Large Language Models have Intrinsic Self-Correction Ability