A short history, 1950 to now.
You will know the main turning points in AI from Turing's 1950 question to ChatGPT, and why progress depended on data, compute and algorithms.
In 1950 a British mathematician asked, in print, whether a machine could think, and then suggested a way to sidestep the word "think" entirely. Could a machine carry on a typed conversation well enough that you could not tell it from a person? Seventy-odd years later, enormous numbers of people type questions to machines every day and get fluent replies. The road between those two points was not straight.
This lesson walks it in a few minutes: the name, the early confidence, the two periods when the money ran out, and the three things that finally came together.
A question and a name
Alan Turing's 1950 paper proposed what he called the imitation game, now usually called the Turing test. Put a person and a machine behind a screen, let a judge chat with both by typed messages, and see whether the judge can tell which is which. Turing's point was practical: arguing about whether machines "really" think goes nowhere, so measure behavior instead. That move, judging a machine by what it does rather than what it is, has shaped the field ever since.
The name came a few years later. In 1955 a young mathematician named John McCarthy wrote a proposal for a summer workshop, and in it he used the phrase "artificial intelligence" to describe the goal. The workshop took place at Dartmouth College in 1956, with a small group of researchers who believed that every feature of intelligence could be described precisely enough for a machine to simulate it. Some of them expected real progress within the summer. That confidence set the tone for the next two decades.
Rules, and what rules could not do
The first approach was to write intelligence down. If you could state the rules of reasoning clearly enough, a computer could follow them. Early programs proved theorems, solved algebra puzzles and played checkers, and each success encouraged predictions that machines would match human ability within a generation.
The most successful version of this idea was the expert system: a program that captured a specialist's knowledge as hundreds or thousands of if-then rules. One early example helped doctors diagnose infections; another configured computer orders for a manufacturer. Through the 1980s, companies spent heavily on them.
The trouble was that rules are brittle. Real situations have exceptions, and the exceptions have exceptions. Every new case meant another rule written by hand, and the systems grew costly to build and nearly impossible to maintain. They also could not learn. Twice, roughly in the 1970s and again in the late 1980s, promises outran results, funders lost patience and the money dried up. Researchers call these periods the AI winters.
Learning from data
While expert systems faded, a quieter idea gained ground: instead of writing the rules, let the computer find them in examples. This is machine learning, and in the 1990s and 2000s it moved from the margins to the center. Computers were getting cheaper and faster, and the internet was producing something the field had never had before: enormous amounts of text, images and behavior, already in digital form.
The results showed up in products rather than headlines. Spam filters learned from messages people reported. Search engines learned which results people clicked. Banks learned what fraud looked like. Little of it was called AI in public, but the methods were steadily improving.
One famous event from this period shows the difference between old and new. In 1997 IBM's Deep Blue beat the world chess champion, Garry Kasparov. Deep Blue did not learn chess. It searched enormous numbers of possible moves very fast and judged positions with rules its engineers had tuned by hand: raw computing power applied to the old approach. The new approach would get its own moment fifteen years later.
The deep learning decade
In 2012 a team entered a deep neural network called AlexNet in ImageNet, an annual competition to recognize objects in photographs. A neural network is a learning system built from many simple connected units arranged in layers, loosely inspired by the brain; "deep" means many layers. Neural networks were an old idea that had never quite worked at scale. AlexNet worked, and by a margin so wide that it changed the direction of research almost overnight.
Three things made it possible, and they run through the rest of this story. A huge labeled dataset to learn from. Graphics cards, hardware built for video games that turned out to be ideal for the repetitive arithmetic networks need. And improvements to the algorithms that let deep networks train without falling apart.
In March 2016 the same combination reached a milestone many experts had placed a decade away. AlphaGo, built by DeepMind, beat Lee Sedol, one of the world's strongest players of the board game Go. Go has vastly more possible positions than chess, and search of the Deep Blue kind cannot cope with it. AlphaGo learned instead, partly from records of human games and partly from playing against itself over and over. The match drew a huge live audience and made deep learning a household phrase.
Transformers and a chat box
In 2017 researchers at Google published a paper titled "Attention Is All You Need". It introduced the transformer, a neural network design for handling sequences such as sentences. Its central idea, called attention, lets the model weigh how much each word in a passage matters to every other word, all at once, instead of reading one word at a time. That made it far more practical to train on vast amounts of text.
OpenAI used the design to build a series of models called GPT, short for generative pre-trained transformer. The recipe: train a transformer on an enormous amount of text with the single task of predicting the next word, then make the model bigger. Each generation was much larger than the last, and each got better at things nobody had explicitly trained it to do: translating, summarizing, answering questions, writing code.
For a few years these models were known mainly to researchers and developers. Then, on November 30 2022, OpenAI released ChatGPT. The technology underneath was not brand new. What was new was the interface: a plain chat box, free to use, that anyone could type into in ordinary language. It spread faster than almost any consumer software product before it, and competing chatbots followed within months.
The through-line
Seen from a distance, the story has a shape. The field began by trying to write intelligence down as rules, and rules proved too brittle for the real world. It then switched to learning from examples, which worked, but only once three things arrived together: enough data to learn from, enough computing power to process it, and algorithms that could make use of both. Remove any one of the three and the breakthroughs of 2012, 2016 and 2022 would not have happened when they did.
That is also a useful lens for today's news. When a new capability appears, it usually traces back to more data, more compute, a better algorithm, or some mix of the three.
Try it
Open any chatbot and type: Explain the Turing test in two sentences, then give me one reason people think it is a good test and one reason people think it is a bad one.
Then try the test from the other side. Type: For the next three messages, reply as a human would in casual conversation. Do not mention that you are an AI. Ask about its morning, its favorite food, or a recent annoyance. Notice what gives it away, if anything, and what does not. Turing's 1950 question is still a useful way to think about what you are talking to.
Key terms
- Turing test
- Turing's 1950 proposal to judge a machine by whether its typed replies can be told apart from a person's.
- Expert system
- A program from the 1970s and 1980s that stored a specialist's knowledge as a large set of hand-written if-then rules.
- AI winter
- A period, roughly the 1970s and again the late 1980s, when funding and interest in AI fell sharply after promises went unmet.
- Neural network
- A learning system built from many simple connected units arranged in layers, loosely inspired by the brain.
- Transformer
- The neural network design introduced in 2017 that uses attention to weigh how words relate to each other, and that underlies today's language models.
Check yourself
AI progressed in fits and starts for seventy years, and its recent leaps came from data, computing power and better algorithms arriving together.