This seems to only really work in languages that allow null variables/timestamps. I wouldn't really want to have to do comparators to the default value of a timestamp.
I do see a different issue, though: The article indeed seems to make no distinction between an absent value and a default timestamp of 0. That limits your database to more or less "now". You cannot really store things about the past. Someone might take such a pattern and fixate it into some kind of library. If then someone else tries to store data from 1970, things can get ... interesting.
Unfortunately, datetime takes 8 bytes vs the 4 for a timestamp.
Assuming the timestamp represent a change of state in a contemporary application, I would expect 1970-01-01 0:00:00Z (UNIX epoch +0 seconds) to be unambiguous enough (But that's definitely an engineering constraint to maintain awareness of)