Imagine a language with only a handful of speakers left, teetering on the brink of extinction. Now, imagine artificial intelligence stepping in to help preserve it. That's the fascinating premise behind new research focusing on extremely low-resource languages, specifically those in the Finno-Ugric family like Võro, Livonian, and Komi. These languages, spoken by communities around the Baltic Sea and Ural Mountains, face a significant digital hurdle: the lack of data needed to train powerful AI models. Researchers tackled this challenge head-on by building LLMs (large language models) almost from scratch. They gathered every scrap of digital text available, from Wikipedia dumps to scraped newspaper articles. Even Seto, a closely related dialect, was included to bolster the meager data pool. For Livonian, with a mere 30 speakers, they pieced together sentence fragments and parallel translations to give the AI a fighting chance. The team used a two-stage training process. First, they trained a base model on related, higher-resource languages like Estonian and Finnish. Then, they fine-tuned this model on the precious Finno-Ugric data, carefully balancing the mix to avoid bias toward the more common languages. They even experimented with different data repetition techniques to maximize the impact of the limited information. The results? Surprisingly positive. The AI models showed a nascent understanding of these endangered languages, demonstrating the potential of AI to become a vital tool for language preservation. They were tested on various tasks like question answering, topic classification, and even translation. While not perfect, the models performed comparably to existing commercial systems in some areas and even outperformed them in others, particularly in capturing the natural flow and nuances of Komi. Perhaps most importantly, human evaluations revealed that these AI-generated texts felt authentic to native speakers, a crucial step towards making these models genuinely useful for the communities they aim to serve. However, the research isn't without its limitations. The small datasets introduce uncertainties, and evaluating AI's performance in such low-resource settings is tricky. There's also the inherent ethical consideration: ensuring the AI doesn't inadvertently create harmful content or misrepresent these fragile languages. The journey of preserving language with AI is just beginning, and this research provides a crucial stepping stone. Future efforts will focus on creating more comprehensive benchmarks and expanding this approach to other endangered languages worldwide. The ultimate goal? To give these languages a voice in the digital age, ensuring they are not lost to the sands of time.