Smiling AI Hypercomputers

The event was held in the large auditorium of the main LMU building, which was designed in the Art Nouveau style. Slides about next generation AI were displayed on a wall that also prominently featured naked Greek gods. Despite the venue's suboptimal acoustics, a similar contrast could be heard between the invited speakers, who came from very diverse backgrounds and spoke on a broad range of topics. While some presentations were primarily promotional in nature, highlighting company products with terms such as "AI Hypercomputer," others offered more detailed insights into the methodologies behind these technologies.

Some presentations were not relevant to civil society, such as a presentation on revenue management at SIXT about what to show customers during the selection process when renting a car. The diversity of this event was demonstrated when such talks were followed by a talk modeling extreme weather events in the wake of climate change. This was accomplished with the SMILE (Single-Model Initial-condition Large Ensemble) approach, which combines multiple simulations with different initial conditions to account for the chaotic nature of climate systems. A deep learning model was trained using high-resolution climate data, allowing a better understanding of the patterns of extreme weather events that are now occurring more frequently. Although we were reminded of the urgency of tackling climate change, the only mention of deep learning's high energy consumption was in this presentation.

Why sharing is caring, but not necessarily daring

In a session that also focused on solutions to climate change, we learned about Federated Machine Learning (FML) and its applications in renewable energy production and medical research. FML is a decentralized approach to training machine learning models, where data remains local, and only model updates—such as weights—are shared with a central global model. This ensures data privacy and security while enabling collaborative learning across multiple systems.

For example, local instances could include wind turbines, where FML can optimize parameters like orientation to maximize energy output, or hospitals, where patient data can be analyzed to improve diagnostics and treatments without exposing sensitive information. It was suggested that modern hospitals should establish dedicated data science departments to fully leverage the potential of such technologies.

Initiatives like the German Portal for Medical Research Data are improving the availability, resolution, and quality of health data. This can drive progress in data-intensive approaches, such as speeding up the diagnosis of rare diseases, advancing precision medicine, and enabling genome-wide association studies (GWAS). GWAS are observational studies designed to identify associations between genetic variants and diseases, which provides opportunities to develop targeted treatments.

The potential of data in medicine shows that sharing data is important. In an interesting talk about data protection laws in Germany and the EU, data was compared to property, which also comes with obligations to society according to the German constitution. Data protection laws were likened to traffic regulations: just as a digitalized society cannot function without data sharing, it also cannot function without proper regulations to ensure privacy and security.

A talk about advances in Open Source Intelligence (OSINT) provided a different perspective on publicly available data, mainly on social media. The speaker highlighted how generative AI enables the transformation of unstructured data—such as text, images, and videos—into actionable insights that can support law enforcement efforts. However, I would add that these same capabilities can also be exploited by malicious actors. On a more positive note, a different speaker recounted the story of the Panama Papers, a leaked data source containing large amounts of text that required hundreds of journalists to analyze in 2016. Today, LLMs could considerably speed that up this analysis.

In a talk by a conflict researcher, it was demonstrated that leveraging publicly available data does not always require the use of generative AI. Since commercial satellite imagery is often too expensive for research and civil society organizations, the researchers in this project turned to publicly available satellite imagery provided by the European Space Agency. Using this data, they developed a model capable of reliably detecting explosion-related building destruction. The model's effectiveness was validated using the 2020 explosion in the port of Beirut as a case study. This tool can also be applied to independently assess the impacts of conflicts, such as the destruction caused by Russia's invasion of Ukraine.

How to use generative AI to generate value

The buzzword of the event was undoubtedly AI, primarily associated with generative models. Image generation was highlighted as a versatile tool, applied in both industrial and fashion contexts. In the fashion industry, for example, image generation can be guided by clothing features, such as color, that have been identified as successful. A significant focus was also placed on language models, with multiple mentions of custom chatbots based on Retrieval-Augmented Generation (RAG), designed to assist employees in various tasks.

Finetuning open language models, such as the LLaMA models by Meta, has become a common practice, allowing these models to be tailored for specific tasks or domains, thereby improving their performance and relevance in specialized applications. For certain specialized use cases, smaller language models can also be employed effectively. Another key topic was quantization, a technique that reduces the computational and memory demands of language models, making them more efficient and accessible without a significant loss in accuracy. Additionally, data curation, particularly through the generation of synthetic data, emerged as a promising approach to enhance model performance and address challenges posed by limited or incomplete datasets. However, despite the ability of models to learn even from limited data, one speaker advocated for maintaining traditional machine learning practices, such as train/test splits.

Language models are increasingly being integrated into agentic systems, which involve multiple specialized models working collaboratively or equipping models with external tools that have compatible interfaces. The concept of "Large Action Models" was introduced, combining symbolic reasoning with language models to improve their ability to execute complex, multi-step tasks.

In conclusion, the German Data Science Days 2025 offered an interesting glimpse into the diverse applications and implications of AI and data science across industries and research. While the event was largely shaped by the ongoing hype around AI, it also provided valuable lessons on how companies derive real value from AI applications and tackle the challenges these technologies present. It is essential however to translate business logic into solutions that actually drive positive change. Furthermore, as somewhat freed from business logic, civil society has a responsibility to reflect on potential negative impacts of new technologies, especially if other societal actors do not live up to their responsibilities. To end on a hopeful note, the conference reaffirmed that data is one powerful tool to enable society to tackle many problems.