Trustworthy AI is a systems problem.
The model is only one part. I trust the system more when it has clear limits, shows its work, can recover, and someone owns the last action.
Read the full note ↗Field notes
Ideas tested against the work.
Practical writing about AI systems, product judgment, automation, and what changes when a prototype has to operate reliably.
Publishing the record changes how you build.
A live record forces clearer data handoffs, visible caveats, reproducible updates, and a sharper distinction between an outcome and a story about that outcome.
That pressure is useful. It moves credibility away from confident language and toward work people can examine for themselves.
Small products sharpen judgment.
A focused consumer product exposes decisions that broad technical demos often avoid: who is the user, what friction matters, what emotional context is easy to mishandle, and which features should never be built.
Constraint is not the enemy of ambition. It is often where product judgment becomes visible.