TL;DR:
- Silo AI introduces multilingual LLM Poro 34B, addressing language bias in AI.
- Poro 34B exhibits best-in-class performance for low-resource languages while excelling in English.
- Future expansion includes support for Nordic languages.
- SiloGen partners with LAION to add multimodal capabilities to Poro 2.
- Poro 2 will be freely available under the Apache 2.0 License, encouraging innovation.
Main AI News:
A year after the introduction of ChatGPT by OpenAI, which brought terms like “foundational model,” “LLM,” and “GenAI” into the mainstream, the anticipated benefits of generative AI remain disproportionately skewed toward English speakers. While there are over 7,000 languages spoken worldwide, most large language models (LLMs) excel primarily in English. This imbalance poses a significant threat, as it has the potential to exacerbate language bias and limit access to knowledge, research, innovation, and competitive advantages for businesses.
In November, Finland’s Silo AI took a groundbreaking step by unveiling its multilingual open European LLM, Poro 34B, developed in collaboration with the University of Turku. Poro, which translates to “reindeer” in Finnish, underwent training on Europe’s most powerful supercomputer, LUMI, located in Kajani, Finland. Notably, LUMI operates on AMD architecture, distinguishing itself from the prevailing trend of LLM training on Nvidia systems.
Accompanying Poro 1’s launch, Silo AI introduced a research checkpoint program that plans to release checkpoints as the model progresses, with the first three points announced alongside the model’s debut. Recently, SiloGen, a subsidiary of Silo AI, reported achieving over 50% model completion and has now unveiled the next two checkpoints in the program. These five checkpoints collectively establish Poro 34B as a best-in-class performer for low-resource languages like Finnish, surpassing alternatives such as Llama, Mistral, FinGPT, and others, all while maintaining high performance in English.
Sampo Pyysalo, a Research Fellow at TurkuNLP, anticipates that the model will reach full training completion in the coming weeks. The next phase of development will see support added for additional Nordic languages, including Swedish, Norwegian, Danish, and Icelandic.
Peter Sarlin, co-founder and CEO of Silo AI, emphasized the importance of aligning language models with European values, culture, and languages for Europe’s digital sovereignty. He expressed pride in Poro’s exceptional performance in handling low-resource languages like Finnish and noted the natural progression toward expanding support to Nordic languages.
Furthermore, SiloGen has embarked on training Poro 2 in collaboration with the non-profit organization LAION (Large-scale Artificial Intelligence Open Network). This partnership aims to introduce multimodality to the model, encompassing both text and vision data.
Sarlin emphasized the significance of extending Poro’s capabilities to incorporate vision, recognizing the immense potential for generative AI to consolidate diverse data modalities. LAION shares a similar commitment to advancing the field of machine learning for the greater good. To align with Silo AI’s mission of democratizing AI models and LAION’s goal of increasing access to large-scale ML models and datasets, Poro 2 will be made freely available under the Apache 2.0 License. This decision enables developers to build proprietary solutions on top of Poro 2, fostering innovation and accessibility.
Established in 2017 as “Europe’s largest private AI lab,” Silo AI has aimed to become a flagship for AI innovation in Europe. Headquartered in Helsinki, Finland, the company focuses on developing AI-driven solutions and products that empower smart devices, autonomous vehicles, industry 4.0, and smart cities. With a workforce exceeding 300 employees, Silo AI has expanded its presence to Sweden, Denmark, the Netherlands, and Canada, further cementing its role as a frontrunner in AI research and development.
Conclusion:
Silo AI’s release of Poro 34B and its commitment to extending language support, in collaboration with LAION for multimodal capabilities, signifies a significant leap towards democratizing AI. This development opens doors to more inclusive AI innovation and market growth, aligning with the evolving demands of a diverse global landscape.
