Biographies · Computation, superintelligence and human control
When David Banks visited I. J. Good in December 1993, books, letters and calculations occupied almost every surface of his office. A small clearing on the desk remained for writing. The crowded room belonged to a mathematician whose work stretched from wartime codebreaking to a question that still troubles artificial intelligence: what happens when a machine becomes capable of designing its successor? Banks’s interview, published in 1996.
Importance of his work for artificial intelligence
Imagine a school robot that can suggest changes to its own instructions. One change makes it sort objects faster. The next gives it permission to move through a shared corridor. Who checks that the improvement is real, that the robot still respects the rules and that someone can stop it? This example is much simpler than Good’s imagined machine, but it introduces the distinction between increasing capability and retaining control.
Good gave an early, explicit formulation of the possibility that an exceptionally capable machine could design still better machines. That prospect connects his work with the history of AI research and with questions of supervision. His contribution invites a further question for every reader: what evidence would be needed before accepting such a prediction? His argument, section 2, p. 33.
A child who wanted to work it out
Irving John Good was born in London in 1916 and was usually known as Jack. In his recollection, a childhood illness left him time to investigate the square root of two. He found his own way to a result mathematicians had known for centuries: it cannot be expressed as a ratio of whole numbers. The interest of the episode is the child’s independent reasoning. It is a personal recollection, rather than a surviving examination of his childhood work. Good’s account of his childhood.
At Cambridge, his mathematical training included work with A. S. Besicovitch and G. H. Hardy. He received his doctorate in 1941. Chess was another serious interest: he had won the Cambridgeshire championship in 1939. His early publications concerned pure mathematics, including continued fractions and measurable functions. The later public image of a man predicting intelligent machines can obscure that mathematical foundation. His curriculum vitae · His early publications.
Working with Turing under wartime pressure
In 1941, Good joined the codebreakers at Bletchley Park. In Hut 8 he worked with Alan Turing on German naval Enigma. Later, under Max Newman, he worked on the different teleprinter ciphers associated with the Colossus machines. These were collective undertakings involving mathematicians, engineers and operators. Good helped develop mathematical methods for using the machines effectively; Tommy Flowers led the engineering team. GCHQ’s historical account · Good’s account of the teams.
That setting makes a useful distinction visible. A calculation can be executed correctly while the conclusion drawn from incomplete evidence remains uncertain. Codebreaking required both computational procedures and judgments about which possibilities the available evidence supported. Good’s interest in that second problem became a defining part of his career. GCHQ records that he continued to champion Bayesian reasoning throughout his life. Evidence and cryptanalysis.
What had not yet been observed
In 1953, Good published work on estimating population frequencies from a sample, developing ideas associated with Turing. A basic difficulty is easy to picture: a class records birds visiting a garden, but some species appear only once and others never appear during the observation period. The recorded counts do not describe the entire population perfectly. Good’s paper examines estimation under such incomplete observation, including the treatment of rare and unobserved categories. The original 1953 paper.
The example also exposes the limit. An estimate of what has not been seen depends on assumptions about the observations and their relationship to the wider world. It is not a deductive guarantee that a particular unseen event will occur. A reader can recognise the importance of this mathematical work while asking how its assumptions are justified and how its errors are assessed. Those questions should remain visible whenever an AI system draws conclusions from examples.
Making a calculation more efficient
Good also contributed to efficient Fourier computation. His 1958 paper, The Interaction Algorithm and Practical Fourier Analysis, belongs to the development of fast Fourier methods. In his later interview, he distinguished his approach from the method published by Cooley and Tukey in 1965. The history involves several contributors and different mathematical constructions. The 1958 paper · His explanation, pp. 15–16.
This provides a helpful contrast within his own work. A method for calculating a transform can be specified and its output checked against the mathematical operation it implements. A prediction about the future of intelligence needs additional evidence about machines, resources and behaviour. Both deserve examination, but they require different kinds of support.
The machine that could design another machine
In Speculations Concerning the First Ultraintelligent Machine, Good defined a prospective machine that would greatly exceed every human intellectual activity. Since designing machines is one such activity, he envisaged a succession of improved machines and an intelligence explosion. His famous conclusion about humanity’s last necessary invention was conditional on retaining control. The paper explicitly offered speculation, without an engineering blueprint. The original argument, p. 33.
The article occupies pages 31–88 of volume 6 of Advances in Computers. Good’s own bibliography dates it to 1965, as does the copyright notice on the surviving reprint; the publisher’s current catalogue labels the volume 1966. The first-page note identifies a draft completed in April 1963 and revised in May 1964. Author’s bibliography, entry 397 · Three-page reprint extract · Publisher’s record.
A forecast that missed its century
Good considered it more probable than not that an ultraintelligent machine would be built within the twentieth century. No machine meeting that description had been demonstrated by the century’s end. The expected timetable failed. That outcome does not, by itself, settle whether such a machine could ever be built, but it belongs beside the celebrated prediction. Conclusions, p. 78.
The deeper difficulty is logical. A machine that exceeds human designers need not be able to improve indefinitely upon itself. A first successful improvement does not establish that every subsequent improvement will be possible, larger or quicker. Each step might demand more computation, physical experiments, manufacturing or time to detect mistakes. The argument identifies a consequential possibility; it does not demonstrate an unavoidable sequence.
There is a second gap between capability and control. A system’s ability to produce a clever design does not establish that its account of the design is complete, that its objectives remain acceptable or that its operators can interrupt it. Returning to the school robot, an impressive new movement is only one part of the assessment. The rules governing where it can go, the record of what changed and a tested means of stopping it matter as well. These are criteria for evaluating the proposal, not achievements that the historical article demonstrated.
Kubrick, Virginia and a life of questions
Good’s ideas reached beyond mathematical journals. Stanley Kubrick engaged him as a technical consultant on HAL 9000 for 2001: A Space Odyssey. Virginia Tech preserves a photograph of him at Hawk Films in 1966. The consultancy brought his scientific interests into a different kind of work, where the relationship between people and an intelligent machine became a story for a cinema audience. The consultancy · Virginia Tech’s photographic record.
His postwar appointments included Manchester, GCHQ, the Admiralty Research Laboratory, Trinity College, Oxford, and the Atlas Computer Laboratory at Chilton. In 1967 he moved to Virginia Tech, where he became a University Distinguished Professor and retired formally in 1994. He died on 5 April 2009. Career record · Departmental memorial.
Reading Good with both curiosity and discipline
Bayesian reasoning remained central to Good’s scientific outlook. It addresses uncertainty through explicit assumptions and calculations whose suitability must themselves be examined. A procedure can be reproducible while its conclusion remains conditional on an imperfect model of the world. Scientific scrutiny asks whether the assumptions are appropriate, whether the calculations are sound and whether the conclusions survive testing.
Good’s place in the history of AI includes mathematical results, practical wartime work and bold conjectures. Keeping those categories distinct makes his story more useful. A young reader can learn from his willingness to ask an ambitious question while also learning to demand evidence for the answer.
Today, his argument provides historical context for the questions considered in this collection’s biographies of Leopold Aschenbrenner and Ilya Sutskever. A projected increase in intelligence brings the same practical demands into view: who can examine the evidence, who authorises the next step and who remains able to stop it?
Milestones
- 1916Born in London.
- 1941Completes his Cambridge doctorate and joins Bletchley Park.
- 1953Publishes his paper on population frequencies and probability estimation.
- 1958Publishes his work on practical Fourier analysis.
- 1965Date given in his bibliography for the ultraintelligent-machine essay.
- 1967Joins Virginia Tech.
- 1994Becomes professor emeritus.
- 2009Dies on 5 April.
Further reading
David L. Banks, A Conversation with I. J. Good (1996) · Speculations Concerning the First Ultraintelligent Machine, complete article · Good’s curriculum vitae · His bibliography.
Historical sources checked on 21 November 2026.
Liked this post? Follow this blog to get more.