Wait, are we actually applying human neurological assessment tools to transformer-based language models and conflating architectural/training differences between model iterations with biological cognitive decline?