AI for EverybodyLearn the intelligence era with better judgment.

Your progress: 0 complete

Glossary My Map AccountQuick answers Choose my route History of AI Sources About this platform

You are here

Current learning moveOrient

See what changed, what did not, and why each milestone matters.

Do nowStart with the foundations

Your guide

History of AI

Your learning move

OrientClarify a term, source, person, or event.

Why this page exists

Timeline prevents current AI from appearing as an event without history. Every milestone separates change, continuity, significance, and debate.

What you will leave with

Twenty-one sourced milestones show what changed, what did not, and how people, concepts, institutions, and governance connect.

Before you begin

You can enter at any era. Begin with foundations if you want to follow the field’s development.

Do this now

Choose one era, then open a milestone to inspect context, debate, and sources.

Do now on this pageStart with the foundationsUnderstand the field’s formation before current generative systems.Another valid routeExplore people and institutionsFollow contributions and disagreements through contextual profiles.

Account

Connected history

AI did not begin with chatbots

A critical timeline of what changed, what did not, and the people, concepts, and sources connected to each milestone.

Five eras, relationships that continueMilestones connect to the concepts, profiles, and sources that let you examine them.

Foundations and symbolic systems
Critique, revival, and statistical learning
Data, deep learning, and transformers
Scaling, governance, and public generative AI
Agentic systems and operational governance

Search the timeline21 / 21

Historical era

Foundations and symbolic systems

From early computational ideas to the organization of an AI research field.

1943–1948

Foundations · foundations

Electronic computing, neuron models and cybernetics converge

Between 1943 and 1948, electronic computation, simplified mathematical neurons and feedback-oriented cybernetics created several foundations later absorbed into AI.

Examine this milestone

What changed

Large electronic computers made previously impractical calculations feasible; formal neuron models connected logic and biological inspiration; cybernetics supplied a language of control, communication and feedback.

What did not change

These strands were not one unified AI field, did not yield general intelligence, and relied on large teams, wartime institutions and ideas extending well before the named publications.

Why it matters

The period explains why AI has always drawn from computing, mathematics, neuroscience, engineering and social questions rather than one invention.

Debate or caveat

Calling this the “birth of AI” erases parallel work and the women and technicians who programmed early machines. The 1943 neuron model is an abstraction, not evidence that brains and computers work identically.

Algorithm Neural network

A Logical Calculus of the Ideas Immanent in Nervous ActivityThe Bulletin of Mathematical Biophysics / Springer Nature A Short History of the Second American RevolutionUniversity of Pennsylvania Cybernetics: Or Control and Communication in the Animal and the MachineMIT Press

Continue on AI for Everybody

1950-10-01

1950 · foundations

Turing reframes a philosophical question as an observable game

Alan Turing’s paper replaced the poorly specified question “Can machines think?” with an imitation-game setup and examined learning machines, objections and limits.

Examine this milestone

What changed

It gave researchers and the public a concrete behavioral thought experiment for discussing machine intelligence and treated learning as a serious possibility.

What did not change

The paper did not define intelligence once and for all, prove that fluent imitation equals understanding, or establish today’s popular pass/fail “Turing test” mythology.

Why it matters

Reading the original reveals a more qualified argument than the slogan and introduces questions about evaluation that remain unresolved.

Debate or caveat

Later accounts often flatten several versions of the game into a single benchmark. Behavioral indistinguishability measures an interaction under chosen conditions; it does not reveal consciousness or guarantee reliability.

Alan Turing Artificial intelligence Evaluation

Computing Machinery and IntelligenceMind / Oxford University Press

Continue on AI for Everybody

1956

1956 · field formation

A workshop helps name and organize an AI research field

The Dartmouth Summer Research Project brought researchers together around the proposal that aspects of learning and intelligence might be described precisely enough for machines to simulate.

Examine this milestone

What changed

The label “artificial intelligence” and a shared research agenda helped form institutions, funding programs and research communities.

What did not change

The workshop did not create every underlying idea, establish consensus on intelligence, or represent all relevant disciplines and communities.

Why it matters

Naming a field changes what can be funded, taught and debated; it also shapes which histories become visible.

Debate or caveat

Dartmouth is a field-formation milestone, not a lone moment of invention. The term appeared in the 1955 proposal, while the workshop met in 1956, and many proposed problems remained open.

John McCarthy Marvin Minsky Artificial intelligence Algorithm

The Early Years (1940–1959)Dartmouth College Artificial Intelligence at DartmouthDartmouth College

Continue on AI for Everybody

1958

1958 · field formation

The perceptron makes learning from examples experimentally visible

Frank Rosenblatt’s perceptron work demonstrated a trainable classifier whose weights could change after errors, giving early neural-network research a concrete algorithm and machine project.

Examine this milestone

What changed

A system could adjust from labeled examples rather than depend entirely on hand-written decision rules.

What did not change

A single-layer perceptron could solve only limited classes of problems; it did not see or understand the world as headlines suggested.

Why it matters

The perceptron introduced a durable idea—learning parameters from examples—and an equally durable lesson about separating demonstrations from forecasts.

Debate or caveat

Contemporary publicity credited a machine with sweeping future abilities. Later criticism identified real limits, but the common story that one book alone “killed neural networks” also oversimplifies funding and research shifts.

Neural network Training Classification

The Perceptron: A Probabilistic Model for Information Storage and Organization in the BrainPsychological Review / American Psychological Association Professor’s perceptron paved the way for AI — 60 years too soonCornell University

Continue on AI for Everybody

1965–1976

1965–1976 · symbolic systems

Symbolic programs bring expert knowledge into narrow domains

Projects such as DENDRAL and MYCIN encoded domain knowledge, rules and inference procedures to support scientific or medical reasoning in bounded settings.

Examine this milestone

What changed

AI research showed that carefully structured knowledge could deliver useful, inspectable behavior in specialized tasks and encouraged an expert-systems industry.

What did not change

These systems did not acquire broad common sense, learn freely from raw experience, or remove the need for expert judgment and careful validation.

Why it matters

They establish a distinct lineage from statistical machine learning and show why “AI” has referred to different technical approaches across time.

Debate or caveat

Descriptions of “expert-level” performance depend on task, evaluation and deployment context. Medical prototypes, for example, should not be retroactively described as routine autonomous clinical systems.

Algorithm Explainability Model

Edward A. Feigenbaum Collection — DENDRAL Original 1965Stanford University What is an Expert System?Stanford Institute for Human-Centered AI

Continue on AI for Everybody

Historical era

Critique, revival, and statistical learning

Limits become visible while new methods recenter learning and generalization.

1973–1980

1973–1980 · contraction and critique

Expectations, evaluations and funding contract unevenly

After ambitious predictions met technical limitations and skeptical reviews, some AI funding and institutional support contracted in the UK and United States.

Examine this milestone

What changed

The period exposed gaps between demonstrations, generalization claims and deliverable systems; some research directions lost resources or changed labels.

What did not change

AI research did not stop everywhere, and one report did not single-handedly produce a universal winter. Work continued in applications, theory, robotics and other regions.

Why it matters

The episode is an early warning that research trajectories are shaped by claims, evaluation, politics and funding—not technical capability alone.

Debate or caveat

“AI winter” is a retrospective umbrella with disputed boundaries and causes. It can conceal national differences and continuing work, so the range is intentionally approximate.

Artificial intelligence Evaluation

AI in the UK: ready, willing and able? — paragraphs 23–24UK House of Lords Select Committee on Artificial Intelligence

Continue on AI for Everybody

1986-10-09

1986 · neural revival

Backpropagation makes multilayer representation learning practical again

Rumelhart, Hinton and Williams demonstrated an effective procedure for adjusting weights through multiple layers so hidden units could learn useful internal features.

Examine this milestone

What changed

Researchers gained a clearer, reusable way to train multilayer neural networks and renewed experimental work on learned representations.

What did not change

Backpropagation did not eliminate optimization difficulties, data needs or computational limits, and related differentiation ideas had earlier histories.

Why it matters

Modern deep learning depends heavily on gradient-based training, so this paper is a key bridge—but the bridge involved algorithms, data, hardware and many research groups.

Debate or caveat

Textbook histories sometimes present a single 1986 discovery. The authors popularized a powerful formulation and evidence; earlier work and parallel developments still matter.

Geoffrey Hinton Neural network Training Parameters Deep learning

Learning representations by back-propagating errorsNature

Continue on AI for Everybody

1987–1993

1987–1993 · contraction and critique

The expert-systems market contracts and another “winter” is named

Costs of maintaining brittle knowledge bases, specialized hardware shifts and unmet commercial expectations contributed to a downturn in expert-system businesses and AI investment.

Examine this milestone

What changed

The contraction redirected attention toward methods that learned from data, probabilistic reasoning and less brittle software practices.

What did not change

Expert systems did not vanish, every AI market did not collapse simultaneously, and useful rule-based systems continued under other names.

Why it matters

This second contraction helps learners recognize recurring cycles in which narrow success is generalized into a market promise and later judged against that promise.

Debate or caveat

The boundaries differ by country, company and research area. A boom-and-bust analogy can illuminate incentives, but it should not be used to predict that current systems must repeat an identical cycle.

Artificial intelligence Generalization Evaluation

Implications for AI Research: Applying Lessons from the Expert Systems Boom and Bust to the Current Large-Language Model BoomAAAI Edward A. Feigenbaum Collection commentaryStanford University

Continue on AI for Everybody

1990s

1990s · statistical learning

Statistical learning moves prediction and generalization to the center

Across the 1990s, probabilistic models, support-vector machines, decision trees, ensembles and benchmark-based evaluation made data-driven prediction a larger part of AI research.

Examine this milestone

What changed

Researchers increasingly compared methods on held-out data and formalized the problem of performing well beyond the training examples.

What did not change

Statistical success did not make data neutral, guarantee causal understanding or remove the need for domain knowledge and feature design.

Why it matters

Many current ideas—training/test splits, generalization, uncertainty and empirical comparison—became central practice in this period.

Debate or caveat

There is no single date when “AI became statistical.” Different communities had used probability and pattern recognition much earlier, and symbolic, neural and statistical work overlapped.

Machine learning Generalization Evaluation Dataset

Support-vector networksMachine Learning / Springer Nature

Continue on AI for Everybody

Historical era

Data, deep learning, and transformers

Datasets, compute, and architectures change the scale of AI systems.

2009

2009 · data and compute scaling

ImageNet turns large labeled datasets into shared vision infrastructure

ImageNet organized millions of web images into a large hierarchy and later supported a common large-scale recognition challenge.

Examine this milestone

What changed

Computer-vision teams could train and compare systems on a shared dataset and benchmark, making data scale an explicit research resource.

What did not change

The dataset did not represent the visual world neutrally or completely; labels, collection practices and benchmark targets carried social and technical assumptions.

Why it matters

ImageNet shows that breakthroughs can depend as much on curated data and evaluation infrastructure as on a new model.

Debate or caveat

Crowd labeling and web collection created privacy, representation and category problems later investigated by the project and critics. A benchmark win does not equal robust vision in real settings.

Fei-Fei Li Dataset Label Classification Benchmark

ImageNet: A Large-Scale Hierarchical Image DatabaseIEEE Conference on Computer Vision and Pattern Recognition About ImageNetImageNet project / Stanford Vision Lab

Continue on AI for Everybody

2012

2012 · deep learning

A deep convolutional network changes the computer-vision benchmark

Krizhevsky, Sutskever and Hinton’s ImageNet system combined convolutional networks, large labeled data, GPUs and training techniques to improve image-classification results sharply.

Examine this milestone

What changed

Deep neural networks became a compelling default direction for computer vision and attracted wider investment in data, accelerators and representation learning.

What did not change

The result did not solve visual understanding, remove dataset bias or prove that one architecture would dominate every task.

Why it matters

It is a clear example of a systems breakthrough: algorithmic ideas became transformative when matched with data, compute, implementation and evaluation.

Debate or caveat

“Deep learning began in 2012” is false. The methods had long lineages; 2012 changed their empirical and institutional position. ImageNet accuracy also measured a narrow classification task.

Geoffrey Hinton Deep learning Neural network Accelerator Benchmark

ImageNet Classification with Deep Convolutional Neural NetworksNeurIPS

Continue on AI for Everybody

2013–2014

2013–2014 · deep learning

Learned vectors and sequence-to-sequence models reshape language processing

Efficient word embeddings made linguistic patterns usable as vectors, while sequence-to-sequence neural models learned mappings between variable-length text sequences.

Examine this milestone

What changed

Language systems could learn more of their representations and transformations from data instead of relying only on hand-built features or phrase pipelines.

What did not change

Word vectors lost context and word order; recurrent sequence models struggled with long dependencies, data coverage and opaque errors.

Why it matters

These developments form a direct technical bridge from statistical language processing to transformers and modern generative language systems.

Debate or caveat

Vector analogies became famous but can exaggerate semantic understanding and reproduce corpus stereotypes. Translation scores do not imply faithful meaning in every language or domain.

Yoshua Bengio Embedding Token Model Training

Distributed Representations of Words and Phrases and their CompositionalityNeurIPS Sequence to Sequence Learning with Neural NetworksNeurIPS

Continue on AI for Everybody

2017

2017 · foundation model architecture

The transformer replaces recurrence with attention for sequence modeling

The transformer architecture used self-attention and feed-forward layers instead of recurrent or convolutional sequence processing in its core translation design.

Examine this milestone

What changed

Training could be parallelized more effectively and relationships across positions could be modeled in a flexible way, enabling later scaling across language and other modalities.

What did not change

The original paper addressed machine translation, not every modern capability; attention did not create factuality, reasoning guarantees or unlimited context.

Why it matters

Most current large language models descend from transformer ideas, making architecture—not only data size or interface—a key part of the story.

Debate or caveat

Eight authors are listed, and the paper itself builds on earlier attention, embeddings and encoder-decoder work. “Attention is all you need” is a title, not proof that other components or constraints disappeared.

Embedding Large language model Context window

Attention Is All You NeedNeurIPS

Continue on AI for Everybody

Historical era

Scaling, governance, and public generative AI

General-purpose models advance while the first international frameworks emerge.

2019-05-22

2019 · governance emerges

The OECD adopts the first intergovernmental AI standard

OECD members adopted principles connecting trustworthy AI with human rights, democratic values, transparency, robustness, accountability and inclusive benefit.

Examine this milestone

What changed

AI governance gained a shared intergovernmental vocabulary that later influenced national and international policy work.

What did not change

The principles were recommendations, not a global law or technical assurance scheme, and adoption did not prove implementation.

Why it matters

The milestone shows governance developing alongside technical scaling rather than appearing only after generative AI reached the public.

Debate or caveat

The OECD updated the principles in 2024 for general-purpose and generative systems. Their reach is broad but not universal, and interpretation depends on institutions, measurement and enforcement.

AI literacy Fairness Transparency Human oversight

OECD AI PrinciplesOECD Scoping the OECD AI PrinciplesOECD

Continue on AI for Everybody

2020-05-28

2020 · foundation model scaling

A much larger language model shows broad prompt-based task adaptation

GPT-3 demonstrated that a scaled autoregressive language model could perform many tasks from instructions or examples in its input without task-specific weight updates.

Examine this milestone

What changed

Prompting and in-context examples became a practical interaction pattern, and model scale became a major research and investment axis.

What did not change

The system still failed on tasks, inherited web-data problems, required substantial compute and did not learn new weights from a user’s prompt.

Why it matters

The paper helps explain why one general model could underpin many applications and why interfaces later centered on natural-language instructions.

Debate or caveat

This is one influential scaling milestone, not proof that larger always means better or that OpenAI alone created large language models. Parameter count is not a complete capability or social-value measure.

Large language model Foundation model Prompt Compute

Language models are few-shot learnersOpenAI Language Models are Few-Shot LearnersarXiv

Continue on AI for Everybody

2021-11-23

2021 · governance emerges

UNESCO adopts a global recommendation on AI ethics

UNESCO’s General Conference adopted a normative recommendation covering human rights, inclusion, environmental effects, data governance, education, culture and other policy areas.

Examine this milestone

What changed

AI governance broadened beyond safety and economic growth to a cross-sector agenda for member states, including practical policy actions.

What did not change

A UNESCO recommendation is not directly enforceable national law, and shared adoption does not ensure equal capacity or consistent implementation.

Why it matters

The milestone brings global, cultural and environmental dimensions into the timeline and provides a reference beyond technology-producing countries.

Debate or caveat

Its language is intentionally broad and must be translated into laws, budgets, institutions and evaluation. “Ethical AI” cannot be verified by citing principles alone.

Fairness Privacy Transparency Human oversight

Recommendation on the Ethics of Artificial IntelligenceUNESCO

Continue on AI for Everybody

2022-11-30

2022 · public generative ai

Conversational generative AI reaches a mass public interface

ChatGPT’s research preview put a capable instruction-following language model behind a simple dialogue interface that people could try without building software.

Examine this milestone

What changed

Follow-up conversation and plain-language prompting made generative AI accessible to far more students, workers, creators and organizations, accelerating adoption and scrutiny.

What did not change

The interface did not make outputs verified, give the model human understanding or originate generative modeling, instruction tuning or chat systems by itself.

Why it matters

The event explains why AI suddenly felt new to many people even though its model, data and research lineages were older.

Debate or caveat

A product launch is only one marker of a broader shift involving multiple laboratories, open communities and image-generation systems. Access was also uneven across language, geography, disability and connectivity.

Generative AI Prompt Hallucination Large language model

Introducing ChatGPTOpenAI

Continue on AI for Everybody

2023

2023 · public generative ai

General-purpose model families combine text with other modalities

Systems announced during 2023 connected language-model interfaces with image input and, across model families, wider combinations of text, image, audio or video.

Examine this milestone

What changed

A single model interface could interpret or generate across more media types, expanding accessibility possibilities and application scope.

What did not change

Multimodal did not mean human-like perception, equal skill in every modality or reliable grounding; product access often lagged research announcements.

Why it matters

The shift changes what counts as an AI input and output and makes system-layer explanations more important than the label of one model.

Debate or caveat

This was a transition, not a single-day invention. Capability claims came largely from producing organizations, and evaluation remained uneven across languages, disabilities, cultures and real-world conditions.

Multimodal model Foundation model Large language model

GPT-4OpenAI Gemini: A Family of Highly Capable Multimodal ModelsGoogle DeepMind

Continue on AI for Everybody

Historical era

Agentic systems and operational governance

Tools, multimodality, risk frameworks, and binding rules move AI closer to real-world effects.

2023-01-26

2023 · operational governance

NIST publishes a voluntary framework for operational AI risk management

AI RMF 1.0 organized risk work around governing, mapping, measuring and managing AI systems across their lifecycle.

Examine this milestone

What changed

Organizations gained a common, use-case-agnostic vocabulary for connecting trustworthiness goals with roles, evidence and actions.

What did not change

The framework was voluntary, not a certification, law or proof that a system was safe; implementation quality remained decisive.

Why it matters

It marks a move from high-level principles toward repeatable organizational risk processes that product, procurement and oversight teams can use.

Debate or caveat

As of 14 July 2026, NIST says AI RMF 1.0 is being revised. The 2023 version must therefore be cited by version and used with current law, domain evidence and later profiles.

AI incident High-risk use Human oversight Evaluation

Artificial Intelligence Risk Management Framework (AI RMF 1.0)NIST AI Risk Management FrameworkNIST

Continue on AI for Everybody

2023–2024

2023–2024 · agentic systems

Language models begin using structured tools and agent loops

Research patterns such as ReAct, product function-calling interfaces and emerging connection protocols made it easier to combine a model with retrieval, code, APIs and multi-step action loops.

Examine this milestone

What changed

Developers could build systems that observe results, choose tools and act through software interfaces rather than only generate text.

What did not change

A model connected to tools did not become an autonomous, trustworthy worker; permissions, state, recovery, security and human review remained application responsibilities.

Why it matters

The period marks the practical separation between a model and an agentic system assembled around it.

Debate or caveat

“Agent” became a marketing label applied to very different systems. Tool access increases impact and attack surface; demonstrations do not establish long-horizon reliability, and internal reasoning traces are not guaranteed explanations.

Agent Tool use Function calling Permission

ReAct: Synergizing Reasoning and Acting in Language ModelsICLR / OpenReview Function calling and other API updatesOpenAI Introducing the Model Context ProtocolAnthropic

Continue on AI for Everybody

2024-08-01

2024 · binding governance

The EU AI Act enters into force with phased obligations

Regulation (EU) 2024/1689 entered into force on 1 August 2024, establishing a binding, risk-based legal framework with rules for prohibited practices, high-risk systems, transparency and general-purpose models.

Examine this milestone

What changed

AI governance moved beyond voluntary principles in a major jurisdiction, with duties assigned across providers, deployers and other actors.

What did not change

Most obligations did not apply at once, the law did not cover every AI harm, and compliance does not guarantee a beneficial or error-free system.

Why it matters

The Act affects product design, documentation, literacy and oversight far beyond Europe because many systems and organizations cross borders.

Debate or caveat

As of 14 July 2026, application remains phased and the official Commission timeline includes later dates and simplification work. Always check the legal text, actor role, system category and current implementing measures.

AI literacy High-risk use Transparency Human oversight

AI Act — regulatory framework and application timelineEuropean Commission Regulation (EU) 2024/1689EUR-Lex

Continue on AI for Everybody

Loading page content...

…