In an article by Logi Einarsson, Minister of Culture, Innovation and Universities, here on Vísir on 25 March, it was stated that discussions are underway with a leading international company on the development of an Icelandic artificial intelligence language model. The collaboration aims to specially train a large language model that would be hosted in Iceland. Particular attention would be paid to the Icelandic language and culture during its training. The plans have been associated with “Sovereign AI” and have parallels in other countries that are concerned about their security and independence from foreign tech giants, particularly American ones.

We, the undersigned, who have worked for years on the development of language technology and artificial intelligence for Icelandic at the company Miðeind, welcome these plans. We believe they are in the interest of national security and can also, if successful, become a catalyst for domestic knowledge and skills in this important and rapidly growing field. We also believe there are considerable opportunities to improve services and increase efficiency, both in the public and private sectors, if a domestic AI that is proficient in the Icelandic language and knowledgeable about Icelandic culture and history is available on favourable terms.

We point out that a better consensus could be reached on the contribution of texts and other material for the training of such a model on domestic terms, rather than for foreign technology companies. It should also be remembered that it is not a given that foreign models will improve in Icelandic. There are examples of well-known models repeatedly regressing between versions, according to Miðeind’s benchmark tests.

Finally, it should be mentioned that the use of closed foreign AI models (such as ChatGPT, Claude, Gemini and Grok) is not free. A considerable “AI tax” is already leaving the country, and it is foreseeable that this will increase rapidly in the coming years with the increased use of AI tools.

With “Sovereign AI”, the aim is not to create a model that would be competitive in all areas with leading foreign models; such a goal is unrealistic. It can still be expected that programmers will use, for example, Claude Opus or similar models to help them build software, and that the general public will have the ChatGPT app on their phones. But it is a realistic goal that an Icelandic model, hosted in a secure domestic environment, could become the best available model for understanding and generating Icelandic text, and for having solid knowledge of Icelandic culture, laws, values and history. These features are sufficient to make the model the best available option for a wide range of applications, for example in public administration and in the health, education and judicial systems, not to mention tasks in the field of security and defence.

Some applications of language technology and artificial intelligence are and will be of a sensitive nature. They involve working with confidential information such as medical records and transcripts of police interviews. It is therefore important that the handling and processing of data is fully in accordance with European regulations on data protection and data security – and meets our own requirements that sensitive data be stored and processed within the country and not sent via submarine cables for processing in various countries.

In our opinion, the most successful way to build a sovereign AI is to find a supplier and partner for the project that meets, among others, the following conditions:

• The supplier already has a large, highly competitive and sufficiently powerful large language model, preferably under open-source software licences, which would be used as the basis for an Icelandic solution.

• The base model is trained in accordance with European requirements and reflects European values.

• The supplier is willing to work with domestic parties on the continued pre-training of the model on Icelandic data. It should be borne in mind that training large language models with hundreds of billions of parameters is a combination of science and art and is by no means an easy task. Teams with experience in this are not a dime a dozen globally (and are very busy, where they can be found).

• The supplier is willing to train and assist domestic teams so that they can subsequently manage continued development, specialisation and maintenance themselves.

• The supplier can provide the hardware for pre-training, as training a large model is extremely computationally intensive, and access to significant computing power can be difficult to obtain. It is not necessary for such training to take place domestically.

• It is possible to run the model domestically in as many instances as deemed necessary, including in specialised versions, such as within the judicial system or in the field of security and defence.

. Please note that significantly less hardware is required to fine-tune or run a fully trained model in use than to pre-train it.

• It is possible, and planned for from the outset, to update fine-tuned and specially trained models when new and improved base models are released by the supplier, without necessarily having to start the fine-tuning from scratch.

• Arrangements are made so that even if the collaboration with the supplier were to be terminated for any reason, it would be possible to continue with the model independently, or to switch the base model with minimal effort.

• It is an advantage if the supplier, in addition to the model itself, can provide various types of supporting software, such as a training framework, a chat environment, an API server with associated security and access control, etc. Also, if the supplier can provide multimodal models if and when needed, such as for speech recognition and synthesis, optical character recognition, and image processing.

• It is important that the supplier is not subject to the jurisdiction of states that are feared to be able – covertly or overtly – to restrict or shut down the processing of important Icelandic AI tasks by decree if international crises were to arise.

If these conditions are met, which we believe is entirely possible, there is every chance that a project on Icelandic sovereign AI will deliver the intended results: improved services, efficiency gains, support for the Icelandic language and culture, the building of knowledge in a key area with derivative innovation effects, and increased security in turbulent times.

The authors are the CEO and Chairman of the Board of Miðeind ehf.