Thinking out loud, in public.
Essays, research notes, and weekly logs — where ideas get tested before they become papers. Writing is how I find out what I actually think.
On the cost of an explanation
An explanation no one has time to read is not an explanation. A piece on what we actually owe the expert in the room — and why the right unit of interpretability is the decision, not the model.
"Interpretability isn't a property of the model. It's a property of the moment a human has to act."
All writing
12 piecesOn the cost of an explanation
What we owe the expert in the room — and why time is the real constraint on interpretability.
Augmentation is a design stance, not a hedge
Why "augment, don't replace" is a concrete set of choices — about interface, authority, and accountability.
Synthetic data is a modeling assumption, not a free lunch
Every generated sample encodes a belief about the world. We should say which belief, out loud.
Week 41 — three dead ends and one good question
A log of what didn't work, which is usually where the signal is.
Calibration is a UX problem
A probability is only useful if the person reading it knows what to do with it.
What agents are actually for
Autonomy is not the goal. Leverage with legibility is.
More in the archive — full posts coming as the site grows.