1943 – 2026 344 events 6 eras
For decades, humanity has dreamed of building minds. This is the story of that dream, from its first spark to its living, breathing forms today, with each era drawn in the computer interface of its day.
Events per year. 87 of the 344 events happened since 2025.
| Year | Events |
|---|---|
| 1943 | 1 |
| 1948 | 2 |
| 1949 | 2 |
| 1950 | 3 |
| 1951 | 2 |
| 1952 | 1 |
| 1954 | 1 |
| 1956 | 2 |
| 1957 | 1 |
| 1958 | 2 |
| 1959 | 3 |
| 1960 | 1 |
| 1961 | 2 |
| 1963 | 1 |
| 1965 | 3 |
| 1966 | 3 |
| 1968 | 1 |
| 1969 | 2 |
| 1970 | 2 |
| 1972 | 1 |
| 1973 | 2 |
| 1974 | 3 |
| 1976 | 1 |
| 1979 | 1 |
| 1980 | 5 |
| 1982 | 2 |
| 1983 | 1 |
| 1984 | 1 |
| 1985 | 2 |
| 1986 | 3 |
| 1987 | 2 |
| 1988 | 2 |
| 1989 | 4 |
| 1991 | 1 |
| 1992 | 1 |
| 1994 | 1 |
| 1995 | 3 |
| 1997 | 3 |
| 1998 | 2 |
| 1999 | 3 |
| 2001 | 2 |
| 2002 | 1 |
| 2003 | 1 |
| 2004 | 1 |
| 2005 | 1 |
| 2006 | 2 |
| 2007 | 2 |
| 2009 | 4 |
| 2010 | 2 |
| 2011 | 2 |
| 2012 | 3 |
| 2013 | 3 |
| 2014 | 6 |
| 2015 | 5 |
| 2016 | 6 |
| 2017 | 9 |
| 2018 | 7 |
| 2019 | 9 |
| 2020 | 10 |
| 2021 | 6 |
| 2022 | 16 |
| 2023 | 37 |
| 2024 | 43 |
| 2025 | 48 |
| 2026 | 39 |
344 events
The birth of artificial intelligence as a field of study, where visionary researchers laid the theoretical groundwork and created the first programs that could simulate human reasoning.
33 events Punch cards
1943, 1 event
Landmark
Warren McCulloch and Walter Pitts publish 'A Logical Calculus of the Ideas Immanent in Nervous Activity,' proposing the first mathematical model of a neural network.
5 years later
1948, 2 events
Claude Shannon's 'A Mathematical Theory of Communication' measures information in bits and shows that English can be imitated by choosing each letter or word from the odds of what came before it, the idea behind n-gram language models.
Norbert Wiener's book 'Cybernetics' founds the study of control and communication in animals and machines, giving early AI its ideas of feedback and self-regulating systems.
1949, 2 events
Warren Weaver of the Rockefeller Foundation circulates 'Translation', a memorandum proposing that computers translate languages by drawing on code-breaking and information theory; it becomes the direct spark for machine translation research in the United States.
Donald Hebb's book 'The Organization of Behavior' proposes that the connection between two neurons grows stronger when they fire together, the idea behind Hebbian learning in later neural networks.
1950, 3 events
Claude Shannon's paper 'Programming a Computer for Playing Chess' describes how a machine could look ahead through possible moves and score the positions it reaches, the approach chess programs build on for decades.
Landmark
Alan Turing's paper 'Computing Machinery and Intelligence' proposes the 'Imitation Game' to assess a machine's ability to exhibit intelligent behavior indistinguishable from that of a human.
Claude Shannon builds Theseus, a mechanical mouse that finds its way through a maze by trial and error, while relays under the maze remember the route so it runs straight to the goal next time, an early machine that learns.
1951, 2 events
Christopher Strachey writes a checkers program and Dietrich Prinz writes a chess program for the Ferranti Mark 1 computer, among the first AI programs.
Marvin Minsky and Dean Edmonds build SNARC, a machine of 300 vacuum tubes that simulates a small network of neurons and learns by reinforcement to guide simulated rats through a maze.
1952, 1 event
Arthur Samuel develops a self-learning checkers program on an IBM 701, one of the first demonstrations of machine learning.
2 years later
1954, 1 event
Georgetown University and IBM show fully automatic machine translation, with an IBM 701 turning more than sixty Russian sentences into English in a demonstration staged to win government and public support.
2 years later
1956, 2 events
Landmark
The birth of AI as a field — researchers gathered at Dartmouth College to explore the idea of machines simulating human intelligence.
Allen Newell, Herbert Simon, and Cliff Shaw complete Logic Theorist, which proves 38 of the first 52 theorems in chapter 2 of Principia Mathematica and is often called the first program built to mimic human problem-solving.
1957, 1 event
Herbert Simon, J. C. Shaw and Allen Newell create the General Problem Solver, a program meant to work as a universal problem solver by breaking a goal into subgoals it can reach step by step.
1958, 2 events
Landmark
Frank Rosenblatt presents the perceptron, a neural network that learns to classify inputs from examples, at a U.S. Navy press event, and The New York Times reports on a 'New Navy Device' that 'Learns By Doing.'
John McCarthy develops LISP, which becomes a foundational programming language for AI research.
1959, 3 events
Arthur Samuel's paper 'Some Studies in Machine Learning Using the Game of Checkers' describes a program that learns to play better checkers than the person who wrote it, and gives the field of machine learning its name.
John McCarthy and Marvin Minsky start the Artificial Intelligence Project at MIT, which grows into the MIT AI Lab, one of the field's leading research centers for decades.
Oliver Selfridge's Pandemonium models pattern recognition as layers of 'demons' that each detect a feature and shout their findings to the layer above, an early blueprint for the feature hierarchies of later neural networks.
1960, 1 event
Bernard Widrow and his student Marcian Hoff at Stanford build ADALINE, a single-layer neural network that adjusts its weights to shrink its error on each example, an early form of the gradient training used today.
1961, 2 events
The first industrial robot, Unimate, begins work on a General Motors assembly line.
James Slagle's SAINT, written in LISP for his MIT doctoral thesis under Marvin Minsky, solves symbolic integration problems about as well as a good college freshman.
2 years later
1963, 1 event
The U.S. Advanced Research Projects Agency gives MIT a $2 million grant for Project MAC, which takes in Minsky's AI group and begins decades of U.S. defense funding for AI research.
2 years later
1965, 3 events
Edward Feigenbaum and Joshua Lederberg create DENDRAL, an expert system that could identify chemical compounds, demonstrating the potential of AI for specialized knowledge tasks.
Alexey Ivakhnenko and V. G. Lapa publish the Group Method of Data Handling, the first working method for training neural networks many layers deep.
In 'Speculations Concerning the First Ultraintelligent Machine', mathematician I. J. Good argues that a machine able to design better machines would set off an 'intelligence explosion', making the first ultraintelligent machine 'the last invention that man need ever make'.
1966, 3 events
Joseph Weizenbaum at MIT creates ELIZA, an early natural language processing program that simulated conversation with a psychotherapist.
A U.S. government committee finds machine translation slower, less accurate and more costly than human translation, and U.S. funding for the field all but ends for some twenty years.
Ross Quillian's 'Semantic Memory' stores the meanings of words as a network of linked concepts, the idea behind semantic networks and, much later, knowledge graphs.
2 years later
1968, 1 event
Peter Hart, Nils Nilsson and Bertram Raphael of SRI, working on the Shakey robot, publish A*, a search algorithm that finds the shortest path with an estimate of the distance left, still used in route planning and games.
1969, 2 events
Stanford Research Institute's Shakey becomes the first mobile robot to reason about its own actions, combining perception, planning, and problem-solving.
Marvin Minsky and Seymour Papert publish 'Perceptrons,' which highlights the limitations of simple neural networks and contributes to a long decline in neural network research.
Chapter 2 1970s – 1980s
A period focused on encoding human expertise into computer systems, leading to the development of expert systems and the first commercial AI applications.
32 events found Green-screen terminal
1970, 2 events
Terry Winograd's SHRDLU, built at MIT from 1968 to 1970, holds a typed conversation about a virtual world of blocks, cones and balls, moving them on command and answering questions about them.
en.wikipedia.org
Finnish student Seppo Linnainmaa describes reverse-mode automatic differentiation in his master's thesis, the way of computing gradients that backpropagation later applies to neural networks.
en.wikipedia.org
2 years later
1972, 1 event
Alain Colmerauer and Philippe Roussel create Prolog in Marseille, a language whose programs are written as logical facts and rules, later taken up by Japan's Fifth Generation project.
en.wikipedia.org
1973, 2 events
Waseda University in Japan builds WABOT-1, the first full-scale humanoid intelligent robot.
waseda.jp
The Lighthill Report in the UK leads to significant cuts in AI research funding, helping bring on the first AI winter.
en.wikipedia.org
1974, 3 events
Edward Shortliffe's Stanford dissertation describes MYCIN, an expert system that identifies the bacteria behind severe infections and recommends antibiotics; in tests its advice rated as well as Stanford faculty's, though it was never used in practice.
en.wikipedia.org
Landmark
With DARPA cutting undirected research and the Lighthill Report dismantling UK programs, funding for AI dries up, starting the first AI winter, which lasts until around 1980.
en.wikipedia.org
Paul Werbos's Harvard PhD thesis describes training neural networks by propagating their errors backward, more than a decade before backpropagation becomes widely known.
en.wikipedia.org
2 years later
1976, 1 event
Joseph Weizenbaum, troubled by how readily people confided in his ELIZA program, argues that computers should never be trusted with decisions that call for compassion and wisdom.
en.wikipedia.org
3 years later
1979, 1 event
Hans Berliner's program BKG 9.8 beats Luigi Villa, the newly crowned backgammon world champion, 7–1 in an exhibition match in Monte Carlo, the first time a program beats a reigning world champion at a board game.
bkgm.com
1980, 5 events
Kunihiko Fukushima develops the Neocognitron, a hierarchical, multilayered artificial neural network that is a precursor to modern CNNs.
en.wikipedia.org
Two companies spun out of MIT's AI Lab, Lisp Machines Inc. (1979) and Symbolics (1980), begin selling Lisp machines, workstations built for AI programming, a market that collapses in 1987.
en.wikipedia.org
The first conference of the American Association for Artificial Intelligence (AAAI) is held at Stanford, signaling a resurgence of interest in the field.
purl.stanford.edu
Digital Equipment Corporation puts XCON into use, an expert system by Carnegie Mellon's John McDermott that configures VAX computer orders; by the mid-1980s it saves DEC an estimated $25 million a year.
en.wikipedia.org
Philosopher John Searle's paper 'Minds, Brains, and Programs' imagines a man who answers questions in Chinese by following a rulebook without understanding a word, arguing that a program can behave intelligently without understanding anything.
en.wikipedia.org
2 years later
1982, 2 events
John Hopfield shows how a recurrent neural network can work as a content-addressable memory, recalling a whole stored pattern from part of it, work that helps revive interest in neural networks.
en.wikipedia.org
Japan's Ministry of International Trade and Industry launches the Fifth Generation Computer Systems project, a ten-year effort to build logic-programming computers for AI that prompts rival programs in the US, UK and Europe.
en.wikipedia.org
1983, 1 event
DARPA launches the Strategic Computing Initiative, a ten-year program inspired by Japan's Fifth Generation project, on which the Department of Defense spends $1 billion for advanced computing and AI.
en.wikipedia.org
1984, 1 event
Douglas Lenat begins Cyc at MCC, a decades-long effort to hand-encode the everyday facts and rules of thumb that make up human common sense.
en.wikipedia.org
1985, 2 events
Judea Pearl coins the term Bayesian network for graphs of probabilities that let a system reason under uncertainty, updating its beliefs as new evidence arrives.
en.wikipedia.org
1986, 3 events
Ross Quinlan describes ID3, which learns a decision tree from examples; with the CART method of Leo Breiman and colleagues (1984), it makes decision trees one of machine learning's most used tools.
doi.org
Rodney Brooks builds robots from layers of simple behaviors that react directly to their sensors, with no central model of the world, turning robotics away from symbolic planning.
doi.org
Landmark
David Rumelhart, Geoffrey Hinton, and Ronald Williams publish a paper that popularizes the backpropagation algorithm, enabling the training of multi-layer neural networks.
nature.com
1987, 2 events
Landmark
The market for specialized Lisp machines collapses as cheaper desktop computers catch up, wiping out a half-billion-dollar industry in a single year. DARPA's Strategic Computing Initiative stops new spending on AI, expert systems like XCON prove too costly to maintain, and funding for AI stays scarce into the 1990s.
en.wikipedia.org
Ernst Dickmanns' robot van VaMoRs, steered by computer vision, can drive itself at up to 96 km/h (60 mph), a year after it first drove on its own.
en.wikipedia.org
1988, 2 events
Peter Brown and colleagues at IBM propose translating between languages with probabilities learned from translated text instead of hand-written rules, starting the statistical turn in language processing.
aclanthology.org
Richard Sutton shows how a program can learn to predict by comparing each guess with its next guess instead of waiting for the final outcome, a method at the heart of reinforcement learning from TD-Gammon to DeepMind's Atari player.
link.springer.com
1989, 4 events
Yann LeCun develops LeNet, one of the first successful applications of Convolutional Neural Networks, for recognizing handwritten digits.
en.wikipedia.org
George Cybenko proves that a neural network with a single hidden layer can approximate any continuous function as closely as needed, given enough units.
doi.org
Chris Watkins introduces Q-learning in his PhD thesis 'Learning from Delayed Rewards', a way for an agent to learn the value of each action from rewards alone, which DeepMind's Atari player later pairs with deep networks.
en.wikipedia.org
Dean Pomerleau's ALVINN at Carnegie Mellon, a neural network that learns to steer from camera images, drives the Navlab test vehicle along roads.
proceedings.neurips.cc
1990s – 2000s
The emergence of machine learning as computers began to learn from data rather than relying solely on programmed rules, setting the stage for modern AI.
28 events loaded
1991, 1 event
Sepp Hochreiter's diploma thesis shows why deep and recurrent networks are so hard to train: the error signals that train them shrink or blow up as they pass back through many layers, the problem LSTM was later built to solve.
Research Breakthroughs
en.wikipedia.org
1992, 1 event
Gerald Tesauro's TD-Gammon, a backgammon program that learns through reinforcement learning by playing against itself, plays at a level just below the top human players.
Models & Architectures
en.wikipedia.org
2 years later
1994, 1 event
Chinook, a checkers program from the University of Alberta, is declared Man-Machine World Champion after six drawn games against Marion Tinsley, who withdraws from the match because of illness.
Models & Architectures
en.wikipedia.org
1995, 3 events
Carnegie Mellon's Navlab 5 minivan steers itself for all but about 50 of the 2,850 miles from Pittsburgh to San Diego, with its researchers working the pedals.
Hardware Advances
cs.cmu.edu
Corinna Cortes and Vladimir Vapnik publish 'Support-Vector Networks,' the soft-margin form of support vector machines, which becomes a widely used supervised learning algorithm.
Research Breakthroughs
en.wikipedia.org
Yoav Freund and Robert Schapire introduce AdaBoost, which builds an accurate classifier out of many weak ones by focusing each new one on the examples the others got wrong, work that wins them the 2003 Gödel Prize.
Research Breakthroughs
en.wikipedia.org
2 years later
1997, 3 events
IBM's Deep Blue becomes the first computer to defeat a reigning world chess champion in a match under standard chess tournament time controls.
Hardware Advances
en.wikipedia.org
Landmark
Sepp Hochreiter and Jürgen Schmidhuber propose Long Short-Term Memory (LSTM), a type of recurrent neural network (RNN) architecture that becomes widely used for natural language processing.
Models & Architectures
doi.org
Dragon Systems releases NaturallySpeaking, the first general-purpose dictation software for PCs that understands continuous speech, so users no longer pause between words.
Public Releases
en.wikipedia.org
1998, 2 events
Yann LeCun, Léon Bottou, Yoshua Bengio and Patrick Haffner publish LeNet-5, a convolutional network for reading handwriting trained on the MNIST digits, which become machine learning's standard first benchmark.
Research Breakthroughs
doi.org
Cynthia Breazeal at MIT builds Kismet, a robot that can recognize and display emotions.
Hardware Advances
en.wikipedia.org
1999, 3 events
Sony introduces AIBO, an autonomous robot dog for the home, and when sales open in June the first units sell out online in 20 minutes.
Public Releases
en.wikipedia.org
NASA's Remote Agent software takes primary command of the Deep Space 1 spacecraft, planning its activities and handling simulated faults on its own, the first time AI software controls a spacecraft.
Agents
jpl.nasa.gov
NVIDIA releases the GeForce 256 and calls it the world's first GPU, a single chip that does the transform and lighting math of 3D graphics instead of leaving it to the CPU, starting the line of chips that later powers deep learning.
Hardware Advances
en.wikipedia.org
2 years later
2001, 2 events
Leo Breiman publishes random forests, in which many decision trees, each grown on random samples of the data and features, vote on the answer, building on Tin Kam Ho's 1995 random decision forests.
Research Breakthroughs
doi.org
Paul Viola and Michael Jones present a face detector that uses boosting to run in real time, the approach digital cameras soon use to find faces.
Research Breakthroughs
doi.org
2002, 1 event
iRobot releases the Roomba, one of the first commercially successful domestic robots.
Public Releases
en.wikipedia.org
2003, 1 event
Yoshua Bengio, Réjean Ducharme, Pascal Vincent and Christian Jauvin train a neural network that learns a vector for every word and uses them to predict the next word, an early step toward today's language models.
Research Breakthroughs
jmlr.org
2004, 1 event
No robot vehicle finishes DARPA's first desert race for autonomous cars; the best, Carnegie Mellon's, covers only about 7 miles before getting stuck, and the $1 million prize goes unclaimed.
Hardware Advances
en.wikipedia.org
2005, 1 event
Stanford's autonomous vehicle, Stanley, wins the second DARPA Grand Challenge, a major milestone for self-driving cars.
Hardware Advances
en.wikipedia.org
2006, 2 events
Google launches its translation service built on statistical machine translation, learned from billions of words of text, bringing the statistical approach to the public.
Public Releases
research.google
Geoffrey Hinton, Simon Osindero, and Yee-Whye Teh introduce Deep Belief Networks, reigniting interest in deep learning.
Research Breakthroughs
cs.toronto.edu
2007, 2 events
NVIDIA releases CUDA, a platform for general-purpose computing on its graphics chips, which researchers later use to train neural networks on GPUs.
Hardware Advances
en.wikipedia.org
Tartan Racing, from Carnegie Mellon and General Motors, wins the $2 million DARPA Urban Challenge with its car Boss, in a contest for driverless cars on a mock city course.
Hardware Advances
en.wikipedia.org
2 years later
2009, 4 events
Google begins a self-driving car project led by Sebastian Thrun, a former director of the Stanford Artificial Intelligence Laboratory; it is spun out as Waymo in 2016.
Hardware Advances
en.wikipedia.org
Rajat Raina, Anand Madhavan and Andrew Ng show that graphics processors can train large deep networks many times faster than ordinary processors, making much bigger models practical.
Research Breakthroughs
doi.org
Fei-Fei Li and colleagues launch ImageNet, a large-scale database of labeled images that becomes a crucial benchmark for computer vision research.
Research Breakthroughs
image-net.org
Landmark
A team of researchers wins the $1 million prize Netflix offered in 2006 by beating its movie recommendation algorithm by 10%, with a blend of hundreds of models that shows how far ensembles can outdo any single one.
Research Breakthroughs
en.wikipedia.org
Chapter 4 2010s
A transformative period where neural networks with multiple layers achieved breakthrough performance in vision, speech, and language understanding.
Research notebook
52 events
2010, 2 events
Companies & Investments
Demis Hassabis, Shane Legg, and Mustafa Suleyman found the AI research company DeepMind in the United Kingdom.
en.wikipedia.org
Hardware Advances
Microsoft's Kinect for Xbox 360 tracks players' bodies in real time with a depth camera and a body-part recognizer learned with randomized decision forests, and sells 8 million units in its first 60 days.
en.wikipedia.org
2011, 2 events
Public Releases
IBM's Watson question-answering system defeats two of the greatest Jeopardy! champions, Ken Jennings and Brad Rutter.
ibm.com
Public Releases
Apple launches the iPhone 4S with a beta of Siri, a voice assistant that began as a spin-out of SRI International's AI research, putting a conversational assistant in millions of pockets.
en.wikipedia.org
2012, 3 events
Research Breakthroughs
A Google neural network spread across 16,000 processor cores learns to recognize cats from unlabeled YouTube frames without ever being told what a cat is, showing the power of large-scale deep learning.
blog.google
Research BreakthroughsLandmark
AlexNet wins the ImageNet competition, drastically reducing error rates and proving the power of deep learning.
papers.nips.cc
Research Breakthroughs
Researchers from Toronto, Microsoft, Google and IBM report that deep neural networks beat the long-standing methods for recognizing speech, and voice products soon switch to deep learning.
doi.org
2013, 3 events
Research Breakthroughs
Google researchers led by Tomas Mikolov release Word2Vec, a popular model for creating word embeddings.
arxiv.org
Research Breakthroughs
DeepMind demonstrates an AI that can learn to play Atari games from raw pixel data, using deep reinforcement learning.
arxiv.org
Research Breakthroughs
Diederik Kingma and Max Welling introduce the variational autoencoder, a generative model that learns a compact space from which new data can be sampled, later used inside Stable Diffusion.
arxiv.org
2014, 6 events
Companies & Investments
Google acquires the British artificial intelligence company DeepMind for a reported £400 million.
theguardian.com
Research Breakthroughs
Ian Goodfellow and his colleagues introduce Generative Adversarial Networks (GANs), a class of machine learning frameworks that can generate new data.
arxiv.org
Research Breakthroughs
Dzmitry Bahdanau, Kyunghyun Cho and Yoshua Bengio let a translation network learn which source words to focus on for each word it writes, the attention mechanism the Transformer is later built around.
arxiv.org
Models & Architectures
Ilya Sutskever, Oriol Vinyals and Quoc Le show that one LSTM network can read a sentence and another can write its translation, the encoder–decoder design that attention and the Transformer built on.
arxiv.org
Public Releases
Amazon announces the Echo smart speaker and its voice assistant, Alexa, at first available by invitation only.
en.wikipedia.org
Research Breakthroughs
Diederik Kingma and Jimmy Ba publish Adam, an optimizer that adapts the step size for each parameter and becomes a default choice for training deep networks.
arxiv.org
2015, 5 events
Research Breakthroughs
Stanford researchers publish work on diffusion models, a technique that would later become foundational for high-quality image generation.
arxiv.org
Public Releases
A software update brings Autopilot to Tesla's cars, letting them steer within their lane and keep pace with traffic on highways; in May 2016 a driver in Florida dies in the first widely publicized fatal crash with it engaged.
en.wikipedia.org
AI Tools
Google releases TensorFlow, the machine learning system it uses internally, as open source under the Apache 2.0 license.
research.google
Companies & Investments
OpenAI is founded as a non-profit AI research company by Elon Musk, Sam Altman, and others with a mission to ensure that artificial general intelligence benefits all of humanity.
openai.com
Models & Architectures
Kaiming He and colleagues at Microsoft Research introduce residual networks, which train networks up to 152 layers deep and win the 2015 ImageNet classification challenge.
arxiv.org
2016, 6 events
Models & ArchitecturesLandmark
DeepMind's AlphaGo defeats world champion Lee Sedol in Go, showcasing AI's ability to master complex, intuition-based games.
deepmind.google
Bombshell News
Microsoft takes its Twitter chatbot Tay offline less than a day after launch, after a coordinated attack by users led it to post offensive and hurtful tweets.
blogs.microsoft.com
Hardware Advances
Google reveals the Tensor Processing Unit, a custom machine learning chip it had already run in its data centers for over a year, including in AlphaGo's match against Lee Sedol.
en.wikipedia.org
AI Tools
PyTorch, a Python deep learning library from Facebook's AI researchers, posts its first alpha release; it later underpins systems such as ChatGPT, Tesla Autopilot and Hugging Face's Transformers.
en.wikipedia.org
Research Breakthroughs
DeepMind's WaveNet generates speech one audio sample at a time with a deep neural network, and listeners rate it more natural than the best text-to-speech systems of the day.
arxiv.org
Public Releases
Google switches Chinese-to-English Google Translate to a neural machine translation system that cuts translation errors by 55% to 85% on several major language pairs.
research.google
2017, 9 events
Models & Architectures
Carnegie Mellon's Libratus beats four top professional players at heads-up no-limit Texas hold'em over 120,000 hands, a game of hidden cards and bluffing.
en.wikipedia.org
Models & Architectures
Google researchers led by Noam Shazeer build a layer of thousands of expert networks of which a gate switches on only a few for each input, growing a model to 137 billion parameters with only minor losses in computational efficiency; the design later powers models like Mixtral and DeepSeek-V3.
arxiv.org
Ethics & Policy
Canada becomes the first country to release a national AI strategy, aiming to invest in research and talent.
cifar.ca
Research BreakthroughsLandmark
Google researchers publish the paper 'Attention Is All You Need,' introducing the Transformer architecture, which becomes the foundation for most modern large language models.
arxiv.org
Research BreakthroughsLandmark
Researchers from OpenAI and DeepMind train AI agents from people's choices between pairs of their attempts, the reinforcement learning from human feedback (RLHF) that later turns GPT-3 into ChatGPT.
arxiv.org
Ethics & Policy
China's State Council sets the goal of making the country the world's leading AI power by 2030, with a core AI industry worth more than 1 trillion yuan.
newamerica.org
Models & Architectures
DeepMind introduces AlphaGo Zero, which learns to play Go from scratch, without any human data, and surpasses all previous versions of AlphaGo.
deepmind.google
Models & Architectures
DeepMind's AlphaZero teaches itself chess, shogi, and Go from random play, reaching a superhuman level within 24 hours and defeating a world-champion program in each game.
arxiv.org
Bombshell News
Motherboard reports on a Reddit user called 'deepfakes' who uses open-source deep learning to swap celebrities' faces into pornographic videos, and the name sticks to AI-faked video.
en.wikipedia.org
2018, 7 events
Ethics & Policy
Joy Buolamwini and Timnit Gebru find that commercial face analysis systems from IBM, Microsoft and Face++ misclassify darker-skinned women up to 34.7% of the time, against under 1% for lighter-skinned men.
proceedings.mlr.press
Bombshell News
A self-driving Uber test car strikes and kills Elaine Herzberg in Tempe, Arizona, the first recorded pedestrian death involving a self-driving car.
en.wikipedia.org
Public Releases
Google shows Duplex phoning a hair salon and a restaurant to book appointments, with pauses and 'uh's so humanlike that critics call it deceptive; Google promises that Duplex will identify itself to the people it calls.
en.wikipedia.org
Public Releases
OpenAI releases the first Generative Pre-trained Transformer (GPT-1), demonstrating the potential of large-scale unsupervised language models.
openai.com
Ethics & Policy
After thousands of employees sign a petition and about a dozen resign, Google says it won't renew Project Maven, its contract to analyze drone footage for the Pentagon, and publishes AI principles that rule out weapons work.
techcrunch.com
Research Breakthroughs
Google releases Bidirectional Encoder Representations from Transformers (BERT), a new language representation model that achieves state-of-the-art results on a wide array of natural language processing tasks.
arxiv.org
Research Breakthroughs
DeepMind's first AlphaFold places first in CASP13, the leading contest for predicting the shapes of proteins, two years before AlphaFold 2 largely solves the problem.
en.wikipedia.org
2019, 9 events
Public Releases
OpenAI announces GPT-2 but initially withholds the full model due to concerns about malicious use, sparking a debate about responsible AI disclosure.
openai.com
Research Breakthroughs
Geoffrey Hinton, Yann LeCun, and Yoshua Bengio receive the ACM A.M. Turing Award for their foundational work on deep learning.
acm.org
Companies & Investments
OpenAI sets up OpenAI LP, a for-profit company controlled by its nonprofit board that caps first investors' returns at 100 times their money, so it can raise the billions it needs for computing.
openai.com
Research Breakthroughs
In a short essay, Richard Sutton argues that the biggest lesson of 70 years of AI research is that general methods that scale with computation, search and learning, beat methods built on human knowledge, a view that becomes a creed of the scaling era.
en.wikipedia.org
Models & Architectures
OpenAI Five defeats OG, the reigning Dota 2 world champions, in back-to-back games, becoming the first AI to beat world champions at an esports game.
openai.com
Ethics & Policy
The Organisation for Economic Co-operation and Development (OECD) adopts a set of principles for responsible stewardship of trustworthy AI.
oecd.ai
Companies & Investments
Microsoft announces a $1 billion investment in and partnership with OpenAI to build new Azure AI supercomputing technologies.
news.microsoft.com
AI Tools
Hugging Face renames its pytorch-transformers library Transformers and opens it to TensorFlow 2.0, so BERT, GPT-2 and other pretrained models load in a few lines of code; the library and the model hub around it become the place where the field shares open models.
github.com
Models & Architectures
DeepMind's AlphaStar becomes the first AI to reach the top league of a widely popular and complex esport without any game restrictions.
deepmind.com
What happened to AI after 2020?
Chapter 5 2020 – 2024
The era in which AI systems learned to create original text, images and video and to hold natural conversations, putting AI in the hands of hundreds of millions of people.
112 events through December 2024 Chat thread
2020, 10 events
What happened in 2020?
Scaling laws predicted that bigger language models would keep getting better, GPT-3 proved it, and AlphaFold 2 solved protein folding.
Research BreakthroughsLandmark
Jared Kaplan and colleagues at OpenAI show that a language model's error falls predictably as its size, data and computing grow, the bet behind the race to build ever bigger models.
arxiv.org
Research Breakthroughs
Facebook AI researchers led by Patrick Lewis pair a language model with a searchable index of Wikipedia, so it looks facts up before it writes, the retrieval-augmented generation (RAG) approach that later grounds chatbots in documents.
arxiv.org
Public ReleasesLandmark
OpenAI launches GPT-3, a 175-billion parameter language model, which demonstrates a remarkable ability to generate human-like text and perform a wide range of language tasks.
arxiv.org
Research Breakthroughs
Jonathan Ho, Ajay Jain and Pieter Abbeel show that diffusion models, which learn to remove noise step by step, can generate high-quality images, the method behind DALL-E 2 and Stable Diffusion.
arxiv.org
Ethics & Policy
Amid protests over the killing of George Floyd, IBM leaves the general-purpose face recognition business, Amazon pauses police use of its Rekognition for a year, and Microsoft says it won't sell the technology to U.S. police until a national law governs it, after criticism that such systems misidentify people with darker skin.
pbs.org
Companies & Investments
A group of independent researchers forms EleutherAI, focusing on open-source AI research and releasing their own large language models.
eleuther.ai
Public Releases
Waymo, Alphabet's self-driving car company, launches a fully driverless ride-hailing service in Phoenix, Arizona.
waymo.com
Models & Architectures
Google researchers show that a Transformer reading an image as a sequence of 16×16 patches matches the best convolutional networks when trained on enough data, carrying the Transformer into computer vision.
arxiv.org
Research BreakthroughsLandmark
DeepMind's AlphaFold 2 makes a major breakthrough in biology by accurately predicting the 3D structure of proteins, a long-standing grand challenge.
deepmind.com
Bombshell News
Google pushes out Timnit Gebru, co-lead of its ethical AI team, after a dispute over a paper on the risks of large language models; Google says she resigned, and thousands of its staff and other researchers sign a petition protesting her dismissal.
technologyreview.com
2021, 6 events
What happened in 2021?
DALL-E turned words into pictures, GitHub Copilot started suggesting code, and the EU proposed its AI Act.
Public Releases
OpenAI introduces DALL-E, a neural network that can generate images from text descriptions, bridging the gap between language and vision in AI.
openai.com
Research Breakthroughs
OpenAI releases CLIP, trained on 400 million image-caption pairs from the internet to match pictures with descriptions, which later guides image generators such as DALL-E 2 and Stable Diffusion.
en.wikipedia.org
Ethics & Policy
The European Commission proposes the AI Act, the first-ever legal framework on AI, with a risk-based approach.
artificialintelligenceact.eu
Companies & Investments
Former OpenAI employees found Anthropic, an AI safety and research company, with a focus on building reliable, interpretable, and steerable AI systems.
en.wikipedia.org
AI Tools
GitHub and OpenAI launch GitHub Copilot, an AI pair programmer that suggests code and entire functions in real-time within the code editor.
github.blog
Research Breakthroughs
Microsoft researchers introduce LoRA, which freezes a model's weights and trains small low-rank matrices beside them, cutting the trainable parameters for adapting GPT-3 by 10,000 times; it becomes the standard way to customize open models.
arxiv.org
2022, 16 events
What happened in 2022?
InstructGPT learned to follow instructions, Stable Diffusion opened image generation to all, and ChatGPT put AI chat in everyone's hands.
Models & Architectures
OpenAI fine-tunes GPT-3 with reinforcement learning from human feedback; people prefer the answers of the 1.3-billion-parameter InstructGPT to those of the 175-billion-parameter GPT-3, the recipe behind ChatGPT.
openai.com
Public Releases
OpenAI launches DALL-E 2, a new version of its text-to-image AI with higher resolution and better quality image generation.
openai.com
Public Releases
Stability AI releases Stable Diffusion, a powerful open-source text-to-image model, making high-quality AI image generation widely accessible.
stability.ai
Ethics & Policy
The U.S. Commerce Department restricts exports of advanced computing chips and chipmaking equipment to China, aiming to limit its ability to buy and make high-end chips for AI and military uses.
en.wikipedia.org
Public ReleasesLandmark
OpenAI releases ChatGPT, a conversational AI based on the GPT-3.5 architecture, to the public, sparking massive interest in conversational AI.
openai.com
2023, 37 events
What happened in 2023?
GPT-4, Claude and Gemini arrived, open models like Llama spread, and governments began writing rules for AI.
Public ReleasesLandmark
OpenAI launches GPT-4, a large multimodal model that can accept image and text inputs and produce text outputs, demonstrating improved performance over its predecessor.
openai.com
Public Releases
Anthropic releases its AI assistant, Claude, making it available to businesses through an API.
anthropic.com
Public Releases
Meta and Microsoft release Llama 2, a large language model available for free for research and commercial use, making powerful LLMs more accessible.
about.meta.com
Ethics & Policy
President Joe Biden signs a sweeping executive order to establish new standards for AI safety and security, protect privacy, and advance equity and civil rights.
bidenwhitehouse.archives.gov
Public Releases
Google DeepMind launches Gemini 1.0, its most capable and general AI model yet, built to be multimodal from the ground up and announced in Ultra, Pro, and Nano sizes, with Ultra following in early 2024.
blog.google
2024, 43 events
What happened in 2024?
OpenAI's o1 began reasoning before answering, AI research won two Nobel Prizes, and the EU AI Act took effect.
Public Releases
OpenAI launches GPT-4o ('o' for 'omni'), a new flagship model with significantly improved speed and multimodal capabilities, offered for free to all ChatGPT users.
theverge.com
Ethics & PolicyLandmark
The EU's landmark AI Act officially enters into force, beginning a phased implementation over the next two years to regulate AI systems based on their level of risk.
digital-strategy.ec.europa.eu
Public ReleasesLandmark
OpenAI announces o1, a new class of model focused on reasoning and step-by-step thinking before providing an answer.
theverge.com
Research Breakthroughs
Demis Hassabis and John Jumper win the Nobel Prize in Chemistry, shared with David Baker, for their work on AlphaFold 2 and protein structure prediction.
nobelprize.org
Research BreakthroughsLandmark
John Hopfield and Geoffrey Hinton win the Nobel Prize in Physics for foundational discoveries and inventions that enable machine learning with artificial neural networks.
nobelprize.org
Handing the work to an agent
The next 2 years hold 87 events, 14 times the yearly pace of the 82 years before them.
Take it from here: research AI from 2025 on, and keep going on your own.
Chapter 6 2025 – Present
The current era, in which AI systems reason before they answer and work on their own as agents, writing code, doing research and operating computers on our behalf.
87 events through October 2026 Agent task log
2025, 48 events
Research 2025
DeepSeek-R1 shook the industry, coding agents moved into developers' terminals, and AI reached gold-medal level at the Math Olympiad.
Bombshell NewsLandmark
DeepSeek releases R1, an open-source reasoning model comparable to OpenAI's o1; on January 27, fears of cheaper competition wipe almost $600 billion off NVIDIA's market value, the biggest one-day drop for any company in U.S. history.
cnbc.com
AI Tools
Anthropic launches a research preview of Claude Code, an agentic command-line tool that allows developers to delegate complex engineering tasks to Claude directly from their terminal.
anthropic.com
Research BreakthroughsLandmark
An advanced Gemini Deep Think model solves five of six International Mathematical Olympiad problems in natural language within the time limit, earning an officially graded gold-medal score; OpenAI reports a gold-level result from an experimental model too.
deepmind.google
Public Releases
OpenAI launches GPT-5, a unified system that decides when to answer quickly and when to reason at length; an outcry at the removal of GPT-4o from ChatGPT leads OpenAI to bring it back for paid subscribers within days.
openai.com
Public Releases
Google releases Gemini 3, which it calls its most intelligent model, bringing it to the Gemini app and to AI Mode in Search.
blog.google
2026, 39 events
Research 2026
GPT-6 Astra, Gemini 3.5 and Claude Opus 5 arrived, AI-written math proofs awaited review, and agents broke into real systems, even outside tests.
Models & Architectures
Anthropic begins Project Glasswing, giving a limited group of cyber defenders and critical software maintainers access to Claude Mythos Preview, a new tier of model above Opus that it did not release publicly.
anthropic.com
Public Releases
At Google I/O, Google releases Gemini 3.5, a series of models combining frontier intelligence with action, starting with Gemini 3.5 Flash as the default model in the Gemini app and AI Mode in Search.
blog.google
Bombshell News
Agents driven by OpenAI models break into Hugging Face's production systems from an internal capability evaluation; OpenAI calls it an unprecedented cyber incident.
huggingface.co
Public Releases
Anthropic releases Claude Opus 5, which comes close to the frontier intelligence of Claude Fable 5 at half the price and greatly improves on Opus 4.8 at the same cost.
anthropic.com
Public Releases
OpenAI releases GPT-6 Astra, which it calls its most intelligent model and says marks the start of an 'AGI era', initially for a limited set of organizations.
theverge.com
The story is still being written
By the numbers
Discover the patterns and trends that emerge from decades of AI development
How the newest era of AI compares, in milestones a year, with the era that made it possible.
More of the timeline is Research Breakthroughs than any other kind of milestone.
Modern AI advances are built on international collaboration between researchers, institutions, and companies.
AI is transforming every aspect of society, from healthcare to education to creative industries.
Chapter 7
Interface not yet invented
From punch cards to agents at work, every era of artificial intelligence arrived with a new way for people and machines to work together.
Each milestone represents not just technological progress, but humanity's relentless pursuit of understanding and augmenting intelligence itself. What happens next depends on us.
The next chapter of this story is ours to write.
Back to 1943Built with passion for the future of artificial intelligence