QualityTraining data and the judgement gap
Why the next leap in model quality will not come from more tokens, but from more disagreement — and how to elicit the right kind of disagreement from experts.
May 20267 min readRead
Blog
Notes from the network
Essays on training data, quality, process, and the practical work of building reliable AI.
Latest posts
What we are writing about
One or two essays a week from our team and senior experts.
Quality
ProcessWhat rubrics can — and cannot — do
Rubrics are how we make judgement teachable. They are also the easiest thing to over-engineer. A short essay on the line between calibration and bureaucracy.
May 20266 min readRead
EarningsPaying experts fairly is harder than it looks
A breakdown of how we set rates: market signals, project stakes, tier multipliers, and the deliberate decisions we made to avoid race-to-the-bottom dynamics.
Apr 20269 min readRead
SafetyRed-teaming medical LLMs without harming patients
How we structure red-team programmes for clinical reasoning models: gating, wellbeing, escalation paths, and the protocols that keep adversarial work safe.
Apr 202611 min readRead
StrategyThe case for vertical expertise in AI training
Generalist annotators have a ceiling. We argue for a sharper division: subject-matter specialists for substance, linguists for tone, calibrated reviewers for both.
Mar 20268 min readRead
ProcessA month on the feedback loop
What changed when we forced every accepted task to carry a rubric note back to the expert. Spoiler: dispute rate dropped, tier progression accelerated.
Mar 20265 min readRead
Have something to say?
Pitch us an essay
Guest posts from Lona experts and serious practitioners are welcome.