The First AI Boom

The Road to Artificial Intelligence — Part 6 of 15

WHEN AI STARTED TO LOOK POSSIBLE

By the early 1960s, artificial intelligence had become more than a field name and a collection of foundational ideas. Laboratories, funding programs, programming tools, and interactive computing systems gave researchers new opportunities to build working demonstrations.

Computers began doing things that seemed astonishing. They solved algebra problems written in a limited form of English, carried on conversations, answered questions about a miniature world of blocks, planned routes for a mobile robot, and helped chemists reason about molecular structures.

These systems were genuine achievements. They also created a recurring problem in the history of artificial intelligence: observers could mistake a demonstration in a carefully bounded environment for evidence of a much broader kind of intelligence.

The first AI boom, then, is not simply a story of success or failure. It concerns the relationship between what a system could actually do, what its creators hoped it might do, and what observers inferred from impressive demonstrations.

A new kind of machine

By the early 1960s, artificial intelligence was becoming an institutional field.

At MIT, Project MAC—short for "Multiple Access Computer" and "Machine-Aided Cognition"—was founded on July 1, 1963. MIT's coordinated AI research began with the Artificial Intelligence Project in 1959; that group joined Project MAC in 1963 and split from it in 1970 to form the Artificial Intelligence Laboratory. At Stanford, John McCarthy initiated an AI project in 1962, obtained ARPA support beginning in 1963, and established the Stanford Artificial Intelligence Laboratory in 1965. Edinburgh's AI research began with a small group established by Donald Michie in 1963; the university created the Experimental Programming Unit in 1965 and the Department of Machine Intelligence and Perception in 1966.

These laboratories benefited from an expanding computing infrastructure. A prototype of the Compatible Time-Sharing System, developed at the MIT Computation Center under Fernando Corbató with Marjorie Merwin Daggett and Robert Daley, was working on an IBM 709 by November 1961. Time-sharing allowed several people to interact with a computer through terminals rather than waiting for one long batch job to finish.

The change gave researchers a practical, interactive environment for experimenting with programs, representations, language, and problem-solving. The claim that time-sharing transformed AI research is an interpretation rather than a simple measurement, but its importance is clear: researchers could increasingly treat computers as interactive research instruments rather than distant calculation machines.

The Advanced Research Projects Agency's Information Processing Techniques Office, created in 1962 under its first director, J. C. R. Licklider, supported work in time-sharing, computer graphics, networking, and artificial intelligence. Its portfolio helped connect ambitious research groups and computing projects. Such support, however, did not remove AI's central difficulty: a program that succeeded in one carefully described setting might fail completely when the setting changed.

GPS and the promise of general procedures

The General Problem Solver, or GPS, developed by Allen Newell and Herbert Simon, became an influential expression of early AI's ambitions.

GPS was designed both as a problem-solving program and as a possible computational model of human thought. It used heuristic methods, including means–ends analysis: compare the current state with a desired state, identify the differences, and choose actions that might reduce them.

The idea was powerful because it suggested that a common problem-solving procedure could work across several tasks. GPS, however, operated within a "narrow but significant problem domain." It depended on formal descriptions of problems and on knowledge supplied in a form the program could manipulate.

A general method applied to formally represented problems was still different from general intelligence in the open-ended human sense.

Language in restricted forms

Other systems explored a question that interested many researchers: could computers understand ordinary language?

Daniel Bobrow's STUDENT accepted a restricted subset of English and converted algebra word problems into internal representations and equations. Within that narrow domain, it could solve problems that appeared linguistic on the surface.

STUDENT's achievement was real. It showed that language could be connected to formal reasoning when the vocabulary, grammar, and subject matter were constrained. Broad comprehension was another matter. The system did not navigate the full ambiguity, background knowledge, and flexibility of everyday language.

The same lesson appeared more dramatically in Joseph Weizenbaum's 1966 program ELIZA. Its DOCTOR script used keywords and predetermined transformations to produce replies that resembled conversation. It had no general model of the speaker or the world.

Nevertheless, some users responded as though it understood them. The gap between mechanism and impression became an enduring theme of AI: people can attribute understanding to a system based on the form of its responses, even when the underlying process is far more limited.

SHRDLU's miniature world

Terry Winograd's SHRDLU took the idea further.

SHRDLU processed English dialogue, answered questions, executed commands, maintained a model of a blocks world, and reasoned about plans and actions. A user could ask it to manipulate objects in the simulated environment, and the program could relate language to the objects and operations represented there.

It was a substantial integration of capabilities. Language, representation, inference, planning, and action were connected in a single system.

The world, however, was deliberately small and formal. The program's apparent understanding depended heavily on having the possible objects, properties, and actions carefully defined in advance. SHRDLU demonstrated what could be achieved inside a well-structured micro-world—not unrestricted understanding of the physical or social world.

Translation and the danger of a compelling demonstration

Machine translation offered an early example of how a small demonstration could generate large expectations.

In January 1954, the Georgetown–IBM experiment translated a limited set of Russian sentences into English using approximately 250 words and six grammatical rules. The demonstration was important, but it did not represent general translation ability.

Its influence extended beyond the system's actual capabilities. The experiment helped foster expectations that machine translation might soon become a practical replacement for human translators.

The 1966 ALPAC report challenged those expectations. Rather than declaring machine translation impossible or ending research into language, it judged existing general machine translation inadequate for the uses evaluated and recommended continued support for computational linguistics and machine-assisted translation.

The report contributed to a retreat from large-scale expectations for general machine translation. Contemporary reporting indicates that the National Science Foundation discontinued direct support through one office, although language research continued in other forms.

The episode illustrates a recurring pattern: a technical demonstration attracts optimism, while later evaluation reveals the distance between a controlled result and robust real-world capability.

Seeing the world

Computer vision produced similar tensions.

Seymour Papert's 1966 "Summer Vision Project" proposed breaking visual perception into problems such as figure–ground analysis, region description, and object identification. The memo expressed considerable optimism about decomposing vision into manageable subproblems.

Contrary to the popular version of the story, it did not literally claim that human vision could be solved in one summer. The more important historical point is that vision appeared tractable when described as a collection of identifiable computational tasks.

Perception proved difficult to reduce to a sequence of isolated operations. Recognizing objects depends on context, prior knowledge, changing viewpoints, incomplete information, and the physical structure of the world. The project's ambitions were not achieved in the straightforward way that early optimism implied, and computer vision remained a difficult research problem.

Shakey enters the physical world

From 1966 through 1972, the Artificial Intelligence Center at the Stanford Research Institute—now SRI International—conducted the Shakey mobile-robot project.

Shakey integrated perception, planning, navigation, and physical action in a controlled environment. It could reason about tasks, plan sequences of actions, and move through its surroundings.

Participants and later institutional histories widely describe Shakey as an early mobile robot integrating AI planning with physical action. Its achievement was important because it showed that symbolic reasoning could be linked to a physical machine.

Shakey nevertheless operated in a limited and structured environment. Its autonomy differed from the open-ended autonomy associated with modern robots operating in unpredictable homes, streets, or workplaces. Its significance lies both in the integration it achieved and in the boundaries that made that integration possible.

Planning becomes explicit

Richard Fikes and Nils Nilsson's STRIPS formalized planning through operators with preconditions and effects.

An action could be represented in terms of what had to be true before it occurred and how it changed the state of the world afterward. This framework connected theorem proving, search, planning, and robot control.

STRIPS did not solve planning in general. Search could become expensive, and the system depended on an explicit representation of the relevant world. Even so, it provided a durable language for describing action and change—one that influenced later planning systems and robotics.

Knowledge-intensive expertise

DENDRAL applied AI techniques to a specialized scientific domain: organic chemistry.

Rather than relying solely on general reasoning procedures, DENDRAL used detailed chemical knowledge and heuristic search to generate explanatory hypotheses about molecular structures. Its performance demonstrated the value of encoding expertise in a domain-specific system.

DENDRAL became one of the earliest major examples of what would later be called expert systems.

The lesson was significant. Progress did not always require making a system broadly intelligent. It could also come from giving a system deep, carefully organized knowledge of one field.

The limits of perceptrons

In 1969, Marvin Minsky and Seymour Papert published Perceptrons: An Introduction to Computational Geometry.

The book analyzed important limitations of particular perceptron architectures. It showed that some apparently simple tasks lay beyond the capabilities of those models without additional structure.

It did not prove that all neural networks were useless, and the book alone should not be said to have "killed" neural-network research. Its influence on research priorities and funding was part of a broader story that also included hardware limitations, competing approaches, and changing research priorities.

Still, the episode reinforced a broader shift in emphasis. Symbolic representations, search, planning, and knowledge engineering appeared to offer promising ways of building systems that could reason in restricted domains.

The Connection to AI

One of early AI's most important contributions was a collection of working demonstrations showing that computers could manipulate symbols, solve formal problems, process restricted language, plan actions, and support scientific reasoning.

These systems established several enduring ideas:

• Problems could be represented in forms a machine could manipulate.
• Heuristics could guide search through large spaces of possibilities.
• Language could be linked to formal structures.
• Planning could be described through goals, preconditions, and effects.
• Specialized knowledge could produce useful expert performance.
• Physical action could be connected to symbolic reasoning.

They also revealed a persistent limitation. Their apparent intelligence depended on boundaries: restricted language, formal problem descriptions, miniature worlds, controlled environments, or large amounts of human-provided knowledge.

These early systems did not lead directly to today's AI, but many of the questions they raised remain relevant. Today's systems differ greatly in data, computing power, learning methods, and scale. The old question nevertheless remains relevant: is a system demonstrating broad competence, or performing impressively within conditions that make the task manageable?

Looking Ahead

The first AI boom established laboratories, research methods, and systems whose influence extended well beyond this early period. It also widened the gap between compelling demonstrations and reliable performance outside carefully structured settings.

The next article, “When AI Hit the Wall,” examines what happened as technical limits became harder to ignore, critical evaluations accumulated, and funding priorities changed—contributing to the uneven downturn later remembered as the first AI winter.

SOURCES

• Newell, Allen; Simon, Herbert A. "GPS: A Program That Simulates Human Thought." 1961. Carnegie Mellon Digital Collections.
• Bobrow, Daniel G. "Natural Language Input for a Computer Problem Solving System." MIT Artificial Intelligence Memo 66, 1964. MIT DSpace.
• Weizenbaum, Joseph. "ELIZA—A Computer Program for the Study of Natural Language Communication Between Man and Machine." Communications of the ACM, 1966.
• Winograd, Terry. "Procedures as a Representation for Data in a Computer Program for Understanding Natural Language." MIT Artificial Intelligence Technical Report 235, 1971. MIT DSpace.
• Papert, Seymour. "The Summer Vision Project." MIT Artificial Intelligence Memo 100, 1966. MIT DSpace.
• National Research Council / Automatic Language Processing Advisory Committee. Language and Machines: Computers in Translation and Linguistics. 1966.
• Fikes, Richard E.; Nilsson, Nils J. "STRIPS: A New Approach to the Application of Theorem Proving to Problem Solving." Artificial Intelligence, 1971.
• Minsky, Marvin; Papert, Seymour. Perceptrons: An Introduction to Computational Geometry. MIT Press, 1969.
• Feigenbaum, Edward; Buchanan, Bruce; Sutherland, Georgia; associated collaborators. "Heuristic DENDRAL: A Program for Generating Explanatory Hypotheses in Organic Chemistry." 1969.
• Hutchins, W. John. "The Georgetown-IBM Experiment Demonstrated in January 1954." Proceedings of AMTA, 2004.
• MIT Computer Science and Artificial Intelligence Laboratory. "CSAIL 20/60 Timeline." Institutional history; accessed September 28, 2026.
• MIT News. "MIT's Computer Science, AI Labs Merge." March 21, 2003.
• McCarthy, John; Earnest, Lester; Reddy, D. Raj; Vicens, Pedro J. "A Computer with Hands, Eyes, and Ears." Stanford Artificial Intelligence Laboratory report, 1974; includes an institutional history of Stanford's AI project.
• Nilsson, Nils J. The Quest for Artificial Intelligence. Cambridge University Press, 2010.
• University of Edinburgh, School of Informatics. "History of Artificial Intelligence at Edinburgh." Institutional archive; accessed September 28, 2026.
• Walden, David; Van Vleck, Tom, eds. The Compatible Time-Sharing System (1961–1973): Fiftieth Anniversary Commemorative Overview. IEEE Computer Society, 2011.
• Corbató, Fernando J.; Merwin-Daggett, Marjorie; Daley, Robert C. "An Experimental Time-Sharing System." Proceedings of the 1962 Spring Joint Computer Conference, 1962, pp. 335–344.
• Defense Advanced Research Projects Agency. "Information Processing Techniques Office." Institutional history; accessed September 28, 2026.
• Nilsson, Nils J., ed. Shakey the Robot. SRI International, Technical Note 323, 1984.
• SRI International. "Shakey the Robot." Institutional history; accessed September 28, 2026.