Douglas B. Books
What Kind of AI Is That? · read online

Back Matter — The Snapshot Statement · Sources · The Map on One Page

The Snapshot Statement

This book is a map of kinds, and a map is only as good as its date. The types themselves — talkers, image models, librarians, and the rest — change slowly. The examples do not. Every named model, version, license, and event in this book is therefore stated as of November 2026, and each was checked against the maker's own pages, papers, model cards, or repositories, where those exist.

Where no primary source could be reached, the rule was simple: the text says less rather than more. A few dates rest on consistent secondary reporting and are stated at the precision it supports — "early 2026," not a day. Items that could not be verified at all were left out, and three are worth naming here, because the leaving-out is part of the record:

The online editions of this book are updated in place as the roster moves. Each carries a version line and a date; when a named model is retired, renamed, or replaced, the update says so, and the dated editions stand as the record of their moment.

Sources

Credit where the record came from. The roster in this book rests on the makers' own publications, grouped by kind. Secondary reporting was used only where noted in the text.

Large language models

Reasoning and multimodal

Image models

Video models

Audio models

Embedding and retrieval models

Small and on-device models

The frontier shelf

The layer

The Map on One Page

The nine kinds

  1. 1. Large language models — text in, text out: the talkers, and the kind everything else starts from.
  2. 2. Reasoning models — models that work a problem in steps before they answer.
  3. 3. Image models — pictures from words: generation, editing, and holding a subject consistent.
  4. 4. Video models — moving pictures from words and stills, a few seconds at a time.
  5. 5. Audio models — sound in three directions: speech out, transcription in, and music.
  6. 6. Multimodal models — native in more than one medium at once: they see, hear, and talk in a single model.
  7. 7. Small and on-device models — types built to run on a phone or a single machine: private, offline, and instant.
  8. 8. Embedding and retrieval models — the librarians: meaning turned into numbers, so anything can be searched by sense.
  9. 9. The frontier shelf — kinds still forming: world models that simulate a place, and robot-action models that move a body in one.

Two melts, recorded as of November 2026. Reasoning has dissolved from a kind into a mode — a setting inside the flagship talkers, covered in its chapter as the record of that change. Multimodal is becoming a property — something every flagship now has, rather than a shelf of its own. Kinds can melt. The map records it when they do.

Not a kind: the layer. A loop — tools, memory, turns, and permissions wrapped around a model until it can do a job — is a system built on a model, not a type of model. It is covered in Part Three, where it belongs.

The five questions, for any new model: What kind is it? What goes in, and what comes out? Are the weights open or closed — and under what license? How big is it, and what does it run on? Who made it, and when?

Names will change by the next printing. The questions will still sort them.

← Part Three — ChoosingContentsAbout the Art · About the Author →