Currently, np.empty(..., dtype=object) fills the array with None, which is certainly a very reasonable (though undocumented, afaict) behavior. Is this guaranteed? (Note that the other "more obvious" way of creating such an array, np.fill(..., None), does not work, see #6366.)