What Is AI? What Are LLMS?
Prof. Klauck began by explaining the basic classification:  Artificial intelligence (AI) is a branch of computer science that deals with the development of systems capable of solving tasks that typically require human intelligence based on data. A special area of this is generative AI – systems that can generate new content such as texts or images based on large amounts of data. LLMs, such as GPT (Generative Pre-trained Transformer), are language models that specialize in precisely this ability.

The presentation provided an insight into the architecture of such models: Central building blocks are so-called embeddings, which convert language into a form that can be used by computers, and the transformer mechanism with encoder-decoder structures, which enables context-based language understanding and text generation. Depending on the training objective, a distinction is made between basic models (e.g., for text completion), instruction models (e.g., for task processing), and chat models (e.g., for conversation).

LLMs in Everyday Scientific Life
The second part of the webinar focused on practical benefits. LLMs offer “quick wins” in many areas, whether in text summarization, code generation, creating technical documentation, or as a brainstorming aid. The potential for materials science applications is particularly evident when it comes to extracting knowledge from unstructured sources: tables, data sheets, free text, or structural diagrams can be structured, analyzed, and placed in new contexts using LLMs.

Domain-specific models such as MatSciBERT and MechGPT were also presented. These are trained for specific tasks such as named entity recognition or hypothesis formation and use corpora from materials research. They can thus help to gain insights more quickly, for example in the field of materials design or experiment planning.

Limitations and Challenges
Despite their impressive capabilities, the webinar also made it clear that LLMs are not all-rounders. Inaccuracies, bias due to training data, a lack of genuine language understanding, and technical limitations (such as the context window) remain challenges, especially in precise scientific environments. In addition, fine-tuning is complex and requires considerable resources.

The question of the use of AI agents was also discussed: When is it worthwhile to use autonomous systems with decision-making freedom – and when is it not? Clear recommendations were made here for data-intensive, dynamic work areas with high automation potential, such as research or software development. On the other hand, their use is less useful for simple or highly regulated tasks, as well as tasks that require deep domain knowledge, human intuition, or empathy.

Outlook: From Webinar to In-Depth Training
The webinar was part of a new training program offered by the DGM. The training course “mit eigenen Daten chatten” on LLMs with a stronger practical focus will be offered for the first time on 11 – 14 November 2025. The topic also ties in perfectly with the DGM conference AI MSE (18 – 19 November), which focuses on the use of AI in materials science. The training course is therefore well suited as preparation for the conference and offers the opportunity to further develop your own use cases in a targeted manner.