Writing on systems, AI, and shipping software

Most sites reach for a server, a database, and a runtime they never needed. Here is the case for shipping HTML you can cache forever — and the narrow set of cases where dynamic actually earns its keep.

Distributed systems don't fail the way single machines do. They fail partially, ambiguously, and at the worst possible moment. A practical taxonomy of the failures that will find you, and the postures that survive them.

I spent the last quarter turning a single overworked assistant into a fleet of forty-odd narrow agents. Here is what actually changed — and the handful of decisions I'd make again on day one.

Training gets the headlines, but inference is where the money quietly goes. A look at what actually drives the bill — context length, cache hit rate, batching — and the levers that move it.

A week of deliberately high throughput taught me where velocity actually comes from — and that the bottleneck is almost never typing speed. Notes on small diffs, fast review, and the discipline that makes speed safe.