Simon Willison

Highlight

Today’s standout post is Don’t classify. Hallucinate!, which showcases Doug Turnbull’s clever “hypothetical classification” technique. By letting LLMs freely hallucinate tags first and then mapping them to an existing taxonomy using vector embeddings, it offers a brilliant way to bypass context window limitations for massive classification schemas.

Posts

Don’t classify. Hallucinate! · Source Simon discusses Doug Turnbull’s innovative method for tagging older blog posts when your taxonomy is too large to fit in a single LLM prompt context. Instead of feeding all 1,856 of Simon’s existing tags to the model, the model is prompted to invent or “hallucinate” appropriate classifications from scratch based on a structural example. These generated tags are then mapped back to the actual taxonomy by finding the closest matches using vector embeddings against the existing corpus.

llm-gemini 0.33 · Source This update to Simon’s llm-gemini CLI plugin introduces support for the newly launched Gemini 3.7 Flash model, alongside 3.6 Flash, 3.5 Flash-lite, and newer embedding models. Built for compatibility with LLM 0.32, the plugin now enables reasoning traces and server-side tools. Simon tests the model’s capabilities by having it generate complex SVGs of pelicans riding bicycles, which revealed an interesting cross-browser quirk where Safari successfully rendered the drawing while Chrome and Firefox hid the pelican due to an empty filter element.

sqlite-utils 4.2 · Source Simon has shipped a feature-packed update to sqlite-utils focusing heavily on refining the table.transform() functionality. This utility allows complex table schema modifications in SQLite—which doesn’t natively support many ALTER TABLE operations—by transparently copying data into a newly constructed table. The 4.2 release does a significantly better job preserving edge cases like check constraints, unique constraints, and column comments, though a quick follow-up 4.2.1 was released shortly after to patch a crashing bug.

Project Pulse

It is a big day for Simon’s local developer toolkit, with significant releases across his core SQLite database management utility and LLM CLI ecosystem, alongside a practical exploration of embedding-based text classification.

🧩 If you’d like, we can write a quick Python script to prototype Doug Turnbull’s hypothetical classification technique using some of your own sample texts and a mock tag set.