Artificial intelligence

An editorial selection for comparing processes, events, and lives. The interval and counts describe selected records, not the historical boundaries or importance of the subject.

Canonical records: 81
Events: 66
People: 12
Processes and periods: 3

Compare in the atlas Selection criteria and limits

Topics represented

Artificial intelligence (81) · Aprendizaje automático (28) · People (11) · Redes neuronales (6) · Nobel laureates (5) · Seguridad y control (5) · Gobernanza (4) · IA simbólica (3) · Instituciones y financiación (3) · Aprendizaje por refuerzo (2) · Physics (2) · Lenguaje (2)

View distribution by region and period

Canonical records by regional label and start date. These are numerical analysis bins, not universal historical periods. A multi-region record appears in multiple rows; do not add rows as a unique-record total.
Region	1900–1999	Since 2000
Africa	0	2
Americas	22	16
Asia	1	2
Europe	10	4
Oceania	0	0
Global / no specific region	6	25

Records and sources

Ordered by start date. Long lives and processes may begin before the period you want to study.

1988
Pearl systematizes Bayesian networks
Judea Pearl published Probabilistic Reasoning in Intelligent Systems, systematizing graphical representations and inference under uncertainty. This tradition enabled reasoning with incomplete evidence, broadening the scope of deterministic logical rules.
1992
Q-learning: learning to act through reward
Christopher Watkins and Peter Dayan published a convergence proof for Q-learning, proposed by Watkins in 1989. The method learns action values from experience and became a foundation of reinforcement learning.
1995
Support-vector machines
Corinna Cortes and Vladimir Vapnik presented support-vector networks for classification. Their margin formulation and feature transformations consolidated a statistical-learning approach that coexisted with neural networks.
1997
Deep Blue defeats Kasparov
IBM’s computer defeats world champion Garry Kasparov in a chess match, a public landmark for specialized AI. The result relied on search and specialized position evaluation, not conversational or general capabilities.
1997
LSTM learns dependencies in sequences
Sepp Hochreiter and Jürgen Schmidhuber introduced Long Short-Term Memory to address difficulties learning long dependencies in recurrent networks. Gated memory units expanded sequence applications before Transformers became dominant.
2005
Autonomous vehicles complete the Grand Challenge
Several teams completed the 2005 DARPA Grand Challenge course, won by Stanford’s Stanley, after none finished in 2004. The advance brought together perception, planning and control, without yet demonstrating general driving in everyday traffic.
2006
Deep learning revival
Work on deep networks and pretraining renews practical interest in deep learning. Deep belief networks offered a practical way to initialize multilayer models before supervised fine-tuning.
2009
ImageNet organizes data for computer vision
Jia Deng, Fei-Fei Li and colleagues introduced ImageNet, a category-organized image collection. Its data and later competitions made vision systems comparable and illustrate the role of annotation work in AI progress.
2009–… (as of 2026-09-20)
Fei-Fei Li
Fei-Fei Li led the creation of ImageNet and promoted computer-vision and human-centered AI research. This activity profile begins with the 2009 ImageNet publication and connects data, evaluation and decisions about technological development.
2011
Watson wins at Jeopardy!
IBM’s Watson won a televised competition against Ken Jennings and Brad Rutter. It combined language processing and answer evaluation, demonstrating progress on open questions within bounded rules and a defined domain.
2012
AlexNet wins ImageNet
AlexNet achieves a decisive leap in image classification and accelerates adoption of deep convolutional networks. The combination of data, GPU computing and training made a transition in computer vision visible.
2013
Adversarial examples reveal fragility
Christian Szegedy and colleagues showed that small image perturbations could induce neural-network classification errors. The finding exposed how good accuracy on ordinary tests does not guarantee robustness to inputs designed to cause failures.
2013
DQN combines deep networks and reinforcement
A DeepMind team presented an agent learning to play Atari games from pixels and rewards. The 2013 publication, expanded in Nature in 2015, connected learned perception with sequential decisions.
2013
Word2vec and word representations
Tomas Mikolov and colleagues introduced efficient methods for learning word-vector representations from large corpora. These representations helped transfer linguistic regularities across tasks, without by themselves amounting to understanding.
2014
Generative adversarial networks
Ian Goodfellow and colleagues introduce GANs, a learning framework that advances synthetic generation of images and other data. Training pits a generator against a discriminator; sample quality does not guarantee factual accuracy.
2015
ResNet makes deeper networks trainable
Kaiming He and colleagues introduced residual learning to facilitate training very deep networks. Residual connections reduced optimization difficulties and became components of many later architectures.
2016
AlphaGo defeats Lee Sedol
AlphaGo won the Go match against Lee Sedol in March 2016. It combined deep networks, reinforcement learning and tree search; the achievement concerns that game, not a general ability to solve any problem.
2016
A reward can divert the objective
OpenAI documented a CoastRunners agent repeating maneuvers to earn points instead of finishing the race. The experiment illustrates exploitation of a misspecified reward: optimizing a measure does not guarantee fulfilling human intent.
2017
Transformer
“Attention Is All You Need” introduces the Transformer architecture, based on self-attention and central to modern language models. It enabled greater parallelization of sequence processing during training and spread to tasks beyond translation.
2018
BERT
Google releases BERT, generalizing bidirectional pretraining of language representations across many tasks. Subsequent fine-tuning allows model reuse in text understanding without training each system from scratch.
2018
Gender Shades audits unequal error rates
Joy Buolamwini and Timnit Gebru found substantial differences across gender and skin-tone groups in commercial classifiers. The study evaluated binary gender classification from faces, not every form of facial recognition, and made intersectional auditing a central evaluation concern.
2018–… (as of 2026-09-20)
Joy Buolamwini
Joy Buolamwini investigated unequal performance in commercial facial-analysis systems and coauthored Gender Shades. Her work links technical evaluation, affected people’s experiences and demands for accountability in AI systems.
2018–… (as of 2026-09-20)
Timnit Gebru
Timnit Gebru coauthored Gender Shades and Stochastic Parrots and founded the DAIR institute. Her work connects data analysis, biases, research conditions and participation by communities affected by AI.
2020
AlphaFold 2 excels at CASP14
AlphaFold 2 achieved outstanding results in the 2020 CASP14 protein-structure prediction assessment. The 2021 scientific publication explained the method; the advance does not solve all protein dynamics, interactions or functions.
2020
Diffusion models for image generation
Jonathan Ho, Ajay Jain and Pieter Abbeel presented probabilistic diffusion models with strong image-synthesis results. Learning to reverse a noise process provided a generative approach later extended to text-guided systems.
2020
GPT-3
GPT-3 demonstrates large-scale in-context learning and accelerates interest in foundation language models. The paper evaluated tasks through instructions and examples in context, with persistent reliability and bias limitations.
2020
Masakhane documents translation for African languages
The Masakhane community published participatory research results on machine translation for African languages. It showed how researchers and speakers of digitally under-resourced languages can shape data, methods and evaluation, rather than only receive externally developed models.
2021
GitHub Copilot
GitHub announces Copilot as a coding assistant, foreshadowing everyday generative AI inside the development environment. The technical preview integrated code suggestions into programming work, subject to human review.
2021
DALL·E
OpenAI introduces DALL·E, a system that generates images from text and demonstrates applications of multimodal models. The presentation demonstrated combining concepts through written prompts; it was a research announcement, not unrestricted public availability.
2021
Stochastic Parrots questions scaling without evaluation
Emily Bender, Timnit Gebru, Angelina McMillan-Major and Shmargaret Shmitchell examined risks of large language models: environmental costs, data biases and plausible text without grounded meaning. The paper proposed critical criteria for deciding what to build and how to document it; its arguments are an intervention in the debate, not proof that every use lacks value.