Juho Snellman is a smart cookie.

SBCL is a large, complex beast: I've talked enough about how hard it is for a system to give the right answer, but when people start wanting it to give the right answer quickly, well... Juho Snellman has single-handedly found improvements to string hashing routines, to poor scaling behaviour with multiple threads, and now bignum division. It appears not to solve the performance problems that we're having with GSharp (which are probably algorithmic in nature: why are we doing arbitrary-precision calculations, anyway), but I should hope that some of the graphs in Andreas Fuchs' boinkmarks (once the nightly run has completed) will look interesting.