In July 1958, in a room in Buffalo, New York, a five-tonne computer the size of a wall was fed a stack of cards. Some were marked on the left, some on the right. Nobody told the machine which was which. It guessed, was told each time whether the guess had landed, and adjusted itself a little. After about fifty cards, it could sort them on its own. A man named Frank Rosenblatt had built a machine that did something no calculator before it could: it learned from examples.
He called it the Perceptron, and he had built it loosely on an idea borrowed from biology — the neuron, a single brain cell. A brain cell takes in signals from its neighbours and only fires off its own signal when those add up to enough. Rosenblatt copied that one small move in wiring and maths — and the machine that resulted could be taught rather than programmed.
The press lost its head. The New York Times reported on a "Navy device" that "learns by doing" — an "embryo of a computer" that, the Navy expected, would one day "walk, talk, see, write" and even be conscious of its own existence. The New Yorker called it the first serious rival to the human brain ever built. A machine had nudged a few numbers, and the world heard: thinking machines are here.
The first Perceptron wasn't only a program. The Mark I Perceptron was a physical machine: a grid of light sensors looking at a card, wired to banks of dials whose settings were the "weights" — and small electric motors turned those dials as the machine learned. Its weights were literally adjusted by motor. It survives today in the Smithsonian.
The idea of an artificial neuron is older than the Perceptron itself. In 1943 — before any computer you could program existed — two researchers, Warren McCulloch and Walter Pitts, worked out on paper how a network of simple on-or-off "neurons" could, in principle, carry out logic. The brain-cell-as-tiny-switch idea was sitting in a maths paper for fifteen years before a machine ran it.
Strip away the word "brain" and a single artificial neuron is almost embarrassingly simple. It does four things, in order, and that is genuinely all:
A few facts come in as numbers. Is it cloudy? Is rain forecast? Each one is an input.
Each input is multiplied by a weight — a number that says how much that fact matters. A big weight, a loud vote.
All the weighted inputs are summed into a single total. One number now stands for the whole situation.
If that total passes a line called the threshold, the neuron fires: yes. If not, it stays quiet: no.
That's it. Inputs in, each scaled by its weight, summed, and checked against a threshold — out comes a single decision. Picture the most everyday choice: do I take an umbrella? "Rain forecast" deserves a heavy weight; "it's a bit cloudy" a light one. Add the votes, and if they clear your threshold for bothering, you grab the umbrella. You have run a neuron in your head, without the maths.
A real neuron in your head is wildly more intricate than this — it's a living cell with branching arms called dendrites bringing signals in, and it passes its signal to the next cell across a tiny gap called a synapse. You're carrying around roughly 86 billion of them. The artificial version keeps only the single idea — add the inputs, fire past a threshold — and throws the rest of the biology away.
Here is that whole little machine as a picture. Read it left to right: two facts come in, each gets multiplied by its own weight, the results are added up in the middle, the total is checked against a threshold — and a single answer comes out the other side.
Tap each card — what does that part of the neuron actually do?
A "neural network" sounds grand, but at its simplest it's many of these tiny neurons wired together — the output of one becoming an input to the next. No single neuron is clever. The behaviour you'd call intelligent comes from a great many simple parts connected up. You'll open one of those networks properly in Lesson 7.
Now the part that made the headlines: learning. For a neuron, learning means adjusting the weights until the answers come out the way the examples ask for. Let's tune our umbrella neuron by hand. The inputs are cloudy? (1 for yes, 0 for no) and rain forecast? (1 or 0). The threshold is fixed at 2: if the weighted total reaches 2, the neuron says take the umbrella. Reveal each step and predict the answer before you move on.
In Lesson 1, Babbage's engine did precisely what its cards told it — step by step, no more, no less. Here, nobody wrote the steps for "take an umbrella". We gave examples and let the machine find its own weights. That shift — from being told the rules to finding them from examples — is the door into everything that comes next, and Lesson 6 walks straight through it.
The weights don't have to be whole numbers, and the nudges are usually tiny — a real network might shift a weight by a thousandth after each example. Learning isn't one big leap; it's a slow drift of thousands of little dials, each barely moving, until the whole thing quietly lands in the right place.
The promise was enormous. The reality had a hard edge. In 1969, two respected researchers — Marvin Minsky and Seymour Papert — published a careful book called Perceptrons, and they proved, with plain mathematics, that a single perceptron could not learn certain very ordinary patterns. The famous one is called XOR — a rule that says "yes when one thing is true, or the other, but not when both are, and not when neither is." Try as you might, you cannot set two weights and one threshold to capture it. One neuron isn't enough.
We now know the answer was to stack neurons into layers — and that's exactly where this story goes later. But at the time, the limit landed like a verdict. Funding for this kind of research dried up. Excitement curdled into doubt. The field slid into what's now called an AI winter: a long cold stretch when the grand promises of the 1950s were quietly filed away, and the words "artificial intelligence" became almost embarrassing to say. There were two such winters — through much of the 1970s, and again in the late 1980s.
"The first serious rival to the human brain."
Soon to walk, talk, see, write — and know it exists.
Thinking machines: basically here.
It learned to sort cards left from right.
A single one couldn't even learn XOR.
The real breakthroughs were fifty years away.
Here is the thinking-move this whole lesson exists to teach you, Florence — and it may be the most useful habit you can carry into a future full of AI: tell the hype from the reality. When something new and astonishing appears, two questions cut through almost all the noise. What did it actually do, exactly? And what is being promised on top of that? The Perceptron genuinely learned — that part was real and important. The talking, conscious machine was a promise stacked on top, and the promise was decades early. Learning to hold those two apart is how you stay clear-eyed.
The "winter" was real, but the idea never fully froze. Rosenblatt himself died young in 1971 and didn't live to see it — yet the very weight-nudging he pioneered is, in spirit, still how today's systems learn. Sometimes a thing isn't off the mark; it's only early, waiting for the rest of the world (faster machines, more data) to catch up. That's the second quiet lesson here, and worth keeping: persistence through the winters.
Back in Lesson 1, Ada Lovelace drew a firm line: a machine, she said, "can do whatever we know how to order it to perform" — it follows instructions, it doesn't originate. For a hundred years that felt airtight. And for Babbage's engine, it was exactly right: every step came from a card.
But sit with what the Perceptron did. Nobody wrote the rule for "left card" or "right card". Nobody could have — they didn't know the rule themselves. The machine found its own weights from examples. So did someone "order it" to sort cards? In a sense, yes — they built it and showed it examples. In another sense, the actual rule it ended up using was something it arrived at, not something a person wrote down. That's a genuinely new wrinkle in Ada's neat line.
It's worth being careful here, not breathless — this is the very thinking-move from the last page, turned on ourselves. A perceptron isn't dreaming up ideas; it's tuning numbers to fit examples we chose, toward a goal we set. That's a long way from "originating" anything. But it's also not plainly "doing what it's told" in the way a clockwork engine does. The truth lives in the awkward middle — and noticing that, rather than reaching for the dramatic version, is exactly the habit this subject is trying to build in you.
The book that helped trigger the winter, Perceptrons, wasn't an attack out of spite — Minsky and Papert genuinely admired the idea. The cruel irony is that the very next year, 1970, Minsky won the Turing Award, computing's highest honour. The argument that cooled the field was made by one of the field's own giants — which is often how science actually moves: not by outsiders, but by insiders unafraid to say "show me the proof."
Fresh one. The line the weighted total has to cross before a neuron fires "yes" is called the ______. What's that word?
Fresh one. Cloudy = 0, rain forecast = 1, with w1 = 1 and w2 = 2. Work out (0×1) + (1×2). What total?
In 1958 the newspapers said a machine that had learned to sort cards would soon walk, talk and think. The learning was real; the rest was a promise that took fifty years to even start coming true. You're growing up surrounded by big claims about AI — in adverts, in the news, from companies with something to sell. So make your case: when someone announces a remarkable new AI, how would you work out what's genuinely real and what's hype? There's no single right method. Give your way of telling them apart, and use an example — the Perceptron, or something you've seen yourself.
Start with your test in one sentence — how you'd tell real from hype. Then show it working on one example. A single example, used well, beats a long list.
strong You led with a real test, not a vague feeling — and that's the spine of clear thinking about any claim. Splitting what a thing actually does from what's being promised on top of it is precisely the move this lesson was reaching for, and you put it in your own words.
try this The place to push is the example. You name one — now make it carry more weight: walk through the exact moment where the real capability stops and the promise begins. Pinning that line down is where the argument earns its keep.
to add One question worth holding for next time: who gains if you believe the claim? Asking who benefits from the hype is a quiet, powerful test — and it pairs neatly with the one you already wrote.
He was born in 1928 and grew up close to New York City. He trained not as an engineer but as a psychologist — and that's the key to him. He wasn't really trying to build a clever gadget; he was asking a deep question about the brain: what is the smallest set of things a brain has to do to manage the astonishing things it does? The Perceptron was his attempt at an answer, built in metal.
He was, by every account, a man of huge and restless curiosity — he built an observatory in his garden, took up mountaineering, loved music and politics. He believed in the Perceptron completely, rated its importance "a 10 out of 10", and kept faith in it even as the funding and the headlines drained away and his rival Minsky's book took hold.
He died young — he drowned in a sailing accident in 1971, on his 43rd birthday, deep in the AI winter, never knowing he'd be proved right. Because he was. The weight-nudging idea he championed sat in the cold for decades, then thawed and became the beating heart of modern AI. Today there's an IEEE Frank Rosenblatt Award in his name. He had the Perceptron right all along. He was early — by about fifty years.
Rosenblatt's follow-up machine was built to recognise speech, and he named it Tobermory — after a talking cat in a short story. He was trying to give a machine ears and a voice in the early 1960s. That dream, too, would have to wait decades — and you'll meet the machines that finally cracked language in Lesson 9.
A short, beautiful film from 3Blue1Brown. Don't worry about the maths that flickers past — watch for the one idea you already know: tiny neurons, each adding up weighted inputs, wired together in layers. Seeing many of them light up at once is the picture Lesson 7 will open up properly.
For the human story behind the science — how a dream cooled into an "AI winter" and then thawed — the Royal Institution's free public talks on the history of AI (search their channel for "the history of AI" or "machine learning") are calm, careful and made for a general audience. Good for ages 13 and 15 alike, and a nice one to watch with Dad.
You saw the Perceptron learn to sort cards in 1958, and watched the headlines promise a thinking machine that was fifty years early. You took a neuron apart — inputs, weights, a sum, a threshold — and tuned one by hand until a single nudge flipped its answer. You met the limit Minsky and Papert found, and the long winter that followed. And you practised the habit that matters most for everything ahead: telling the real thing from the hype. The Perceptron could learn — but how does a machine actually learn from thousands of examples, well enough to know a cat when it sees one? That's exactly where we go next, Florence. Lesson 6.