A historic collaboration: Switzerland’s own version of ChatGPT

As easy to use as ChatGPT, but with answers that are traceable and ensure confidentiality. This is what the founders of AlpineAI are promising with their offer for companies in Switzerland: SwissGPT is being designed to meet the highest international requirements and, in addition to ethical and legal standards, also to take account of the specific needs of multilingual European companies.

At present, the vast majority of so-called large language models (LLMs) such as ChatGPT are being developed in the US and China. The Swiss alliance, which is made up of researchers from leading AI labs at ETH Zurich, the ZHAW, the Lucerne University of Applied Sciences and Arts and other universities as well as the private sector, wants to improve Europe’s position in this promising market. The aim of SwissGPT is to pave the way for a distinct Swiss solution that is compatible with the EU. LLMs are mathematical models that predict and generate sequences of words according to the laws of probability. 

“While LLMs are nothing new from a scientific standpoint, they make artificial intelligence useful, and they do so for a wide range of user groups,” said Thilo Stadelmann, the co-founder of AlpineAI and Head of the ZHAW Centre for Artificial Intelligence, at a recent media conference. The LLM behind SwissGPT is based on the Falcon model, whose roots can be traced back to the Swiss Federal Institute of Technology Lausanne (EPFL). This model features among the leaders in the most important ranking lists. Compared to the LLMs evaluated as part of a Stanford study, during which the models’ output, copyright compliance and environmental footprint were assessed, it performs excellently.

In a first step, AlpineAI will provide companies with secure access to the most widely used LLMs, such as ChatGPT, which offers the anonymisation of prompts with heightened privacy requirements. “You can then have the model summarise a client’s history or quarterly figures, for example, without the prompt or the answer subsequently floating around the Internet,” says Thilo Stadelmann. Hosted in Switzerland, the core product SwissGPT is designed to be able to access sensitive, internal company data.

One possible direct application of SwissGPT is said to be the AI-supported answering of e-mails, with users being able to choose what to do with a message – either ignore it, reply to it with a standard sentence or send a detailed response. The model then suggests alternative content. The aim is for further products to also be used by the scientific community, in particular.

According to CEO Pascal Kaufmann, the financing of AlpineAI has been secured and the first client enquiries have already been received. The AI start-up is now looking for specialists in areas such as LLMs and machine learning. Finding these experts is considered to be the big challenge.