Skip to content
Multigrid

The history of AI

What the founding documents of artificial intelligence actually said, separated from the folklore that grew up around them.

The history of artificial intelligence is unusually badly served on the web, and for a specific reason. It is evergreen, it is heavily searched, and it is cheap to write from other people’s summaries — so the same dates, the same quotations and the same causal stories have been copied across thousands of pages, acquiring errors as they go. Turing gets credited with a pass mark he never set. Minsky and Papert get blamed for a claim their book does not make. The Dartmouth proposal gets described as promising human-level machines within a decade, which is not what it says.

These twenty pages go back to the proposal, the paper, the report or the press release, and separate what the document says from what everyone now believes it says. Where a date is certain it is given. Where only a decade or an ordering is certain, the page says so rather than inventing a year, and where a famous quotation cannot be verified it is paraphrased and labelled as a paraphrase. That makes for slightly duller prose than the version with an invented Turing epigram, and it makes for a page you can cite.

A companion page, the single-page timeline from Dartmouth to now, runs the same material chronologically if that is what you came for. This cluster goes one level deeper on each episode.

The Dartmouth Workshop and the Naming of AI

What the 1955 Dartmouth proposal actually promised, who wrote it, who turned up, and why calling it a conference is misleading.

9 min read

The Turing Test: What Turing Wrote, and What People Think He Wrote

Turing's 1950 paper proposed an imitation game and made one dated prediction. Neither is the pass/fail benchmark the phrase now describes.

10 min read

The Perceptron, and the Book That Stopped It

Rosenblatt's 1958 perceptron, the Mark I machine, and what Minsky and Papert's 1969 book actually proved — which is narrower than its reputation.

10 min read

ELIZA and the First Chatbot Illusion

How Weizenbaum's 1966 program worked, why the DOCTOR script is not the same thing as ELIZA, and why its author spent the next decade arguing against it.

9 min read

The AI Winters: What Froze and Why

Two funding collapses, traced to the specific reports and specific broken promises that caused them, and what carried on working through both.

11 min read

Expert Systems: The Boom Everyone Forgot

Why rule-based systems became a real industry in the mid-1980s, what specifically broke them, and where the technology quietly survives.

10 min read

Backpropagation's Long Road to Acceptance

Six independent discoveries between 1960 and 1986, in order, and why the 1986 paper is the famous one despite not being first.

10 min read

Deep Blue and What Beating Kasparov Proved

The 1996 and 1997 matches, the machine's actual architecture, and why the lesson most people drew from it was the wrong way round.

10 min read

ImageNet and the 2012 Moment

AlexNet's 2012 result, the error rates it beat, and why the dataset built between 2007 and 2009 was the harder half of the achievement.

10 min read

AlphaGo, Move 37, and What Self-Play Changed

The 2016 Lee Sedol match, the one-in-ten-thousand move, and the four properties of Go that made the result possible and kept it from generalising.

10 min read

The Long History of Machine Translation

Seventy-five years from Weaver's 1949 memorandum to neural systems, with each era's headline result stated inside its own evaluation.

11 min read

word2vec and the Idea That Meaning Has Coordinates

The 2013 papers, the king-minus-man example, and the evaluation detail that makes the famous analogy result weaker than it looks.

9 min read

How the Transformer Paper Came About

Attention Is All You Need was a translation-throughput paper. What it proposed, what it reported, and how much of a modern transformer is not in it.

10 min read

GPT-1 to GPT-3: What Each One Added

Three papers between 2018 and 2020, and the one specific idea each contributed — read from the papers rather than from the announcements.

10 min read

The Launch That Was Meant to Be a Research Preview

ChatGPT launched on 30 November 2022 on a model that was already public. What was new was the interface — and the growth figures need their sources.

9 min read

The Open-Weights Wave and How It Started

Open weights did not begin with the LLaMA leak in March 2023. The chronology from BERT to DeepSeek-R1, with each licence named.

10 min read

Symbolic AI and Connectionism: A Fifty-Year Argument

The intellectual history of the field's central disagreement, from a shared 1943 ancestor to the hybrid systems both sides now build.

11 min read

The Chatbots Before ChatGPT

PARRY, A.L.I.C.E., SmarterChild, Tay and the rest — what each one was built from and the specific thing each failed at.

10 min read

Ten AI Predictions That Were Confidently Wrong

Eleven dated predictions with their authors, what actually happened, and the two failure modes that account for nearly all of them.

11 min read

The Hardware Accident That Made Deep Learning Possible

How a chip designed to draw triangles for video games became the substrate for machine learning, with the arithmetic that explains why it fits.

11 min read

Other topics