From computing’s foundations to today’s frontier—and the possibilities of general intelligence and superintelligence.
Established systems for language, vision, prediction and decision-making.
Broadly capable intelligence: a research goal with contested definitions.
Hypothetical intelligence surpassing humans across most domains—not an established achievement.
Ada Lovelace’s notes on the Analytical Engine describe an algorithm and consider machines manipulating symbols, including music.
Why it matters A conceptual foundation for general-purpose computing—not an AI system.
Concept / research goalScience MuseumAlan Turing describes an abstract computing machine and the limits of computability.
Why it matters Provides the mathematical foundation for programmable computers.
Turing Digital ArchiveWarren McCulloch and Walter Pitts model networks of neurons using mathematical logic.
Why it matters Connects neural activity to computation and inspires artificial neural networks.
Bulletin of Mathematical BiophysicsDonald Hebb proposes that connections strengthen when neurons repeatedly activate together.
Why it matters An influential hypothesis for learning through changes in connection strength.
Concept / research goalThe Organization of BehaviorComputing Machinery and Intelligence reframes machine intelligence through a conversational imitation game.
Why it matters Introduces an influential behavioural test, not a universal definition of intelligence.
Concept / research goalMind · Alan TuringSamuel develops a checkers-playing program and later experiments with learning from experience.
Why it matters Early evidence that a program could improve its game-playing performance.
IBM
The Dartmouth summer research project brings together researchers pursuing machine intelligence; its 1955 proposal uses the term artificial intelligence.
Why it matters AI becomes a named research field.
DartmouthFrank Rosenblatt develops a trainable perceptron for pattern recognition.
Why it matters An early learning model that helps establish neural-network research.
Perceptron historyJohn McCarthy develops LISP, a programming language designed around symbolic expressions.
Why it matters Becomes a central tool for AI research for decades.
LISP historyGood speculates that an ultraintelligent machine could design better machines, creating a feedback loop.
Why it matters A foundational superintelligence hypothesis, not evidence of an achieved system.
Concept / research goalI. J. Good · speculationJoseph Weizenbaum publishes ELIZA, a text-processing program famous for simulating a therapist.
Why it matters Shows how easily people attribute understanding to scripted dialogue.
Communications of the ACMMinsky and Papert analyse limitations of particular perceptron architectures.
Why it matters Highlights the importance of architecture; the findings do not rule out all neural networks.
PerceptronsStanford researchers begin developing MYCIN to recommend treatments for bacterial infections using rules.
Why it matters Demonstrates specialised expert reasoning; it was a research system, not routine clinical deployment.
MYCIN historyA critical UK research review contributes to funding cuts, alongside unmet expectations and technical limitations.
Why it matters AI progress is uneven, with setbacks as well as breakthroughs.
AI winter historyDigital Equipment Corporation deploys a rule-based system for configuring computer orders.
Why it matters Shows commercial value in narrow, knowledge-rich domains.
XCON historyRumelhart, Hinton and Williams demonstrate learning internal representations using backpropagation.
Why it matters Popularises a method for training multilayer networks; earlier contributions predate this paper.
NatureThe Lisp-machine market contracts and many expert-system investments disappoint.
Why it matters Exposes maintenance costs and the gap between demonstrations and scalable products.
AI winter historyHochreiter and Schmidhuber introduce LSTM networks to improve learning over long sequences.
Why it matters Influences speech recognition, translation and sequence modelling.
Neural ComputationIBM’s chess computer wins a six-game match against the reigning world champion.
Why it matters A landmark in specialised search and evaluation, not general intelligence.
IBMLeCun and colleagues describe convolutional networks for document recognition.
Why it matters Establishes influential patterns for modern computer vision.
Proceedings of the IEEEFive vehicles complete the desert course, following a 2004 contest with no finishers.
Why it matters Accelerates research into autonomous navigation.
DARPA Grand ChallengeThe edited volume Artificial General Intelligence collects research on broadly capable machine intelligence.
Why it matters Helps consolidate AGI as a research community; general intelligence remains a goal.
Concept / research goalSpringerResearchers introduce ImageNet, a large labelled image dataset.
Why it matters Shared datasets and competitions make progress in vision easier to compare.
ImageNetIBM Watson defeats champions Ken Jennings and Brad Rutter in a televised quiz competition.
Why it matters Demonstrates question answering combining language processing and information retrieval.
IBMA GPU-trained deep convolutional network wins the ImageNet competition by a substantial margin.
Why it matters Triggers widespread adoption of deep learning in image recognition.
NeurIPSNick Bostrom’s Superintelligence examines possible paths to advanced intelligence and control risks.
Why it matters Makes long-term alignment a prominent debate; the scenarios are hypothetical.
Debate / safety initiativeOxford University PressGoodfellow and colleagues propose competing generator and discriminator networks.
Why it matters A major advance in learning to generate realistic synthetic data.
arXivOpenAI announces a research organisation focused on beneficial artificial intelligence.
Why it matters Adds a major institution to the pursuit of general-purpose AI.
Concept / research goalOpenAIDeepMind’s AlphaGo defeats the Go champion four games to one.
Why it matters Combines learned policies and value estimates with search in a difficult game.
Google DeepMind
Researchers introduce the Transformer architecture, dispensing with recurrence in their sequence model.
Why it matters Attention-based architectures become central to many modern language and multimodal models.
arXivOpenAI demonstrates language-model pre-training followed by supervised fine-tuning.
Why it matters Helps establish a general-purpose approach to language tasks.
OpenAIGoogle researchers introduce bidirectional pre-training for language representations.
Why it matters Strong transfer across language-understanding benchmarks broadens the impact of Transformers.
arXivOpenAI initially withholds the largest GPT-2 model because of concerns about misuse.
Why it matters Model release decisions become part of the AI safety debate.
OpenAIThe Language Models are Few-Shot Learners paper describes a 175-billion-parameter model.
Why it matters Shows that prompting can adapt a large model to many tasks without task-specific training.
arXivDeepMind reports highly accurate protein-structure predictions in the CASP14 assessment.
Why it matters A transformative advance in structural biology, while many biological challenges remain.
Google DeepMindOpenAI demonstrates a model producing images from natural-language descriptions.
Why it matters Brings text-to-image generation to wider attention.
OpenAIDeepMind presents a generalist agent trained on text, images, games and robot-control tasks.
Why it matters An experiment in shared multi-task models—not a demonstration of AGI.
Concept / research goalarXivStability AI releases Stable Diffusion model weights to the public.
Why it matters Makes powerful image generation widely runnable and adaptable.
Stability AIOpenAI releases a conversational interface trained to follow instructions with human feedback.
Why it matters Brings generative AI into everyday public and workplace use.
OpenAIOpenAI announces GPT-4, supporting text and image inputs with text outputs.
Why it matters Advances broad task performance; benchmark results do not establish AGI.
OpenAIOpenAI announces research into steering and controlling systems much more capable than humans.
Why it matters An institutional safety initiative, not evidence that superintelligence exists.
Debate / safety initiativeOpenAICountries at the UK AI Safety Summit agree on the need to understand and address frontier AI risks.
Why it matters Establishes an important international framework for safety cooperation.
UK GovernmentOpenAI previews a model generating videos from text prompts.
Why it matters A significant step in generative video; the initial announcement is a research preview.
OpenAIOpenAI announces a model trained across text, vision and audio, with capabilities rolled out in stages.
Why it matters Moves toward lower-latency interaction across modalities.
OpenAIThe European Union’s risk-based AI law enters into force, with obligations phased in over time.
Why it matters Creates a major legal framework for AI development and deployment.
European CommissionOpenAI introduces models trained to spend more computation working through problems before answering.
Why it matters Highlights inference-time computation as another route to improved task performance.
OpenAIJohn Hopfield and Geoffrey Hinton receive the Nobel Prize in Physics.
Why it matters Recognises foundational discoveries enabling machine learning with artificial neural networks.
Nobel PrizeDavid Baker is recognised for protein design; Demis Hassabis and John Jumper for protein-structure prediction.
Why it matters Acknowledges AI’s growing contribution to scientific discovery.
Nobel PrizeDeepSeek releases R1 and smaller distilled models, with its main model under an MIT licence.
Why it matters Broadens access to reasoning-model research and deployment.
DeepSeekOpenAI introduces GPT-5 as a system routing requests between fast and deeper-reasoning models.
Why it matters Reflects convergence between conversational AI and task-oriented reasoning; it is not a verified AGI milestone.
OpenAIAnthropic releases Sonnet 4.5 with coding, computer-use and agent-development improvements.
Why it matters Shows the growing emphasis on models operating tools and completing multi-step tasks.
AnthropicOpenAI announces GPT-6.1 Sol, emphasising coding, computer use and professional workflows at lower cost than its Astra model.
Why it matters Illustrates the continued focus on tool-using AI and cost-efficient deployment; capability claims are the developer’s, not evidence of AGI.
OpenAIOpenAI announces wider GPT-6 availability and Intelligent UI, allowing answers to include interactive visual interfaces.
Why it matters Expands AI interaction beyond text into task-specific visual experiences; this release is not a declaration of AGI.
OpenAIIntelligence that substantially exceeds human capabilities across most important domains. Its feasibility, timing and safe development remain unresolved.
No verified achievement or forecast date is assigned here. Historical SI entries record ideas and safety work, not an existing superintelligent system.
Oxford University PressSource-linked, curated history—not an exhaustive event log. Dates use source-supported precision. AGI and SI concepts are distinguished from established breakthroughs; new milestones are curated separately from the automatic news feed.