Dharma Insights — Operational№ 114 · AI Systems
← The Signal№ 114 · AI Systems · October 13, 2025 · 1 min read

Delethink

Modern LLMs can “think” for long chains, but traditional methods are quadratically expensive, making long reasoning slow and memory-hungry. Delethink changes the game: it chunks reasoning into fixed-size states and…

Modern LLMs can “think” for long chains, but traditional methods are quadratically expensive, making long reasoning slow and memory-hungry.

Delethink changes the game: it chunks reasoning into fixed-size states and passes a short summary (Markovian state) between chunks. This decouples reasoning length from context size, enabling long-horizon thinking without ballooning compute.

The result? Smaller models match or surpass traditional long reasoning performance, scale to 128K+ tokens, and maintain constant memory usage.

This is a glimpse into the future of efficient, scalable AI reasoning — thinking longer doesn’t have to cost more.

View all signals →