Regensburg 2025 – scientific programme

Parts | Days | Selection | Search | Updates | Downloads | Help

AKPIK: Arbeitskreis Physik, moderne Informationstechnologie und Künstliche Intelligenz

AKPIK 4: Focus: Applications of Deep Neural Networks

AKPIK 4.1: Invited Talk

Tuesday, March 18, 2025, 14:00–14:30, H5

The Scaling of Intelligence: From Transformers to Agentic AI — •Oliver Mey — Vodafone Tech Innovation Center, Dresden, Germany

The 2024 Nobel Prize in Physics recognized fundamental contributions to artificial intelligence and highlighted its profound impact on all disciplines, including physics. Generative AI has become a central tool in science and beyond, and understanding its underlying principles, the forces driving its rapid progress, and its emerging applications opens the door to new scientific breakthroughs and transformative innovations. We trace the evolution from Moore's Law to the scaling principles that enable today's large-scale AI models. At the heart of this transformation lies the Transformer architecture, the foundation of large-scale language models (LLMs) that generate coherent, context-aware text. These models are evolving into multimodal systems that seamlessly integrate text, images and other data types, greatly expanding their capabilities. Retrieval-augmented generation (RAG) extends LLMs with dynamic memory, enabling access to external information. In parallel, new concepts for task-dependent scaling of computations allow LLMs to distribute computational effort based on task complexity, increasing their efficiency in reasoning and adaptive problem solving. These advances pave the way for AI systems that act as collaborative agents and are capable of context-aware, goal-oriented interactions. In this talk, I will provide an overview of these developments and discuss them in the context of their broader implications, setting the stage for further specialized discussions.

Keywords: Generative AI; Large Language Models (LLMs); Retrieval-Augmented Generation (RAG); Agentic AI; Reasoning Models