Python's reliability here comes because it is a much less flexible language in some ways. If you write your own array type in python, and pass it into tensorflow, you would expect it to error. If you do the same thing in Julia, you would expect it to work.