Especially since GADTs don't do too well with type inference. My take on "premature optimization is the root of all evil" is that I will take any "cheap" premature optimization which doesn't impact readability, but I'd like to be sure there is an actual performance improvement first.