snapshot
An image-first English learning app. Upload a picture, extract a word with AI, and keep the result in history, check-ins, and word books.
段晓坤
Building products, writing code, reading papers.
Most days I work on product ideas, read papers, and turn repetitive tasks into small tools.
Taste classifies, and it classifies the classifier. Pierre Bourdieu
About
Work, not persona.
I work across product, code, and research notes. If a workflow repeats often enough, I usually automate it.
Selected Work
Palo Wire
An AI-native source node for tracking high-signal Silicon Valley tech and VC writing, built for downstream agents instead of human browsing.
hammerspoon-config
My Hammerspoon setup for macOS. It handles windows, quick capture, app switching, and a few annoying little chores.
Writing
Current writing
-
A shorter entry point to the full robustness essay, reorganized into five linked posts for easier reading and sharing.
-
A short opening on the central paradox: high-performing systems can still fail under tiny, highly optimized perturbations.
-
A theory-first explanation of Bayes error, gradients, and smooth loss landscapes as the foundations of robustness.
-
From FGSM and PGD to 3D perception, adversarial viewpoints, and attacks on explanation systems.
-
A compact map of adversarial training, data-centric methods, certification, and efficient purification.
-
Why robustness becomes a systems problem once model architecture, deployment context, and asymmetric error costs enter the picture.