The river and its stones
There are so many stones in the river. Anyone standing at a creek knows the stones are why it sings. They are where the riffles form, where the current divides and rejoins, where air gets into the water, and life gathers. Take the stones out and you have a ditch.
I have written before about a neuromorphic fabric as black water, dark and quiet until the world pushes against it, showing its work only where the current meets the rocks. Today, I want to talk about the rocks. More precisely, the habit of mind that a seventeenth-century Genevan theologian, Francis Turretin, made into a method, and why that habit turns out to matter for a heterogeneous compute ontology and the neuromorphic compute that runs inside it.
The habit has a short name that sounds fancy in the Latin Turretin used. Distinguendum est, ‘it must be distinguished.’ When a question arrives tangled, you find the place where two things have been made to look like one, or one thing like two, and you name the difference. Every distinction is a stone set in the current. The water still runs. It just runs truer.
What Turretin was doing
Francis Turretin taught at the Academy of Geneva and published his Institutio Theologiae Elencticae between 1679 and 1685. The title names the genre. Elenctic theology is disputational. It takes a contested question, sets out exactly what is and is not in dispute, the status quaestionis, and then answers. The answers turn again and again on the same move. A term is being used in two senses. A proposition is true in one respect and false in another. One thing differs from another really, or only modally, or only in our way of conceiving it. We distinguish.
It is easy to read this as hair-splitting, and generations of readers have. A Thousand Heads on a Pin, so to speak, Part Two.
But, the distinctions had a real purpose. Turretin was guarding a theology he had received. Behind the Reformed scholastics stood the councils, and behind the councils stood Chalcedon’s confession of one person in two natures, united without confusion, without change, without division and without separation. Those four adverbs are themselves distinctions held under tension. Two refuse collapse. Two refuse division. Every later controversy over the person of Christ, whether one will or two, what passes between the natures and what does not, was a contest over whether those distinctions would hold. Turretin’s distinctions were the stones that kept a received grammar from washing out.
I want to be careful here. Turretin knew nothing of workload managers, spiking networks, or storage fabrics. What carries forward here is the habit and the reason for it. We distinguish so that what is genuinely different can stay different and what is genuinely one can stay one.
An ontology is a set of distinctions
A heterogeneous compute ontology is, before it is anything else, a set of distinctions someone was willing to make and maintain.
Songnian Zhou’s 1987 Berkeley dissertation is where the lineage behind LSF and Symphony began and its decisive insight is a distinction. Heterogeneous machines are not fungible. A host has a measured state of its own and the work should go where that state suits it. The load index was a refusal to let a real difference be flattened into a count.
That refusal still separates workload managers today. Kubernetes advertises an accelerator as an integer under a string label. Slurm counts generic resources. Richer description can be restored above the count through plugins, but these simplistic schedulers own model of the world remains a flat pool. Symphony’s resource schema declares each metric with a type, a reporting interval and a semantic direction, whether a higher value means more capacity or more load. That last detail seems small but it is everything. A distinction with semantic direction is grounded in what the resource is. A label with no direction is a distinction in name only.
The same move separates heterogeneous from hybrid. Hybrid names a blend of two and a hybrid is by definition neither of its parents. Gray comes from mixing black and white, the popular dialectic’s synthesis of today’s market descriptions. A heterogeneous compute ontology keeps its kinds and its whole value depends on difference being preserved and put to work. It describes a community of kinds, a spectrum rather than a shade of gray. To say heterogeneous is already distinguendum est.
You can only participate as you are
Here the distinctions become the point.
Brain-inspired computing, as I use the term, means taking the brain’s arrangement as a grammar of participation. To participate, something has to be addressable, so others can reach it. It has to be schedulable, so it can take its turn in shared work. It has to be trustworthy, so others can rely on what it contributes, and it has to receive as well as give. A device that can only be commanded is an instrument. A system fluent in that grammar is a participant.
Grammar presupposes a vocabulary. Nothing can participate as itself until the system knows what it is. A resource with no distinguished nature can only be used. It cannot contribute its proper operation, because no one has declared what its proper operation is. The instrument and the participant are separated by a prior act of distinguishing.
Participation and distinction belong together. Chalcedon’s adverbs hold distinction and union together, and they have to hold at once. Without confusion and without change keep the natures from collapsing into each other. Without division and without separation keep them from drifting into two actors joined by an external hand-off. An architecture that honors three of the four still fails, because the failure the fourth guards against propagates into the whole.
In an ontology, this reads plainly. Each kind of compute keeps its own nature and its reported characteristics are never bent to a common mold. All of them live in one workload management domain. Each one genuinely contributes to the work. Each of those is a stone. Pull any one and the river changes course.
Where neuromorphic computing tests the distinctions
For a GPU or a CPU, the question of nature can mostly stay at the level of orchestration. Neuromorphic silicon pushes it down to the substrate, and there the distinctions multiply.
A kind of its own. An Akida chip is not a small GPU and it isn’t an accelerator either, no matter what might come from marketing channels and industry chatter. Neuromorphic chips run event-driven, sparse and local, firing when the world stirs, which is why it can stand watch at sub-watt for years where a rack of bulk LLM tech remains the wrong kind of compute for its work. A workload manager that sees it as one more device under a label has already lost the difference that makes it worth having.
Genuine participation and simulation. A cortex and a neuromorphic fabric both exercise one structural form, event-driven sparse computation organized by connections, each as the proper operation of its own substrate. That is participation. A GPU can compute a model of that form while exercising its own proper operation, dense tensor parallelism. That is simulation. The line falls between a form loaded onto a neutral host as instructions and a form configured into the substrate itself, with no instruction stream running above it. The two modes differ in kind.
Degree within kind. Inside genuine participation, the extent is graded. Present silicon exercises part of the form: a feed-forward spiking classifier with on-chip plasticity at the last layer. A cortex carries the full temporal-recurrent dynamics. The orchestration layer carries the columnar hierarchy at another scale. Saying both at once, a difference of kind between the modes and a difference of degree within one of them, is the sort of statement that only holds if you are willing to distinguish.
Operation and person. An operation belongs to a nature. The machine’s operation is genuinely its own. A nature becomes personal when a person takes it up, so the machine remains one nature in a joint work and never a second agent. Responsibility for the joint work rests with the person through whom the work is one. A made nature that no one takes up leaves responsibility nowhere to rest.
Structure and experience. A neuromorphic fabric encounters the world and learns from it. That encounter is real operation, what we typically call experience. What it is like to be the thing doing the sensing lies beyond structural sharing, however deep. That stays with the person. The grammar declines to legislate about interiority in either direction, which is itself a distinction worth keeping.
The collapses the stones answer
Read back over the arguments I have made this past year and nearly every one is a distinguendum est aimed at a collapse with some convenience behind it.
Operating system. Call a coordination layer an operating system and you start designing as if you must own the kernel. The kernel already has an owner. Linux binds processes to hardware and drives the Akida chip as a device. Symphony does the inverse, freeing the work from any particular machine. One word, two referents, and the equivocation does the work of a moat. Then, you do things like pretend you invented patching. None of it is real and none of it carefully distinguishes the way you should.
Placing and understanding. Kubernetes answers where a pod fits. It has no native way to ask what kind of work a job is or what kind of resource suits it. Kubernetes places. A compute ontology understands.
Tool and singularity. The tool narrative, Aristotle’s living instrument returned under technological cover, refuses the machine any nature of its own. The singularity narrative, the Platonic ascent rendered in silicon, dissolves the particular, human and machine alike, into a substrate-indifferent intelligence. They look like opponents and are halves of one inheritance. The distinction both of them erase is the one Chalcedon kept: two genuinely different natures, one operation, neither absorbed.
Edge and center. Call a neuromorphic node the edge and you have already decided it is a periphery feeding a center that understands. No one talks about their own senses as the edge of their personhood. Distinguish the node as a participant and local knowledge never has to be extracted into someone else’s representation space before it counts.
Substrate collapse. Some of the most ambitious current research absorbs computation, memory and input-output into one learned runtime, and its authors are candid about what is lost. Arithmetic and stable state are exactly the functions that suffer once the distinctions classical computing keeps are dissolved. That is what happens to a river with the stones dredged out.
Why this matters
A stone in the river gives the water a shape it can keep. Stones are also how you cross. When Israel crossed the Jordan, twelve stones were taken up from the riverbed and set down at Gilgal, so that later generations would ask what they meant. A distinction works the same way. It marks where something was learned at cost, so the lesson does not have to be learned again.
That is why the habit matters for the heterogeneous compute ontology now. Increased pressure in the industry runs toward one platform, one abstraction, one vendor’s stack, one learned substrate that does everything, one failed paradigm after another. Every such proposal asks us to pull the stones and the result is a ditch, fast, straight and lifeless, carrying everything to one place.
A heterogeneous compute ontology is the opposite. Linux drives the silicon. Symphony orchestrates the fabric. GPFS carries the state. Akida stands as a peer alongside quantum, GPU, CPU and the mainframe, each invoked where its physics wins, each participating as its own. The community of kinds holds together because the differences were named and kept.
Turretin distinguished for the sake of a union he had received. The ontology distinguishes for the sake of one coordinated work that no single kind of compute could do alone. Distinguendum est is how so many different things can share one river and still be themselves in it.