The history of artificial intelligence How intelligence became a machine problem.

Wiener treated organisms and machines as systems that sense outcomes, communicate information, and correct their behavior through feedback.

Cybernetics offered an embodied account of intelligent behavior that developed alongside, and sometimes in tension with, symbolic AI.

Shannon described how a computer could explore a game tree, stop before exhaustive search became impossible, and evaluate the positions it reached.

Search plus a hand-built evaluation function became one of AI's most durable engineering patterns, from early chess programs to Deep Blue.

Turing proposed the imitation game as a way to discuss thinking machines through observable conversation, then examined objections, learning, and the education of a child machine.

The paper shifted debate from an elusive definition of thought toward behavior that people could inspect and challenge.

The Dartmouth proposal grouped language, abstraction, problem solving, neural networks, and self-improvement under the name artificial intelligence.

It gave a loose collection of research programs a shared identity, an ambitious agenda, and a place to argue about what intelligence required.

Logic Theorist represented proofs as a search among symbolic expressions and used heuristics to avoid exploring every possible derivation.

It showed a running computer program performing a task associated with mathematical reasoning and made heuristic search central to early AI.

LISP represented programs and symbolic expressions with the same list structures, supporting recursion, dynamic experimentation, and programs that manipulate other programs.

Its fit with symbolic computation made LISP the main working language of many AI laboratories for decades.

McCarthy proposed a system that would accept declarative statements about the world, draw consequences from them, and improve its behavior when given new facts.

The paper placed explicit knowledge representation and common-sense reasoning near the center of symbolic AI.

General Problem Solver compared a current state with a goal, selected an operator that might reduce the difference, and created subgoals when that operator could not yet be used.

Means-ends analysis became a durable account of planning and a serious attempt to separate a problem-solving strategy from any one puzzle.

A Unimate arm entered service at a General Motors plant, replaying stored motions to handle hot die-cast parts in a dangerous industrial setting.

It moved programmable machines from laboratory discussion into repeatable physical work and helped establish industrial robotics as a field and business.

DENDRAL generated and tested candidate molecular structures by combining chemical constraints, mass-spectrometry evidence, and rules supplied by specialists.

The project showed that deep, carefully encoded domain knowledge could outperform a shallow claim to generality on a real scientific task.

Robinson joined resolution with a practical unification procedure, allowing a program to prove first-order statements by deriving a contradiction from their negation.

The method gave automated reasoning a compact logical foundation and influenced theorem provers, deductive databases, and logic programming.

ELIZA transformed typed input with scripts and pattern-matching rules. Its DOCTOR script reflected statements back in the style of a nondirective therapist.

The program became an early lesson in human-computer interaction: a sparse conversational mechanism could invite users to attribute far more understanding than it possessed.

The ALPAC committee found no immediate prospect that general-purpose machine translation would compete with human translation and called for stronger evaluation and basic research.

Its recommendations sharply changed US support for machine translation and made evaluation, cost, and real demand unavoidable parts of the technical story.

Shakey combined camera input, a model of its surroundings, route finding, logical planning, and low-level actions in one mobile robot system.

The project forced separate AI techniques to work together in a physical environment and produced influential methods in search, planning, and computer vision.

A* ranked possible paths by combining the cost already paid with a heuristic estimate of the cost still required to reach the goal.

The algorithm exposed a precise tradeoff between problem knowledge and search effort, and it became foundational in planning, robotics, routing, and games.

SHRDLU accepted instructions and questions about a simulated blocks world while sharing one representation across parsing, dialogue, deduction, planning, and action.

It showed how coherent language behavior could emerge when every object, relation, and possible action belonged to a small, fully specified world.

STRIPS represented actions through conditions and effects, then searched for a sequence of actions that transformed an initial world state into a goal state.

Its compact planning language shaped decades of work on automated planning and remains visible in later planning formalisms.

MYCIN used backward-chained rules to recommend antimicrobial treatments, attached certainty factors to uncertain conclusions, and could explain parts of its reasoning.

Its architecture influenced expert-system shells, knowledge acquisition, rule-based explanation, and the design of later decision-support systems.

The Marseille team built a language in which facts and rules could be executed through unification and proof search, drawing on theory developed through collaboration with Kowalski.

Prolog became a major language for symbolic AI, natural-language processing, theorem proving, expert systems, and computer-science education.

Lighthill argued that many general AI claims had not survived combinatorial growth and the complexity of real environments. The review contributed to a harsher funding climate in Britain.

The report made scaling, evaluation, and the gap between a controlled demonstration and a dependable system central policy questions.

Minsky proposed structured frames whose slots could hold expected participants, default values, procedures, and links to related situations.

The proposal influenced knowledge-representation languages, schema theories, semantic networks, object-centered models, and later ontology work.

R1, later called XCON, configured VAX computer orders by applying production rules elicited from Digital Equipment Corporation specialists.

Its practical value gave expert systems a persuasive commercial success and helped drive corporate investment in knowledge engineering.

Japan's Fifth Generation Computer Systems project pursued parallel machines built around logic programming and knowledge processing rather than conventional sequential computing.

The effort turned AI into industrial policy and prompted strategic computing programs in Britain, the United States, and Europe.

Cyc began as a long-term effort to hand-encode the ordinary facts, categories, exceptions, and contexts that specialized programs usually leave unstated.

The project made the scale of common-sense knowledge visible and turned brittleness and knowledge acquisition into explicit engineering problems.

Brooks organized robot control as interacting layers of behavior. Higher layers could suppress or redirect lower-level outputs without relying on one complete symbolic world model.

The architecture helped establish behavior-based robotics and renewed attention to intelligence grounded in real-time perception and action.

Brittle rule bases, high maintenance costs, cheaper general-purpose hardware, unmet expectations, and changing government priorities weakened the commercial AI market.

The correction pushed work toward commodity computers, probabilistic reasoning, statistical methods, and narrower systems whose value could be measured more directly.

Pearl developed a framework in which directed graphs represented probabilistic dependencies and local computations updated beliefs as evidence changed.

Bayesian networks gave AI a principled language for uncertain diagnosis, prediction, evidence combination, and conditional independence.

Tesauro trained a neural network to evaluate backgammon positions by playing hundreds of thousands of games against itself and updating predictions through temporal differences.

The program became an early, persuasive example of reinforcement learning, neural networks, and self-play producing high-level strategic play.

Wooldridge and Jennings organized intelligent-agent research around theories, architectures, and programming languages, with autonomy, social ability, reactivity, and initiative as recurring properties.

The agent view treated an AI program as an actor situated in an environment, pursuing goals and sometimes coordinating with other actors rather than only returning an answer.

Deep Blue won its six-game rematch against Garry Kasparov by combining parallel search, a detailed evaluation function, specialized chess hardware, databases, and expert preparation.

The result showed how far a carefully engineered special-purpose system could go in a domain long treated as a public test of intelligence.

Monte Carlo Localization represented a robot's uncertain position with many sampled hypotheses and updated them as movement and sensor readings arrived.

It helped make probabilistic reasoning practical for mobile robots whose sensors were noisy and whose location could not be known exactly.

No vehicle finished DARPA's 2004 desert course. Eighteen months later, Stanford's Stanley won a 132-mile autonomous race and four other vehicles also completed it.

The challenge forced perception, localization, planning, learning, and control to operate as one dependable real-time system.

ImageNet organized millions of human-labeled images through WordNet concepts, giving computer vision a dataset and later a challenge at unprecedented scale.

It made data construction, annotation labor, shared categories, and public measurement central infrastructure for progress in visual AI.

Watson defeated two leading Jeopardy champions by analyzing each clue, generating candidate answers, gathering evidence, estimating confidence, and responding under time pressure.

The project demonstrated how many specialized language, retrieval, reasoning, and systems techniques could be assembled into one measurable question-answering machine.

The SuperVision entry won the 2012 ImageNet classification challenge by a wide margin with a deep convolutional network trained on graphics processors.

The visible benchmark result persuaded much of computer vision that learned representations, large labeled datasets, and accelerator hardware could beat hand-engineered feature pipelines.

DQN combined convolutional perception, Q-learning, and experience replay to learn game control directly from screen pixels and score changes.

The result joined learned perception with sequential action and helped restart broad interest in deep reinforcement learning.

Generative adversarial networks trained a generator against a discriminator in a coupled game, so each model changed the learning problem faced by the other.

GANs opened a major route to learned image synthesis and made generative modeling a central branch of modern AI research.

AlphaGo won four of five games against Lee Sedol by combining policy and value networks, reinforcement learning, and Monte Carlo tree search.

The system showed how learned judgment and explicit search could work together in a decision space long considered beyond brute-force game programs.

Libratus finished a 120,000-hand heads-up no-limit Texas Hold'em contest ahead of four professional players by computing strategies, solving subgames during play, and repairing weaknesses.

Poker forced an AI system to reason about hidden information, strategic deception, and an opponent whose private state could not be observed.

The Transformer processed sequences through attention and feed-forward layers without recurrent or convolutional sequence layers, first demonstrating the design on machine translation.

Its parallel training and flexible context mixing made much larger language models practical and later supported vision, audio, multimodal, and scientific systems.

BERT pretrained a bidirectional Transformer encoder on unlabeled text, then adapted that representation to many language tasks with small task-specific output layers.

Pretrained language representations became a standard starting point for search, classification, extraction, question answering, and other applied AI systems.

OpenAI announced GPT-2 but initially released only a smaller model, then published progressively larger weights while studying misuse, replication, detection, and stakeholder response.

Model access, misuse evaluation, documentation, and release strategy entered mainstream AI research as questions in their own right.

GPT-3 used a 175-billion-parameter autoregressive model to attempt translation, question answering, classification, writing, and other tasks from instructions or examples placed in its context.

Prompting became a general way to operate one pretrained system across many tasks without training a new set of weights for each one.

At the CASP14 blind assessment, AlphaFold2 produced highly accurate protein-structure predictions across difficult targets, followed by a detailed peer-reviewed system account.

The result established an AI system as a serious scientific tool and soon made large numbers of predicted structures available to researchers.

CLIP learned a shared representation from 400 million image-text pairs and used natural-language labels to transfer across many visual recognition tasks.

Language became a flexible interface for visual classification, and the shared representation became useful inside later multimodal and image-generation systems.

Latent diffusion moved the diffusion process into a compressed learned representation and used cross-attention to condition generation on text and other inputs.

The approach reduced the computational burden of high-resolution synthesis and helped make text-guided image generation broadly deployable.

ReAct prompted language models to alternate between written reasoning traces and actions such as consulting a knowledge source or interacting with a task environment.

It supplied a clear pattern for tool-using language-model agents and tied claims about reasoning to actions that changed or inspected an external state.

OpenAI released ChatGPT as a free research preview with a dialogue interface, instruction-following behavior, and a feedback channel for strengths, failures, and unsafe outputs.

Large language models moved from papers and developer tools into a direct, iterative interface that millions of people could test in ordinary work.

Regulation (EU) 2024/1689 created a risk-based legal framework covering prohibited practices, high-risk systems, transparency, governance, enforcement, and general-purpose AI models.

AI development and deployment became tied to a broad statutory regime whose requirements reach across providers, deployers, importers, and public authorities.

OpenAI released o1-preview, a model trained with reinforcement learning to spend additional computation on multi-step work before returning an answer.

Capability was presented not only as a consequence of larger pretraining, but also as a function of post-training and computation spent while solving a problem.

DeepSeek published the R1 technical report, model repository, R1 and R1-Zero weights, and several distilled models focused on mathematics, code, and multi-step reasoning.

A competitive reasoning-model family became downloadable and reusable beyond a small group of proprietary providers, drawing attention to reinforcement learning and distillation.