Unpacking the ingredients of atomic representations: machine learning force fields and beyond