Once a dangerous AI model is released it cannot be taken back, whereas development can be restarted after improving safeguards