Method and limits

The selection rules were registered before the list was made; dates come from the originals and lineage lines from citation data. This page records the procedure and the limits it carries.

Selecting papers — registered 2026-11-03

Six reference works were fixed first: Nilsson, The Quest for Artificial Intelligence (2010); Russell & Norvig, Artificial Intelligence: A Modern Approach, 4th ed. (2020); Goodfellow, Bengio & Courville, Deep Learning (2016); Sutton & Barto, Reinforcement Learning, 2nd ed. (2018); Prince, Understanding Deep Learning (2023); Jurafsky & Martin, Speech and Language Processing, 3rd ed. draft. Four inclusion rules apply.

• R1 Textbooks. At least two of the six cite the original publication. (122 papers)
• R2 Lineage. At least three already-selected papers cite it (OpenAlex or the citing paper's reference list). (38)
• R3 Awards. The Turing Award or Nobel Prize citation names the achievement. (15)
• R4 Recent (2021 onward). Peer-reviewed publication or public review record, original text checked. (28)

The result is 191 papers. Textbook matching compared bibliography text by title, author and year automatically, with samples checked by hand. Nilsson's book keeps its references in chapter endnotes, so the whole text (the author's public PDF, Internet Archive capture of 2026-10-11) was used as the comparison text; a work named by title or by author and year in the body counts as cited. Schmidhuber's "Annotated History of Modern AI and Deep Learning" (2022) is not a reference work here; it is used only as one side in the priority boxes.

Dates

A point sits at the first public date: the earliest documented technical report, thesis, arXiv v1 or conference presentation. Journal issue dates, conference dates and reprint years are listed on the card. OpenAlex publication years were not used because they mix in reprint and conference dates (23 of 234 candidates differed).

Checking the originals

Each card's quotation and page were read from the stated version (journal, conference, or arXiv version number). By status: 166 verified, 14 partially checked, 11 not accessible. "Not accessible" means no legitimate public copy was found; those cards carry no quotation.

Lineage lines

Three kinds: in OpenAlex's reference list (425), not in OpenAlex but seen in the citing paper's own reference list (683), and stated in a textbook without a confirmed citation (0, dotted). OpenAlex alone leaves large gaps: among the candidates, 30% of pre-1990 records and 40% of arXiv records had zero references.

Priority disputes

Where "who was first" is disputed (backpropagation, convolutional networks, deep network training, residual connections), the boxes place both sides' original bibliographies and sentences side by side and make no ruling. Secondary sources that do rule are introduced as claims, with their source.

Bias

The reference set is English-language textbooks, and the resulting bias is not hidden. Early neural-network work from the Soviet Union and Japan (Ivakhnenko, Amari, Fukushima) and Linnainmaa in Finland sometimes fails the two-textbook rule; such papers appear in the priority boxes. The country distribution of author affiliations can be computed from this version's papers.csv and the OpenAlex records.

Sources and terms

• OpenAlex — bibliographic data, citations, citation counts, yearly counts. Data CC0. Citation counts appear only with the date read (2026-11-03).
• arXiv — metadata (title, authors, v1 date) CC0. PDFs are linked, not hosted.
• Crossref — DOI checks.
• Wikidata (Turing Award laureates, CC0) and the Nobel Prize API (CC0).
• AI Index 2019, chapter 2 attendance data; Our World in Data (CC BY); proceedings.neurips.cc.

Text on this site is CC BY-NC 4.0, code PolyForm Noncommercial 1.0.0, and the data compilation CC BY 4.0 (sources CC0 or CC BY). Paper texts are not redistributed; only short quotations and links.