MIT License

You don't need 10,000 examples to make a model sound like you. The real number is closer to 100 — if the edits are opinionated.

A shopping bot that holds one cart total turns out to contain every hard problem in agent state management.

Small models plus strict JSON schemas should be a safe bet. A 15,000-generation study found the opposite.

There's no tool for finding an LLM's verbal tics — you have to build one. Here's how keyness analysis works.

A 744B-parameter model drops quietly on Hugging Face. Is "agentic" a real model category, or just a deployment pattern?

DeepSeek quietly routed Pro traffic to Flash — and that routing change says everything about its two-endpoint strategy.