AI Development Timeline

Kurt Gödel published his incompleteness theorems, proving that any sufficiently powerful formal system contains propositions that cannot be proven or disproven within that system. This discovery not only shocked the mathematical community but also had profound implications for computing theory and artificial intelligence.

Alonzo Church and Alan Turing independently proposed the famous Church-Turing Thesis. This thesis asserts that any effectively computable function is Turing-computable. This formalizes the intuitive notion of "computability" and became the foundation of computer science.

Alan Turing introduced the concept of the Turing machine in his landmark paper "On Computable Numbers, with an Application to the Entscheidungsproblem". This abstract computational model can simulate any mathematically computable process.

Alonzo Church proposed Lambda Calculus, a formal system for function definition and application. Lambda Calculus is regarded as the theoretical foundation of functional programming. The original 1936 lambda calculus was untyped; Church added simple types in 1940 ("A Formulation of the Simple Theory of Types"), creating one of the earliest explicitly-typed systems.

Warren McCulloch and Walter Pitts published "A Logical Calculus of the Ideas Immanent in Nervous Activity", introducing the first artificial neuron model (McCulloch-Pitts neuron). This model simplified neurons to binary threshold switches, laying the foundation for later neural network research.

Norbert Wiener published "Cybernetics: Or Control and Communication in the Animal and the Machine", founding the new discipline of cybernetics. The study of how systems regulate themselves through feedback mechanisms had profound impact on AI and robotics.

Claude Shannon published "A Mathematical Theory of Communication", laying the foundation for information theory. Concepts like entropy, coding, and information quantity became core tools for understanding data and processing information.

In "Computing Machinery and Intelligence," Turing replaced the ambiguous question "Can machines think?" with the imitation game, originally a three-role game (a man, a woman, and an interrogator) where Turing proposed substituting a machine for one of the players. It is a behavioral criterion, not proof that a machine is conscious.

In 1955, John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon used the term "artificial intelligence" in the proposal for the Dartmouth Summer Research Project and outlined a summer 1956 research program.

The summer 1956 Dartmouth workshop helped establish "artificial intelligence" and its research agenda as the identity of an emerging academic field; the term had already appeared in the project’s 1955 proposal.

Allen Newell and Herbert Simon developed the "Logic Theorist", an early practical AI program. This program could prove theorems from Principia Mathematica and is widely discussed as one of the earliest AI programs. Newell and Simon received the Turing Award in 1975.

Frank Rosenblatt proposed the Perceptron model, an influential early artificial neural-network model. The Perceptron is a simple linear binary classifier that learns to classify input data by adjusting weights.

John McCarthy developed Lisp at MIT, an early high-level programming language closely associated with AI research. Lisp’s flexibility and powerful symbolic processing made it the language of choice for AI research.

Newell and Simon developed the General Problem Solver (GPS) from 1957 to 1959, with collaborators including J. C. Shaw. It was the first AI program to attempt to simulate human problem-solving by separating the problem-solving technique from the knowledge of the specific problem domain, using goal decomposition and sub-goal search (means-ends analysis).

Frank Rosenblatt proved the convergence property of the perceptron learning algorithm, showing that if data is linearly separable, the perceptron will definitely converge.

Unimate became the first robot to work on an assembly line. General Motors installed it at its Trenton, New Jersey plant in 1961 to unload a die-casting press, taking the hot cast parts off the machine. Unimation built these units for work that was difficult, dangerous or monotonous for people.

In his 1963 MIT doctoral thesis, Larry Roberts described a system for interpreting scenes of simple polyhedral solids from images, an important early foundation of three-dimensional computer vision.

MIT launched Project MAC in 1963 under Robert Fano to study time-sharing, computation, and machine-aided cognition. The MIT Artificial Intelligence Project, founded in 1959, later participated in it; the two were not a single laboratory founded in 1959.

Joseph Weizenbaum created ELIZA at MIT, one of the earliest chatbots. ELIZA simulated a psychotherapist’s conversation through pattern matching, sparking deep reflection on human-computer interaction.

SRI International developed Shakey, an early autonomous mobile robot that could perceive its environment, plan paths, and execute tasks. Shakey combined computer vision, NLP, and planning algorithms.

Marvin Minsky and Seymour Papert published "Perceptrons," a rigorous analysis of what single-layer perceptrons can and cannot represent. The book is often associated with the later decline in interest in neural networks, but funding shifts also reflected limited computing power, unmet expectations, and broader policy decisions.

Alain Colmerauer, Philippe Roussel, and their collaborators developed the first version of Prolog in Marseille, applying predicate logic to programming and advancing logic programming and symbolic AI.

British scientist James Lighthill submitted a pessimistic report on AI research to the UK government, leading to significant funding cuts.

Marvin Minsky published the "Frame Paper", proposing Frame theory for knowledge representation. This theory influenced later object-oriented programming and knowledge representation methods.

Stanford University developed MYCIN, an expert system for medical diagnosis. It demonstrated the practicality of expert systems, though never used clinically due to liability issues.

DEC deployed XCON (eXpert CONfigurer), an expert system for configuring VAX computer systems. Development began in 1978 at CMU; initial deployment was 1979, with full operational status by 1980 — one of the first commercially successful expert systems, saving the company millions annually.

Following critical assessments and funding cuts in the 1970s, parts of AI research entered a downturn. "First AI winter" is a retrospective periodization whose timing and severity varied by country and subfield.

Japan announced the Fifth Generation Computer Project (FGCS) at a JIPDEC / MITI conference in October 1981, a 10-year national initiative to develop computers capable of logical reasoning and knowledge processing. The project was formally launched with the founding of ICOT (Institute for New Generation Computer Technology) in April 1982.

David Rumelhart, Geoffrey Hinton, and Ronald Williams published on the backpropagation algorithm, making training multi-layer neural networks possible.

In the late 1980s, expectations that had run ahead of results went unmet. Lisp machine makers lost out to general-purpose workstations, several AI companies closed, and DARPA’s funding for basic AI research fell between 1987 and 1989.

In the 1990s, statistical methods began to dominate AI research. Hidden Markov Models (HMMs) in speech recognition, Naive Bayes and decision trees in machine learning, all achieved good results.

The US military used AI-based planning and logistics systems in the Gulf War, demonstrating AI’s potential in real-world applications.

Tim Berners-Lee created the World Wide Web. The internet’s proliferation changed human lifestyle and provided unprecedented data resources for AI.

The downturn eased gradually rather than ending on a single date. Many researchers turned to more modest, achievable goals, while renewed work on neural networks and probabilistic methods laid the foundation for a new era.

Vladimir Vapnik et al. proposed Support Vector Machine (SVM), a powerful classification and regression method. SVM achieved excellent performance in text classification and handwritten recognition.

IBM’s Deep Blue defeated world chess champion Garry Kasparov, demonstrating AI’s ability to surpass human capabilities in well-defined tasks.

Sepp Hochreiter and Jürgen Schmidhuber proposed LSTM, a special type of RNN capable of learning long-term dependencies. LSTM achieved great success in speech recognition and language modeling.

Yann LeCun developed LeNet-5, a CNN for handwritten digit recognition. LeNet-5 was among the first deep learning models deployed at scale commercially — used by AT&T/NCR for check reading starting in 1993 and by the US Postal Service for ZIP code recognition from the late 1990s.

Yoshua Bengio, Réjean Ducharme, Pascal Vincent, and Christian Jauvin proposed a neural probabilistic language model that jointly learns distributed word representations and word-sequence probabilities to address the curse of dimensionality. The model is feed-forward, not recurrent.

Beginning in 2004, a CIFAR program brought Geoffrey Hinton, Yann LeCun, Yoshua Bengio, and collaborators into regular exchange. It helped coordinate research that contributed to the deep-learning revival; it was not a single universal starting point.

Geoffrey Hinton and his colleagues proposed Deep Belief Networks (DBN), with a fast, greedy algorithm that trains them one layer at a time. This layer-wise pre-training helped revive interest in deep neural networks.

Researchers began using GPUs to accelerate deep learning training. GPU’s parallel computing made training large neural networks possible.

The ImageNet team published its large-scale hierarchical image database paper in 2009, organizing labeled images with the WordNet hierarchy. ImageNet and its later recognition challenge became important infrastructure for large-scale computer-vision training and evaluation.

Microsoft Research and the University of Toronto pioneered the first major deep-learning breakthrough in speech recognition (Dahl, Yu, Deng 2011; Mohamed, Dahl, Hinton 2012), achieving significant word-error-rate reductions. Google deployed DNN/LSTM-based speech recognition in production starting in 2012–2015.

Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton’s team entered AlexNet in the 2012 ImageNet competition (ILSVRC) and won by a landslide: a top-5 error of 15.3%, against 26.2% for the runner-up.

Tomas Mikolov released Word2Vec, a neural network model for learning word embeddings. Word2Vec enabled computers to understand semantic relationships between words.

Ian Goodfellow proposed GAN, one of the most important innovations in deep learning. GAN generates extremely realistic images through adversarial training.

Kaiming He et al. proposed ResNet, solving the training difficulty of deep networks through skip connections. ResNet enabled training hundreds of layers.

DeepMind’s AlphaGo defeated world Go champion Lee Sedol 4-1. Go was considered one of the most difficult board games for AI.

Google proposed the Transformer architecture, a revolutionary breakthrough in NLP. Transformer is the foundation of modern large language models.

OpenAI released GPT-3, a large language model with 175 billion parameters. GPT-3 demonstrated astonishing language generation capabilities, sparking widespread discussion about AGI.

OpenAI released DALL-E (image generation) and CLIP (image-text contrastive learning). These models demonstrated AI’s ability to understand and generate multimodal content.

OpenAI released Codex, an AI model that understands and generates code. GitHub Copilot became the first widely-used AI programming assistant.

In November 2022, OpenAI released ChatGPT, the first mass-market generative-AI conversational assistant. ChatGPT reached 100 million users in two months, becoming the fastest-growing consumer product in history.

Stability AI released Stable Diffusion, an image generation model whose weights can be freely downloaded, under a license that restricts certain uses. Open-source community activity rapidly popularized image generation technology.

OpenAI released GPT-4, a multimodal model capable of processing images and text. GPT-4 showed near-human performance on professional and academic benchmarks.

Anthropic released Claude, an AI assistant focused on safety and helpfulness. Claude’s design reflects emphasis on AI safety—AI should be helpful, harmless, and honest.

Meta released LLaMA, a large language model under a non-commercial research license. Although LLaMA has relatively small parameter size, its excellent performance promoted the development of open-weight LLMs.

The AI Agent concept began to receive widespread attention. Agents are AI systems capable of autonomously planning and executing complex tasks.

In February 2024, Google introduced Gemini 1.5 with a mixture-of-experts architecture and an experimental context window of up to one million tokens, demonstrating retrieval and reasoning across long documents, codebases, and video.

Anthropic released the Claude 3 series, including Opus, Sonnet, and Haiku. Claude 3 Opus performed competitively with GPT-4 on several benchmarks including MMLU and GPQA, while trailing GPT-4 on others.

Meta released the Llama 3 family with open weights, including 8B and 70B parameter models.

The European Union enacted the AI Act, the first comprehensive horizontal regulation of artificial intelligence. The Act established a risk-based classification and strict requirements for high-risk AI systems.

OpenAI released the o1 model, a large model with reasoning capabilities. o1 performs excellently on scientific and programming tasks, demonstrating "thinking" ability.

In October 2024, Anthropic introduced a public beta of computer use for an upgraded Claude 3.5 Sonnet. Developers could direct Claude to inspect a screen, move a cursor, click, and type, making it one of the first frontier models publicly offered with general graphical-interface control.

Anthropic released MCP (Model Context Protocol), an open standard protocol for connecting AI models with external tools and data sources. MCP is becoming the "USB-C" of AI applications.

By 2025, AI agents matured from research demos to production tools. Coding agents, browser-use agents, and multi-step task agents became standard offerings, with OpenAI, Anthropic, Google, and open-source projects all shipping capable systems.

Chinese AI lab DeepSeek released R1, an open-weight reasoning model, with code and weights under the MIT license, whose performance rivaled OpenAI’s o1. The release demonstrated that frontier capabilities could be achieved outside the US AI establishment and triggered a global re-evaluation of AI competition.

OpenAI, Oracle, SoftBank, and the UAE’s MGX jointly announced that they intend to invest $500 billion over four years to build AI infrastructure in the United States, including multiple large-scale data centers.

Anthropic released Claude 3.7 Sonnet, which it described as the first hybrid reasoning model generally available: users can switch between instant answers and extended thinking in the same model.

In March 2025, Anthropic announced a $3.5 billion funding round at a $61.5 billion valuation, making it one of the highest-valued AI startups in the world and reflecting further capital concentration in the AI industry.

Meta released the natively multimodal open-weight Llama 4 Scout and Maverick models and previewed Behemoth, a teacher model that was still training and was not released. Scout and Maverick use mixture-of-experts architectures.

Anthropic released the Claude 4 family (Opus 4 and Sonnet 4), introducing Extended Thinking capability with significant improvements in programming, reasoning, and long-horizon tasks. Claude 4 consolidated Anthropic’s position in the enterprise AI market.

On 2 August 2025, the EU AI Act’s governance rules and obligations for general-purpose AI models became applicable. Other provisions follow separate transition dates, so this milestone is not a blanket start of full enforcement.

OpenAI released GPT-5, a unified flagship model that integrates the reasoning capabilities of the o-series with the conversational abilities of the GPT series. GPT-5 set a new frontier in programming, mathematics, and reasoning while supporting longer contexts and tool use.

OpenAI completed the recapitalization of its corporate structure. The for-profit business became a public benefit corporation named OpenAI Group PBC, and the OpenAI Foundation — the nonprofit — continues to control it under the same mission. OpenAI says the recapitalization followed nearly a year of discussions with the California and Delaware Attorneys General.

Through the first half of 2026, agent capabilities moved out of demonstrations and into shipped enterprise products: coding, support, and office-automation agents became documented features, and tool-calling and multi-agent infrastructure matured around them. How far adoption actually reached is not settled. Published surveys report very different production rates because they do not agree on what counts as an agent, so this entry records the direction of travel rather than a level.

Through the first half of 2026, open-weight releases from DeepSeek, Alibaba Qwen, Moonshot, Zhipu, and Meta were published at a scale previously seen only in closed systems, with weights and model cards downloadable rather than described. Whether any of them match the closed frontier depends on which benchmark index is used and how it is weighted, so this entry records what was released and under what licence, not a ranking.

Nature published Co-Scientist, a multi-agent system built on Gemini that formulates novel research hypotheses for experimental verification. The paper reports validation across three biomedical applications: drug repurposing, discovery of novel targets, and explaining mechanisms of antimicrobial resistance. The authors present it as an assistant to researchers rather than a replacement, which makes it peer-reviewed evidence for AI participating in hypothesis generation rather than only in analysis.

Reporting on June 11, 2026 said Meta had begun unwinding its acquisition integration with AI agent company Manus, halting data sharing and separating operations. The report linked the move to Chinese regulatory requirements; later legal and operational status should be checked against updated disclosures.

Anthropic launched safeguarded Claude Fable 5 for general use and the less-restricted Mythos 5 for vetted Project Glasswing partners on June 9, 2026. On June 12, U.S. export controls required restrictions on foreign nationals; unable to verify nationality in real time, Anthropic suspended both models globally. The controls were lifted on June 30. Fable 5 returned globally on July 1, while Mythos 5 initially returned for approved U.S. organizations.

On June 13, 2026, a coalition of US state attorneys general led by New York opened an investigation into OpenAI and issued a subpoena. The subpoena demands documents relating to advertising placement, user engagement and retention, model sycophancy, consumer and health data handling, and protections for minors and the elderly. OpenAI said it would cooperate with the attorneys general and noted that ChatGPT has built stronger safeguards for minors.

On 29 June 2026, Samsung and SK Hynix announced multi-year semiconductor investment plans totaling 800 trillion won and involving four memory fabs in southwestern South Korea. Dollar conversions vary with exchange rates, while scope, approvals, and execution remain subject to review.

On July 1, 2026 Google released Gemini Spark — its 24/7 agentic AI assistant — for macOS, with real-time tracking and broader app integration. Spark can take persistent action across a user’s digital life, marking Google’s push to bring agentic AI onto the consumer desktop.

In July 2026 Moonshot AI published open weights for Kimi K3 on Hugging Face. The model card records 2.8 trillion total parameters with 104 billion activated, a 1,048,576-token context window, and the Kimi K3 License. By total parameter count it was among the largest openly downloadable language models at its release.

On 1 July 2026, Cloudflare announced separate controls for Search, Agent, and Training traffic, and said that from 15 September new Cloudflare domains would block Training and Agent crawlers by default on ad-bearing pages while allowing Search by default; site owners can change the settings, and multi-purpose crawlers follow the strictest applicable rule.

On July 2, 2026 Microsoft announced Microsoft Frontier Company, a new operating business focused on enterprise AI deployments using Microsoft’s existing AI tools. The project is backed by a $2.5 billion Microsoft investment and 6,000 industry and engineering experts — joining Amazon and OpenAI in launching a dedicated AI deployment organization.

On July 9, 2026, after ending a limited preview, OpenAI made the GPT-5.6 family generally available: Sol, Terra, and Luna. During the preview, OpenAI worked closely with expert organizations and trusted partners.

On August 2, 2026 the EU AI Act became generally applicable, and the AI Office together with national authorities took up implementation, supervision, and enforcement. The same revision deferred the heaviest high-risk duties: stand-alone Annex III systems apply from December 2, 2027, and AI embedded in regulated products from August 2, 2028. Prohibited practices had already applied since February 2, 2025, and general-purpose model obligations since August 2, 2025.

On September 2, 2026 ARC Prize published verified ARC-AGI-3 results for GPT-6 Astra across twelve harness configurations. On the same Semi-Private evaluation set the model scores 62.71% with the Standard harness, which lets it carry forward notes it chooses to keep, and 98.55% with a Provider Adapter harness, which preserves reasoning state between requests and compacts long conversations. Astra used fewer actions than the median tested human on 96% of levels. The figure measures a model together with its scaffolding, which is why both are recorded here.

On September 3, 2026 OpenAI published a system card placing GPT-6 Astra at the Critical threshold for cyber capability under the company’s own Preparedness Framework, the first OpenAI model at that level. The card describes a model that can find previously unknown security flaws and develop ways to exploit them across many well-protected systems without a person guiding each step, and the sensitive capabilities are gated behind a trusted-access programme. The classification is OpenAI’s assessment of its own model against its own framework, and the card does not use the words AGI or general intelligence.