Charts are everywhere, conveying complex information at a glance. But for AI, deciphering these visual puzzles is a real challenge. Large Language Models (LLMs) like GPT-4 show promise, but how can we transfer this ability to smaller, more resource-efficient Multimodal LLMs (MLLMs)? Researchers have developed a clever technique called Code-as-Intermediary Translation (CIT) that uses code as a bridge between visuals and text. Imagine translating a chart's visual structure into a language LLMs understand. This method allows researchers to create synthetic datasets of charts and questions, effectively teaching MLLMs to 'see' and 'reason.' The results are impressive, with MLLMs showing significant improvements not just in chart understanding, but also in general mathematical reasoning. This approach suggests a future where even lightweight AI can interpret complex visual data, opening doors to applications in education, data analysis, and beyond. However, challenges remain, including extending this method to natural images and ensuring the complete accuracy of the synthetic data. The journey to truly intelligent AI that perceives the world as we do is ongoing, but CIT represents a significant step forward.