> How is mathml, 10 years later helping a scientist produce a substantial
> text? A book, a 50 page handout for students?
It's probably not helping a lot in the production (unless it's a collaborative work with some people using tex, some MS Word, some openoffice, some... and you need a common communication format).
So let's assume for the sake of argument that the book is written in LaTeX.
In 1985 the author would have been happy to get that LaTeX typeset on to paper and the book published, but these days many people still want that, but also want online versions, version that meet legal or moral requirements for accessible renderings to audio or braille, versions that dynamically reflow in a web browser...
Conversion to mathml either offline with things like tex4ht or latexml or for more constrained input within the browser with asciimathml or mathjax is eminently possible and doesn't require you to throw away the latex skills.
> How can one integrate the millions of documents available at http://arxiv.org/
there are existing projects underway to convert sections of that to mathml
using latexml as having xhtml (or html5) versions of documents have many advantages when offered as an addition rather than a replacement for the exactly faithful but statically typeset TeX renderings to pdf.