Language at the edge, logic at the core
I write when a technical subject stays with me long enough to become worth clarifying: language models, observability, embeddings, memory, and the boundary between narrative and proof. The goal is simple: keep the useful parts, remove the fog.
No publishing schedule. Just subjects that remain interesting after the first explanation.
Newest first
Renaming JSON keys can look like a harmless refactor. In a preregistered experiment with TypeSafe’s Jev, it changed the application’s behavior: with identical descriptions and the same routing rule, the human-review rate rose from 45.8% to 70.8% on cases selected near the threshold.
When model judgment controls a consequence-bearing value, language becomes part of the system contract. This essay shows why JSON improves legibility but not enforcement, and why the boundary has to be pinned, versioned, tested, and enforced outside the model.
A step-by-step inside-the-box explanation of what is cached, what is not, and why reuse is limited. Especially relevant for multiple questions over the same large document.
A practical argument for why LLMs are exceptional at explanation but unreliable as causal engines. Includes concrete failure modes: arithmetic, long context decay, prompt bias, and the cost of fake reasoning.
A step-by-step inside-the-box explanation of how meaning emerges from training, why dimensions have no names, and why rotated spaces still think the same.