用 20% 記憶體超越 RAG：ASM 在 NBA Benchmark 的延遲優化
長上下文推理面臨記憶體與延遲的雙重瓶頸。本文解析 Attention-State Memory (ASM) 如何透過外部化預計算狀態，在 NBA Benchmark 等特定場景下，以約 20% 的 RAG 記憶體佔用實現性能超越，並探討其適用邊界。
RAG
Surpassing RAG with 20% Memory: ASM Latency Optimization on the NBA Benchmark
Long-context inference faces two constraints at once: memory use grows with the prefix, while latency rises at every decode step. Attention-State Memory (ASM) offers a training-free alternative by externalizing precomputed attention states into a lightweight lookup-based memory. On the NBA Benchmark, it exceeded full-attention RAG performance using about 20% of the memory. This article explains ASM’s hierarchical lookup and online-softmax merge, then maps the boundaries that matter in deployment: query-distribution stability, prefix updates, offline construction, and codebook-size tuning.
Not Every Company Should Build RAG—Here’s How to Know
Your boss asks when you can have an AI-powered knowledge base. This article gives you a framework: a 9-point readiness checklist, a decision matrix, and scripts for explaining—in language executives understand—why doing it now might be right or wrong.
不是每間公司都該做 RAG——這是判斷標準
老闆問你「我們什麼時候能有 AI 知識庫」？這篇給你一個框架：9 項就緒度檢核、決策矩陣、以及怎麼用老闆聽得懂的語言，說明為什麼「現在做」可能是對的——或錯的。
5 Pitfalls You’ll Hit in Your First Month Building RAG
Assigned to build a RAG PoC? This is not about choosing vector DBs. It is about the 5 pitfalls you will definitely hit: chunking strategy, data quality, success metrics, maintenance costs, and user expectations. Lessons paid for in production outages.
RAG 專案的第一個月，你會踩的 5 個坑
被指派做 RAG PoC？這篇講的不是怎麼選 vector DB，而是你第一個月一定會踩的 5 個坑：chunking 策略、資料品質、成功定義、維運成本、用戶期望。都是花錢買來的教訓。