Deep Dive: Why Attention Mechanisms Are Hard to Replace

akbirthko · x · 2026-08-14

Amidst discussions sparked by recent papers from Kimi and DeepSeek regarding attention mechanisms (like SWA vs. KDA), a researcher analyzes why explicit addressable history as the primary memory mechanism is unlikely to be replaced in the near term.

The core argument is that as long as LLMs rely primarily on Maximum Likelihood Estimation (MLE) over internet data to acquire domain knowledge, attention variants are still performing the same job. Even if they achieve log-depth with respect to sequence length (e.g., associative scan or chunkwise parallelism), they remain giant constant-depth circuits with respect to real-world time.

Only after moving past internet-based pre-training and entering serial scaling phases (like RL) will models genuinely need to tackle complex issues like context compaction, state tracking, and continual learning. Until then, the end-user experience remains that of a finite token buffer being filled and compacted.

Original post →

More from Research

Research channel →