The summer of 1956 at Dartmouth College marked a turning point in human history—the official birth of artificial intelligence as a scientific field. What began as an ambitious proposal by John McCarthy, Marvin Minsky, Claude Shannon, and Nathaniel Rochester evolved into three decades of remarkable innovation that transformed theoretical concepts into working systems.
1951-1952: Machine Learning and Hardware Neural Networks
Arthur Samuel's Checkers Program (IBM): Developed at IBM Poughkeepsie Laboratory, first running on IBM 701 in 1952. Used minimax search with alpha-beta pruning, reducing search space by 99%. Implemented "rote learning" storing 53,000 positions and generalization learning adjusting 16 weight parameters. By 1955, the program could beat Samuel himself. In 1962, it defeated Connecticut state champion Robert Nealey. Samuel coined "machine learning" in his 1959 paper, defining it as "field of study that gives computers the ability to learn without being explicitly programmed." The program demonstrated that machines could improve through experience, fundamental to modern AI.
SNARC Neural Network (1951): Built by Marvin Minsky and Dean Edmonds at Princeton, the Stochastic Neural Analog Reinforcement Calculator used 3,000 vacuum tubes and 40 Hebb synapses (potentiometers adjusted by clutched motors). Each synapse had probability of reinforcement based on success/failure. Simulated rats learning to navigate mazes, demonstrating reinforcement learning in hardware. This was Minsky's doctoral project, inspiring his later work on perceptrons and AI.
Herbert Robbins' Stochastic Gradient Descent (1951): With Sutton Monro, developed the Robbins-Monro algorithm for stochastic approximation, providing mathematical foundation for iterative optimization under noise. The update rule: θn+1 = θn - αn∇f(θn) + noise, where learning rate αn decreases over time. This became fundamental to training neural networks, enabling optimization with noisy gradients from mini-batches.
1954: Early Natural Language Processing
Georgetown-IBM Experiment (January 7, 1954): Led by Leon Dostert (Georgetown) and Cuthbert Hurd (IBM), demonstrated at IBM World Headquarters, New York. Translated 60+ Russian sentences to English using IBM 701 with 250-word vocabulary and 6 grammar rules. Sample translation: "Mi pyeryedayem mislyi posryedstvom ryechyi" → "We transmit thoughts by means of speech." Generated headlines in major newspapers with predictions of solving translation in 3-5 years. Though overly optimistic, it launched machine translation research and secured significant government funding. The system used dictionary lookup and word reordering rules, lacking syntactic or semantic understanding.
Self-Organizing Systems (1954): Wesley Clark and Belmont Farley at MIT Lincoln Laboratory built the first self-organizing system, demonstrating pattern recognition through adaptive networks. Used analog computers to simulate neurons with adjustable connections, showing emergence of feature detectors through experience.
1956: The Dartmouth Conference - AI's Official Beginning
Dates: June 18 - August 17, 1956
Location: Dartmouth College, Hanover, New Hampshire
Organizers: John McCarthy (Dartmouth), Marvin Minsky (Harvard), Claude Shannon (Bell Labs), Nathaniel Rochester (IBM)
Funding: $7,500 from Rockefeller Foundation
Attendees: Including Allen Newell, Herbert Simon, Arthur Samuel, Oliver Selfridge, Ray Solomonoff
Proposal: "We propose that a 2-month, 10-man study of artificial intelligence be carried out... The study is to proceed on the basis of the conjecture that every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it."
Outcomes: First use of term "Artificial Intelligence" (coined by McCarthy), established AI as legitimate academic field, created research agenda including automatic computers, programming languages, neural networks, theory of computation, self-improvement, abstractions, and creativity. Though no breakthroughs occurred during the conference itself, it created the community and vision that drove subsequent AI development.
1955-1957: Early AI Programs
Logic Theorist (1955-1956): Created by Allen Newell, Herbert Simon, and J.C. Shaw at RAND Corporation and Carnegie Tech. Written in IPL (Information Processing Language), proved 38 of first 52 theorems in Whitehead and Russell's Principia Mathematica, with shorter proof for theorem 2.85. Used heuristic search with backward chaining, means-ends analysis, and pruning. First program to prove mathematical theorems, demonstrating machines could perform tasks requiring "intelligence." Journal of Symbolic Logic rejected a paper co-authored with the program.
General Problem Solver (GPS) (1957): Evolution of Logic Theorist by same team, designed as universal problem solver. Used means-ends analysis: identify differences between current and goal states, find operators reducing differences, and apply recursively. Could solve diverse problems: Tower of Hanoi, cryptarithmetic, logical proofs, and word problems. Introduced planning and subgoal decomposition but failed on complex real-world problems due to combinatorial explosion. Influenced STRIPS planner and modern planning algorithms.
1957-1958: The Perceptron
Developer: Frank Rosenblatt (Cornell Aeronautical Laboratory)
Theory Published: 1957, "The Perceptron: A Perceiving and Recognizing Automaton"
Implementation: 1958 on IBM 704, Mark I Perceptron hardware built 1960
Architecture: Single layer of artificial neurons with adjustable weights, threshold activation function, and supervised learning rule: Δw = α(target - output)×input. Could learn linearly separable patterns.
Mark I Specifications: 400 photocells arranged in 20×20 grid, weights implemented as potentiometers adjusted by electric motors, could distinguish simple shapes and letters after training.
Rosenblatt's Claims: Predicted perceptrons would eventually be able to recognize people, call out their names, translate languages, and write music. Navy funded research for automatic target recognition.
1969 Critique: Minsky and Papert's "Perceptrons" proved single-layer limitations (couldn't learn XOR), showing need for multiple layers but suggesting these were untrainable. This contributed to first AI Winter, though multilayer networks with backpropagation later solved these limitations.
1958-1959: LISP and MIT AI Lab
LISP Programming Language (1958): Created by John McCarthy at MIT, first implemented by Steve Russell on IBM 704. Revolutionary features included symbolic expression (S-expressions) as universal data structure, recursive function definitions, conditional expressions (IF-THEN-ELSE), automatic memory management with garbage collection, and interpretation allowing runtime code modification. Became dominant AI programming language for 30+ years. Innovations influenced all modern languages. Dialects included Common Lisp, Scheme, Clojure. Enabled rapid prototyping and symbolic reasoning essential for early AI research.
MIT AI Laboratory (1959): Founded as "Artificial Intelligence Project" by McCarthy and Minsky within Research Laboratory for Electronics. Initial funding of $25,000 from MIT, major DARPA support from 1963 ($2.2 million/year by 1970s). Became center of AI research with projects including MACSYMA (symbolic math), SHRDLU (natural language), vision systems, and robotics. Developed ITS (Incompatible Timesharing System) and EMACS editor. Split in 1970s between Minsky's symbolic AI and Patrick Winston's approach. Alumni founded major AI companies and research centers worldwide.
1960s: Growth and Early Applications
J.C.R. Licklider's "Man-Computer Symbiosis" (1960): ARPA/IPTO director envisioned interactive computing where humans and computers work together, each doing what they do best. Funded interactive computing research leading to time-sharing, computer graphics, and eventually the Internet.
SAINT (1961): James Slagle's Symbolic Automatic INTegrator at MIT could solve symbolic integration problems at calculus student level, using heuristic search and pattern matching, demonstrating AI for mathematical problem-solving.
UNIMATE (1961): First industrial robot installed at GM, developed by Unimation (George Devol and Joseph Engelberger). Weighed 4,000 pounds, used hydraulic actuators, programmable with magnetic drum memory. Beginning of industrial automation.
ELIZA (1964-1966): Joseph Weizenbaum's natural language processing program at MIT simulated Rogerian psychotherapist using pattern matching and substitution. Example:- User: "I am unhappy"- ELIZA: "Why do you think you are unhappy?"Simple technique created illusion of understanding, raising questions about nature of intelligence and human-computer interaction.
DENDRAL (1965-1968): Edward Feigenbaum, Bruce Buchanan, Joshua Lederberg at Stanford. First expert system for identifying molecular structures from mass spectrometry data. Encoded expertise of chemists in rules, performed better than graduate students, used generate-and-test paradigm. Launched expert systems field.
Shakey the Robot (1966-1972): SRI International's first mobile robot integrating perception, planning, and action. Used TV camera for vision, laser rangefinder, bump sensors, radio link to DEC PDP-10 and PDP-15 computers. Implemented STRIPS planner, A* search algorithm. Could navigate rooms, push blocks, turn lights on/off. Cost $750,000/year to develop.
Semantic Networks (1966): Ross Quillian's model for human memory using nodes (concepts) and labeled edges (relationships), enabling inference through network traversal. Influenced knowledge representation and modern knowledge graphs.
Hidden Markov Models (1966): Leonard Baum and colleagues at IDA developed statistical models for sequential data with hidden states and observable outputs. Became fundamental for speech recognition, natural language processing, and bioinformatics.
Mac Hack VI (1967): Richard Greenblatt's chess program at MIT, first to play tournament chess, achieving 1400 rating. Beat philosopher Hubert Dreyfus who had claimed computers couldn't play good chess.
1970s: Expert Systems and First AI Winter
PROLOG (1972): Alain Colmerauer and Philippe Roussel at University of Marseille created logic programming language based on first-order predicate calculus. Declarative programming with facts, rules, and queries, using unification and backtracking for inference. Became major AI language, especially in Europe and Japan.
MYCIN (1972-1976): Edward Shortliffe at Stanford created expert system for diagnosing bacterial infections and recommending antibiotics. Used 600 rules with certainty factors, explanation facility showing reasoning, and performed at expert physician level (69% accuracy vs 65% for human experts). Never clinically deployed due to liability concerns but inspired numerous expert systems.
Lighthill Report (1973): James Lighthill's report to UK Science Research Council criticized AI's failure to deliver promises, identifying combinatorial explosion, lack of common sense reasoning, and toy problem syndrome. Led to dramatic funding cuts in UK, contributing to "AI Winter" (1974-1980).
Frames Theory (1975): Marvin Minsky proposed knowledge representation using frames (structured records) with slots and default values, enabling inheritance and procedural attachment. Influenced object-oriented programming and modern knowledge representation.
FREDDY Robot (1973-1976): University of Edinburgh's assembly robot could identify and assemble wooden models using vision system. Took 10 minutes to recognize objects, demonstrating difficulty of real-world perception.
Fifth Generation Computer Project(1982): Japan's 10-year initiative to create parallel inference machines for AI, investing $850 million. Though failing to achieve goals, spurred international AI research competition.