Optimum number of single points of failure — LessWrong