So basically it's a kind of overfitting with pattern matching features? This doesn't undermine the power of LLMs but it is great to study their limitations.