Both the European Union and its individual member states hope to become more digitally autonomous. However, for AI implementations, they rely more often than not on non-European companies. A sovereign digital infrastructure should prevent this external dependence. Some nations have already taken some steps towards this, but when can we refer to a fully-fledged “Sovereign AI Infrastructure”?
The push for European digital independence has been on the rise for several years. Simply creating a sovereign cloud service is a difficult task. Microsoft and AWS, for example, attempt to provide this, but are of limited use to Europe. After all, they are non-European companies which may be forced by the US to hand over European data. As AI moves from promising innovation to actual implementation, Europe will have to avoid having this dependence on others in this area. It will have to protect the data that’s generated here as much as possible.
But what is needed to do this? Sovereign AI can be promoted in many ways, but a full-fledged infrastructure for it has quite a few requirements. Namely, it means control over training data, software, applications within sectors and the hardware needed to run it. This is important because only then can it be guaranteed that AI applications can be steered in the right direction with European and national legislation. As a prime example, the Netherlands has already explained how it hopes to achieve this in a general sense with GenAI.
A government-wide AI vision
Over a month ago, the outgoing Dutch cabinet unveiled a “government-wide vision for generative AI.” The emphasis is on a desire to invest in AI innovations so that these develop in a way desirable to the Netherlands. The opacity of current state-of-the-art implementations of GPT-4, for example, makes it impossible to guarantee that the data used and the resulting LLM behaviour is as desired. Granted, there are open-source options (such as Meta’s Llama 2) that offer finer controls, but complete autonomy is lacking. Time’s ticking, as local officials are already showing that it’s not an option to simply not use GenAI. In the absence of an AI solution to aid productivity, they are covertly turning to ChatGPT.
Motivated to solve this, last year several Dutch initiators already presented GPT-NL, a homegrown language model. A public LLM, trained and developed in Dutch, could be a solid basis for government-wide implementations of GenAI. A crucial aspect of this is that the training data can be fully monitored. Dutch-language sources should be at the heart of GPT-NL’s development, with clear copyright protection and possibly compensation for the use of news articles or archival material. GenAI shows time and again that the data used has a huge impact on the quality of the outputs, so homegrown training data is highly desirable. In the long run, this also helps safeguard languages that aren’t globally used, which other examples show already.
At the European level, there are already companies taking this into account. For example, France’s Mistral AI already supports French, English, German, Spanish and Italian. It ensures
Also read: LLM for Europe: Mistral AI puts Europe on the AI map
In a practical sense, an AI model trained on this data can be used extensively by government agencies and others. Examples include tools allowed for civil servants that can provide increased productivity. AI can help to search documents or make extensive records comprehensible in a short time. Such AI implementations will be familiar to many readers, but for now are not easy to reconcile with confidential documents. Those who do so now therefore have no idea whether the parameters and training data used fit the kind of AI models that meet national values. There are also no guarantees, for example, when using ChatGPT, that the data does not end up with OpenAI. In short: there’s no verifiability, and as a result no protection of sensitive data, meaning no sovereignty.
What exactly these national values are, of course, is not easy to answer. What we can say is that common legal definitions of security, justice and sustainability within a specific nation may differ greatly from the ideals of OpenAI, Microsoft or AWS. In fact, the possibility that new legislation may be needed does not matter. The point is that Europe and its member statesmust be able to take a monitoring role to test AI software, data and applications.
Infrastructure: also challenging, but a lesser problem
GenAI can quickly have steep system requirements, spearheaded by low latency. Energy consumption can be significant too, which adds to the difficulty in implementing AI everywhere. A ubiquitous approach will be needed, especially if Nvidia is to be believed. That company is the leading facilitator of the AI revolution by a vast margin, with its powerful GPUs for AI training, fine-tuning and inferencing.
Nvidia talks about future “AI Factories” that will be needed around the world. Examples include data centers specifically designed to accelerate AI, with a greater focus on AI accelerators and a modular design where hardware can be quickly replaced. Indeed, new generations of GPUs and other AI hardware are expected to make giant strides in efficiency and speed. A sovereign infrastructure must grow with this acceleration to remain relevant.
Nvidia CEO Jensen Huang is a great evangelist of this vision. He recently attended the World Governments Summit in Dubai, emphasizing that every country needs sovereign AI. That’d be favourable to Nvidia’s checkbook, but it’s certainly a defensible position to take. Either way, Huang is busy convincing countries, including through visits to leaders in Canada, France, India, Japan, Malaysia, Singapore and Vietnam. “It’s not that costly, it is also not that hard,” Huang believes. “The first thing that I would do, of course, is I would codify the language, the data of your culture into your own large language model.”
So even the top executive of the premier AI hardware company advocates an approach that instead puts software at the center. Building a sovereign AI infrastructure ensures that one’s own culture, expertise and values are central. Some nations are at the forefront of this, meaning they may be in a position to fully utilize AI the soonest.
2 days ago
