model collapse

Compiled with AI assistance and selectively reviewed. Please consult the cited passages; the books remain the authoritative source. About this companion

Editorial synthesis · provisional

Definition

Model collapse names the recursive degradation the corpus associates with models learning from material generated by predecessor models. The archive increasingly records prior machine output rather than independent encounters with the world. AB grounds its illustration in a cited nine-generation experiment; the broader danger of a self-sealing archive is a philosophical extension, not a demonstrated fate of every synthetic-data method.

Explanation

AB reports a sequence that begins with English church architecture and ends with a list of jackrabbits. Its note identifies Shumailov and colleagues’ 2024 experiment, keeping the anecdote attached to a named publication. The assessment records what the source reports rather than claiming a new replication.

The accompanying argument is that AI increasingly helps select the archive on which later systems are formed. Repeated projections can lose correction from the practices that generated the original traces. This connects recursive training to provenance and selection without identifying every reuse of generated data with inevitable failure.

RB3 widens the issue beyond training quality: a poorly grounded symbolic system can still reorganize bodies and institutions through decisions. It also retains the possibility of design that keeps the archive porous to correction. PB’s fictional recursive echo should remain literary context rather than experimental evidence.

Scope and attribution

Retain the source-reported empirical illustration and mark the wider diagnosis as an extension.

AB — The Actual Book

The church-to-jackrabbit illustration has an explicit experimental citation.

Occurrence locators: 151, 152; Notes: 245.

RB3 — The Real Book, Volume III: Murmurations

Recursive archive selection motivates a demand for world-coupled correction.

Occurrence locators: 153.

Background used in the editorial explanation

Named theories and mechanisms are distinguished from neighboring background, homonyms and the author’s own synthesis. Reproduced AI passages retain their separate voice.