隨筆

A couple weeks ago I wrote a "llm_grep" script that basically asks an LLM to grep something. For this script to be useful in a broad context, ideally you'd want the model to be a local one, otherwise who knows what weird stuff you'd accidentally send out as training data to LLM providers... So I tried it with a bunch of small-ish local models, but they didn't quite work. (TODO: point to the test prompt files) Until Qwen released Qwen3-30B-A3B-Instruct. This model is great. Long context, fast, and generally performant. It might become my go-to model replacing Gemma-3-27B