AI did not appear from nowhere. Russell and Norvig identify the contributing disciplines, and the questions each brought. PHILOSOPHY (from ~428 BC) Can formal rules be used to draw valid conclusions? How does the mind arise from a physical brain? Where does knowledge come from, and how does it lead to action? KEY IDEAS: Aristotle's SYLLOGISMS — the first formal system of reasoning; DUALISM vs MATERIALISM; EMPIRICISM (knowledge from sensory experience); the PHYSICAL SYMBOL SYSTEM HYPOTHESIS that a physical system manipulating symbols can exhibit intelligence. MATHEMATICS (from ~800) What are the formal rules to draw valid conclusions? What can be computed? How do we reason with uncertain information? KEY IDEAS: LOGIC (Boole, Frege); ALGORITHMS (al-Khwarizmi); COMPUTABILITY and the CHURCH-TURING thesis; GÖDEL's INCOMPLETENESS THEOREM (in any consistent formal system powerful enough for arithmetic there are true statements it cannot prove); INTRACTABILITY and NP-completeness — the distinction between what is computable in principle and what is computable in practice; PROBABILITY (Cardano, Bayes). ECONOMICS (from 1776) How should we make decisions to maximise payoff? How should we do so when others may not go along? How should we do so when the payoff is far in the future? KEY IDEAS: UTILITY THEORY; DECISION THEORY (= probability + utility); GAME THEORY (von Neumann and Morgenstern); OPERATIONS RESEARCH; MARKOV DECISION PROCESSES and SATISFICING (Simon: real decision-makers choose "good enough" rather than optimal, because optimising costs more than it returns). NEUROSCIENCE (from 1861) How do brains process information? KEY IDEAS: neurons as the units of processing; the enormous parallelism of biological computation; the observation that brains and computers do very different things well. PSYCHOLOGY (from 1879) How do humans and animals think and act? KEY IDEAS: BEHAVIOURISM (study only observable stimulus and response) giving way to COGNITIVE PSYCHOLOGY (the brain as an information-processing device); the idea that a knowledge-based agent can be studied as a computational process. COMPUTER ENGINEERING (from 1940) How can we build an efficient computer? Without the machine, none of the rest is testable. The exponential growth in available computation is a large part of why methods invented in the 1980s only became practical after 2010. CONTROL THEORY AND CYBERNETICS (from 1948) How can artefacts operate under their own control? KEY IDEAS: HOMEOSTASIS; feedback; the design of systems that minimise an objective function over time — which is the same mathematics as reinforcement learning. LINGUISTICS (from 1957) How does language relate to thought? KEY IDEAS: Chomsky's critique of behaviourism, showing it could not explain the creativity of language; formal GRAMMARS; COMPUTATIONAL LINGUISTICS.

THE GESTATION (1943–1955) 1943 McCULLOCH AND PITTS publish a model of the artificial neuron — the mathematical starting point of neural networks (ACtE0906 topic 1). 1949 HEBB proposes a learning rule for modifying connection strengths — Hebbian learning. 1950 TURING's "Computing Machinery and Intelligence" introduces the imitation game, machine learning, genetic algorithms and reinforcement learning, all in one paper. 1951 Minsky and Edmonds build SNARC, the first neural network computer. THE BIRTH (1956) THE DARTMOUTH WORKSHOP, organised by John McCarthy with Minsky, Shannon and Rochester. McCarthy COINED THE TERM "ARTIFICIAL INTELLIGENCE" here — the standard answer to "when did AI begin as a field?" EARLY ENTHUSIASM AND GREAT EXPECTATIONS (1952–1969) the LOGIC THEORIST (Newell and Simon) proving theorems from Principia Mathematica the GENERAL PROBLEM SOLVER (GPS) — the first program to imitate the structure of human problem solving 1958 McCarthy defines LISP, the dominant AI language for thirty years, and proposes the ADVICE TAKER — a program embodying general knowledge of the world 1959 Samuel's CHECKERS program learns to play better than its author, which is a genuine landmark: the program improved beyond the knowledge put into it 1965 Robinson's RESOLUTION method for logical inference 1965 Weizenbaum's ELIZA — a pattern-matching "therapist" whose users attributed understanding to it, which is a lasting lesson about how readily people over-attribute intelligence A DOSE OF REALITY, AND THE FIRST WINTER (1966–1974) THE THREE CAUSES OF FAILURE, and they are worth learning because they recur: 1. programs succeeded on small examples and FAILED TO SCALE. Early methods were "toy" — the search space grew exponentially and there was no way to control it. 2. the problems being attacked were INTRACTABLE, and the theory of NP-completeness (1971) explained why: it was not a matter of faster machines. 3. some basic structures were fundamentally LIMITED. Minsky and Papert's "Perceptrons" (1969) proved a single-layer perceptron cannot represent XOR (ACtE0906 topic 5), and neural-network funding collapsed for a decade — even though the multilayer solution was already conceivable and only lacked a training algorithm. 1973 the LIGHTHILL REPORT in Britain led to almost all AI funding being cut; US funding followed. THE FIRST AI WINTER. KNOWLEDGE-BASED SYSTEMS (1969–1979) The response: abandon general-purpose weak methods for DOMAIN-SPECIFIC KNOWLEDGE. DENDRAL inferring molecular structure from mass spectrometry — the first successful knowledge-intensive system MYCIN diagnosing blood infections, performing at specialist level, and introducing CERTAINTY FACTORS to handle uncertainty THE LESSON: knowledge beats cleverness. A weak inference engine with strong domain knowledge outperformed a strong engine with none. AI BECOMES AN INDUSTRY, AND THE SECOND WINTER (1980–1993) R1/XCON at Digital saved a reported tens of millions of dollars a year configuring computer orders Japan's FIFTH GENERATION project; expert-system companies proliferated; the industry grew from a few million to billions of dollars THEN THE COLLAPSE (from ~1987): expert systems proved EXPENSIVE TO MAINTAIN, BRITTLE outside their narrow domain, and unable to learn. Specialised LISP machines were undercut by ordinary workstations. THE SECOND AI WINTER. THE RETURN OF NEURAL NETWORKS (1986–) BACKPROPAGATION was reinvented and popularised (Rumelhart, Hinton, Williams), solving the multilayer training problem that had stalled the field since 1969 — seventeen years lost to a missing algorithm. AI ADOPTS THE SCIENTIFIC METHOD (1987–) The important shift, and arguably the reason there has been no third winter: the field moved to RIGOROUS EXPERIMENT, shared BENCHMARK DATASETS, statistical rather than ad-hoc methods, and building on existing theory (probability, statistics, control theory) rather than inventing from scratch. Claims became measurable. MILESTONES SINCE 1997 DEEP BLUE defeats Kasparov at chess — by search and hardware, not by learning 2011 IBM WATSON wins Jeopardy! 2012 ALEXNET wins ImageNet by a large margin, beginning the DEEP LEARNING era — enabled by three things arriving together: large labelled datasets, GPU computation, and algorithmic improvements 2016 ALPHAGO defeats Lee Sedol at Go, combining deep networks with Monte Carlo tree search 2017 the TRANSFORMER architecture 2020s large language models THE PATTERN ACROSS BOTH WINTERS, and the reason to know this history: BOTH FOLLOWED PERIODS OF CONFIDENT OVER-PROMISING ABOUT TIMELINES, and both collapses were triggered by the gap between demonstration and deployment rather than by any technique being disproved. The techniques mostly survived and returned when the missing piece — an algorithm, data, or compute — arrived.