Applying superintelligence without collusion — LessWrong