Doug Turnbull's Original Post: When You Want Hallucination, a Tiny Model Is Enough — Because You Need Creativity, Not Accuracy

softwaredoug.com·low signalXBlueskyLinkedInCopy link

The primary source behind Willison's August 14 endorsement argues that the useful pattern for LLM classification at scale is to let the model invent plausible fake entities and then resolve them to real ones, rather than trying to make the model accurate up front. Turnbull's sharpest point is economic: because you are not relying on the LLM for correctness, only for close-enough language generation, you can run the step on very small and cheap models. That inverts the usual instinct to reach for a frontier model whenever classification quality matters.