EU Allocates Funds for European AI

3 Min. Reading Time

The EU Commission is promoting the development of European AI with up to 54 million Euro. While this may seem like a drop in the ocean compared to the sums being invested internationally in AI, it is part of a larger EU-wide effort.

Teuken-7B, the AI language model released as open source at the end of November 2024, which was developed with significant contributions from the Fraunhofer Institutes IAIS and IIS and supports 24 European official languages, could be the starting point for a European artificial intelligence of its own.

A similar plan is now being pursued by another consortium of companies, research and educational institutions, and supercomputing centers. “We will build new next-generation language models from scratch and make them available to all citizens, businesses, and public administrations,” quotes Handelsblatt co-project leader Peter Sarlin.

Sarlin, with his Finnish company Silo AI, has already worked on open, multilingual AI models and sees the current project as a continuation of that work.

What are 54 million compared to 500 billion?

The large-scale initiative, which is running under the name “OpenEuroLLM,” is primarily intended to prevent Europe from falling further behind internationally in AI. However, the funding promised by the EU Commission of up to 54 million Euro over three years seems rather meager in comparison to the 500 billion dollars from the American Stargate project.

In Europe, however, the priority is to overcome the dependence on English as the most reliable input and output language. Developers, mostly based in the USA, such as OpenAI, Anthropic, and Google, are focused on English for purely economic reasons, as it is the most widely spoken language with over 1.5 billion people, including those who speak it as a second language.

In comparison, investing in Estonian, with around 1.2 million native and second-language speakers, hardly seems worthwhile. Even Turkish is more widely spoken in the EU than many of the 24 official languages.

Many Languages, Many Challenges

For Peter Sarlin, the “OpenEuroLLM” project is like a “moonshot.” The large language models based on it are to be trained directly in 35 languages, covering not only the languages of EU accession candidates but also some regional languages like Basque. The goal is for the European language models to perform just as well with Estonian and Lithuanian as they do with English, German, or French.

The major challenge, however, is not just that the multilingual European language models need to learn a vast amount of vocabulary and grammar rules. The bigger difficulty lies in the fact that fewer text sources and other resources are available for less commonly spoken and written languages.

The Handelsblatt points to other multilingual AI projects, such as the Viking models from Peter Sarlin’s company Silo AI. These can process not only English and programming languages but also Nordic languages like Finnish and Icelandic. The development team, now part of AMD, has already announced an AI model for all European languages.

Another project, called Eurolingua, aims to train AI models for all official European languages in collaboration with the Fraunhofer Institute for Intelligent Analysis and Information Systems (IAIS) and AI Sweden. Teuken-7B is also heading in this direction.

According to Sarlin, the “OpenEuroLLM” project surpasses all previous initiatives. Five companies and ten universities and research institutions from across Europe are working on the project, including some from Germany, such as the Fraunhofer Society, the Ellis Institute, the University of Tübingen, the Heidelberg AI startup Aleph Alpha, and the Bremen AI specialist Ellamind.

“By comparison, an investment in Estonian, with around 1.2 million native and second-language speakers, hardly seems worthwhile.”

Limited Resources? Deepseek Offers Hope

The supercomputing power is set to come from the EuroHPC joint venture in Barcelona and Jülich, North Rhine-Westphalia. However, it remains uncertain whether the EU Commission’s allocated funding of up to 54 million Euro will be sufficient. The development of the ChatGPT-based model GPT-4 is reported to have cost around 75 million Euro, while Anthropic CEO Dario Amodei mentioned that developing Claude Sonnet 3.5 also cost tens of millions of dollars.

The Chinese startup Deepseek offers hope, having reportedly spent only around 5.6 million dollars or 5.45 million Euro to develop a partially superior AI model. According to Jan Hajic from Charles University in Prague, the available computing capacity in Europe is sufficient to keep pace with the largest commercial language models.

“If it’s true that we can train AI with significantly less computing power, that’s good for Europe,” says the computational linguist. The consortium he is part of plans to release the results of their joint development, including code and accompanying research material, as open source in the future, protecting them from being monopolized by individual companies.

Frequently Asked Questions

Editor’s Picks

• Funding slows down the Mittelstand: Why doing it simply is faster
• Managed Services: Why the Mittelstand outsources IT instead of building it up
• NIS2 implementation: What the Mittelstand still needs to do

More from the MBF Media Network

cloudmagazin | SecurityToday | Digital Chiefs

Source title image: Unsplash / Mika Baumeister