An embedding is a series of numbers that have been gradually shifted to better fit some purpose. The gradients tell me that if I increase the first number of embedding X a little, the model will perform better, so I do.