The tension between successfully removing harmful knowledge and maintaining useful model capabilities.