organized, globally coordinated wave dynamics is consciousness. Not a necessary substrate for it, not a mechanism that happens to accompany it — the thing itself. Nowhere else in the paper is that condition narrowed. There's no requirement that the coordinated activity represent anything in particular, no requirement that it model the system generating it, no requirement that it be recursive in any sense. Global coherence, on its own, does the whole job.

This is the same move RSP has already had to answer twice. Against Integrated Information Theory, RSP's existing reply is that "a thermostat with nonzero Φ is conscious under IIT; under RSP it is not" — because IIT is architecture-neutral, indifferent to what is being integrated as long as enough of it is integrated. Against Global Neuronal Workspace theory, RSP's reply is that GNW "treats all information equally once broadcast," while RSP insists the content has to be structured through core affect and schemas before integration means anything phenomenally. Miller, Brincat, and Roy's coherence criterion is a third variant of the same pattern, just paid for in a different currency: instead of a bit-integration measure or a broadcast event, the sufficient condition is wave-based coherence. But the objection transfers cleanly. A network of coupled oscillators — entrained, globally synchronized, "organized" and "integrated" in exactly the sense the paper describes — would satisfy their stated criterion for consciousness while containing nothing that models itself, predicts its own states, or represents anything about the system doing the coordinating. RSP would deny that system is conscious. The paper gives no resource for denying it either.

Notably, the paper cites Riddle and Schooler's "nested observer model" as a close relative — oscillatory dynamics imposing "a hierarchy of consciousness" — and explicitly declines the observer-system framing: "we emphasize the very act of top-down organization in generating conscious integration," rather than positing a separate system that renders states conscious. That's a coherent position to take. It is also precisely the position RSP argues doesn't work, because it collapses the distinction between a system being organized and a system representing itself as the thing being organized. Those are not the same fact about a system, and only the second one, on RSP's account, produces something it is like to be that system.

What RSP Adds

What RSP supplies, in other words, is exactly the condition the paper's own sufficiency claim is missing: a reason global coordination would produce a first-person point of view rather than just a well-synchronized system. RSP's answer is architecturally specific — it isn't integration in general, it's a particular recursive act, the system generating a prediction of itself as a social object, folded back into its own predictive hierarchy. That's a stronger, more falsifiable claim than "organized enough," and it comes with a developmental story — core affect has to stabilize before schemas can be commissioned, schemas before recursive self-modeling — that has no counterpart anywhere in "Analog Cognition and Consciousness."

What Miller, Brincat & Roy Give RSP

The debt runs the other way too, and it's a real one. RSP's appendixes are unapologetically formal about the computations involved — generative models, precision weighting, gain matrices, contraction mappings — and almost entirely silent about the physical medium carrying them out. That's a legitimate division of labor for a philosophy paper grounding a hard-problem argument in mathematics rather than wet biology, but it leaves an obvious question open: what, physically, gates a prediction, routes an error signal, or determines which of several overlapping neural populations gets to express a representation at a given instant? Miller, Brincat, and Roy answer that question with more mechanistic detail than RSP has ever offered — ephaptic fields, cross-frequency "stencils," traveling waves shifting population activity between coding subspaces, all backed by primate single-unit and LFP recordings, not just modeling. They even sketch a concrete worked case: the intersection of waves induced by an object's contours in visual cortex, they propose, could compute a "grassfire transform" — a classical image-processing operation that extracts an object's medial axis from its boundary — through wave interference alone, with no change to the underlying synaptic wiring. That's not a metaphor for computation; it's a specific, testable claim about what cortex might be solving and how, and RSP's appendixes have nothing at that level of concreteness anywhere in them. If RSP's Level 2-to-3a gating (which schema gets to structure which concept, in which context) needs a physical implementation, spatial computing theory is a serious, testable candidate for supplying one. That's a genuine gift, independent of whether their consciousness claim survives.

Classical Battlegrounds

Here the comparison runs thin, because the paper simply doesn't engage the classical battlegrounds. There is no discussion of zombies, no mention of qualia as a category, no reference to Mary's Room or the Chinese Room, anywhere in the text. This isn't a criticism so much as a scope observation: "Analog Cognition and Consciousness" is a neuroscience paper about a computational mechanism, dressed in the language of consciousness at the end, not a philosophy-of-mind paper trying to dissolve the explanatory gap. Its single sentence on why coherent integration should produce experience — that it "naturally links and influences widespread activity" — is doing the entire philosophical job the paper asks of it, and it isn't built to survive the pressure a philosopher would put on it. RSP's dissolution argument, that phenomenology just is recursive self-modeling rather than something caused by it, is answering a question this paper never quite poses.

The Honest Disagreement

Strip away the parts that dissolve into complementarity and one real disagreement is left standing: is phenomenal consciousness a matter of degree of organization, or a matter of a specific kind of content? Miller, Brincat, and Roy's account is, in the end, gradable along a single axis — how coherent, how globally integrated is the wave state right now. RSP's account is not gradable along that axis alone; a system can be highly coherent and still lack the one ingredient RSP treats as constitutive, a self-model built from social prediction. This isn't a case where more data resolves things by itself. It's a disagreement about what kind of fact consciousness is, and no amount of additional evidence about wave coherence, on its own, would settle it either way — you'd need evidence that specifically targets whether coherence without recursive self-modeling ever produces experience, and neither paper currently supplies that.

Where This Leaves Things

None of this diminishes what the mixed-selectivity and ephaptic-coupling evidence establishes. The cortex really does seem to run on flexible, wave-coordinated routing rather than fixed neuron-to-function assignment, and that fact needs to be true of whatever theory of consciousness ultimately turns out to be correct. It just isn't, on its own, the fact that makes a brain state feel like something to be in. Coherence can be necessary without being sufficient — and the paper's own framework, ironically, has the resources to say so, if it ever asks what, besides coordination, the coordinated activity is coordinating around.