About this dictionary

Preface

In 1993 IBM published the tenth edition of its Dictionary of Computing. It defined about 18,000 terms, and many carried a small letter telling you which standards body stood behind the definition. It gave an industry a shared vocabulary at the moment that industry was remaking the world.

Artificial intelligence has reached that moment, and its vocabulary moves faster than any printed book. In 2023 people rarely called the software around an AI agent a harness; by 2026 researchers were publishing papers on what exactly counts as one. Vibe coding went from a single social media post to everyday speech within a year. Older words such as inference and bias took on new senses along the way. Even the name of the field is now in play: in September 2026 a US executive order directed federal agencies to say "Super Intelligence" where they had said "artificial intelligence."

The Last Dictionary records that language in 2,709 entries, from Bayes' theorem and the perceptron to agents and superintelligence. It is written for three readers at once. The newcomer gets a plain-words definition at the top of every entry. The practitioner gets precise senses, formulas and primary sources. The decision-maker gets labels for what is emerging or contested, and markers for terms that laws and standards define.

Every entry also carries a birth certificate: who introduced the word, when and where, stated only as firmly as the evidence allows. Read in order, those dates tell the history of the field one word at a time. The timeline lays them out.

The name is a pun. Last as in final, the way Fermat's Last Theorem was the last of Fermat's claims to be settled; and also the latest, and meant to last. The entry for the book itself says a little more.

This first edition was researched and drafted with AI assistance and then fact-checked entry by entry by reviewers who had not written the entries. It will still contain errors. A dictionary of a field this young has to be corrected and reissued, as IBM's was through ten editions. That is why its source is open: every entry can be corrected by anyone.

How to read an entry

Every entry follows the same anatomy. Here is the entry for harness, a word whose meaning has changed three times, with each part numbered.

1. 1
   Headword. The word as it is most often written.
2. 2
   Part of speech and other names. Noun, verb, adjective, abbreviation, phrase or proper noun, with a simple respelling where the reading is not obvious. Other names and abbreviations follow Also and work in search.
3. 3
   Domains and labels. The home domain comes first, then neighboring domains, then labels such as emerging or contested.
4. 4
   As a model reads it. The headword split into tokens by the open o200k tokenizer: how a language model actually sees the word.
5. 5
   In plain words. One or two sentences anyone can follow. Start here.
6. 6
   Senses. The precise definitions, numbered when a word has several meanings, each with its context.
7. 7
   Example. The word used the way practitioners use it.
8. 8
   Origin and birth year. Who introduced the word, when and where, stated with the confidence the evidence supports. The year places the word on the timeline.
9. 9
   Usage note. Common confusions, disputes and advice on usage.
10. 10
    Related. See also points to neighboring ideas. Contrast with points to opposites and easily confused terms, a convention borrowed from the IBM Dictionary of Computing.
11. 11
    Sources. Primary sources for the origin and the definition, with links where available.
12. 12
    Law and standards. Where a statute or standard defines the term, the instrument and article. Time-sensitive entries also carry an as of date.

Labels and abbreviations

Parts of speech

Labels

Levels

Families and domains

Cross-references

Also lists other names and abbreviations. See also points to related entries. Contrast with points to opposites and easily confused neighbors. Mentioned in lists the entries that link to this one. Links inside a definition show a plain-words preview when you hover over them.

Law and standards

IBM's dictionary marked definitions taken from American (A) and international (I, T) standards. This dictionary keeps the idea for the AI era. Where a statute or standard defines a term, such as the EU AI Act, ISO/IEC 22989 or the NIST AI Risk Management Framework, the entry names the instrument and the article, paraphrased rather than quoted.

Editorial principles

1. Two layers, one truth. Every entry opens with a plain-words definition, then gives the technical one. Neither contradicts the other.
2. Every word has a birth certificate. We say coined by only when the evidence shows first use; otherwise introduced, popularized or earliest known use.
3. Senses, not soup. Words with several meanings, such as harness, kernel, bias, token and inference, get numbered senses with their contexts.
4. Describe usage and label it honestly. We record how words are used, including slang, and mark what is emerging, contested, historical or deprecated.
5. Map disagreements; don't settle them. On AGI, emergence, consciousness and risk, entries set out the positions and who holds them.
6. Standards-aware. Where law or a standard defines a term, the entry says so.
7. Dated, because the field moves. Anything that can change carries an as of date. The dictionary is issued in numbered editions.
8. Connected. Every entry links to related and contrasting entries. There are no dead ends.
9. Original and safe. Definitions are written fresh, not copied. Entries on attacks and misuse explain concepts and defenses, never procedures.

Lineage

Four computing dictionaries shaped this one. Each solved a different problem.

Modern AI glossaries also informed the scope: Google's Machine Learning Glossary, NIST's The Language of Trustworthy AI (AI 100-3, 2023) and the international standard ISO/IEC 22989:2022 on AI concepts and terminology. None of them covers the whole field, from classical statistics to agents, safety, law and slang, with origins for every term. That gap is why this dictionary exists.

Edition and citation

First Edition, October 2026, 2,709 entries. Time-sensitive facts are current to the as of date shown in each entry.

The whole dictionary is free to download as JSON and CSV, and every entry is a small file in the source repository. The text is licensed under Creative Commons Attribution 4.0; the site's code is MIT licensed.

To cite an entry, use Copy citation on the entry, which produces a reference like: The Last Dictionary, 1st ed. (October 2026), s.v. “harness,” lastdictionary.com/harness.

How this edition was made

A master list of about 1,860 planned headwords was divided among forty contributing editors, with one owner for every word; the editors added some 850 more as they worked. Each editor researched and drafted entries against a common style guide, checking origins, dates and recent facts against primary sources and flagging every claim they could not confirm.

A separate fact-checker, who had not written the entries, then reviewed every entry in every domain, resolving each flag and correcting errors. In fast-moving domains such as law, organizations and the newest vocabulary, a third or more of the entries needed a correction. Three further auditors then re-checked 166 randomly chosen entries. About one entry in twelve still had a factual error, usually a minor one, and about one in twenty had a passage that followed its source too closely. Every problem they found was corrected. A final originality sweep then compared each entry with the best-known published wording on its subject, such as encyclopedia openings, paper abstracts, documentation and legal texts, and rewrote more than a thousand passages that followed a source too closely, or credited the source by name. Similar errors surely remain in an edition of this size.

The edition was compiled with Claude, an AI model made by Anthropic.

Colophon

Set in Newsreader, a typeface by Production Type made for reading on screens, whose optical sizes sharpen as the type grows, with DM Mono for labels and figures. The mark is the small square that ends a mathematical proof. Tokens are shown with the open o200k tokenizer through the gpt-tokenizer library. Formulas are rendered as MathML with KaTeX. The site is built with Next.js into plain static pages, one per word.

Sponsor

The Last Dictionary is sponsored by Plane, the work infrastructure for humans and agents: Projects, Knowledge and Service. The dictionary stays free to read, free to download and open to correction.

Contribute

The dictionary is maintained in the open at github.com/vamsi/lastdictionary. Each entry is one small YAML file, so a correction is a one-file pull request, and a new word is a new file. Found an error? Open an issue with the entry and a source, or fix it directly. The contributing guide explains the fields, the style rules and how to check your work before you send it.

The site is built to WCAG 2.2 AAA where that is sensible; the accessibility statement says what was tested, what the exceptions are, and how to report a problem.