Honestly, I think it's an insane waste of computing power.  Even just considering GPUs/ASICs, which are limited to parallelizable computations -- there are a LOT of good scientific applications that people have figured out ways to parallelize.  (See, for example, the grid computing project Boinc: http://boinc.berkeley.edu/)  Why did the Bitcoin creators pick solving SHA-256 hashes over, for example, SETI@home?  SETI@home predates Bitcoin.

Primecoin and its derivatives are at least doing something sort of useful.  But still, finding Cunningham chains is about as close to the line of uselessness as you can get without falling over it...