Safe probability manipulation, superweapons, and stable self-improvement research — LessWrong