Posts
Industrial machine learning has some particular challenges such as scarce labels, skewed data, costly exploration and poor observability. Not all is lost, though: there are a number of techniques to tackle them.
A field guide to how machine learning models fail in practice — from problem-definition exploits to distribution shift — and how to guard against each.
Data only has value in combination with a problem — why the amount of data you need depends on model complexity, not on collecting as much as possible.
Why 'coding agent' undersells the job, and what it means to direct rather than micromanage AI-assisted development.
