When does adding more people reliably make a system better? — LessWrong