Blog
Notes on (explainable) AI, computer vision, open data and mapping, with occasional excursions into whatever's coming to my mind.
Rooftop Solar Needs an Open Registry
Rooftop solar is a blind spot in energy data. Why I think it needs an open, regularly updated registry, how DeepPVMapper and PV Check contribute to it, and what 213 volunteers checking detections make possible.
Read the article →If in Doubt, Say So
Ten volunteers vote on each PV detection. Where they hesitate matches where the model is less reliable, and the crowd is more cautious than a confident detector: it rejects a few detections the detector is sure about.
Read the article →Navier–Stokes, Scaling Laws, and Peak Tokens
OpenAI's Navier–Stokes result quietly admits that user prompts may be training data too. A look at why tokens — not energy or money — could be AI's binding constraint, and why that last reservoir is set up to run dry.
Read the article →Hidden in Plain Sight
Something visible from the sky or the street isn't made safe by hiding the dataset — that only decides who gets to see it. Why openness is the robust choice for mapping infrastructure like rooftop solar.
Read the article →The Blind Men and the Solar Panel
Reconciling grid connection data, remote sensing, and OpenStreetMap: why comparing PV maps means agreeing on what a PV installation even is.
Read the article →What Does a Solar Panel Detector Actually See?
Using wavelet decomposition to explain what a deep learning model actually detects, and why DeepPVMapper turned out to be a grid detector.
Read the article →