Taming AI Bots: Controllability of Neural States in Large Language Models