Download The Language of Shapes [1990/2026]

Introduction

An Amateur Sees Around a Corner

KRISTEN M. FABIANS

The essay that follows began as a letter, written in 1990 to the editors of Scientific American[1] in answer to two articles in that January’s issue on the old question, “Can machines think?” Its author was not a scientist and says so repeatedly, with the theatrical humility of a man sure he is about to overstay his welcome. What programming he knew he had taught himself as a boy, in BASIC, on a Commodore VIC-20 — a toy computer you plugged into the television, a few kilobytes of memory, a screen twenty-two characters wide. He had set it aside years before; by the time he wrote this, at twenty, he was a poet, reading Rimbaud, Pound, and Sylvia Plath. That is the whole of his laboratory: a machine he had half-forgotten and a head full of poems.

And yet — this is the reason to reprint it — the essay keeps arriving at ideas the field would spend the next thirty years formalizing, and arriving at them from near-total isolation, with none of the vocabulary. So the introduction has two honest jobs: to mark the places where a twenty-year-old with a VIC-20 seems to have seen around a corner, and, just as carefully, to mark the places where he was rebuilding wheels others had already turned. Both are true, and the essay is better read with both in mind than mistaken for pure prophecy.

Start with what it is really about, since the title question is a feint. The essay says “think,” but nearly everything it cares about is alive. Its imaginary machine has a “life instinct.” It has a “death residue” — a memory, passed to the next generation, of how its predecessor died. It has a “panic” mode, a dread of its own decay, an impulse to read water as a threat. This is not a theory of problem-solving; it is a biology. The author is not after a better calculator. He wants an organism that happens to be made of code — and most of what looks prescient follows from that one swerve.

Take the framing argument first — built on Hugh Kenner’s notion of literary “counterfeiting,” and on a mechanical duck. The Turing test, it says, measures the wrong thing: whether we can be fooled, and we are easily fooled — by ducks, by anonymous novels, by (the essay’s best image) a balloon on an invisible string that rises because a man on a roof is pulling it. What the test misses is whether the behavior connects to anything inside. That worry now has a name — the symbol grounding problem — coined the very same year, 1990, by the cognitive scientist Stevan Harnad, whom the author cannot have read. Picture learning Chinese from a Chinese-to-Chinese dictionary, each word defined only by other words you also cannot read: you could shuffle the symbols forever and never touch the world. A machine that only manipulates symbols is trapped in that dictionary. Most people use the point to say machines cannot think; the author takes it as an instruction — then build the grounding — which is roughly the turn the field itself would make, from symbol-shuffling toward machines wired to sight, sound, and action.

His grounding is “shapes”: a geometric inner language in which things are patterns and two things are alike when their patterns overlap. His example — a capital “I” beside a lowercase “i” as little grids of dots — is charming and would not survive a tilted photograph, but the bet beneath it won. Today’s systems turn every word and image into a long list of numbers, a point in a vast space, and read meaning as distance, so that “she looks like Madonna” is a short hop between two points. Engineers call the lists embeddings, and their space has thousands of dimensions rather than the two or three we can picture; but the intuition is the author’s exactly — meaning as nearness, thought as the geometry of shapes. The reigning view in 1990 was the reverse: intelligence as hand-written rules, a tidy stack of if-this-then-that. He bet against it, and it lost.

Then the part I find most striking, because he clearly did not know it was. His machine holds an expectation — he calls it “sanity” — and a fresh perception may only nudge that expectation, never overturn it, so the machine keeps its shape unless a signal is too alien to absorb, at which point it is “confused”; and confusion, he says, is what forces it to think. That runs almost verbatim from the physicist Hermann von Helmholtz, who held that perception is unconscious inference, the brain’s best guess at a world it never meets directly, down to the current theory called predictive processing: the mind as a prediction engine that tells itself what it expects and treats the senses as a correction. The “wall of sanity” is what that theory calls a prior; “confusion” is prediction error; and the claim that surprise, not correctness, drives thought is its claim too. He got there by watching himself perceive.

A few more, quickly. His “hypothetical thinking” — a faculty that runs experiments internally, working out how acid rain corrodes a church dome “without the burden of time,” without waiting for weather — is what we now call a world model: an inner simulator a system can plan inside, imagining consequences before acting. It is among the liveliest problems in the field, and it separates a creature that merely reacts from one that can rehearse.

His chapter on “beauty” is bolder. He imagines the machine pulled toward regular, symmetrical shapes, the pull both pleasurable and useful for making new ideas. That beauty might be measured — that it lives in a ratio of order to complexity — is an old dream; the mathematician George Birkhoff tried to write it as an equation in the 1930s, beauty as order over complexity. The author’s twist is modern: the attraction to compressible pattern is not ornament but a drive, a motor for learning. That is close to what the AI researcher Jürgen Schmidhuber later formalized as artificial curiosity — a system rewarded not by any outside prize but by finding fresh regularities it can compress, so that beauty and discovery become one event. He calls it an “addiction” to certain shapes; Schmidhuber calls it intrinsic motivation. The same machine.

And the closing insistence on embodiment and mortality — that a real mind needs a body, and needs to fear its own death — lands near ideas that took decades to come back. The machine has a “death residue,” a “panic,” an inherited memory of how it died before; it reads water as a threat the way an animal reads a predator. For most of the field’s life this sounded like poetry. But self-preservation as a by-product of having any goal at all is now a live concern in AI-safety work, and the deeper thought has a distinguished recent champion: Geoffrey Hinton, an architect of the machine learning that made all this real, now argues for mortal computation — intelligence bound to one imperfect, perishable piece of hardware, uncopyable, the way a brain is, rather than the endlessly duplicable software we run. The morbid, embodied machine of 1990 is, in that light, early.

One anticipation the field bet against deserves its own line. The essay insists the machine’s inner life is private — its shapes its own, its mission “to us, irrelevant,” its language as opaque to us as a speech of fluttering eyelids would be to a Frenchman. That is an uncanny preview of the great headache of modern interpretability research — the effort to read what actually happens inside these systems — which keeps finding that the useful representations really are a private tangle of superimposed shapes, nothing like human-readable, and that prying them open is a discipline in itself. He proposed as an ideal the very illegibility researchers now treat as the problem. The solipsistic machine whose purposes are “irrelevant to us” is, depending on your mood, the freest thing in the essay or the first faint sketch of the alignment nightmare.

So far, so flattering; here is the other half of the honesty. None of these ideas was new to the world in 1990, and a literary reader deserves the names. That intelligence must be embodied, not a matter of rules, the philosopher Hubert Dreyfus had pressed for two decades, turning Heidegger and Merleau-Ponty against the AI labs in What Computers Can’t Do (1972), ridiculed before he was vindicated. That a machine shuffling symbols understands nothing is John Searle’s “Chinese Room” of 1980 — a man sealed in a room, following rules to answer Chinese notes he cannot read — and Searle is named in the essay, so no claim there. That each creature inhabits its own sealed world of meaning, a cow’s significances not a bull’s, is the biologist Jakob von Uexküll’s Umwelt, from 1934; his emblem was a tick that knows existence through three signals and will wait years for one. That another mind’s interior may be closed to us for good is Thomas Nagel’s, from his 1974 essay on what it is like to be a bat — we cannot know. And the deepest move of all, the machine that keeps itself alive and owes its meanings to nothing outside, was already Humberto Maturana and Francisco Varela’s autopoiesis[2]: a living thing defined as a system that continuously makes itself and, in the same act, brings forth its world. If “genius” means first to think it, then no — and so is nearly everyone.

But that is the wrong measure, and the right one is where the essay earns its keep. The point is not that each brick was new; it is that a twenty-year-old with no training and a toy computer rederived the bricks and stacked a startling number of them into one structure. To reinvent symbol grounding, unconscious inference, world models, and machine curiosity out of BASIC and introspection is more telling than any citation. It is the Ramanujan move — the self-taught mathematician arriving alone at results Europe already held, no less remarkable for it. One should be precise about what that earns him: anticipation is not advancement, and these are seeds, not results — no mathematics, no mechanism. What it reveals is a mind that throws off the right abstractions unbidden, which is rarer than being early to any one of them.

There is exactly one place the essay is flatly wrong, and it is the honest kind of wrong. The author dismisses speed — “I don’t believe speed has much to do with the ‘life’ of the computer” — and shrugs at faster hardware. The last fifteen years have argued the reverse: that raw scale, poured into general methods, tends to flatten the hand-built cleverness he prizes. (Richard Sutton gave this its blunt name, “the bitter lesson.”) He was on the losing side of it. And yet, if you grant that his true subject is life rather than intelligence, the claim recovers. Life is indifferent to speed: a tardigrade suspends itself for years, a Greenland shark idles toward four centuries, a bristlecone pine is a going concern at five thousand, and none forfeits being alive by being slow. He was meeting a claim about capability with an intuition about life, and half-talking past his opponents without knowing it.

Which returns me, briefly, to the poetry. The essay’s deepest claim is barely about computers at all. Its machine’s meanings are “irrational, and, to us, irrelevant”; its inner language is personal; a real mind, it insists, owes us no legibility whatever. That is a claim about individuation — about a thing that resolves its own shape out of a field of tension and answers to nothing outside itself — and it is close to what the French philosopher Gilbert Simondon spent a career elaborating: that we misread being when we treat things as finished objects rather than ongoing processes, and that even machines have a “mode of existence” of their own. It is also, nearly word for word, a description of a poem that refuses to be decoded. The author would spend the decades after this essay making that argument in earnest — in animated poems, in electronic literature, in the criticism of Word Toys: Poetry and Technics — but it is already here, in 1990, dressed as engineering: the seam where his two vocations touched before he had a name for the join.

One housekeeping note. The essay’s final page is lost — it breaks off mid-sentence. Rather than leave it hanging I have supplied an imagined close, bracketed and pitched in the essay’s own voice, so nothing invented is taken for the original. Everything outside the brackets is the author’s; everything inside them is a forgery, offered in the same spirit as the duck.

---

[1]The two articles were John Searle’s “Is the Brain’s Mind a Computer Program?” and Paul and Patricia Churchland’s “Could a Machine Think?”, whose disagreement frames the question this essay answers. —Ed.

[2]From the Greek autos, self, and poiesis, making: a system that continually produces itself. See Humberto Maturana and Francisco Varela, Autopoiesis and Cognition (1980). —Ed.