This is a visualizer of the loss surface for softmax. The weights for each class are shown with purple dots; it's assumed that bias is zero. The larger light-purple dot is the weight vector of the "true" class. Class weights are normalized. Each of the class weights are draggable. The embedding z is also draggable; its loss value is shown. Moving it does not affect the overall error surface, which is drawn in the background in pen. Lower loss is dark, and higher loss is light.
Here is a (advanced, much math) video about softmax cross entropy loss for training neural networks: https://www.youtube.com/watch?v=PHP8beSz5o4