what sign up
nixm nixm

← learn

A short history of AI: twelve moments that made the field

October 14, 2026

AI is seventy-five years old and has been "about to arrive" for most of them. Alan Turing posed the question in 1950; the field was named at Dartmouth in 1956; early optimism collapsed into two "AI winters"; a chess computer beat the world champion in 1997; deep learning took over vision in 2012 and Go in 2016; the transformer arrived in 2017; and ChatGPT in November 2022 turned a research field into a daily tool. The current chapter is agents and the argument about general intelligence.

1950: the question

Alan Turing publishes "Computing Machinery and Intelligence", asks whether machines can think, and proposes the imitation game, now the Turing test: if a judge cannot tell the machine from a person in conversation, call it thinking. The test is argued about to this day; the question stuck.

1956: the name

A summer workshop at Dartmouth College, organised by John McCarthy, uses the phrase "artificial intelligence" and sets the agenda: reasoning, language, learning. The founders expected major progress within a generation. The field has been alternately embarrassed and vindicated by that expectation ever since.

1960s–1970s: logic, and the first winter

Early systems reasoned with symbols and rules: proving theorems, solving puzzles, holding stilted conversations (ELIZA, 1966, a pattern-matching "therapist" that people confided in anyway). Then the limits showed: the methods did not scale to the messy real world, funding was cut, and the mid-1970s brought the first AI winter.

1980s: expert systems, and the second winter

Companies built "expert systems" that encoded a specialist's rules by hand and sold well for a while. Maintaining thousands of hand-written rules proved impossible, the market collapsed, and the late 1980s brought winter two. The lesson both winters taught: hand-written rules do not scale. The alternative, learning from data, was waiting.

1997: chess

IBM's Deep Blue beats world champion Garry Kasparov. It worked by brute-force search and hand-tuned evaluation, not learning, but it ended the argument about whether machines could beat humans at something that looked like thinking. The goalposts moved to "but it cannot do X".

2000s: the quiet machine-learning decade

With the internet producing data at scale, statistical methods beat rule-based ones at spam filtering, search ranking, recommendations and speech recognition. AI stopped being called AI and started being called "the algorithm". It ran your inbox before it could write your email.

2012: deep learning takes vision

A deep neural network (AlexNet) wins the ImageNet image-recognition competition by a margin that stuns the field, trained on graphics chips. Within three years, deep networks beat every other method at seeing and hearing. The decades-old neural network idea finally had the data and the compute it needed.

2016: Go

DeepMind's AlphaGo beats Lee Sedol, one of the best Go players alive, in a game long thought too intuitive for machines. It learned partly by playing itself. The move 37 in game two, which experts called creative, is the moment many researchers date the current era from.

2017: the transformer

A paper titled "Attention Is All You Need" introduces the transformer, a network design where every word can attend to every other. It scales in a way earlier language models did not. Every major language model since, including the one behind ChatGPT, is a transformer. The LLM page explains what they became.

2020–2022: scale, then the public

GPT-3 in 2020 shows that a transformer trained on enough text can write passable prose and code without being taught either. Image generators arrive in 2022. On 30 November 2022 OpenAI releases ChatGPT, a chat interface on a tuned model, and it becomes the fastest-adopted consumer application to that point. AI stops being a research story and becomes a daily tool, a classroom problem and a boardroom priority within months.

2023–2024: the race, and the rules

Every major company ships a model; models gain vision, voice and tools; the first serious regulation, the EU AI Act, passes in 2024; the training-data lawsuits begin; and the argument about jobs, safety and general intelligence moves from conferences to parliaments. The ethics page covers what that argument is about.

2025 onward: agents, and the open question

Models that act rather than answer, carrying out multi-step tasks with tools, become the focus, along with models that make video with sound and run on local devices. The future page sorts what is likely from what is hype, and the AGI page covers the question the field has carried since 1956: whether general intelligence is years away, decades away, or a different thing from what we are building.

The pattern across seventy-five years

Overpromise, winter, quiet progress, breakthrough, repeat. The breakthroughs came from scale and data, not from cleverer rules, and the last ten years of them came from the same idea, the neural network, that the field had twice abandoned. One modern footnote: the data that made the breakthroughs possible was produced by people, and the question of what AI products do with what you give them is the live one of this chapter. It is also why nixm keeps no transcript and trains on nothing you say.

frequently asked

When was AI invented?

The question was posed by Alan Turing in 1950 and the field was named at the Dartmouth workshop in 1956. The methods behind modern AI, neural networks, date from the same era but only worked at scale from about 2012.

Who is the father of AI?

John McCarthy coined the term and organised the 1956 Dartmouth workshop; Alan Turing posed the founding question in 1950. Marvin Minsky, Geoffrey Hinton, Yann LeCun and Yoshua Bengio are also often named for later eras.

What were the AI winters?

Two periods, the mid-1970s and the late 1980s, when early promises failed, funding collapsed and the field stalled. Both followed approaches based on hand-written rules, which did not scale.

When did ChatGPT come out?

On 30 November 2022. It put a chat interface on a tuned language model and became the fastest-adopted consumer application to that point.

What was the biggest breakthrough in AI history?

Most researchers point to deep learning winning ImageNet in 2012 and the transformer design in 2017. Together they produced the language and image models in use today.

What is the current era of AI?

Generative models that make text, images and video, and agents that carry out multi-step tasks with tools, alongside the first regulation and the debate about general intelligence.

try it now — no account, no transcript kept. even the signal dissolves in seven days.

seed a signal →

Enter your email and receive a link to sign in. No passwords, just a secure signal to connect.

why connect

  • 20 free credits to start
  • your signals, remembered
  • buy more credits anytime
  • image & video generation features
or
+