Last week, the annual collaboration meeting of TrustLLM, a research project funded by the European Union's Horizon research program, took place over 3 days. This is a collaborative project involving eleven partners from across Northern Europe, including the University of Iceland and Miðeind ehf., as well as institutions such as Fraunhofer in Germany and AI Sweden in Sweden, to name just a few. The project's goal is to develop a large, open, and reliable artificial intelligence language model for Germanic languages, including Icelandic.

The collaboration meeting covered a diverse range of topics related to the development of such a model, including data collection, quality assessment, and training methodologies. There was also an emphasis on deepening cooperation between partners and strengthening relationships.

Two of our experts from Miðeind's technical team, Garðar Ingvarsson Juto and Haukur Barri Símonarson, gave a joint presentation at the meeting where they demonstrated how conventional training of large language models systematically discriminates between languages. They proposed a new method to measure and correct this bias, which particularly affects highly inflected languages like Icelandic.

Þorvaldur Páll Helgason, Chief Technology Officer of Miðeind, also participated in panel discussions about challenges and opportunities in developing artificial intelligence for smaller languages. The topics of discussion included trust in artificial intelligence, methodologies for working with limited data, the societal impact of technological solutions not supporting users' native languages, and a general vision for the future in this field.