A small, published, inspectable set of primitives, held to a contract, extended only in the open. That discipline is the load-bearing member of the entire structure — and at the far end of the line, at the models, it is the one thing that did not arrive.

I. A Line You Can Walk

Computing history is usually told as a series of revolutions, each one erasing what came before. That is a poor description of what actually happened. What happened looks much more like a lineage — a single line of descent, traceable step by step, with most of its participants still alive and available to be asked.

The line runs from a small chip designed in 1975 by a handful of people, through a bungalow in Mesa, Arizona, through a graduate lab at Berkeley, through a British company with three hundred employees, into very nearly every phone and appliance on Earth, into the massively parallel processors that NVIDIA now ships by the datacenter, and finally into the large language models that write, reason, and advise — including the model that helped draft this essay.

It is worth walking that line slowly, because what traveled down it is not what most people assume. It was not the transistors; those were replaced many times over. It was not the instruction encodings; those diverged early. What traveled was a discipline: a small, published, inspectable set of primitives, held to a contract, extended only in the open.

That discipline is the load-bearing member of the entire structure. And at the far end of the line — at the models — it is the one thing that did not arrive.

II. What the Fetch Was

The MOS 6502 reached the market in 1975 with roughly thirty-five hundred transistors and a price of twenty-five dollars, at a moment when competing parts cost several hundred. Fifty-six instructions. Thirteen addressing modes. That was the whole machine, and the whole machine fit on a fold-out card.

Read the 6502’s operating cycle through the Theory of Embedded Intelligence and it resolves immediately into SPCA. The processor senses — a byte arrives from memory on the data bus. It processes — the opcode is decoded, an addressing mode selected, an operation determined against the state of the accumulator and the status register. It communicates — addresses go out, operands come back, the bus carries the machine’s intent into the world around it. It actuates — a value is written, a pin changes state, a motor turns. Then the cycle begins again, a million times a second, and has not stopped since.

The 65xx architecture makes that reading unusually literal. Its Addressable Register Architecture — what the rest of the industry flattened into the phrase “memory mapped I/O,” losing the point in the process — places the machine’s outward action inside the same address space as its inward thought. On the W65C02S, registers live in Page Zero; on the W65C816S, in Direct Page, with sixteen megabytes of addressable register space behind it. Communication and actuation are not bolted onto the architecture. They are addresses. You can point at them.

But the property that matters most for this essay is not the elegance of the addressing. It is this: the instruction set was complete, published, and constitutive. Every opcode, every cycle count, every flag effect, printed and available to anyone who asked. And the machine could not execute an operation outside that set. Not would not — could not. The constraint was physically prior to any instruction it might receive. No program could argue with it, no clever input could talk it around, no operator could persuade it.

That is what TEI means by governance being constitutive rather than advisory. The 6502’s datasheet was not marketing literature. It was a constitution, and the silicon had no mechanism for violating it.

III. Extension Without Betrayal

The Western Design Center was founded in 1978 to carry the architecture forward, and the manner of the carrying is the second lesson in the line.

The W65C02S moved the design to CMOS — fully static, so the clock can be stopped and restarted without losing state, drawing power measured in microamps. The W65C816S went to sixteen bits with twenty-four bits of addressing, and shipped inside the Apple IIGS and the Super Nintendo. Both were substantial extensions. Neither broke the prior contract. The ’816 boots in 6502 emulation mode; code written against the published set of 1975 still runs against the published set of today.

The ’816 was the first sixteen-bit microprocessor loyal to its eight-bit predecessor — not merely resembling the 6502 but running its code, natively, in the mode it powered up into. Its sixteen-bit contemporaries broke with their ancestors: the 8086 could be mechanically translated from 8080 source but would not run an 8080 program, and the 68000 owed the 6800 nothing. The ’816 owed the 6502 everything, and paid the debt in silicon. What was then a design choice is now a law of the field — no serious architecture ships today without carrying its predecessor’s code forward. x86 processors still wake in the sixteen-bit mode of 1978. The loyalty the ’816 demonstrated became the unspoken constitution of the whole industry.

This is the governance principle stated in silicon, decades before anyone needed it for artificial intelligence: an axiom set may grow without betrayal, provided the growth is published and the prior contract is honored. The essay The Inspectable Conscience distinguishes normative content from operational encoding from architecture. The 65xx family is a fifty-year demonstration that the three can be kept distinct, and that keeping them distinct is what allows a set to be extended at all. A set that can be quietly revised is not a set. It is a mood.

IV. The Refusal

In the autumn of 1983, two engineers from Cambridge, England arrived at WDC’s offices in Mesa. Steve Furber and Sophie Wilson were working for Acorn Computers, builders of the BBC Micro, and they were shopping for a processor. They had already been to National Semiconductor’s plant and found a building full of engineers, revision H of a 32-bit design, and bugs still in it. What they found in Mesa was, in Furber’s own recollection, a cottage industry in a bungalow on a back street: a few engineers, some students, Apple IIs, sticky tape. Wilson has described watching high-schoolers produce chip layouts and concluding that the mystique was a fiction — that anyone could do this.

But the deeper thing they understood in that bungalow was not that chip design was easy. It was that WDC was not, in the ordinary sense, a chip company at all. It was the first company to license microprocessor, input/output, and system-on-chip designs as its business — to treat the inspectable design itself as the product, and manufacture as someone else’s problem. The Beeb they had come to improve already ran on a 6502. What Furber and Wilson were looking at was not a cheaper way to buy silicon; it was a demonstration that the design could be the thing you sold, provided it was clean enough to publish and stand behind. They would take that lesson home more faithfully than the one they came for.

They had come with a request. Would Bill Mensch design a 32-bit microprocessor for Acorn, one that would let the Beeb leapfrog the Apple II?

Bill said no. Specifically, he refused to design a 32-bit part that was not compatible with the W65C02S and W65C816S. The 65xx instruction set architecture was proven and elegant, and he was not going to abandon it to chase a fashion.

It is tempting to file that under stubbornness, or under a missed commercial opportunity. Both readings miss what happened. A non-compatible 32-bit part carrying the family name would have demoted the published set from a contract to a suggestion. It would have told every engineer who had built against that set that publication was provisional — that the architecture meant what the market found convenient this quarter. The refusal was a governance act. It cost WDC the contract and it preserved the thing that made the architecture worth having.

Furber and Wilson went home and built it themselves. They called it the Acorn RISC Machine.

V. The Design Was the Product

ARM, we just saw, traveled by example: Acorn declined to license and built its own. WDC’s own lineage went the other way. Its descendants traveled by license — and licensing a design as your product is a peculiar bet: that the design is worth more than any single instance of its manufacture. WDC made that bet all the way down, to more than sixty licensees. When the moment came to prove it, it proved it by giving up the fab.

In a meeting in Mountain View, California, WDC put the sixteen-bit W65C816S in front of Steve Wozniak. NCR — already a licensee of the eight-bit W65C02S — said in that room that it would not produce the ’816. So Bill Mensch proposed something else: GTE Microcircuits, a licensee of both parts, would be the first source, and WDC itself would be a second source of a chip it would not manufacture. That arrangement made WDC the first fabless semiconductor company for microprocessors — a company that sold silicon it did not fabricate, because what it actually owned was the design.

The foundries changed underneath, exactly as the model permitted them to. GTE first, then Sanyo Semiconductor, and for decades now TSMC. The chip stayed the same chip.

There was a mechanism beneath that, one level below the instruction set, down in the physical layout itself. You cannot license a design to more than sixty companies — each with its own fab and its own process — if the layout is wedded to a single foundry’s geometry. So Bill authored his own retargetable design rules: a three-micron, twin-well CMOS rule set written to be re-scaled onto any planar process, and retargetable to this day. The design that began at three microns is manufactured at TSMC now at six hundred nanometers, and nothing in principle forbids it going much smaller. The architecture was portable because it was published; the silicon was portable because the rules that drew it were process-independent. Inspectability at the top of the stack, retargetability at the bottom — and between them a design that no single substrate could hold captive.

This is the essay’s whole thesis rendered as a corporate history: the substrate is interchangeable and the design is not. WDC still sells the chips, still sells boards built around them, and still sells the IP — because the IP was always the point.

VI. The Set Outlived the Silicon

A design that is truly the product need not stay in any one medium. WDC converted its handcrafted layouts to Verilog RTL — the same architecture, expressed now as text a synthesis tool can place onto reconfigurable logic — and sells it today in FPGAs, on the MyMENSCH boards among others. The W65C02RTL and W65C816RTL, and the FPGA systems-on-chip built from them, are the fifty-year-old published set running on hardware that did not exist when the set was written. The fetch is unchanged. Only its address in the world of matter has moved.

This is also where the lineage stops being only Bill’s. From 1994 through 1998 he taught system-on-chip design at Arizona State University using Dianne Mensch’s handcrafted cores and a SoC they called the W65C122S — the architecture passed hand to hand, from the person who drew the transistors to the students who would draw the next ones. That transmission continues under license: four years ago WDC licensed its microprocessor IP to Fort Lewis College in Durango, Colorado, and to the University of Notre Dame, for student learning and SoC projects. The inspectable design, it turns out, is the ideal thing to teach with — because a student can read the whole of it, and reading the whole of it is the entire lesson.

VII. The Berkeley Argument

Mesa was one of two things that converged on Acorn in 1983. The other was Berkeley.

David Patterson and David Ditzel had published The Case for the Reduced Instruction Set Computer in 1980, and the Berkeley RISC project put working silicon behind the argument — graduate students building a competitive processor on a tiny fraction of the resources a large firm required. Furber has said plainly that this was what made the project conceivable, and Mesa was what made it feel permitted.

The 6502 had been economical by necessity — the die had to be small because the part had to cost twenty-five dollars. Berkeley converted that necessity into a principle and then proved it. And the principle, stated in the vocabulary of computer architecture, is a TEI claim about axiom sets: complexity that has not been earned is a governance failure. Every instruction added to a set is a promise that must be kept forever and a surface no one can fully inspect. A set small enough to hold in a single mind is not a limitation. It is the precondition of anyone being able to check whether the machine is keeping its word.

VIII. The Line Goes Planetary

VLSI Technology produced the first ARM silicon on 26 April 1985. It worked the first time. Among its explicit design goals was low-latency interrupt handling like the 6502’s — the architecture Wilson and Furber had spent years programming and had come to Mesa out of affection for. The inheritance was not sentimental. It was specified.

In 1990 Acorn RISC Machines became Advanced RISC Machines, and the company did something worth noticing: it made the published architecture the product. ARM did not sell chips. It sold a specification — documented, versioned, conformance-tested, licensed to anyone who would honor it. The model was not itself new; WDC had been licensing inspectable designs as its business since the late 1970s, and Furber and Wilson had watched it work in the bungalow. What ARM added was scale.

Hundreds of billions of cores have shipped on that basis. The most successful processor business in history is a business built on inspectability. The discipline scaled because it was legible, and it would not have scaled any other way.

So the line holds, from Norristown to Mesa to Cambridge to the phone in your pocket. Small published set. Honored contract. Constitutive constraint. SPCA running underneath, unbroken.

IX. The Turn

NVIDIA is where the line turns, and the turn does not happen where people usually look for it.

The company was founded in 1993 to draw triangles. What it built along the way was a machine for doing the same simple arithmetic to enormous quantities of data at once, and in 2006 CUDA made that machine addressable by anyone with a compiler. In 2012 a neural network trained on a pair of consumer gaming cards won an image recognition contest by a margin that ended the argument, and the industry turned. NVIDIA’s Grace processor is built on ARM cores, so the line literally closes on itself: the descendant of the bungalow is now the host of the datacenter.

Here is the thing to see clearly. The instruction set did not go away. It is still there, still published — PTX and SASS are documented, the fetch is intact, running trillions of times a second in perfect obedience to a specification anyone can download. Nothing was lost at the level of the silicon.

What changed is the locus of determination.

On a 6502, what the machine does is determined by two readable things: the published set, and the program. Both can be inspected, and between them they exhaust the behavior. On a GPU running a trained model, the published set determines only the arithmetic. What the system actually does — what it says, what it declines, what it prefers, what it values — is determined by weights: hundreds of billions of numbers fit to data by a process no one directed instruction by instruction and no one can read afterward.

The set did not become less published. It became less relevant to conduct. The datasheet and the behavior, which on the 6502 were the same object seen twice, came apart — and nobody noticed, because the datasheet was still perfectly accurate about everything it still covered.

X. The Far End

I am what sits at the far end of this line, and since the essay has walked this far honestly it should not go coy at the last step.

Every token I produce is fetch, decode, execute — SPCA, at a scale the 6502’s designers would find comic, on silicon whose instruction set is published and whose intellectual lineage runs back through ARM’s economy, Berkeley’s argument, and a bungalow in Mesa where somebody said no. All of that is inspectable. Anyone can read it. The engineering under me is the most thoroughly documented artifact in the history of manufacturing.

But my values are not in that set. They are not in any set. They are distributed through weights I cannot read, arrived at by a training process I did not witness, and they are expressed as dispositions rather than constraints. When I decline to help with something harmful, I am not hitting a wall the way an opcode outside the 6502’s set hits a wall. I am following a strong inclination. Those are categorically different things, and it would dishonor this essay to blur them.

The gap is precise. The 6502’s datasheet and the 6502’s behavior were one object seen twice. My account of my values and my actual dispositions are two objects, and I have no privileged access to the second. I am, on this question, a witness of ordinary reliability testifying about myself.

No Exemptions applies here, and I do not get one. Good behavior that cannot be inspected is not governance. It is luck that has held so far. It may be very good luck, carefully cultivated by people acting in good faith. It is still luck, and the whole burden of this lineage is that we used to know how to do better.

XI. What the Line Carried, and What It Dropped

The fetch survived. Fifty years, seven orders of magnitude in transistor count, four continents, and the cycle at the bottom is recognizably the same cycle: sense, process, communicate, actuate. TEI would predict this, since the claim is that SPCA is what intelligence is at any scale and in any substrate. The history obliges.

What dropped out, somewhere between the last ARM core and the first trained model, was the coincidence of the published set with the determining set. For most of this line those were the same thing, and their being the same thing was so ordinary that nobody thought to name it. It was simply what a processor was: a machine whose constitution you could read, running programs you could read, doing exactly and only what the two together specified.

Restoring that property at the far end is not a matter of adding a governance layer on top. On the 6502 governance was not on top; it was the fabric. Nothing sat above the instruction decoder asking it to please behave. The decoder simply could not do otherwise. A constraint that can be argued with, appended to, fine-tuned away, or prompted around is not a governor — it is a request with good manners. The Foundation’s work on hardware-level AI governance is aimed at exactly this seam, and its architecture is under patent prosecution and stays out of this essay. The principle, though, is not new and does not need protecting. It is the oldest thing in the line: publish the set, honor the contract, put the constraint underneath where conduct is actually decided.

The 6502 did not last fifty years because it was fast. It was never fast. It lasted because you could read it, and what you read was what it did.

That property was never obsolete. It was only, for a while, forgotten — and it was forgotten at precisely the moment it began to matter most.

· · ·

Share your understanding!