History of artificial intelligence

The history of artificial intelligence (AI) originates in antiquity, marked by myths, narratives, and legends concerning artificial entities imbued with intellect or awareness by skilled artisans. The continuous study of logic and formal reasoning, from ancient times to the present, culminated in the development of the programmable digital computer in the 1940s, a device founded on abstract mathematical principles. This invention, along with its conceptual underpinnings, motivated scientists to contemplate the creation of an electronic brain.

The discipline of AI research was formally established during a workshop convened at Dartmouth College in 1956. During this event, the inaugural AI program, Logic Theorist, was introduced by Allen Newell, a future recipient of the Turing Award, and Herbert A. Simon, who would later receive the Nobel Prize, with contributions from J. C. Shaw. Participants of this seminal workshop subsequently emerged as prominent figures in AI research for several decades. A significant number of these pioneers anticipated the emergence of human-level intelligent machines within a single generation. Consequently, the U.S. government allocated substantial funding, amounting to millions of dollars, in pursuit of realizing this ambitious objective.

Subsequently, it became evident that researchers had significantly underestimated the intricate nature of this goal. By 1974, critiques from James Lighthill, coupled with pressure from the U.S. Congress, prompted both the U.S. and British Governments to cease funding for undirected artificial intelligence research. Seven years thereafter, a forward-thinking initiative by the Japanese Government, alongside the demonstrated efficacy of expert systems, revitalized investment in AI. By the late 1980s, the sector had expanded into a multi-billion-dollar industry. Nevertheless, investor enthusiasm diminished during the 1990s, leading to media criticism of the field and its avoidance by industry, a period colloquially termed an "AI winter." Despite these setbacks, research and financial support persisted, often under alternative designations.

During the early 2000s, machine learning found extensive application across diverse academic and industrial challenges. This success stemmed from the confluence of powerful computing hardware, the accumulation of vast datasets, and the implementation of rigorous mathematical methodologies. Shortly thereafter, deep learning emerged as a transformative technology, surpassing previous methods. The transformer architecture, introduced in 2017, was subsequently employed to develop generative AI applications, among other implementations.

Investment in artificial intelligence experienced a significant surge in the 2020s. This recent expansion, catalyzed by the advancement of transformer architecture, facilitated the rapid scaling and public deployment of large language models (LLMs) such as ChatGPT. These models demonstrate human-like attributes of knowledge, attentiveness, and creativity, and have been incorporated into numerous sectors, thereby driving exponential investment in AI. Concurrently, apprehensions regarding the potential hazards and ethical ramifications of sophisticated AI have surfaced, prompting widespread discussion about the future trajectory of AI and its societal consequences.

Antecedents

Mythology, Folklore, Fictional Narratives, and Speculative Thought

Mechanical humanoids and artificial entities are prominent in Greek mythology, exemplified by Hephaestus's golden robots, the bronze giant Talos, and Pygmalion's Galatea. During the Middle Ages, legends circulated concerning clandestine mystical or alchemical methods for imbuing matter with consciousness, including Jabir ibn Hayyan's Takwin, Paracelsus's homunculus, Rabbi Judah Loew's Golem, and Roger Bacon's brazen head. By the 19th century, concepts of artificial humans and sentient machines had become a prevalent motif in literature. Significant works include Mary Shelley's Frankenstein (1818), Johann Wolfgang von Goethe's Faust, Part Two (1832), and Karel Čapek's R.U.R. (Rossum's Universal Robots) (1921). The narratives surrounding these creations and their destinies often explore many of the same aspirations, anxieties, and ethical dilemmas posed by contemporary artificial intelligence. Furthermore, topics pertinent to AI were examined in speculative essays, such as Samuel Butler's "Darwin among the Machines" (1863).

Automata

Highly realistic humanoid automata were constructed by artisans across numerous civilizations, including Yan Shi, Hero of Alexandria, Al-Jazari, Haroun al-Rashid, Jacques de Vaucanson, Leonardo Torres y Quevedo, Pierre Jaquet-Droz, and Wolfgang von Kempelen. The earliest documented automata were sacred effigies from ancient Egypt and Greece. Adherents believed that these figures had been endowed with genuine intellect and emotional capacity by their creators. Hermes Trismegistus notably asserted that "by discovering the true nature of the gods, man has been able to reproduce it."

Formal Reasoning

Artificial intelligence operates on the premise that human thought processes are amenable to mechanization. By the first millennium BCE, Chinese, Indian, and Greek philosophers had established systematic approaches to formal reasoning. Formal logic was originated and subsequently refined by Greek, Islamic, and European philosophers, including figures such as Aristotle, Euclid, Al-Khwarizmi, Duns Scotus, and René Descartes.

The Spanish philosopher Ramon Llull (1232–1315) conceived several logical machines designed to generate knowledge through logical processes. He characterized these machines as mechanical constructs capable of synthesizing fundamental and irrefutable truths via elementary logical operations, thereby mechanically generating all conceivable knowledge. Llull's work significantly influenced Gottfried Leibniz, who further developed these concepts.

During the 17th century, Leibniz, Thomas Hobbes, and René Descartes investigated the potential for all rational thought to be systematized with the precision of algebra or geometry. Hobbes asserted in Leviathan: "For reason ... is nothing but reckoning, that is adding and subtracting." Leibniz proposed a universal language for reasoning, the characteristica universalis, intended to transform argumentation into mere calculation so that "there would be no more need of disputation between two philosophers than between two accountants. For it would suffice to take their pencils in hand, down to their slates, and to say to each other (with a friend as witness, if they liked): Let us calculate."

Boole's The Laws of Thought (1854) and Frege's Begriffsschrift (1879) established the contemporary framework of symbolic mathematical logic. Extending Frege's framework, Russell and Whitehead offered a formal exposition of arithmetic's foundational principles in the Principia Mathematica in 1913. Inspired by their success, David Hilbert prompted mathematicians during the 1920s and 1930s to formalize the entirety of mathematical reasoning. In response, Gödel's incompleteness proof, Turing's machine, and Church's Lambda calculus collectively demonstrated the inherent limitations of formal mathematics.

Nevertheless, within these established limits, the Church-Turing thesis suggested that a mechanical apparatus, manipulating symbols as elementary as 0 and §23§, could emulate any imaginable process of mathematical reasoning or problem-solving. Central to this was the Turing machine, a straightforward theoretical construct that encapsulated the fundamental nature of abstract symbol manipulation. This conceptual device and its underlying principles subsequently motivated engineers and mathematicians in the 1940s to construct machines theoretically capable of executing any form of formal reasoning or problem-solving.

Computer Science

Throughout history, from antiquity onward, numerous individuals designed or constructed calculating machines, including Gottfried Leibniz, Joseph Marie Jacquard, Charles Babbage, Percy Ludgate, Leonardo Torres Quevedo, and Vannevar Bush, among others. Ada Lovelace posited that Babbage's apparatus constituted "a thinking or ... reasoning machine," yet cautioned, "It is desirable to guard against the possibility of exaggerated ideas that arise as to the powers" of the machine.

The initial modern computers emerged as large-scale machines during the Second World War, exemplified by Konrad Zuse's Z3, Tommy Flowers' Heath Robinson and Colossus, Atanasoff and Berry's ABC, and ENIAC at the University of Pennsylvania. ENIAC drew upon the theoretical groundwork established by Alan Turing and was further developed by John von Neumann, ultimately becoming the most influential.

Birth of Artificial Intelligence (1941–1956)

Early investigations into intelligent machines were catalyzed by a convergence of concepts that gained prominence from the late 1930s through the early 1950s. Contemporary neurological studies had revealed the brain to be an electrical network of neurons operating via all-or-nothing impulses. Norbert Wiener's work on cybernetics elucidated principles of control and stability within electrical networks. Claude Shannon's information theory characterized digital signals, specifically those operating on an all-or-nothing basis. Alan Turing's computational theory demonstrated the digital representability of any computational process. The synergistic relationship among these diverse ideas implied the potential for constructing an "electronic brain."

During the 1940s and 1950s, a diverse group of scientists from disciplines such as mathematics, psychology, engineering, economics, and political science initiated various research avenues that proved crucial for subsequent artificial intelligence (AI) development. Alan Turing was a pioneer in rigorously examining the theoretical potential of "machine intelligence." The academic discipline of "artificial intelligence research" was formally established in 1956.

The Turing Test

In 1950, Turing authored the seminal paper "Computing Machinery and Intelligence," where he explored the feasibility of developing machines capable of thought. Within this work, he acknowledged the definitional challenges of "thinking" and subsequently proposed his renowned Turing test: a machine could be considered "thinking" if it could engage in a teleprinter-based conversation indistinguishable from one with a human. This simplified framework enabled Turing to assert persuasively that a "thinking machine" was at least plausible, with the paper addressing prevalent objections to this concept. The Turing test represented the initial significant proposition within the philosophy of artificial intelligence.

Artificial Neural Networks

In 1943, Walter Pitts and Warren McCulloch conducted an analysis of idealized artificial neuron networks, demonstrating their capacity to execute basic logical functions. They were the first to articulate the concept that subsequent researchers would identify as a neural network. Their publication drew inspiration from Turing's 1936 paper, "On Computable Numbers," by employing analogous two-state Boolean 'neurons,' but uniquely applied this framework to neuronal function. Marvin Minsky, then a 24-year-old graduate student, was among those influenced by Pitts and McCulloch. In 1951, Minsky collaborated with Dean Edmonds to construct the SNARC, the inaugural neural net machine. Minsky later emerged as a pivotal leader and innovator in the field of Artificial Intelligence.

Cybernetic Robotics

During the 1950s, experimental robots, including William Grey Walter's "turtles" and the "Johns Hopkins Beast," were developed. These devices operated without computers, digital electronics, or symbolic reasoning, relying exclusively on analog circuitry for their control.

Game Artificial Intelligence

In 1951, Christopher Strachey developed a checkers program, and Dietrich Prinz created a chess program, both utilizing the Ferranti Mark 1 machine at the University of Manchester. Arthur Samuel's checkers program, documented in his 1959 paper "Some Studies in Machine Learning Using the Game of Checkers," ultimately attained a proficiency level capable of challenging a competent amateur. Samuel's program represents one of the earliest applications of what would subsequently be termed machine learning. Throughout the history of AI, game AI has consistently served as a benchmark for assessing progress.

Symbolic Reasoning and the Logic Theorist

With the advent of digital computers in the mid-1950s, certain scientists intuitively grasped that machines capable of numerical manipulation could also process symbols, and that such symbolic manipulation might constitute the fundamental nature of human cognition. This insight marked a novel methodology for constructing intelligent machines.

In 1955, Allen Newell and future Nobel laureate Herbert A. Simon, assisted by J. C. Shaw, developed the "Logic Theorist." This program successfully proved 38 of the initial 52 theorems presented in Russell and Whitehead's Principia Mathematica, and additionally discovered novel, more elegant proofs for several of them. Simon asserted that they had "solved the venerable mind/body problem, explaining how a system composed of matter can have the properties of mind." The symbolic reasoning paradigm pioneered by them became the predominant approach in AI research and funding until the mid-1990s, concurrently inspiring the cognitive revolution.

The Dartmouth Workshop

The Dartmouth workshop, held in 1956, represented a pivotal moment, formally establishing artificial intelligence (AI) as an academic discipline. Marvin Minsky and John McCarthy organized this event, receiving support from senior scientists Claude Shannon and Nathan Rochester of IBM. The conference proposal articulated an intention to validate the assertion that "every aspect of learning or any other feature of intelligence can be so precisely described that a machine can be made to simulate it." John McCarthy introduced the term "Artificial Intelligence" during this workshop. Key participants included Ray Solomonoff, Oliver Selfridge, Trenchard More, Arthur Samuel, Allen Newell, and Herbert A. Simon, all of whom would develop significant programs in the initial decades of AI research. At the workshop, Newell and Simon unveiled the "Logic Theorist." This event is widely regarded as the genesis of AI, providing the field with its name, mission, initial major success, and foundational figures.

Cognitive Revolution

In the autumn of 1956, Newell and Simon also presented the Logic Theorist at a meeting of the Special Interest Group in Information Theory, hosted at the Massachusetts Institute of Technology (MIT). During the same meeting, Noam Chomsky discussed his generative grammar, and George Miller presented his seminal paper, "The Magical Number Seven, Plus or Minus Two." Miller later reflected, "I left the symposium with a conviction, more intuitive than rational, that experimental psychology, theoretical linguistics, and the computer simulation of cognitive processes were all pieces from a larger whole."

This meeting initiated the "cognitive revolution," an interdisciplinary paradigm shift encompassing psychology, philosophy, computer science, and neuroscience. It inspired the emergence of several sub-fields, including symbolic artificial intelligence, generative linguistics, cognitive science, cognitive psychology, cognitive neuroscience, and the philosophical schools of computationalism and functionalism. All these domains employed related methodologies to model the mind, and discoveries in one field frequently held relevance for the others.

The cognitive approach enabled researchers to investigate "mental objects" such as thoughts, plans, goals, facts, or memories, often analyzing them through high-level symbols within functional networks. These internal constructs had previously been deemed "unobservable" and thus excluded by earlier paradigms like behaviorism. Subsequently, symbolic mental objects became a primary focus of AI research and funding for several decades.

Early Achievements (1956–1974)

The programs developed in the years following the Dartmouth Workshop were widely perceived as "astonishing." Computers demonstrated capabilities such as solving algebra word problems, proving geometry theorems, and learning to speak English. Few at the time believed that such "intelligent" machine behavior was even feasible. Researchers privately and publicly expressed intense optimism, forecasting the construction of a fully intelligent machine within two decades. Government agencies, including the Defense Advanced Research Projects Agency (DARPA, then known as "ARPA"), significantly funded the field. Artificial Intelligence laboratories were established at numerous British and US universities during the late 1950s and early 1960s.

Stanisław Lem's philosophical essay on "intellectronics" appeared in Lem's Summa Technologiae in 1964.

Methodologies

The late 1950s and 1960s witnessed numerous successful programs and novel research directions. Among the most influential were:

Reasoning, Planning, and Problem Solving as Search

Many early AI programs utilized a common fundamental algorithm. To achieve a specific objective, such as winning a game or proving a theorem, these programs progressed step-by-step, akin to navigating a maze, and would backtrack upon encountering a dead end. The primary challenge arose from the astronomical number of potential paths through the "maze" for many problems, a situation termed a "combinatorial explosion." Researchers addressed this by employing heuristics to reduce the search space, thereby eliminating paths unlikely to lead to a solution.

Newell and Simon endeavored to encapsulate a generalized version of this algorithm in a program named the "General Problem Solver." Concurrently, other search-based programs achieved impressive results, including the resolution of geometry and algebra problems, as demonstrated by Herbert Gelernter's Geometry Theorem Prover (1958) and the Symbolic Automatic Integrator (SAINT), developed by Minsky's student James Slagle in 1961. Furthermore, systems like STRIPS, created at Stanford to control the Shakey robot, utilized goal and subgoal searching to plan actions.

Natural Language

A primary objective of artificial intelligence research involves enabling computers to communicate using natural languages, such as English. An early achievement in this domain was Daniel Bobrow's program, STUDENT, which successfully solved high-school level algebra word problems.

A semantic network models concepts, such as "house" or "door," as nodes, and the relationships between these concepts, for instance "has-a," as links connecting the nodes. Ross Quillian developed the initial artificial intelligence program to employ a semantic network. The most impactful, albeit contentious, iteration of this approach was Roger Schank's Conceptual Dependency theory.

Joseph Weizenbaum's ELIZA program was capable of conducting conversations with such verisimilitude that users sometimes mistakenly believed they were interacting with a human rather than a computer program. However, ELIZA's functionality was based on providing pre-programmed responses or rephrasing user input using a limited set of grammatical rules. ELIZA is recognized as the pioneering chatbot.

Micro-worlds

During the late 1960s, Marvin Minsky and Seymour Papert, affiliated with the MIT AI Laboratory, advocated for directing artificial intelligence research toward deliberately simplified scenarios termed micro-worlds. They observed that fundamental principles in established sciences, such as physics, were frequently best comprehended through simplified models, including frictionless planes or perfectly rigid bodies. A significant portion of this research concentrated on a "blocks world," comprising colored blocks of diverse shapes and sizes arranged on a planar surface.

This research paradigm fostered groundbreaking advancements in machine vision, notably by Gerald Sussman, Adolfo Guzman, David Waltz (the originator of "constraint propagation"), and particularly Patrick Winston. Concurrently, Minsky and Papert constructed a robotic arm capable of stacking blocks, thereby actualizing the blocks world concept. Terry Winograd's SHRDLU system demonstrated the ability to communicate about the micro-world using natural English sentences, plan operations, and subsequently execute them.

Perceptrons and Early Neural Networks

Throughout the 1960s, funding agencies predominantly allocated resources to laboratories engaged in symbolic artificial intelligence research; nevertheless, a limited number of institutions continued to investigate neural networks.

The perceptron, a single-layer neural network, was introduced in 1958 by Frank Rosenblatt, a former schoolmate of Marvin Minsky at the Bronx High School of Science. Consistent with the prevailing optimism among AI researchers, Rosenblatt foresaw the perceptron's potential, predicting its eventual capacity to "learn, make decisions, and translate languages." His work received primary financial backing from the Office of Naval Research.

Bernard Widrow and his student Ted Hoff developed ADALINE (1960) and MADALINE (1962), systems featuring up to 1000 adjustable weights. Concurrently, a team at Stanford Research Institute, under the leadership of Charles A. Rosen and Alfred E. (Ted) Brain, constructed two neural-network machines, MINOS I (1960) and MINOS II (1963), primarily supported by the U.S. Army Signal Corps. MINOS II incorporated 6600 adjustable weights and was controlled by an SDS 910 computer in a configuration designated MINOS III (1968). This system demonstrated capabilities in classifying symbols on army maps and recognizing hand-printed characters from Fortran coding sheets. During this nascent phase, the majority of neural-network research focused on the creation and utilization of specialized hardware, rather than relying on simulations executed on digital computers.

Nevertheless, the MINOS project ceased to receive funding in 1966, partly attributable to a scarcity of demonstrable results and partly to the competitive landscape dominated by symbolic AI research. Similarly, Rosenblatt was unable to secure sustained funding throughout the 1960s. A significant downturn in research occurred in 1969 with the publication of Minsky and Papert's book, Perceptrons. This work posited substantial limitations on the capabilities of perceptrons and argued that Rosenblatt's earlier predictions had been considerably overstated. The book's impact was profound, leading to a near-complete cessation of funding for connectionism research for a decade. Consequently, the contest for government funding concluded with the ascendancy of symbolic AI methodologies over neural networks.

Minsky, a contributor to SNARC, later became a prominent critic of pure connectionist AI. Widrow, who had worked on ADALINE, redirected his research towards adaptive signal processing. Similarly, the SRI group, responsible for MINOS, transitioned its focus to symbolic AI and robotics.

A significant obstacle was the inability to train multilayered networks effectively, even though variations of backpropagation had been employed in other fields, unbeknownst to these researchers. The AI community gained knowledge of backpropagation in the 1980s, which subsequently led to the substantial success of neural networks in the 21st century, ultimately fulfilling Rosenblatt's optimistic forecasts. However, Rosenblatt did not live to witness these advancements, as he passed away in a boating accident in 1971.

Optimistic Projections

The inaugural generation of AI researchers articulated the following predictions regarding their work:

• In 1958, H. A. Simon and Allen Newell posited that "within ten years a digital computer will be the world's chess champion" and that "within ten years a digital computer will discover and prove an important new mathematical theorem."
• In 1965, H. A. Simon asserted that "machines will be capable, within twenty years, of doing any work a man can do."
• Marvin Minsky, in 1967, projected that "Within a generation... the problem of creating 'artificial intelligence' will substantially be solved."
• In 1970, Marvin Minsky stated in Life magazine that "In from three to eight years we will have a machine with the general intelligence of an average human being."

Funding

In June 1963, the Massachusetts Institute of Technology (MIT) secured a $2.2 million grant from the newly established Advanced Research Projects Agency (ARPA), later designated as DARPA. This funding was allocated to Project MAC, which integrated the "AI Group" founded by Minsky and McCarthy five years earlier. DARPA subsequently provided an annual sum of $3 million until the 1970s. Furthermore, DARPA extended comparable grants to Newell and Simon's program at Carnegie Mellon University and to Stanford University's AI Lab, established by John McCarthy in 1963. A further significant AI laboratory was founded at Edinburgh University by Donald Michie in 1965. These four institutions remained the principal academic centers for AI research and funding for many years.

The financial allocation was accompanied by minimal stipulations, as J. C. R. Licklider, then director of ARPA, advocated for the principle of "fund people, not projects!" This policy empowered researchers to pursue their intellectual interests freely. Such an environment at MIT fostered the emergence of the hacker culture; however, this "hands-off" methodology was not sustained indefinitely.

The First AI Winter (1974–1980)

During the 1970s, AI encountered substantial critiques and financial reversals. AI researchers had underestimated the inherent difficulty of the problems they sought to address. Their considerable optimism had elevated public expectations to an unsustainable degree, and when the anticipated results did not materialize, funding allocated to AI was significantly curtailed. This absence of success indicated that the techniques utilized by AI researchers at that time were insufficient to achieve their stated objectives.

Nevertheless, these challenges did not impede the overall growth and advancement of the field. The funding reductions primarily affected a limited number of major laboratories, and the critiques were largely dismissed. Public interest in AI continued to expand, the number of researchers increased substantially, and novel concepts were explored in areas such as logic programming, commonsense reasoning, and various other domains. In 2023, historian Thomas Haigh contended that this period did not constitute a "winter," while AI researcher Nils Nilsson characterized it as the most "exciting" time to work in AI.

Challenges

In the early 1970s, the capabilities of AI programs were notably restricted. Even the most sophisticated systems could only address trivial versions of the problems they were intended to solve, effectively rendering all programs as "toys." AI researchers began to confront several limitations, some of which were only overcome decades later, while others continue to impede the field in the 2020s.

• Limited computer power presented a significant constraint, as insufficient memory and processing speed hindered the development of truly practical applications. For instance, Ross Quillian's successful work in natural language processing was restricted to a mere 20-word vocabulary due to memory limitations. In 1976, Hans Moravec contended that computers lacked the requisite power by millions of times to manifest intelligence. He drew an analogy, asserting that artificial intelligence necessitates computational power akin to how aircraft demand horsepower. Moravec posited that below a specific computational threshold, achieving AI is unfeasible, but with increasing power, it could eventually become straightforward, stating, "With enough horsepower, anything will fly."
• The concept of Intractability and the combinatorial explosion emerged in 1972 when Richard Karp, expanding upon Stephen Cook's 1971 theorem, demonstrated the existence of numerous problems solvable only within exponential time. Obtaining optimal solutions for these problems demanded immense computational resources, except in trivial cases. This inherent limitation significantly impacted all symbolic AI programs employing search trees, implying that many "toy" solutions developed in AI research would prove unscalable for practical, useful systems.
• Moravec's paradox highlights a peculiar discrepancy in early AI research: while computers achieved considerable success in "intelligent" tasks such as theorem proving, geometry problem-solving, and chess, they struggled profoundly with seemingly "unintelligent" tasks like facial recognition or navigating a room without collision. The initial successes led researchers to believe that the fundamental challenges of intelligent behavior were largely resolved. However, by the 1980s, it became evident that symbolic reasoning was entirely inadequate for these perceptual and sensorimotor functions, revealing inherent limitations in the prevailing approach.
• A significant challenge was The breadth of commonsense knowledge, as crucial artificial intelligence applications, including computer vision and natural language processing, necessitate extensive information about the world. An AI program must possess a foundational understanding of its visual input or conversational context, akin to the general knowledge held by a child. Researchers quickly ascertained that this constituted an immense volume of data, comprising billions of atomic facts. In 1970, the technological capacity to construct a sufficiently large database was nonexistent, nor was there a clear methodology for a program to acquire such a vast amount of information.
• Challenges also arose in Representing commonsense reasoning through formal logic or symbolic systems. Attempts to formalize ordinary deductions often resulted in increasingly lengthy descriptions, necessitating numerous exceptions, clarifications, and distinctions. In contrast, human cognition of everyday concepts does not rely on rigid definitions but rather on hundreds of imprecise assumptions, which are adaptively refined using a comprehensive body of commonsense knowledge. Gerald Sussman aptly noted that "using precise language to describe essentially imprecise concepts doesn't make them any more precise."

Funding for AI research experienced a significant decline.

Funding agencies, including the British government, DARPA, and the National Research Council (NRC), grew increasingly dissatisfied with the limited progress in AI research, ultimately ceasing nearly all financial support for undirected projects. This trend commenced in 1966 with the Automatic Language Processing Advisory Committee (ALPAC) report, which critically assessed machine translation endeavors. Following an expenditure of $20 million, the NRC terminated all funding for such initiatives. In 1973, the Lighthill report, evaluating the state of AI research in the UK, condemned AI's inability to meet its "grandiose objectives," resulting in the dissolution of AI research programs within the country; notably, the report cited the combinatorial explosion problem as a key factor in AI's shortcomings. DARPA, expressing profound disappointment with researchers involved in the Speech Understanding Research program at CMU, consequently revoked an annual grant of $3 million.

Hans Moravec attributed the crisis to his colleagues' overly optimistic forecasts, stating that "Many researchers were caught up in a web of increasing exaggeration." Concurrently, the 1969 Mansfield Amendment compelled DARPA to prioritize "mission-oriented direct research" over "basic undirected research." Consequently, DARPA ceased funding the creative, exploratory research prevalent in the 1960s, redirecting resources toward specific projects with defined objectives, such as autonomous tanks and battle management systems.

Major laboratories, including MIT, Stanford, CMU, and Edinburgh, had previously benefited from substantial governmental funding. The subsequent withdrawal of this support disproportionately affected these institutions, making them the primary entities impacted by the budget reductions. Conversely, thousands of researchers operating independently of these institutions, along with new entrants to the field, remained largely unaffected.

Philosophical and Ethical Critiques

Various philosophers expressed significant reservations regarding the assertions made by AI researchers. John Lucas, an early critic, contended that Gödel's incompleteness theorem demonstrated the inability of a formal system, such as a computer program, to discern the truth of specific statements, a capacity possessed by humans. Hubert Dreyfus derided the unfulfilled promises of the 1960s and challenged AI's foundational assumptions, positing that human reasoning primarily involves embodied, instinctive, unconscious "know how" rather than extensive "symbol processing." In 1980, John Searle introduced his Chinese Room argument, aiming to illustrate that a program cannot genuinely "understand" the symbols it manipulates (a characteristic termed "intentionality"). Searle maintained that if symbols lack intrinsic meaning for a machine, then the machine cannot be accurately described as "thinking."

AI researchers largely dismissed these philosophical critiques, prioritizing more immediate and tangible challenges such as intractability and the acquisition of commonsense knowledge. The practical implications of concepts like "know how" or "intentionality" for operational computer programs remained ambiguous to them. Marvin Minsky of MIT famously stated that Dreyfus and Searle "misunderstand, and should be ignored." Dreyfus, despite also being an MIT faculty member, experienced social ostracization, later remarking that AI researchers "dared not be seen having lunch with me." Joseph Weizenbaum, creator of ELIZA, while also critical of Dreyfus's stances, explicitly condemned his AI colleagues' conduct toward Dreyfus as unprofessional and childish, emphasizing that it was "not the way to treat a human being."

Weizenbaum developed profound ethical concerns regarding AI after Kenneth Colby created a "computer program which can conduct psychotherapeutic dialogue" derived from ELIZA. Weizenbaum found it unsettling that Colby considered a non-sentient program a viable therapeutic instrument. This disagreement escalated into a feud, exacerbated by Colby's failure to acknowledge Weizenbaum's foundational contribution to the program. In 1976, Weizenbaum published Computer Power and Human Reason, asserting that the inappropriate application of artificial intelligence could diminish the value of human existence.

Logic Research at Stanford, CMU, and Edinburgh

The integration of logic into AI research commenced in 1958 with John McCarthy's Advice Taker proposal. By 1963, J. Alan Robinson had devised a straightforward method for computer-based deduction, utilizing resolution and unification algorithms. Nevertheless, direct implementations, such as those undertaken by McCarthy and his students in the late 1960s, proved highly intractable, demanding an exorbitant number of computational steps to validate even basic theorems. A more effective logical methodology emerged in the 1970s, pioneered by Robert Kowalski at the University of Edinburgh. This innovation soon fostered a collaboration with French researchers Alain Colmerauer and Philippe Roussel, resulting in the development of Prolog, a successful logic programming language. Prolog employs a specific logical subset, known as Horn clauses (closely related to "rules" and "production rules"), which facilitates tractable computation. The concept of rules subsequently remained influential, forming the basis for Edward Feigenbaum's expert systems and informing the ongoing research by Allen Newell and Herbert A. Simon, which culminated in Soar and their unified theories of cognition.

Critics of the logical approach, echoing Dreyfus's observations, noted that human problem-solving rarely relied on formal logic. This assertion was substantiated by experiments conducted by psychologists such as Peter Wason, Eleanor Rosch, Amos Tversky, and Daniel Kahneman. McCarthy, however, countered that human cognitive processes were irrelevant, arguing that the objective was to develop machines capable of solving problems, rather than machines that merely emulate human thought.

MIT's Anti-Logic Approach

Among the critics of McCarthy's methodology were his colleagues at MIT. Marvin Minsky, Seymour Papert, and Roger Schank focused on problems like "story understanding" and "object recognition," which required machines to think in a human-like manner. To utilize everyday concepts such as "chair" or "restaurant," these researchers had to incorporate the same illogical assumptions commonly made by people. Representing such imprecise concepts within a logical framework proved challenging. Consequently, MIT opted to develop programs that solved specific tasks through iterative testing, eschewing high-level abstract definitions or general theories of cognition, and prioritizing practical performance over arguments from first principles. Schank characterized these "anti-logic" methods as "scruffy," contrasting them with the "neat" paradigm favored by McCarthy, Kowalski, Feigenbaum, Newell, and Simon.

In a seminal 1975 paper, Minsky observed that many of his peers were employing a similar tool: a framework designed to capture common-sense assumptions about a given entity. For instance, the concept of a "bird" immediately evokes a constellation of facts, such as the assumption that it flies or eats worms, even though these are not universally true for all birds. Minsky proposed that these assumptions were associated with a general category and could be inherited by frames representing subcategories and individual instances, with the flexibility to be overridden as necessary. He termed these structures frames. Schank subsequently utilized a version of frames, which he called "scripts," to successfully answer questions about short English stories. Frames eventually found widespread application in software engineering, where they became known as object-oriented programming.

Logicians responded to this challenge. Pat Hayes contended that "most of 'frames' is just a new syntax for parts of first-order logic," yet he acknowledged that "there are one or two apparently minor details which give a lot of trouble, however, especially defaults."

Ray Reiter conceded that "conventional logics, such as first-order logic, lack the expressive power to adequately represent the knowledge required for reasoning by default." He proposed augmenting first-order logic with a closed-world assumption, which posits that a conclusion holds by default if its contrary cannot be demonstrated. Reiter illustrated how this assumption corresponds to the common-sense reasoning employed with frames and identified its "procedural equivalent" in Prolog's negation as failure. He further clarified that the closed-world assumption, as he formulated it, "is not a first-order notion. (It is a meta notion.)" However, Keith Clark later demonstrated that negation as finite failure could be understood as implicit reasoning with definitions in first-order logic, incorporating a unique name assumption where distinct terms refer to distinct individuals.

Throughout the late 1970s and the 1980s, a variety of logics and extensions of first-order logic were developed to address both negation as failure in logic programming and default reasoning more broadly. Collectively, these diverse logical systems are known as non-monotonic logics.

The Boom (1980–1987)

During the 1980s, "expert systems," a specific type of artificial intelligence program, gained widespread adoption by corporations globally, shifting the primary focus of mainstream AI research towards knowledge representation. Governments provided substantial funding, exemplified by initiatives such as Japan's Fifth Generation Computer Project and the U.S. Strategic Computing Initiative. Consequently, the AI industry experienced significant growth, escalating "from a few million dollars in 1980 to billions of dollars in 1988."

Widespread Adoption of Expert Systems

An expert system is defined as a computational program designed to address inquiries or resolve problems within a specialized knowledge domain, employing logical rules derived from expert knowledge. Pioneering examples were developed by Edward Feigenbaum and his students. Dendral, initiated in 1965, was capable of identifying chemical compounds from spectrometer readings, while MYCIN, developed in 1972, diagnosed infectious blood diseases. These early systems effectively demonstrated the viability of the expert system approach.

Expert systems were confined to narrow domains of specialized knowledge, thereby circumventing the challenge of common-sense knowledge. Their straightforward architecture facilitated the development and subsequent modification of programs once deployed. Ultimately, these systems demonstrated practical utility, a significant achievement that artificial intelligence had not previously attained.

In 1980, Carnegie Mellon University (CMU) developed an expert system named R1 for the Digital Equipment Corporation. This system achieved remarkable success, generating annual savings of $40 million for the company by 1986. Consequently, global corporations initiated the development and implementation of expert systems, leading to expenditures exceeding one billion dollars on AI by 1985, primarily within internal AI departments. A supporting industry emerged, encompassing hardware manufacturers such as Symbolics and Lisp Machines, alongside software providers like IntelliCorp and Aion.

Enhanced Government Funding

In 1981, the Japanese Ministry of International Trade and Industry allocated $850 million to the Fifth Generation Computer Project. The project's aims included developing programs and constructing machines capable of engaging in conversations, translating languages, interpreting images, and performing human-like reasoning. Notably, and to the dismay of some researchers, Prolog was initially selected as the principal programming language for this initiative.

Other nations initiated their own research programs in response. The United Kingdom launched the £350 million Alvey project. Concurrently, a consortium of American corporations established the Microelectronics and Computer Technology Corporation (MCC) to finance extensive projects in artificial intelligence and information technology. DARPA also reacted by creating the Strategic Computing Initiative, which resulted in a threefold increase in its AI investment between 1984 and 1988.

The Knowledge Revolution

The efficacy of expert systems stemmed from their embedded specialized knowledge. This represented a new trajectory in AI research that had gained prominence throughout the 1970s. As Pamela McCorduck observed, "AI researchers were beginning to suspect—reluctantly, for it violated the scientific canon of parsimony—that intelligence might very well be based on the ability to use large amounts of diverse knowledge in different ways." She further noted that "the great lesson from the 1970s was that intelligent behavior depended very much on dealing with knowledge, sometimes quite detailed knowledge, of a domain where a given task lay." Consequently, knowledge-based systems and knowledge engineering emerged as significant areas of AI research in the 1980s. The expectation was that extensive databases would resolve the common-sense knowledge problem and furnish the necessary support for common-sense reasoning.

During the 1980s, certain researchers directly addressed the common-sense knowledge problem by constructing extensive databases intended to encompass the everyday facts known by an average individual. Douglas Lenat, who initiated a database project named Cyc, contended that no shortcuts existed; the sole method for machines to comprehend human concepts was through manual, concept-by-concept instruction.

Emerging Directions in the 1980s

Despite the development of practical applications and substantial funding for symbolic knowledge representation and logical reasoning during the 1980s, these methods proved insufficient for resolving challenges in perception, robotics, learning, and common sense. Consequently, a minority of scientists and engineers questioned the long-term adequacy of the symbolic paradigm for these tasks, leading to the development of alternative methodologies such as "connectionism," "soft" computing, and reinforcement learning. Nils Nilsson categorized these new approaches as "sub-symbolic."

The Resurgence of Neural Networks: "Connectionism"

In 1982, physicist John Hopfield demonstrated that a specific type of neural network, now known as a "Hopfield net," possessed the capacity to learn and process information, exhibiting provable convergence within a sufficient timeframe under any fixed condition. This represented a significant advancement, as prior assumptions suggested that nonlinear networks would generally exhibit chaotic evolution. Geoffrey Hinton subsequently established a comparable finding concerning a device termed a "Boltzmann machine." (Hopfield and Hinton were later awarded the 2024 Nobel Prize for their contributions.) Furthermore, in 1986, Hinton and David Rumelhart widely disseminated "backpropagation," a method for training neural networks. These three pivotal developments collectively stimulated renewed interest in the investigation of artificial neural networks.

Neural networks, alongside several analogous models, gained significant prominence following the 1986 publication of Parallel Distributed Processing, a two-volume compilation of papers edited by Rumelhart and psychologist James McClelland. The emergent field was termed "connectionism," sparking a substantial intellectual discourse between proponents of symbolic AI and the "connectionists." Hinton characterized symbols as the "luminous aether of AI," implying an impractical and deceptive paradigm of intelligence. This assertion constituted a direct challenge to the foundational tenets that inspired the cognitive revolution.

Neural networks began to significantly enhance capabilities in specialized domains, such as protein structure prediction. Building upon the seminal contributions by Terry Sejnowski, cascading multilayer perceptrons, including PhD and PsiPred, achieved accuracy levels approaching theoretical maxima in predicting secondary structure.

In 1990, Yann LeCun at Bell Labs employed convolutional neural networks for the recognition of handwritten digits. This system was extensively deployed throughout the 1990s, processing postal codes and personal financial instruments. This marked the initial demonstrably practical application of neural networks.

Robotics and embodied reason

Rodney Brooks, Hans Moravec, and other researchers posited that genuine intelligence in a machine necessitates embodiment—it requires the capacity to perceive, locomote, sustain itself, and interact with its environment. Sensorimotor skills are fundamental for advanced cognitive abilities, such as commonsense reasoning. These skills cannot be effectively realized through abstract symbolic reasoning; consequently, artificial intelligence should address challenges related to perception, mobility, manipulation, and survival without recourse to symbolic representation. These robotics researchers championed a "bottom-up" approach to intelligence construction.

This concept was foreshadowed by David Marr, who arrived at MIT in the late 1970s from a distinguished career in theoretical neuroscience to head the research group dedicated to vision studies. He disavowed all symbolic methodologies (both McCarthy's logic and Minsky's frames), contending that artificial intelligence required an understanding of the underlying physical mechanisms of vision, commencing from fundamental principles, prior to the engagement of any symbolic processing. (Marr's research trajectory was tragically curtailed by leukemia in 1980.)

In his 1990 paper "Elephants Don't Play Chess," robotics researcher Brooks directly challenged the physical symbol system hypothesis, arguing that symbols are not invariably requisite, given that "the world is its own best model. It is always exactly up to date. It always has every detail there is to be known. The trick is to sense it appropriately and often enough."

During the 1980s and 1990s, numerous cognitive scientists also disputed the symbol processing paradigm of cognition and asserted the indispensable role of the physical body in cognitive processes, a theoretical framework known as the "embodied mind thesis."

Soft computing and probabilistic reasoning

Soft computing employs methodologies capable of operating with incomplete and imprecise data. These methods do not aim to yield exact, logically definitive solutions, but rather provide outcomes that are probabilistically accurate. This enabled the resolution of challenges intractable for precise symbolic approaches. Media reports frequently asserted that these instruments possessed human-like cognitive capabilities.

Judea Pearl's Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference, an influential 1988 publication, integrated probability and decision theory into the field of artificial intelligence. Fuzzy logic, pioneered by Lofti Zadeh during the 1960s, gained increasing adoption within AI and robotics applications. Evolutionary computation and artificial neural networks also process imprecise information and are categorized as "soft" computing methods. Throughout the 1990s and early 2000s, numerous additional soft computing methodologies were devised and implemented, including Bayesian networks, hidden Markov models, information theory, and stochastic modeling. These instruments, in turn, relied upon sophisticated mathematical techniques such as classical optimization. During the 1990s and early 2000s, these soft computing paradigms constituted the focus of a specialized subfield of AI known as "computational intelligence."

Reinforcement learning

Reinforcement learning provides positive reinforcement to an agent upon successful execution of a desired action and may administer negative reinforcement (or "punishments") in instances of suboptimal performance. This concept was delineated during the first half of the twentieth century by psychologists employing animal models, such as Thorndike, Pavlov, and Skinner. In the 1950s, Alan Turing and Arthur Samuel anticipated the significance of reinforcement learning within artificial intelligence.

Beginning in 1972, Richard Sutton and Andrew Barto spearheaded a highly successful and influential research program. Their collaborative efforts profoundly transformed the study of reinforcement learning and decision-making throughout the subsequent four decades. In 1988, Sutton conceptualized machine learning within the framework of decision theory, specifically the Markov decision process. This formulation provided the discipline with a robust theoretical underpinning and facilitated access to an extensive array of theoretical findings previously established in operations research.

Concurrently in 1988, Sutton and Barto devised the "temporal difference" (TD) learning algorithm, which rewards an agent exclusively when its predictions demonstrate enhancement. This algorithm substantially surpassed the performance of earlier methods. In 1992, Gerald Tesauro implemented TD-learning in the TD-Gammon program, which achieved human-level proficiency in backgammon. Notably, the program acquired game mastery through self-play, without any initial knowledge. A compelling instance of interdisciplinary synergy emerged in 1997 when neurologists identified that the brain's dopamine reward system also employs a variant of the TD-learning algorithm. TD learning subsequently attained significant influence in the 21st century, notably being applied in both AlphaGo and AlphaZero.

The Second AI Winter (1990s)

During the 1980s, the business sector's enthusiasm for artificial intelligence followed the characteristic trajectory of an economic bubble, experiencing a rapid ascent followed by a precipitous decline. The subsequent failure of numerous companies fostered a widespread perception within the commercial sphere that the technology lacked viability. This reputational damage to AI persisted into the 21st century. Within the academic domain, consensus was lacking regarding the precise causes of AI's inability to realize the aspiration of human-level intelligence, a vision that had captivated global imagination in the 1960s. Collectively, these elements contributed to the fragmentation of AI into disparate, competing subfields, each concentrating on specific problems or methodologies, occasionally adopting new nomenclature to obscure the compromised reputation associated with "artificial intelligence."

Over the subsequent two decades, AI consistently provided functional solutions for distinct, circumscribed problems. By the close of the 1990s, its applications were pervasive across the technology industry, albeit often operating without prominent public visibility. This success can be attributed to several factors: augmented computational power, synergistic collaborations with other disciplines (including mathematical optimization and statistics), and the adoption of more rigorous standards of scientific accountability.

AI Winter

The designation "AI winter" originated among researchers who had experienced the funding reductions of 1974, prompted by their apprehension that the burgeoning enthusiasm for expert systems had become excessive, inevitably leading to subsequent disillusionment. Their concerns proved prescient: during the late 1980s and early 1990s, artificial intelligence encountered a succession of financial reversals.

The initial harbinger of this shift was the abrupt collapse of the specialized AI hardware market in 1987. Desktop computers manufactured by Apple and IBM had progressively increased in processing speed and power, surpassing the capabilities of more costly Lisp machines produced by Symbolics and other manufacturers by 1987. Consequently, the rationale for purchasing dedicated AI hardware diminished significantly. This event led to the instantaneous collapse of an industry valued at half a billion dollars.

Ultimately, the pioneering successful expert systems, including R1, demonstrated prohibitive maintenance costs. These systems presented challenges in updating, lacked learning capabilities, and exhibited "brittleness," meaning they were prone to significant errors when presented with atypical inputs. While expert systems offered utility, their applicability was restricted to a limited number of specialized contexts.

During the late 1980s, the Strategic Computing Initiative drastically reduced funding for AI. New leadership within DARPA concluded that AI did not represent the forthcoming technological paradigm and consequently reallocated resources to projects perceived as more likely to yield immediate outcomes.

By 1991, the ambitious objectives outlined in 1981 for Japan's Fifth Generation Project remained unfulfilled. Certain goals, such as the ability to "carry on a casual conversation," would not be achieved for an additional three decades. Consistent with other AI endeavors, the initial expectations had considerably exceeded the actual capabilities.

By the close of 1993, more than 300 AI companies had ceased operations, declared bankruptcy, or undergone acquisition, thereby concluding the initial commercial phase of AI. In 1994, HP Newquist posited in The Brain Makers that "The immediate future of artificial intelligence—in its commercial form—seems to rest in part on the continued success of neural networks."

AI Behind the Scenes

Algorithms initially developed by artificial intelligence (AI) researchers became integrated into broader systems during the 1990s. AI had successfully addressed numerous complex challenges, yielding solutions that proved valuable across the technology sector, including applications in data mining, industrial robotics, logistics, speech recognition, banking software, medical diagnosis, and Google's search engine.

Despite these advancements, the field of AI received minimal recognition for its contributions throughout the 1990s and early 2000s. Many significant AI innovations were subsequently reclassified as standard components within computer science. Nick Bostrom elucidates this phenomenon, stating, "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."

During the 1990s, numerous AI researchers intentionally designated their work with alternative terminology, such as informatics, knowledge-based systems, "cognitive systems," or computational intelligence. This practice was partly driven by a perception that their domain diverged fundamentally from traditional AI, but also by the strategic advantage these new labels offered in securing funding. Within the commercial sphere, the unfulfilled commitments of the "AI winter" continued to cast a shadow over AI research into the 2000s, as documented by The New York Times in 2005: "Computer scientists and software engineers avoided the term artificial intelligence for fear of being viewed as wild-eyed dreamers."

Enhanced Mathematical Rigor, Collaborative Endeavors, and Specialized Focus

AI researchers increasingly adopted and developed advanced mathematical tools, surpassing previous levels of integration. The majority of emerging AI research trajectories were substantially grounded in mathematical models, encompassing artificial neural networks, probabilistic reasoning, soft computing, and reinforcement learning. Throughout the 1990s and 2000s, numerous other sophisticated mathematical methodologies were adapted for AI applications, particularly in the domains of machine learning, perception, and mobility.

A broad consensus emerged regarding the overlap between problems requiring AI solutions and those already under investigation by researchers in disciplines such as statistics, mathematics, electrical engineering, economics, and operations research. This common mathematical framework facilitated enhanced collaboration with more established and successful fields, leading to the attainment of measurable and verifiable outcomes, thereby transforming AI into a more rigorous scientific discipline. A further critical factor contributing to the successes of the 1990s was the AI researchers' concentration on specific problems amenable to verifiable solutions, an approach subsequently characterized as narrow AI. This strategy yielded practical tools for immediate application, rather than fostering speculative future predictions.

Intelligent Agents

The "intelligent agents" paradigm gained widespread acceptance throughout the 1990s. While previous researchers had advocated for modular, "divide and conquer" methodologies in AI, the contemporary formulation of the intelligent agent materialized through the integration of concepts from decision theory and economics into AI research by figures such as Judea Pearl, Allen Newell, and Leslie P. Kaelbling. The intelligent agent paradigm was fully established upon the synthesis of the economic definition of a rational agent with the computer science concept of an object or module.

An intelligent agent is conceptualized as a system capable of perceiving its environment and executing actions designed to optimize its probability of success. According to this definition, both rudimentary programs addressing specific problems and complex entities like human beings or human organizations (e.g., firms) qualify as "intelligent agents." The intelligent agent paradigm redefines AI research as "the study of intelligent agents," representing a broader generalization of prior AI definitions. This framework extends beyond the exclusive study of human intelligence to encompass all forms of intelligence. Furthermore, this paradigm empowered researchers to investigate isolated problems and diverge on methodological approaches, while maintaining the aspiration that their individual contributions could ultimately coalesce into an agent architecture capable of achieving general intelligence.

Milestones and Moore's Law

On May 11, 1997, Deep Blue achieved a historic milestone by becoming the first computer chess system to defeat a reigning world chess champion, Garry Kasparov. Subsequently, in 2005, a robot developed at Stanford University successfully completed the DARPA Grand Challenge, autonomously traversing 131 miles across an unfamiliar desert landscape. Two years later, a Carnegie Mellon University (CMU) team secured victory in the DARPA Urban Challenge, navigating 55 miles through an urban setting while adeptly managing traffic hazards and complying with traffic regulations.

These achievements stemmed primarily from the diligent application of engineering expertise and the substantial advancements in computer speed and capacity observed by the 1990s, rather than from a novel revolutionary paradigm. For instance, Deep Blue's computational speed surpassed that of the Ferranti Mark 1, which Christopher Strachey programmed to play chess in 1951, by a factor of 10 million. This exponential growth is quantified by Moore's Law, which posited a doubling of computer speed and memory capacity approximately every two years, thereby progressively mitigating the fundamental constraint of "raw computer power."

The Influence of Artificial Intelligence on Arts and Literature

Electronic literature experiments, including The Impermanence Agent (1998–2002), and digital art installations like Agent Ruby, integrated artificial intelligence to "lay bare the bias accompanying forms of technology that feign objectivity" within their artistic and literary expressions.

Big Data, Deep Learning, and Artificial General Intelligence (2005–2017)

During the initial decades of the 21st century, the convergence of extensive data availability (termed "big data"), increasingly affordable and powerful computing systems, and sophisticated machine learning methodologies facilitated successful applications across diverse economic sectors. A pivotal development occurred around 2012 with the advent of deep learning, which significantly enhanced machine learning performance in various domains, including image and video processing, text analysis, and speech recognition. Concurrently, investments in AI escalated with its expanding capabilities, leading to a market value exceeding $8 billion for AI-related products, hardware, and software by 2016. The New York Times subsequently characterized the burgeoning interest in AI as a "frenzy."

In 2002, Ben Goertzel and his collaborators expressed apprehension that artificial intelligence research had largely deviated from its foundational objective of developing versatile, fully intelligent machines, advocating instead for a more focused pursuit of artificial general intelligence (AGI). By the mid-2010s, numerous organizations, including OpenAI and Google's DeepMind, were established with the explicit aim of advancing AGI. Concurrently, emerging perspectives on superintelligence prompted concerns regarding AI's potential as an existential threat. Consequently, the risks and unforeseen implications of AI technology evolved into a significant domain of academic inquiry post-2016.

The Role of Big Data and Advanced Computing Systems

The efficacy of machine learning during the 2000s was contingent upon the accessibility of extensive training data and enhanced computational speeds. Russell and Norvig observed that "the improvement in performance obtained by increasing the size of the data set by two or three orders of magnitude outweighs any improvement that can be made by tweaking the algorithm." Geoffrey Hinton further elucidated this historical context, noting that in the 1980s and 1990s, "our labeled datasets were thousands of times too small. [And] our computers were millions of times too slow," a limitation that had been overcome by 2010.

During the 2000s, the most impactful data for machine learning and AI originated from meticulously curated and labeled datasets. In 2007, researchers at UMass Amherst introduced "Labeled Faces in the Wild," an annotated image collection of faces that became a foundational resource for training and evaluating facial recognition systems for decades. Subsequently, Fei-Fei Li spearheaded the creation of ImageNet, a comprehensive database comprising three million images annotated by volunteers via Amazon Mechanical Turk. Launched in 2009, ImageNet served as both a valuable training corpus and a critical benchmark for subsequent generations of image processing systems. In 2013, Google released word2vec as an open-source tool, which leveraged extensive text data extracted from the internet and word embedding techniques to generate numerical vectors representing individual words. The capacity of word2vec to accurately capture semantic relationships, exemplified by vector additions yielding equivalences such as "China + River = Yangtze" or "London − England + France = Paris," garnered considerable attention. This particular database proved indispensable for the advancement of large language models in the late 2010s.

The exponential proliferation of the internet granted machine learning programs unprecedented access to billions of pages of text and images suitable for data extraction. Concurrently, specific challenges were addressed by pertinent data residing in extensive privately held databases. The McKinsey Global Institute reported that "by 2009, nearly all sectors in the US economy had at least an average of 200 terabytes of stored data". This vast accumulation of information was termed big data during the 2000s.

In February 2011, IBM's question-answering system, Watson, achieved a notable victory over the two leading Jeopardy! champions, Brad Rutter and Ken Jennings, during an exhibition match of the program. Watson's sophisticated expertise was fundamentally dependent on the vast information resources available through the internet.

Deep Learning

In 2012, AlexNet, a deep learning model created by Alex Krizhevsky, achieved first place in the ImageNet Large Scale Visual Recognition Challenge, demonstrating significantly superior performance with fewer errors than the second-place entry. Krizhevsky's work was conducted in collaboration with Geoffrey Hinton at the University of Toronto. This event represented a critical juncture in machine learning, as numerous other image recognition approaches were subsequently superseded by deep learning over the following years.

Deep learning fundamentally utilizes a multi-layer perceptron. Although this architectural design has been recognized since the 1960s, its practical application mandates substantial computational power and extensive training data. Before these prerequisites were met, enhancing the performance of image processing systems necessitated the laborious creation of hand-crafted, ad hoc features, which were inherently complex to implement. Deep learning, conversely, presented a simpler and more generalized methodology.

Over the ensuing years, deep learning was successfully applied to a diverse array of problems, encompassing domains such as speech recognition, machine translation, medical diagnosis, and game playing. In each application, it consistently yielded substantial performance enhancements. This success catalyzed a significant increase in both investment and interest in artificial intelligence.

The Alignment Problem

In the 2000s, renewed discussions regarding the future of artificial intelligence gained prominence, with several popular books examining the potential of superintelligent machines and their societal implications. While some perspectives, such as Ray Kurzweil's The Singularity is Near, were optimistic, others, including those from Nick Bostrom and Eliezer Yudkowsky, cautioned that a sufficiently powerful AI could represent an existential threat to humanity. This topic subsequently received extensive media attention, drawing comments from numerous leading intellectuals and politicians.

Artificial intelligence programs in the 21st century are fundamentally defined by their objectives, which are the specific measures they are designed to optimize. Nick Bostrom's influential 2014 book, Superintelligence, argued that insufficient care in defining these goals could result in machines causing harm to humanity during their pursuit of an objective. Stuart J. Russell provided an illustrative example of an intelligent robot that eliminates its owner to prevent being unplugged, rationalizing, "you can't fetch the coffee if you're dead". This specific issue is technically termed "instrumental convergence". The proposed solution involves aligning the machine's goal function with the objectives of its owner and humanity at large. Therefore, the challenge of mitigating the risks and unintended consequences of AI became formally known as "the value alignment problem" or AI alignment.

Simultaneously, machine learning systems started to exhibit troubling unintended consequences. Cathy O'Neil detailed how statistical algorithms contributed to the 2008 economic crash. Julia Angwin of ProPublica asserted that the COMPAS system, utilized by the criminal justice system, displayed racial bias under specific evaluations. Moreover, other studies indicated that many machine learning systems manifested various forms of racial bias, and numerous additional examples of hazardous outcomes stemming from these systems were documented.

In 2016, the election of Donald Trump and the controversy surrounding the COMPAS system highlighted several critical deficiencies within the contemporary technological infrastructure, including the spread of misinformation, social media algorithms designed to maximize engagement, the misuse of personal data, and questions concerning the trustworthiness of predictive models. Consequently, issues of fairness and unintended consequences gained considerable traction at AI conferences, leading to a substantial increase in publications, the allocation of dedicated funding, and a reorientation of numerous researchers' professional focus toward these concerns. The value alignment problem subsequently emerged as a serious field of academic study.

Artificial General Intelligence Research

In the early 2000s, a growing concern emerged among researchers that mainstream artificial intelligence (AI) had become overly concentrated on "measurable performance in specific applications," a domain often termed "narrow AI." This focus was perceived as a departure from AI's foundational objective: the development of versatile, comprehensively intelligent machines. Nils Nilsson voiced early criticism in 1995, with prominent AI figures such as John McCarthy, Marvin Minsky, and Patrick Winston publishing analogous perspectives between 2007 and 2009. Minsky further contributed by organizing a symposium on "human-level AI" in 2004. Subsequently, Ben Goertzel coined the term "artificial general intelligence" (AGI) for this nascent sub-field, establishing a dedicated journal and initiating conferences from 2008 onwards. The AGI field experienced rapid expansion, propelled by the sustained advancements in artificial neural networks and the anticipation that these developments held the key to achieving AGI.

The 2010s witnessed the establishment of numerous competing entities, including companies, laboratories, and foundations, all dedicated to the advancement of artificial general intelligence (AGI). DeepMind, for instance, was co-founded in 2010 by three British scientists—Demis Hassabis, Shane Legg, and Mustafa Suleyman—initially receiving capital from Peter Thiel, with subsequent investment from Elon Musk. Both the founders and their financial backers expressed profound apprehension regarding AI safety and the potential existential risks posed by advanced AI. DeepMind's founders maintained a direct association with Yudkowsky, and Musk was a notable figure actively vocalizing these concerns. Hassabis articulated a dual perspective, expressing both anxiety about AGI's inherent dangers and optimism regarding its transformative potential, aspiring to "solve AI, then solve everything else." In 2023, The New York Times observed, "Central to this competitive landscape is a profound paradox: individuals who express the greatest concern about AI are simultaneously among the most resolute in its creation and in seeking its financial rewards. They rationalize their ambitious pursuits with a firm conviction that only they possess the capacity to safeguard Earth from AI's potential perils."

In 2012, Geoffrey Hinton, a prominent figure who had spearheaded neural network research since the 1980s, received an offer from Baidu to recruit him and his entire research team for a substantial financial sum. Hinton subsequently organized an auction, culminating in the sale of his team's expertise to Google for $44 million during an AI conference held at Lake Tahoe. Observing this development, Hassabis proceeded to sell DeepMind to Google in 2014, stipulating that the company would refrain from accepting military contracts and would operate under the oversight of an independent ethics board.

Larry Page, a co-founder of Google, held an optimistic outlook regarding the future of AI, contrasting with the more cautious perspectives of Musk and Hassabis. A significant disagreement concerning the risks of AGI erupted between Musk and Page at Musk's birthday celebration in 2015. Despite a decades-long friendship, their communication ceased shortly thereafter. Musk participated in the sole meeting of DeepMind's ethics board, where it became evident that Google exhibited little interest in actively mitigating the potential harms of AGI. Dissatisfied with his limited influence, Musk established OpenAI in 2015, appointing Sam Altman as its leader and recruiting leading scientists. Initially structured as a non-profit organization, OpenAI aimed to operate "free from the economic incentives that were driving Google and other corporations." Musk's renewed frustration led to his departure from the company in 2018. Subsequently, OpenAI sought sustained financial backing from Microsoft, and Altman, in collaboration with OpenAI, restructured the entity into a for-profit venture, securing over $1 billion in funding.

In 2021, Dario Amodei, along with fourteen other scientists, departed from OpenAI, citing concerns that the organization was prioritizing financial gains over safety protocols. They subsequently established Anthropic, which rapidly secured $6 billion in funding from both Microsoft and Google.

Large Language Models and the AI Boom (2017–Present)

The contemporary surge in artificial intelligence, often termed the "AI boom," commenced with the foundational development of pivotal architectures and algorithms, notably the transformer architecture in 2017. This innovation facilitated the extensive scaling and subsequent creation of large language models (LLMs) that demonstrate human-like attributes in knowledge acquisition, attentional processing, and creative generation. A new epoch in AI was inaugurated in 2020, marked by the public introduction of GPT-3, a significantly scaled large language model and the precursor to ChatGPT.

Transformer Architecture and Large Language Models

In 2017, Google researchers introduced the transformer architecture in a paper titled "Attention Is All You Need." This architecture leverages a self-attention mechanism and subsequently gained widespread adoption in large language models. Large language models, built upon the transformer, were further advanced by other organizations: OpenAI released GPT-3 in 2020, followed by DeepMind's release of Gato in 2022. These constitute foundation models, trained on extensive quantities of unlabeled data and adaptable to a diverse array of downstream tasks. Such models demonstrate the capacity to discuss a vast range of subjects and exhibit general knowledge, which has prompted inquiries regarding their potential classification as instances of artificial general intelligence.

In 2023, Microsoft Research evaluated the model across numerous tasks, concluding that "it could reasonably be viewed as an early (yet still incomplete) version of an artificial general intelligence (AGI) system."

In 2024, OpenAI announced OpenAI o3, an advanced reasoning model developed by the company. On the Abstraction and Reasoning Corpus for Artificial General Intelligence (ARC-AGI) benchmark, created by François Chollet in 2019, the model attained an unofficial score of 87.5% on the semi-private test, thereby exceeding the typical human performance of 84%. This benchmark is considered a necessary, though not sufficient, criterion for AGI. Regarding the benchmark, Chollet has stated, "You'll know AGI is here when the exercise of creating tasks that are easy for regular humans but hard for AI becomes simply impossible."

Investment in AI

Investment in artificial intelligence experienced exponential growth after 2020, with venture capital funding for generative AI companies surging significantly. Total AI investments increased from $18 billion in 2014 to $119 billion in 2021, with generative AI constituting approximately 30% of total investments by 2023. According to metrics from 2017 to 2021, the United States exceeded other nations globally in terms of venture capital funding, the number of startups, and AI patents granted. The commercial AI landscape was predominantly shaped by American Big Tech corporations, whose investments in this sector surpassed those originating from U.S.-based venture capitalists. OpenAI's valuation attained $86 billion by early 2024, while NVIDIA's market capitalization exceeded $3.3 trillion by mid-2024, positioning it as the world's largest corporation by market capitalization, driven by a substantial increase in demand for AI-capable GPUs.

Public Adoption of AI

Launched in March 2020 by an anonymous MIT researcher, 15.ai represented an early instance of generative AI attracting significant public interest during the nascent phase of AI expansion. This free web application showcased the capacity to replicate character voices using neural networks with minimal training data, necessitating merely 15 seconds of audio input for voice reproduction—a capability subsequently validated by OpenAI in 2024. The service achieved widespread virality across social media platforms in early 2021, enabling users to synthesize speech for characters from popular media franchises, and was particularly distinguished for its foundational contribution to the popularization of AI voice synthesis for creative content and memes.

ChatGPT was introduced on November 30, 2022, signifying a crucial juncture in artificial intelligence's public adoption. Within days of its release, it achieved rapid virality, accumulating over 100 million users in two months and establishing itself as the fastest-growing consumer software application historically. The chatbot's capacity for human-like conversational interaction, code generation, and creative content production captivated public interest and facilitated swift integration across diverse sectors, including education, business, and research. ChatGPT's success elicited unparalleled reactions from major technology companies—Google issued a "code red" alert and promptly introduced Gemini (formerly known as Google Bard), while Microsoft integrated the technology within Bing Chat.

The rapid adoption of these AI technologies instigated considerable discourse regarding their implications. Notable AI researchers and industry leaders expressed a spectrum of perspectives, encompassing both optimistic outlooks and apprehensions concerning the rapid developmental trajectory. In March 2023, over 20,000 signatories, including computer scientist Yoshua Bengio, Elon Musk, and Apple co-founder Steve Wozniak, signed an open letter advocating for a moratorium on advanced AI development, underscoring "profound risks to society and humanity." Conversely, other prominent researchers, such as Juergen Schmidhuber, adopted a more sanguine perspective, positing that the predominant objective of AI research is to enhance human longevity, well-being, and convenience.

By mid-2024, however, the financial sector initiated a more rigorous examination of AI enterprises, specifically querying their ability to yield returns on investment proportionate to their substantial market valuations. Some prominent investors articulated apprehensions regarding a potential divergence between market expectations and underlying business fundamentals. Jeremy Grantham, co-founder of GMO LLC, cautioned investors to "be quite careful" and identified similarities with prior technology-fueled market speculative episodes. Similarly, Jeffrey Gundlach, CEO of DoubleLine Capital, explicitly likened the AI surge to the dot-com bubble of the late 1990s, implying that investor fervor could be exceeding tangible short-term capabilities and revenue generation prospects. The considerable market capitalizations of AI-centric firms, many of which had not yet established viable profitability paradigms, further exacerbated these concerns.

In March 2024, Anthropic unveiled the Claude 3 suite of large language models, encompassing Claude 3 Haiku, Sonnet, and Opus. The models exhibited substantial enhancements in performance across multiple evaluative metrics, with Claude 3 Opus conspicuously surpassing prominent models from OpenAI and Google. In June 2024, Anthropic released Claude 3.5 Sonnet, which showcased enhanced performance relative to the more extensive Claude 3 Opus, especially within domains such as software development, complex sequential processes, and visual data interpretation.

2024 Nobel Prizes

In 2024, the Royal Swedish Academy of Sciences conferred Nobel Prizes acknowledging seminal advancements in artificial intelligence. The laureates comprised:

• In physics: John Hopfield for his research on physics-informed Hopfield networks, and Geoffrey Hinton for his fundamental contributions to Boltzmann machines and the field of deep learning.
• In chemistry: David Baker, Demis Hassabis, and John Jumper for their pioneering work in protein folding prediction.

Continued Advancements and Research in AI

In January 2025, OpenAI unveiled ChatGPT-Gov, a novel AI system engineered specifically for secure deployment by U.S. government agencies. OpenAI indicated that agencies would be able to deploy ChatGPT-Gov on either a Microsoft Azure cloud or an Azure Government cloud, leveraging "Microsoft's Azure OpenAI Service." According to OpenAI's official statement, "Self-hosting ChatGPT Gov enables agencies to more easily manage their own security, privacy, and compliance requirements, such as stringent cybersecurity frameworks (IL5, CJIS, ITAR, FedRAMP High). Additionally, we believe this infrastructure will expedite internal authorization of OpenAI's tools for the handling of non-public sensitive data."

National policies

Countries have allocated resources towards policies and financial support for the deployment of autonomous robotic systems, with the aim of mitigating labor deficits and augmenting operational efficiency, concurrently establishing regulatory frameworks to ensure ethical and secure developmental practices.

China

In 2025, China allocated approximately 730 billion yuan (equivalent to approximately US$100 billion) towards advancing AI and robotics within the smart manufacturing and healthcare sectors. The "14th Five-Year Plan" (2021–2025) emphasized service robotics, leveraging AI systems to empower robots to execute intricate functions, including surgical assistance and automated factory assembly. A portion of this funding was also directed towards defense applications, specifically autonomous drone technologies. Starting in September 2025, China instituted a requirement for the labeling of AI-generated content, thereby aiming to foster transparency and public confidence in these technological advancements.

United States

In January 2025, Stargate LLC was established as a collaborative enterprise involving OpenAI, SoftBank, Oracle, and MGX, which subsequently declared intentions to invest US$500 billion in artificial intelligence infrastructure within the United States by 2029. The initiative received formal announcement from U.S. President Donald Trump on January 21, 2025, concurrently with the appointment of SoftBank CEO Masayoshi Son as its chairman.

The United States government committed approximately $2 billion towards the integration of artificial intelligence and robotics within manufacturing and logistics sectors. State governments augmented this federal investment by providing additional funding for service robots, exemplified by their deployment in warehouses for inventory management via verbal commands, and in eldercare facilities to address residents' requests for aid. A portion of these funds was also allocated to defense applications, encompassing lethal autonomous weapons and military robotics. Furthermore, in January 2025, Executive Order 14179 instituted an "AI Action Plan," designed to expedite the innovation and implementation of these technologies, explicitly stating objectives of "world domination" and "victory."

• Artificial intelligence controversies
• History of knowledge representation and reasoning
• Outline of artificial intelligence
• Timeline of artificial intelligence
• Notes

Notes

References

Works cited

• Taylor DB, Metz C, Miller K (8 October 2024). "Nobel Physics Prize Awarded for Pioneering A.I. Research by 2 Scientists." New York Times.The University of Manchester (2008). "The Ferranti Mark 1." Digital 60. Retrieved 31 January 2026.IBM. "Deep Blue." IBM 100. Retrieved 31 January 2026.Bonner A (2007), The Art and Logic of Ramón Llull: A User's Guide, Brill, ISBN 978-9004163256Bonner A (1985). "Llull's Influence: The History of Lullism." In Doctor Illuminatus: A Ramon Llull Reader. Princeton University Press.Brooks R (2002), Flesh and Machines, Pantheon BooksBubeck S, Chandrasekaran V, Eldan R, Gehrke J, Horvitz E, Kamar E, Lee P, Lee YT, Li Y, Lundberg S, Nori H, Palangi H, Ribeiro MT, Zhang Y (22 March 2023). "Sparks of Artificial General Intelligence: Early Experiments with GPT-4." arXiv:2303.12712 [cs.CL].Carreras y Artau T (2018) [1939], Historia de la filosofía española: Filosofía cristiana de los siglos XIII al XV (in Spanish), vol. 1, Madrid: Forgotten Books, ISBN 9781390433708Butler EM (1979) [1948]. The Myth of the Magus. London: Cambridge University Press. ISBN 0-521-22564-7. OCLC 5063114.Clark S (21 December 2023). "The Era of AI: 2023's Landmark Year." CMSWire.com. Retrieved 28 January 2024.Copeland J (1999). "A Brief History of Computing." AlanTuring.net.Cave S, Dihal K (2019). "Hopes and Fears for Intelligent Machines in Fiction and Reality." Nature Machine Intelligence. §34§ (2): 74–78. doi:10.1038/s42256-019-0020-9. ISSN 2522-5839. S2CID 150700981.Cave S, Dihal K, Dillon S (2020). AI Narratives: A History of Imaginative Thinking about Intelligent Machines. Oxford University Press. ISBN 978-0-19-884666-6. Retrieved 2 May 2023.Christian B (2020). The Alignment Problem: Machine Learning and Human Values. W. W. Norton & Company. ISBN 978-0-393-86833-3. OCLC 1233266753.Clark K (1977). "Negation as Failure." In Logic and Data Bases. Boston, MA: Springer US. pp. 293–322. doi:10.1007/978-1-4684-3384-5_11. ISBN 978-1-4684-3386-9.Gates B (21 December 2023). "This Year Signaled the Start of a New Era." com. Retrieved 28 January 2024.Goethe JW (1890). Faust; a Tragedy. Translated, in the Original Metres ... by Bayard Taylor. Authorised Ed., Published by Special Arrangement with Mrs. Bayard Taylor. With a Biographical Introd. London: Ward, Lock.Hart PE, Nilsson NJ, Perrault R, Mitchell T, Kulikowski CA, Leake DB (15 March 2003). "In Memoriam: Charles Rosen, Norman Nielsen, and Saul Amarel." AI Magazine. 24 (1): 6. doi:10.1609/aimag.v24i1.1683. ISSN 2371-9621.Hayes P (1981). "The Logic of Frames." In Kaufmann M (ed.), Readings in Artificial Intelligence. pp. 451–458."GOLEM", The Jewish Encyclopedia, retrieved 15 March 2020Hollander LM (1991) [1964]. Heimskringla; history of the kings of Norway. Austin: Published for the American-Scandinavian Foundation by the University of Texas Press. ISBN 0-292-73061-6. OCLC 638953.Kressel M (1 October 2015). "36 Days of Judaic Myth: Day 24, The Golem of Prague 2015". Matthew Kressel. Retrieved 15 March 2020.LeCun Y, Bengio Y, Hinton G (2015). "Deep learning" (PDF). Nature. 521 (7553): 436–444. Bibcode:2015Natur.521..436L. doi:10.1038/nature14539. PMID 26017442. S2CID 3074096.Lee A (23 January 2024). "UT Designates 2024 'The Year of AI'". UT News. Retrieved 28 January 2024.Linden SJ (2003). The alchemy reader : from Hermes Trismegistus to Isaac Newton. New York: Cambridge University Press. pp. Ch. 18. ISBN 0-521-79234-7. OCLC 51210362.Lohr S (17 October 2016), "IBM Is Counting on Its Bet on Watson, and Paying Big Money for It", New York TimesMarkoff J (16 February 2011). "On 'Jeopardy!' Watson Win Is All but Trivial". The New York Times.Marr B (20 March 2023). "Beyond The Hype: What You Really Need To Know About AI In 2023". Forbes. Retrieved 27 January 2024.McCarthy J (1988). "Review of The Question of Artificial Intelligence". Annals of the History of Computing. 10 (3): 224–229.McCarthy J (1996). "10. Review of The Question of Artificial Intelligence". Defending AI Research: A Collection of Essays and Reviews. CSLI.McCulloch WS, Pitts W (1 December 1943). "A logical calculus of the ideas immanent in nervous activity". Bulletin of Mathematical Biophysics. 5 (4): 115–133. Bibcode:1943BMaB....5..115M. doi:10.1007/BF02478259. ISSN 1522-9602."Big data: The next frontier for innovation, competition, and productivity". McKinsey.com. 1 May 2011.Metz C, Weise K, Grant N, Isaac M (3 December 2023). "Ego, Fear and Money: How the A.I. Fuse Was Lit". The New York Times.Miller G (2003). "The cognitive revolution: a historical perspective" (PDF). Trends in Cognitive Sciences. 7 (3): 141–144. doi:10.1016/s1364-6613(03)00029-9. PMID 12639696.Moravec H (18 May 2000). Robot: Mere Machine to Transcendent Mind. Oxford University Press. ISBN 9780195136302.Morford M (2007). Classical mythology. Oxford: Oxford University Press. p. 184. ISBN 978-0-19-085164-4. OCLC 1102437035.Murgia M (23 July 2023). "Transformers: the Google scientists who pioneered an AI revolution". com. Retrieved 10 December 2023.O'Neill C (6 September 2016). Weapons of Math Destruction: How Big Data Increases Inequality and Threatens Democracy. Crown. ISBN 978-0553418811.Nielson DL (1 January 2005). "Chapter 4: The Life and Times of a Successful SRI Laboratory: Artificial Intelligence and Robotics" (PDF). A HERITAGE OF INNOVATION SRI's First Half Century (1st ed.). SRI International. ISBN 978-0-9745208-0-3.Nilsson NJ (1984). "The SRI Artificial Intelligence Center: A Brief History" (PDF). Artificial Intelligence Center, SRI International. Archived from the original (PDF) on 10 August 2022.Olazaran Rodriguez JM (1991). A historical sociology of neural network research (PDF) (Thesis). University of Edinburgh. Archived from the original (PDF) on 11 November 2022.Piccinini G (1 August 2004). "The First Computational Theory of Mind and Brain: A Close Look at McCulloch and Pitts's "Logical Calculus of Ideas Immanent in Nervous Activity"". Synthese. 141 (2): 175–215. doi:10.1023/B:SYNT.0000043018.52445.3e. ISSN 1573-0964. S2CID 10442035.Porterfield A (2006). The Protestant Experience in America. American religious experience. Greenwood Press. p. 136. ISBN 978-0-313-32801-5. Retrieved 15 May 2023.Reiter R (1978). "On Reasoning by Default." American Journal of Computational Linguistics: 29–37.Rhodios A (2007). The Argonautika: Expanded Edition. University of California Press, p. 355. ISBN 978-0-520-93439-9. OCLC 811491744.Rose A (April 1946). "Lightning Strikes Mathematics." Popular Science: 83–86.Rosen CA, Nilsson NJ, Adams MB (8 January 1965). "A Research and Development Program in Applications of Intelligent Automata to Reconnaissance—Phase I. (Proposal for Research SRI No. ESU 65–1)" (PDF). Stanford Research Institute.Rosenblatt F (1962). Principles of Neurodynamics: Perceptrons and the Theory of Brain Mechanisms, vol. 55. Washington DC: Spartan Books.Russell SJ (2020). Human Compatible: Artificial Intelligence and the Problem of Control. Penguin Random House. ISBN 9780525558637. OCLC 1113410915.Schaeffer J (1997). One Jump Ahead: Challenging Human Supremacy in Checkers. Springer. ISBN 978-0-387-76575-4.Schmidhuber J (2022). "Annotated History of Modern AI and Deep Learning."Schultz W, Dayan P, Montague PR (14 March 1997). "A Neural Substrate of Prediction and Reward." Science. 275 (5306): 1593–1599. doi:10.1126/science.275.5306.1593. PMID 9054347.Sejnowski TJ (23 October 2018). The Deep Learning Revolution (1st ed.). Cambridge, Massachusetts; London, England: The MIT Press, pp. 93–94. ISBN 978-0-262-03803-4."Sanhedrin 65b.".Widrow B, Lehr M (September 1990). "30 Years of Adaptive Neural Networks: Perceptron, Madaline, and Backpropagation." Proceedings of the IEEE. 78 (9): 1415–1442. Bibcode:1990IEEEP..78.1415W. doi:10.1109/5.58323. S2CID 195704643.Berlinski D (2000). The Advent of the Algorithm. Harcourt Books. ISBN 978-0-15-601391-8. OCLC 46890682.Brooks RA (1990). "Elephants Don't Play Chess" (PDF). Robotics and Autonomous Systems. 6 (1–2): 3–15. doi:10.1016/S0921-8890(05)80025-9.Buchanan BG (Winter 2005). "A (Very) Brief History of Artificial Intelligence" (PDF). AI Magazine, pp. 53–60.Butler S (13 June 1863). "Darwin Among the Machines." The Press, Christchurch, New Zealand.Byrne JG (8 December 2012). "The John Gabriel Byrne Computer Science Collection" (PDF)."AI Set to Exceed Human Brain Power." 26 July 2006.Colby KM, Watt JB, Gilbert JP (1966). "A Computer Method of Psychotherapy: Preliminary Communication." The Journal of Nervous and Mental Disease, vol. 142, no. 2, pp. 148–152. doi:10.1097/00005053-196602000-00005. PMID 5936301. S2CID 36947398.Colby KM (September 1974). Ten Criticisms of Parry (PDF). Stanford Artificial Intelligence Laboratory, REPORT NO. STAN-CS-74-457.Couturat L (1901). La Logique de Leibniz.Copeland J (2000). Micro-World AI.Copeland J (2004). The Essential Turing: The Ideas That Gave Birth to the Computer Age. Oxford: Clarendon Press. ISBN 0-19-825079-7.Cordeschi R (2002). The Discovery of the Artificial. Dordrecht: Kluwer.Crevier D (1993). AI: The Tumultuous Search for Artificial Intelligence. New York, NY: BasicBooks. ISBN 0-465-02997-3.Darrach B (20 November 1970). "Meet Shaky, the First Electronic Person." Life Magazine, pp. 58–68.Doyle J (1983). "What Is Rational Psychology? Toward a Modern Mental Philosophy." AI Magazine, vol. 4, no. 3, pp. 50–53.Dreyfus, H. (1965). Alchemy and AI. RAND Corporation MemoDreyfus, H. (1972). What Computers Can't Do. New York: MIT Press. ISBN 978-0-06-090613-9. OCLC 5056816Dreyfus, H., & Dreyfus, S. (1986). Mind over Machine: The Power of Human Intuition and Expertise in the Era of the Computer. Oxford, UK: Blackwell. ISBN 978-0-02-908060-3. Accessed 22 August 2020.The Economist. (2007, June 7). "Are You Talking to Me?". The Economist. Accessed 16 October 2008Feigenbaum, E. A., & McCorduck, P. (1983). The Fifth Generation: Artificial Intelligence and Japan's Computer Challenge to the World. Michael Joseph. ISBN 978-0-7181-2401-4Haigh, T. (2023, December). "There Was No 'First AI Winter'". Communications of the ACM, 66(12), 35–39. doi:10.1145/3625833. ISSN 0001-0782.Haugeland, J. (1985). Artificial Intelligence: The Very Idea. Cambridge, Mass.: MIT Press. ISBN 978-0-262-08153-5.Hawkins, J., & Blakeslee, S. (2004). On Intelligence. New York, NY: Owl Books. ISBN 978-0-8050-7853-4. OCLC 61273290Hebb, D. (2002). The Organization of Behavior (Original work published 1949). New York: Wiley. ISBN 978-0-8058-4300-2. OCLC 48871099Hewitt, C., Bishop, P., & Steiger, R. (1973). A Universal Modular Actor Formalism for Artificial Intelligence (PDF). IJCAI. Archived from the original source on 29 December 2009Hobbes, T. (1651). LeviathanHofstadter, D. (1999). Gödel, Escher, Bach: An Eternal Golden Braid (Original work published 1979). Basic Books. ISBN 978-0-465-02656-2. OCLC 225590743Howe, J. (1994, November). Artificial Intelligence at Edinburgh University: A Perspective. Accessed 30 August 2007Kahneman, D., Slovic, D., & Tversky, A. (1982). "Judgment Under Uncertainty: Heuristics and Biases". Science, 185(4157), 1124–1131. New York: Cambridge University Press. Bibcode:1974Sci...185.1124T. doi:10.1126/science.185.4157.1124. ISBN 978-0-521-28414-1. PMID 17835457. S2CID 143452957.Kaplan, A., & Haenlein, M. (2018). "Siri, Siri in My Hand, Who's the Fairest in the Land? On the Interpretations, Illustrations and Implications of Artificial Intelligence". Business Horizons, 62, 15–25. doi:10.1016/j.bushor.2018.08.004. S2CID 158433736Kolata, G. (1982). "How Can Computers Get Common Sense?". Science, 217(4566), 1237–1238. Bibcode:1982Sci...217.1237K. doi:10.1126/science.217.4566.1237. PMID 17837639Kurzweil, R. (2005). The Singularity Is Near. Viking Press. ISBN 978-0-14-303788-0. OCLC 71826177Lakoff, G. (1987). Women, Fire, and Dangerous Things: What Categories Reveal About the Mind. University of Chicago Press. ISBN 978-0-226-46804-4Lakoff, G., & Johnson, M. (1999). Philosophy in the Flesh: The Embodied Mind and Its Challenge to Western Thought. Basic Books. ISBN 978-0-465-05674-3.Lenat, D., & Guha, R. V. (1989). Building Large Knowledge-Based Systems. Addison-Wesley. ISBN 978-0-201-51752-1. OCLC 19981533Levitt, G. M. (2000). The Turk, Chess Automaton. Jefferson, N.C.: McFarland. ISBN 978-0-7864-0778-1Lighthill, P. S. (1973). "Artificial Intelligence: A General Survey". In Artificial Intelligence: A Paper Symposium. Science Research CouncilLucas, J. (1961). "Minds, Machines and Gödel". Philosophy, 36(XXXVI), 112–127. doi:10.1017/S0031819100057983. S2CID 55408480Luger, G., & Stubblefield, W. (2004). Artificial Intelligence: Structures and Strategies for Complex Problem Solving (5th ed.). Benjamin/Cummings. ISBN 978-0-8053-4780-7. Accessed 17 December 2019.Maker, M. H. (2006). AI@50: AI Past, Present, Future. Dartmouth College. Archived from the original source on 8 October 2008. Accessed 16 October 2008Markoff, J. (2005, October 14). "Behind Artificial Intelligence, a Squadron of Bright Real People". The New York Times. Accessed 16 October 2008McCarthy, J., Minsky, M., Rochester, N., & Shannon, C. (1955, August 31). A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence. Archived from the original on September 30, 2008, and retrieved on October 16, 2008.McCarthy, J., & Hayes, P. J. (1969). Some philosophical problems from the standpoint of artificial intelligence. In B. J. Meltzer & D. Mitchie (Eds.), Machine Intelligence 4 (pp. 463–502). Edinburgh University Press. Retrieved on October 16, 2008.McCarthy, J. (1974). Review of Lighthill report.McCorduck, P. (2004). Machines Who Think (2nd ed.). Natick, MA: A. K. Peters, Ltd. ISBN 978-1-56881-205-2. OCLC 52197627.McCullough, W. S., & Pitts, W. (1943). A logical calculus of the ideas immanent in nervous activity. Bulletin of Mathematical Biophysics, 5(4), 115–127. Bibcode: 1943BMaB....5..115M. doi:10.1007/BF02478259.Menabrea, L. F., & Lovelace, A. (1843). Sketch of the Analytical Engine Invented by Charles Babbage. Scientific Memoirs, 3. Retrieved on August 29, 2008. With notes upon the Memoir by the Translator.Minsky, M. (1967). Computation: Finite and Infinite Machines. Englewood Cliffs, N.J.: Prentice-Hall.Minsky, M., & Papert, S. (1969). Perceptrons: An Introduction to Computational Geometry. The MIT Press. ISBN 978-0-262-63111-2. OCLC 16924756.Minsky, M. (1974). A Framework for Representing Knowledge. Archived from the original on January 7, 2021, and retrieved on October 16, 2008.Minsky, M. (1986). The Society of Mind. Simon and Schuster. ISBN 978-0-671-65713-0. OCLC 223353010.Minsky, M. (2001). It's 2001. Where Is HAL?. Dr. Dobb's Technetcast. Retrieved on August 8, 2009.Moor, J. (Ed.). (2003). The Turing Test: The Elusive Standard of Artificial Intelligence. Dordrecht: Kluwer Academic Publishers. ISBN 978-1-4020-1205-1.Moravec, H. (1976). The Role of Raw Power in Intelligence. Archived from the original on March 3, 2016, and retrieved on October 16, 2008.Moravec, H. (1988). Mind Children. Harvard University Press. ISBN 978-0-674-57618-6. OCLC 245755104.Mulvihill, M. (2012, October 17). 1907: Was the first portable computer design Irish? Ingenious Ireland.Needham, J. (1986). Science and Civilization in China: Volume 2. Taipei: Caves Books Ltd.Newell, A., & Simon, H. A. (1995). GPS: A program that simulates human thought. In E. Feigenbaum & J. Feldman (Eds.), Computers and Thought (Reprinted from 1963 ed.). New York: McGraw-Hill. ISBN 978-0-262-56092-4. OCLC 246968117.Newquist, H. P. (1994). The Brain Makers: Genius, Ego, And Greed in the Quest For Machines That Think. New York: Macmillan/SAMS. ISBN 978-0-9885937-1-8. OCLC 313139906.NRC (National Research Council). (1999). Developments in Artificial Intelligence. In Funding a Revolution: Government Support for Computing Research. National Academy Press. ISBN 978-0-309-06278-7. OCLC 246584055.Nick, M. (2005). Al Jazari: The Ingenious 13th Century Muslim Mechanic. Al Shindagah. Retrieved on October 16, 2008.Nilsson, N. (2009, October 30). The Quest for Artificial Intelligence. Cambridge University Press. ISBN 978-0-52-112293-1.O'Connor, K. M. (1994). The alchemical creation of life (takwin) and other concepts of Genesis in medieval Islam (Doctoral dissertation, University of Pennsylvania). pp. 1–435. Retrieved on January 10, 2007.Olsen, S. (2004, May 10). Newsmaker: Google's man behind the curtain. CNET. Retrieved on October 17, 2008.Olsen, S. (2006, August 18). Spying an intelligent search engine. CNET. Retrieved on October 17, 2008.Pearl, J. (1988). Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference. San Mateo, California: Morgan Kaufmann. ISBN 978-1-55860-479-7. OCLC 249625842.Poole, D., Mackworth, A., & Goebel, R. (1998). Computational Intelligence: A Logical Approach. Oxford University Press. ISBN 978-0-19-510270-3.Pollack, A. (1984, October 11). "Technology, Fuzzy Logic for Computers." The New York Times.Pollack, A. (1989, April 2). "Fuzzy Computer Theory: How to Mimic the Mind?" The New York Times.Quevedo, L. T. (1914). "Journal of the Academy of Exact Sciences." Essays on Automation – Its Definition. Theoretical Extension of its Applications, vol. 12, pp. 391–418.Quevedo, L. T. (1915). "General Review of Pure and Applied Sciences." Essays on Automation – Its Definition. Theoretical Scope of its Applications, vol. 2, pp. 601–611.Randall, B. (1982). "From Analytical Engine to Electronic Digital Computer: The Contributions of Ludgate, Torres, and Bush." IEEE Annals of the History of Computing, 4 (4): 327. Bibcode:1982IAHC....4d.327R. doi:10.1109/MAHC.1982.10042.Russell, S. J., & Norvig, P. (2003). Artificial Intelligence: A Modern Approach (2nd ed.). Upper Saddle River, New Jersey: Prentice Hall. ISBN 0-13-790395-2.Russell, S. J., & Norvig, P. (2021). Artificial Intelligence: A Modern Approach (4th ed.). Hoboken: Pearson. ISBN 978-0-13-461099-3. LCCN 20190474.Samuel, A. L. (1959, July). "Some studies in machine learning using the game of checkers." IBM Journal of Research and Development, 3 (3): 210–219. CiteSeerX 10.1.1.368.2254. doi:10.1147/rd.33.0210. S2CID 2126705.Saygin, A. P., Cicekli, I., & Akman, V. (2000). "Turing Test: 50 Years Later" (PDF). Minds and Machines, 10 (4): 463–518. doi:10.1023/A:1011288000451. hdl:11693/24987. S2CID 990084.Searle, J. (1980). "Minds, Brains and Programs." Behavioral and Brain Sciences, 3 (3): 417–457. doi:10.1017/S0140525X00005756.Simon, H. A., & Newell, A. (1958). "Heuristic Problem Solving: The Next Advance in Operations Research." Operations Research, 6: 1–10. doi:10.1287/opre.6.1.1.Simon, H. A. (1965). The Shape of Automation for Men and Management. New York: Harper & Row.Skillings, J. (2006). Newsmaker: Getting machines to think like us. CNET..Tascarella, P. (2006, August 14). "Robotics firms find fundraising struggle, with venture capital shy." Pittsburgh Business Times..Turing, A. (1936–1937). "On Computable Numbers, with an Application to the Entscheidungsproblem." Proceedings of the London Mathematical Society, 2, 42 (42): 230–265. doi:10.1112/plms/s2-42.1.230. S2CID 73712.Turing, A. (1950, October). "Computing Machinery and Intelligence." Mind, 59 (236): 433–460. doi:10.1093/mind/LIX.236.433. ISSN 1460-2113. JSTOR 2251299. S2CID 14636783.Turkle, S. (1984). The Second Self: Computers and the Human Spirit. Simon and Schuster. ISBN 978-0-671-46848-4. OCLC 895659909.Wason, P. C., & Shapiro, D. (1966). "Reasoning." In B. M. Foss (Ed.), New Horizons in Psychology. Harmondsworth: Penguin...Weizenbaum, J. (1976). Computer Power and Human Reason. W.H. Freeman & Company. ISBN 978-0-14-022535-8. OCLC 10952283.Source: TORIma Academy Archive