Definition
Plain language
A score for how often a model's answers are ones nobody else would give.
As stated in the literature
Per-model diversity metric measuring, in bits, how unlikely each of a model's single-word answers is under the pooled distribution of all other models' answers, averaged across prompts.
Why it matters: It lets researchers put a single number on how off-the-crowd a model tends to be, making it easy to compare which models play it safe and which take risks.
For example, a model that keeps answering "quince" and "kumquat" when others say "apple" would earn a high Mustard Quotient.
Heard on the show
“But the scoring gives each model one number he nicknames the Mustard Quotient.”Episode 219 — Forty-Four AI Models, One Word, And The Newest Ones Conform Most