Skip to content
Multigrid

AI Timeline Interactive

Drag a cursor year across seventy years of AI milestones and see the gaps between them computed.

Milestones on this timeline by 2026
21 of 21

8 of the 21 fall in 2012 or later: 38% of the list packed into 15% of the 73 years it spans.

AI milestones on a linear time axis195019701990201020302026shaded: the longest gap in the list, 9 years
Every event on one axis. Filled marks are at or before the cursor; hollow ones are still ahead of it. The long empty stretches are the periods usually called AI winters — they are gaps in the list, not a label applied to it.
  • 1950Turing asks whether machines can think—
  • 1956The Dartmouth workshop names the field+6y
  • 1958Rosenblatt's perceptron+2y
  • 1966ELIZA+8y
  • 1969Minsky and Papert publish Perceptrons+3y
  • 1973The Lighthill report; UK funding is cut+4y
  • 1980Expert systems reach production use+7y
  • 1986Backpropagation is popularised+6y
  • 1987The LISP machine market collapses+1y
  • 1989Convolutional networks read handwritten digits+2y
  • 1997Deep Blue beats Kasparov+8y
  • 2006Deep belief networks revive deep learning+9y
  • 2009ImageNet is released+3y
  • 2012AlexNet wins ImageNet+3y
  • 2014Generative adversarial networks+2y
  • 2016AlphaGo beats Lee Sedol+2y
  • 2017The transformer paper+1y
  • 2018BERT+1y
  • 2020GPT-3; AlphaFold 2 at CASP14+2y
  • 2022ChatGPT is released+2y
  • 2023GPT-4+1y
Events on the list
21
At or before the cursor
21
Longest gap between two of them
9 years (1997 → 2006)
Within ten years of the cursor
5
Years since the previous one
3
Mean gap across the whole list
3.6 years
What this assumes: the years are the ones commonly cited for each event and one or two could reasonably be argued a year either way — publication, presentation and public release are frequently different dates, and the list uses whichever is normally quoted. More importantly, the list is a selection, and every statistic the tool computes is a statistic about that selection rather than about the field: add ten entries from the 1990s and the “longest gap” changes. That is the point of the input box. Use it as a shape, not as evidence.

A timeline as a list of dates tells you almost nothing. A timeline with the gaps computed tells you two things immediately. The first is that the longest quiet stretches sit exactly where the funding collapses were — the years after the Lighthill report and after the LISP machine market went — which is the honest version of the claim that AI has had winters: not a label somebody applied, but a visible absence of entries anyone thought worth listing.

The second is compression. Nearly two in five of the entries here land in the last dozen years of a span running back to 1950. Some of that is real acceleration and some of it is recency bias in what a list like this chooses to include, and the tool cannot separate them for you. What it can do is let you test the claim: add the events you think are missing from the 1970s and 1990s, and see whether the shape survives your own corrections. If it does not, the shape was about the list.

The other use for the cursor is perspective on the current moment. Park it at 1970 and the field is six years past Dartmouth with real money and a decade of disappointment ahead; park it at 2010 and the transformer is seven years away and nobody in the list knows it. Every position on that slider felt, to the people standing on it, like the present.

The reasoning behind this

Related tools

AI Timeline Interactive · Multigrid