Definition
Plain language
The lead author of the well-known study showing AI trained on AI output degrades over generations.
As stated in the literature
First author of the 2024 Nature paper on model collapse, demonstrating progressive distribution narrowing and tail loss when generative models are recursively trained on their own outputs.
Why it matters: It gave the field a concrete name and demonstration for the risk of training AI on AI-generated content, which is now a routine concern when gathering data from the open web.
For example, the study associated with this name showed that generative models trained again and again on their own output gradually lose the rare cases at the edges of the data.
Heard on the show
“… The dominant story so far, from Shumailov and colleagues at Nature in twenty-twenty-four, locates collapse in training: when you train new …”Episode 073 — When Three LLMs Talk to Each Other, Their Ideas Quietly Stop Moving