There’s something very Italian about an AI introduced in huge words and then stumbled upon Super Mario. A kind of ribbon cutting with the municipal band, the smiling mayor, the promise of the future and, two minutes later, the microphone whistling. Emma, the Italian AI developed by Egomniaarrived online with a high ambition: to talk about technological sovereignty, linguistic models in Italian, independence from large foreign suppliers. Then users started asking her normal questions. And the screenshot festival has begun.
The full name of the project refers to a family of Italian linguistic models. On the technical data sheet of Emma-5 published on Hugging Face, the model is described as an experimental decoder-only LLM developed by Egomnia, designed for the Italian language, with moderate contextual understanding and generation of medium complexity content. The same documentation also clarifies the boundaries: no critical applications, no advanced multi-step reasoning, no use in sensitive medical, legal or financial fields. Simply translated: Emma was born small, light, experimental. The problem comes when a national event stage is built around a small thing.
The datasheet already says so
To make the story more interesting there is a detail that tends to disappear in memes: Emma-5 is not presented in its technical documentation as a model capable of competing with ChatGPT, Gemini or Claude. The card talks about 2,048 context tokens, SentencePiece BPE tokenizer, GPT decoder-only architecture, model exported to ONNX and weight of 2.46 GB. The indicated pretraining corpus is approximately 54.36 GB raw, equal to approximately 10.8 billion tokens. They are numbers of a compact model, suitable for light contexts and small corporate chatbots, rather than a generalist assistant to be put in front of the whole Internet with the tricolor hat on his head.
The same page indicates among the limitations the possibility of producing imprecise or incomplete content, sensitivity to ambiguous or long prompts and reasoning capabilities that are still limited compared to large models. In short, Emma had already written the leaflet. Except that the public read the packaging.
And there were much stronger words on the packaging. In the poster shown in the model sheet we talk about Italian technological sovereigntyof artificial intelligence as a critical infrastructure for the economic, cultural and democratic future of a nation, of models developed elsewhere according to logics and priorities not always aligned with the Italian and European context. Serious concepts, even shareable. Precisely for this reason they require suitable products, tests and precautions.
The meme runs faster than the model
@minnovo_media Since yesterday Italy has its own artificial intelligence. Her name is Emma, and when asked which stadium Milan and Inter play in she replied that Milan is in Serie A and Inter in Serie B. ⠀ It was created by Egomnia, a SME listed on Piazza Affari led by Matteo Achilli, who gave the model the name of his daughter. Emma is presented with a manifesto on «Italian technological sovereignty»: a model trained only in Italian, hosted in Italy, compliant with GDPR and AI Act. ⠀ The problem is that Emma-5 has 550 million parameters and a 2,048 token context window. To clarify: serious models start from a few billion parameters and up. So, in addition to the stadium, Emma claims that the mayor of Milan is Matteo Achilli, that Cesare Pavese died in the Appio district and that 10×10 is 4. ⠀ Online immediately became meme material, and in part it fits. But there is a less funny point. Emma is not “broken”: she does exactly what a model of that size can do. The real mistake is selling it as the technological redemption of a country. ⠀ And here comes the bitter part. While we laugh at Emma, we continue to depend on American and Chinese models, because in Europe there are few real alternatives. The paradox is that in Italy something serious already exists: Minerva della Sapienza, Model Italy by iGenius. Emma’s risk isn’t making us laugh. It’s making millions of people believe that “Italian AI” means this, discrediting an idea that deserved to be taken seriously.
♬ original audio – minnovo_media
The web, of course, made the web. He took the wrong answers, packaged them, relaunched them, commented on them, transformed them into jokes. In the space of a few hours, Emma has become the “Italian AI” that gets video games, elementary calculations and the most well-known cultural references wrong. At a certain point the topic stopped being technical and became almost anthropological: how much can we load a still immature technology with national rhetoric before the internet dismantles it with screenshots? But above all, how dangerous can it become?
@ballaranigianluigi Emma AI: why Italian artificial intelligence has become a meme (and what it really teaches us) In recent days Emma, the new artificial intelligence model developed by Egomnia, has ended up at the center of an avalanche of memes after some responses went viral. The temptation is to stop at the jokes. In my opinion it would be a mistake. This story reminds us of something important: building a competitive language model is incredibly difficult. It is not enough to say “we have an Italian AI” because the comparison automatically becomes with ChatGPT, Claude and Gemini. Emma is still an experimental project, with well-defined technical limits. The problem isn’t that a prototype makes mistakes: it’s normal. The problem arises when the public perceives it as a mature product and puts it on the same level as the best models in the world. Italy has high-level researchers and engineers. For this reason, it would be a shame if this story sent the wrong message across. It makes us understand how complex it is to build reliable artificial intelligence. How do you see it: was it mainly a technical or communication problem? #AI #ArtificialIntelligence #Emma #Egomnia #ChatGPT Claude Gemini LLM MachineLearning Tech Innovation Italy Startup Technology
♬ original sound – Gigi B
Meanwhile, the official website has changed its tone. Emma’s page now thanks for the more 60,000 chats with Emma-5 and explains that the release was for exploratory and experimental purposes. It adds that the usage that emerged was not fully in line with the intended objectives of the test and that availability was temporarily suspended. In its place appears an application to become a tester of Emma-6.
This step is almost more important than funny answers about video games. He says the public launch produced data, reactions, chaos, useful material and a lot of irony. He also says that an experimental model, let loose in front of an audience accustomed to interrogating enormous systems, is judged as if it were already a mature product. And there the distance becomes brutal.
Sovereignty needs fewer trumpets
The theme ofItalian artificial intelligence deserves respect. Having skills, models, datasets, infrastructures and companies capable of working on AI in Italy and Europe is a concrete, industrial, cultural issue. Reducing everything to laughter would be convenient, and also a little lazy. But there is another mistake, equal and opposite: using big words to cover the limits of a still fragile instrument.
Emma can also be a useful experiment. It can be used to collect data, gain experience, build internal skills, and prepare more solid subsequent models. The documentation, in various places, clearly suggests this. The short circuit arises when an experiment is perceived as the symbol of national recovery. At that point every wrong answer weighs double. A mistake about Mario becomes a meme. A mistake about Zelda becomes a joke. An error on the video game crisis becomes the perfect summary of an overly charged debut.
Italian AI will need many things: research, funds, patience, independent evaluations, transparency on limits, realistic use cases, fewer slogans hanging over models that are still being tested. Emma, for now, has done only one thing with great effectiveness: to remind us that the technological future is not declared in a manifesto. Let’s try it. And when Super Mario makes a mistake, it shows immediately.
