An Italian AI's Plan to Beat Silicon Valley's $100 Million Scientists

In Rome, we are digging again, but this time into the depths of neural networks. Translated , an Italian company among the most advanced in the world in AI-powered machine translation, is leading an ambitious European project to overcome the limitations of current large-scale linguistic models (LLM), which are the basis of tools such as ChatGpt , Gemini, and Claude.

Computer scientist Marco Trombetti , who founded Translated in 1999 together with linguist Isabelle Andrieu , hosted 70 European researchers at his company's headquarters who will work on DVPS , a new initiative of the Horizon Europe program that allocates around 1 billion euros per year for AI, with a focus on basic research, industrial innovation and practical applications.

The European heart of the DVPS project

Translated will coordinate the project supported by an investment of 29 million euros. “Twenty leading organizations” are participating - companies but also universities - from nine European countries. The name DVPS - a Latin acronym for Diversibus Vilis Plurima Solvo , or “through different paths, I solve multiple problems” - winks at Silicon Valley's passion for the Roman world, but the soul of this research work is entirely European.

The project started in Rome, in the Pi Campus located south of the capital, a small ecosystem of villas where startups and venture capitalists regularly meet.

The DVPS team intends to build a new AI model that can “learn from real-world interactions by combining linguistic, visual, and sensor data.”

Beyond LLMs: A New Paradigm

“Large language models have been a game changer, but we are starting to see their limits, both in terms of their architecture and in how they learn from static content, which is created by humans and only available in the digital world. To evolve, AI must interact with the real world in real time. With DVPS, we enable machines to grow by interacting with reality and exchanging knowledge with each other instantly,” said Trombetti, one of the Italian entrepreneurs closest - in spirit, attitude and personal relationships - to Silicon Valley .

In recent years, some of the symbolic names of global AI have passed through Pi Campus: from Lukasz Kaiser , one of the authors of the paper Attention Is All You Need which contributed to the birth of ChatGpt, to - just a few days ago - Jonathan Cohen , Head of AI Software at Nvidia , the company that with its graphics cards made the revolution of generative artificial intelligence possible.

The Foundations of DVPS

Translated recorded a turnover of 69.2 million euros in 2024. In 2021, it closed a round of 30 million euros led by the investment firm Ardian. The company, based in Rome, has 250 employees. Trombetti is its CEO.

With the company’s profits, Translated’s founders started Pi Campus, which they call “a venture capital firm investing in applied AI.”

The Roman campus is a place to nurture talent : an artificial intelligence school, Pi School, has been created inside, which selects the best profiles every year to invite them to solve real problems “proposed by leading companies or growing startups”.

The DVPS project, which is described as “one of the most important European investments in AI research,” is therefore built on solid foundations. The dream is to create advanced AI systems that understand context better than LLMs can.

The idea is to develop a general model that can then be “verticalized” in at least three sectors: medicine, environment and linguistic translations .

In the language space, where Translated has a strong position, DVPS will integrate visual input, spatial audio and contextual information to correctly identify the speaker and provide more accurate translations.

How to win in the AI ​​championship

The project will start with an initial experimental phase, supported by European funding of 29 million euros, of which 4 million will be allocated to computing capacity .

If the model developed in this phase gives promising results, the goal will be to raise 100 million euros within the following year to build intermediate-scale models. At that point, if the project is confirmed among the best in this new category, it will be possible to aim for an investment of one billion euros and entry into the global competition of artificial intelligence.

But how do you win AI championships?

“You start by doing research, then you build a small model, with 5-7 billion parameters,” explains Trombetti. “These models cost about a million. Those who emerge in this category do so by spending little time and resources. Among the small ones, which is the best? Even if it is inferior to a 100 billion model, it wins in its class. It’s like winning the “provincial” championships: then you move on to the “regional” championships, collect 100 million and develop a much larger model. If you win in that category too, you can invest billions and aim for the “Olympics.” We are motivated to get there, but we know that we have to go through all the steps. And we are certainly not the ones who ask for a billion without having first won the minor championships.”

The strength of Europe? Seeing the problems

What are the chances of success for DVPS? “20%,” Trombetti replies. Why so low? “We will not use the current LLM architecture,” explains the Translated CEO. “We will not follow the DeepSeek approach, which simply introduced an improvement on an existing system. Our bet is different: we focus on basic research, we want to invent new ways to tackle these problems and build a model with a completely new architecture. The probability of success is lower, but if we succeed, we will not make a 101% step forward, we will make a 200% leap.”

DVPS's mission is complex, if we consider that even on the other side of the ocean, in Silicon Valley, the best minds in the world are working to overcome the current limitations of LLMs, in particular the lack of data on which to train models. But Big Tech has billions to invest, and immense computing power. Their success rates, unlike Translated and its European fellow travelers, are much higher. So how can you compete with so many Goliaths, all focused on the same goal?

“We know these people, we talk to them,” Trombetti says. “I talked to Cohen, to Lukasz, to Ilya Sutskever, co-founder of OpenAI . They have incredible technical and theoretical understanding. But there is one thing they don’t have: they can’t understand real problems. We have been working for years with translators and linguists, people who experience the problem of language every day. And language is probably the most complex and human thing that a machine can try to understand. When a translator works with an AI system, it doesn’t accept approximations. If the machine “hallucinates” or makes a mistake, it notices it immediately. And it forces us to understand why it made a mistake. This exposes us to a level of precision and truth about the problem that many researchers don’t see. Lucasz, every time I talk to him, is surprised. Because we see problems that they don’t see. And this gives us an advantage: they have won so far by brute force, but we understand where the machine really fails.”

But according to Trombetti there is another reason why it is worth trying .

“The gap between open source models and closed models [like ChatGpt and Gemini, ed.] is narrowing,” he says. “And while funding is starting to run out, ingenuity is becoming a decisive competitive asset again. For three years, the one with the most computing power won. Then DeepSeek arrived: with ten times fewer resources and a bit of ingenuity, it took a step forward. And this opens a door for us. If you have a scientist worth a hundred million , but he hasn’t understood the problem, you’re in trouble. If you have a scientist worth a million, but he really knows the problem, he can win. This is our real advantage. And then there’s another thing: we can afford a 20% success rate. They can’t. They have to win every time. We, on the other hand, can take the risk. And that’s where something new can be born.”

Get the machine out of the computer

The goal of Translated and all the entities participating in the DVPS is to “get the machine out of the computer and start making it interact with the physical world”.

“To do this,” says Trombetti, “it must be able to process all the sensors. Let’s think about an autonomous car: it has cameras, radar, lidars, distance meters. Today, all this data must be pre-processed by hand. We must explain to the machine what each input means, simplifying a lot. But if we want to give it a thousand sensors and have it manage complex information – like in medicine, where the data is heterogeneous – we can no longer code everything by hand. The machine must be able to read the information directly, byte by byte, and assign a meaning to it autonomously. This requires parallelism, byte-level learning, but above all a paradigm shift: up to now we have trained machines only with historical data. But humans learn above all by interacting with the world, by having experiences: learning by doing. For this reason we must “open the door”: let the machine out, have it have experiences, and develop an architecture that allows it to self-train while it acts. This is the next step.”

AGI needs a change

A fundamental step, among other things, to arrive at a superior artificial intelligence, which many call “general” (AGI), and which could one day surpass human cognitive capabilities in many areas.

“Today I am certain that with the current architecture we cannot get to AGI. There is too much to learn in the physical world and in the future to really generalise. The current approach is not enough”.

The new model will be “open”, but not too much

The efforts of DVPS, says Trombetti, will be totally open as far as research papers are concerned. While the first model created will be “ open weight ”: it means that its weights (i.e. the parameters learned during training) are publicly accessible.

And why won't it be an open source AI?

"We have to be careful about how much to share: we haven't decided yet - says Trombetti -. The biggest risk is that, once again, in Europe a lot of research is done and money is invested, while then it is the American researchers who take the results, publish first and propose alternative projects, reaping the benefits. For this reason we have to carefully evaluate where to draw the limit. It is a strategic choice: if the project goes very well, we can afford to close ourselves off a little more; if instead the progress is slower, then it is better to be more open, more collaborative and look for new opportunities".