Германии и Европе необходимо разрабатывать собственные передовые модели искусственного интеллекта

Viewpoint Why we in Germany and Europe need to build our own open frontier models The world’s leading AI models – also known as frontier models – are currently being developed in the US and China. Christian Cyron and Jenia Jitsev explain why this situation entails significant risks and missed opportunities. To what extent should Europe be dependent on others when it comes to a key technology like AI? This question is particularly relevant with regard to so-called frontier foundation models – models that are universally applicable to a wide variety of tasks and could fundamentally transform research and industry. Only those who can develop them independently can ensure their technological sovereignty – developing specialized models that can handle only narrowly defined tasks is not enough. Some consider investing in proprietary frontier AI models to be neither necessary nor worthwhile given the availability of “open-weight” models. With these models, the trained model parameters (“weights”) are publicly available and can be downloaded, run on one’s own servers, and customized. Using such models from the US and China is legitimate and important. Nevertheless, there are at least four good reasons for Europe to pursue research on and develop its own frontier models. Ensuring access to the most advanced capabilities and their further development Whether the weights of an AI model are published is decided anew by the publishers with each release. They themselves or political decision-makers can simply choose not to. We have recently experienced both scenarios, albeit only temporarily. Furthermore, there is as yet no evidence that an older, freely available generation of models can be adapted to match the performance of the next generation. Consequently, if Europe were ever denied access to the next generation of AI models, it would be cut off from the most advanced problem-solving capabilities in all areas of research, industry, and society. This would not only create a nearly impossible to overcome disadvantage, but also – given the rapid pace of technological development – pose an existential risk. Understanding risks: How frontier models behave largely depends on their training data and methods. Deducing this from the final parameters (weights) alone is extremely difficult, if possible at all. To assess risks, we must therefore have a complete understanding of how the models were developed – which is impossible with today’s leading models. Frontier models will be deeply embedded in the economy and public administration by the 2030s. As such, thoroughly assessing them represents a fundamental duty of care. Fulfilling this duty requires a detailed understanding that can only be gained by those who have built at least some frontier models themselves. Seizing opportunities: Those who build frontier models are shaping the next generation of technology and helping to determine what it will look like and when it will arrive. By adopting these models early on, developers can set the agenda and withhold them for as long as they wish. In contrast, those who rely exclusively on models created by others must make do with whatever is made available to them. Many opportunities presented by a new technology first become apparent at its forefront. Building its own frontier models would therefore enable Europe not only to take advantage of the AI era but also to actively shape it. Understanding the future How a robust AI foundation model is created is itself the subject of basic research. But such research is only possible if there is an open pipeline for building such models. Currently, most experts outside the frontier labs cannot fully understand exactly what is happening. This leads to a troubling situation: frontier models are arguably among the greatest breakthroughs of the 21st century. As such, understanding their implications should be a top priority for public research. However, doing so remains infeasible without an open training pipeline that grants a broad community access to all relevant information. For at least these four reasons, Germany and Europe need to be capable of building their own frontier models. Anything else would entail significant risks and missed opportunities. Above all, however, this is the only way we can understand and help shape how AI transforms our society in the years and decades to come.