Context rot
The way a model's answers get worse as its input grows longer, often well before the context window is full.
A model can accept a long input without handling all of it well. Researchers testing 18 models found that performance drops as inputs grow longer, even on simple tasks and well within the limit. Older work found models handle facts at the start and end of a long text better than those in the middle, hence the nickname "lost in the middle".
Long conversations suffer in a similar way. In one large study, top models did 39% worse on average when a task was given across several messages instead of one. They commit to early assumptions and keep building on them, even after you've corrected course.
For example
2 hours into planning a wedding in one chat, the assistant suggests a venue you ruled out in the first 10 minutes.
Put it to use
Where you'll meet context rot in ChatGPT, Gemini and Claude, how to use it there, and the mix-up most people make.
Daily Bytes Plus is in the Android app.