He never met her eyes. He watched the bronze, and the bronze was enough
Of all the monsters, only Medusa could not be fought. Her weapon was not her claws or her wings — it was being seen. One direct look and the looker was stone. So when Perseus went to kill her, Athena gave him a strange weapon for a duel: a shield polished until it worked as a mirror. He walked into her cave backwards, watching the fight in the bronze — her reflection, flat, reversed, slightly the wrong colour — and struck by it, and it was enough. Here is the detail everyone forgets: he won without ever seeing her. Everything he knew about where she stood and how she moved, he built from what a curved sheet of metal happened to show him.
That is not a metaphor for how a machine sees. That is how a machine sees. No machine has ever met the world; a camera hands it a flat grid of numbers, and everything it will ever know it must build from arithmetic on that grid. This piece opens the grid up: what is actually in it, how edges and parts and whole things are conjured out of it, the strange dust that can fool it completely — and then the part of the story that is our own working day at Phonetico: teaching the shield to read the Ge'ez fidel off pages that nobody, anywhere, ever typed.
Written for ages 12+, and for any adult who has been told a computer “recognised” their face and never once been told what, exactly, did the recognising. The real engineering words are all at the bottom.
the claim
Seeing is not receiving
Your own eyes feel like windows: open them and the world pours in, cups and faces and doorways, already sorted and named. They are not windows. The back of your eye sends your brain a storm of tiny electrical sparks, and every cup and face and doorway you have ever “seen” was assembled from sparks, behind your forehead, without you being told. You live inside a finished reconstruction and you have never once been shown the raw material.
The machine's situation is the same, but stripped so bare it becomes honest. A digital camera divides its view into a grid of tiny squares and, for each square, measures three things: how much red light fell there, how much green, how much blue. Three numbers per square. A photo from your phone is about twelve million squares — call it thirty-six million numbers — and that is the entire delivery. No depth arrives. No objects arrive. No names arrive. Somewhere in those thirty-six million numbers there may be your grandmother's face, but nothing in the grid says so, any more than the bronze of the shield said Gorgon. The shield did not show Perseus a monster. It showed him bronze, arranged. The whole field called computer vision is one question asked for seventy years: how do you get from bronze, arranged to there she is — strike now?
One more thing the myth quietly gets right. On the road to Medusa, Perseus met the Graeae — three ancient sisters who shared a single eye, passing it from hand to hand. A camera lives the Graeae's condition permanently: one eye, no partner. You have two eyes, set a little apart, and the small disagreement between them is where your feeling of depth comes from. A single camera has no disagreement to consult. Its world genuinely is the reflection in the shield — flat, painted on glass — and any depth a machine reports, it inferred. It was never given it.
one picture, from a handful of squares to nine thousand. drag the gold square and read the numbers
Each pixel is the average of all the light that fell in its little square — the instrument computes it that honest way, nine samples per square — which is why small pictures look soft rather than jagged. And the numbers panel is the last time this page will show you the raw grid. Everything after this is built on top of it, and nothing after this ever sees anything else.
step one
An edge is a disagreement
So you hold nine thousand numbers. What is the first honest move? Not “find the pot” — nothing in the grid knows what a pot is. The first move ever found, and still the first move today, is almost insultingly small: subtract neighbours. Take each pixel and ask how much it disagrees with the pixel next door. Where the answer is near zero, the world is flat there — sky agreeing with sky, clay agreeing with clay. Where the answer is large, something ends: the rim of the pot against the morning light. The world draws its own outlines, if you know to subtract.
The tool that does the asking is a stencil: a tiny grid of numbers, usually three by three, that is slid across the whole picture and asked the same small question at every neighbourhood — multiply the nine pixels under me by my nine numbers and add it up. One stencil, one question, nine thousand askings, and out comes a new picture: a map of everywhere the question was answered loudly. A stencil of all positive numbers smooths. A stencil that is negative on the left and positive on the right finds every upright edge in the picture at once.
For half a century, engineers designed these stencils by hand — the famous edge stencil in the instrument below carries the name of the man who tuned it in 1968 — and hand-built vision spent decades stuck roughly where hand-built speech was stuck in piece one, and for the same reason. The neural turn was the same turn it always is in this series: stop typing the stencils in. Give the machine a pile of pictures and let it learn which stencils are worth asking. And here is the finding that still makes engineers quiet: networks trained on nothing but the pile, told nothing about edges, grow edge stencils in their first layer, every single time. We know because you can open a trained network and look at them. The disagreement trick is not a human convention. It is apparently what seeing is made of, whoever builds it.
one stencil, nine numbers, slid over every neighbourhood. watch the answers become a map
A real first layer holds not one stencil but dozens — edges at every angle, spots, little gratings — all asked everywhere at once. And the stencil is the reason this machinery is called a convolutional network: convolution is simply the mathematician's word for “slide a stencil and add things up.” You have now done the thing the name is about.
step two
Glints, outlines, parts, things
One layer of edges is not sight. A map of disagreements still contains no pot, no face, no Gorgon. The move that turned this machinery from a curiosity into the thing that now reads your photos is the stacking: the second layer runs stencils over the first layer's maps, not over the picture. The first layer answered “is there an edge here, and leaning which way?” The second layer, reading those answers, can ask something no single pixel could ever support: “do an upright edge and a lying edge meet here?” — a corner. A curve. The end of a stroke. The third layer, reading the second, asks about loops and crossings and hooks; layers above that, about whole parts; until near the top a layer is asking, in effect, spout? — and nothing anywhere in the stack ever saw “the whole picture.” Every layer only ever answered small questions about the layer below. Sight, built this way, is questions stacked on questions, and the bottom of the stack is subtraction.
And the top of the stack is an old friend. By the final layer, the picture has been boiled down to a list of numbers that no longer describes pixels at all — it describes what was found. It is a place on a map: the word-map of piece two, the meadow of voices of piece eight, the meaning-map of piece nine, now wearing its fourth coat. Pictures of the same thing land close together on it, whatever the lighting, whatever the angle, however worn the ink. Recognition — the entire imposing word — is nothing but nearness: the machine stores where known things landed, and answers with whatever landed closest. One trick. The whole series keeps arriving at it because the field kept arriving at it.
three letters of the fidel, read layer by layer. then shake the pen and watch the stack degrade honestly
The letters are ሀ, ለ and መ — from the top of the traditional order every Ethiopian schoolchild chants, ha-hu. The real reading machines differ from this toy in depth and in one enormous way: their stencils and their map were learned from the pile, not fixed by us. What does not differ is the shape: layers of small questions, then nearness at the top.
the trouble
The dust only the shield can see
Now for the two ways this kind of sight fails, because they matter more than the success stories, and because both are already out in the world misreading things. The first is the strange one. The machine's certainty is arithmetic on numbers — you have now watched every stage of it — and arithmetic can be aimed at. It turns out there exist changes to a picture so small your eye cannot use them — a grain of brightness here, a grain there, spread thinly across thousands of pixels — that are nevertheless arranged to push the picture's place on the map straight toward a different answer. Not noise: a direction. Scatter the very same grains at random and nothing happens; the instrument below lets you prove that with one button. The danger was never the amount of dust. It is the aim.
A machine that meets an aimed picture is Perseus looking up from the shield. It does not become confused — confusion would be survivable. It becomes certain: frozen into the wrong answer at full confidence, without one flicker of doubt, because as far as its arithmetic can tell, nothing whatsoever is wrong. You have met this creature twice before — the Chimera of piece four, the confident nonsense of piece nine — and here it is with eyes: fluent, confident, wrong, now in the visual world, where it can be aimed on purpose by someone who wants a face missed or a sign misread.
The second failure needs no attacker, which makes it commoner and quieter. Ask what the machine actually learned from its pile, and the honest answer is: the cheapest thing that worked. If every cow in the pile stood on green grass, the machine may have learned green, not cow — perfect scores all through training, right up until the first cow on a beach. Piece five showed you the bed the pile builds; piece seven showed you a learner reading its contract like a lawyer. This is both at once: the pile set the exam, and the machine found the loophole. It did not learn what you meant. It learned what the pile allowed.
dust your eye cannot use, aimed along a direction. then scatter the same dust and watch nothing happen
This is a laboratory toy of a real and demonstrated attack: researchers have flipped serious vision systems with stickers on a road sign and with glasses frames on a face. The machine in this instrument is the simplest of the family — nearest stored picture — but the deep stacks fall for the same trick, because they run on the same arithmetic. Defences exist, and it remains an open fight, and the honest sentence is: nobody has closed it.
our job
Teaching the shield to read
So what is this machinery for, told from our corner of the world? Pages. Ge'ez, the script this series keeps returning to, has been printed for five centuries and written for far longer — books, newspapers, church manuscripts, school readers — and the overwhelming majority of it exists only as ink. No file. No text. To a language machine, an unscanned page does not exist; and pieces two and nine showed you what runs on text: Echo's whole education, and every stone in Hermes' roads. Locked in those pages is exactly the pile the Ethiopian languages are missing.
Reading a page with a camera is everything in this piece performed in sequence: the page arrives as a grid of numbers, the stencils raise the strokes out of the paper, the layers assemble strokes into letters, and each letter is answered by nearness on the map. The machinery has a dull name — optical character recognition — and a beautiful job: it turns ink back into language, at a speed no human copyist could survive, piece one's Sisyphus problem all over again.
And the pile decides, one last time. Train the reader on clean printed type — the easy pile to make — and it will read clean printed type superbly, and fall apart on the faded, the stained, the smudged: precisely the pages most worth saving, because the oldest pages are the ones no second copy exists of. Procrustes does not spare books. The fix is the same fix it has been all series: the pile must contain the world as it actually is. Worn pages, taught on purpose. The instrument below runs that exact experiment, and the difference it computes is not a special effect.
forty-two letters on a worn page, one reading machine, two piles. widen the pile and count again
This is our actual working day, miniaturised. The trade even has a name for wearing clean pages on purpose to widen a pile — augmentation — and piece six's warning applies in full: the worn pages you test on must never be the worn pages you trained on, or the score is Narcissus at the pool. The instrument obeys; its test wear uses seeds the pile never saw.
the last thing
The head in the bag
The myth does not end at the cave, and the ending is the part to carry out of this piece. Perseus did not leave the head where it fell. He carried it away in a leather bag, and it never stopped working: he used it — on a sea monster, on a tyrant king — and every use was a choice made by the hand that held the bag. In the end Athena fastened the Gorgon's face to her own shield, the aegis, where it stared out of the armour of a goddess. The most dangerous gaze in the world ended up installed.
That is where machine sight stands now. It is out of the cave and fastened to the world's equipment: it reads tumours off scans that tired human eyes would miss, sorts a billion photos, steers cars around children, and — the use this piece was written for — it can hand a language its own literature back, page by page, at a price small languages can finally afford. And the very same stack, unchanged, follows one face through a crowd of thousands, and does it carrying every unfairness its pile carried, briskly and politely and without the smallest flicker of doubt — you met the bed it was built on in piece five. The machinery cannot tell the difference between reading a manuscript and reading a crowd. It is bronze, arranged. Only the hand holding the shield knows which job it is doing, and the hand, not the bronze, is where the responsibility lives.
So end with the honest version of the whole picture. A machine will never meet the world; neither, if we are strict about it, do you — you also live by a reflection your nervous system polishes for you. What the machine's reflection lacks is not sharpness. It is stakes: nothing in the arithmetic knows what a face costs, or what a page is worth. People know that. Which is why the answer to “should this machine be looking at that?” is never inside the machine — and why every one of these pieces ends at the same place: the ideas are explainable, the machinery is buildable, and the choices, all of them, are still ours.
the glossary
The words the engineers use
| the polished shield | the camera sensor; the image plane |
| bronze, arranged | a digital image: a pixel grid, three channels (RGB) |
| three numbers per square | pixel intensity values, 0–255 per channel |
| each square averages the light that fell in it | sensor integration; anti-aliasing by area sampling |
| the Graeae's single eye | monocular vision: a flat projection, depth not received |
| two eyes disagreeing slightly | stereo disparity, the basis of depth perception |
| an edge is a disagreement | an intensity gradient |
| the stencil | a convolution kernel (filter) |
| sliding the stencil everywhere | convolution |
| the 1968 stencil with a man's name | the Sobel operator (hand-designed feature era) |
| stencils grown, not typed | learned features; first-layer filters of a CNN |
| layers reading layers | a deep convolutional network; feature hierarchy |
| less fussy about exactly-where | pooling; translation tolerance |
| the place on the map | an embedding (the image representation) |
| recognition is nearness | classification in a learned representation space |
| the aimed dust | an adversarial perturbation / adversarial example |
| amount harmless, aim fatal | robustness to random noise vs adversarial directions |
| green means cow | shortcut learning; spurious correlations |
| the pile | the training data distribution |
| reading ink back into language | optical character recognition (OCR) |
| wearing clean pages on purpose | data augmentation |
| test wear the pile never saw | train/test separation (piece six, forever) |
| boxes around things · naming every pixel | object detection · semantic segmentation |
| the head, installed | deployed vision systems; surveillance and its governance |
One sentence to take away
A machine never meets the world — it receives a flat grid of numbers and builds everything from arithmetic on them, edges out of disagreements, parts out of edges, recognition out of nearness on a learned map — and it does this exactly as well as its pile of pictures allows, can be flipped by dust aimed along directions no human eye can use, and reads exactly the pages, and the faces, that the hand holding it points it at.
Perseus never looked directly at her. The shield was enough — but the shield never chose the target. The hand did.
the sources
Where these ideas come from
The shield is Ovid's; the arithmetic on the bronze is from these papers. Each note says what the paper actually established.
- Y. LeCun, L. Bottou, Y. Bengio & P. Haffner, “Gradient-based learning applied to document recognition”, Proc. IEEE (1998). Stencils grown, not typed — decades early: convolutional networks trained end to end, reading millions of bank cheques commercially while the field still carved its stencils by hand.
- A. Krizhevsky, I. Sutskever & G. Hinton, “ImageNet Classification with Deep Convolutional Neural Networks”, NeurIPS (2012). AlexNet: the layered readers, on GPUs, taking the benchmark's error from 28.2% down to 16.4% — about a 40% cut, in one competition, and the result that ignited everything the modern shield can do.
- C. Szegedy et al., “Intriguing properties of neural networks” (2013) and I. Goodfellow, J. Shlens & C. Szegedy, “Explaining and Harnessing Adversarial Examples” (2015). The aimed dust, discovered and then explained: imperceptible changes that flip a confident answer, and why amount is harmless while aim is fatal.
- R. Geirhos et al., “Shortcut learning in deep neural networks”, Nature Machine Intelligence (2020). Green-means-cow, unified: networks latch onto whatever cue happens to work on the pile — texture, background, position — and fail the moment the world stops cooperating.
- S. Rijhwani, A. Anastasopoulos & G. Neubig, “OCR Post Correction for Endangered Language Texts”, EMNLP (2020). The torn page, in the literature: general-purpose OCR collapses on scanned books in low-resource languages, and a small corrective model trained on scarce data cuts the errors by a third — the published cousin of our own working day.