You are the operator at the Mark I Perceptron's console.
On 7 July 1958, in a room at the U.S. Weather Bureau in Washington, a 29-year-old Cornell psychologist named Frank Rosenblatt fed punch cards into an IBM 704 — a vacuum-tube machine the size of a room.
Each card was marked on the left side or the right side. Nobody had told the program which was which. It guessed. When it guessed wrong, a rule nudged its numbers a little in the direction of being less wrong. After about fifty cards, it stopped guessing wrong.
That is a small thing and an enormous thing at the same time. Nobody had written the rule for "left." The machine had found it.
The next morning's New York Times ran the story under the headline "NEW NAVY DEVICE LEARNS BY DOING." The article's first sentence is one of the most famous sentences in the history of artificial intelligence:
"The Navy revealed the embryo of an electronic computer today that it expects will be able to walk, talk, see, write, reproduce itself and be conscious of its existence."
— The New York Times, 8 July 1958. Note the careful wording: the reporter attributes the expectation to the Navy. Who exactly promised what is still argued about. Also worth noticing: sixty-eight years later, most of that sentence is still not true.
The thing on the punch cards was software. What you are standing in front of now is the hardware: the Mark I Perceptron, assembled at the Cornell Aeronautical Laboratory in Buffalo during 1959 and shown to reporters on 23 June 1960. It is the size of two large wardrobes, and it is, in a real sense, a machine that changes its own mind.
The Mark I sees through 400 photocells mounted in a 20 × 20 array, sitting where the film would go in a modified view camera. Point the camera at a card under floodlights, and each cell either fires or doesn't. That's it. No edges, no shapes, no objects — just four hundred yes-or-no answers.
On the cabinet, a 20 × 20 panel of lamps showed the operators exactly what the machine was seeing. Here you can draw straight into its eye. Try it. Drag on the grid (or use the arrow keys and Space).
Every yes-or-no lamp above is one number the machine can use. Four hundred numbers is the entire world it lives in. Everything it will ever "understand" has to be built out of them.
Here is the whole trick, and it is small enough to write on your hand. The machine's memory is a set of weights — numbers stored as the physical position of potentiometer knobs, each turned by its own little electric motor. To decide, it multiplies every lit input by its weight, adds everything up, and answers RIGHT if the total is above zero, LEFT if it isn't.
When it answers wrong, the motors for every lit input turn a fraction of a revolution — toward the answer that would have been right. When it answers right, nothing moves at all. That is the entire learning algorithm. Rosenblatt's error-correction rule:
d = the right answer, y = what it actually said, xi = 1 if that photocell was lit, η = how far each motor turns. If it was right, d − y = 0, nothing moves, and the room goes quiet.
Honest footnote: the real Mark I did not wire its knobs straight to the photocells. The 400 cells fed 512 randomly-wired “association units,” and the 512 motor-driven knobs sat on those units' outputs — hence 512 little motors. The console here wires one knob straight to each of the 400 cells, plus one “bias” knob that acts like a cell that is always lit. Simpler, but the learning rule and the arithmetic are exactly Rosenblatt's, computed live in your browser.
Every mistake adds one step. When the machine has found a rule that works, the staircase goes flat — and stays flat. That flat line is what "it learned" actually looks like from outside.
Push the Sabotage slider up on the console above and run it again. Now some of the labels you are giving it are lies. The staircase never flattens, and the memory map never resolves into a clean picture. A model is only ever as honest as the answers it was handed — which is a fact about 2026 as much as about 1960.
Same arithmetic, new job.
Honest footnote.
Strip the card down to two facts: is there a mark on the left? is there a mark on the right? Now ask a different question — not "which side," but "is exactly one side marked?" A human child solves this instantly. Your perceptron will never solve it, and not because it needs more practice.
Below, the four possible cards sit at the four corners of a square. The machine's decision is a straight line across that square — everything on one side gets one answer. Drag the weights by hand and try to draw a line with the two blue corners on one side and the two red corners on the other. Then let the machine try.
You have already met this wall: choose “Exactly one side marked?” on the console above and run 200 cards. The mistakes never stop.
Marvin Minsky and Seymour Papert proved this properly in a 1969 book called Perceptrons. Their real result is deeper than the famous version: a whole family of questions — parity, and telling whether a shape is connected — needs feature detectors that look at more and more of the image at once as the image grows. No amount of training fixes it.
The myth is that they declared neural networks dead. They didn't. What they actually wrote, on page 231, is that they considered it "an important research problem to elucidate (or reject) our intuitive judgment that the extension is sterile." That is a bet, offered openly, with an invitation to prove it wrong.
It took seventeen years. The missing piece wasn't the idea of extra layers — Rosenblatt had devoted two entire sections of his 1962 book to multi-layer and cross-coupled perceptrons. The missing piece was a way to work out which knob to blame when a deep network gets something wrong.
That method is backpropagation, made famous by Rumelhart, Hinton and Williams in 1986. Tick the hidden-layer box above and you are running it: two hidden units each draw their own line, the output unit combines them, and the impossible question falls apart in a few seconds of arithmetic. Everything you have used an AI for this year is a very large descendant of that box.
Frank Rosenblatt did not see it. He drowned in a boating accident on Chesapeake Bay on 11 July 1971 — his forty-third birthday.
Swap the photocells for a camera or a sensor, swap 400 knobs for millions of numbers, and you have the systems quietly working in shops, plants and job sites today. Every one of them was taught the way you taught the console: examples, mistakes, small nudges.
On production lines, vision systems learn from labelled photos to flag a bad weld, a cracked tile or a missing bolt — the same “is this card A or B?” question, asked thousands of times an hour.
Camera systems can be trained to notice a missing hard hat or harness, or a person stepping into a machine's blind spot, and sound a warning. They are only as good as the examples they were shown.
Vibration and temperature sensors on pumps, compressors and HVAC units feed models that learn what “healthy” looks like, so a technician gets a heads-up before a bearing fails.
Someone has to label the examples — and it is usually the person who knows the trade. An experienced inspector marking “good” and “bad” is doing exactly what you did with the teaching buttons. Remember the sabotage slider: sloppy labels make a sloppy machine.
You need: nine small cups or squares of paper in a 3 × 3 grid, a pile of paperclips, and index cards.
On each card, shade in some of a 3 × 3 grid to make either a T or an L. Make twenty. Put one paperclip in each cup to start — that's your memory, and it currently believes nothing.
To answer a card: count the paperclips in the cups matching the shaded squares. Ten or more, say "T." Fewer, say "L."
To learn: have someone flip the card. If you were right, do nothing. If you said "L" but it was a T, add a paperclip to every cup whose square was shaded. If you said "T" but it was an L, remove one from each. Then go to the next card.
Keep a tally of mistakes. Somewhere in the second or third pass through the deck, the tally stops growing — and if you look down at your cups, the cups that make a T-shape will be visibly fuller than the rest. You will have built a picture of the letter T out of paperclips, without ever deciding to. That is the same picture the weight map draws upstairs.
Each input gets a number. Multiply, add up, check the sign: that is the whole decision.
Compare its answer with the right one. Right: nothing moves. Wrong: that gap is the signal.
Nudge every weight that took part toward the right answer. Repeat. If one straight line can do the job, it will find it — and if none can, it never will.
Seven questions. One try each. No pressure — the Navy is only expecting a machine that becomes conscious of its existence.