How the virtual AI controls itself — LessWrong