Local LLM Lab.
3 posts in this series. All writing →
An Alphabetical File Listing Outlasted Most of My Ideas
What to ask before you fund a context platform or task router. On a rotated benchmark an unranked file listing beat the context maps I built for small local models, and even a hindsight oracle over the candidate strategies failed the funding gate for a router.
Cost-Aware Model Selection: When Local Beats Frontier
The most capable model is not always the correct choice. A routing policy that defaults bounded tasks to a local engine, with a CLI you can run yourself, and two real bugs found while proving the policy worked as documented.
LLM-as-Judge, and How Not to Fool Yourself
A model judging another model's output sounds confident by default. Confident is not the same as calibrated. The measurement methodology that tells the difference, and a bug the methodology caught in its own tool.