Image: Marcin Wichary, CC BY 2.0, via Wikimedia Commons
The closing unit of the Foundation asks what the methods of the first seven units add up to. It reads Turing's 1950 paper and scores his imitation game with real numbers, tells the history of the two traditions and the two winters as a story of money, hardware, and data, and shows why 2012 was different, with the AlexNet numbers and the compute that has grown by eight orders of magnitude since. It presents the Chinese room argument and its replies fairly, examines what AlphaFold and the antibiotic halicin did and did not do, states the four questions that divide the field today without a verdict, and ends with a map of the eight units and the one equation under them.
Preview: the beginning of the unit
The seven units before this one taught methods: how to search, how to reason with probabilities, how to fit a line, how to train a network. This closing unit of the Foundation steps back and asks the questions that the methods raise. What would it mean for a machine to be intelligent? Why did a field with such confident founders run out of money twice? What happened in 2012 that had not happened in 1958 or 1986? And what did the models that fold proteins and find antibiotics do, exactly? The unit gives the answers that are settled, presents the disputes as disputes, and ends with a map of what the Foundation has taught.
1 What intelligence could mean for a machine
In October 1950 the journal Mind printed a paper by Alan Turing (1912-1954), shown in Figure 1, with the title “Computing Machinery and Intelligence”. Its first sentence proposes to consider the question whether machines can think, and the same paragraph gives the question up, because the words “machine” and “think” have no agreed meaning. In their place Turing proposed a game, and the game has shaped the discussion ever since.
Figure 1. Alan Turing at the age of 16, in a passport photograph of 1928. Twenty-two years later he replaced the question whether a machine can think by a game that can be played and scored. Credit: Wikimedia Commons, public domain.
Turing designed the game so that only behaviour counts. The typed channel removes the voice, the face, and the body, and the comparison with a hidden person removes the need to define intelligence in advance. Whatever the person does that the interrogator notices, the machine must do too.
The paper also made a prediction, and it is worth reading exactly. Turing wrote that in about fifty years it would be possible to program computers with a storage capacity of about \(10^{9}\) binary digits to play the game so well that an average interrogator would have no more than a 70 per cent chance of making the right identification after five minutes of questioning. A storage of \(10^{9}\) binary digits is about 125 megabytes, because a byte holds 8 binary digits, and a telephone with 128 gigabytes holds a thousand times as much. The success rate he asked for is modest, because the machine only has to fool the interrogator in 30 per cent of the games, and a machine that could not be told from a person at all would be identified correctly only half the time, by chance.