At Rise of AI 2024 in Berlin, a panel moderated by Larissa Holzki (Handelsblatt) with Vanessa Cann (KI Bundesverband), Peter Sarlin (CEO and co-founder, Silo AI) and Nicolas Flores-Herr (Fraunhofer IAIS) discusses whether Europe can build competitive large language models against OpenAI and Microsoft. The panelists agree OpenAI dominates industry and political mindshare with Microsoft's backing, but argue the best model is not necessarily the biggest, and that Europe should focus on smaller, specialized, multilingual and industrial models rather than a general-purpose chatbot. Sarlin presents Silo AI's newly released Viking model for Nordic languages, arguing GPT-4 speaks Finnish or Swedish like a translated American model. Flores-Herr describes Fraunhofer's openGPT-X project and a new EuroHPC-funded initiative training models in 45 languages. Cann warns Europe risks becoming merely a deployment market for US models unless companies rethink their business models, not just bolt AI onto existing products, and shares experience collaborating with a Shanghai AI team. The panel closes on cautious hope, tempered by fear that Microsoft becomes Europe's default answer.
Key takeaways
- Peter Sarlin argues GPT-4 effectively speaks Finnish and Swedish 'like an American' because its reasoning is built in English and translated rather than native to the language.
- Vanessa Cann says GDPR has discouraged European companies from using industrial machine data even though it is usually not personal data.
- Fraunhofer's new EuroHPC-funded project, running on the MareNostrum 5 supercomputer, trains models in 45 languages including Bavarian, ending in May 2025.
- Silo AI released its Viking model for Nordic languages at the panel, following an earlier Finnish-English-and-code model, without sacrificing English performance.
- Vanessa Cann warns that Germany's roughly 400-member AI association network and 600 AI startups risk Europe becoming a deployment market for foreign models, with Microsoft as the default answer, rather than a genuine 'deployment champion'.
Detailed notes
Reacting to OpenAI's latest model
Asked about a new, more conversational GPT-4 release, Cann says OpenAI and Microsoft together dominate the conversation with corporates and politicians she meets, and that Europe is eager for its own players even as OpenAI also excels at marketing. Flores-Herr says ChatGPT proved what was possible and made AI a household topic, prompting even his parents to ask what Germany and Europe are doing in response.
Is AI progress accelerating or gradual?
Sarlin argues that scientific progress in AI has been gradual over the past 20 to 40 years, while what is accelerating is the engineering and funding applied to that research; he says current transformer-based, token-prediction models are not a path to AGI, whatever people say about it. Flores-Herr recalls that when Fraunhofer proposed the openGPT-X project to train German foundation models from scratch, the team never anticipated a moment like ChatGPT's arrival, and expects another such leap could still come.
Why smaller, specialized European models
Cann argues the best LLM is not necessarily the biggest one: her own work on industrial models found that large general-purpose models often do not fit industry needs, which instead require multimodal data such as tabular data and knowledge graphs, an area where smaller, focused European models can compete. She also says GDPR has wrongly discouraged use of machine data that is not personal data, and that boards now take AI seriously but often lack the change management to overcome employee resistance. Sarlin says GPT-4 can answer in Finnish or Swedish, but effectively speaks it like an American because its reasoning is built in English and translated, not native to how Finns communicate; Silo AI's newly released Viking model, following an earlier Finnish-English-code model, targets Nordic languages without sacrificing English performance, with a pan-European model in training, because value should stay with companies that control performant regional models rather than flow to closed foreign providers.
Building sovereign models: openGPT-X and beyond
Flores-Herr describes a new EuroHPC-funded project, run on the MareNostrum 5 supercomputer in Barcelona, training both large and small models covering 45 languages including Bavarian, designed so countries like Malta or Portugal can continue training the models on their own data; the project runs until May 2025. He argues Europe cannot build an economy on an API hosted elsewhere and must own the full value chain, comparing a country that only supplies the car radio while the car itself comes from abroad. Sarlin describes Silo AI's model-as-service platform, built on open-source models, giving companies ownership and control of their own models with production tooling, guardrails and retrieval-augmented generation.
AGI, priorities, and the ingredients for competitiveness
Asked whether reaching AGI just needs more compute or new architectures, Flores-Herr says both are needed, and urges Europe to act boldly now on multimodal models, hallucination detection and tool learning, which he expects to show returns within two years. Sarlin frames AI competitiveness around three ingredients: data, compute and talent, where he says Europe is strong, but argues the bigger gap is in scalable software products and distribution that can turn models into real value, warning that good European models will otherwise also benefit better-positioned North American software companies.
Europe's position, deployment risk, and China
Cann outlines a European AI landscape with high openness toward AI, a manufacturing-heavy industry, Germany's AI association network of about 400 companies and 600 AI startups, and a European AI Forum of national partners, but warns most companies are only adding AI to existing processes rather than rethinking their business models, unlike an example she gives of Amazon considering becoming a predictive shipping rather than shopping company. She fears Europe is gambling its future on deployment of foreign models rather than becoming a genuine deployment champion, and voices worry that Microsoft has become the default answer many companies reach for. Asked about her earlier startup's collaboration with a team in Shanghai, she says she still sees potential in EU-China cooperation despite IP and data security risks that required careful safeguards, though her own venture ultimately failed due to internal team commitment issues rather than the China partnership itself.
Quotes
GPT-4 is speaking Swedish and Finnish like an American.
Peter Sarlin
We cannot make an economy that is based on an API that is somewhere in other countries.
Nicolas Flores-Herr
I really do hope that the answer is not Microsoft, and that's my biggest anxiety right now.
Vanessa Cann
Questions answered in this talk
Why does Peter Sarlin say Europe needs its own language models instead of just using GPT-4?
According to Peter Sarlin at Rise of AI 2024, GPT-4's reasoning is built in English and translated into other languages, so it effectively speaks Finnish or Swedish 'like an American' rather than reflecting how those languages and cultures actually communicate, which is why Silo AI built the Viking model for Nordic languages.
What project is Fraunhofer IAIS working on to build European foundation models?
According to Nicolas Flores-Herr at Rise of AI 2024, after Fraunhofer's openGPT-X project trained German foundation models from scratch, a new EuroHPC-funded project on the MareNostrum 5 supercomputer trains large and small models covering 45 languages, including Bavarian, ending in May 2025.
What is Europe's biggest risk in the AI race according to Vanessa Cann?
According to Vanessa Cann at Rise of AI 2024, Europe's biggest risk is gambling on becoming merely a deployment market for foreign models rather than a true deployment champion, since most companies are only adding AI to existing processes instead of rethinking their business models, with Microsoft often becoming the default answer.