Hello
RSS FeedI'm Rooz. I work on infrastructure and operations: networks, cloud platforms, and the machinery underneath that keeps services running when nobody is watching. This is where I write up the things I had to work out the hard way.
Recent Posts
The restraint of a model that only decides
Jev does not write anything, cannot be talked into trying, and ships with a published list of its own weaknesses. I have not admired a piece of AI engineering this much in a while, and the reasons are mostly about what it refuses to do.
Token cost as a design constraint
Caching, batching and routing are the three levers on an inference bill. They are worth roughly ten times, two times and up to a hundred times respectively, and almost everybody reaches for them in the wrong order.
When inference costs more than the engineer
An hour of engineer time buys roughly four million output tokens at top tier rates. Working out where that stops being true, and why the answer has almost nothing to do with which model you picked.
What AI has not changed
Writing code was never the bottleneck in delivering software. Four pieces into an argument about what the tools changed, here is the larger part of the job they left exactly where it was.
The cost of a wrong suggestion
Time saved by accepting a suggestion is immediate, visible and easy to count. The cost of accepting a wrong one is delayed, diffuse and lands on somebody else. Guess which one ends up in the productivity deck.
Tests are the specification now
When an implementation can be regenerated in a minute, it stops being the valuable artefact. The statement of what the system should do is what survives, and most codebases have never written one down properly.