Evaluating an AI model's performance requires subjective 'feel' from practical use, as benchmark scores can be misleading