Writing
Writing
Short essays working out, in prose, the arguments the case studies only imply. Where AI weather models fail, why they fail there, and what an honest evaluation of the failure looks like.
-
Why AI weather models miss the extremes
Regression-trained weather emulators regress towards the mean, which thins their forecast tails by construction. This is a short account of why that happens, how to read the deficit directly off the fitted tail, and what a calibrated tail evaluation reveals once you stop scoring the average day.
Read → -
Defending a negative result
A positive result obtained under pressure can rest on one lucky split; a negative result attacked from three independent directions and left standing is the more robust piece of science. This is why Cascade's central finding, a measured boundary of predictability, is defended harder than the result that worked.
Read →
Forthcoming
- What operational deployment asks of research software
- Recasting air–sea coupling from a correlation into a mechanism
- Spectral blur, and the skill it costs downstream