Don't classify, hallucinate! A clever LLM tagging trick using embeddings
Simon Willison · rss · 2026-08-15
Simon Willison shares Doug Turnbull's neat solution: let the LLM generate tags without seeing the existing vocabulary, then use vector embeddings to match them to the closest real tags. This solves the problem of too many tags to feed to an LLM at once.
More from Research
- INSIDE Framework Accepted to COLM 2026: Modeling Student Reasoning — alexisjross · 2026-08-15
- dots3 preview: Long-horizon agency in real life — otarU · 2026-08-15
- New metric suggests dense models benefit significantly at bs=1 — teortaxesTex · 2026-08-15
- MONA: Myopic Optimization Mitigates Multi-step Reward Hacking in RL — sebkrier · 2026-08-15
- Sébastien Bubeck's book on Convex Optimization available on ChapterPal — burkov · 2026-08-15
- Benchmarking Qwen2.5-32B Quantization: Custom AD-IQ3_S Beats Community by 33% — Top-Eye-8104 · 2026-08-15