Imagine an AI that can answer your questions, no matter the language or subject. That's the goal of a new research project called M2QA (Multi-domain Multilingual Question Answering). Researchers have created a massive dataset of over 13,500 questions and answers, spanning three languages (German, Turkish, and Chinese) and three distinct topics (news, product reviews, and creative writing). This dataset is designed to challenge current AI models and push the boundaries of cross-lingual and cross-domain question answering. The team tested several leading AI models, including large language models (LLMs) like GPT-3.5, Aya 23, and Llama, as well as more traditional models like XLM-R. The results? While LLMs showed promise, there's still a long way to go. Interestingly, the study revealed that Chinese posed a greater challenge than German or Turkish, likely due to its unique linguistic structure and lack of whitespace between words. Furthermore, performance varied across different subjects, highlighting the difficulty of creating an AI that's truly knowledgeable across the board. This research underscores the importance of diverse datasets like M2QA in pushing the limits of AI understanding. By training models on a variety of languages and topics, researchers hope to create AI systems that are more robust, adaptable, and ultimately, more useful in our multilingual world. One intriguing finding was the impact of typography. When spaces were added between words in Chinese text, the performance of certain models actually improved, suggesting that even visual elements can play a role in AI's ability to understand language. The next steps? The researchers plan to expand M2QA with more languages and topics, particularly focusing on low-resource languages that often get overlooked in AI research. They also aim to explore new methods for transferring knowledge between languages and domains, in hopes of closing the gap between AI and human-level understanding.