A Dutch counterpart to ChatGPT: good idea?

TNO is receiving a 13.5 million euro government grant to develop GPT-NL, the first open AI model from Dutch soil. The aim: to be less dependent on the big US tech companies. The development comes after the rapid rise of AI applications such as OpenAI's ChatGPT.

Founded by Selmar Smit (TNO), GPT-NL aims for more transparency: both the source data and the algorithm will be public. The language model could also be used by the Dutch government, for example for translating complex tax documents into more accessible language.

However, there are questions about the adequacy of funding, as large AI companies invest many billions in their own AI technology. TNO is working with the Netherlands Forensic Institute and SURF on the project, whose first applications are expected by the end of next year.

Razo van Berkel, AI engineer at Centric, is in favour of developing a Dutch AI model. "About time something like this came about. It helps to attract AI talent. Besides, it's good to show the world that we believe it's important too. In Europe, of course, the call to be less dependent on US tech companies has been heard for some time."

Why would we want to be less dependent on the Americans?

"Well, in the Netherlands and Europe, we think very differently about privacy. This was evident recently, when Meta launched its new chat app Threads, which can’t be used over here for that very reason. One of the objections to OpenAI is that we don't know what data sources they use to train their system. Even in America itself, there are now lawsuits about the origin of that data. In Europe, of course, we look at that even more strictly."

What does it take to build GPT-NL?

"Mainly a lot of technical research, for example into the architecture for this kind of language model. But data also needs to be collected and processed, in the right way and with the necessary checks. In that too, TNO's approach differs greatly from OpenAI."

Microsoft and Google are investing billions in AI. So, how far will TNO get with 13.5 million?

"Further than you might think. OpenAI is a pioneer, they are the first to do this. The market leader often has to invent everything themselves and make large initial investments. Those who start with AI now don't start empty-handed. A lot of knowledge and blueprints for the architecture are already available. And server farms can be rented. So, other parties can build on what is already there. That makes things easier."

Europe is coming up with AI legislation. Why a GPT-NL and not, say, a GPT-EU?

"That is possible in the long run and maybe even desirable. The issue is, dozens of languages are spoken in Europe. Developing a model that can handle all languages well is a huge task. AI that is usable within the Netherlands only needs to know things about the Netherlands. An EU-wide language model is much more complex. Although European cooperation and funding would open up a lot of possibilities."

The new AI model is not here yet. What can enterprises and developers do to still work securely with ChatGPT?

"OpenAI offers paid enterprise environments, closed containers that should guarantee that nothing happens to your data and that others cannot access it. You then run your own instance of GPT in there. That way, you should be able to use their AI model securely for business. Of course, this requires you to trust that promise."