A mathematical neuron
An early mathematical account of how networks of idealized neurons could carry out logical operations.
Warren S. McCulloch and Walter PittsRead entry
A history through research and practice
The ideas that shaped
artificial intelligence.
Trace the publications behind the breakthroughs, from mathematical neurons to reasoning models. Open any entry for the research, its significance, and its limits.
1943–2025
67 selected worksOriginal sources linked throughout
Before the research · 1816–1899
Precursor to AI
Literary and philosophical precursors
The Sandman, Erewhon, and Moxon’s Master explored attachment to artificial beings, machines surpassing humanity, and the loss of human control. Follow those ideas and Turing’s later reference to Butler.
Read the articleExplore the wider cultural history in the separate books and films timeline, from antiquity to 2024.
Explore by AI topic
67 entries across all topics.
No entries match this topic. Choose another topic or clear the filter.
Foundations & symbolic beginnings
11 entriesLearning leaves a lasting neural trace
Cell assemblies and activity-dependent changes connected learning with the organization of neural circuits.
Donald O. HebbRead entryThe imitation game
Turing reframed the question of machine intelligence through an operational test based on conversation.
Alan M. TuringRead entryThe Logic Theory Machine
Heuristic search made symbolic theorem proving a concrete computational task.
Allen Newell and Herbert A. SimonRead entryLearning with the perceptron
A trainable model connected adaptive weights with pattern recognition.
Frank RosenblattRead entryLisp and symbolic computation
Recursive operations on symbolic expressions provided a foundation for Lisp.
John McCarthyRead entrySIR stores facts and reasons over them
A small language system built an internal model from statements and searched it to answer later questions.
Bertram RaphaelRead entrySTUDENT turns word problems into equations
A restricted English interface translated algebra story problems into equations and solved for unknowns.
Daniel G. BobrowRead entryResolution gives theorem proving a common rule
Resolution combined substitution and logical inference in a method designed for computer theorem proving.
J. A. RobinsonRead entryELIZA and conversational patterns
Scripted pattern matching produced a convincing surface of conversation.
Joseph WeizenbaumRead entrySemantic memory as a network of concepts
Linked concepts let a memory support questions that were not specified when its information was stored.
M. Ross QuillianRead entry
Knowledge, memory & learning
23 entriesRecall extends through time
A temporal analogy to holography explored how familiar sequences might be recognized and recalled.
H. C. Longuet-HigginsRead entryQA1 and QA2 derive answers from stored facts
Question answering combined stored facts, relevant retrieval, and deductions that were not explicitly stored.
C. Cordell Green and Bertram RaphaelRead entryMapping the limits of perceptrons
A mathematical analysis clarified limitations of restricted perceptron architectures.
Marvin Minsky and Seymour PapertRead entryRemembering a story through relations
A system stored the relations in an English story so that it could answer questions about the story later.
Alan L. Tharp and Gilbert K. KruleeRead entryNew language connects to stored knowledge
The Teachable Language Comprehender related new text to a semantic network and represented what it learned in the same form.
M. Ross QuillianRead entryQA3 constructs solutions through proofs
An extended resolution procedure used proofs to construct solutions to formally described problems.
Cordell GreenRead entryPLANNER makes reasoning procedures programmable
A problem-solving language connected assertions and goals with procedures that controlled inference.
Carl HewittRead entryCARPS reasons over calculus word problems
Structured descriptions of objects helped a program interpret a limited class of calculus rate problems.
Eugene CharniakRead entrySAMENLAQ II reasons with a persistent memory network
An interactive system stored facts and relation definitions in a network, with memories that could be saved to disk.
Stuart C. Shapiro and George H. WoodmanseeRead entrySHRDLU grounds language in a small world
Language, reasoning, and action were integrated in a constrained blocks world.
Terry WinogradRead entryExpert knowledge becomes explicit rules
Early MYCIN research explored rule-based advice in a tightly defined specialist domain.
Edward H. Shortliffe and colleaguesRead entryPersistent activity as a memory mechanism
A mathematical neural model explored conditions under which patterns of activity could persist.
W. A. LittleRead entryRetrieving a memory from a cue
A systematic account connected stored associations, cue-based recall, and adaptive mathematical models.
Teuvo KohonenRead entryFinding information by its content
Storage and retrieval were organized around matching content rather than specifying a memory address.
Teuvo KohonenRead entryMemory as a network attractor
Recurrent networks linked associative memory with the dynamics of physical systems.
John J. HopfieldRead entryLearning internal representations
Backpropagation demonstrated how hidden units could learn useful internal representations.
David E. Rumelhart, Geoffrey E. Hinton and Ronald J. WilliamsRead entryLearning while a sequence unfolds
Dynamic networks extended error-based learning from static patterns to time-varying streams.
A. J. Robinson and F. FallsideRead entryPrior activity becomes context
Feeding earlier hidden activity back into a network created a learned representation of temporal context.
Jeffrey L. ElmanRead entryThe expressive power of a hidden layer
Under stated conditions, feedforward networks can approximate broad classes of functions.
Kurt Hornik, Maxwell Stinchcombe and Halbert WhiteRead entryRecurrent state learns a grammar
A simple recurrent network learned temporal context sufficient to recognize a studied finite-state grammar.
Axel Cleeremans, David Servan-Schreiber and James L. McClellandRead entryActivity traces preserve temporal information
Self-connected hidden units integrated a sequence while local traces supplied information needed for learning.
Michael C. MozerRead entryTraining recurrent memory in real time
Real-time recurrent learning was applied to recognizing sequential structure in finite-state grammars.
Anthony W. Smith and David ZipserRead entryThe AI Toy
Three commercial magazine articles brought neural-network learning to the Commodore 64, giving readers a practical baseline for what AI could do on a home computer.
Kevin E. MartinRead entry
Statistical learning to deep vision
7 entriesSupport-vector networks
Margin-based classification offered a powerful approach to nonlinear learning.
Corinna Cortes and Vladimir VapnikRead entryLong Short-Term Memory
Memory cells and gates addressed difficulties in learning long-range dependencies.
Sepp Hochreiter and Jürgen SchmidhuberRead entryConvolutional networks read documents
Convolutional networks and gradient-based training connected representation learning with document recognition.
Yann LeCun, Léon Bottou, Yoshua Bengio and Patrick HaffnerRead entryTraining deeper networks layer by layer
Greedy layerwise learning provided an effective training strategy for deep belief networks.
Geoffrey E. Hinton, Simon Osindero and Yee-Whye TehRead entryImageNet expands the data frontier
A large, organized image database created new opportunities for visual recognition research.
Jia Deng and colleaguesRead entryA shared toolkit for machine learning
A consistent software interface made established learning methods easier to use and compare.
Fabian Pedregosa and colleaguesRead entryDeep convolutional learning at ImageNet scale
GPU-trained convolutional networks substantially improved large-scale image classification.
Alex Krizhevsky, Ilya Sutskever and Geoffrey E. HintonRead entry
Deep learning takes shape
7 entriesEfficient word representations
Efficient training made useful continuous word vectors available at large scale.
Tomas Mikolov and colleaguesRead entryGenerative adversarial learning
A generator and discriminator learned through an adversarial training objective.
Ian Goodfellow and colleaguesRead entryDropout regularizes neural networks
Randomly omitting units during training helped reduce overfitting.
Nitish Srivastava and colleaguesRead entryDeep reinforcement learning from pixels
A deep Q-network learned control policies from game images and rewards.
Volodymyr Mnih and colleaguesRead entryResidual connections make depth easier to train
Residual learning enabled effective optimization of much deeper image-recognition networks.
Kaiming He, Xiangyu Zhang, Shaoqing Ren and Jian SunRead entryNeural networks meet tree search in Go
Policy and value networks strengthened search in a difficult board game.
David Silver and colleaguesRead entryTensorFlow scales the learning system
A dataflow system supported machine-learning workloads across different computing devices.
Martín Abadi and colleaguesRead entry
The Transformer era
4 entriesAttention Is All You Need
The Transformer replaced recurrence with attention in a sequence-to-sequence architecture.
Ashish Vaswani and colleaguesRead entryBERT learns bidirectional language representations
Masked pretraining helped a bidirectional Transformer transfer to many language tasks.
Jacob Devlin, Ming-Wei Chang, Kenton Lee and Kristina ToutanovaRead entryGPT-2 explores zero-shot task transfer
A language model performed several tasks without task-specific parameter updates.
Alec Radford and colleaguesRead entryPyTorch bridges flexibility and performance
An imperative programming style supported flexible research with efficient tensor computation.
Adam Paszke and colleaguesRead entry
Scale, transfer & generation
5 entriesLanguage-model scaling becomes measurable
Empirical relationships connected language-model loss with parameters, data, and compute.
Jared Kaplan and colleaguesRead entryGPT-3 learns tasks from examples in context
A large language model performed many tasks using demonstrations supplied in its prompt.
Tom B. Brown and colleaguesRead entryDenoising diffusion produces high-quality images
Learning to reverse gradual noising became an effective image-generation method.
Jonathan Ho, Ajay Jain and Pieter AbbeelRead entryCLIP connects images with natural language
Contrastive learning from image-text pairs enabled transfer through natural-language descriptions.
Alec Radford and colleaguesRead entryDiffusion moves into a compressed latent space
Performing denoising in a learned latent space reduced the computational burden of image synthesis.
Robin Rombach and colleaguesRead entry
Instructions, feedback & reasoning
4 entriesWorked reasoning examples guide prompting
Prompts containing intermediate reasoning examples improved performance on several reasoning benchmarks.
Jason Wei and colleaguesRead entryHuman feedback shapes instruction following
Supervised examples and human preference feedback improved how language models followed instructions.
Long Ouyang and colleaguesRead entryZero-shot step-by-step prompting
A simple step-by-step instruction improved reasoning results without worked examples in the prompt.
Takeshi Kojima and colleaguesRead entryPrinciples guide AI feedback
Written principles guided self-critique and AI-generated feedback during training.
Yuntao Bai and colleaguesRead entry
Multimodal & reasoning models
6 entriesLLaMA examines efficient foundation models
A family of language models emphasized training efficiency and publicly available training data.
Hugo Touvron and colleaguesRead entryGPT-4 reports broader multimodal capability
A large model accepted image and text inputs and generated text outputs.
OpenAIRead entryMistral 7B focuses on efficient attention
Grouped-query and sliding-window attention supported an efficient language-model architecture.
Albert Q. Jiang and colleaguesRead entryMixtral activates a sparse set of experts
Sparse routing selected two of eight feedforward experts per token at each layer.
Albert Q. Jiang and colleaguesRead entryGemini 1.5 studies long multimodal context
Long-context experiments included text, video, and audio within a multimodal model family.
Gemini TeamRead entryDeepSeek-R1 explores reinforcement learning for reasoning
Reinforcement learning improved reasoning benchmarks, with distinct training paths for R1-Zero and R1.
DeepSeek-AIRead entry
Reading the timeline
Follow the evidence.
Dates belong to publications
Years refer to the specific work linked in each entry. For papers first released on arXiv, the timeline uses the first submission year and notes a later conference when applicable. An invention, a software release, and a paper may have different dates.
Claims stay within the source
Entries summarize a contribution and its limits. Journal papers, conference papers, preprints, technical reports, a foundational monograph, and a commercial magazine series with programs are labeled separately. Being a primary source does not mean every claim has been independently reproduced.
This is a curated history of AI research and practice, not an exhaustive catalogue or a ranking of current models. Era names are editorial navigation, not universally agreed scientific periods. Entries within a year are not necessarily ordered by month.
The selection builds on the MTS research-paper timeline and Tracing the Evolution of Artificial Intelligence, with dates and descriptions checked against the original publications and authoritative records linked on each detail page. The latter review is a preprint; neither overview substitutes for the original evidence.
Source review: 16 September 2026. Historical coverage ends in 2025. External links may lead to a publisher record, an open manuscript, or a PDF; some publishers restrict full-text access.