My thoughts too. Simulated annealing and similar strategies seem to intuitively be better is most cases where traditional gradient methods aren't applicable. I can imagine a handful of cases where genetic algorithms MIGHT be better, but even then I am not fully convinced and it just feels gimmicky.