I'm wondering if the solution to the Sufficiently Smart Compiler is to have the Sufficiently Stupid Language. Automated parallelism becomes easier if you only give users parallelizable operators. That sounds terrible, but maybe it isn't that bad (I don't know). Perhaps users should always be working on sets of data (a number doesn't a exist, just a set with one number in it), then the compiler really just has to be smart enough to turn that set operator into just an operation on an integer. That is, deparallelize it. Is this why array languages tend to work well for these things?