2006: Celebrating 75 years of AI - History and Outlook: the Next 25 Years

Zuse and Turing. In 1936 Alan Turing [37] introduced the Turing machine to reformulate Gödel’s results and Alonzo Church’s extensions thereof. TMs are often more convenient than Gödel’s integer-based formal systems, and later became a central tool of CS theory. Simultaneously Konrad Zuse built the first working program-controlled computers (1935)(1936)(1937)(1938)(1939)(1940)(1941), using the binary arithmetic and the bits of Gottfried Wilhelm von Leibniz (1701) instead of the more cumbersome decimal system used by Charles Babbage, who pioneered the concept of program-controlled computers in the 1840s, and tried to build one, although without success. By 1941, all the main ingredients of ‘modern’ computer science were in place, a decade after Gödel’s paper, a century after Babbage, and roughly three centuries after Wilhelm Schickard, who started the history of automatic computing hardware by constructing the first non-program-controlled computer in 1623.

In the 1940s Zuse went on to devise the first high-level programming language (Plankalkül), which he used to write the first chess program. Back then chess-playing was considered an intelligent activity, hence one might call this chess program the first design of an AI program, although Zuse did not really implement it back then. Soon afterwards, in 1948, Claude Shannon [33] published information theory, recycling several older ideas such as Ludwig Boltzmann’s entropy from 19th century statistical mechanics, and the bit of information (Leibniz, 1701).

Relays, Tubes, Transistors. Alternative instances of transistors, the concept pioneered and patented by Julius Edgar Lilienfeld (1920s) and Oskar Heil (1935), were built by William Shockley, Walter H. Brattain & John Bardeen (1948: point contact transistor) as well as Herbert F. Mataré & Heinrich Walker (1948, exploiting transconductance effects of germanium diodes observed in the Luftwaffe during WW-II). Today most transistors are of the field-effect type à la Lilienfeld & Heil. In principle a switch remains a switch no matter whether it is implemented as a relay or a tube or a transistor, but transistors switch faster than relays (Zuse, 1941) and tubes (Colossus, 1943;ENIAC, 1946). This eventually led to significant speedups of computer hardware, which was essential for many subsequent AI applications.

The I in AI. In 1950, some 56 years ago, Turing invented a famous subjective test to decide whether a machine or something else is intelligent. 6 years later, and 25 years after Gödel’s paper, John McCarthy finally coined the term “AI”. 50 years later, in 2006, this prompted some to celebrate the 50th birthday of AI, but this chapter’s title should make clear that its author cannot agree with this view-it is the thing that counts, not its name.

Roots of Probability-Based AI. In the 1960s and 1970s Ray Solomonoff combined theoretical CS and probability theory to establish a general theory of universal inductive inference and predictive AI [35] closely related to the concept of Kolmogorov complexity [14]. His theoretically optimal predictors and their Bayesian learning algorithms only assume that the observable reactions of the environment in response to cer-tain action sequences are sampled from an unknown probability distribution contained in a set M of all enumerable distributions. That is, given an observation sequence we only assume there exists a computer program that can compute the probabilities of the next possible observations. This includes all scientific theories of physics, of course. Since we typically do not know this program, we predict using a weighted sum ξ of all distributions in M, where the sum of the weights does not exceed 1. It turns out that this is indeed the best one can possibly do, in a very general sense [11,35]. Although the universal approach is practically infeasible since M contains infinitely many distributions, it does represent the first sound and general theory of optimal prediction based on experience, identifying the limits of both human and artificial predictors, and providing a yardstick for all prediction machines to come.

AI vs Astrology? Unfortunately, failed prophecies of human-level AI with just a tiny fraction of the brain’s computing power discredited some of the AI research in the 1960s and 70s. Many theoretical computer scientists actually regarded much of the field with contempt for its perceived lack of hard theoretical results. ETH Zurich’s Turing award winner and creator of the PASCA

…(Full text truncated)…