Writing
Long-form on AI, cloud, strategy, & technology.
8 pieces · by topic
- Evaluations are not testsA test tells you whether the code did what you wrote. An evaluation tells you whether the system is still good enough. Confusing the two produces suites that pass while the product gets worse.
- Why AI pilots die in model risk reviewSR 11-7 wasn't written for systems that change weekly. Here's the tiered validation pattern that gets agents through second-line review.
- The AI skills gap is a job-design problemEveryone is hiring for the scarcest skill on the org chart. The role that decides whether a system survives its first year is the one nobody has written a job description for.
- The AI centre of excellence is a transition stateCentralising capability is the right first move and the wrong steady state. The question worth arguing about is what the CoE is supposed to make itself unnecessary at.
- Your model is only as auditable as its lineageThe question "what data is this built on" gets asked at review time, and it cannot be answered retroactively. By then the honest answer is a shrug with a confidence interval.
- Your cloud bill is a design review you keep skippingCost anomalies are almost never pricing problems. They are architecture decisions arriving late, in a format nobody on the engineering team reads.
- The business case that survives second lineMost AI business cases are built on hours saved and die the moment somebody prices the controls. The number that matters is the one nobody put in the model.
- Retrieval is not a knowledge baseTeams ship retrieval as though it were search with better manners. It is a data product, and it fails in the ways data products fail — quietly, and mostly on freshness.