Part of Temporal Constructs Appendix, which contains details behind the research: dates, quotations, descriptions of instruments, and sources.

General definition

The field of computer science whose systems perform tasks associated with intelligence, among them reasoning, learning and the use of language (Russell and Norvig). Its systems divide into those whose behaviour follows rules written down by a person and those whose behaviour is learned from data (Pasquale 2015). The name has been in use since 1955, and it has covered different work in each decade since.

Relation to the practice

The doctorate works inside this field and draws a line through it. It uses agents that act on stated rules in the symbolic domain, and it excludes models trained on finished music. A rule-based system can be read in full and a learned one cannot (Legibility).

Name and first use

The phrase artificial intelligence was written down in a proposal of 31 August 1955, by John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon, for a summer research project at Dartmouth College. The proposal asked for “a 2 month, 10 man study of artificial intelligence” and stated the conjecture the study would proceed on: “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”. The meeting was held in the summer of 1956. Those who came included Allen Newell and Herbert Simon, Oliver Selfridge, Ray Solomonoff and Trenchard More, and the beginning of the field is dated to that summer.

The phrase was new in 1955, and the question it named was older. Alan Turing had asked in 1950 whether machines could think, and had set that question aside in favour of a game of imitation, in which a person and a machine are distinguished by a questioner who cannot see either. Warren McCulloch and Walter Pitts had described the neuron as a logical device in 1943, and Norbert Wiener had published Cybernetics in 1948. Three of the proposal’s seven headings concern machines that learn, and one of them, neuron nets, names the line of work that returned to prominence sixty years later.

Name through the decades

YearsWhat the name coveredWhat carried it
1956 to the early 1970sreasoning, as the manipulation of symbols by stated rulesthe Logic Theorist, the General Problem Solver, LISP
the 1970sthe same claim, and a withdrawal of fundingthe Lighthill report of 1973, the first winter of 1974 to 1980
the 1980sknowledge, and systems that hold an expert’s rulesMYCIN, XCON, the Fifth Generation project from 1982
1986 to the 1990slearning from data, by networks trained on examplesbackpropagation in 1986, the second winter of 1987 to 1993
the 2000s and 2010sstatistical learning over large collections of dataAlexNet in 2012, AlphaGo in 2016
2017 to the presentmodels that generate text, images and soundthe transformer in 2017, the large language model

Name in the MIDI era

The years this doctorate calls the MIDI era, about 1983 to the late 1990s, are the same years in which the academic field’s centre of gravity was knowledge: systems that held an expert’s rules, sold as expert systems, and the Fifth Generation project begun in Japan in 1982. Funding for that work was withdrawn between 1987 and 1993, and many researchers then called their work by other names, among them informatics, knowledge-based systems, cognitive systems and computational intelligence. A New York Times report in 2005 said that computer scientists and software engineers had avoided the term artificial intelligence for fear of being viewed as wild-eyed dreamers.

In music software the word intelligent meant something narrower and more concrete. Laurie Spiegel gave the phrase intelligent instruments public currency with Music Mouse in 1986, and it named programs that answer a performer with musical decisions of their own: M and Jam Factory from Intelligent Music, HMSL, and Max, which turned programming into a visual form of play. These systems were built on rules, probabilities and patterns written by their authors, and they could be read in full because memory was expensive and the machines were small. Nothing in that period learned from a collection of recorded music, so the intelligence of the MIDI era was legible by construction, and the line this doctorate draws through the field runs along the same period.

Name in use now

The OECD’s Recommendation on Artificial Intelligence defined an AI system in 2019 as “a machine-based system that can, for a given set of human-defined objectives, make predictions, recommendations, or decisions influencing real or virtual environments”. The Council revised that sentence on 8 November 2023: “An AI system is a machine-based system that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments.” The European Union’s AI Act of 2024 built Article 3 on the revised wording and added two properties: the system is designed to operate with varying levels of autonomy, and it may exhibit adaptiveness after deployment.

The objectives may now be implicit, where the 2019 wording required them to be human-defined, and the outputs may be content, which names what generative models produce. A model trained on a large collection of music and asked for a piece of it falls inside the 2023 definition and outside the 2019 one.

Model families

A rule-based system can be inspected, because its behaviour comes from rules that were written down. A system trained on data can be tested and observed, and its internal states cannot be read as rules. The two traditions have names of their own: symbolic AI, called good old-fashioned AI by John Haugeland in 1985, and connectionism, which builds on networks of simple units. Both are in use, and most working systems now combine them.

Use of the name

The doctorate’s rules for its own use of these systems are on Guidelines for the use of AI. They exclude the closed services, keep the models local where the work allows, and keep the data in the practice’s hands. The reasons for those rules, and the places where the practice falls short of them, are recorded on that page. Models that produce sound, images or video from a prompt are described on Generative models.

Unforeseen

Here the unforeseen situations, sometimes difficult, sometimes unexpectedly interesting, will be described as the project runs. Many of these unforeseen situations have already occurred and have been recorded. They will be added in the time to come.