To protect privacy and interpretability we need pre-training, fine-tuning and deployment of models without permission. https://arxiv.org/abs/2404.15834