The Offline-First Doctrine

Walk through almost any modern system that claims to be intelligent and you will find the same hidden clause in its design. It assumes the link is always there. The model lives in a data centre, the device in your hand is a window, and the conversation between them is the product. While the connection holds, the experience feels effortless. The assistant answers, the camera recognises, the document is summarised. None of it advertises the dependency, because the dependency is

The common way to handle disconnection is to bolt on a fallback. The real system lives in the cloud, and when the cloud is unreachable a shrunken offline mode steps in, usually a cache of yesterday's answers and an apology. This is better than nothing, but it betrays the underlying belief that offline is the degraded case. I want to invert that belief. In an offline-first architecture, local is the default and the network is the enhancement, not the other way around.

A doctrine that only makes engineering sense is a hobby. To be worth adopting, offline-first has to make economic sense as well, and the honest starting point is that it costs more up front. Owning hardware capable of local inference is a real expense. Specialising and serving your own models is a real effort. Anyone who tells you sovereignty is free is selling you something. So the question is not whether it costs, but whether what it buys is worth the price over the life of