Can we change an AI model's architecture without retraining it from scratch?