Humboldt-Universität zu Berlin - Artificial Intelligence at HU Berlin

AI at HU Berlin

Humboldt-Universität zu Berlin promotes the use of AI in research, teaching and administration. The Computer and Media Service (CMS) is building infrastructure based on open source products that can be used in a low-threshold and data protection-compliant manner. The services currently consist of Large Language Models, the High Performance Computing Cluster and a JupyterHub. The AI guideline will be gradually developed towards an AI policy for the HU.

Generative AI tools[1], such as OpenAI's ChatGPT or Google's Gemini, are developing rapidly and are increasingly being used for more and more purposes - including in a professional context. As always with new technologies, we as individuals and as an institution need to consider how we can integrate these developments constructively and securely into our everyday working lives. The benefits, consequences and risks for use in research, teaching and administration are only partially foreseeable and it is to be expected that the speed of developments in the field of AI applications will tend to increase in the near future. In addition, as a university, we have to take into account the requirements of data protection and data security and fulfill an educational task.


Initial measures at the HU

  • Provision of data protection-compliant large language models (Large Language Models, LLM) for use in research, teaching and administration
  • HPC@HU for research (High Performance Computing Cluster)
  • Development of a JupyterHub and its integration into the HU's digital teaching and learning landscape (HDL3) incl. interface to the HU Moodle
  • Development of AI guidelines that will be developed towards an AI policy for the HU.

Furthermore, the establishment of experimental environments for RAG[2] and LoRA[3] is being planned.


Data protection-compliant AI Large Language Models (LLM)

CMS hosts several LLMs for general use at the HU. They can only be accessed in the HU network or via VPN.

In contrast to other commercial services, prompts are not saved after the chat is closed and are not used for further development of the model. Personal information or personal data therefore remains confidential, as no one outside the current chat has access to it. Logging does not take place.

Here you will find further information about the Large Language Models.


HPC@HU - High Performance Computing Cluster for research

With HPC @ HU, the CMS of the HU is launching an innovative high-performance computing offer to make hardware and software for high-performance computing easily accessible to all researchers, scientists and teachers at the HU in a resource-saving manner.

Interested parties can currently apply for test access to the "Azimuth" platform. Please contact for this. HPC@HU is scheduled to go live in 2024. HPC@HU is currently only accessible from the HU VPN.


JupyterHub in teaching (JH)

JupyterHub is an open source tool that allows students, teachers and researchers to run interactive Jupyter notebooks on a powerful HU environment without additional installations for prototyping, data analysis and data visualization in the browser. To use the JupyterHub, you need an HU account. The JupyterHub is freely accessible from the web.




You can find more information on the individual infrastructure offerings under "More about LLM, HPC & JH".


[1] Artificial intelligence (AI) refers to algorithms that imitate human cognitive abilities. Generative AI in this sense refers to the use of AI to analyze patterns and structures in existing data in order to generate new content in various formats such as text, music, computer code, images or videos.

[2] RAG, or Retrieval Augmented Generation, is a technique that combines the capabilities of a pre-trained large language model with an external data source. This approach combines the generative power of LLMs like GPT-3 or GPT-4 with the precision of specialized data search mechanisms, resulting in a system that can offer nuanced responses. (,3/14/2024)

[3] LoRA (Low-Rank Adaptation of Large Language Models) is a popular and lightweight training technique that significantly reduces the number of trainable parameters. It works by inserting a smaller number of new weights into the model and only these are trained. This makes training with LoRA much faster, memory-efficient, and produces smaller model weights (a few hundred MBs), which are easier to store and share. (