How we got here

1950: the question

The mathematician Alan Turing asked, in a 1950 paper, whether a machine could ever hold a conversation well enough that you could not tell it from a person. He proposed a test: talk to something through a screen, and if you cannot tell whether it is human, that is good enough. People argued about it for decades. Today you can run the test on your phone.

1956: the name

A summer workshop at Dartmouth College gathered a handful of researchers who believed thinking could be described precisely enough for a machine to do it. The proposal used the phrase "artificial intelligence" and the name stuck. They thought it would take a summer. It took seventy years.

1960s to 1990s: the long road

The early decades produced ideas, then disappointment, then more ideas. A program called ELIZA, in 1966, imitated a therapist by reflecting your words back at you, and people found it unsettlingly convincing. Then funding dried up when promises were not kept. Researchers call those stretches "AI winters". There were two.

Along the way the field split into two camps. One tried to write down the rules of thinking by hand. The other, called machine learning, tried to let the machine find patterns in examples instead. For a long time neither camp could do much that mattered outside a lab. In 1997 a chess computer called Deep Blue beat the world champion, which made headlines but was built for exactly one game.

2012: the machine learns to see

The turning point came from the second camp. Researchers had been working on "neural networks", programs loosely inspired by how brain cells connect, which learn by adjusting millions of tiny internal numbers until their output matches the examples. In 2012 one of these networks won an image recognition contest by a margin nobody expected. It could tell a cat from a dog from a photo, learned entirely from examples.

Two things made it possible: far more examples, thanks to the internet, and far more computing power, thanks to the graphics chips built for video games. Those same two ingredients power everything that came after.

2017: the transformer

In 2017 a paper from Google introduced a design called the transformer. The technical details do not matter here. What matters is that it let a network pay attention to every word in a passage at once, and it scaled: make it bigger, feed it more text, and it kept getting better with no sign of stopping.

That word, "scaled", is why the last few years happened.

2018 to 2022: bigger and bigger

A young company called OpenAI took the transformer and trained it on more and more of the internet. Each version was much larger than the last. GPT-2 in 2019 could write a plausible paragraph. GPT-3 in 2020 could write a plausible essay, a poem, working code. It was available to developers, and mostly developers noticed.

November 2022: everyone notices

On 30 November 2022 OpenAI put a chat window in front of one of these models and called it ChatGPT. It reached a hundred million users in two months, faster than any product before it. For the first time a person with no technical skill could sit down and talk to the thing, and the thing talked back, about anything.

That date is the line between "AI research" and "AI in everyone's life".

2023 to 2024: the field fills in

Within months there was competition. Anthropic released Claude. Google released Gemini. Meta released models with open weights that anyone could download and run. Microsoft put the technology into Office and Windows. Models learned to look at images, then to hear and speak. The context window, the amount a model can hold in mind at once, grew from a few pages to whole books.

2025 to today: from talking to doing

The most recent shift is the one this site exists for. Chatbots answered questions. Now they take actions: read your inbox, check your calendar, search the web, place an order you approved, and run on a schedule while you sleep. The industry calls these agents. Late in 2024 a standard called MCP appeared that lets any model plug into any tool the same way, and by 2026 every major AI company supports it. That is why one set of instructions can run in Claude, ChatGPT, Grok or Cursor with the same connectors.

Models kept improving too. In 2026 alone, every major company shipped a new generation, and models that "think" through a problem step by step before answering became normal.

What to remember

• The idea is seventy years old. The thing on your phone is about four.
• Two ingredients made it possible: the internet as a textbook, and gaming chips as a brain.
• November 2022 is when it stopped being research.
• The current chapter is about agents: AI that acts, not just answers. That is the chapter you are in.