What we checked
Where the old stories and the numbers in this guide came from. Each chapter also carries its own short list at the bottom of the page.
What the guide says
The last column is how far we got. Read the original means we opened that paper or record ourselves; checked via a quotation means we confirmed it from another work that reproduced the original wording.
| Chapter | What the guide says | Source | How far we got |
|---|---|---|---|
| 1.1 | In the 1980s people typed in thousands of rules by hand, and factories and companies really ran on themchecked via a quotation | 16 | checked via a quotation |
| 1.1 | As the rules piled up, keeping them grew bigger than writing themchecked via a quotation | 17 | checked via a quotation |
| 1.2 | A checkers program started beating the person who built it — after eight or ten hours of machine-playing timeread the original abstract only · pp. 601–602 | 6 7 | read the original abstract only · pp. 601–602 |
| 1.3 | The phrase "artificial intelligence" was first used in 1955, in a plan built around a single summerread the original full text | 5 | read the original full text |
| 1.3 | The idea of computing by imitating nerve cells dates from the 1940schecked via a quotation | 1 | checked via a quotation |
| 1.3 | A machine built on that idea came in the 1950schecked via a quotation | 8 | checked via a quotation |
| 1.3 | The learning machine made of matchboxes is from the 1960schecked via a quotation | 10 | checked via a quotation |
| 1.3 | A way to teach many layers stacked together spread in the 1980schecked via a quotation | 29 | checked via a quotation |
| 1.3 | Seven prediction cards — things people actually said about what would happen whenchecked via a quotation | 23 3 24 8 25 26 5 | checked via a quotation |
| 2.1 | The first digital photograph, in 1957, was 176 squares on a side — about thirty thousand in allread the original full text | 20 | read the original full text |
| 2.2 | People designed a whole new typeface just so machines could read itchecked via a quotation | 15 | checked via a quotation |
| 2.2 | Reading handwriting stayed unsolved for more than forty years — commercial readers of printed text arrived in the mid-1950s, handwritten postal addresses were read for real in 1997checked via a quotation | 28 27 | checked via a quotation |
| 2.3 | The method published in 2001 made face-finding about fifteen times faster — the paper reports 15 frames a second on 384×288 images. That the previous method took about a second a frame is our own arithmetic from those two figuresread the original abstract and conclusion | 19 | read the original abstract and conclusion |
| 2.3 | The little box around faces appeared on cameras in the mid-2000schecked via a quotation | 21 | checked via a quotation |
| 2.4 | The first speech recogniser, in 1952, knew ten digits — and only in one set voicechecked via a quotation | 4 | checked via a quotation |
| 2.4 | 1990s dictation software made you read set sentences aloud before it would workchecked via a quotation | 22 | checked via a quotation |
| 3.1 | About seventy years ago came a machine taught by cards instead of written ruleschecked via a quotation | 8 | checked via a quotation |
| 3.2 | A machine imitating the brain was announced in 1958, to enormous expectationschecked via a quotation | 8 | checked via a quotation |
| 3.2 | Eleven years later two researchers proved that a single straight line can't separate even a very simple arrangementchecked via a quotation | 13 | checked via a quotation |
| 3.2 | How to teach those extra layers was worked out in the 1980schecked via a quotation | 29 | checked via a quotation |
| 3.4 | A machine that learned without a computer — 304 matchboxes and coloured beadschecked via a quotation | 10 | checked via a quotation |
| 4.1 | When computers first handled letters, people settled on 128 of themchecked via a quotation | 9 | checked via a quotation |
| 4.1 | A way to hold every letter in the world took about thirty years to arrivechecked via a quotation | 9 18 | checked via a quotation |
| 4.2 | Opening a book at random and writing down the next word was enough to produce something that looked like Englishread the original §3, pp. 7–8 | 2 | read the original §3, pp. 7–8 |
| 4.4 | A 1966 program that did nothing but find a keyword and ask backread the original p. 36 | 11 | read the original p. 36 |
| 4.4 | His secretary asked him to leave the room — he wrote it down himselfread the original p. 477 | 12 | read the original p. 477 |
| 4.4 | Its author went on to become one of AI's sharpest criticschecked via a quotation | 14 | checked via a quotation |
| 6.1 | In 1950 a mathematician asked how we would know whether a machine thinkschecked via a quotation | 3 | checked via a quotation |
| 6.1 | Telling a person from a machine by conversation alone has now become very hard — in a study where people talked for five minutes to a human and a machine side by side, the machine was picked as the human more often than the human wasread the original abstract only | 30 | read the original abstract only |
What the labs are made of
| Chapter | Lab | What it's made of |
|---|---|---|
| 2.2 | H1 | Weights we trained ourselves. We drew the training characters too |
| 2.2 | P2·P3 | Photos that need no credit line (CC0 · public domain · Pexels) |
| 2.2 | P3 | The response maps come from running the photos through a pre-trained network. We don't ship the weights |
| 2.3 | H11 | OpenCV's frontal face detector (BSD-3 · Apache-2.0). We didn't touch the numbers |
| 2.3 | H11·P4 | Three stand-in videos we made ourselves. The screen says they were made with AI |
| 2.3 | P4 | MediaPipe face, hand and pose (Apache-2.0), bundled so nothing is fetched from elsewhere |
| 2.4 | H12·P5 | The prepared voices are computer-made, and the screen says so |
| 3.1 | P6 | MediaPipe only turns the photo into numbers. The last layer — the one that decides what it is — you train that one here |
| 3.3 | P8 | Sixty photos gathered with backgrounds deliberately split (Pexels), checked by eye |
| 4.1 | P12 | Cutting rules we built ourselves from the old stories used in 4.2 |
| 4.2 | H5·P13 | Out-of-copyright old stories — Bang Jeong-hwan, Joseph Jacobs, Hans Christian Andersen, Kusuyama Masao |
| 4.2 | P13 | A small language model we trained on those stories. No borrowed weights |
| 4.3 | P14 | Numbers only, pulled from a multilingual language model. The screen says a person picked one map out of 144 |
| 5.1 | P18·P19 | Output from a picture machine we trained on pictures we drew in code |
| 5.3 | P21 | Reuses the pictures from 5.1. Nothing new was made |
| 5.3 | H9 | The photos laid over each other are public-domain and CC0 files already in use |
What we measured
Not literature — numbers we took ourselves. Most of them the labs put on screen as they run.
| Chapter | Lab | What we measured |
|---|---|---|
| 2.3 | H11 | We timed the stretch where the 2001 method loses the face, then made sure the lab runs through it |
| 2.4 | H12·P5 | Measured by enrolling the ready-made voice and testing all nine. Korean: the enrolled voice six times, another voice twice. English: nine and four. Japanese: eight and two. Spanish: nine and four — English digit words sound so unlike each other that the gap narrows |
| 3.3 | P8 | Taught on set A it gets 9 of 12; taught on set B, all 12. The three it misses are always the swapped-background photos. The lab shows this on screen |
| 4.1 | P12 | How differently a language gets cut up is something the lab counts on the spot and shows you |
| 4.2 | H5·P13 | Counting three words back, 33% of what it writes is copied straight from this book. The trained side: 0%. (The Korean storybook comes out at about 60% — those tales repeat more. Japanese counts single characters, and eight characters back gives 54%; the trained side is 0% there too) |
| 4.3 | P14 | Six of the seven sentences we score come out right. The eighth splits people too, so it isn't scored. The number on screen is a different one — how often you and the machine agreed |
| 5.1 | P18·P19 | The screen says it's a 32-square machine, which is why you can see the blocks |
Parts 5 and 6 have almost nothing to cite. What to hand to an AI, how your own data gets used, which rules you'll hold yourself to — no paper has the answer written down. Which is why those chapters ask rather than answer.
Sources
- McCulloch & Pitts, "A Logical Calculus of the Ideas Immanent in Nervous Activity", Bulletin of Mathematical Biophysics 5 (1943) checked via a quotation doi:10.1007/BF02478259
- Shannon, "A Mathematical Theory of Communication", Bell System Technical Journal 27 (1948) read the original §3, pp. 7–8 doi:10.1002/j.1538-7305.1948.tb01338.x
- Turing, "Computing Machinery and Intelligence", Mind 59(236) (1950) checked via a quotation doi:10.1093/mind/LIX.236.433
- Davis, Biddulph & Balashek, "Automatic Recognition of Spoken Digits", Journal of the Acoustical Society of America 24(6) (1952) checked via a quotation doi:10.1121/1.1906946
- McCarthy, Minsky, Rochester & Shannon, "A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence" (1955) read the original full text open
- Samuel, "Some Studies in Machine Learning Using the Game of Checkers", IBM Journal of Research and Development 3(3) (1959) read the original abstract only doi:10.1147/rd.33.0210
- Samuel, "Some Studies in Machine Learning Using the Game of Checkers. II — Recent Progress", IBM Journal of Research and Development 11(6) (1967) read the original pp. 601–602 doi:10.1147/rd.116.0601
- "New Navy Device Learns By Doing", The New York Times (1958-07-08) checked via a quotation open
- ASA X3.4-1963 — American Standard Code for Information Interchange checked via a quotation open
- Michie, MENACE (1961) · "Experiments on the Mechanization of Game-Learning", The Computer Journal 6(3) (1963) checked via a quotation doi:10.1093/comjnl/6.3.232
- Weizenbaum, "ELIZA — A Computer Program For the Study of Natural Language Communication Between Man And Machine", Communications of the ACM 9(1) (1966) read the original p. 36 doi:10.1145/365153.365168
- Weizenbaum, "Contextual Understanding by Computers", Communications of the ACM 10(8) (1967) read the original p. 477 doi:10.1145/363534.363545
- Minsky & Papert, Perceptrons (MIT Press, 1969) checked via a quotation open
- Weizenbaum, Computer Power and Human Reason (W. H. Freeman, 1976) checked via a quotation open
- OCR-A (American Type Founders, 1968) · ANSI X3.17-1981 · ISO 1073-1 checked via a quotation open
- McDermott, "R1: A Rule-Based Configurer of Computer Systems" (1980) · Barker & O'Connor, Communications of the ACM 32(3) (1989) checked via a quotation doi:10.1145/62065.62067
- Soloway, Bachant & Jensen, "Assessing the Maintainability of XCON-in-RIME" (1987) checked via a quotation open
- The Unicode Standard, Version 1.0 (1991) checked via a quotation open
- Viola & Jones, "Rapid Object Detection using a Boosted Cascade of Simple Features", CVPR 2001 read the original abstract and conclusion doi:10.1109/CVPR.2001.990517
- NIST, "First Digital Image" · "Fiftieth Anniversary of First Digital Image Marked" (2007) read the original full text open
- Fujifilm, "FinePix S6500fd" press release and specifications (2006-07-13) — the first Fujifilm camera with face detection checked via a quotation open
- IBM ViaVoice user guide — "Enrollment" (reading set sentences to train the recogniser on one voice), 1990s editions checked via a quotation
- Simon, The New Science of Management Decision (Harper & Row, 1960 · 1965) checked via a quotation open
- George Johnson, "To Test a Powerful Computer, Play an Ancient Game" (interview with Piet Hut), The New York Times (1997-07-29) checked via a quotation open
- Moravec, Mind Children (Harvard University Press, 1988) checked via a quotation open
- Hofstadter, Gödel, Escher, Bach (Basic Books, 1979) checked via a quotation open
- United States Postal Service, 1997 Annual Report, p.43 — the Remote Computer Reader, built on University at Buffalo (CEDAR) handwritten address interpretation checked via a quotation open
- OCR trade histories on the first commercial installations (Shepard / Intelligent Machines Research Corp. at Reader's Digest, mid-1950s) — decade only; no primary source located checked via a quotation
- Rumelhart, Hinton & Williams, "Learning representations by back-propagating errors", Nature 323 (1986) checked via a quotation doi:10.1038/323533a0
- Jones & Bergen, "Large Language Models Pass the Turing Test", arXiv:2503.23674 (2025-03-31) read the original abstract only open