AI

Starting a thread of work costs almost nothing now. Switching between them costs exactly what it always did, and that's the bill nobody put on a dashboard.

Flat-rate AI taught us to burn tokens like they were free. Metered billing puts a price back on the sloppy prompt, and that visibility is a feature.

Autonomous agent loops don't remove you from the work, they relocate you — onto the spec you write up front and the big diff you review but never watched grow.

The slop flooding Steam isn't a tool problem, it's a taste problem. AI didn't lower the floor for making games — it removed the excuse for a bad one.

Privacy was never the only reason to run a model locally — resilience is. A fourth case for local AI, made concrete by 18 days of a frontier model going dark.

AI collapsed the cost of writing code and changed nothing about the cost of owning it. Velocity is a development-time number. The bill is a lifetime one.

AI gave me the throughput to cover a role I don't normally hold — but not the judgment that role is made of. A release, an incident, and a skill I'd outsourced.

Three cases where running a model on your own machine wins, and the longer list where it loses.

Why scraping didn't die when LLMs arrived, and how I decide between Beautiful Soup and Scrapy for a given job.