1. 1950 — Turing asks the question
In a paper, Alan Turing proposes: let us not argue about whether a machine thinks, because we cannot define that. Let us ask instead whether a person can tell a machine apart from another person in a conversation.
That turn is still decisive today. Ever since, we have measured machine intelligence by performance rather than by internal state — every modern benchmark is a descendant of it.
2. 1956 — the name is born
At a summer research meeting at Dartmouth College, the phrase "artificial intelligence" is used officially for the first time. The participants are optimistic: they believe the bulk of the problem can be solved over one summer.
It could not. But the field got its name, and the research began.
3. 1974–1993 — the AI winters
The same thing happens twice: grand promises, results that fail to arrive, funding that dries up. Researchers start avoiding the term in their grant applications.
Those two decades hold one lesson that still applies. The systems of the time — so-called expert systems — worked by having people write the rules down by hand: "if the patient has a fever and a cough, then…". This worked until the number of rules reached a few hundred. After that it became unmanageable: every new rule contradicted three others.
It was this dead end that pushed the field away from writing rules and towards learning from examples.
4. 1997 — Deep Blue beats Kasparov
A machine defeats the world chess champion. Huge news, but it is worth knowing what actually happened: Deep Blue did not learn. It calculated its way through an enormous number of positions by brute force, according to evaluation rules written by people.
That difference matters. Deep Blue was the peak of the old direction, not the start of the new one. Anyone who wanted a machine that played chess well could build one. Anyone who wanted a machine that worked chess out for itself had another two decades to wait.
5. 2012 — deep learning starts working
At the results announcement of an image recognition contest, it turns out that a many-layered neural network is dramatically better than every previous method.
The idea was not new — neural networks date back to the sixties. Three things had changed: the internet had produced enough images and text to train on, graphics cards had become cheap enough to do the arithmetic, and a handful of practical techniques had come together that made deep networks trainable at all.
From here on, features no longer had to be described by people — the model finds them. That is what Machine Learning, Deep Learning, Generative AI is about.
6. 2017 — the transformer
A research paper introduces a new architecture that throws out the earlier word-by-word crawl. Instead, every word can look at every other word and decide which ones matter to it.
This is what made longer texts manageable and — at least as importantly — parallel training possible. With the earlier methods, training was tied to sequence; with the transformer, thousands of graphics cards could work on the same model at once.
From then on the size of models and the computation spent on them grew by orders of magnitude year after year. Every current language model is a descendant of that architecture — the modern meaning of the word "AI" was essentially born in 2017.
7. 2022 — the general public arrives
At the end of November a free, chat-based service appears and gains a hundred million users in two months.
Technologically, no breakthrough had occurred: the model was an existing variant of an existing architecture. The interface was the novelty. Until then this required programming knowledge; after it, you only needed to be able to type. That moment is not a story about technology but about access.
What has happened since
Four directions in which today's systems differ from those of 2022:
- Multimodality. Models handle not only text but images, audio and video, in the same train of thought.
- Open weights. Downloadable models you can run on your own servers have appeared, with a shrinking gap behind the closed frontier.
- Thinking effort. Having the model work something out at length before answering was first a separate model category and is now a dial.
- Agents. Systems no longer only answer, they operate tools and inspect their own results. This is the current direction — and What Is an AI Agent? goes into it properly.
What to take away
Two patterns repeat across these seventy years, and both still hold.
Hand-written rules always hit a wall. After a few hundred of them they become unmanageable. This is as true of a fifty-rule prompt today as it was of the expert systems of the eighties.
Breakthroughs are rarely new ideas. The neural network is sixty years old, yet the current boom came from there being enough data and cheap enough computation for it. Which is why it is worth being sceptical when someone announces a "revolutionary new approach": real change usually comes not from the idea but from something suddenly becoming affordable. Applying the same yardstick to the years ahead, where AI is heading separates four measured trends from four claims with no evidence behind them.