Institutions from 12 countries, in a Chilean artificial intelligence-driven initiative, are working together to develop LatamGPT, the first large-scale language model trained with representative data from Latin America and the Caribbean.
This is a collaborative, free and open source project, led by the National Center for Artificial Intelligence (CENIA) and supported by the Ministry of Science, Technology, Innovation and Knowledge of Chile, which seeks to generate its own capabilities in artificial intelligence with a regional approach.
Through a network of universities, research centers, public agencies and experts from across the region, LatamGPT is being built collectively to respond to a common challenge: to reduce technological dependence on models developed mainly in the northern hemisphere, in contexts that are alien to the Latin American reality, and to move towards a contextualized AI at the service of people.
"It is a collective and open effort, which is a reflection of what we hope this language model will be. When we talk about AI, it has to project the world we are, its diversity. And in the case of Latin America, not only to speak Spanish or Portuguese, but to understand our idiosyncrasy, to contribute from the culture and from the vision of the world," said the Chilean Minister of Science, Aisén Etcheverry.
LatamGPT is trained with data from the region itself, collected by local institutions. This allows for a more accurate representation of the cultural and social particularities of Latin America and the Caribbean.
"One of the problems we have in Latin America and the Caribbean is that the LLMs (Large Language Model) we use are quite limited in the knowledge they have about the region. It is important that here we can develop capabilities to have some independence and make decisions on how this technology impacts society. So far we do not have a regional language model," explained Alvaro Soto, director of CENIA.
One of the project's priorities is to promote the preservation of native languages, incorporating data in indigenous languages and local dialects. For example, there are already pilots with the Rapa Nui and Mapudungún languages.
The model is expected to have concrete applications in key areas such as education, health, public services or economic development, with versions adapted to different needs and technical capabilities.
The model is trained on a supercomputing infrastructure at centers such as the Universidad de Tarapacá in Chile.
