Overfitting

Learning the training data instead of the underlying task — great scores in practice, poor on new data. Its LLM-era cousin: prompts tuned to death on five examples.