Large language models (LLMs) have revolutionized how we interact with technology, but their prowess has largely been confined to English. What about languages with fewer resources and different linguistic structures? Researchers tackled this challenge head-on in a new study focusing on adapting Llama, a powerful open-source LLM, to understand and generate Persian. Their multi-stage approach involved expanding Llama’s vocabulary with Persian tokens, pre-training it on a massive dataset of Persian text and code, and then fine-tuning it with bilingual English-Persian instruction data. The results reveal a fascinating interplay between cross-lingual knowledge transfer and the importance of a model's initial strengths. While incorporating Persian improved Llama’s performance on Persian classification tasks like sentiment analysis and textual entailment, it sometimes hindered its English text generation abilities. This suggests a delicate balance—adapting to a new language can broaden an LLM’s horizons but also requires careful calibration to avoid diluting its existing skills. The study's findings highlight the complexities of making AI truly inclusive. While direct translation of instructions and inputs boosted performance on some tasks, it wasn’t a universal solution. For complex tasks like summarization, aligning the input language with the model's strongest language proved more effective. Furthermore, when training data is scarce, an LLM’s pre-existing proficiency in the target language becomes even more critical. This research offers valuable insights into the challenges and opportunities of extending LLMs to a wider range of languages. It underscores the need for innovative training strategies and more diverse datasets to build AI that truly reflects the richness and diversity of human language. The journey to create globally accessible AI is ongoing, and this study provides a significant step forward in unlocking the potential of LLMs for languages like Persian.