Apertus critics “misunderstand” Swiss AI model

Apertus performs ridiculously? 'It was misinterpreted' Keystone-SDA

Critics of the Swiss artificial intelligence model Apertus have "misinterpreted" its objectives, according to an eminent computer scientist.

+Get the most important news from Switzerland in your inbox

Developed by the Federal Institutes of Technology in Lausanne and Zurich, as well as the Swiss Centre for Scientific Computing (CSCS) in Lugano, Apertus is an open-source, multilingual and 100% Swiss AI model. Presented on September 2, it has aroused enormous interest, but several experts have also criticised its performance.

+ Fact and fiction about the Swiss AI model Apertus

“Above all, I note that many people have misinterpreted the primary objective of this project,” Philippe Cudré-Mauroux, professor of computer science at the University of Fribourg said in an interview in La Liberté. “The fact that several sites use the Apertus models to create a conversational agent has created confusion.”

“Contrary to what these initiatives suggest, the Federal Institutes of Technology and the CSCS have not created a Swiss ChatGPT capable of answering any kind of question. It is therefore crucial to distinguish Apertus from its derivatives in the form of chatbots.”

Cudré-Mauroux said Apertus should be distinguished from chat LLMs like ChatGPT. “It is a generic language model (LLM) that needs to be adapted to perform specific tasks. It is intended for researchers, companies or administrations,” he said. “Chatbots have been created from Apertus, but these remain secondary.”

Transparent and ethical

Currently, any interested person or entity can download Apertus from the Hugging Face platform, which hosts other open artificial intelligence models. Different versions are currently available: some with eight billion parameters learnt during training and some larger ones containing 70 billion.

But GPT 3, the previous version of ChatGPT, had 175 billion parameters. “With 70 billion Apertus is already a good base that can allow, after adaptation, the automation of many administrative tasks or the creation of other AIs,” said Cudré-Mauroux. “And let’s not forget: its main advantage is its openness and ethics. To my knowledge, we are dealing with one of the first 100% transparent models’.

“Most American and Chinese AIs are very opaque,” Cudré-Mauroux added, both in terms of the data used by their teams to train, and in relation to the orientation of results on ethical, social or political grounds. “The fact that Apertus makes its entire learning method public is reassuring for researchers and local companies that are attentive to their data and use of AI”.

More funding needed

“In terms of the amount of data used for training and the size of the model, Apertus cannot compete with the LLMs of OpenAI or other large companies,” the expert points out. Its results will therefore be less accurate and some applications may lead to hallucinations. It is necessary to work on new versions.

“But I would like to remind you of one thing: Apertus is not a finished product to be used as it is. It is an interesting first step, light years away from the ‘Far West’ of international big tech.”

“If Switzerland stays put, our self-determination is clearly threatened in the long term’.

“Currently, Switzerland allocates little funding to the development of artificial intelligence, unlike most of our industrialised neighbours,” Cudré-Mauroux concludes.

What is your opinion? Join the debate:

Translated from Italian by DeepL/mga

How we work

We select the most relevant news for an international audience and use automatic translation tools to translate them into English. A journalist then reviews the translation for clarity and accuracy before publication.  

Providing you with automatically translated news gives us the time to write more in-depth articles. The news stories we select have been written and carefully fact-checked by an external editorial team from news agencies such as Bloomberg or Keystone.

If you have any questions about how we work, write to us at [e-mail]