
Context windows and performance degradation

Working Effectively with Legacy Code and AI Coding Assistant - Michael Feathers
Topic 17 of 22
Published Oct 14, 2024 · Clip 45:19–47:14 (1:55)
Michael explains context window limitations: inference is roughly quadratic in memory use, so long sessions degrade—models forget things given minutes earlier—a phenomenon he calls performance degradation. He notes newer models like Gemini claim near-infinite context, but he expects attention limits to persist, comparing it to human cognitive science: attention is relatively fixed, and understanding what's in attentive focus is key to being effective with AI.