Transformer-basierte Large-Language-Models verstehen
This course provides a conceptual understanding of modern large language models. It focuses on the Transformer architecture and its key components, such as attention and positional encoding. With this knowledge, your will gain insight into the inner workings of current models from OpenAI and DeepSeek and understand how they function.
Overview
Large language models are now central to many AI applications, from integrated chatbots to AI agents. But how do these systems actually work?
In this online course, you’ll receive a clear introduction to the fundamentals of modern language models. The focus is on the Transformer architecture, which forms the basis of today’s large language models. Your will become familiar with key concepts such as attention, positional encoding, and the modular structure of the architecture.
The course emphasizes a clear conceptual understanding and deliberately avoids programming and complex mathematical derivations. The most important ideas are explained step by step using clear examples.
Building on this foundation, we will then examine current models from OpenAI and DeepSeek, looking at how they are structured, how they have evolved in recent years, and how they differ from one another.
Which topics will be covered?
- Fundamentals of modern large language models and their role in current AI applications
- Structure and functioning of the Transformer architecture
- Structure and development of modern model families such as GPT and DeepSeek
What will I achieve?
By the end of the course, learners will be able to:
- explain how large language models operate and what they are utilized for.
- describe the key components of the Transformer architecture.
- understand the development and structure of current model families such as GPT or DeepSeek.
- compare different LLM architectures based on fundamental criteria.
Which prerequisites do I need to fulfill?
A basic understanding of neural networks or an introductory course on topics such as tokenization, embeddings, or chatbots is helpful but not required
The underlying project was funded by the Federal Ministry of Research, Technology and Space under the funding code “KI-Servicezentrum Berlin-Brandenburg” 16IS22092. Responsibility for the content of this page remains with the author.
