Written down, so it reads like common sense.
Notes on technical leadership, delivery and applied AI, from twenty years of building software where it had to actually work.
The model was never the hard part
Creating AI feature that writes to a few thousand people. No quality difference between models at five times the price, prompt rules obeyed about half the time... if you're lucky.
The AI Act deadline that moved isn't the one that matters
The expensive obligation slipped sixteen months. A cheap one arrived exactly on time, applies to software you already shipped, and falls due in December with nobody assigned to it.
Can your AI vendor prove what their agent can reach?
Three organisations had frontier models try to get out of an evaluation environment this July. Two of them lost. The difference was not the model, and I think that is the useful lesson.
You still can't outsource responsibility
Hertz sued Accenture for $32m after letting go of the people who could have told it the work was wrong. Teams are now doing the same thing to themselves, and there's nobody left to sue.
Five years on, some people still live in goddamn vans
I argued in 2021 that public sector software costs what it does because government can't choose its users. That's still true. What's changed is that the private sector's exemption is running out.