There's a problem here in going from their specific case to where this may make sense, to the more general case they're trying to make, where it doesn't.

For instance, if you're mapping disk space to memory addresses and have large blocks of data, this would mean that every time you needed to update a single byte, you'd either have to duplicate the entire block (possibly several mega- or giga-bytes) or give up linear access. Yuck. There are a lot of use cases that would absolutely be miserable for.

You can get away with such things a lot further if you're implementing these things at the file-system level.