I think the short answer is that it's just a massive amount of difficult work to get one of these high-performance GC algorithms implemented, and it has to be h

I think the short answer is that it's just a massive amount of difficult work to get one of these high-performance GC algorithms implemented, and it has to be highly customized to the language implementation. Also, the original implementations of Python and Ruby are so slow in general that it's not clear that improving the GC is the best use of one's time. In the open-source Common Lisp world, CMUCL and SBCL have for years used Douglas Crosher's multigenerational collector. Guido and Matz would have done well to use it too -- it's much better, probably, than either of theirs. It's still far behind the JVM, but the point is, Python and Ruby didn't necessarily need the very latest research; they just needed 1995 technology instead of 1975 technology.