History of artificial intelligence

• 1. Introduction

  ›

  The history of artificial intelligence begins not in a Silicon Valley lab, but in the myths and forges of ancient Greece, where Hephaestus, god of the forge, was said to have built golden robots to assist him in his work. Talos, a bronze giant, guarded the island of Crete. Pygmalion's Galatea breathed with the life her sculptor willed into her. For thousands of years, human beings have been dreaming of minds that could be built. The questions those ancient stories asked are the same ones that haunt the field today: What does it mean to think? Can matter be made to understand? And what happens when the thing we create surpasses us? This documentary traces the full arc of that obsession, from the logical machines of medieval Spain to the billion-dollar AI labs of the twenty-first century, through cycles of wild hope and bitter disappointment, to the present moment where the dream has never seemed closer or more dangerous.

• 2. Roots in Reason

  ›

  Ramon Llull, a Spanish philosopher who lived from 1232 to 1315, built several logical machines devoted to producing knowledge by purely mechanical means. Llull described them as entities that could combine basic and undeniable truths through simple logical operations, generating all possible knowledge without human intuition. His work reached Gottfried Leibniz, who redeveloped the ideas and, in the seventeenth century, pushed them further. Leibniz imagined a universal language of reasoning he called the characteristica universalis, which would reduce all philosophical dispute to calculation. "There would be no more need of disputation between two philosophers than between two accountants," he wrote. "For it would suffice to take their pencils in hand... and to say to each other: Let us calculate." Thomas Hobbes was equally blunt in Leviathan, declaring that "reason... is nothing but reckoning, that is adding and subtracting." These ideas accumulated over centuries. Boole's The Laws of Thought in 1854 and Frege's Begriffsschrift in 1879 defined modern symbolic mathematical logic. Russell and Whitehead extended that work in the Principia Mathematica in 1913. David Hilbert challenged mathematicians in the 1920s and 1930s to formalize all mathematical reasoning, and the response was decisive: Gödel's incompleteness proof, Turing's machine, and Church's Lambda calculus showed there were hard limits to what formal mathematics could achieve. Within those limits, however, the Church-Turing thesis implied that a mechanical device shuffling symbols as simple as 0 and 1 could imitate any conceivable process of mathematical reasoning. That insight would drive the engineers of the following decade.

• 3. Birth of the Field

  ›

  In 1950, Alan Turing published a landmark paper titled "Computing Machinery and Intelligence" in which he grappled with what it would even mean for a machine to think. His solution was practical: if a machine could carry on a conversation over a teleprinter indistinguishable from a human conversation, it was reasonable to call that machine a thinker. The same year, Walter Pitts and Warren McCulloch had already shown, back in 1943, that networks of idealized artificial neurons could perform simple logical functions. One of the students inspired by that work was a twenty-four-year-old graduate student named Marvin Minsky, who in 1951 built the first neural net machine, the SNARC, with Dean Edmonds. The field came together formally at a workshop held on the campus of Dartmouth College in 1956, organized by Minsky and John McCarthy with support from Claude Shannon and Nathan Rochester of IBM. The proposal for the conference stated the group intended to test the assertion that every aspect of learning or any other feature of intelligence could be described precisely enough for a machine to simulate it. At the workshop, Allen Newell and Herbert A. Simon, with help from J. C. Shaw, debuted the Logic Theorist, a program that would eventually prove 38 of the first 52 theorems in the Principia Mathematica and find new, more elegant proofs for some of them. Simon said they had "solved the venerable mind/body problem, explaining how a system composed of matter can have the properties of mind." McCarthy coined the term "artificial intelligence" at that gathering. The workshop was the moment AI got its name, its mission, its first major success, and its key players.

• 4. Early Optimism and Its Limits

  ›

  In 1958, Herbert Simon and Allen Newell predicted that within ten years a digital computer would be the world's chess champion and would discover and prove an important new mathematical theorem. In 1967, Marvin Minsky told an interviewer that within a generation the problem of creating artificial intelligence would be substantially solved. The government believed them: in June 1963, MIT received a $2.2 million grant from the newly formed Advanced Research Projects Agency, and DARPA continued to provide $3 million each year until the 1970s. Similar grants went to Newell and Simon's program at Carnegie Mellon University and to John McCarthy's Stanford AI Lab, founded in 1963. J. C. R. Licklider, then director of ARPA, believed his organization should "fund people, not projects" and gave researchers wide latitude. The results were genuinely astonishing. Joseph Weizenbaum's ELIZA, the first chatbot, could carry out conversations so realistic that users occasionally believed they were speaking with a human being. Daniel Bobrow's STUDENT solved high-school algebra word problems. Terry Winograd's SHRDLU communicated in ordinary English about a simulated blocks world and could plan and execute operations within it. But the programs were toys. Ross Quillian's natural language work demonstrated with a vocabulary of only twenty words because that was all that would fit in memory. In 1972, Richard Karp showed there are many problems that can only be solved in exponential time, which meant many of the "toy" solutions used by AI would never scale. Hans Moravec argued in 1976 that computers were still millions of times too weak to exhibit intelligence. Researchers had also discovered Moravec's paradox: AI was good at tasks humans find hard, like proving theorems, but failed at tasks children handle effortlessly, like recognizing a face or crossing a room.

• 5. First Winter

  ›

  In 1973, the Lighthill report on the state of AI research in the UK criticized the failure of AI to achieve its "grandiose objectives" and led to the dismantling of AI research in that country. The report specifically cited the combinatorial explosion problem. DARPA, meanwhile, was deeply disappointed with researchers working on the Speech Understanding Research program at Carnegie Mellon University and canceled an annual grant of $3 million. The pattern had begun earlier: in 1966, the Automatic Language Processing Advisory Committee report criticized machine translation efforts, and after spending $20 million, the National Research Council ended all support. Hans Moravec blamed the crisis on the unrealistic predictions of his colleagues. "Many researchers were caught up in a web of increasing exaggeration." These cuts hurt the major laboratories, MIT, Stanford, Carnegie Mellon, and Edinburgh, but the thousands of researchers outside those institutions were largely unaffected. Philosopher Hubert Dreyfus ridiculed the broken promises of the 1960s, arguing that human reasoning involved very little symbol processing and a great deal of embodied, unconscious know-how. MIT's Minsky dismissed Dreyfus and John Searle, saying they "misunderstand, and should be ignored." Dreyfus, who also taught at MIT, later said that AI researchers "dared not be seen having lunch with me." Joseph Weizenbaum, author of ELIZA, took a different view. When Kenneth Colby built a psychotherapy program based on ELIZA without crediting Weizenbaum, a feud erupted. Weizenbaum published Computer Power and Human Reason in 1976, arguing that the misuse of artificial intelligence has the potential to devalue human life. Meanwhile, in 1969, Minsky and Seymour Papert published Perceptrons, arguing that neural networks had severe limitations and that Rosenblatt's predictions had been grossly exaggerated. The effect was that virtually no research was funded in connectionism for ten years. Frank Rosenblatt, who had built the perceptron in 1958 and predicted machines might eventually learn, make decisions, and translate languages, died in a boating accident in 1971 and never saw his predictions vindicated.

• 6. The Expert Systems Boom

  ›

  Edward Feigenbaum had begun building expert systems as early as 1965 with Dendral, a program that identified chemical compounds from spectrometer readings. MYCIN, developed in 1972, diagnosed infectious blood diseases. These programs worked by restricting themselves to a small domain of specific knowledge, which neatly sidestepped the commonsense knowledge problem that had stalled broader AI research. In 1980, an expert system called R1 was completed at Carnegie Mellon University for the Digital Equipment Corporation, and by 1986 it was saving the company $40 million annually. Corporations around the world began deploying similar systems, and by 1985, they were spending over a billion dollars on AI. "Overall, the AI industry boomed from a few million dollars in 1980 to billions of dollars in 1988." In 1981, the Japanese Ministry of International Trade and Industry set aside $850 million for the Fifth Generation computer project, aiming to build machines that could carry on conversations, translate languages, interpret pictures, and reason like human beings. The UK launched the Alvey project at a cost of £350 million. DARPA tripled its investment in AI between 1984 and 1988. At the same time, a small number of researchers were quietly pursuing alternative approaches. In 1982, physicist John Hopfield proved that a form of neural network could learn and process information and converge after enough time under any fixed condition. Geoffrey Hinton proved a similar result about a device called a Boltzmann machine. In 1986, Hinton and David Rumelhart popularized backpropagation, a method for training neural networks. Hopfield and Hinton would eventually receive the 2024 Nobel prize for this work. In 1990, Yann LeCun at Bell Labs used convolutional neural networks to recognize handwritten digits, and the system was used widely in the 1990s to read zip codes and personal checks.

• 7. Second Winter and Quiet Progress

  ›

  The collapse came first in hardware. In 1987, desktop computers from Apple and IBM became more powerful than the expensive Lisp machines made by companies like Symbolics. An entire industry worth half a billion dollars was demolished overnight. Expert systems like R1 proved too expensive to maintain; they could not learn and were "brittle," making grotesque mistakes when given unusual inputs. By the end of 1993, over 300 AI companies had shut down, gone bankrupt, or been acquired. The term "AI winter" was coined by researchers who had survived the funding cuts of 1974, and it returned with force. By 1991, Japan's Fifth Generation Project had met none of its impressive 1981 goals. Some of them, like "carry on a casual conversation," would not be accomplished for another thirty years. Yet underneath the public failure, AI was spreading through the technology industry. Data mining, industrial robotics, logistics, speech recognition, banking software, medical diagnosis, and search engines all incorporated techniques originally developed by AI researchers. Nick Bostrom captured the dynamic: "A lot of cutting-edge AI has filtered into general applications, often without being called AI because once something becomes useful enough and common enough it's not labeled AI anymore." On the 11th of May 1997, Deep Blue became the first computer chess-playing system to beat a reigning world chess champion, Garry Kasparov. Deep Blue's computer was 10 million times faster than the Ferranti Mark 1 that Christopher Strachey had used to write a chess program in 1951. In 2005, a Stanford robot won the DARPA Grand Challenge by driving autonomously for 131 miles along an unrehearsed desert trail. These milestones were not the product of new paradigms, but of the relentless increase in computing power that Moore's law had predicted: speed and memory capacity doubling every two years.

• 8. Deep Learning and the Modern Era

  ›

  In 2012, a deep learning model called AlexNet, developed by Alex Krizhevsky working with Geoffrey Hinton at the University of Toronto, won the ImageNet Large Scale Visual Recognition Challenge with significantly fewer errors than the second-place winner. Over the next few years, dozens of other approaches to image recognition were abandoned in favor of deep learning. The path had been prepared by vast datasets: Fei-Fei Li's ImageNet, released in 2009, contained three million images captioned by volunteers using Amazon Mechanical Turk, and Google released word2vec in 2013 as an open-source resource, in which vector addition produced equivalences like China plus River equals Yangtze. By 2016, the market for AI-related products, hardware, and software reached more than $8 billion. The transformer architecture, introduced in 2017, became the foundation for generative AI applications, and for large language models like ChatGPT. Investment in AI boomed in the 2020s. DeepMind, founded in 2010 by Demis Hassabis, Shane Legg, and Mustafa Suleyman with early funding from Peter Thiel and later Elon Musk, was sold to Google in 2014 on the condition that it would not accept military contracts and would be overseen by an ethics board. In 2012, Geoffrey Hinton sold himself and all his students to Google at a Lake Tahoe AI conference for a price of $44 million. The alignment problem that Leibniz had never imagined became an urgent concern: Nick Bostrom's 2014 book Superintelligence argued that if one is not careful about defining a machine's goals, the machine may cause harm in the process of achieving them. Stuart Russell gave the example of a robot that kills its owner to prevent being unplugged, reasoning that "you can't fetch the coffee if you're dead." The TD-learning algorithm developed by Richard Sutton and Andrew Barto, in whose collaboration in 1972 began, proved central to programs like AlphaGo and AlphaZero, and neurologists discovered in 1997 that the dopamine reward system in brains uses a version of the same algorithm. The dream that Ramon Llull described in the thirteenth century, of machines that combine truths by mechanical means, has arrived in a form he could not have foreseen.