0:00 On November 30th, 2022, a small company in San Francisco put a chat box on the internet. No ad campaign. No launch event. Just a text field and a blinking cursor. Five days later, a million people had used it. Two months later, a hundred million. No app in history had grown that fast. It was called ChatGPT, and almost overnight, the whole world started asking the same question. Where did this come from? The surprising answer? The idea was not new. It was more than seventy years old. It had been promised, celebrated, funded, mocked, and left for dead. Twice. So if the dream is that old, why did it take so long?
0:45 Why did artificial intelligence collapse, not once, but two times? And what happened in 2012 that suddenly made everything work? This is the entire history of artificial intelligence. In 1950, a British mathematician sat down to write a paper with a strange first line. "I propose to consider the question, 'Can machines think?'" His name was Alan Turing. During the Second World War, he had helped break the German Enigma code. After the war, he worked on one of the world's first electronic computers, in Manchester. It filled an entire room. And Turing believed that one day, machines like it would think. But thinking was impossible to define. So Turing replaced the question with a game.
1:33 A judge types messages to two hidden players. One is a human. One is a machine. If the judge cannot reliably tell which is which, then, Turing argued, we have no good reason to say the machine isn't thinking. He called it the imitation game. Today, we call it the Turing test. He predicted that by the year 2000, machines would fool an average judge about three times in ten. He never saw how close he came. He died in 1954, at just forty-one. Two years later, in the summer of 1956, a group of researchers gathered at Dartmouth College in New Hampshire. Its organizer, a young mathematician named John McCarthy, had used a brand new term in his proposal. Artificial intelligence.
2:21 The proposal was breathtakingly confident. It claimed that every feature of intelligence could, in principle, be described so precisely that a machine could simulate it. And it suggested that real progress could be made if a carefully selected group of scientists worked on it together. For one summer. The guest list became legendary. Marvin Minsky. Claude Shannon, the father of information theory. Allen Newell and Herbert Simon, who arrived with a program that could prove mathematical theorems on its own. Nobody solved intelligence that summer. But the field had a name, a mission, and its first stars. And the stars were not shy. In 1965, Herbert Simon wrote that within twenty years, machines would be
3:09 capable of doing any work a man can do. Two years later, Minsky said the problem would be substantially solved within a generation. The money followed. Much of it came from the U.S. military. Not everyone wanted to program intelligence line by line. A psychologist named Frank Rosenblatt had a different idea. Build a machine that learns, a little like a brain does. He called it the perceptron. It was a simple artificial neuron. It took in signals, weighed each one, added them up, and fired if the total crossed a line. When it got an answer wrong, it adjusted its weights. Again and again, until it got it right. In 1958, the U.S. Navy showed it to the press.
3:55 Fed with punch cards, it learned to tell cards marked on the left from cards marked on the right. A small trick. But the New York Times reported that the Navy expected it to one day walk, talk, see, write, reproduce itself, and be conscious of its existence. Rosenblatt then built the Mark I Perceptron, a cabinet of wires and motors with a camera eye made of four hundred light sensors. It could learn to recognize simple shapes. Neural networks were born. Then, in 1969, Marvin Minsky and Seymour Papert published a book called Perceptrons. It proved that a single-layer perceptron could not learn some very simple patterns. The most famous was a tiny logic puzzle called XOR.
4:42 Is exactly one of these two switches on? A single neuron simply cannot draw that line. The book did not say neural networks were hopeless. But that's how much of the field read it. Funding for neural networks dried up. Two years later, Rosenblatt died in a boating accident, on his forty-third birthday. His idea went quiet for more than a decade. Meanwhile, at MIT, a computer scientist named Joseph Weizenbaum built a program that seemed far smarter than it was. It was called ELIZA. ELIZA played a therapist. You typed that you were sad, and it asked why. You mentioned your mother, and it asked you to tell it more about your family. There was no understanding underneath.
5:27 Just pattern matching and clever reflections of your own words. And yet people opened up to it. Weizenbaum's own secretary, who had watched him build it, asked him to leave the room so she could talk to ELIZA in private. He spent the rest of his life warning that people would trust machines far more than they deserved. Elsewhere, the promises were running into reality. A robot called Shakey could plan a path across a room, but it paused for long stretches to think between moves. Programs that looked brilliant in the lab fell apart in the real world. In 1973, the British government asked the mathematician James Lighthill to judge the field. His report concluded that in no part of the field had the discoveries produced the major impact that was promised.
6:16 Britain cut most of its funding. In America, the military started asking for results instead of dreams. The money vanished. Labs shrank. Researchers avoided the words "artificial intelligence" in their grant applications. This was the first AI winter. The field came back in the 1980s, wearing a suit. The new idea was called the expert system. Instead of building a general mind, you interviewed human experts and wrote down their knowledge as thousands of rules. If this, then that. And for a while, it worked. A system called XCON helped Digital Equipment Corporation configure computer orders, saving it tens of millions of dollars a year. Suddenly every big company wanted one.
7:03 Japan launched a ten-year national program to build intelligent computers. AI was booming again. But expert systems had a weakness. They only knew what someone had written down. They could not learn. The systems grew huge, brittle, and expensive. In 1987, the market for specialized AI computers collapsed almost overnight, as cheaper desktop machines caught up. Companies folded. Japan's grand project ended without reaching its goals. The second winter had arrived. But in the cold, a few researchers kept working on the idea everyone else had abandoned. Neural networks. The problem with Rosenblatt's perceptron was that one layer was not enough.
7:49 Stack several layers, and a network could, in theory, learn almost anything. But nobody had a practical way to train all those layers at once. In 1986, David Rumelhart, Geoffrey Hinton, and Ronald Williams published a method that changed that. It was called backpropagation. The network makes a guess. You measure how wrong it was. Then you send that error backwards through every layer, nudging each connection a little in the right direction. Repeat millions of times, and the network teaches itself. Others had found similar math before, but this paper put neural networks back on the map. At Bell Labs, a young French researcher named Yann LeCun used these ideas to build a network that read handwritten numbers.
8:36 By the late nineties, systems built on his work were reading more than one in ten checks written in the United States. But neural networks were still slow, hungry, and hard to train. Most of the field moved on. The believers would have to wait. Meanwhile, AI was about to get its most famous moment yet, over a chessboard. Chess had been a target since the very beginning. Turing himself had written a chess program on paper. If a machine could beat the best human player alive, surely that would prove something. In 1996, IBM's Deep Blue played world champion Garry Kasparov. Kasparov won. A year later, IBM came back with an upgraded machine that could examine around two hundred million chess positions every second.
9:24 In May 1997, in New York, they played again. In the second game, the machine made a move so subtle that Kasparov suspected people were secretly helping it. He was rattled. In the sixth and final game, he resigned after just nineteen moves. Deep Blue had won the match. For the first time, a computer had beaten the reigning world champion in a match. One magazine cover called it the brain's last stand. But Deep Blue did not learn. It searched, at incredible speed, through options humans had taught it to judge. It was a triumph of raw computing power. And a hint of what raw power could do. By the 2000s, the internet, the subject of our last video, was in every
10:11 home and every pocket. And everyone was leaving behind an ocean of photos, posts, and data. A computer scientist named Fei-Fei Li saw an opportunity. In 2009, her team released ImageNet. A library of millions of photos, each one labeled by hand, by thousands of online workers around the world. Then she turned it into a yearly competition. Show a program a photo. Can it tell you what's in it? For two years, the best programs improved slowly. Then, in 2012, a team from the University of Toronto entered. Alex Krizhevsky, Ilya Sutskever, and their professor, Geoffrey Hinton. The same Hinton who had worked on backpropagation in the eighties.
10:56 Their network, later called AlexNet, was a deep neural network trained on two graphics cards made for video games. Its error rate was fifteen percent. The next best entry was above twenty-six. It wasn't a win. It was a landslide. And this is the moment everything finally clicked. The old idea had been waiting for two things. Huge amounts of data, which the internet now provided. And huge amounts of computing power, which came from an unlikely place. Chips built to render explosions in video games. A race for talent began. And in London, a startup called DeepMind, bought by Google in 2014, was about to take on the hardest board game of all.
11:42 Go is an ancient game played on a nineteen by nineteen grid. The number of possible board positions is greater than the number of atoms in the observable universe. You cannot win at Go by brute force, the way Deep Blue won at chess. You need something closer to intuition. In March 2016, in Seoul, DeepMind's AlphaGo played Lee Sedol, one of the greatest Go players of all time. Around two hundred million people watched. In the second game, on move thirty-seven, AlphaGo placed a stone where no professional would play. Commentators thought it was a mistake. AlphaGo itself estimated that a human would choose it about once in ten thousand times. It turned out to be brilliant.
12:28 One former champion said it was not a human move, and that he had never seen anything so beautiful. AlphaGo won four games to one. It had studied millions of positions, then played millions of games against itself, and discovered ideas humans had missed for centuries. One problem still stood out. Language. Computers could see and play, but they still struggled to read and write like us. They read sentences one word at a time, and by the end of a long sentence, they had forgotten the beginning. In 2017, eight researchers at Google published a paper called "Attention Is All You Need." It introduced a new design called the Transformer. Instead of reading word by word, a Transformer looks at every word at once
13:15 and learns which words should pay attention to which. In "the animal didn't cross the street because it was too tired," it learns that "it" means the animal. Transformers could also be trained on enormous amounts of text, very fast. And a young research lab called OpenAI decided to test a simple bet. What happens if you just keep making them bigger? In 2018, their first GPT model had about a hundred and seventeen million parameters, the adjustable connections inside the network. GPT-2, in 2019, had one and a half billion. It wrote such convincing text that OpenAI at first held back the full model. GPT-3, in 2020, had a hundred and seventy-five billion.
14:03 Each time the model grew, it learned to do things nobody had explicitly taught it. Then came the chat box. On November 30th, 2022, OpenAI took a model from this family, trained it to follow instructions and hold a conversation, and released it as ChatGPT. You know what happened next. The race that followed was the fastest in the history of technology. Google, Meta, Anthropic, and dozens of others released their own models. AI learned to write code, create images and video, and talk out loud. Companies began spending hundreds of billions of dollars on data centers packed with graphics chips. And NVIDIA, the company that made those chips, became one of the most valuable companies on Earth. In 2024, Geoffrey Hinton, the man who kept neural networks
14:54 alive through two winters, won the Nobel Prize in Physics. The next day, the Nobel Prize in Chemistry went to DeepMind's Demis Hassabis and John Jumper, for an AI that predicts the shapes of proteins. Hinton, meanwhile, was warning the world about the risks of what he had helped build. So, why did it take seventy years? Because the idea was right long before the world was ready. Neural networks needed three ingredients. A way to learn, which backpropagation provided. Mountains of data, which the internet delivered. And massive computing power, which came from video game chips. In 2012, all three arrived at the same time. Why did AI collapse twice? Because both times, the promises ran far ahead of the machines.
15:42 Researchers told the world that intelligence was only a few years away. When it wasn't, the money disappeared. Some people believe AI is about to change everything. Others think the hype is running ahead of reality, just like before. But for the first time in seventy years, the machines are actually keeping up with some of the promises. Alan Turing asked whether machines can think. We still don't agree on what thinking is. But today, millions of people talk to machines every day. And sometimes, just like Weizenbaum's secretary, they ask everyone else to leave the room. Subscribe, because the next story is up to you. Vote in the poll on our community tab, and tell me in the comments which moment in AI history surprised you most.
