Signals
Shorter pieces on method, position and the occasional correction of something the field believes without checking. Research reports what we found. Signals report what we make of it, and the two are kept apart on purpose.
Adoption is the wrong thing to measure
Use is near universal and tells you nothing. The question that matters is what stops working when you switch it off.
A category map cannot survive a participant funding it
One paid box does not corrupt that box. It corrupts every other box, because the reader now has to ask which ones were paid for.
The ceiling is arithmetic, not opinion
Decompose the work, find the share the system actually touches, and the maximum holds regardless of how good the model is.
Most enterprise AI is failing a supervisory guidance from 2011
Model risk management predates this wave by more than a decade. In financial services it is still doing most of the real work.
The most quoted number in enterprise AI rests on one study
Unreplicated, inconsistently attributed, and treated as settled. That is a fair description of the whole evidence base.
Working AI is boring AI
Card fraud and industrial inspection have run continuously for decades. Neither uses a language model. Neither calls itself AI.
Six pieces of research a week. One email on Sunday.
An organisation mapped, a failure documented, a category defined, an obligation explained, a company running AI in production, and an investor mapped by position. Free, and the archive is public.