5th AI Winter School of the Austrian Academy of Sciences (ÖAW) 2026
This resource brings together three sessions from the 5th AI Winter School of the Austrian Academy of Sciences (ÖAW) and the Machine Learning Thematic Platform MLA²S, held from 16–20 February 2026 in Vienna. The AI Winter School was designed as an introductory training event for participants who did not yet have in-depth knowledge of machine learning or artificial intelligence, with a particular focus on supporting scholars from the humanities. It aimed to provide foundational knowledge, practical orientation, and critical perspectives on AI and ML, while creating a space for exchange between researchers, students, and staff from different disciplinary backgrounds.
The Winter School was primarily open to people affiliated with, or working closely with, institutes of the Austrian Academy of Sciences, including MSc and PhD students, PostDocs, faculty, and staff. External participants were also able to attend where places were available. As part of the wider programme, the school introduced participants to key concepts, methods, and debates in AI, ranging from hands-on approaches to AI-assisted research and software development to broader questions of open language models, bias, and responsible use of machine learning.
The sessions included in this resource offer selected insights into the themes addressed during the Winter School. They are intended as an accessible entry point for learners, researchers, and practitioners who want to better understand how AI tools can be used, developed, and critically assessed in academic and applied contexts.
The first session, ‘Vibe Coding & Promptotyping’ by Christopher Pollin, consists of two videos and introduces current approaches to AI-assisted software development. It explores methods such as Vibe Coding, Context Engineering, and Promptotyping, showing how researchers without advanced programming skills can use Large Language Models to develop data-driven tools, interfaces, models, and workflows. The session highlights the potential of generative AI as a collaborative tool in the research process.
The second session, ‘Open EuroLLM’ by Patrick Rarivoson, focuses on developments in the field of open European Large Language Models. It addresses the significance of open language technologies for research, innovation, transparency, and digital sovereignty, and situates current AI developments within broader European technological and societal contexts.
The third session, ‘Bias in ML’ by Philip Winter, provides an accessible introduction to bias in machine learning. Beginning with core concepts of model training and learning paradigms, the session examines how social and technical biases can arise in modern deep learning applications. It also discusses potential real-world consequences and outlines practical strategies for identifying and mitigating bias.
Together, these recordings document selected contributions from the AI Winter School 2026 and provide a compact resource for anyone interested in foundational AI knowledge, AI-assisted research practices, open language models, and the responsible use of machine learning.
Target Audience
This resource is aimed at researchers, students, educators, and professionals in the digital humanities, humanities, social sciences, and related fields who are interested in understanding and applying artificial intelligence in research contexts. It is particularly relevant for learners who want to explore the practical use of Large Language Models, AI-assisted software development, open European AI infrastructures, and responsible machine learning. No advanced technical background is required, although a basic interest in digital methods, data-driven research, or AI applications will be helpful.
Learning Outcomes
After engaging with the videos, learners will be able to:
- Explain and critically evaluate the functioning, capabilities, and limitations of Large Language Models and machine learning systems.
- Apply AI-assisted methods to research contexts.
- Assess the ethical, social, legal, political, and sustainability implications of AI technologies.
1.Vibe Coding & 'Promptotyping' - Part 1
This session by Christopher Pollin introduces vibe coding and Promptotyping as ways of using large language models to support research workflows, especially in the Digital Humanities and cultural heritage. The speaker shows how AI can help generate code, build small research tools, process data, and create visualisations.
At the same time, the talk emphasises that AI does not replace research expertise or professional software development. Its outputs must be checked, validated, and used critically. The concept of asymmetric amplification highlights that AI can strongly enhance the work of those with the skills, resources, and infrastructure to use it effectively.
Learning Outcomes
After watching this video, learners will be able to:
-
Explain vibe coding and Promptotyping as AI-supported approaches to creating code, prototypes, and research tools.
-
Identify key risks of AI-assisted coding.
-
Reflect on how AI can amplify research practices unevenly.
-
2.Vibe Coding & 'Promptotyping' - Part 2
This video explains the basic principles behind large language models and connects them to practical prompting and context engineering. The speaker introduces key concepts such as next-token prediction, tokenisation, embeddings, context windows, pre-training, post-training, hallucination, sycophancy, and the limits of current models. A central point is that language models do not “understand” text in a human sense, but operate through statistical patterns, compressed knowledge, and semantic relationships learned from large training datasets.
The talk also shows why data formats, prompt structure, and context management matter in practice. Lightweight and structured formats such as CSV and Markdown can be more efficient than complex formats like Excel or Word, because they use fewer tokens and provide clearer structure. The video ends by discussing broader limitations and future directions, including world models, synthetic environments, AI agents, and the question of whether AI systems might eventually support or automate more complex forms of research.
Learning Outcomes
After watching this video, learners will be able to:
- Explain core concepts of large language models.
- Apply basic context engineering principles.
- Critically assess the limitations of language models.
3.Open EuroLLM
Artificial intelligence has become a key factor in scientific innovation, industrial competitiveness, and public-sector transformation. At the same time, it raises urgent questions around ethics, sustainability, cultural bias, transparency, and technological sovereignty.
In this talk, Patrick Rarivoson from the French AI company LightOn discusses Europe’s efforts to develop trustworthy and sovereign AI infrastructures. Drawing on LightOn’s experience in photonic hardware, large language models, OCR, retrieval models, and industrial AI applications, he introduces major European initiatives such as OpenEuroLLM, LLMs4EU, and AI-on-Demand / DeployAI.
A central focus is OpenEuroLLM, which aims to create open, AI Act-compliant language models for European languages and cultures. Rather than simply supporting multilingual vocabulary and grammar, the project seeks to address deeper cultural representation and bias, especially for low-resource languages. Patrick Rarivoson highlights how open collaboration, transparent data practices, European supercomputing infrastructure, and responsible model evaluation can help build AI systems that align with European values while remaining competitive on a global scale.
Learning Outcomes
After watching this video, learners will be able to:
- Critically assess the importance of European AI sovereignty.
- Analyse challenges of bias, multilingualism, cultural representation, and sustainability in Large Language Models.
- Describe and evaluate approaches to trustworthy AI applications.
4.Bias in ML
This talk offers a rigorous yet accessible introduction to the topic of bias in machine learning systems. It begins with a concise overview of the fundamentals of machine learning, covering key concepts such as model training, learning paradigms, and typical areas of application in contemporary ML approaches. Building on this foundation, the talk examines different forms of social and technical bias that may arise, particularly in modern deep learning applications.
Special attention is given to the real-world implications of such biases. In addition to societal risks and potential forms of discrimination, the talk also addresses technical consequences, including limited generalizability, inaccurate predictions, and reduced model reliability. It concludes by presenting practical strategies for identifying, assessing, and mitigating bias, thereby supporting a more responsible and reflective use of machine learning technologies.
Learning Outcomes
After watching this video, learners will be able to:
- Explain basic principles of modern deep learning and machine learning bias.
- Identify and distinguish common sources and types of bias in machine learning.
- Critically evaluate the ethical, social, and legal consequences of biased AI systems.