Expected Attention: KV Cache Compression by Estimating Attention | Dark Hacker News