The Optimal Architecture for Small Language Models | Dark Hacker News