Simon Willison describes a creative approach to blog tagging where LLMs hallucinate novel tags rather than matching against an existing vocabulary, then vector embeddings are used to map those hallucinated tags to the closest real tags in the corpus. Doug Turnbull proposed this two-step method, which sidesteps the limitation of feeding large tag lists into LLMs.
Background
Willison has over 1,800 blog tags, making direct classification impractical for LLMs. This technique leverages generative flexibility combined with semantic similarity matching via embeddings.
- Source
- Simon Willison
- Published
- Aug 15, 2026 at 05:54 AM
- Score
- 5.0 / 10