Fine-tuning large language models (LLMs) like ChatGPT or Bard is like teaching a star student a new subject—they excel, but sometimes forget their core values. This 'catastrophic forgetting' in AI means that while an LLM becomes great at a specific task (like medical advice or writing code), it might also become more likely to generate harmful or inappropriate content, undoing its initial safety training. Researchers have been grappling with this safety vs. performance trade-off, often resorting to complex and resource-intensive methods like adding more safety data during fine-tuning. A new study from National Taiwan University and Intel Labs has found a surprisingly simple yet effective way to boost both safety and performance: merging the model's 'before' and 'after' states. Imagine blending the LLM's original, safety-conscious self with its new, specialized expertise. This merging technique essentially acts as a refresher course on safety while preserving the newly acquired skills. The researchers tested this approach across various LLMs, tasks (reasoning, medical advice, code generation, and tool usage), and merging techniques. They found that merging significantly reduced harmful outputs by up to 30% while maintaining or even improving performance on the intended tasks. This is a major step forward because it provides an efficient solution to a pervasive problem in LLM development. While the results are encouraging, some limitations exist. The study primarily focused on benign datasets, and the effectiveness on other domains or larger models needs further exploration. Also, the safety evaluation method used, while cost-effective, may not capture all nuances of harmful content. This research opens promising avenues for improving LLM deployment. Future studies could explore the effect of this merging technique on bias mitigation and delve into more granular safety analysis. The simplicity and effectiveness of the method make it a potential game-changer in making LLMs both smarter and safer.