What is AI, and where did it come from?

What is AI, really

"Artificial intelligence" is the name we give to computer systems that perform tasks we usually think of as requiring human intelligence — understanding language, recognising images, playing complex games, writing essays, diagnosing illness, driving cars. The phrase has been around since 1956 and has meant slightly different things in different decades, but that core idea has stayed steady.

Here is the most important thing to internalise upfront. Modern AI does not "think" the way a human thinks. It does not have feelings, intentions, or self-awareness — at least not in any way that anyone can demonstrate scientifically. What it has, in 2026, is an extremely sophisticated ability to pattern-match: to look at a piece of input (a question, an image, a piece of code, a clinical scan) and produce a useful output (an answer, a description, a fix, a diagnosis) based on what it has learned from enormous amounts of data.

The fact that the outputs are often very good — sometimes indistinguishable from what an expert human would produce — is what makes AI feel magical. The fact that the system does not actually understand anything in the human sense is what makes AI sometimes fail in surprising ways. Both of these things are true at the same time. Holding both in your mind is the start of AI literacy.

A short history of the idea — and why now

The idea of intelligent machines is much older than computers. Greek myths had them. Frankenstein had them. But the modern, scientific version of the idea begins with one person and one question.

1950 — Alan Turing's question

The British mathematician Alan Turing, in a 1950 paper called "Computing Machinery and Intelligence", asked: Can machines think? He realised the question was poorly defined and reformulated it. Instead of asking whether machines can think, he proposed asking whether a machine could play a conversational game so well that a human judge could not tell whether they were talking to a person or a computer. This is now called the Turing Test. In 1950, the answer was clearly no. In 2026, the answer is increasingly murky.

1956 — Dartmouth

In the summer of 1956, a small group of researchers — John McCarthy, Marvin Minsky, Claude Shannon, Nathaniel Rochester, and others — met at Dartmouth College in New Hampshire for a workshop. McCarthy coined the term "artificial intelligence" for the funding proposal. The workshop produced no breakthroughs, but the name stuck. The field has had a birthday party ever since.

1956 to about 1974 — The first golden age

For roughly twenty years, AI researchers built programs that could play checkers, prove mathematical theorems, and solve simple word problems. Optimism ran high. Minsky famously predicted in 1970 that "in three to eight years we will have a machine with the general intelligence of an average human being." He was wrong by at least fifty years.

1974 to about 1980 — The first AI winter

It turned out that the techniques of the first golden age did not scale. They worked in tiny "toy" problems and broke down in real-world messiness. Funders lost patience. Money dried up. Researchers left the field. This was the first AI winter — a period when the phrase "artificial intelligence" was associated with broken promises and funding cuts.

1980 to about 1987 — The expert systems boom

A different approach — expert systems, where engineers manually encoded the rules of a particular domain (medical diagnosis, geological survey, factory configuration) — found commercial use. Companies invested. Some of those systems are still in production. But expert systems were brittle: they could not handle situations the engineers had not anticipated, and they could not learn from new data.

1987 to about 2012 — The second AI winter, then the quiet recovery

Expert systems failed to live up to their hype. Another winter. But underneath, two things were happening that would change everything. First, computing power was doubling every two years, on schedule, as it had done since the 1960s. Second, a small group of researchers — notably Geoffrey Hinton, Yann LeCun, and Yoshua Bengio — kept working on a quietly unfashionable technique called neural networks, inspired by the structure of the brain. They were ignored for most of two decades.

2012 — The deep-learning revolution

In 2012, a neural network designed by Hinton's group at the University of Toronto won a major image-recognition competition called ImageNet by a startling margin. It was the moment when neural networks went from "an interesting niche" to "the most important technique in AI". The next ten years saw neural networks transform image recognition, speech recognition, translation, and game-playing. Hinton, LeCun, and Bengio won the Turing Award — the Nobel Prize of computer science — for this work in 2018.

2017 — The transformer moment

In 2017, eight researchers at Google published a short paper titled "Attention is All You Need". It introduced a new neural-network design called the transformer. The paper was technical and at the time did not look revolutionary. But the transformer turned out to be the foundation for almost every major AI system that followed — including the chatbots and image-generators that you use today.

2022 — ChatGPT

On 30 November 2022, OpenAI released ChatGPT — a transformer-based chatbot — to the public. It reached 100 million users in two months, faster than any consumer product in history. For most of the world, this was the moment AI stopped being "something researchers do" and became "something I can use right now in a web browser". Everything since has been, in some sense, the world figuring out what to do with that.

2023 to today — The race

Since ChatGPT, every major technology company has been in a furious race. OpenAI, Anthropic, Google DeepMind, Meta, xAI, and a smaller number of Chinese labs (DeepSeek, Alibaba, Baidu) have released increasingly capable models almost every month. The capabilities have moved from "useful chatbot" to "writes professional-grade software", "passes graduate-level exams", "produces publication-quality video from a sentence", and, increasingly, "carries out multi-step tasks on a computer without supervision". The race is still going. Lesson 4 is about where it might lead.

So why now

One question worth pausing on. Researchers worked on AI for sixty years before the recent explosion. Why did the breakthrough happen in the last decade rather than earlier? Three answers, all true at once.

Compute. Modern AI models require an enormous amount of computation to train. Until the 2010s, the necessary computing power did not exist at any reasonable price. Specialised AI hardware (especially GPUs from a company called Nvidia) made it affordable.

Data. Modern AI models also require enormous amounts of training data. The internet — particularly the era of cheap digital cameras, smartphones, social media, and ubiquitous text on websites — produced more data in the 2010s and 2020s than existed in all of prior human history.

Algorithms. Specifically, the transformer (2017) and the techniques developed around it — pre-training on huge datasets, fine-tuning for specific tasks, reinforcement learning from human feedback — turned out to be the key. None of these would have worked without compute and data; but compute and data alone would not have been enough either.

It was all three together. This is also why the future trajectory (Lesson 4) depends on all three together — and on whether any of them runs out.

Looking ahead

Lesson 2 leaves history behind and looks at the present. We will walk through the eight categories of AI capability that actually work today — language, vision, code, science, healthcare, creative work, robotics, and decision-support — with concrete examples for each, and an honest read on which are mature, which are emerging, and which are oversold.