Interesting. I wonder if this is related to the model architecture and attention mechanism.

The author seems to be implying it could be: "Even a single mention of ‘code enhancement suggestions’ in our instructions seemed to hijack the model’s attention"