Blog
Notes from building a data layer that maintains itself. RSS
We gave a small model an ontology. It mostly ignored it.
A 0.8B model with an enterprise-style catalog in its prompt picked the right type 4.8% of the time. Training took it to 77.4%. Then the new types arrived, and it fell to 1.7%.
SPARQL, Without the Semantic Web Sermon
What SPARQL actually is, the two things it does that SQL and Mongo can't, and the five places it will make you miserable. With figures you can drag.
How to Lie to Yourself With an Eval Judge
Three ways our eval harness produced wrong accuracy numbers: a judge that couldn't see correct answers, a heuristic that accepted almost anything, and weeks of measuring code we don't ship.