This is a dumb question, I'm sure: Does the loss function have to be a function of the derivative somehow? Or does being differentiable just allow gradient des

This is a dumb question, I'm sure: Does the loss function have to be a function of the derivative somehow? Or does being differentiable just allow gradient descent learning to work?