Don't Classify, Hallucinate: A New Approach to Tagging with LLMs
On August 14, 2026, Simon Willison posted on his blog about innovative approaches to content tagging using large language models (LLMs). He outlined the complications of managing 1,856 different tags on his site. Doug Turnbull suggested a novel method where the LLM creates tags independently, enhancing flexibility. By applying vector embeddings, the system can identify the most relevant existing tags from its database. Turnbull illustrated this with a hierarchical tag structure example. Additionally, Willison offered a monthly $10 sponsorship option for personalized email updates on LLM advancements, showcasing creative applications of LLM capabilities for improved content organization.
Key facts
- Simon Willison published a blog post on 14th August 2026 titled 'Don't classify. Hallucinate!'.
- Willison's blog has 1,856 tags.
- Doug Turnbull proposed a solution for tagging using LLMs.
- The method involves generating novel tags without prior vocabulary knowledge.
- Vector embeddings are used to find existing tags closest to the generated ones.
- Turnbull's example prompt includes sample tag structures for furniture and home goods.
- The post is a link post by Willison.
- Sponsors at $10/month receive a curated email digest of LLM developments.
Entities
—