A chronology, 1936–2025
The History of Artificial Intelligence in Primary Sources
38 milestones, from Turing’s 1936 paper on computable numbers to reasoning models in 2025. Each entry names its authors and institution, states in two sentences what the work established and what limitation followed, and links to the original publication.
1936–1955
Foundations and Formalisms
Formal models of computation, neural logic, information and machine play are set out before the field has a name.
-
1936
On Computable Numbers, with an Application to the Entscheidungsproblem
Defined an abstract machine model of computation and used it to prove that the Entscheidungsproblem has no general algorithmic solution. The model settles what can be computed in principle but says nothing about the resources, or the methods, by which a machine could exhibit intelligent behaviour.
TheoryProceedings of the London Mathematical Society, series 2, vol. 42, 1936–37, pp. 230–265doi:10.1112/plms/s2-42.1.230
-
1943
A Logical Calculus of the Ideas Immanent in Nervous Activity
Modelled neurons as binary threshold units and showed that networks of such units can realise the expressions of propositional logic. The units had fixed connections, so the model described what a network could compute and not how a network could learn.
TheoryBulletin of Mathematical Biophysics 5(4), pp. 115–133doi:10.1007/BF02478259
-
1948
A Mathematical Theory of Communication
Defined entropy as a measure of information and proved limits on the rate at which a message can be sent reliably over a noisy channel. The theory treats information independently of meaning, so questions of semantics and reasoning lay outside it.
TheoryBell System Technical Journal 27(3) and 27(4), pp. 379–423 and 623–656doi:10.1002/j.1538-7305.1948.tb01338.x
-
1950
Computing Machinery and Intelligence
Proposed replacing the question whether machines can think with an imitation game, and answered a series of anticipated objections to the possibility of machine intelligence. The paper offers a behavioural criterion and a sketch of a learning machine, but no working program.
TheoryMind 59(236), pp. 433–460doi:10.1093/mind/LIX.236.433
-
1950
Programming a Computer for Playing Chess
Described how a computer could play chess by scoring positions with an evaluation function and searching ahead through possible moves. The number of possible games makes exhaustive search impracticable, so the paper turns to selective search guided by heuristics.
Systems and BenchmarksPhilosophical Magazine, series 7, 41(314), pp. 256–275doi:10.1080/14786445008521796
1956–1973
Symbolic Reasoning and Early Learning
The field is named, programs prove theorems, play games and hold conversations, and the first sustained critiques of its methods and results appear.
-
1956
A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence
Proposed a summer study of the conjecture that every aspect of learning or any other feature of intelligence can in principle be described so precisely that a machine can be made to simulate it, and used the term artificial intelligence for the field. The proposal expected a significant advance within a single summer, an expectation the following decades did not bear out.
TheoryProposal dated 31 August 1955 for the 1956 summer projectwww-formal.stanford.edu archive copy
-
1956
The Logic Theory Machine
Described a program that proved theorems of propositional logic from Principia Mathematica by heuristic search in place of exhaustive enumeration of proofs. Its heuristics were designed for that one domain, and carrying the approach to other problems required further programs.
Systems and BenchmarksIRE Transactions on Information Theory 2(3), pp. 61–79doi:10.1109/TIT.1956.1056797
-
1958
The Perceptron: A Probabilistic Model for Information Storage and Organization in the Brain
Introduced the perceptron, a network of units with modifiable connection weights that learns to classify patterns from examples. A single layer of adjustable weights can represent only linearly separable classifications, a restriction later analysed in detail.
ArchitecturePsychological Review 65(6), pp. 386–408doi:10.1037/h0042519
-
1959
Some Studies in Machine Learning Using the Game of Checkers
Reported checkers-playing programs that improved with experience, through rote learning of positions and adjustment of the weights of a scoring polynomial. The candidate features were chosen by hand, and the learning was confined to a single game.
LearningIBM Journal of Research and Development 3(3), pp. 210–229doi:10.1147/rd.33.0210
-
1966
ELIZA: A Computer Program for the Study of Natural Language Communication between Man and Machine
Described a program that held typed conversations by matching keywords in the user’s text and applying transformation rules, in a script modelled on a Rogerian psychotherapist. The paper itself notes that explaining the program’s inner workings removes the appearance of understanding that it can create.
Systems and BenchmarksCommunications of the ACM 9(1), pp. 36–45doi:10.1145/365153.365168
-
1969
Perceptrons: An Introduction to Computational Geometry
Gave a mathematical analysis of what single-layer perceptrons can and cannot compute, including results on parity and connectedness. The analysis left open how far multilayer networks could extend those limits, a question taken up again when multilayer training methods appeared.
TheoryMIT Press, 1969; reissued by MIT Press, 2017doi:10.7551/mitpress/11301.001.0001
-
1973
Artificial Intelligence: A General Survey
Surveyed AI research for the United Kingdom Science Research Council and judged that discoveries in the field had not yet produced the major impact that had been promised. The report is commonly cited in accounts of the reduced support for AI research in the United Kingdom that followed.
TheoryArtificial Intelligence: a paper symposium, Science Research Council, 1973chilton-computing.org.uk archive copy
1980–2011
Knowledge, Search and Statistical Representation
Learning rules for multilayer networks, probabilistic reasoning and specialised game-playing machines develop alongside the first large labelled datasets.
-
1982
Neural Networks and Physical Systems with Emergent Collective Computational Abilities
Showed that a network of symmetrically connected threshold units has an energy function that decreases as the network updates, so that stored patterns act as stable states from which a partial pattern is completed. Storage capacity is only a fraction of the number of units, and the network can also settle into spurious states.
ArchitectureProceedings of the National Academy of Sciences 79(8), pp. 2554–2558doi:10.1073/pnas.79.8.2554
-
1986
Learning Representations by Back-Propagating Errors
Described back-propagation, a procedure that adjusts the weights of a multilayer network by propagating error derivatives backward from the output, and showed hidden units learning useful internal representations. The authors note that the error surface may contain local minima, so gradient descent is not guaranteed to find the best solution.
LearningNature 323, pp. 533–536doi:10.1038/323533a0
-
1988
Learning to Predict by the Methods of Temporal Differences
Introduced temporal-difference methods, which learn to predict by moving each estimate toward the estimate that follows it, and analysed their convergence for a restricted case. The analysis covered linear predictors only, and convergence under more general function approximation remained to be established.
LearningMachine Learning 3(1), pp. 9–44doi:10.1007/BF00115009
-
1988
Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference
Presented Bayesian networks, a graphical representation of conditional independence among variables, together with message-passing methods for updating beliefs under uncertainty. Exact message passing is efficient on tree-structured networks, and networks with loops require more costly or approximate methods.
TheoryMorgan Kaufmann, 1988doi:10.1016/C2009-0-27609-4
-
1989
Backpropagation Applied to Handwritten Zip Code Recognition
Applied a convolutional network with shared, constrained weights and training by back-propagation to the recognition of handwritten zip code digits. The network handled small digit images and a dataset of a few thousand examples, far from the scale of general image recognition.
ArchitectureNeural Computation 1(4), pp. 541–551doi:10.1162/neco.1989.1.4.541
-
1997
Deep Blue
Described the IBM chess machine, built from parallel hardware with custom chess chips, that defeated world champion Garry Kasparov in a six-game match in May 1997. The system’s strength came from chess-specific hardware and a hand-tuned evaluation function, and it did not transfer to other tasks.
Systems and BenchmarksArtificial Intelligence 134(1–2), 2002, pp. 57–83doi:10.1016/S0004-3702(01)00129-1
-
1997
Long Short-Term Memory
Introduced long short-term memory, a recurrent unit with gated memory cells that keep error signals from vanishing or exploding over long time lags. Because each step depends on the previous one, computation along a sequence cannot be parallelised across time steps.
ArchitectureNeural Computation 9(8), pp. 1735–1780doi:10.1162/neco.1997.9.8.1735
-
1998
Gradient-Based Learning Applied to Document Recognition
Showed that multilayer networks trained by gradient descent, including the convolutional network LeNet-5, could form the core of document recognition systems, with results on the MNIST handwritten digit set. The approach depended on labelled data and computation that were not widely available when the paper appeared.
ArchitectureProceedings of the IEEE 86(11), pp. 2278–2324doi:10.1109/5.726791
-
2009
ImageNet: A Large-Scale Hierarchical Image Database
Described a database of labelled images organised according to the WordNet hierarchy and collected with crowdsourced annotation. Its labels cover object categories only, so the resource measures recognition within a fixed vocabulary of classes.
Systems and BenchmarksIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2009, pp. 248–255doi:10.1109/CVPR.2009.5206848
2012–2021
Deep Learning, Scale and Structure
Networks trained on graphics processors and large datasets displace hand-built features, and generative, game-playing and attention-based models follow.
-
2012
ImageNet Classification with Deep Convolutional Neural Networks
Trained a deep convolutional network on graphics processors and won the 2012 ImageNet large-scale recognition challenge with a top-5 error of 15.3%, against 26.2% for the runner-up. The network had 60 million parameters, took five to six days to train on two GPUs, and relied on dropout to limit overfitting.
ArchitectureAdvances in Neural Information Processing Systems 25 (NeurIPS 2012)papers.nips.cc
-
2013
Efficient Estimation of Word Representations in Vector Space
Proposed two shallow architectures, continuous bag-of-words and skip-gram, that learn word vectors from large text corpora at low cost, and showed that vector offsets capture syntactic and semantic relations. Each word receives a single vector regardless of context, so different senses of a word are not separated.
LearningICLR 2013 workshoparXiv:1301.3781
-
2013
Auto-Encoding Variational Bayes
Introduced a stochastic variational inference method whose reparameterised lower-bound estimator allows models with continuous latent variables to be trained by gradient descent. The reparameterisation applies to continuous latent variables, so discrete latent variables need other estimators.
LearningICLR 2014arXiv:1312.6114
-
2014
Generative Adversarial Nets
Proposed training a generative model against a discriminator in a two-player minimax game, so that the generator learns without an explicit likelihood. The authors note that the generator’s distribution has no explicit representation and that the two networks must be kept in step during training.
LearningAdvances in Neural Information Processing Systems 27 (NeurIPS 2014)arXiv:1406.2661
-
2015
Human-Level Control through Deep Reinforcement Learning
Combined Q-learning with a deep convolutional network to learn control policies from raw pixels and game score, reaching a level comparable to a professional human tester across 49 Atari 2600 games. Performance was weak on games that need long-term planning, such as Montezuma’s Revenge.
LearningNature 518, pp. 529–533doi:10.1038/nature14236
-
2016
Mastering the Game of Go with Deep Neural Networks and Tree Search
Combined deep neural networks trained by supervised learning and reinforcement learning with Monte Carlo tree search, and defeated the European champion Fan Hui five games to zero. The policy network was first trained on records of expert human games, and the system was designed for Go.
Systems and BenchmarksNature 529, pp. 484–489doi:10.1038/nature16961
-
2017
Attention Is All You Need
Introduced the Transformer, a sequence model built on attention alone, without recurrence or convolution, and reached 28.4 BLEU on the WMT 2014 English-to-German translation task. Self-attention compares every position with every other, so its cost grows quadratically with sequence length.
ArchitectureAdvances in Neural Information Processing Systems 30 (NeurIPS 2017)arXiv:1706.03762
-
2017
Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm
Showed that one reinforcement-learning algorithm, trained by self-play from the rules alone and with no human game data, could reach superhuman strength in chess, shogi and Go. The method depends on a simulator of the game rules, so it applies only where such a simulator exists.
LearningPreprintarXiv:1712.01815
-
2018
BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
Pre-trained a deep bidirectional Transformer encoder with masked language modelling and, after fine-tuning, reported new state-of-the-art results on eleven natural language processing tasks. The model produces representations rather than free-form text, and each task needed fine-tuning on labelled data.
LearningNAACL-HLT 2019arXiv:1810.04805
-
2020
Language Models Are Few-Shot Learners
Trained a 175-billion-parameter autoregressive language model and evaluated it on many tasks given only instructions or a few examples in the prompt, without gradient updates. The paper reports weaknesses on some tasks, including natural language inference, and discusses data contamination and bias.
Systems and BenchmarksAdvances in Neural Information Processing Systems 33 (NeurIPS 2020)arXiv:2005.14165
-
2021
Highly Accurate Protein Structure Prediction with AlphaFold
Predicted protein structures from amino-acid sequence with accuracy competitive with experiment in most cases in the 14th Critical Assessment of Structure Prediction. The method predicts single static structures and reports a per-residue confidence measure that marks regions where predictions are unreliable.
Systems and BenchmarksNature 596, pp. 583–589doi:10.1038/s41586-021-03819-2
2022–2025
Reasoning, Multimodal and Formal Systems
Models are tuned with human feedback, scaling behaviour is quantified, and systems are trained to spend computation on reasoning before they answer.
-
2022
Training Language Models to Follow Instructions with Human Feedback
Fine-tuned GPT-3 with supervised learning on human demonstrations and reinforcement learning from human preference rankings, and found that outputs of a 1.3-billion-parameter InstructGPT model were preferred to those of the 175-billion-parameter GPT-3. The models still made simple mistakes, and the preferences encoded were those of a small group of labelers and researchers.
LearningAdvances in Neural Information Processing Systems 35 (NeurIPS 2022)arXiv:2203.02155
-
2022
Training Compute-Optimal Large Language Models
Trained over 400 language models to estimate compute-optimal training, found that model size and training tokens should grow in equal proportion, and showed the 70-billion-parameter Chinchilla outperforming the 280-billion-parameter Gopher. The prescription needs training sets far larger than those then in use and addresses training compute, not the cost of running a model.
TheoryPreprintarXiv:2203.15556
-
2024
Solving Olympiad Geometry without Human Demonstrations
Combined a language model trained on synthetic proofs, without human demonstrations, with a symbolic deduction engine and solved 25 of 30 olympiad-level geometry problems, against 10 for the previous best method. The system covers Euclidean plane geometry only and relies on a purpose-built symbolic engine and synthetic data generator.
Systems and BenchmarksNature 625, pp. 476–482doi:10.1038/s41586-023-06747-5
-
2024
Scaling LLM Test-Time Compute Optimally Can Be More Effective than Scaling Model Parameters
Studied how extra computation spent at inference, through search against a verifier or revision of answers, trades off against model size, and found that a compute-optimal allocation improved efficiency by more than four times over a best-of-N baseline. How much inference computation helps depends on problem difficulty, and on the hardest problems additional pretraining was more effective.
TheoryPreprintarXiv:2408.03314
-
2024
OpenAI o1 System Card
Reported models trained with large-scale reinforcement learning to reason through a chain of thought before answering, and described the safety evaluations run on them. The card describes evaluations rather than methods, and the model, its training data and its reasoning traces were not published.
Systems and BenchmarksPreprintarXiv:2412.16720
-
2025
DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning
Showed that reinforcement learning applied directly to a base model, without supervised fine-tuning, elicited long chain-of-thought reasoning, and released the resulting model weights. The first model, DeepSeek-R1-Zero, produced hard-to-read output that mixed languages, which the full DeepSeek-R1 pipeline addressed with cold-start data and further training stages.
LearningPreprintarXiv:2501.12948
No milestone matches these filters.