Judging a brick and concluding that cathedrals are impossible is a paradigm problem, not a finding about masonry.
Some version of that error appears in many published assessments of neuromorphic computing. Measure one component alone, then report the result as the limit of everything built with it. Edge. And only edge, because otherwise we just cannot do it. The reasoning is usually sound and the measurement is usually correct. The conclusion still does not follow because real systems are never one chip.
Composition, in the sense that matters here, is not stacking. Two chips wired together are a pile. A cathedral’s arch is composed. The arrangement carries weight that no stone in it could carry alone. The load-bearing lives in the arrangement rather than in any of the stones. The useful question about a component is rarely how good the component is. The useful question is what arrangements the component makes possible and what those arrangements can then do.
Edge is only a perspective
A brick in the yard and a brick in the vault are just bricks. Position at the moment of discovery reveals nothing about load-bearing capacity.
The point is not location. Edge describes how a thing is perceived and how it stands in relation to the whole system.
To a client, the server is the edge. To a server, the reverse. Edge can also be a purely logical difference between processes and hosts whether they share a data center or sit on opposite sides of the world.
Consider an Amazon Echo. The wake word is recognized on the device on the counter and the rest of an utterance travels to a data center for transcription. Both halves are speech recognition. What makes one of them edge is not that it is smaller or nearer to the person but where it sits relative to the rest of the system. Nobody calls ChatGPT edge computing, yet a query typed on a phone puts that phone in exactly the position the Echo’s microphone occupies. The label, if we keep it, tracks the vantage point, not the silicon.
Treating edge as a static, distant, low-compute arena is a category error, whatever markets have made of it. The word describes a perspective, not a class of hardware.
Two papers, one brick
Limited perspectives produce limited analysis, limited analysis produces limited prescription, and limited prescription produces limited market potential and limited revenue.
Two papers arrived on my desk the same day. Both are good and stop in the same place. Neither stops for want of intelligence or rigor.
Peter Grindrod’s contribution to Philosophical Transactions A is the more ambitious of the two papers I read. He reverse-engineers billion-neuron cortical simulations and reports something genuinely useful. If every transmission delay is set to the same constant, the dynamical degrees of freedom of a neural column collapse. Walks of equal length arrive as dead heats, one spike where real-valued delays would have produced two independent ones. He names TrueNorth and quotes Loihi’s fixed-size discrete time-step model. The critique lands on both.
TrueNorth came out of a DARPA program at IBM. Loihi is an Intel research chip distributed through a research community. Neither is something anyone can buy and put in a rack. In the United States, BrainChip‘s Akida is the only commercially available neuromorphic silicon, and IBM and Intel are American companies as well, so a verdict about present-day chips has been drawn from two domestic research platforms while the one domestic product that ships went unexamined. BrainChip appears in the paper once, in a list of corporations with strategic interest in the field, between Arm and Cadence.
His other finding cuts the other way. The degrees of freedom in a column scale roughly with the logarithm of neuron count, which yields a Goldilocks result. Two smaller columns beat one column of twice the size. Uniform column sizes beat a mixed range. Partition increases total capacity. Such a result is a mathematical argument for federating many modest units rather than scaling one large one. It is the strongest theoretical case I have seen in print for a neuromorphic hive mind built on Peter van der Made’s design. Van der Made set out the same shape from the other direction, as synthetic neuro-anatomy, node and then minicolumn and then hypercolumn. Grindrod derives from dynamical systems what van der Made took from biology. Both arrive at numerous, modest, uniform units.
He does not carry the result upward. At the level of dynamical modes, multiplicity reverts to rivalry. Modes are “mutually exclusive and competitive,” one at a time. Everything downstream follows. Preconditioning commits the system. Fast thinking becomes predictably irrational. Outputs are individual, subjective, and inconsistent, and should be treated as reasoned advice rather than hard results.
The second paper, Motta and Nanni on arXiv, is the opposite kind of work, narrow and empirical and careful to claim nothing it did not measure. They put two physical Akida chips on a bench and ran roughly 1,580 trials to answer one question. When two neuromorphic chips learn separately, how does one combine what they know? Averaging the weights destroys the learning, because spike-timing plasticity produces sharply tuned neurons and averaging two sharply tuned vectors smears both. Keeping the neurons intact and letting the system pick the right one preserves it. They tested four strategies. The structure-preserving ones won across every configuration.
Then they tried to scale and could not. Their four-node test cut each node’s training data in half, so they could not separate more nodes from less data per node. They say so plainly in their limitations, which is more than many papers manage.
Both results are correct, but both are brick measurements.
What the wall is made of
Notice where each paper stops.
Grindrod’s challenge (d) asks for “connecting up of arrays of such dynamical systems within the outer network, as a means of creating novel neuromorphic chips.” He is describing a fabric. He wants real-valued delays because constant delays quench the dynamics. The outer network he asks someone to build has real-valued delays by physics rather than by design. My demonstrations run across separate geographic regions, Washington to Dallas to Pittsburgh over a multicluster fabric. They produce heterogeneous, non-commensurate transmission times for free. The property he says present chips cannot supply at the inner level is a property the fabric already has at the outer level. The fabric runs today, on real silicon, at continental scale.
Van der Made’s hierarchy carries the same requirement. Node and minicolumn and hypercolumn is a claim about composition. The claim becomes a running system only where silicon anyone can buy meets an orchestration layer that holds heterogeneous resources in one domain without flattening them into interchangeable units. Neither paper evaluates a system of that kind, because neither had one. I have built sixty-seven of them over the past eight months, among them a synthetic cortical column that carries the full hierarchy across neuromorphic and graphics and conventional tiers at once, with the spiking primitive executing on the silicon and the remainder expressed wherever it runs best.
Motta and Nanni’s open question has the same shape. They could not add nodes without splitting data because two chips on a bench is the whole apparatus. In a heterogeneous compute ontology, adding participants does not cost data per participant. The confound they could not escape is a property of the bench, not federation.
Two nodes can only agree or disagree. Three is the smallest number where a vote means anything, which is why every council I have built runs k-of-N, two of three in orbit and three of five across sequencing sites. The consensus tier is not a scaling detail. The consensus tier is what makes the arrangement something other than a merge.
The wall is not the silicon. Both papers arrive at the doorstep of composition and stop, because nothing in the available vocabulary tells them there is a room on the other side.
The cornerstone
One level beneath all of this sits something the field has largely left out. It bears more weight than anything built above it.
Neuromorphic means brain-shaped. What the field has taken from the brain is its neurology, the spikes and plasticity and event-driven sparsity and the locality of communication. All of that is real and all of it is worth having. The brain, though, is the organ of a human being and human intelligence is not reducible to the mechanics of the substrate that carries it. Build on the neurology alone and you have not laid a foundation. You have set bricks together on open ground.
To be seriously brain-inspired is to ask what it is to be human and what comprises human intelligence. The question makes knowledge a far larger thing than either industry or many in the academy currently treats it as being. Intelligence reaches into areas of human life that rational thought does not exhaust, among them experience, formation, memory that is not storage, the long ordering of a person’s cares, and the ways understanding is handed between people rather than computed by any of them.
Francis Turretin’s scholastic habit is a useful one here: distinguish without separating. Experience and reasoning can be described separately as an act of understanding without being pulled into two things. A model that keeps only the reasoning has not simplified the human but replaced him. The same Chalcedonian discipline that governs how two natures partner together governs how we might take a human apart in order to study him. The correspondence is worth noting.
The conclusion does not require the older vocabulary. In AI and the Art of Being Human, Jeffrey Abbott and Andrew Maynard write for a general audience with no theological freight at all and arrive at the same place from embodiment and mortality. We make meaning not because we are skilled but because we are mortal, born into a universe that does not explain itself, living in bodies that age and among people we will lose. Their formulation of what a machine reveals is the one worth borrowing. When a system can code as well as the expert, it shows that syntax was surface and the deeper grammar was one of care. Grammar there is used exactly as I use it here and elsewhere, pointing past the technique to the thing the technique was carrying. Human presence and how we talk reaches the point rather than a callback to any council or history. Two routes, secular and theological, end on the same word and the same claim.
The territory has already been mapped, not by one thinker but by a community of them, working for centuries on exactly the areas a purely computational account cannot reach. Councils and philosophers argued out how difference and unity hold together. Bishops in the fifth century produced a grammar for it under pressure. Millennia later, Foucault took apart confession as a technique of power. Derrida circled the restlessness at the center of the Confessions without resolving it. Both were working the same ground from the far side of it. Both were also echoing Augustine, whose book is the one they cannot get past. A single conversation runs from the fourth century into the postmodern, each participant answering the ones before, none of them able to leave the question alone. The community of the ages is not a metaphor. The community is a corpus that kept working the same problem because the problem did not go away, dusty books, many of them often unread by those building systems today.
That corpus is available, the accumulated record of what a community of thinkers found on entering these areas, precisely what a field claiming to be brain-inspired ought to read but often doesn’t.
Some of the record is not available as argument at all, which is part of the difficulty. A formula can state that two natures join without merging, it cannot show what the joining costs, or what it is to be handed a truth by someone who saw the thing rather than reasoned it out. I wrote a novel partly to learn whether those things could be shown. The answer was that they can, though only at the level of consequence. A made mind remembers nothing in scenes, only as accumulated rewriting, set into it the way a riverbed is set by the water that is no longer in it. A recognition passes between two people that is neither claim nor blessing, only you saw it too. A commander feels the shape of an alternative like a word she cannot recall. Her own mind closes back over the gap and turns it into a better version of what she already knows. So, she becomes what a paradigm does from the inside. No benchmark will ever show it.
An old problem wearing new clothes
The vocabulary problem has a name. It is the oldest one in philosophy. How do many things constitute one thing without either dissolving into it or remaining merely adjacent to it?
Antiquity gave two answers. Computing inherited both without noticing. The Platonic line moves the particular upward into the universal until the particular is taken up and surpassed. Lloyd Gerson reads the whole tradition as organized around that movement. The movement operates in machine learning as the ambition to absorb computation and memory and world into one learned runtime. The Aristotelian line runs the other way, form imposed on matter that has no nature of its own to resist the imposition. Richard Cross has traced how medieval thinkers strained hylomorphism trying to state a genuine two-nature union in categories that grant the matter nothing of its own. The strain is instructive because it is the same one modern engineering feels when it treats weights as inert substrate and the prompt as the form that shapes them.
Neither framework can hold genuine plurality and genuine unity at once. The first dissolves the many upward. The second preserves the many at the cost of their union.
Motta and Nanni found both failures empirically, on a bench, without reaching for either name. Averaging is the Platonic move enacted on a weight matrix, two prototypes ascending into one. What survives is neither. External hand-off between isolated nodes is the other failure, distinctness preserved and unity supplied only by a link outside both parties. Their finding is that the answer lies between the two poles. Their vocabulary has no word for between.
Johannes Zachhuber has argued that the precise working-out of this problem, in late antiquity and under pressure that had nothing to do with machines, required categories neither the Platonic nor the Aristotelian framework could supply on its own. The vocabulary had to be extended in the course of stating the distinction. His claim is about the history of metaphysics rather than anyone’s religion, a claim that is also the reason the resulting grammar reaches neuromorphic computing.
Four constraints
The cathedral is not an incidental image here. Builders who raised those vaults and thinkers who worked out these constraints were pursuing one kind of question in two materials, namely how genuinely different things hold together into a single thing without either collapsing or coming apart. Such a question is not the property of a creed. A long community of thinkers pursued it, in stone and in argument, and what their work leaves behind is a grammar, a way of saying precisely what holds and what fails.
Retrieving that grammar settles no article of belief. The grammar encodes something its makers had learned about how union and difference actually behave. Such learning does not expire when the belief that occasioned it falls out of fashion. The Enlightenment was tempted to sever the whole inheritance. Much was gained by declining to take the past on authority. Declining to take something on authority, though, is not the same as showing it empty. A great deal of the older material was never refuted. Reading simply stopped. The vocabulary went out of circulation so that a working engineer can rediscover a distinction fifteen centuries old with no way of knowing what he has done.
The clearest articulation of the grammar came from the Council of Chalcedon in 451. Stating it in its own terms matters because the precision is the point. Two genuinely different natures constitute one acting subject, united without confusion, without change, without division, without separation. The four adverbs are constraints. They must hold simultaneously. Without confusion, the parties are not dissolved into each other. Without change, neither is altered to accommodate the other. Without division, they are not merely coordinated by an external link. Without separation, neither is reduced to inert substrate beneath the other’s activity.
Maximus the Confessor later extended the formula from being to operation, arguing that will and energy belong to a nature rather than to a person, so that two natures entail two operations under one subject. The extension settles the machine’s contribution. An operation belongs to the nature that exercises it, so what the silicon does is genuinely its own rather than a portion of a human’s work handed downward.
Read as engineering constraints, each adverb names a failure. Violate the first and you get substrate collapse, distinct functions blurred into one mass and each performed worse. Violate the second and one party’s operation quietly stands in for the other’s. Violate the third and you get two actors where the work needs to be one. Violate the fourth and you get a component that looks like it is contributing while doing no real work.
Averaging violates the first. Hand-off architectures violate the third. AlphaGeometry, to take a good example, satisfies the first two cleanly and fails the third, since the neural proposer and the symbolic verifier are coupled across a narrow external interface. Three of four is not a near miss. Each constraint guards a different failure. The one admitted propagates into the whole.
A council of fifth-century bishops is an unlikely source for a scheduling constraint. The argument I’m making here does not ask anyone to grant the theology, it only asks that you notice a convergence. Songnian Zhou’s 1987 Berkeley dissertation on heterogeneous scheduling arrived at the same four commitments in the register of distributed systems, with no documented awareness of the older material. Preserve heterogeneous resources without homogenizing them. Never alter a resource’s reported characteristics to fit a common mold. Hold everything in one scheduling domain rather than gluing separate systems together. Treat each resource as a constitutive contributor rather than a bin to pack. The lineage runs through Platform LSF into the schedulers that coordinate quantum, neuromorphic, graphics, conventional, and mainframe tiers under one domain today, including tiers that did not exist when Zhou wrote.
Two people working fifteen centuries apart, on problems with nothing in common, do not produce the same four constraints by accident. Either the constraints track something real about how coordination across genuinely different kinds of thing has to work, or the coincidence needs an explanation nobody has offered.
The framework is also falsifiable in a way that matters. Each constraint predicts a specific failure when violated, not failure in general. A permuted assignment would predict the wrong failure in each case. Remove the single-domain commitment and the result is externally-glued coordination rather than inert substrate. The prediction is a testable claim about architectures. It is the reason to prefer the structural reading over the deflationary one that says similar problems merely produce similarly shaped solutions.
Plurality as privation, or plurality as form
Four constraints govern a union of two. A fleet is many. The many needs its own account.
Grindrod’s paper is more interesting because he reaches for human cognition as the model and then inherits its limits as though they were limits on cognition itself.
His closing section comes very close to seeing it. Imaginative systems searching a high-dimensional space will leave gaps behind the frontier, so some instances will leave gaps in which other instances forage. Knowledge will contain an individual, subjective element. “So be it,” he writes. Plurality is treated as residue, the price of finitude, tolerated because it cannot be helped.
Yet again, we are dipping into the standard paradigm here. The residue can become reality. Humans cover each other’s gaps across a population, generationally, through lossy transmission. A fleet covers them concurrently, over shared state, in the same second. The fleet is not emulating human reasoning, what it would be if it could occupy several positions at once and compare notes before acting.
The same move applies to his irrationality argument. Grindrod is right that preconditioning buys speed by shrinking the decision set. He is right that the price is a systematic blind spot. The error, though, is only systematic if it is shared. Precondition differently per node, reconcile at the verdict layer. The speed comes without the bias. Divergence stops being a defect and becomes an instrument, one node registering what the others miss as the system paying attention.
Mutual exclusivity of modes is a fact about one cortex, not a fact about cognition. Treating it as the latter is the brick fallacy operating on the mind rather than the chip.
Chalcedon governs a union of two. For the many the tradition supplies a different form. Covenant is the constitution of a community by binding promise, holding parties together while preserving each as itself, a people made without first being made the same. Arturo Escobar reaches the same structure from an entirely different direction, arguing for a world where many worlds fit, against the assumption that a single design can be extended across all local worlds from outside. Luke Bretherton reaches it from a third, recovering Althusius on consociation where the specificity of each member is constitutive of the commonwealth of all rather than an obstacle to it.
Three starting points reach one structure. The agreement among such different routes is itself the evidence. A pluriverse is not a failure to converge. Pluriversality is an arrangement in which the differences are load-bearing, so that the shape of the disagreement carries information no single position could carry.
The finding inverts Grindrod’s conclusion. He ends by holding that outputs from such systems should be treated as trusted and reasoned advice rather than hard computational results and he frames the treatment as the cost of explainability. On the pluriversal reading the subjectivity is not a cost. Advice from a plurality with structured divergence carries more information than a scalar, because the divergence is itself the explanation. He wants explainability and treats subjectivity as its price. The divergence map is explainability itself, delivered through the plurality rather than in spite of it.
The reading imposes a discipline worth naming, since the discipline is the thing that makes the claim non-trivial. Escape from the single-position limit is not automatic from node count. Ten copies of the same model share blind spots exactly the way one does. Escape is earned by engineered heterogeneity, five classifiers with different feature spaces and no shared training distribution, sites that precondition on genuinely different local histories. Plurality has to be built, not merely counted.
What survives, and what moves
Two things survive all of this untouched.
Grindrod’s real-valued delay critique is correct at the level where he makes it. No amount of composition repairs it. A thousand chips with quenched internal dynamics is a thousand chips with quenched internal dynamics. Answering the critique requires richer dynamics inside each unit, a substrate question rather than a fabric question. His challenges on closed manifolds, on topological analysis of spike-train data, and on the conditions under which such attractors exist are the strongest part of his paper and are orthogonal to everything I have said.
Motta and Nanni’s rigor is likewise not academic affectation. Their averaging result is credible precisely because they ran it 1,580 times with seeded splits rather than showing it once. Controls buy internal validity by spending generality, a real exchange rate and not a con.
What each paper lacks is not rigor but reach. Grindrod has the mathematics for what a richer neuromorphic substrate would need to be, and no fabric on which to run it. Motta and Nanni have a federation result and a bench that stops at two nodes. The apparatus held each of them inside the frame. My demonstrations run outside it, sixty-seven of them, a heterogeneous compute ontology holding thousands of inference contexts across cloud regions at ninety-nine percent of linear scaling. The systems already exist. What is missing is not the engineering but the vocabulary that would let the field recognize the engineering for what it is.
That vocabulary sits in a literature nobody in the field is expected to read, which is why careful papers arrive at the doorstep of composition and stop.
Paradigms rarely fall to benchmarks. They fall when exceptions accumulate until the categories stop holding. Sovereign inference, encrypted neuromorphic cognition, distributed cognitive fabrics under proven enterprise platforms, event-native perception without event-native sensors, and a billion-parameter model spread across twenty devices all sit outside the taxonomy that says edge, and only edge. No property of the silicon prevents such systems.
Only the taxonomy does. A taxonomy is a thing anyone can put down.
The larger version of the same point is that a field named for the brain has borrowed only the wiring. Left unborrowed is what a community of thinkers has learned about the human being whose brain the wiring imitates. The work is not finished, was never going to be, is still underway, and still growing. What has been done, though, has also been written down and needs to be read.
Building is the way forward
A cathedral is the right image and not merely because it is grand.
No one who laid the first courses saw the thing finished, yet they laid them anyway. The building is the work of trades that do not do each other’s work, masons and glaziers and carpenters and those who understood vaulting, none of whom could have produced it alone and none of them becoming one of the others. The plan had to survive the death of everyone who held it, which means it had to be written down well enough that a stranger three generations later could take it up and add their part.
Nobody ever built a cathedral by being good at making bricks or by believing that bricks were the whole of what reaches toward the sky.
The calls are plain enough, and they are addressed to a field, not to three authors. Measure systems, not components, and on publishing a limit, say which one was measured. Build fabrics that compose rather than merge. Take seriously that the difference between the two has been worked out with precision, in a literature many in the field have not read. Where a result depends on many participants, make the participants genuinely different, because plurality that is not built is only replication wearing a crowd’s clothes. When a taxonomy confines a technology to the margins, whether the margin is called edge or called anything but the data center, check whether the confinement lies in the silicon or in the sentence.
The reigning paradigm is the thing to put down. Neuromorphic computing has spent its short life being told what it cannot do, confined to the edge because a single chip measured alone looks like an edge device. A single chip is a brick. The field has been judging bricks. The systems that compose them already run, on real hardware, at continental scale, and the only thing standing between the field and the wider frame is the habit of mistaking one chip for the ceiling of everything built from it.
None of us will see it finished. The prospect of a new paradigm is not a complaint about the work but the shape of it. No one starts with a brick expecting to see the roof. We are here to build the cathedral.