There is an artificial intelligence that speaks Italian. Think in Italian. Composes sentences, elaborates answers, reasons in Italian. It’s still little known, but it’s there. It is coming to life in the corridors of university departments, in the research centers of companies, of startups, in virtuous examples of public administration. A movement from below. Lively. Made up of women and men with cutting-edge skills. Working on tools capable of providing Italy with the assets necessary to compete in the challenge posed by the new generative AI technologies.
The government’s priority. Butti: “We need an Italian ChatGpt”
Artificial Intelligence has become a priority for the government. The Prime Minister, Giorgia Meloni, announced investments of one billion to develop Italian projects. The Undersecretary of the Prime Minister’s Office, Alessio Butti, reiterated the importance for Italy of having an Italian ChatGpt, to meet the security and privacy needs requested by Italian companies. A real need. Heard by PA and companies. Which is drawing new trajectories in the evolution of innovation: no longer subjected to the large international players, but created in various countries to respond to their own needs while respecting their own culture and traditions.
“There are at least a couple of reasons to prefer an Italian generative AI. One is data control and sovereignty, which is fundamental for companies. The other is the performance on the language, given that today the chatbots in circulation are trained in English, so they think in English, and then translate”. Giovanni Germani is a data scientist at Fastweb. The company is developing generative AI in Italian, training a model from scratch. The question of language is not secondary. Because a generative AI like Chatgpt is the result of long training done by these machines on written texts. 98% of these texts are in English. Italian and other languages have a few zeros left. If an AI is trained in English, it thinks in English. “Training a generative AI from scratch in Italian is the only way to prevent it from thinking in English. And thinking in Italian is important for a company that perhaps wants a chatbot capable of responding to Italian customers by understanding exactly what they want to say, responding with correct and precise texts”, continues Germani to La Stampa.
Fastweb is aiming for an Italian AI. The role of Cineca for research
Fastweb will soon inaugurate a data center in Lombardy. A supercomputer capable of training the first LLM in Italian will be installed there. He will be given texts only in Italian: texts from Wikipedia, journalistic sources, social media, seeking agreements with the various bodies. But it is not the only private company to move into the LLM field. iGenius, founded in Milan in 2016 by Uljan Sharka, last month announced the birth of a project to create an LLM in Italian developed with Cineca, the inter-university consortium made up of universities and public bodies. It’s not a casuality. In Bologna Cineca has one of the most powerful computers in the world. His name is Leonardo. To get an idea of how powerful it is, just think that in one hour it is able to make calculations that would take the most advanced laptop on the market 920 years of work. The project is co-financed by the European Commission and the Ministry of University and Research and is allowing Italy to scale up artificial intelligence at a good pace.
“The importance of this supercomputer for Italy is enormous,” explains Alessandra Poggiani, general director of Cineca, to La Stampa. “Especially because it is public and free. All Italian universities and research groups have free access to Leonardo. The European Commission has ascertained that Italy is the country where the national scientific community uses supercomputing the most for its research,” she adds. Leonardo was used by Mistral, the French generative AI startup that went from zero to 2 billion in valuation in 4 months. And Leonardo is used by many Italian research projects on generative AI. Projects that can become businesses.
University projects: Llamantino (Bari) Cerbero (Pisa)
Llamantino is one of them. It was developed by a research group from the Computer Science department of the University of Bari. “We wanted to create an Italian alternative to ChatGpt. Indeed, something that was in contrast to ChatGpt, because we believe in the open model of artificial intelligence, not a closed one like that of OpenAI”. Lucia Siciliani is a researcher. You wrote Llamantino’s project with your colleagues, which was then trained on Italian texts thanks to the Leonardo supercomputer. “Today Llamantino is a research project, but we know that it can be applicable to industrial settings,” she adds.
The Llamantino team: Lucia Siciliani, Pierpaolo Basile, Giuseppe Fiameni, Elio Musacchio, Giovanni Semeraro
From Bari to Tuscany. Cerberus is an LLM project of the University of Pisa. It was created by a research group led by Federico Gatolo. “We recently released our second version. It’s a Chatgpt, but Italian. Hopefully it will be as good as Mistral, but with a completely open license. Cerberus has the characteristic of being trained on a high-quality dataset. Therefore the choice of the Italian texts on which he was trained had to respect very high parameters from a linguistic point of view”. Cerberus is available online. Already compared to ChatGpt, greater attention to language and precision in syntax seems evident.
Federico Gatolo and Mario Cimino, creators of Cerberus from the University of Pisa
At La Sapienza Faun, Camoscio and Dante. And Zefiro, born in PagoPa
In Rome, La Sapienza University has several projects in development: Faun, Camoscio and Dante. Projects with their own specificity, but basically they answer the same question: “Training in Italian is more sensitive to small variations in the language. We have conducted studies and know that English LLMs do not have the right sensitivity to understand potentially offensive sentences in other languages. If the understanding of a term is wrong, its communication is also wrong”, explains Fabrizio Silvesti, professor of AI at the department of computer engineering. Faun and Dante, the latest to arrive, came to life there.
Fabrizio Silvestri’s team: from left Cesare Campagnano, Fabrizio Silvestri, Andrea Bacciu and Giovanni Trappolini.
It uses seven billion parameters and is based on Mistral as its base model. But there is also a piece of Pa who is working on his LLM. PagoPA in particular, where Alessandro Ercolani, developer, put on Zefiro: “I created it during the last Christmas holidays, we would like to use it in PagoPa so as not to rely on foreign big players. He converses with customers, he does it in Italian, he is trained in Italian. But what is most important for us is that it embraces the principles we have set ourselves as a company: adopting open source as a philosophy. We believe that the development of these technologies should be open and shared with everyone. Only in this way can it do the good of everyone.” A movement from below. With its own vision of the world and of artificial intelligence. Italian.