The first two decades of AI were a time of excitement, generous government funding and bold predictions. Researchers built programs that proved theorems, played games and held simple conversations. Many believed human-level intelligence was only a few decades away.
Symbolic AI
Most early AI followed what is now called the symbolic approach, sometimes nicknamed “good old-fashioned AI.” Its central idea was that intelligence consists of manipulating symbols according to rules, as in logic or mathematics. If knowledge could be written down as symbols and rules, a computer could reason with it.
In 1956, Allen Newell, Herbert Simon and Cliff Shaw created the Logic Theorist, often called the first AI program. It proved many theorems from a famous work of mathematical logic, Whitehead and Russell's Principia Mathematica, and for one theorem found a more elegant proof than the original. They followed it with the General Problem Solver, meant to solve any problem that could be stated formally. In 1958, John McCarthy created the programming language LISP, which became the main language of AI research for decades.
Bold predictions
Success bred confidence. In 1957, Herbert Simon predicted that within ten years a computer would be world chess champion. In 1965 he predicted that within twenty years machines would be capable of doing any work a human can do. Marvin Minsky, a founder of the MIT AI laboratory, made similar forecasts. These predictions proved far too optimistic. A computer did not beat the world chess champion until 1997, and general human-level intelligence remained out of reach for decades.
The perceptron
A different approach drew on the neural network idea. In 1958, the psychologist Frank Rosenblatt, working at Cornell with funding from the US Navy, built the perceptron, a machine that could learn to recognize simple patterns by adjusting the strengths of connections between artificial neurons based on its errors. The press was dazzled. A 1958 New York Times report described the Navy's expectation that such a machine would eventually walk, talk, see, write, reproduce itself and be conscious of its existence.
Machine learning gets a name
In 1959, Arthur Samuel at IBM described a program that played checkers and improved by playing against itself, eventually beating its creator. Samuel used the term “machine learning” for giving computers the ability to learn without being explicitly programmed. His program was one of the earliest demonstrations that a machine could get better through experience.
ELIZA
Between 1964 and 1966, the MIT computer scientist Joseph Weizenbaum created ELIZA, a program that held text conversations. Its most famous script, DOCTOR, imitated a psychotherapist by turning users' statements into questions: if a user typed “I am unhappy,” ELIZA might reply, “Why do you say you are unhappy?” The program understood nothing; it used simple pattern matching. Yet many users became emotionally engaged. Weizenbaum reported that his own secretary asked him to leave the room so she could talk to ELIZA privately.
Weizenbaum was disturbed. In his 1976 book Computer Power and Human Reason, he warned that people were too ready to attribute understanding to machines, and that some tasks involving human judgment and care should never be handed to computers. The tendency to read human qualities into simple programs is now called the ELIZA effect.
Robots and microworlds
From 1966 to 1972, researchers at the Stanford Research Institute built Shakey, the first mobile robot that could reason about its own actions, planning routes and pushing blocks around a set of rooms. At MIT, a program called SHRDLU, created around 1970, could understand and carry out typed commands about a simulated world of colored blocks. These systems worked well in simplified “microworlds” but struggled in the messy real world.
The limits appear
In 1969, Marvin Minsky and Seymour Papert published Perceptrons, a mathematical analysis showing that simple single-layer perceptrons could not learn certain basic functions, such as the logical operation called exclusive or. Many readers concluded that neural networks were a dead end, and funding for them dried up for over a decade, even though multi-layer networks could, in principle, overcome these limits.
Worked example: why were the predictions wrong?
Early AI researchers were brilliant. Why did they misjudge the difficulty so badly?
- Easy for humans, hard for machines. Researchers found that tasks hard for people, like proving theorems, were relatively easy for computers, while tasks easy for a child, like recognizing a face or walking across a room, were extremely hard. This became known as Moravec's paradox.
- Common sense. Human reasoning relies on vast unspoken knowledge about the world. Writing it all down as rules proved impossible.
- Computing power. Computers of the 1960s were millions of times slower, with far less memory, than those of today.
- Data. Learning systems needed large amounts of examples, which did not yet exist in digital form.
The lesson applies to many new technologies: early demonstrations in controlled settings can make a goal seem close when the hardest problems have not yet appeared. A sensible attitude to bold forecasts, then and now, is to ask what has actually been shown, under what conditions, and what still has to be solved.