Chapter 2
From hand-written rules to machines that learn
The chapter in a minute
- In 1950 Turing asked whether machines can think, and a 1955 proposal gave the field the name 'artificial intelligence'.
- Rule-based systems follow rules people wrote by hand, so they handle only the cases those rules cover.
- Expert systems such as MYCIN put an expert's knowledge into rules, but they were slow to build, costly to keep up and easily caught out by the unexpected. Hopes and funding fell more than once, each time for several reasons.
- Machine learning finds patterns in examples, so a spam filter that learns can be trained again when junk mail changes.
- The 1997 chess win came mostly from fast search and hand-written chess knowledge, and it did only one thing.
Try it now
Rules against examples
0 of 5 steps done
Step 1 of 5
Write five rules for sorting your mail into keep and bin. Use the form 'If this, then that'. Declan's first rule is 'If it says free, bin it'. His second is 'If it's from someone I know, keep it'.
Tools and settings, as of a date
The book points to the ‘Timeline’ section of the companion page. Here it is in full, with the date each detail was checked. Open it on the companion page
Chapters 2 and 3 both point here. Chapter 2 points to "further reading for this chapter". Chapter 3 points to "the names behind each milestone and further reading". Both are on this one section.
Chapter 2: further reading with a free copy
These are the sources that have an official or author-posted free copy. University course copies were used for checking only and are not linked.
- The Dartmouth proposal, 31 August 1955 (McCarthy, Minsky, Rochester and Shannon). Reprinted in AI Magazine 27(4), 15 December 2006, DOI 10.1609/aimag.v27i4.1904. McCarthy's own web copy (opens in a new tab).
- Nils J. Nilsson, "John McCarthy, 1927 to 2011: A Biographical Memoir" (National Academy of Sciences, pages 4 and 5): PDF from the National Academy of Sciences (opens in a new tab).
- The Lighthill report, "Artificial Intelligence: A General Survey" (Part I dated July 1972; published early 1973), in the Chilton Computing archive (opens in a new tab).
- Buchanan and Shortliffe (eds.), Rule-Based Expert Systems: The MYCIN Experiments of the Stanford Heuristic Programming Project (1984). Out of print; chapters posted free by the editor (opens in a new tab).
- Jon Agar, "What is science for? The Lighthill report on artificial intelligence reinterpreted", The British Journal for the History of Science 53(3), 10 July 2020, pp. 289 to 310, DOI 10.1017/S0007087420000230 (opens in a new tab). Free on the publisher site.
- IBM, "Deep Blue" (opens in a new tab) history page (used for the match month only).
Checked: 2 November 2026 (every link works).
Chapter 3: the names behind the milestones
Chapter 3 names no people or products in print. Here are the names and the papers. Every link was checked on 2 November 2026.
| Milestone | Who and what | Where to read it |
|---|---|---|
| 1943, a simple mathematical model of a neuron | Warren S. McCulloch and Walter Pitts, "A logical calculus of the ideas immanent in nervous activity" | First published in The Bulletin of Mathematical Biophysics 5(4), December 1943, pp. 115 to 133, DOI 10.1007/BF02478259 (opens in a new tab); reprinted in Bulletin of Mathematical Biology 52(1/2), 1990 |
| 1958, the perceptron described | Frank Rosenblatt, "The perceptron: A probabilistic model for information storage and organization in the brain". Described in 1958, not built then. | Psychological Review 65(6), 1958, pp. 386 to 408, DOI 10.1037/h0042519 (opens in a new tab) |
| 1986, training networks with hidden layers made widely known | David E. Rumelhart, Geoffrey E. Hinton and Ronald J. Williams, "Learning representations by back-propagating errors", published 9 October 1986 | Nature 323(6088), pp. 533 to 536, DOI 10.1038/323533a0 (opens in a new tab) |
| 2012, a deep network wins the image-recognition challenge | Alex Krizhevsky, Ilya Sutskever and Geoffrey E. Hinton, "ImageNet classification with deep convolutional neural networks". The network is known as AlexNet. Top-5 error 15.3% against 26.2% for the second-best entry. | NeurIPS 2012; republished in Communications of the ACM 60(6), 2017. The name and the original 2012 code: Computer History Museum, "CHM Releases AlexNet Source Code" (opens in a new tab), 20 March 2025, released with Google. |
| Around 1990, translation moves from rules to statistics | Peter F. Brown and colleagues, "A Statistical Approach to Machine Translation" | ACL Anthology (opens in a new tab) |
| Mid 2010s, neural translation | Bahdanau, Cho and Bengio (2014), arXiv 1409.0473 (opens in a new tab), then Google's Neural Machine Translation system (GNMT), Wu and colleagues (2016), arXiv 1609.08144 (opens in a new tab) | |
| 2017, the transformer design | Vaswani and colleagues, "Attention Is All You Need", first posted 12 June 2017 | arXiv 1706.03762 (opens in a new tab) |
| 2020, very large language models learn from a few examples | Brown and colleagues, "Language Models are Few-Shot Learners", first posted 28 May 2020. The paper's model is GPT-3 (OpenAI), which the abstract describes as having 175 billion parameters. | arXiv 2005.14165 (opens in a new tab) |
| Late 2022, a chat assistant reaches the public | OpenAI, "Introducing ChatGPT", 30 November 2022 | OpenAI (opens in a new tab) |
| 2016, a program beats a top Go professional (optional milestone) | Google DeepMind's AlphaGo: a 4 to 1 win in Seoul in March 2016. The date and score come from the developer's own page, which carries no date of its own, so read them as the developer's account. | Google DeepMind, AlphaGo (opens in a new tab) |
Checked: 2 November 2026. DeepMind's page still gives the 4 to 1 score and March 2016 in Seoul.
Chapter 3: how quickly use grew
- Pew Research Center, 25 June 2025. 34% of U.S. adults said they had ever used ChatGPT, roughly double the share in summer 2023. Survey of 5,123 U.S. adults, 24 February to 2 March 2025, American Trends Panel, recruited by random sampling of residential addresses and weighted to the U.S. adult population. One product, "ever used", U.S. only. Pew, 25 June 2025 (opens in a new tab).
- Pew Research Center, 17 June 2026 (newer). 44% of U.S. adults now say they have used ChatGPT, up from 34% in 2025, 23% in 2024 and 18% in 2023. About half (49%) say they use AI chatbots of any kind. Survey of 5,119 U.S. adults, 17 to 23 February 2026. Pew notes the 2026 question was worded differently: earlier years asked only people who had heard of ChatGPT, and the 2026 question also asked about other chatbots, so the years don't line up exactly. Pew, "Americans and AI 2026" (opens in a new tab).
Checked: 2 November 2026.
Chapter 3: what assistants cost today
Chapter 3 says what assistants cost today is on this page. The current plans and monthly prices for four assistants, as shown on each provider's own pricing page on 2 November 2026, are under "Is it free?" in the Everyday AI section above.
From the book
A system that learns from examples saves people from writing every rule by hand. What it learns comes from the examples it was given, so the examples matter as much as the method.
Sources
- What is science for? The Lighthill report on artificial intelligence reinterpreted (opens in a new tab) · Agar, Jon, 2020
- Rule-Based Expert Systems: The MYCIN Experiments of the Stanford Heuristic Programming Project (opens in a new tab) · 1984
- Deep Blue (opens in a new tab) · Campbell, Murray and Hoane, A. Joseph and Hsu, Feng-hsiung, 2002
- Dartmouth Summer Research Project and the Origin Myth of Artificial Intelligence (opens in a new tab) · Choi, Sokion, 2023
- Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act) (opens in a new tab) · European Parliament and Council of the European Union, 2024
- How the AI Boom Went Bust (opens in a new tab) · Haigh, Thomas, 2024
- Deep Blue (opens in a new tab) · IBM, 2026
- Spam does not bring us joy: ridding Gmail of 100 million more spam messages with TensorFlow (opens in a new tab) · Kumaran, Neil, 2019
- Artificial Intelligence: A General Survey (opens in a new tab) · Lighthill, James, 1973
- A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence, August 31, 1955 (opens in a new tab) · McCarthy, John and Minsky, Marvin L. and Rochester, Nathaniel and Shannon, Claude E., 2006
- The Dark Ages of AI: A Panel Discussion at AAAI-84 (opens in a new tab) · McDermott, Drew and Waldrop, M. Mitchell and Chandrasekaran, B. and McDermott, John and Schank, Roger, 1985
- The Logic Theory Machine: A Complex Information Processing System (opens in a new tab) · Newell, Allen and Simon, Herbert A., 1956
- John McCarthy, 1927-2011: A Biographical Memoir (opens in a new tab) · Nilsson, Nils J., 2012
- A Bayesian Approach to Filtering Junk E-Mail (opens in a new tab) · Sahami, Mehran and Dumais, Susan and Heckerman, David and Horvitz, Eric, 1998
- Some Studies in Machine Learning Using the Game of Checkers (opens in a new tab) · Samuel, A. L., 1959
- A model of inexact reasoning in medicine (opens in a new tab) · Shortliffe, Edward H. and Buchanan, Bruce G., 1975
- Computing Machinery and Intelligence (opens in a new tab) · Turing, A. M., 1950
- ELIZA: A Computer Program for the Study of Natural Language Communication Between Man and Machine (opens in a new tab) · Weizenbaum, Joseph, 1966