Comparative tests of Anthropic's Opus 4.8 show "High" reasoning effort outperforms "Max" setting due to context compaction issues
LightCell's Danielle Fong warns context compaction degrades long-term recall
——0——
you can in fact frame this as a compression problem where a generator learns to summarize some prior sequence in such a way that minimizes the conditional distribution drift (as measured by kldiv) instead of bolting on a summary prompt posthoc and hoping it tracks salience
kalomaze@kalomazeit seems crazy to me how uninterested ant is as an org when it comes to compaction as a methodology
7:52 PM · May 28, 2026 · 14.6K Views
7:54 PM · May 28, 2026 · 1.5K Views