Imagine translating English into Icelandic, a language as rich and complex as the sagas it holds. That’s the challenge the Árni Magnússon Institute for Icelandic Studies (AMI) tackled for the World Machine Translation (WMT24) competition. Their approach? A blend of cutting-edge AI, clever data filtering, and a dash of old-school linguistic know-how. Instead of relying solely on massive AI models, which can be resource-intensive, AMI used four smaller, specialized translation models working in concert. Think of it like an orchestra: each instrument plays its part, and together they create a richer, more nuanced sound. Training these AI ‘instruments’ required a carefully curated dataset of English-Icelandic text. AMI aggressively filtered out 'noise' like inaccurate translations or misaligned text. They also tapped into the power of large language models (LLMs) to generate additional synthetic data, effectively expanding their training material. One surprising finding was that adding more data didn’t always lead to better results. Low-quality data, like rough machine translations, could actually harm the system’s performance, highlighting the need for precise, high-quality training data. AMI's approach proved competitive, showcasing the effectiveness of combining smaller, efficient models with strategic use of LLMs and rigorous data cleaning. This project not only advanced Icelandic translation but also provided valuable insights into the complex relationship between AI models and the data that fuels them. As researchers continue refining these techniques, we can anticipate even greater leaps in machine translation for Icelandic and other less-resourced languages, bridging cultures and unlocking knowledge across linguistic divides.