Don't classify. Hallucinate!
Blog · Watch/Read
- Simon Willison highlights Doug Turnbull’s “Don’t classify. Hallucinate!” method for tagging content, which addresses the challenge of using LLMs with an excessively large existing tag vocabulary.
- The core idea is to prompt an LLM to generate novel, imagined tags or classifications (e.g., hierarchical paths like “Furniture / Living Room Furniture / Coffee Tables”) for content, without providing the existing tag list.
- These “hallucinated” tags are then mapped to the closest concrete tags in the existing corpus using vector embeddings, effectively bypassing token limits and enabling efficient tagging of large, untagged archives.