The Limits of Self-Improving in Large Language Models | Dark Hacker News