“We are at the very forefront of research”

Professor Kersting, the EU is planning to invest around 200 billion euros in AI, according to an announcement made after the AI summit in Paris. Will Germany and Europe then be able to keep up with the big players - the US and China?

The figures are indeed pretty impressive. There does appear to be greater willingness now in Europe to take the risk and invest massively in a technology despite nobody knowing exactly how it will evolve in future. I regard this as a positive step, though I do find Macron’s “Plug, Baby, Plug”, which essentially is just an advert for nuclear-powered AI factories, somewhat inappropriate. After all, the large-scale AI models still consume far too much electricity. We urgently need to think about how to build such models in a more efficient and resource-friendly manner.

Does this mean that foundation models, i.e. AI systems that can do all kinds of things and are trained on huge datasets, are not the right way forward? 

No, that’s not what I am saying. We have to keep the discussion open as we still know too little about which models of which size will best help us tackle which problems. Demis Hassabis, the CEO of Google DeepMind and a Nobel Prize laureate, recently announced a plan to create a digital simulation of a human cell. It would of course be ridiculous if he were not able to use a large-scale foundation model to do so. However, we should question whether building ever bigger models is the best approach. Ultimately, it is the concrete benefits that count. And we are seeing that more and more companies want highly specific solutions for their problems: How can I automate responses to customer enquiries? How can I verify forms? How can I reformat them so that they can be processed by a particular software? 

Arthur Mensch from the French AI start-up Mistral says that European AI’s strength lies precisely in developing compact models tailored to the local needs of businesses. Is Europe perhaps not in fact lagging as far behind as people keep claiming?

Not at all. I’d even go so far as to say that we are not lagging behind at all when it is only a question of scale, i.e. size. This is something we are now seeing with the plan to step up military spending: Europe is financially stronger than it is frequently claimed. What has been lacking to date is the will. And we can see from developments in China that a resolute will to embrace risk has enabled the country to reach a similarly high level as the US in a relatively short period of time. What is now becoming apparent, however, is something that Arthur Mensch has also alluded to, namely that the quality of outcomes is not necessarily dependent on the size of the models alone. It is perfectly possible to accomplish the same with much smaller systems. This involves dividing the task up into different parts. For example, calculating the sum of two plus two requires far less capacity than finding a solution to the climate crisis.

Isn’t that precisely the idea behind the Chinese AI model DeepSeek? 

Correct, though I’d be cautious about automatically viewing this as a blueprint for the success of smaller models made in Europe. DeepSeek is run by a company with around 200 employees that is financed by a hedge fund. Even small models need to be trained, and this is where we in Europe still lack any really efficient infrastructure. If the processing power of US systems is compared with that of supercomputers such as JUPITER in Jülich, LUMI in Finland or LEONARDO in Italy, the United States is still ahead of us.

And yet you still claim that Europe is keeping up.

Yes, because we have certain things that others don’t have. Such as valuable data from industry, for example. We are still at the very forefront of research and train top-notch people. And we are also ahead of the US and China in another respect: we are currently learning better and more quickly how to translate AI systems into concrete applications. There are various exciting examples of this in Germany alone.

Which do you have in mind?

Take Celonis, a TU Munich spin-off that is currently estimated to be worth over ten billion euros. Celonis helps firms analyse and optimise their business processes. Another example is Black Forest Labs - an AI start-up that leads the world in AI image generation. Munich-based Helsing is one of Europe’s fastest-growing defence technology companies and currently valued at five billion euros. And the Heidelberg start-up Aleph Alpha is successfully developing AI applications for organisations, companies and official authorities.

Yet even Aleph Alpha has not managed to develop a foundation model that can compete with industry leaders such as OpenAI, Anthropic, Google or Meta.

This is because the investment volumes are simply on a different scale: we see sums of around half a billion euros here, whereas major players in the US can access multiple tens of billions in investment. This doesn’t necessarily mean that the strategy of focusing on smaller and more specific models is wrong. Ultimately, it’s a question of monetising such models - turning them into concrete applications, in other words. And this is a path that I believe we are following more consistently in Germany and Europe than elsewhere.

Do our high regulatory standards also pose an obstacle to Europe? Is Europe making life difficult for itself with its AI Act?

That’s not how I see it. It would be negligent not to sensibly regulate a technology like AI. In the medium term the US will also realise this, though just now it boasts about giving free rein to development. And yet it is a country with warning signs all over the place, telling people what they shouldn’t do and where danger lurks. This is intended to reduce the risk of liability claims. Yet this is not supposed to apply to AI, despite all the risks that the technology poses? That makes little sense. We should not allow ourselves to be intimated by the US approach. Developments in AI are simply too dynamic and it’s by no means certain that foundation models subject to little or no regulation would ultimately result in the best outcomes. Overregulation is not the way forward either, however.

Could regulated AI systems made in Europe even turn out to be a model for success?

Yes, absolutely. I firmly believe that the demand for well regulated and therefore reliable AI systems will increase in future. Especially if AI in other countries is elevated to the status of a kind of religion in which only a few people end up deciding what others should or shouldn’t believe, we should push back against such moves in Europe. AI is a cultural asset that everyone should be able to access. We don’t need an AI religion, we need AI enlightenment.