He argues in favour of using the expert assembly knownlegde one might have in order to help to enhance the optmization or the {IR to assembly} steps instead of keep repeating this work manually, which is a strong and good argument and he closes saying:
I think it's amazing that once written pseudo-assembly code can be retargetted to any architecture. For me IR has all the advantages of assembly without any of its problems: fast, expressive, retargettable and maintainable.
