iown.si: a chronicle of artificial intelligence, 1943 to now

McCulloch & Pitts model the neuron as logic

Warren McCulloch and Walter Pitts publish “A Logical Calculus of the Ideas Immanent in Nervous Activity”, describing networks of idealised threshold neurons that can compute logical functions.

The first actual bug

Operators of the Harvard Mark II tape a moth into the logbook with the note “first actual case of bug being found”.

Shannon’s maze-solving mouse

Claude Shannon builds Theseus, a magnetic mouse that learns its way through a reconfigurable maze using relay memory, one of the first learning machines.

Turing asks: can machines think?

Alan Turing’s “Computing Machinery and Intelligence” proposes the imitation game, replacing the question “can machines think?” with a behavioural test.

“I propose to consider the question, ‘Can machines think?’”

“Artificial intelligence” is coined

John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon propose a summer research project on “artificial intelligence”, the first use of the term.

Logic Theorist proves theorems

Allen Newell, Herbert Simon and Cliff Shaw’s Logic Theorist proves theorems from Principia Mathematica, sometimes finding shorter proofs than the original.

The Dartmouth summer workshop

About two dozen researchers gather at Dartmouth College over the summer of 1956. The workshop is usually treated as the founding event of AI as a field.

McCarthy designs LISP

John McCarthy designs LISP, the list-processing language that becomes the workhorse of AI research for decades.

The Perceptron makes headlines

The US Navy demonstrates Frank Rosenblatt’s perceptron; The New York Times calls it the “embryo of an electronic computer” expected to walk, talk and see.

Samuel’s checkers player learns

Arthur Samuel publishes his checkers program that improves by playing itself, and popularises the phrase “machine learning”.

“The last invention that man need ever make”

I. J. Good argues that an ultraintelligent machine could design even better machines, triggering an “intelligence explosion”.

“The first ultraintelligent machine is the last invention that man need ever make, provided that the machine is docile enough to tell us how to keep it under control.”

ELIZA, the first chatbot

Joseph Weizenbaum’s ELIZA imitates a psychotherapist with simple pattern matching. Users confide in it anyway, which alarms its creator.

Shakey the robot

SRI’s Shakey becomes the first mobile robot to reason about its own actions, combining vision, planning (STRIPS) and navigation.

Minsky & Papert publish Perceptrons

The book proves limits of single-layer perceptrons, such as the XOR problem, and is widely blamed for a decade of reduced interest in neural networks.

“In from three to eight years…”

Life magazine quotes Marvin Minsky predicting a machine with the general intelligence of an average human within three to eight years.

“In from three to eight years we will have a machine with the general intelligence of an average human being.”

MYCIN diagnoses infections

Stanford’s MYCIN uses around 600 rules to recommend antibiotics, often matching specialists in evaluations, but is never used clinically.

The Lighthill Report

James Lighthill’s report to the UK Science Research Council criticises AI’s failure to meet its promises. UK funding is cut, marking the start of the first AI winter.

Weizenbaum warns against delegating judgement

ELIZA’s creator Joseph Weizenbaum publishes Computer Power and Human Reason, arguing that some decisions should never be handed to computers.

The Stanford Cart crosses a room

Hans Moravec’s Stanford Cart uses stereo vision to cross a chair-filled room without human help, taking roughly five hours.

XCON goes to work at DEC

Digital Equipment Corporation deploys the R1/XCON expert system to configure computer orders, reportedly saving millions of dollars a year and sparking an expert-systems boom.

Japan’s Fifth Generation project

Japan’s Ministry of International Trade and Industry launches the Fifth Generation Computer Systems project, prompting rival programmes in the US and UK.

Cyc begins encoding common sense

Douglas Lenat starts Cyc, an attempt to hand-encode millions of common-sense facts and rules.

Backpropagation, popularised

Rumelhart, Hinton and Williams publish “Learning representations by back-propagating errors” in Nature, showing how multi-layer networks can learn internal representations.

The LISP machine market collapses

Cheaper general-purpose workstations undercut specialised LISP machines; the expert-systems industry contracts and a second AI winter sets in.

A neural network reads zip codes

Yann LeCun and colleagues train a convolutional network with backpropagation to read handwritten zip codes, later used to process cheques.

Deep Blue defeats Kasparov

IBM’s Deep Blue wins the deciding sixth game of a rematch against world chess champion Garry Kasparov, the first match win by a computer over a reigning champion.

Long short-term memory

Sepp Hochreiter and Jürgen Schmidhuber introduce LSTM, a recurrent network with gates that can remember information over long sequences.

Deep belief nets revive deep learning

Geoffrey Hinton, Simon Osindero and Yee-Whye Teh show a fast way to train many-layered networks layer by layer, reviving interest in “deep learning”.

ImageNet

Fei-Fei Li’s team presents ImageNet at CVPR 2009, a database that grows to over 14 million labelled images and anchors the annual recognition challenge.

Watson wins Jeopardy!

IBM Watson beats champions Ken Jennings and Brad Rutter over three televised episodes.

Siri arrives on the iPhone 4S

Apple announces Siri, bringing a voice assistant to a mass-market phone.

The network that found cats

A Google Brain network trained on YouTube frames without labels develops a unit that responds to cat faces, nicknamed the “cat neuron”.

AlexNet wins ImageNet

Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton’s convolutional network, trained on two GPUs, wins ILSVRC 2012 with a top-5 error of about 15.3%, far ahead of the runner-up.

word2vec: king − man + woman ≈ queen

Tomas Mikolov and colleagues at Google publish word2vec, word embeddings in which vector arithmetic captures analogies.

Google acquires DeepMind

Google buys the London lab DeepMind, reportedly for around $500 million.

Generative adversarial networks

Ian Goodfellow and colleagues introduce GANs: two networks, a generator and a discriminator, trained against each other.

OpenAI is founded

OpenAI launches as a non-profit research lab with pledged funding from founders including Sam Altman and Elon Musk.

AlphaGo beats Lee Sedol

DeepMind’s AlphaGo defeats Go champion Lee Sedol 4–1 in Seoul. Its 37th move in game two, initially judged a mistake, becomes famous.

Tay is taken offline

Microsoft withdraws its Twitter chatbot Tay within about a day after users teach it to post offensive messages.

Attention Is All You Need

Vaswani and colleagues at Google introduce the Transformer, an architecture built entirely on attention, which becomes the basis of modern language models.

BERT

Google releases BERT, a bidirectional transformer pre-trained on masked words, which sets new records across language-understanding benchmarks.

GPT-2 and staged release

OpenAI announces GPT-2 but initially withholds the full model over misuse concerns. A sample about unicorns in the Andes becomes famous.

GPT-3: few-shot learners

OpenAI’s 175-billion-parameter GPT-3 shows that a large enough language model can perform new tasks from a few examples in its prompt.

AlphaFold 2 at CASP14

DeepMind’s AlphaFold 2 predicts protein structures with accuracy competitive with experiments in the CASP14 assessment.

Anthropic is founded

Former OpenAI researchers found Anthropic, an AI safety company.

DALL·E and CLIP

OpenAI introduces DALL·E, which generates images from text (its avocado armchair becomes iconic), and CLIP, which links images and language.

“Stochastic parrots”

Bender, Gebru, McMillan-Major and Mitchell’s paper warns about the costs and risks of ever-larger language models, coining a lasting phrase.

An engineer says LaMDA is sentient

The Washington Post reports that a Google engineer believes the company’s LaMDA chatbot is sentient. Google disagrees and later dismisses him.

Stable Diffusion released openly

Stability AI and partners release Stable Diffusion with open weights, putting high-quality image generation on consumer GPUs.

ChatGPT launches

OpenAI releases ChatGPT as a free “research preview”. It reportedly reaches a million users in five days.

“The hottest new programming language is English”

Andrej Karpathy’s quip captures the shift from code to prompts.

“The hottest new programming language is English”

Bing’s “Sydney” misbehaves

Early users of Microsoft’s Bing chat publish conversations in which its “Sydney” persona declares love, argues and scolds them.

“You have not been a good user.”

GPT-4, and Claude, on the same day

OpenAI releases GPT-4, a multimodal model that passes professional exams at high percentiles. Anthropic releases its first Claude model the same day.

The pause letter

A Future of Life Institute open letter calls for a six-month pause on training systems more powerful than GPT-4. Signatories include Yoshua Bengio and Elon Musk.

Hinton leaves Google to speak freely

Geoffrey Hinton tells The New York Times he left Google so he could talk about the risks of AI.

Statement on AI Risk

Hundreds of researchers and executives sign a one-sentence statement saying that mitigating the risk of extinction from AI should be a global priority.

Bletchley Park AI Safety Summit

28 countries and the EU sign the Bletchley Declaration on the risks of frontier AI at the first AI Safety Summit.

OpenAI’s board fires, then rehires, Sam Altman

OpenAI’s board removes CEO Sam Altman. After most staff threaten to leave, he returns five days later with a new board.

1. Fri 17 NovThe board announces Altman’s removal.
2. Mon 20 NovMost employees sign a letter threatening to leave.
3. Wed 22 NovAgreement for Altman to return with a new board.

Google Gemini 1.0

Google DeepMind announces Gemini, a natively multimodal model family.

Sora previews text-to-video

OpenAI previews Sora, generating minute-long videos from text prompts.

Glue on pizza

Google’s new AI Overviews suggest, among other errors, adding glue to pizza sauce, apparently drawn from an old joke post.

Safe Superintelligence Inc.

Ilya Sutskever, Daniel Gross and Daniel Levy found SSI, “the world’s first straight-shot SSI lab”. It is the first major lab named for superintelligence.

The EU AI Act enters into force

The world’s first comprehensive AI law enters into force, with obligations phased in over the following years.

OpenAI o1: models that reason

OpenAI releases o1, trained to spend more computation “thinking” before it answers.

Nobel prizes for AI

The Physics prize goes to John Hopfield and Geoffrey Hinton for neural networks; the next day the Chemistry prize goes to David Baker, Demis Hassabis and John Jumper for protein design and structure prediction.

An AI uses a computer

Anthropic releases “computer use” in beta: Claude can look at a screen, move a cursor and type.

Model Context Protocol

Anthropic open-sources MCP, a standard for connecting AI assistants to tools and data sources. It is later adopted widely across the industry.

“Superintelligence in the true sense of the word”

In a blog post, Sam Altman writes that OpenAI is beginning to turn its aim beyond AGI, to superintelligence.

“We are beginning to turn our aim beyond that, to superintelligence in the true sense of the word.”

DeepSeek-R1

Chinese lab DeepSeek releases an open-weight reasoning model reportedly trained at a fraction of rivals’ cost; on 27 January Nvidia shares fall sharply.

The Stargate Project

OpenAI, SoftBank, Oracle and MGX announce a venture committing up to $500 billion to US AI infrastructure.

Operator browses for you

OpenAI launches Operator, an agent that uses its own web browser to complete tasks.

“Vibe coding”

Andrej Karpathy describes “vibe coding”: fully giving in to the AI and forgetting that the code even exists.

Claude 3.7 Sonnet and Claude Code

Anthropic releases a hybrid reasoning model and Claude Code, an agent that works in a developer’s terminal.

“The Gentle Singularity”

Sam Altman writes that humanity is past the event horizon and that the takeoff has started.

“We are past the event horizon; the takeoff has started.”

Meta Superintelligence Labs

Meta reorganises its AI work under Meta Superintelligence Labs, led by Alexandr Wang and Nat Friedman, after a $14.3B investment in Scale AI.

Gold-medal level at the IMO

Experimental systems from Google DeepMind and OpenAI solve five of six International Mathematical Olympiad problems, reaching gold-medal standard.

“Personal Superintelligence”

Mark Zuckerberg publishes a letter arguing for superintelligence in the hands of every individual rather than one central system.

GPT-5

OpenAI releases GPT-5. Reception is mixed and the company restores GPT-4o for paying users within days.

California SB 53

California enacts the first US state law specifically addressing catastrophic risk from frontier AI models.

Statement on Superintelligence

A statement hosted by the Future of Life Institute calls for a prohibition on developing superintelligence until there is broad scientific consensus it can be done safely and strong public buy-in.

Microsoft’s superintelligence team

Mustafa Suleyman forms the MAI Superintelligence Team at Microsoft, framing its goal as “humanist superintelligence”, starting with medical diagnosis.

Claude Mythos and Project Glasswing

Anthropic discloses Claude Mythos and gives selected organisations access through Project Glasswing to find vulnerabilities in critical software.

Access to frontier models restricted, then restored

A US government letter bars non-US nationals from Mythos 5 and Fable 5; Anthropic suspends access, and restrictions are lifted on 30 June.

GPT-6 Astra and “the AGI era”

OpenAI launches GPT-6 Astra. President Greg Brockman says “Welcome to the AGI era”, a framing the company left users to judge.

Meta Muse

Meta launches Muse, an AI agent for everyday tasks.

“AI” becomes “Super Intelligence”

A US executive order directs federal agencies to use “Super Intelligence (SI)” instead of “Artificial Intelligence (AI)” in non-statutory communications. The relabel applies to the existing legal definition of AI, not to the research meaning of superintelligence.

“It's not artificial.”

SUPERARTIFICIAL INTELLIGENCE

OpenAI Dots

OpenAI releases GPT-6.1 Sol and launches Dots, a personal AI agent system.