Information work is bottlenecked by high-signal context more than by inference (twitter.com)

This kind of explains why LLMs are good at making demos and good at writing code under clear specification, but often break existing features in a large codebase. When codebase become large, the signal-to-noise ratio of context reduces.