A new beta version of an Italian artificial intelligence named Emma-6 has revealed significant limitations during recent testing. Developed by Egomnia, a company founded by entrepreneur Matteo Achilli, the chatbot failed to answer simple questions correctly. When asked what two plus two equals, the model provided conflicting answers of 2, 1, and 0, accompanied by a nonsensical table of numbers. Similarly, when queried about the height of the Eiffel Tower, Emma-6 offered three different figures before attempting a correction that claimed 327 plus 317 equals 327, describing it as exactly double the width. These errors mirror issues seen in the previous version, Emma-5, which went viral for claiming the Parisian landmark could reach one thousand meters in strong winds.
Emma-6 represents a technical upgrade from its predecessor, increasing its parameter count from 550 million to 1.55 billion. The company states that the model was trained on approximately 25 billion tokens, which are text units often corresponding to words or parts of words. While these dimensions are considered small compared to major international models, Egomnia acknowledges that compact models are more susceptible to hallucinations, where the AI generates fabricated information presented as fact. The company views these beta tests as essential for identifying bugs and aims to progressively develop a national technology with user contributions.
The concept of an AI "thinking" in Italian does not imply human-like cognition but rather the processing of mathematical representations optimized for the language. Training a model primarily on Italian data can enhance its ability to recognize specific grammatical nuances, idiomatic expressions, and cultural references. For instance, a specialized model could better distinguish between subtle differences in words like "boh," "mah," and "beh," or understand regional variations in terms like "sticazzi," which conveys indifference in Rome but surprise in Northern Italy. However, this specialization is not an automatic advantage, as large multilingual models also possess strong capabilities in Italian due to the vast amount of data available in other languages.
Beyond conversational abilities, a model trained on Italian documents could offer practical benefits for public administration, healthcare, businesses, and professional studies. Potential applications include interpreting tax regulations, summarizing administrative documents, or guiding citizens through bureaucratic procedures, provided the responses are verified against updated sources. Furthermore, smaller models generally require less memory and computing power, allowing them to run on local servers or devices. This setup can reduce costs and response times while minimizing the need to transmit sensitive data to external services.
The development of national AI models is also driven by strategic reasons related to digital sovereignty. Currently, many European entities rely on technologies controlled by foreign corporations, creating dependencies regarding updates, pricing, and data handling. Possessing proprietary models allows institutions, such as hospitals, to manage critical infrastructure directly without sending confidential information abroad. However, the nationality of the developer is insufficient; ownership, licensing, data sources, and processing locations are equally critical factors in achieving true autonomy.
Emma is part of a broader European effort to create alternatives to American tech giants. Other Italian initiatives include Minerva, developed by Sapienza University of Rome with support from Cineca, and FastwebMIIA by Fastweb, which focuses on enterprises and public administration using local infrastructure. Almawave has also introduced Velvet 25B, a model designed for all 24 official EU languages. Internationally, France's Mistral AI serves as a prominent example of European success in the sector. While Emma-6 currently struggles with elementary operations and cannot yet be considered a reliable tool, these challenges are viewed as part of the necessary learning process for the future of Italian artificial intelligence.
Sources: Focus.it.