Nearly Optimal Attention Coresets | Dark Hacker News