'Self-improving AI, an analysis' [link] — LessWrong