In a groundbreaking move for the AI industry, Silo AI, in collaboration with the University of Turku’s research group TurkuNLP and HPLT, has unveiled Viking 7B, the largest open-source language model encompassing all Nordic languages. This model not only covers Nordic languages but also includes English and several programming languages, marking a significant milestone in multilingual AI applications. The release aims to address linguistic and cultural biases prevalent in existing models, thereby promoting digital inclusivity and sovereignty across Europe. This development is expected to substantially enhance the quality and accessibility of AI-driven services in various linguistic domains.
Silo AI, founded in 2017, is a prominent European AI lab specializing in providing custom AI solutions for various industries. Known for its expertise in machine learning, computer vision, and natural language processing, the company aims to democratize AI by developing robust and inclusive language models. By leveraging advanced AI technologies, Silo AI strives to offer scalable and adaptable solutions that cater to diverse linguistic and cultural needs.
Previous Developments in AI Language Models
Earlier reports have highlighted the dominance of English in AI language models, often leading to biases and limited utility for non-English speaking users. Models like GPT-3 and BERT have primarily focused on high-resource languages, making them less effective for low-resource languages. Compared to these, Viking 7B presents a significant advancement, offering superior performance across multiple languages without compromising on quality. This inclusive approach not only broadens the scope of AI applications but also mitigates the linguistic biases observed in prior models, ensuring a more equitable digital landscape.
Enhanced Performance and Inclusivity
Viking 7B’s training approach mirrors that of its predecessor, Poro, focusing on low-resource languages. It extends to include Danish, Finnish, Norwegian, Icelandic, and Swedish, alongside English and programming languages. The model’s architecture has been updated, ensuring best-in-class performance in Nordic languages while maintaining excellence in English. This balance addresses the so-called ‘curse of multilinguality’ and enables Viking 7B to serve as a robust, inclusive AI tool.
Commitment to Digital Sovereignty
Silo AI emphasizes the importance of digital sovereignty, ensuring that the data used for training Viking 7B accurately represents European languages and cultures. The model is trained on the EuroHPC supercomputer, LUMI, which is notable for its sustainability and computational power. This approach not only enhances the model’s performance but also aligns with European values, promoting a digitally sovereign and culturally inclusive AI infrastructure.
Implications and Future Directions
Viking 7B’s release underscores the potential for AI to bridge cultural and linguistic divides, fostering digital inclusivity. The model’s sensitivity to local values and cultures ensures that AI-driven communication tools serve as connectors rather than dividers. This development is expected to accelerate the adoption of AI applications across Europe, offering valuable solutions for various sectors.
User-Usable Inferences
Silo AI’s Viking 7B is a significant leap towards creating inclusive and high-performing language models that cater to multiple languages and cultural contexts. Unlike previous models that predominantly focused on English, Viking 7B ensures balanced performance across low-resource languages, addressing the inherent biases in AI. The model’s reliance on LUMI, Europe’s most powerful and sustainable supercomputer, further underscores its commitment to digital sovereignty and environmental sustainability. This initiative marks a pivotal step in the evolution of AI, setting a new standard for inclusivity and performance in multilingual AI applications.
