Why I built the Gradient Descent Failure Atlas
Gradient descent is easy to describe mathematically but harder to recognize when a training run is going wrong. I built this atlas to turn common failure modes into small, controlled experiments: the same deterministic data, stable and unstable learning rates, step-by-step playback, and shareable states. The goal was not to build another generic ML playground, but to create something I could use to inspect, explain, and remember why an apparently correct gradient can still produce a failing run.