🧪 Learning Rate Schedules

Most learning-rate explainers only ever plot η against epoch — a 2D curve that never shows what the schedule actually does to training. This simulation puts a weight vector on a real rippled 3D loss surface and runs true gradient descent on it, one epoch at a time, with the step size controlled by whichever schedule you pick: constant, step decay, cosine annealing, or warmup followed by cosine annealing. The instanced trail traces the exact path the optimizer took, and the loss readout is the exact value at the current position — so you can watch, rather than infer, why a schedule that decays too fast gets stuck in the first ripple it meets, why a constant high rate keeps oscillating, and why cosine annealing (with or without warmup) is the schedule most training runs default to.