1950
The imitation game
In Manchester, Alan Turing asks whether a machine could think, and proposes a simple test: could it hold a conversation well enough to pass for human? The question still frames the field today.
The story of AI,1950 to1950
From a question asked in Manchester to machines that act on your behalf. Roll through it one frame at a time and find the moment from a year that matters to you.
Find your year
on the reel.
The reel below is drawn to scale, one sprocket for every year, so the long quiet stretches are as visible as the breakthroughs. Choose a year, or slide to it, and the reel finds the moment the story had reached.
Waiting for a year1950No frame found yetYour moment on the reel.
Choose a year or slide to it and this readout finds the frame the story had reached, then the reel lights it up.
The foundations
1956
A field gets a name
A handful of researchers gather for a summer workshop at Dartmouth under the banner John McCarthy had coined the year before, "artificial intelligence", and set out to make machines that reason, learn and use language. The optimism was boundless.
1958
The perceptron learns
Frank Rosenblatt builds the perceptron, an early model loosely inspired by the neuron that could learn to tell patterns apart. It is the distant ancestor of every neural network since.
1966
ELIZA, an early chatbot
Joseph Weizenbaum writes ELIZA, a program that mimics a therapist by echoing your words back as questions. People confided in it for hours, an early hint at how readily we trust a talking machine.
1970
A machine that understands blocks
Terry Winograd's SHRDLU follows instructions to move blocks around a virtual world and answers questions about them. Impressive in its tiny domain, it also exposed how brittle early language understanding was.
1973
The Lighthill report
Sir James Lighthill tells the British government that AI has failed to deliver on its promises, and funding for the field across the UK is cut back sharply. It is the moment the first AI winter sets in on this side of the Atlantic.
The AI winters
1980
Expert systems go commercial
Businesses pour money into "expert systems" that encode a specialist's know-how as thousands of hand-written rules. For a while, rule-based AI looks like it will run the world.
1986
Backpropagation revived
A landmark paper popularises backpropagation, the algorithm that lets neural networks learn from their mistakes across many layers. The idea would sit quietly until the hardware caught up.
1987
The AI winter bites
The specialist hardware market collapses and the grand promises go unmet, so funding dries up. It is the deepest of the "AI winters", when the field learns humility the hard way.
The deep learning era
1997
Deep Blue beats Kasparov
IBM's Deep Blue defeats world chess champion Garry Kasparov, the first time a computer bests a reigning champion in a full match. It won by brute-force search rather than understanding, but the symbolism landed all the same.
1998
Neural nets read digits
Yann LeCun's convolutional network learns to read handwritten digits well enough to sort real post and cheques. It is a rare, quiet proof that neural networks could do useful work.
2009
ImageNet gives the field a yardstick
Fei-Fei Li and her colleagues release ImageNet, millions of labelled photographs that let anyone measure whether a vision system is genuinely improving. The benchmark quietly sets up everything that follows.
2010
DeepMind founded in London
Demis Hassabis and colleagues start DeepMind with an audacious mission: to solve intelligence, and then use it to solve everything else. British AI research finds a new centre of gravity.
2011
Watson wins Jeopardy!
IBM's Watson beats two champion players at the quiz show Jeopardy!, parsing puns and trivia in real time. Answering natural-language questions on live television felt like a genuine leap.
2012
AlexNet changes everything
A deep neural network called AlexNet wins the ImageNet contest by a stunning margin, trained on graphics cards. It is the spark that lights the deep learning boom we are still living in.
2014
Machines learn to imagine
Ian Goodfellow introduces generative adversarial networks, pitting two neural networks against each other to conjure realistic images from scratch. Machines start to create, not just classify.
2016
AlphaGo topples a legend
DeepMind's AlphaGo beats Go champion Lee Sedol, mastering a game so intuitive it was thought a decade away. Move 37, an alien stroke of strategy, showed the machine had its own kind of insight.
The generative age
2017
Attention is all you need
A Google paper introduces the transformer, a new architecture built on "attention". Unglamorous at the time, it becomes the engine under the bonnet of nearly every modern language model that follows.
2018
Machines pre-trained on language
BERT and the first GPT show that a model trained on a mountain of text can then be pointed at almost any language task. The recipe of pre-train, then adapt, becomes the new standard.
2020
GPT-3 finds its words
A single model with 175 billion parameters writes essays, code and poetry from a short prompt, with no task-specific training. Scale alone, it turned out, bought startling new abilities.
2020
AlphaFold cracks a fifty-year problem
DeepMind's AlphaFold predicts the 3D shape of many proteins with accuracy comparable to experimental methods, a puzzle biologists had wrestled with for half a century. AI delivers a genuine gift to science.
2022
ChatGPT reaches everyone
A free chatbot puts a capable AI in front of the public, and estimates suggest it reaches 100 million monthly active users within two months. Overnight, the whole world is having a conversation about conversation.
2023
AI learns to see and hear
A new wave of multimodal models takes in images, audio and text together, describing a photo or reading a chart as readily as they answer a question. The interface to AI stops being just words.
2024
AI wins the Nobel Prize
Pioneers of neural networks and of AlphaFold are recognised with Nobel Prizes in physics and chemistry. The field the establishment once doubted is now honoured at its very summit.
2025
The rise of the agents
AI stops merely answering and starts doing: planning tasks, using tools and acting on your behalf across software. The question shifts from what AI can say to what it can responsibly be trusted to do.
2026
AI rules start to bite
The EU AI Act enters its broad application and enforcement phase, alongside new transparency obligations for certain AI systems. The story now includes not only what AI can do, but the rules organisations must meet when they deploy it.
Each date marks the milestone, publication, launch, competition or wider shift described; some later frames represent a period rather than one event. Provided by the Institute of AI for interest and learning.
What the timeline
tells us.
Winters come and go
At least two major AI winters followed periods of inflated expectations. Funding fell, but important ideas and methods survived to be picked up in later cycles.
Ideas need the conditions to catch up
Backpropagation was known for decades before neural networks transformed AI. The breakthrough came when better algorithms met larger datasets, cheaper computing and the software to use them.
Scoreboards set the pace
Clear benchmarks make progress visible. Chess, Go, ImageNet and protein-folding challenges gave researchers shared tests that exposed genuine gains, even when no single score captured the whole problem.
Scale bought new abilities
The transformer provided a scalable foundation. Larger and better-curated data, more compute, improved training and new post-training methods turned that foundation into systems with broader capabilities.
Join the Institute of AI
Membership is free. Accreditation is the standard.
More from Get AI
Keep
exploring.
Get AI is a collection of games, experiences, and tools that make AI easier to understand from the Institute of AI.