Writing
Writing
Mostly the economics and mechanics of owning your AI instead of renting it.
What on-prem LLM deployment actually costs once you count engineers, and the volume threshold where it beats hosted.
INT8 vs FP16 on real eval sets — the accuracy you give up and where it matters.
When internal systems are already governed tools, swapping the model underneath gets cheap.
On public data that isn't usable, and what it takes to change that.