On the 8th of July, 1958, the New York Times ran a short, astonishing article. The United States Navy, it reported, had revealed the embryo of an electronic computer that it expected would be able to walk, talk, see, write, reproduce itself, and be conscious of its existence.
Read that sentence again. Walk, talk, see, write, reproduce itself, and be conscious of its existence. In 1958. The machine in question was a tangle of wires and light sensors the size of a room, and the man standing next to it was a Cornell psychologist named Frank Rosenblatt. He called his creation the perceptron.
Now, the easy thing to do with a story like this is to laugh. We know how it turned out. The perceptron did not become conscious. It did not reproduce. The whole episode has gone down as maybe the most famous case of overpromising in the history of technology. But here is the part that is easy to miss, and the part that actually matters: Rosenblatt was not a fraud, and he was not a fool. He had built something genuinely new. For one strange, electric moment, it really did look like the beginning of thinking machines. The tragedy of this chapter is not that a con man fooled the press. The tragedy is that a brilliant man was right too early, and the backlash buried him - and his idea - for almost two decades.
This is the story of artificial intelligence’s first boom, and its first crash. It runs from 1958 to 1973. It is a story about two rival tribes who each thought they were about to crack the human mind, about a machine that fooled people into spilling their secrets, and about how a single book and a single government report managed to freeze an entire field of science.
So what was the perceptron, really? Remember, from the last chapter, the idea that a neuron is just a switch - on or off, fire or don’t fire. Rosenblatt took that idea and added the missing ingredient: learning. His machine did not need to be programmed with the answer. You showed it examples - shapes, letters, patterns - and it adjusted its own internal connections, nudging them up and down, until it could sort them on its own. Nobody told it the rules. It found them. And if you have never thought about why that is amazing, think about it now. Every machine before this did exactly what it was told. This was a machine that learned from experience, the way you do. No wonder the Navy lost its mind.
But Rosenblatt had a rival - a whole rival philosophy, really. While he was training his networks to mimic the brain, another group of researchers was chasing intelligence from the opposite direction. Forget the brain, they said. Forget neurons. Intelligence is about symbols and logic and rules. If you want a machine to be smart, don’t grow it - program it. Two of them, Allen Newell and Herbert Simon, had built a program that could prove mathematical theorems. Another, John McCarthy, invented a programming language called Lisp that would be the tool of the field for thirty years. They called their approach symbolic AI, and for the next quarter century, it would be the establishment.
Its showpieces were these tiny, perfect little worlds. The most famous was a program called SHRDLU, built by Terry Winograd around 1970, that let you have a typed conversation with a computer about a tabletop of colored blocks. Move the red block onto the green one. Why did you move it? And it would answer, sensibly, like it understood. Inside its tiny universe of blocks, it seemed almost intelligent. The catch - and this is the catch that haunts the whole era - was that the universe had to stay tiny. The moment you stepped outside the blocks, into the actual messy world, the magic evaporated.
Which brings us to the strangest character in this chapter: a program named ELIZA. In 1966, an MIT professor named Joseph Weizenbaum wrote a simple little script that imitated a psychotherapist. You typed that you were unhappy; it asked why you were unhappy. It had no understanding of anything. It was a parlor trick, and Weizenbaum knew it - he had built it partly to show how shallow the illusion was. And then something happened that disturbed him for the rest of his life. People fell for it. They poured their hearts out to it. His own secretary, who had watched him build the thing and knew exactly what it was, asked him to leave the room so she could talk to it in private. We have a name for this now. We call it the ELIZA effect: the human readiness to see a mind where there is only a mirror. Keep that in the back of your mind. It is going to come back, decades later, at a scale Weizenbaum could never have imagined.
For a while, the money flowed and the promises piled up. And then, between 1966 and 1973, the bill came due, in three blows.
The first hit machine translation. After years of funding and grand claims about computers that would translate Russian overnight, a 1966 government report - the ALPAC report - concluded there was no useful machine translation on the horizon and recommended pulling the plug. Funding collapsed for nearly twenty years.
The second blow was aimed straight at Rosenblatt. In 1969, two of the symbolic-AI heavyweights, Marvin Minsky and Seymour Papert, published a book called simply Perceptrons. It was a rigorous mathematical takedown, and its killer result was almost insultingly simple: a single-layer perceptron, they proved, could not learn to tell whether exactly one of two inputs was on - a basic logical operation called exclusive-or. A child could do it. Rosenblatt’s celebrated machine could not. Now, the limitation was real, but it applied to the simplest version of the network. The conclusion that spread through the field was much broader and much more damaging: neural networks were a dead end. Don’t waste your career on them. And people listened. The funding dried up, the students drifted away, and the approach went into a deep freeze that would last seventeen years. Two years later, in 1971, Frank Rosenblatt drowned in a boating accident on his forty-third birthday, in the middle of the winter he did not live to see end.
The third blow came from across the Atlantic. In 1973, the British mathematician James Lighthill delivered a report to the UK government declaring that artificial intelligence had basically failed to deliver on any of its grand promises. British AI funding was gutted. That spring, the BBC actually televised a debate, Lighthill against the AI researchers, like a public trial of the entire field. And the verdict, more or less, was guilty.
This is the pattern we now call an AI winter - a collapse of money and belief after a season of hype. It would not be the last. But it was the first, and it was brutal. By 1973, the perceptron was a punchline, machine translation was defunded, and Britain had switched off the lights.
And yet. Here is the thing about that devastating Minsky and Papert critique. They had proved that a network with a single layer was too weak. The obvious fix was to add more layers - to stack them. Everyone knew that. The problem was that nobody knew how to train a stack. There was no method for taking the error at the end and figuring out how to adjust the neurons buried deep in the middle. Without that, deep networks were useless. With it, they would eventually swallow the world.
And in 1974, just as the winter was settling in, a graduate student at Harvard named Paul Werbos worked out exactly that method, and wrote it down in his doctoral thesis. The answer to the whole problem. The key that would one day unlock everything.
Almost nobody read it. It sat on a shelf, in the wrong winter, for more than a decade. How that forgotten idea came roaring back to life - and who rescued it - is where the next chapter begins.