Artificial intelligence beyond language models: Italian-led EU project kicks off

'Diversibus viis plurima solvo' (Dvps), 'through different paths, I solve multiple problems': the innovative project, one of the most ambitious funded by Horizon Europe, which aims to study the next generation of artificial intelligence systems, uses Latin. The aim is to overcome the limits of large language models, ChatGPT-style models, with new models capable of learning from interaction with the real world, combining linguistic and textual data with visual and sensory data, in real time.

Leading the European Dvps project is Translated, an Italian company operating between Rome and Silicon Valley, a pioneer in the use of AI in the field of linguistic solutions in the service of translation, which coordinates a team of seventy researchers, around thirty of whom are Italian, from around twenty leading organisations in nine European countries, supported by an initial funding of EUR 29 million, one of the largest in the EU in the field of artificial intelligence.

Unlike current systems, which learn from representations of the world through text, images and video in a sequential manner, next-generation models are expected to be more like the human approach, based on empirical learning through direct interaction with the physical world, with a multi-sensory process in which all senses process information simultaneously. By integrating linguistic, visual and sensor data, they develop deeper contextual awareness, expanding human capabilities in domains where confidence, accuracy and adaptability are crucial.

"Large language models have marked a turning point, but are beginning to show their limits, with human-guided supervised learning of digital content, based on an architecture that leads machines to follow a sequential process,' explains Translated co-founder and CEO Marco Trombetti. The Dvps project aims at overcoming these limits, on the one hand by providing machines with the ability to process all information in a parallel, simultaneous manner, as all intelligent biological forms do, and then reshaping the learning model on the basis of experience and interaction with the rest of the world, as we humans do on the other hand'.

Today, sensors to provide real-time information about the world around us already exist in abundance, it is a matter of creating models capable of simultaneously and multimodally interpreting all the signals they receive. The project, involving top research institutes such as Oxford University, Alan Turing Institute, École Polytechnique Fédérale