How was AI made?

Three numbersThe shape of the story

Three dates give you the outline before the detail.

1. 1950 — Alan Turing publishes a paper asking whether machines can think, and proposes a test rather than a definition.
2. 1956 — A summer research project at Dartmouth College gives the field its name: artificial intelligence.
3. More than 70 years — of ideas, learning methods, data and computing power piling up to reach today's generative AI.

The timelineIt did not suddenly get clever

Each of these moments mattered for a different reason. Some were ideas; some were engineering; one was mostly about hardware getting fast enough.

Why nowThree things had to arrive together

The ideas behind neural networks are decades old. What changed recently was not mainly the idea.

1. Data — enormous quantities of text and images became available in digital form.
2. Computing power — chips designed for graphics turned out to suit this kind of maths.
3. Method — architectures like the Transformer let all that data and power actually be used.

Often misunderstoodWhat "learning" means here

The everyday word does a lot of hidden work in this topic.

AI was invented by one genius.

It is the accumulated work of very many people across more than seventy years, in mathematics, engineering, linguistics and neuroscience.

AI learns the way a child does.

It adjusts numerical parameters to fit patterns in data. That is genuinely useful and genuinely different from how a child learns.

If it answers confidently, it knows.

Confidence in the output is not evidence about the input. Systems of this kind can produce fluent text that is simply wrong.

Talk about itWhich of these steps was the biggest?

There is no settled answer, which is what makes it a good question. Was the crucial move the idea, the data, or the hardware? Argue for one, then argue for a different one.

SourcesWhere we checked this.

• Computing Machinery and IntelligenceOxford Academic (Mind) — Turing's 1950 paper, used for the imitation game and the framing of the question.
  Document ↗
• A Proposal for the Dartmouth Summer Research Project on Artificial IntelligenceStanford University — The 1955 proposal, used for the origin of the term "artificial intelligence".
  Document ↗
• Deep BlueIBM — Used for the 1997 chess match and how the system worked.
  Document ↗
• ImageNet Classification with Deep Convolutional Neural NetworksNeurIPS — The 2012 result, used for the shift to learned features in image recognition.
  Document ↗
• Attention Is All You NeedarXiv — The 2017 Transformer paper, used for the architecture behind current language models.
  Document ↗