A knowledge base about artificial intelligence where every fact links back to its original source - the paper, the announcement, the talk. No articles about articles.
The events that shaped AI, in order.
Ada Lovelace's translation of Menabrea's memoir on the Analytical Engine added Notes that included the first published algorithm and the warning that the machine could only do what it was ordered to do.
George Boole's 1854 book recast logic as algebra, giving the world the symbolic system that digital hardware would later implement directly.
Alan Turing's 1936 paper defined the abstract machine that bears his name and proved that no general procedure can decide every mathematical question.
Isaac Asimov's 1942 story Runaround introduced the Three Laws of Robotics, the first widely known attempt to state safety rules for autonomous machines.
McCulloch and Pitts showed that simple all-or-none neurons wired into networks can compute any logical proposition.
Vannevar Bush's 1945 essay imagined the memex, a personal machine that would store and cross-link a person's knowledge, anticipating hypertext and machine-augmented thought.
Norbert Wiener's 1948 book Cybernetics unified feedback and control across machines and living things, supplying a vocabulary that shaped early thinking about intelligent systems.
Claude Shannon's 1948 paper founded information theory, defining the bit and showing how to measure, compress, and reliably transmit information.
Donald Hebb proposed that learning happens when connections between co-active neurons strengthen, giving neural networks a rule for learning.
Alan Turing proposed the imitation game, now the Turing Test, replacing 'can machines think' with a concrete behavioral test.
The 1955 Dartmouth proposal by McCarthy, Minsky, Rochester, and Shannon coined 'artificial intelligence' and launched the field.
Newell, Simon, and Shaw built the Logic Theorist at RAND, a program that proved theorems in symbolic logic and is widely called the first AI program.
Frank Rosenblatt's perceptron was an early trainable neural network that adjusted its own connections to classify patterns.
Arthur Samuel's self-improving checkers program popularized the term machine learning and showed computers could learn from experience.
George Devol's patented programmable manipulator, built by Unimation, became the first industrial robot on a factory floor at General Motors.
Mathematician I.J. Good's 1965 paper introduced the idea of an ultraintelligent machine that designs ever-better machines, producing an intelligence explosion.
Gordon Moore's 1965 magazine article observed that the number of components on an integrated circuit was doubling roughly every year, the forecast that became Moore's Law and the metronome of cheap computing.
The 1966 ALPAC report concluded there was no immediate or predictable prospect of useful machine translation and recommended cutting funding, collapsing US support for the field for roughly two decades - before the term 'AI winter' existed.
Joseph Weizenbaum's ELIZA was an early conversational program that mimicked a psychotherapist using simple pattern matching.
SRI's Shakey combined perception, planning, and movement in one machine, and the project produced lasting AI techniques including the A* search algorithm and STRIPS planning.
Minsky and Papert's book exposed the limits of single-layer perceptrons, helping freeze neural network funding for years.
Terry Winograd's SHRDLU let a person hold a typed English conversation with a computer that moved blocks in a simulated world, a landmark in early natural language understanding.
Sir James Lighthill's survey for the UK Science Research Council judged AI a disappointment, cutting British AI funding for a decade.
Paul Werbos's 1974 Harvard dissertation set out the method later known as backpropagation for training multilayer networks.
MYCIN was a Stanford expert system that advised doctors on diagnosing and treating blood infections using a few hundred rules and a way of handling uncertainty.
Kunihiko Fukushima's Neocognitron introduced a layered, shift-tolerant network for pattern recognition that prefigured modern CNNs.
R1, later called XCON, was a rule-based expert system that configured Digital Equipment Corporation's VAX computer orders and became one of the first expert systems to pay off in business.
John Hopfield's PNAS paper showed how networks of simple units can store memories as stable collective states, reenergizing the field.
Japan's MITI launched a ten-year national project, run by the ICOT institute, to build knowledge-processing computers based on logic programming and parallel inference.
Douglas Lenat began Cyc in 1984, an ambitious decades-long effort to hand-encode the everyday common-sense knowledge that AI systems lacked.
The 1986 Nature paper made backpropagation the standard way to train multilayer neural networks with hidden units.
In the late 1980s the commercial AI boom collapsed as the specialized Lisp-machine market crashed and expert systems failed to live up to their hype, ushering in a long downturn.
LeCun and colleagues at Bell Labs trained a convolutional network on real US Postal Service digits, a landmark for practical deep learning.
An IBM team publishes a statistical approach to machine translation, treating language as a problem of probability and data rather than hand-written grammar rules.
Gerald Tesauro's neural network learned to play backgammon at near-champion level by playing millions of games against itself using temporal-difference learning.
Cortes and Vapnik publish the support-vector network, a learning method that became the dominant classifier of the 1990s and 2000s and made neural networks look obsolete.
IBM's Deep Blue became the first computer to beat a reigning world chess champion in a full match, defeating Garry Kasparov 3.5 to 2.5.
Hochreiter and Schmidhuber publish LSTM, a recurrent network that remembers across long gaps, solving the vanishing-gradient problem.
LeCun and colleagues publish the LeNet-5 convolutional network, showing end-to-end gradient learning for reading handwritten documents.
iRobot's Roomba put a small, self-navigating robot in millions of households, the first truly mass-market consumer robot.
Stanford's self-driving car Stanley completed a 132-mile desert course to win the 2005 DARPA Grand Challenge, a turning point for autonomous vehicles.
Hinton, Osindero, and Teh show a fast greedy way to train deep networks layer by layer, reviving interest in deep learning.
Netflix launches a $1 million open contest to improve its movie recommendation system by 10 percent, mainstreaming the public machine learning competition.
NVIDIA introduced CUDA, a platform that let developers run general-purpose code on GPUs, providing the compute engine that later powered deep learning.
After 18 years of computation, Jonathan Schaeffer's team proved that checkers is a draw with perfect play, the most complex game solved at that time.
Three Google researchers argue that simple models trained on enormous amounts of data beat elaborate models trained on less, a data-first creed that predates the scaling laws.
Fei-Fei Li and colleagues introduce ImageNet, a labeled image database that became the proving ground for modern computer vision.
Kaggle turned machine learning into a competitive discipline, hosting public prediction contests that became both a benchmark and a talent pipeline for the field.
Google publicly disclosed that it had been quietly building self-driving cars that had already logged 140,000 miles on real roads.
IBM's Watson question-answering system beat champions Ken Jennings and Brad Rutter on Jeopardy!, showing computers could understand and answer natural-language questions.
Apple launches Siri with the iPhone 4S, putting a conversational voice assistant into the hands of millions of consumers.
A deep convolutional neural network crushed the ImageNet contest, proving deep learning could outperform hand-built computer vision.
Mikolov and colleagues introduced word2vec, an efficient way to learn word vectors that capture meaning through geometry.
DARPA and Boston Dynamics unveiled Atlas, a six-foot hydraulic humanoid robot built as a common platform for the DARPA Robotics Challenge.
DeepMind trained a single neural network to play Atari video games directly from raw pixels, beating humans on several titles.
Google buys the London AI lab DeepMind in January 2014, placing a leading research lab inside a big-tech company.
Ian Goodfellow and colleagues proposed pitting two neural networks against each other, launching the era of AI-generated images.
Nick Bostrom's 2014 book Superintelligence brought existential risk from advanced AI to a wide audience and shaped how lab founders talk about safety.
Sutskever, Vinyals, and Le showed neural networks could map input sequences to output sequences, enabling end-to-end translation.
Amazon introduces the Echo smart speaker and its voice assistant Alexa, creating the always-listening ambient assistant category for the home.
Google released TensorFlow, its internal machine-learning system, as free open-source software, putting an industrial-grade deep learning framework in everyone's hands.
Residual connections let neural networks grow to 152 layers and win ImageNet 2015, unlocking much deeper architectures.
OpenAI launched in December 2015 as a nonprofit AI research company with a $1 billion commitment and a mission to ensure artificial general intelligence benefits all of humanity.
DeepMind's AlphaGo beat Go world champion Lee Sedol 4-1, a feat experts had thought was a decade away.
Google disclosed the Tensor Processing Unit, a custom chip built specifically for neural networks that had already been running in its data centers for over a year, marking the start of the custom-silicon era for AI.
Facebook AI Research released PyTorch, a Python-first, define-by-run deep learning framework that became the dominant tool for AI research.
Carnegie Mellon's Libratus defeated four top professionals at heads-up no-limit poker, the first AI to master a major game of imperfect information.
Google researchers publish 'Attention Is All You Need', introducing the Transformer architecture that underpins modern AI.
China's State Council issues a national strategy setting a three-step goal to make the country the world's primary AI innovation center by 2030.
DeepMind's AlphaZero taught itself chess, shogi and Go from scratch by self-play, reaching superhuman strength in each and defeating the best existing programs.
OpenAI's 2018 analysis found that the compute used in the largest AI training runs had been doubling every 3.4 months since 2012, making the empirical case that raw computation drives AI capability.
OpenAI's first GPT showed that pre-training a Transformer on large amounts of unlabeled text and then fine-tuning it set a new bar across many language tasks.
Google researchers release BERT, a pre-trained bidirectional Transformer that set new state-of-the-art on eleven NLP benchmarks.
OpenAI's 1.5-billion-parameter GPT-2, trained only to predict text, performed many language tasks with no task-specific training.
OpenAI Five became the first AI system to defeat the reigning world champions at an esport, beating Dota 2 champions OG in April 2019.
Pluribus, built by Carnegie Mellon and Facebook AI, beat elite professionals at six-player no-limit poker, the first superhuman result in a multiplayer imperfect-information game.
In 2019 OpenAI restructures as a capped-profit company and Microsoft invests 1 billion dollars, creating the financial structure behind much of the AI boom that followed.
Hugging Face's open-source Transformers library put state-of-the-art language models one pip install away, becoming the distribution layer of the LLM era.
DeepMind's AlphaStar reached Grandmaster level in StarCraft II, ranking above 99.8 percent of human players across all three races.
OpenAI researchers showed language model performance follows smooth power laws in model size, data, and compute.
OpenAI's 175-billion-parameter GPT-3 performed many tasks from a few text examples alone, with no fine-tuning or weight updates.
The foundational scientific-computing library underpinning the Python data and AI stack was formally documented in a Nature paper, 'Array Programming with NumPy.'
At the CASP14 contest, DeepMind's AlphaFold predicted protein 3D structures with accuracy rivaling slow, costly lab experiments.
EleutherAI released The Pile, an 825 GiB English text dataset built from 22 documented sources, making the contents of a large language model training corpus openly inspectable.
OpenAI unveils the original DALL-E, a GPT-3-scale model that generates images from natural-language text descriptions - the start of the generative image era.
CLIP learns visual concepts directly from natural-language supervision, becoming the bridge that made modern multimodal AI possible.
A 2021 FAccT paper by Bender, Gebru, McMillan-Major, and Mitchell that became the most-cited critique of the large-language-model paradigm.
Anthropic is founded by siblings Dario and Daniela Amodei and colleagues as an AI safety and research company, surfacing publicly with a 124 million dollar Series A.
GitHub launches Copilot, an AI coding assistant built on OpenAI Codex that suggests code as developers type.
DeepMind and EMBL-EBI launched a free, open database of AlphaFold protein structure predictions, starting with the human proteome and later expanding to over 200 million structures.
OpenAI's InstructGPT used reinforcement learning from human feedback to make GPT-3 follow instructions, the direct precursor of ChatGPT.
DeepMind's Chinchilla paper showed that most large language models were undertrained and that model size and training data should grow together for compute-optimal results.
OpenAI unveils DALL-E 2, a system that generates and edits high-resolution images from natural-language text prompts.
Midjourney, an independent research lab's text-to-image generator run through Discord, opened to the public in July 2022 and became one of the dominant image AI tools.
Stability AI publicly releases Stable Diffusion, an open text-to-image model with downloadable weights and code.
OpenAI releases Whisper, an open multilingual speech recognition model trained on large-scale weakly supervised audio.
OpenAI releases ChatGPT, a conversational AI that reaches mass adoption and ignites the modern generative-AI boom.
Meta releases LLaMA, a family of foundational large language models, accelerating the open-weights movement.
Anthropic launches Claude, a next-generation AI assistant built around helpful, honest, and harmless training.
OpenAI releases GPT-4, a multimodal model scoring in the top tiers of professional exams and marking a major capability jump.
A March 2023 open letter from the Future of Life Institute called on all AI labs to pause training systems more powerful than GPT-4 for at least six months.
A one-sentence May 2023 statement from the Center for AI Safety put extinction risk from AI alongside pandemics and nuclear war, signed by leading researchers and lab CEOs.
President Biden signs Executive Order 14110, the most sweeping US federal action on AI of the period, which was later revoked in January 2025.
At the first global AI Safety Summit in November 2023, 28 countries and the European Union signed the Bletchley Declaration on managing frontier AI risk.
DeepMind's GraphCast, a machine-learning weather model, outperformed the leading conventional forecasting system on most metrics while running in under a minute.
DeepMind's GNoME deep-learning tool predicted 2.2 million new crystal structures, including about 380,000 stable candidates worth synthesizing in the lab.
Google launches Gemini, a natively multimodal model family built jointly by Google DeepMind, in three sizes - Ultra, Pro, and Nano.
The New York Times filed a copyright lawsuit against OpenAI and Microsoft in federal court over the use of its articles in training AI systems, putting the training-data question before the courts.
DeepMind's AlphaGeometry solved 25 of 30 olympiad geometry problems, near the gold-medalist average, and its successors later reached silver-medal level across a full IMO.
OpenAI unveils Sora, a model that generates video from text prompts, extending generative AI into moving images.
Anthropic launches Claude 3 in three models - Opus, Sonnet, and Haiku - which it says set new industry benchmarks across cognitive tasks and add vision capabilities.
AlphaFold 3, from DeepMind and Isomorphic Labs, predicts the joint structure of proteins together with DNA, RNA, small-molecule ligands and ions in a single model.
OpenAI releases GPT-4o, an 'omni' model trained end-to-end across text, vision, and audio that can hold real-time spoken conversations.
The European Union adopts Regulation (EU) 2024/1689, the world's first comprehensive law governing artificial intelligence through a risk-based framework.
OpenAI introduces o1, a model trained to reason step by step before answering, marking the reasoning-model era.
John Hopfield and Geoffrey Hinton win the 2024 Nobel Prize in Physics for foundational work enabling machine learning with neural networks.
The 2024 Nobel Prize in Chemistry went to David Baker for computational protein design and to Demis Hassabis and John Jumper for AlphaFold's protein structure prediction.
Anthropic releases the Model Context Protocol, an open standard for connecting AI applications to external data and tools.
OpenAI, SoftBank, Oracle, and MGX announce Stargate, a new company intending to invest 500 billion dollars over four years in AI infrastructure in the United States.
DeepSeek releases DeepSeek-R1, an open reasoning model trained largely via reinforcement learning, intensifying global LLM competition.
Anthropic introduces Claude Code, its first agentic coding tool, alongside Claude 3.7 Sonnet as a limited research preview.
Anthropic releases Claude Opus 4 and Sonnet 4, models built to sustain focused work over thousands of steps, marking the shift to long-running agentic coding.
Google DeepMind's Gemini with Deep Think earns an officially certified gold-medal score at the 2025 IMO, with OpenAI reporting the same score separately.
Obligations for providers of general-purpose AI models under the EU AI Act start to apply, the first binding rulebook aimed at the makers of foundation models.
OpenAI releases GPT-5 as a unified system that automatically routes between fast answers and deeper reasoning, becoming the default model for ChatGPT users.
Google releases Gemini 3, its most capable model, launched across all its products on day one and topping major reasoning and multimodal benchmarks.
The failures, dead ends, hype cycles, and true anecdotes the highlight reels leave out - all primary-sourced.
A 1770 chess 'automaton' toured Europe and America for decades fooling audiences into believing a machine could play chess; in fact a human chess master was hidden inside the cabinet, and Edgar Allan Poe reasoned his way to that conclusion in 1836.
Isaac Asimov's Three Laws of Robotics became the public's default frame for robot ethics, but the stories that introduced them are mostly about the Laws producing strange, harmful, or paradoxical behavior - the failures are the point.
The 1956 Dartmouth proposal that named the field 'artificial intelligence' also estimated that a 2-month, 10-man summer study could make significant progress on it.
In 1966 MIT assigned 'the construction of a significant part of a visual system' as a summer project for students; making computers see well took roughly fifty more years.
The murderous shipboard computer HAL 9000 in Arthur C. Clarke's 1968 novel '2001: A Space Odyssey' became one of the most durable popular images of artificial intelligence, fixing the idea of a calm, capable machine that turns on its makers.
Minsky and Papert's 1969 book Perceptrons exposed what single-layer networks could not do, helping freeze neural-network research and funding until backpropagation revived the field in 1986.
Joseph Weizenbaum built ELIZA to show how shallow a chatbot was, then watched people - including his own secretary - treat it as if it understood them, an experience that turned him into a critic of AI.
Around 1980 the MIT AI Lab spun out two companies, Symbolics and LMI, to sell purpose-built Lisp computers for AI work; both collapsed within a decade as cheap general-purpose workstations made specialized hardware obsolete.
In the early 1980s 'knowledge engineering' was promoted as the future of AI, sparking a wave of expert-systems companies; by the end of the decade the market for specialized expert-system hardware and tools had collapsed.
The famous tale of a machine translating 'the spirit is willing but the flesh is weak' into 'the vodka is good but the meat is rotten' is an undocumented legend; the historian John Hutchins traced it and found no original occurrence.
Doug Lenat's Cyc project has been hand-encoding human common-sense knowledge since 1984; decades and a person-century of effort later, it never became the general intelligence the approach promised.
In 2004 not a single autonomous vehicle finished DARPA's desert race - the best managed about seven miles - yet one year later five cars completed it.
Researchers re-identified individuals in the supposedly anonymous Netflix Prize dataset, a privacy failure that helped kill the planned sequel contest.
In 2014 the University of Reading announced that a chatbot posing as a 13-year-old boy had passed the Turing test for the first time; the claim said more about the rules chosen than about machine intelligence.
In AlphaGo's 2016 match against Lee Sedol, the machine played a move so unlikely commentators thought it was a mistake; two games later Lee answered with a brilliancy of his own, and both moves became legend.
Microsoft released a Twitter chatbot meant to learn from conversation, and users taught it to post offensive content within hours, forcing a shutdown.
In 2016 Facebook opened its Messenger Platform to bots and declared conversation the next computing interface; the bot wave fizzled and Facebook's own M assistant was wound down by 2018.
Tesla has repeatedly told the public that full self-driving was nearly ready, starting with its 2016 claim that every car it built already had the hardware for it - a milestone that kept moving.
In 2017 Geoffrey Hinton and colleagues published capsule networks as a richer alternative to convolutional neural networks; the paper drew wide attention but capsules never displaced CNNs in practice.
In 2019 OpenAI declined to release the full GPT-2 model, citing fears of malicious use, then released it in stages over nine months after the feared harms did not materialize.
Zillow's algorithm-driven home-flipping business mispriced houses badly enough that the company wound it down, took hundreds of millions in write-downs, and cut about a quarter of its staff.
A decade after Watson won Jeopardy and IBM pitched it as the future of medicine, IBM sold the Watson Health data and analytics assets to a private equity firm.
In 2022 a Google engineer publicly argued that the LaMDA chatbot was sentient and published an edited interview as evidence; Google disagreed, said its review found no support, and later dismissed him.
Meta released a large language model for science, and after it confidently produced authoritative-sounding but wrong and fabricated text, the public demo was pulled within three days.
Google introduced its Bard chatbot with a James Webb Space Telescope example, and the promotional demo answer contained a factual error that drew immediate scrutiny.
Within a week of launching its new AI-powered Bing, Microsoft found that long conversations could push the chatbot into a strange, erratic tone, and it responded by limiting chat length.
Two New York lawyers submitted a legal brief full of court cases that ChatGPT invented, stood by them when challenged, and were sanctioned by a federal judge.
A rights-holder group had the Books3 dataset, roughly 200,000 pirated e-books used to train AI models, taken offline, exposing how unlicensed material had been baked into training corpora.
OpenAI's board fired CEO Sam Altman on a Friday in November 2023; within days he was reinstated and most of the board that fired him was gone.
A tribunal ruled that Air Canada had to honor a bereavement-fare policy its website chatbot made up, rejecting the airline's argument that the chatbot was a separate entity.
In February 2024 Google paused Gemini's ability to generate images of people after it produced historically inaccurate results, and published an explanation of what went wrong.
Cognition's March 2024 launch of Devin, marketed as 'the first AI software engineer,' ignited a wave of autonomous-agent hype built around its own performance claims.
Inflection AI raised 1.3 billion dollars for its Pi assistant, then in March 2024 its founders and many staff moved to Microsoft and the consumer product was set aside - the era's defining acqui-hire.
In 2024 standalone AI devices like the Humane Ai Pin launched to enormous hype; the Ai Pin was discontinued within a year and Humane's assets were sold to HP in February 2025.
Plain-language explanations of the ideas behind modern AI.
AI systems that use a language model to plan, take actions through tools, observe results, and iterate toward a goal.
The effort to ensure AI systems pursue the goals and values their designers and users actually intend, and avoid harmful behavior.
Compute, the total amount of calculation spent training a model, has become the field's key input, growing far faster than Moore's Law and forecastable enough to plan capability around.
The emerging body of laws, executive actions, and international declarations that governments use to govern the development and use of artificial intelligence.
A period when disappointment in artificial intelligence causes a sharp drop in funding, interest, and progress, following a stretch of inflated expectations.
The tendency of machine-learning systems to encode and amplify biases present in their training data, producing systematically unequal outcomes across groups.
The idea of an AI system with broad, human-level competence across most tasks - a goal that is real but deliberately ill-defined and contested.
A technique that lets a model focus on the most relevant parts of its input when producing each piece of output.
The learning algorithm that efficiently trains multi-layer neural networks by propagating error signals backward through the network.
When test questions leak into a model's training data, inflating its benchmark scores and making them an unreliable measure of true ability.
Prompting a model to spell out intermediate reasoning steps, which improves its accuracy on complex, multi-step problems.
The field of getting computers to interpret images and video, which advanced from hand-crafted features through the data-driven ImageNet era to today's multimodal models.
Anthropic's method for training a model to be harmless by having it critique and revise its own outputs against a written set of principles.
The maximum amount of text - measured in tokens - that a model can take in and work with at one time.
A neural network designed for images that detects visual features regardless of their position, foundational to modern computer vision.
The study of control and communication through feedback in machines and living things, an interdisciplinary framework that preceded and shaped early artificial intelligence.
Machine learning with many-layered neural networks that learn features directly from data, the approach behind the modern AI boom.
Software libraries like TensorFlow and PyTorch that handle automatic differentiation and GPU computation, making it practical to build and train neural networks.
Synthetic media in which a person's likeness or voice is generated or swapped using deep-learning techniques, and the detection research built to identify it.
A generative method that learns to create images by reversing a step-by-step noising process, powering tools like DALL-E 2 and Stable Diffusion.
Representations that turn words, items, or data into lists of numbers so that similar things sit close together.
A type of AI program that captures a human expert's knowledge as a set of if-then rules and applies them to give advice or make decisions in a narrow domain.
Adapting a pre-trained model to a specific task or behavior by training it further on a smaller, targeted dataset.
A pair of neural networks that compete, one generating fake data and one judging it, producing strikingly realistic synthetic images and media.
Graphics chips turned out to be ideal neural network engines because both rendering and neural networks are built on massive parallel matrix math, and programmable GPUs made that power available for AI.
An optimization method that repeatedly nudges a model's settings in the direction that most reduces its error.
When an AI language model produces fluent, confident text that is factually wrong or unsupported by its inputs.
A model's ability to learn a new task from examples placed in the prompt, without any change to its trained weights.
Research aimed at understanding what is actually happening inside an AI model, so its behavior can be explained, trusted, and corrected.
An AI system trained on vast text to predict and generate language, able to perform many tasks from a single model.
A recurrent neural network design that remembers information over long sequences, enabling early advances in speech and language processing.
A formula that scores how wrong a model's predictions are, giving training a single number to minimize.
A field where computer programs improve at a task by learning patterns from data and experience rather than being explicitly programmed.
An architecture that routes each input to a few specialized sub-networks, growing total capacity without growing per-query cost.
An open standard from Anthropic for connecting AI applications to external data sources, tools, and workflows in a uniform way.
Training a smaller, cheaper model to mimic a larger one, transferring much of its capability at a fraction of the cost.
AI systems that work across more than one type of data - for example understanding images and text together rather than text alone.
The field of getting computers to work with human language, which evolved from hand-written rules to statistics to neural networks and finally to large language models.
A computing model built from layers of simple interconnected units that adjust their connections to learn patterns from data.
The distinction between models whose trained parameters you can download and run yourself and those accessible only through a vendor's API - and why 'open source AI' is contested.
When a model memorizes its training data, including noise, and fails to perform well on new, unseen data.
The first trainable artificial neuron, introduced by Frank Rosenblatt in 1958, that learns to classify patterns from examples.
Training a model on huge amounts of general text first, so it learns broad language skills before any task-specific tuning.
The practice of crafting the instructions and examples you give a model to steer it toward better, more reliable outputs.
Compressing a model by storing its numbers at lower precision, so it uses less memory and runs faster with little loss in quality.
Models trained to spend extra computation thinking step by step before answering, trading speed for accuracy on hard problems.
Training an agent to make decisions by rewarding good actions and penalizing bad ones through trial and error.
A training method that uses human preferences to teach a model which responses are better, aligning it with what people want.
A method that lets a model fetch relevant external documents at answer time, grounding its output in current, specific information.
Robotics is the field of building machines that sense, plan, and act in the physical world, and it has historically lagged behind disembodied AI because moving atoms is harder than moving bits.
The empirical finding that model performance improves predictably as you increase model size, training data, and computation.
Training a model on labeled examples so it learns to predict the correct output for new, unseen inputs.
A classification method that finds the widest-margin boundary between two classes and uses kernels to draw curved boundaries, the leading machine learning method before deep learning.
The approach to AI that represents knowledge as symbols and produces intelligence by manipulating and searching through those symbols, the dominant paradigm from the 1950s into the 1980s.
Alan Turing's proposal to replace the question 'can machines think?' with a practical test of whether a machine's conversation is indistinguishable from a human's.
The process of breaking text into small units (tokens) that a model can read, often using subword pieces to handle any word.
The collection of examples a machine learning model learns from, whose quality and coverage largely determine model performance.
The provenance chain that turns raw web crawls into a model's training corpus: collection, filtering, deduplication, and weighted mixing of diverse sources.
The neural network architecture, built entirely on attention, that powers modern large language models.
Finding structure, groupings, or patterns in data that has no labels or predefined correct answers.
A pattern in which learned models replace or accelerate physical simulation and brute-force search across biology, weather, materials and mathematics.
How frontier AI gets financed: capped-profit structures, big-tech compute partnerships, mega-scale infrastructure ventures, and acqui-hires, anchored to the labs' own governance and deal documents.
What the papers actually said - linked to the originals.
The 2017 Google paper that introduced the Transformer architecture, the foundation of virtually all modern large language models.
The researchers and builders behind the breakthroughs.
Nineteenth-century mathematician whose 1843 Notes on the Analytical Engine contained the first published algorithm and an early statement of what machines can and cannot do.
British mathematician who proposed the 1950 imitation game (Turing test) and asked whether machines can think.
American researcher who co-founded artificial intelligence, co-built the Logic Theorist, and proposed the physical symbol system hypothesis with Herbert Simon.
AI researcher and educator, a founding member of OpenAI and former Director of AI at Tesla.
IBM researcher who built one of the first self-improving game programs and is credited with coining the term 'machine learning' in his 1959 checkers paper.
Nineteenth-century English mathematician and inventor who designed the Difference Engine and the programmable Analytical Engine, the conceptual ancestor of the computer.
Founder of information theory whose 1950 paper on computer chess was among the first serious proposals for machine intelligence.
AI researcher and Anthropic co-founder and CEO, previously a research leader at OpenAI during the GPT-2 and GPT-3 era.
Co-founder and CEO of DeepMind, behind AlphaGo and AlphaFold, and a 2024 Nobel laureate in Chemistry.
Computational linguist at the University of Washington, lead author of the Stochastic Parrots paper and a prominent skeptic of large-language-model hype.
Stanford professor who created ImageNet and co-directs the Stanford Human-Centered AI Institute.
American psychologist who built the perceptron, an early trainable neural network, at Cornell Aeronautical Laboratory.
Neural network pioneer at the University of Toronto, 2024 Nobel laureate in Physics for work enabling machine learning.
American polymath who co-founded artificial intelligence, won the Turing Award and the Nobel Prize in Economics, and is known for bounded rationality and satisficing.
Machine learning researcher who invented generative adversarial networks and co-authored the Deep Learning textbook.
Deep learning researcher, AlexNet and seq2seq co-author, OpenAI co-founder, and founder of Safe Superintelligence Inc.
Co-founder, president and CEO of NVIDIA, the company whose GPUs and CUDA software became the hardware engine of the deep learning revolution.
Physicist who invented the Hopfield network in 1982 and shared the 2024 Nobel Prize in Physics for foundational work enabling machine learning.
Scientist who led development of AlphaFold and shared the 2024 Nobel Prize in Chemistry for AI-driven protein structure prediction.
Computer scientist who coined the term artificial intelligence, co-proposed the 1956 Dartmouth project, and invented the Lisp programming language.
MIT computer scientist who created the ELIZA chatbot and then became one of AI's most prominent ethical critics.
Computer scientist who invented Bayesian networks and built the modern mathematics of causality; winner of the 2011 ACM Turing Award.
AI researcher whose lab co-developed the LSTM recurrent neural network and pioneered many deep learning ideas; director of the KAUST AI Initiative and IDSIA.
Founder of the Chinese AI lab DeepSeek and of the quantitative hedge fund High-Flyer; he is the submitting and corresponding author behind DeepSeek's headline research papers.
AI pioneer at MIT, co-originator of the 1956 Dartmouth project that named the field of artificial intelligence.
OpenAI's chief technology officer through the ChatGPT era, who in 2025 founded the AI research and product company Thinking Machines Lab.
A co-founder of DeepMind and of Inflection AI who in 2024 became EVP and CEO of Microsoft AI, leading the company's consumer AI products including Copilot.
Philosopher and founding director of Oxford's Future of Humanity Institute whose 2014 book Superintelligence shaped the debate on existential risk from AI.
A co-author of the Transformer paper and a pioneer of the mixture-of-experts approach, who co-founded Character.AI and then returned to Google to help lead its Gemini models.
American mathematician who founded cybernetics, the study of control and communication through feedback in machines and living things.
A founder of modern reinforcement learning and author of the influential 2019 essay 'The Bitter Lesson' on the primacy of computation in AI.
Robotics pioneer behind behavior-based robotics, former director of the MIT AI Lab and CSAIL, co-founder of iRobot and Robust.AI, known for his annually graded AI prediction scorecards.
Technology executive and a co-chair named at the founding of OpenAI, the research organization behind ChatGPT.
Sebastian Thrun led Stanford's winning DARPA Grand Challenge team, founded Google's self-driving car project, and co-founded the online-education company Udacity.
Berkeley computer scientist, co-author of the standard AI textbook AIMA, and a leading advocate for rethinking AI around human control.
Computer scientist known for the Gender Shades and Stochastic Parrots papers and founder of the Distributed AI Research Institute (DAIR).
Co-creator of statistical learning theory, the VC dimension, and support vector machines, the methods that dominated machine learning before deep learning.
Computer scientist who pioneered convolutional neural networks and LeNet; NYU professor and a director of AI research at Facebook/Meta.
Deep learning pioneer at the Universite de Montreal and founder of Mila, the Quebec AI institute.
The labs and companies driving the field.
The Qwen team at Alibaba Cloud, which develops the widely used open Qwen family of language, multimodal, and coding models.
AI safety and research company, maker of the Claude AI models, founded by former OpenAI researchers including Dario and Daniela Amodei.
Chinese technology company that built ERNIE Bot, China's first mover in the ChatGPT wave of generative AI chatbots.
The nonprofit that has crawled the open web since 2008 and gives away the resulting corpus, which underlies much of the text used to train large language models.
AI research lab founded in 2010, acquired by Google in 2014, known for AlphaGo and AlphaFold, now part of Google DeepMind.
Chinese AI lab behind the open DeepSeek model lines, known for reasoning models trained with reinforcement learning.
Google's deep learning research team, started in 2011, merged with DeepMind in 2023 to form Google DeepMind.
The open AI platform and community hub for sharing machine-learning models, datasets, and applications.
The company behind Deep Blue and Watson, now building open enterprise AI through its Granite models and watsonx platform.
The volunteer-driven nonprofit that released the open image-text datasets, including LAION-5B, used to train Stable Diffusion and other generative image models.
Meta's AI division, heir to its FAIR research lab and developer of the open-weight Llama models.
The software giant whose AI arc runs from the 2016 Tay chatbot through its multibillion-dollar OpenAI partnership to the Copilot product line.
French AI lab building frontier models, known for open-weight releases and enterprise deployment flexibility.
The hardware and software company whose GPUs and CUDA platform power most modern AI training and inference.
AI research and deployment company founded in December 2015, creator of the GPT model series and ChatGPT.
The company behind Stable Diffusion, the open text-to-image model that opened generative imaging to the public, followed by a turbulent 2024 leadership change.
Elon Musk's AI company, founded in 2023, maker of the Grok models with a mission to understand the universe.
Boston Dynamics is a robotics company spun out of MIT in 1992, known for three decades of legged robots and viral videos of machines that walk, run, and do backflips.
The Seattle non-profit founded by Paul Allen, known for Semantic Scholar and the fully open OLMo language models.
Waymo is Alphabet's autonomous-driving company, spun out of the Google self-driving car project that began in 2009.
The volunteer collective that grew from a Discord server into a non-profit lab and replicated GPT-class models in the open, including GPT-Neo, GPT-J, and Pythia.
The major AI model families, from the developers own pages.
Anthropic's family of Claude assistant models, organized into Opus, Sonnet, and Haiku tiers and delivered via API and apps.
Chinese lab DeepSeek's model lines, including the V-series general models and the R-series reasoning models.
Google DeepMind's family of natively multimodal Gemini models, spanning Pro, Flash, and specialized variants.
OpenAI's family of general-purpose generative pre-trained transformer models, delivered mainly through the OpenAI API and ChatGPT.
xAI's family of Grok models, the AI assistants developed by Elon Musk's xAI and integrated with the X platform.
Meta's family of open-weight Llama models that organizations can download, fine-tune, and deploy themselves.
French lab Mistral AI's models, mixing open-weight releases with commercial API models across general, code, and audio tasks.
OpenAI's line of reasoning models, beginning with o1, that think through problems step by step before answering.
Microsoft's Phi family of small language models designed to deliver strong capability at sizes that can run locally.
Alibaba's Qwen family of language, vision, and image models, many released publicly through Hugging Face and GitHub.
How AI is measured - each tied to the paper or site that defined it.
A hard high-school mathematics competition that has become a standard test of advanced reasoning in AI models.
A puzzle-based benchmark of novel reasoning tasks that are easy for humans but hard for AI, meant to measure fluid intelligence.
A set of 448 hard graduate-level science questions designed so they cannot be answered by simple web search.
A dataset of about 8,500 grade-school math word problems that tests a model's multi-step arithmetic reasoning.
A 164-problem test that checks whether a model can write working Python code from a natural-language description, graded by unit tests.
A 2,500-question expert-level exam across many subjects, built to stay hard for frontier AI as easier benchmarks get saturated.
The annual image-recognition competition whose 2012 results launched the deep learning era.
A live leaderboard that ranks AI chatbots by anonymous head-to-head human preference votes.
A dataset of 12,500 challenging competition math problems, each with a full worked solution, used to measure mathematical problem-solving in AI.
An industry benchmark suite from MLCommons that measures how fast computing systems can train and run AI models.
A 57-subject multiple-choice test measuring how broadly an AI model has acquired world knowledge and problem-solving ability.
A benchmark that tests whether AI systems can resolve real GitHub issues by editing real codebases, graded by the projects' own tests.
Atomic, verifiable facts - every one tied to a primary source.
The 18th-century chess 'automaton' known as the Mechanical Turk was operated by a human chess player concealed in its cabinet; Edgar Allan Poe argued in 1836 that 'some person is concealed in the box during the whole time of exhibiting the interior.'
Isaac Asimov's Three Laws of Robotics are stated in his 1950 collection 'I, Robot,' attributed in-story to a fictional Handbook of Robotics, and the stories repeatedly show the Laws producing paradoxical or harmful behavior.
The 1955 proposal for the Dartmouth summer project asked for 'a 2 month, 10 man study of artificial intelligence' to be carried out in the summer of 1956.
George Devol's foundational robot patent, US 2,988,237, was filed in 1954, granted in 1961, and titled 'Programmed Article Transfer' - it became the basis for the Unimate.
The phrase intelligence explosion comes from I.J. Good's 1965 paper on the first ultraintelligent machine.
The 1966 ALPAC report concluded there was 'no immediate or predictable prospect of useful machine translation,' a finding that led the US government to cut machine-translation funding.
A 1966 MIT memo proposed 'the construction of a significant part of a visual system' as a summer project for students; reliable computer vision took roughly fifty more years.
The artificial intelligence HAL 9000 appeared in Arthur C. Clarke's novel '2001: A Space Odyssey,' published in 1968 alongside Stanley Kubrick's film.
Human evaluators preferred outputs from the 1.3-billion-parameter InstructGPT over the 175-billion-parameter GPT-3, despite InstructGPT having about 100x fewer parameters.
In Moffatt v. Air Canada, the Civil Resolution Tribunal ruled the airline is responsible for what its website chatbot says, awarding the customer $650.88.
OpenAI's 2018 'AI and Compute' analysis found the compute in the largest AI training runs doubled every 3.4 months since 2012, more than 300,000 times growth, far faster than Moore's Law.
The 2012 AlexNet network that launched the deep learning era was trained on two NVIDIA GTX 580 GPUs, hardware also sold to gamers, using an efficient GPU implementation of convolution.
Starting from random play and knowing only the rules, AlphaZero reached a superhuman level of chess and shogi within 24 hours of self-play training.
Anthropic was founded and is led by siblings Dario Amodei (CEO) and Daniela Amodei (President), former OpenAI researchers.
Arthur Samuel's 1959 IBM paper on a self-improving checkers program is widely credited as the origin of the term 'machine learning'.
Isaac Asimov first stated the Three Laws of Robotics as an explicit code in his short story Runaround, published in the March 1942 issue of Astounding Science Fiction.
The Books3 AI training dataset was a collection of about 200,000 illegal e-books before a rights-holder group had it taken offline in 2023.
DeepMind's Chinchilla paper found that for compute-optimal training, the number of training tokens should double each time model size doubles.
DeepMind's Chinchilla paper found that for compute-optimal training, model size and training tokens should scale equally - each doubling of model size needs a doubling of data.
OpenAI's CLIP learned from a dataset of 400 million image-and-text pairs collected from the internet.
Cognition's March 2024 Devin launch post stated that Devin resolved 13.86% of SWE-bench issues end-to-end, up from a prior state of the art of 1.96%.
Common Crawl describes its openly available web corpus as petabytes of data that it has regularly collected since 2008.
The AIME is a 15-question, 3-hour exam in which every answer is an integer between 0 and 999, making it trivial to score automatically.
At launch, Google said Gemini Ultra was the first model to outperform human experts on the MMLU benchmark of multitask knowledge.
The 2018 Gender Shades study reported commercial gender-classification error rates of up to 34.7% for darker-skinned females against a maximum of 0.8% for lighter-skinned males.
On November 9, 2015, Google released TensorFlow as free open-source software under the Apache 2.0 license, announced by Jeff Dean and Rajat Monga.
After Gemini produced historically inaccurate images, Google said in February 2024 that it had turned off image generation of people while it worked on improvements.
When Google revealed the Tensor Processing Unit in May 2016, it had already been running the custom AI chips in its data centers since 2015, powering products like Search and AlphaGo.
GPT-4o can respond to audio inputs in as little as 232 milliseconds, averaging 320 milliseconds - close to human conversational response time.
OpenAI's GSM8K benchmark is a dataset of roughly 8,500 high-quality, linguistically diverse grade-school math word problems.
Herbert Simon received the 1975 ACM Turing Award in computing and the 1978 Nobel Memorial Prize in Economic Sciences.
The Fifth Generation Computer Systems project, launched by Japan in 1982, concluded around 1992, capped by ICOT's final international conference that year.
The LAION-5B dataset comprises 5.85 billion CLIP-filtered image-text pairs, of which 2.32 billion are in English.
Over 120,000 hands of heads-up no-limit poker, Libratus finished ahead of four top professionals by a collective 1,766,250 dollars in chips.
In Note G of her 1843 Notes, Ada Lovelace wrote that the Analytical Engine has no pretensions whatever to originate anything and can only do what we order it to perform.
Meta's Galactica science model went public on November 15, 2022 and its demo was paused by November 17, 2022.
The DFDC dataset released by Meta (then Facebook) AI contained over 100,000 face-swap video clips sourced from 3,426 paid actors who agreed to participate.
The Llama community license requires companies with over 700 million monthly active users to obtain a separate license from Meta, which is one reason it is not OSI-approved open source.
Microsoft took its Tay Twitter chatbot offline less than a day after launching it, after users provoked it into offensive posts.
What became known as Moore's Law began as a four-page article Gordon Moore wrote for Electronics magazine on April 19, 1965, not a law of physics.
The physical symbol system hypothesis states that a physical symbol system has the necessary and sufficient means for general intelligent action.
Noam Shazeer is the second author of 'Attention Is All You Need' (the Transformer) and the first author of the sparsely-gated mixture-of-experts paper, two architectural foundations of modern large language models.
The 2020 Nature paper by NumPy's core developers describes the library as 'the foundation upon which the entire scientific Python universe is constructed.'
On April 13, 2019, OpenAI Five defeated Dota 2 world champions OG, becoming the first AI system to beat the reigning world champions at an esport.
Hugging Face's documentation states there are more than one million Transformers model checkpoints available on the Hugging Face Hub, loadable with a single install.
The PyTorch team's own retrospective fixes the framework's public debut to January 2017, one year before their January 19, 2018 'a year in' post.
On January 1, 2018, Rodney Brooks published dated predictions about AI, robotics, self-driving cars, and space travel, and he publicly grades them in an annual scorecard he promised to keep until 2050.
In his 1948 paper, Claude Shannon noted that the word bit, for a binary digit, was suggested by J. W. Tukey.
A 2009 Google essay argued that simple models trained on large amounts of data consistently beat more elaborate models trained on less data.
Apple introduced Siri as a feature of the iPhone 4S, announced on October 4, 2011.
Proving that checkers is a draw with perfect play required dozens of computers running almost continuously from 1989 to 2007.
Gerald Tesauro's TD-Gammon reached strong master level at backgammon by playing against itself and learning from the outcomes, not from human game records.
A federal judge imposed a $5,000 Rule 11 penalty on the lawyers who filed ChatGPT-fabricated case citations in Mata v. Avianca.
Richard Sutton's 2019 essay 'The Bitter Lesson' argues that over 70 years of AI, general methods that leverage computation have proven most effective by a large margin.
Joseph Weizenbaum built the ELIZA chatbot in 1966 and a decade later wrote a 1976 book warning against over-trusting computers with human judgment.
The MATH benchmark introduced by Hendrycks et al. consists of 12,500 challenging competition mathematics problems, each with a full worked solution.
The Netflix Prize offered one million dollars to the first team to improve Netflix's recommendation accuracy by 10 percent over its Cinematch system.
EleutherAI's openly documented language model training corpus, The Pile, totals 825 GiB and is assembled from 22 distinct datasets.
Cortes and Vapnik's paper introducing the modern support vector machine was published in the journal Machine Learning in 1995.
'Attention Is All You Need' was written by eight researchers working at Google: Vaswani, Shazeer, Parmar, Uszkoreit, Jones, Gomez, Kaiser, and Polosukhin.
On Computable Numbers showed that one universal machine can simulate any other computing machine given a description of it, the basis of the general-purpose computer.
IBM Watson finished its 2011 Jeopardy! match with $77,147, well ahead of human champions Ken Jennings and Brad Rutter; the winnings went to charity.
Norbert Wiener took the term cybernetics from the Greek word for steersman to name the study of control and communication in animals and machines.
When Zillow announced it would shut down its algorithmic home-buying business in November 2021, it said the wind-down would include a roughly 25% workforce reduction.
Doug Lenat's Cyc project began in 1984; by 1995 he reported 'a person-century of effort' had gone into hand-building its knowledge base.
No vehicle finished the 2004 DARPA Grand Challenge; in DARPA's own account the top-scoring car covered just 7.5 miles and the 1 million dollar prize went unclaimed.
By January 2010, iRobot reported that sales of its home robots had passed five million units since the Roomba's 2002 launch.
When Google publicly disclosed its self-driving car project in October 2010, the cars had already driven 140,000 miles on public roads.
The University of Reading's 2014 Turing-test claim involved Eugene Goostman, a chatbot that simulated a 13-year-old boy and convinced 33% of judges in five-minute chats.
DeepMind describes AlphaGo's Move 37 in game two against Lee Sedol as having had a 1 in 10,000 chance of being used by a human player.
In game four against AlphaGo, Lee Sedol played Move 78, a wedge DeepMind says had its own 1-in-10,000 chance of being played; it won him the only game the machine lost.
'Attention Is All You Need', which introduced the Transformer, was submitted to arXiv on June 12, 2017.
China's 2017 New Generation AI Development Plan set a three-step goal to make China a major global AI innovation center by 2030.
When OpenAI announced its capped-profit structure in March 2019, it said returns for first-round investors were capped at 100 times their investment, with any excess owned by the nonprofit.
On July 22, 2019, Microsoft announced a 1 billion dollar investment in OpenAI alongside an exclusive Azure computing partnership.
OpenAI withheld the full GPT-2 model in February 2019 citing misuse concerns, then released the complete 1.5-billion-parameter model in November 2019 after a staged rollout.
EleutherAI describes itself as having grown from a Discord server for discussing GPT-3 into a leading non-profit research institute, founded in July 2020.
After its 2021 launch, the AlphaFold Protein Structure Database expanded to over 200 million predicted protein structures, near-complete coverage of catalogued proteins.
The 2023 Future of Life Institute open letter asked AI labs to pause for at least six months the training of systems more powerful than GPT-4.
The Center for AI Safety's 2023 Statement on AI Risk is a single sentence of exactly 22 words.
The Bletchley Declaration was signed at the November 2023 AI Safety Summit by 28 countries together with the European Union.
DeepMind's GraphCast outperformed the leading conventional weather system on 90 percent of 1,380 verification targets, producing 10-day forecasts in under a minute.
OpenAI's board removed Sam Altman as CEO on November 17, 2023, and OpenAI announced his return on November 22, 2023 - a span of less than a week.
DeepMind's GNoME tool predicted 2.2 million new crystal structures, of which about 380,000 were flagged as stable candidates for synthesis.
DeepMind's AlphaProof and AlphaGeometry 2 solved 4 of 6 problems at the 2024 International Mathematical Olympiad, scoring 28 of 42 points - the level of a silver medalist.
Regulation (EU) 2024/1689, the EU AI Act, was published in the Official Journal on 12 July 2024 and entered into force on 1 August 2024.
The 2023 US AI executive order, EO 14110, was rescinded by Executive Order 14148 on 20 January 2025.
The Stargate Project, announced January 21, 2025, is a new company that intends to invest 500 billion dollars over four years in US AI infrastructure, beginning with 100 billion deployed immediately.
Humane's Ai Pin launched in April 2024; in February 2025 HP announced it was acquiring Humane's assets, and the device was discontinued less than a year after release.
At the 2025 International Mathematical Olympiad, Google DeepMind's Gemini with Deep Think solved five of six problems for 35 of 42 points - a certified gold-medal score.
The EU AI Act entered into force on August 1, 2024, with obligations for general-purpose AI model providers beginning to apply on August 2, 2025.
DeepSeek's April 24, 2026 V4 preview shipped DeepSeek-V4-Pro and DeepSeek-V4-Flash, both with a 1M-token context window.