A short honest history

Artificial intelligence feels like it arrived overnight, somewhere around the time chatbots started writing essays. It did not. The field is older than your grandparents' television, and its history is a repeating cycle of big promises, real progress, disappointment, and quiet rebuilding. Knowing that cycle is the best protection against both the hype and the doom you will hear about AI for the rest of your life.

1950: a question about machines

The story starts before the name did. In 1950 the British mathematician Alan Turing, who had spent the war breaking German codes with machines, published a paper asking whether machines could think. Rather than argue over the word think, he proposed a practical test: if a machine could hold a written conversation so well that a person could not reliably tell it from a human, arguing about whether it really thinks becomes beside the point. So the founding question of AI was a question about conversation, which deserves a pause, because conversation is exactly where the technology finally broke through to the public seventy years later.

1956: a name and a bold bet

The term artificial intelligence was coined in 1956, at a summer workshop at Dartmouth College in the United States. A small group of researchers proposed that every feature of intelligence could in principle be described precisely enough for a machine to simulate it. They thought serious progress might take a single summer. It has so far taken seventy years and counting, which tells you something about how hard the problem turned out to be.

Their confidence was not stupid. Computers had just spent a decade doing things people assumed only minds could do, calculating in seconds what took human teams weeks and breaking codes thought unbreakable, so it seemed reasonable that the rest of thinking would fall the same way. What the Dartmouth group misjudged is that the hard part of intelligence is not the part that feels hard. Long division feels difficult and turned out to be trivial for machines. Recognising a friend's face or following a conversation feels effortless and turned out to be enormously difficult, because your brain does that work below the level you can notice, and nobody could write down how.

This puzzle has a name. It is sometimes called Moravec's paradox, after the roboticist who pointed it out, and the reason behind it still shapes what AI can and cannot do today. Evolution spent hundreds of millions of years tuning the parts of your brain that handle seeing, walking and reading a face, so that machinery runs deep and silent and feels like nothing at all. Formal logic and arithmetic are recent additions, a few thousand years old at most, and they feel like effort precisely because your brain was never built for them. So the tasks that feel hard to you are the ones a machine finds easy, and the tasks you do without noticing are the ones that took the field another fifty years to approach.

Rules first, and the winters

Early AI was rule based. Programmers wrote down explicit instructions: if the patient has a fever and a rash, consider these diagnoses. These expert systems worked in narrow settings but were brittle. The real world has more exceptions than any team of humans can type in, and a system with no rule for a situation simply fails. When results fell short of the enormous promises, funding collapsed. This happened badly enough, twice, that researchers gave the droughts a name: AI winters.

The first chatbot belongs to this era, and its story still matters. In the mid 1960s a program called ELIZA imitated a therapist using simple pattern matching, mostly by turning your statement back into a question, so telling it you felt sad produced a prompt asking why you felt sad. There was no understanding anywhere inside it, and its own creator said so bluntly. Yet people confided in it, and some insisted it truly understood them. Researchers named our tendency to read a mind into fluent output the ELIZA effect, and it has never been more relevant, because today's chatbots are incomparably more fluent than ELIZA while still containing nobody.

So why do winters happen at all? The mechanism is ordinary. Research money follows exciting promises, promises get made in the language of what might eventually be possible, and funders hear next year when researchers mean next generation. When next year arrives without the miracle, budgets collapse everywhere at once, including for the careful unglamorous work that was actually making progress. Every technology suffers some version of this cycle, but AI suffers it worse than most, because thinking machines invite bigger promises than almost anything else humans build.

Learning from data

The field recovered by flipping the approach. Instead of writing rules, machine learning lets a system find patterns in examples. Show it thousands of emails marked spam or not spam and it works out the tells itself. Through the 1990s and 2000s this quietly powered search engines, recommendations and fraud detection. Around 2012, a technique called deep learning, which uses layered networks loosely inspired by the brain, suddenly got dramatically better at recognising images, largely because the internet supplied huge amounts of data and gaming hardware supplied the computing power to train on it.

Two famous board game matches mark the shift from rules to learning. In 1997 the IBM machine Deep Blue beat the world chess champion Garry Kasparov, but it had not learned to play. It won largely through raw computing power, searching enormous numbers of possible move sequences and scoring positions with rules that chess experts helped write. The older game of Go resisted that approach for another two decades, because Go has far too many possible games for brute search to cope with. When the AlphaGo system beat Lee Sedol, one of the world's best players, in 2016, it had learned much of its skill from data and from playing millions of games against itself. Handwritten rules cracked chess, learning cracked Go, and that difference is the whole modern story in two matches.

It helps to name what actually changed, because the same three ingredients drive every leap since. The first is data, the raw examples a system learns from, which the internet suddenly supplied in staggering quantity. The second is computing power, the hardware that does the training, which happened to arrive cheaply in the form of graphics chips built for video games. The third is the algorithms, the methods for turning examples into a working model, many of which had existed for decades and were only waiting for the other two to catch up. Hold those three in mind, because whenever you hear about the next AI breakthrough, it is almost always one of the three getting bigger rather than some entirely new kind of thinking.

The chatbot moment

The same recipe of big data plus big computing, applied to text, produced large language models. From late 2022, chatbots built on them reached hundreds of millions of ordinary users, and AI stopped being a research topic and became a household tool. That is genuinely a big deal. It is also, notice, the same pattern as before: a real breakthrough, wrapped in claims that run well ahead of it. The honest position is that both things are true at once.

You will feel this cycle personally, because you are living through the loudest stretch of it so far. Every few weeks a headline promises that AI will end homework, replace teachers, take every job or wake up and turn dangerous. Some of these claims come from people selling something, some from people who are genuinely worried, and almost all of them describe next generation as though it meant next year. The habit worth building now is the historian's habit: when you hear a huge claim about AI, ask what has actually shipped and works today, then treat the rest as a promise that has been made many times before.

The rest of this topic is about that gap between what has shipped and what is promised. The next lesson opens up how machines learn from data, which is the engine underneath every modern breakthrough, and the lessons after it get specific about what a language model really does and where the technology is strong and weak. History hands you the pattern. The rest hands you the mechanism, so that you can judge the claims for yourself instead of taking anyone's word for them.