Published on 6 November, the dataset is accompanied by Itiner-e an interactive online platform that increases the known length of the Roman Empires road system by more than 100,000 kilometres. Co-author and PI of SRL, Tom Brughmans, describes it as a Google Maps for Roman roads, and it is already receiving massive attention. So much so that the servers are struggling to keep up.

Following Thursdays press release, interest has been overwhelming, with the map being viewed around 100,000 times in just the first 24 hours, says CHC developer Peter Vahlstrup, who built the platform.

The groundbreaking work highlights the power of combining social and computational perspectives an approach at the heart of the CORE collective, where interdisciplinary collaboration drives innovation and best practice.

The publication and the new results are the outcome of a collaboration between an international research team from Aarhus University and Universitat Autònoma de Barcelona.

Find the paper here:
de Soto, P., Paout, A., Brughmans, T. et al. Itiner-e: A high-resolution dataset of roads of the Roman Empire. Sci Data 12, 1731 (2025). https://doi.org/10.1038/s41597-025-06140-z

Their talk introduced the concept of culture neurons  sparse neural populations whose activity both predicts and causally influences culturally conditioned behavior, while remaining disentangled from language-specific features. Using MUREL, an 85.2M-token corpus spanning six cultures, they combined localisation analysis with causal interventionsto identify and manipulate these neurons.

Their findings suggest that cultural tendencies in LLMs tend to cluster in upper network layers and can be selectively adjusted with minimal impact on overall model performance. This opens new possibilities for understanding and guiding cultural dynamics in language technologies.

Many thanks to Lukas and Danial for an insightful and boundary-crossing session and for advancing the conversation on culture, computation, and critical AI research.

• Creating and managing local Python packages
• Publishing to GitHub and PyPI
• Writing unit tests and setting up continuous integration
• Collaborating via pull requests and issue tracking
• Generating documentation
• Ensuring clean code through formatting and linting

The session served as a practical introduction to modern development workflows, with a follow-up workshop planned to build on the material covered.

Many thanks to Kenneth for a clear and instructive session and to everyone who participated, both online and in person. We look forward to continuing the conversation around open, collaborative, and sustainable research software.

In the first workshop, participants delved into the unexpected Danish corpus embedded in Googles scans of U.S. academic libraries. This discovery raised important questions about metadata, model training, and the future of Nordic-language texts in the age of national large language models (LLMs).

The second workshop bridged art history and machine learning, as participants experimented with diffusion models and LoRA fine-tuning to engage with Danish Golden Age painting in new computational ways. The session also explored multimodal LLMs and how cultural heritage institutions might critically use image-captioning models to support both accessibility and scholarly description.

A warm thank you to all who joined and especially to Peter Leonard for sparking new conversations at the intersection of culture, computation, and critical reflection.