Fastweb is creating the first LLM (Large Language Model), i.e. a large language model for artificial intelligence, natively trained in the Italian language. The goal is to create AI systems able to understand Italian perfectly, capturing all the nuances and linguistic particularities. This would have advantages both in terms of precision in understanding requests and in terms of protecting and controlling information. For this purpose, Fastweb purchased 31 NVIDIA DGX H100the hardware that will constitute the beating heart of the supercomputer that the company controlled by the Swiss group Swisscom will use to create and power its LLM.
How Fastweb's artificial intelligence will be trained
The models on which they are based i most popular AI systems of the moment – like OpenAI's ChatGPT and Google's Gemini – are natively trained in English and, by interacting with them, you perceive their deep Anglo-Saxon imprint. This can represent a limit for those who use them by issuing commands in other languages (such as Italian), as there is not always a full understanding of the requests made on the user side. These AIs, therefore, are not always able to return outputs that are satisfactory for a culture other than the Anglo-Saxon one. With the new LLM trained natively in Italian, Fastweb wants to fill this “gap”.
Giovanni GermaniManager of Architecture & AI COE at Fastweb, reported:
Our goal is to cover all links in the artificial intelligence value chain, with the exception of chip production. We start from the NVIDIA supercomputer to get to consultancy, passing through the supply of computational power to startups and companies and the creation of tailor-made models.
In order to achieve a result that is qualitatively relevant from a linguistic point of view, Fastweb is collecting a large amount of data with which to train the model from multiple sources: from publishing to social media, via the Web and Public Administration. In this regard, Fastweb's AI manager explained:
The quality of the dataset and the balance of the various sources are fundamental to having an efficient LLM capable of truly speaking in the Italian language.
When it comes to training a large language model and the data that is used to train it, questions are often raised about the respect for user privacy and copyright regulations (both addressed at the heart of the AI Act, the world's first law regulating the use of artificial intelligence). Regarding this issue, the Fastweb manager reassured:
Unlike other models, we will comply with all European regulations on personal data protection, AI safety and copyright.
What are the potentials of an “Italian” AI
Self by 2024 the goal of will be achieved have the first models built from scratch trained natively in the Italian language, the AI landscape for us Italians could change significantly. «At that point» says Giovanni Germani «we will be able to offer companies and the Public Administration an alternative to the current AI suppliers».
This will mean being able to train models with ad hoc customizations by having your data physically remain in Italy (for the record, the data center that will host the Fastweb supercomputer used for training is located in Lombardy), with all the related advantages to the protection and control of information.
Even more interestingly, having an “Italian” artificial intelligence will contribute to the potential birth of new chatbots and virtual assistants capable of communicating naturally with those who speak Italian, having more effective, reliable and precise automatic translation systems, and generate creative content that also makes sense in our language.
