The Beta version of Emma-6 is still not satisfactory and falls into the same errors as the previous one, but we really need AI that thinks in Italian and there are other projects in development.
What is two plus two? The answer is obvious, but not for Emma, an Italian artificial intelligence in development. During a short test of the Beta version of Emma-6, the chatbot responded first 2, then 1 and finally 0, accompanying everything with a very long table of meaningless numbers. It didn’t go any better with the height of the Eiffel Tower: after indicating three different ones, he tried to correct himself by claiming that 327 plus 317 equals 327, commenting: “exactly double the width”.
Emma-5, the previous version, had gone viral for its bizarre answers, including the one according to which the iconic Parisian tower could reach a thousand meters when the wind blew strong. In short, the new version is still “hallucinating”. We are certain that sooner or later the programmers will find the solution and fix every bug (the beta versions, on the other hand, are used for this), but beyond that, Emma raises a series of interesting questions about how such a system works and what advantages it could offer if designed specifically for the Italian language.
A growing AI
Emma is developed by Egomnia, a company founded by the entrepreneur Matteo Achilli. Compared to Emma-5, which had 550 million parameters, Emma-6 has 1.55 billion. Parameters are the numerical values that a neural network changes during training to learn to recognize patterns and produce responses. Their number affects the model’s capabilities, but does not alone determine its intelligence and Egomnia claims to have trained Emma-6 on approximately 25 billion tokens, the units into which texts are divided, often corresponding to words or parts of them. They are very small in size compared to those of the large international models.
The company itself recognizes the limits of compact models, more exposed to so-called “hallucinations”, i.e. invented answers but presented as true. The objective is to progressively develop a national technology, also through the contribution of users who participate in the experimentation.
Thinking in Italian
But can an artificial intelligence really think in our language? Actually, not in the human sense of the word. Language models process mathematical representations of information and, when generating text, calculate which tokens to produce in succession. However, training one mainly in Italian can improve their ability to recognize grammatical peculiarities, idiomatic expressions and cultural references. Think of the nuances between “boh”, “mah” and “beh”, or the famous “sticazzi”, which in its Roman use mainly expresses indifference, while in northern Italy it is often used to express amazement.
Or, again, the differences between bureaucratic, professional and colloquial language.
A specialized model could interpret them more precisely, but this is not an automatic advantage: even large multilingual AIs understand Italian well and can benefit from enormous amounts of information collected in other languages. What makes the difference are the quality of the data, the training and subsequent checks.
Concrete applications
The potential benefits aren’t just about conversation. A model trained on Italian documents could be particularly useful in public administration, healthcare, businesses and professional firms. For example, it could help interpret a tax law, summarize an administrative document or explain to a citizen how to fill out a form, as long as the answers are verified and based on updated sources. Specialization could also reduce the need to adapt systems designed primarily for the English-speaking market.
Another advantage concerns size: relatively small models generally require less memory and computing power, with the ability to run on corporate servers or local devices. This can reduce costs, response times and the need to transmit information to external services. Of course, knowing the language well does not mean automatically knowing the laws or being able to provide reliable advice: in the most delicate sectors, certified data, human controls and, above all, systems capable of admitting when they do not have a certain answer are needed.
Digital sovereignty
There is also a strategic reason to develop national artificial intelligences. Today, many European companies and institutions use technologies controlled by large foreign companies, on which they depend for updates, prices, service availability and data processing. Having your own models can contribute to so-called digital sovereignty, i.e. the ability to autonomously manage the technologies fundamental to the economy and public services. A hospital, for example, may be interested in using AI installed on directly controlled infrastructure, without having to send confidential information to external suppliers.
The nationality of the developer, however, is not enough: the ownership of the systems, the licenses, the data used and the place where it is processed also count. The challenge, therefore, is not only to create a chatbot capable of speaking in our language, but to build scientific skills, computing infrastructures and tools that can be adapted to the needs of the country.
The European challenge
Emma is not the only Italian project. Minerva, developed within the research of Sapienza University of Rome with the support of Cineca infrastructures, is a family of language models trained from scratch.
Fastweb instead created FastwebMIIA, trained mainly on native language texts and intended primarily for businesses and public administrations, with computing infrastructures located in Italy. Another path is the one chosen by Almawave with Velvet 25B, a model with 25 billion parameters designed to process texts in all 24 official languages of the European Union.
Looking beyond the border, the most significant example is Mistral AI, a company founded in Paris in 2023, which has become one of the main European players in the sector thanks to multilingual models aimed at the international market. They are very different initiatives in size and objectives, but they show that the European alternative to the American giants can follow multiple paths, from linguistic specialization to large-scale technological competition. Emma, for now, struggles even with elementary operations, and this prevents us from considering her – at the moment – a reliable tool, but the future of Italian AI also depends on stumbles of this type.
