The signal-to-noise problem: why breaking work down for LLMs is harder than it looks
Last week, I was reviewing an outline I worked on with an LLM agent for a technical article. It looked solid: comprehensive sections, logical flow, all the boxes checked. But something was off. On a hunch, I removed a section that was good (but didn’t quite fit) entirely. The result? Not only did the article flow better, but when I ran it through quality scoring, they all rated the shortened version higher. The very tools that helped created the meandering narrative recognized the improved signal when it was removed. ...