perhaps this approach can lead to better training of foundational models?..