Field NotesHonest essays from researchers about what's working, what isn't, and what we're learning. We publish negative results, retrospectives, and field reports — not just the wins.
A retrospective on tone modeling, why our F1 collapsed in Eastern Province, and the dataset patch that fixed it. The short version: we treated tone as out-of-band when it carries lexical disambiguation for common verb stems. Annotating a 40-hour subset and using tone as an auxiliary CTC target moved Eastern Province WER from 19.1% to 13.6%.
Read the full post →Field deployments humble models fast. Fourteen farmers, two cooperatives, one growing season. Here's what they taught us about UX (the camera shutter sound matters more than we expected), latency (anything over half a second loses them), trust (a model that says 'unsure' is more trusted than one that's always confident), and dawn lighting (it ruins everything).
Read the full post →Calibration matters more than accuracy in clinical decision support. We deliberately route 18% of cases to a clinician review band. The internal critique was that we were 'wasting' high-confidence predictions. The clinical critique — which we agreed with — was that an over-confident model in a low-resource setting causes more harm than a usefully cautious one.
Read the full post →What's signal, what's noise, and where the next gains will come from. We unpack the SahelClim benchmark, why station coverage matters more than fancy architectures right now, and what we'd ask of the next dataset if we could write it from scratch.
Read the full post →What we look for, how we score, and how to make a strong application. We read every application we receive. The strongest ones are clear about a single question, honest about scope, and concrete about the community the work is for. The weakest are sprawling and hedged.
Read the full post →Quantization, pruning, and the unsexy systems work behind a model that runs on a $40 phone. A walk through the cassava classifier stack from TFLite conversion to on-device thermal management. Includes the three things we tried that didn't work.
Read the full post →If we can't re-run your training in a notebook six months later, it didn't happen. The reproducibility checklist we now require on every release, and why we made it stricter than the conferences we publish in.
Read the full post →Notes from the FAccT paper, but for practitioners. What 'community consent' actually looks like when you're collecting 1,200 hours of speech. The forms, the panels, the takedown processes, the things we keep getting wrong.
Read the full post →One email per month. New papers, deployment notes, and open calls. No tracking.
Subscribe